Mercurial > hg > graal-compiler
annotate src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp @ 1837:c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
Summary: Inserted missing yield(check)s in closures used during the work-stealing phase of parallel concurrent marking, a missing synchronous yield-request in the cms perm gen allocation path, and a terminator-terminator for the offer_termination invocation that monitors the yield status of the concurrent marking task. Elaborated some documentation comments and made some task queue termination loop flags configurable at start-up to aid debugging in the field.
Reviewed-by: jmasa, johnc, poonam
author | ysr |
---|---|
date | Wed, 29 Sep 2010 16:17:02 -0700 |
parents | 894b1d7c7e01 |
children | a7214d79fcf1 |
rev | line source |
---|---|
0 | 1 /* |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1521
diff
changeset
|
2 * Copyright (c) 2001, 2010, Oracle and/or its affiliates. All rights reserved. |
0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1521
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1521
diff
changeset
|
20 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1521
diff
changeset
|
21 * questions. |
0 | 22 * |
23 */ | |
24 | |
25 # include "incls/_precompiled.incl" | |
26 # include "incls/_concurrentMarkSweepGeneration.cpp.incl" | |
27 | |
28 // statics | |
29 CMSCollector* ConcurrentMarkSweepGeneration::_collector = NULL; | |
30 bool CMSCollector::_full_gc_requested = false; | |
31 | |
32 ////////////////////////////////////////////////////////////////// | |
33 // In support of CMS/VM thread synchronization | |
34 ////////////////////////////////////////////////////////////////// | |
35 // We split use of the CGC_lock into 2 "levels". | |
36 // The low-level locking is of the usual CGC_lock monitor. We introduce | |
37 // a higher level "token" (hereafter "CMS token") built on top of the | |
38 // low level monitor (hereafter "CGC lock"). | |
39 // The token-passing protocol gives priority to the VM thread. The | |
40 // CMS-lock doesn't provide any fairness guarantees, but clients | |
41 // should ensure that it is only held for very short, bounded | |
42 // durations. | |
43 // | |
44 // When either of the CMS thread or the VM thread is involved in | |
45 // collection operations during which it does not want the other | |
46 // thread to interfere, it obtains the CMS token. | |
47 // | |
48 // If either thread tries to get the token while the other has | |
49 // it, that thread waits. However, if the VM thread and CMS thread | |
50 // both want the token, then the VM thread gets priority while the | |
51 // CMS thread waits. This ensures, for instance, that the "concurrent" | |
52 // phases of the CMS thread's work do not block out the VM thread | |
53 // for long periods of time as the CMS thread continues to hog | |
54 // the token. (See bug 4616232). | |
55 // | |
56 // The baton-passing functions are, however, controlled by the | |
57 // flags _foregroundGCShouldWait and _foregroundGCIsActive, | |
58 // and here the low-level CMS lock, not the high level token, | |
59 // ensures mutual exclusion. | |
60 // | |
61 // Two important conditions that we have to satisfy: | |
62 // 1. if a thread does a low-level wait on the CMS lock, then it | |
63 // relinquishes the CMS token if it were holding that token | |
64 // when it acquired the low-level CMS lock. | |
65 // 2. any low-level notifications on the low-level lock | |
66 // should only be sent when a thread has relinquished the token. | |
67 // | |
68 // In the absence of either property, we'd have potential deadlock. | |
69 // | |
70 // We protect each of the CMS (concurrent and sequential) phases | |
71 // with the CMS _token_, not the CMS _lock_. | |
72 // | |
73 // The only code protected by CMS lock is the token acquisition code | |
74 // itself, see ConcurrentMarkSweepThread::[de]synchronize(), and the | |
75 // baton-passing code. | |
76 // | |
77 // Unfortunately, i couldn't come up with a good abstraction to factor and | |
78 // hide the naked CGC_lock manipulation in the baton-passing code | |
79 // further below. That's something we should try to do. Also, the proof | |
80 // of correctness of this 2-level locking scheme is far from obvious, | |
81 // and potentially quite slippery. We have an uneasy supsicion, for instance, | |
82 // that there may be a theoretical possibility of delay/starvation in the | |
83 // low-level lock/wait/notify scheme used for the baton-passing because of | |
84 // potential intereference with the priority scheme embodied in the | |
85 // CMS-token-passing protocol. See related comments at a CGC_lock->wait() | |
86 // invocation further below and marked with "XXX 20011219YSR". | |
87 // Indeed, as we note elsewhere, this may become yet more slippery | |
88 // in the presence of multiple CMS and/or multiple VM threads. XXX | |
89 | |
90 class CMSTokenSync: public StackObj { | |
91 private: | |
92 bool _is_cms_thread; | |
93 public: | |
94 CMSTokenSync(bool is_cms_thread): | |
95 _is_cms_thread(is_cms_thread) { | |
96 assert(is_cms_thread == Thread::current()->is_ConcurrentGC_thread(), | |
97 "Incorrect argument to constructor"); | |
98 ConcurrentMarkSweepThread::synchronize(_is_cms_thread); | |
99 } | |
100 | |
101 ~CMSTokenSync() { | |
102 assert(_is_cms_thread ? | |
103 ConcurrentMarkSweepThread::cms_thread_has_cms_token() : | |
104 ConcurrentMarkSweepThread::vm_thread_has_cms_token(), | |
105 "Incorrect state"); | |
106 ConcurrentMarkSweepThread::desynchronize(_is_cms_thread); | |
107 } | |
108 }; | |
109 | |
110 // Convenience class that does a CMSTokenSync, and then acquires | |
111 // upto three locks. | |
112 class CMSTokenSyncWithLocks: public CMSTokenSync { | |
113 private: | |
114 // Note: locks are acquired in textual declaration order | |
115 // and released in the opposite order | |
116 MutexLockerEx _locker1, _locker2, _locker3; | |
117 public: | |
118 CMSTokenSyncWithLocks(bool is_cms_thread, Mutex* mutex1, | |
119 Mutex* mutex2 = NULL, Mutex* mutex3 = NULL): | |
120 CMSTokenSync(is_cms_thread), | |
121 _locker1(mutex1, Mutex::_no_safepoint_check_flag), | |
122 _locker2(mutex2, Mutex::_no_safepoint_check_flag), | |
123 _locker3(mutex3, Mutex::_no_safepoint_check_flag) | |
124 { } | |
125 }; | |
126 | |
127 | |
128 // Wrapper class to temporarily disable icms during a foreground cms collection. | |
129 class ICMSDisabler: public StackObj { | |
130 public: | |
131 // The ctor disables icms and wakes up the thread so it notices the change; | |
132 // the dtor re-enables icms. Note that the CMSCollector methods will check | |
133 // CMSIncrementalMode. | |
134 ICMSDisabler() { CMSCollector::disable_icms(); CMSCollector::start_icms(); } | |
135 ~ICMSDisabler() { CMSCollector::enable_icms(); } | |
136 }; | |
137 | |
138 ////////////////////////////////////////////////////////////////// | |
139 // Concurrent Mark-Sweep Generation ///////////////////////////// | |
140 ////////////////////////////////////////////////////////////////// | |
141 | |
142 NOT_PRODUCT(CompactibleFreeListSpace* debug_cms_space;) | |
143 | |
144 // This struct contains per-thread things necessary to support parallel | |
145 // young-gen collection. | |
146 class CMSParGCThreadState: public CHeapObj { | |
147 public: | |
148 CFLS_LAB lab; | |
149 PromotionInfo promo; | |
150 | |
151 // Constructor. | |
152 CMSParGCThreadState(CompactibleFreeListSpace* cfls) : lab(cfls) { | |
153 promo.setSpace(cfls); | |
154 } | |
155 }; | |
156 | |
157 ConcurrentMarkSweepGeneration::ConcurrentMarkSweepGeneration( | |
158 ReservedSpace rs, size_t initial_byte_size, int level, | |
159 CardTableRS* ct, bool use_adaptive_freelists, | |
160 FreeBlockDictionary::DictionaryChoice dictionaryChoice) : | |
161 CardGeneration(rs, initial_byte_size, level, ct), | |
1571
2d127394260e
6916623: Align object to 16 bytes to use Compressed Oops with java heap up to 64Gb
kvn
parents:
1521
diff
changeset
|
162 _dilatation_factor(((double)MinChunkSize)/((double)(CollectedHeap::min_fill_size()))), |
0 | 163 _debug_collection_type(Concurrent_collection_type) |
164 { | |
165 HeapWord* bottom = (HeapWord*) _virtual_space.low(); | |
166 HeapWord* end = (HeapWord*) _virtual_space.high(); | |
167 | |
168 _direct_allocated_words = 0; | |
169 NOT_PRODUCT( | |
170 _numObjectsPromoted = 0; | |
171 _numWordsPromoted = 0; | |
172 _numObjectsAllocated = 0; | |
173 _numWordsAllocated = 0; | |
174 ) | |
175 | |
176 _cmsSpace = new CompactibleFreeListSpace(_bts, MemRegion(bottom, end), | |
177 use_adaptive_freelists, | |
178 dictionaryChoice); | |
179 NOT_PRODUCT(debug_cms_space = _cmsSpace;) | |
180 if (_cmsSpace == NULL) { | |
181 vm_exit_during_initialization( | |
182 "CompactibleFreeListSpace allocation failure"); | |
183 } | |
184 _cmsSpace->_gen = this; | |
185 | |
186 _gc_stats = new CMSGCStats(); | |
187 | |
188 // Verify the assumption that FreeChunk::_prev and OopDesc::_klass | |
189 // offsets match. The ability to tell free chunks from objects | |
190 // depends on this property. | |
191 debug_only( | |
192 FreeChunk* junk = NULL; | |
187 | 193 assert(UseCompressedOops || |
194 junk->prev_addr() == (void*)(oop(junk)->klass_addr()), | |
0 | 195 "Offset of FreeChunk::_prev within FreeChunk must match" |
196 " that of OopDesc::_klass within OopDesc"); | |
197 ) | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
198 if (CollectedHeap::use_parallel_gc_threads()) { |
0 | 199 typedef CMSParGCThreadState* CMSParGCThreadStatePtr; |
200 _par_gc_thread_states = | |
201 NEW_C_HEAP_ARRAY(CMSParGCThreadStatePtr, ParallelGCThreads); | |
202 if (_par_gc_thread_states == NULL) { | |
203 vm_exit_during_initialization("Could not allocate par gc structs"); | |
204 } | |
205 for (uint i = 0; i < ParallelGCThreads; i++) { | |
206 _par_gc_thread_states[i] = new CMSParGCThreadState(cmsSpace()); | |
207 if (_par_gc_thread_states[i] == NULL) { | |
208 vm_exit_during_initialization("Could not allocate par gc structs"); | |
209 } | |
210 } | |
211 } else { | |
212 _par_gc_thread_states = NULL; | |
213 } | |
214 _incremental_collection_failed = false; | |
215 // The "dilatation_factor" is the expansion that can occur on | |
216 // account of the fact that the minimum object size in the CMS | |
217 // generation may be larger than that in, say, a contiguous young | |
218 // generation. | |
219 // Ideally, in the calculation below, we'd compute the dilatation | |
220 // factor as: MinChunkSize/(promoting_gen's min object size) | |
221 // Since we do not have such a general query interface for the | |
222 // promoting generation, we'll instead just use the mimimum | |
223 // object size (which today is a header's worth of space); | |
224 // note that all arithmetic is in units of HeapWords. | |
1571
2d127394260e
6916623: Align object to 16 bytes to use Compressed Oops with java heap up to 64Gb
kvn
parents:
1521
diff
changeset
|
225 assert(MinChunkSize >= CollectedHeap::min_fill_size(), "just checking"); |
0 | 226 assert(_dilatation_factor >= 1.0, "from previous assert"); |
227 } | |
228 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
229 |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
230 // The field "_initiating_occupancy" represents the occupancy percentage |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
231 // at which we trigger a new collection cycle. Unless explicitly specified |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
232 // via CMSInitiating[Perm]OccupancyFraction (argument "io" below), it |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
233 // is calculated by: |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
234 // |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
235 // Let "f" be MinHeapFreeRatio in |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
236 // |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
237 // _intiating_occupancy = 100-f + |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
238 // f * (CMSTrigger[Perm]Ratio/100) |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
239 // where CMSTrigger[Perm]Ratio is the argument "tr" below. |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
240 // |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
241 // That is, if we assume the heap is at its desired maximum occupancy at the |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
242 // end of a collection, we let CMSTrigger[Perm]Ratio of the (purported) free |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
243 // space be allocated before initiating a new collection cycle. |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
244 // |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
245 void ConcurrentMarkSweepGeneration::init_initiating_occupancy(intx io, intx tr) { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
246 assert(io <= 100 && tr >= 0 && tr <= 100, "Check the arguments"); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
247 if (io >= 0) { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
248 _initiating_occupancy = (double)io / 100.0; |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
249 } else { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
250 _initiating_occupancy = ((100 - MinHeapFreeRatio) + |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
251 (double)(tr * MinHeapFreeRatio) / 100.0) |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
252 / 100.0; |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
253 } |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
254 } |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
255 |
0 | 256 void ConcurrentMarkSweepGeneration::ref_processor_init() { |
257 assert(collector() != NULL, "no collector"); | |
258 collector()->ref_processor_init(); | |
259 } | |
260 | |
261 void CMSCollector::ref_processor_init() { | |
262 if (_ref_processor == NULL) { | |
263 // Allocate and initialize a reference processor | |
264 _ref_processor = ReferenceProcessor::create_ref_processor( | |
265 _span, // span | |
266 _cmsGen->refs_discovery_is_atomic(), // atomic_discovery | |
267 _cmsGen->refs_discovery_is_mt(), // mt_discovery | |
268 &_is_alive_closure, | |
269 ParallelGCThreads, | |
270 ParallelRefProcEnabled); | |
271 // Initialize the _ref_processor field of CMSGen | |
272 _cmsGen->set_ref_processor(_ref_processor); | |
273 | |
274 // Allocate a dummy ref processor for perm gen. | |
275 ReferenceProcessor* rp2 = new ReferenceProcessor(); | |
276 if (rp2 == NULL) { | |
277 vm_exit_during_initialization("Could not allocate ReferenceProcessor object"); | |
278 } | |
279 _permGen->set_ref_processor(rp2); | |
280 } | |
281 } | |
282 | |
283 CMSAdaptiveSizePolicy* CMSCollector::size_policy() { | |
284 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
285 assert(gch->kind() == CollectedHeap::GenCollectedHeap, | |
286 "Wrong type of heap"); | |
287 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) | |
288 gch->gen_policy()->size_policy(); | |
289 assert(sp->is_gc_cms_adaptive_size_policy(), | |
290 "Wrong type of size policy"); | |
291 return sp; | |
292 } | |
293 | |
294 CMSGCAdaptivePolicyCounters* CMSCollector::gc_adaptive_policy_counters() { | |
295 CMSGCAdaptivePolicyCounters* results = | |
296 (CMSGCAdaptivePolicyCounters*) collector_policy()->counters(); | |
297 assert( | |
298 results->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, | |
299 "Wrong gc policy counter kind"); | |
300 return results; | |
301 } | |
302 | |
303 | |
304 void ConcurrentMarkSweepGeneration::initialize_performance_counters() { | |
305 | |
306 const char* gen_name = "old"; | |
307 | |
308 // Generation Counters - generation 1, 1 subspace | |
309 _gen_counters = new GenerationCounters(gen_name, 1, 1, &_virtual_space); | |
310 | |
311 _space_counters = new GSpaceCounters(gen_name, 0, | |
312 _virtual_space.reserved_size(), | |
313 this, _gen_counters); | |
314 } | |
315 | |
316 CMSStats::CMSStats(ConcurrentMarkSweepGeneration* cms_gen, unsigned int alpha): | |
317 _cms_gen(cms_gen) | |
318 { | |
319 assert(alpha <= 100, "bad value"); | |
320 _saved_alpha = alpha; | |
321 | |
322 // Initialize the alphas to the bootstrap value of 100. | |
323 _gc0_alpha = _cms_alpha = 100; | |
324 | |
325 _cms_begin_time.update(); | |
326 _cms_end_time.update(); | |
327 | |
328 _gc0_duration = 0.0; | |
329 _gc0_period = 0.0; | |
330 _gc0_promoted = 0; | |
331 | |
332 _cms_duration = 0.0; | |
333 _cms_period = 0.0; | |
334 _cms_allocated = 0; | |
335 | |
336 _cms_used_at_gc0_begin = 0; | |
337 _cms_used_at_gc0_end = 0; | |
338 _allow_duty_cycle_reduction = false; | |
339 _valid_bits = 0; | |
340 _icms_duty_cycle = CMSIncrementalDutyCycle; | |
341 } | |
342 | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
343 double CMSStats::cms_free_adjustment_factor(size_t free) const { |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
344 // TBD: CR 6909490 |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
345 return 1.0; |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
346 } |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
347 |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
348 void CMSStats::adjust_cms_free_adjustment_factor(bool fail, size_t free) { |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
349 } |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
350 |
0 | 351 // If promotion failure handling is on use |
352 // the padded average size of the promotion for each | |
353 // young generation collection. | |
354 double CMSStats::time_until_cms_gen_full() const { | |
355 size_t cms_free = _cms_gen->cmsSpace()->free(); | |
356 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
357 size_t expected_promotion = gch->get_gen(0)->capacity(); | |
358 if (HandlePromotionFailure) { | |
359 expected_promotion = MIN2( | |
360 (size_t) _cms_gen->gc_stats()->avg_promoted()->padded_average(), | |
361 expected_promotion); | |
362 } | |
363 if (cms_free > expected_promotion) { | |
364 // Start a cms collection if there isn't enough space to promote | |
365 // for the next minor collection. Use the padded average as | |
366 // a safety factor. | |
367 cms_free -= expected_promotion; | |
368 | |
369 // Adjust by the safety factor. | |
370 double cms_free_dbl = (double)cms_free; | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
371 double cms_adjustment = (100.0 - CMSIncrementalSafetyFactor)/100.0; |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
372 // Apply a further correction factor which tries to adjust |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
373 // for recent occurance of concurrent mode failures. |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
374 cms_adjustment = cms_adjustment * cms_free_adjustment_factor(cms_free); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
375 cms_free_dbl = cms_free_dbl * cms_adjustment; |
0 | 376 |
377 if (PrintGCDetails && Verbose) { | |
378 gclog_or_tty->print_cr("CMSStats::time_until_cms_gen_full: cms_free " | |
379 SIZE_FORMAT " expected_promotion " SIZE_FORMAT, | |
380 cms_free, expected_promotion); | |
381 gclog_or_tty->print_cr(" cms_free_dbl %f cms_consumption_rate %f", | |
382 cms_free_dbl, cms_consumption_rate() + 1.0); | |
383 } | |
384 // Add 1 in case the consumption rate goes to zero. | |
385 return cms_free_dbl / (cms_consumption_rate() + 1.0); | |
386 } | |
387 return 0.0; | |
388 } | |
389 | |
390 // Compare the duration of the cms collection to the | |
391 // time remaining before the cms generation is empty. | |
392 // Note that the time from the start of the cms collection | |
393 // to the start of the cms sweep (less than the total | |
394 // duration of the cms collection) can be used. This | |
395 // has been tried and some applications experienced | |
396 // promotion failures early in execution. This was | |
397 // possibly because the averages were not accurate | |
398 // enough at the beginning. | |
399 double CMSStats::time_until_cms_start() const { | |
400 // We add "gc0_period" to the "work" calculation | |
401 // below because this query is done (mostly) at the | |
402 // end of a scavenge, so we need to conservatively | |
403 // account for that much possible delay | |
404 // in the query so as to avoid concurrent mode failures | |
405 // due to starting the collection just a wee bit too | |
406 // late. | |
407 double work = cms_duration() + gc0_period(); | |
408 double deadline = time_until_cms_gen_full(); | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
409 // If a concurrent mode failure occurred recently, we want to be |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
410 // more conservative and halve our expected time_until_cms_gen_full() |
0 | 411 if (work > deadline) { |
412 if (Verbose && PrintGCDetails) { | |
413 gclog_or_tty->print( | |
414 " CMSCollector: collect because of anticipated promotion " | |
415 "before full %3.7f + %3.7f > %3.7f ", cms_duration(), | |
416 gc0_period(), time_until_cms_gen_full()); | |
417 } | |
418 return 0.0; | |
419 } | |
420 return work - deadline; | |
421 } | |
422 | |
423 // Return a duty cycle based on old_duty_cycle and new_duty_cycle, limiting the | |
424 // amount of change to prevent wild oscillation. | |
425 unsigned int CMSStats::icms_damped_duty_cycle(unsigned int old_duty_cycle, | |
426 unsigned int new_duty_cycle) { | |
427 assert(old_duty_cycle <= 100, "bad input value"); | |
428 assert(new_duty_cycle <= 100, "bad input value"); | |
429 | |
430 // Note: use subtraction with caution since it may underflow (values are | |
431 // unsigned). Addition is safe since we're in the range 0-100. | |
432 unsigned int damped_duty_cycle = new_duty_cycle; | |
433 if (new_duty_cycle < old_duty_cycle) { | |
434 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 5U); | |
435 if (new_duty_cycle + largest_delta < old_duty_cycle) { | |
436 damped_duty_cycle = old_duty_cycle - largest_delta; | |
437 } | |
438 } else if (new_duty_cycle > old_duty_cycle) { | |
439 const unsigned int largest_delta = MAX2(old_duty_cycle / 4, 15U); | |
440 if (new_duty_cycle > old_duty_cycle + largest_delta) { | |
441 damped_duty_cycle = MIN2(old_duty_cycle + largest_delta, 100U); | |
442 } | |
443 } | |
444 assert(damped_duty_cycle <= 100, "invalid duty cycle computed"); | |
445 | |
446 if (CMSTraceIncrementalPacing) { | |
447 gclog_or_tty->print(" [icms_damped_duty_cycle(%d,%d) = %d] ", | |
448 old_duty_cycle, new_duty_cycle, damped_duty_cycle); | |
449 } | |
450 return damped_duty_cycle; | |
451 } | |
452 | |
453 unsigned int CMSStats::icms_update_duty_cycle_impl() { | |
454 assert(CMSIncrementalPacing && valid(), | |
455 "should be handled in icms_update_duty_cycle()"); | |
456 | |
457 double cms_time_so_far = cms_timer().seconds(); | |
458 double scaled_duration = cms_duration_per_mb() * _cms_used_at_gc0_end / M; | |
459 double scaled_duration_remaining = fabsd(scaled_duration - cms_time_so_far); | |
460 | |
461 // Avoid division by 0. | |
462 double time_until_full = MAX2(time_until_cms_gen_full(), 0.01); | |
463 double duty_cycle_dbl = 100.0 * scaled_duration_remaining / time_until_full; | |
464 | |
465 unsigned int new_duty_cycle = MIN2((unsigned int)duty_cycle_dbl, 100U); | |
466 if (new_duty_cycle > _icms_duty_cycle) { | |
467 // Avoid very small duty cycles (1 or 2); 0 is allowed. | |
468 if (new_duty_cycle > 2) { | |
469 _icms_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, | |
470 new_duty_cycle); | |
471 } | |
472 } else if (_allow_duty_cycle_reduction) { | |
473 // The duty cycle is reduced only once per cms cycle (see record_cms_end()). | |
474 new_duty_cycle = icms_damped_duty_cycle(_icms_duty_cycle, new_duty_cycle); | |
475 // Respect the minimum duty cycle. | |
476 unsigned int min_duty_cycle = (unsigned int)CMSIncrementalDutyCycleMin; | |
477 _icms_duty_cycle = MAX2(new_duty_cycle, min_duty_cycle); | |
478 } | |
479 | |
480 if (PrintGCDetails || CMSTraceIncrementalPacing) { | |
481 gclog_or_tty->print(" icms_dc=%d ", _icms_duty_cycle); | |
482 } | |
483 | |
484 _allow_duty_cycle_reduction = false; | |
485 return _icms_duty_cycle; | |
486 } | |
487 | |
488 #ifndef PRODUCT | |
489 void CMSStats::print_on(outputStream *st) const { | |
490 st->print(" gc0_alpha=%d,cms_alpha=%d", _gc0_alpha, _cms_alpha); | |
491 st->print(",gc0_dur=%g,gc0_per=%g,gc0_promo=" SIZE_FORMAT, | |
492 gc0_duration(), gc0_period(), gc0_promoted()); | |
493 st->print(",cms_dur=%g,cms_dur_per_mb=%g,cms_per=%g,cms_alloc=" SIZE_FORMAT, | |
494 cms_duration(), cms_duration_per_mb(), | |
495 cms_period(), cms_allocated()); | |
496 st->print(",cms_since_beg=%g,cms_since_end=%g", | |
497 cms_time_since_begin(), cms_time_since_end()); | |
498 st->print(",cms_used_beg=" SIZE_FORMAT ",cms_used_end=" SIZE_FORMAT, | |
499 _cms_used_at_gc0_begin, _cms_used_at_gc0_end); | |
500 if (CMSIncrementalMode) { | |
501 st->print(",dc=%d", icms_duty_cycle()); | |
502 } | |
503 | |
504 if (valid()) { | |
505 st->print(",promo_rate=%g,cms_alloc_rate=%g", | |
506 promotion_rate(), cms_allocation_rate()); | |
507 st->print(",cms_consumption_rate=%g,time_until_full=%g", | |
508 cms_consumption_rate(), time_until_cms_gen_full()); | |
509 } | |
510 st->print(" "); | |
511 } | |
512 #endif // #ifndef PRODUCT | |
513 | |
514 CMSCollector::CollectorState CMSCollector::_collectorState = | |
515 CMSCollector::Idling; | |
516 bool CMSCollector::_foregroundGCIsActive = false; | |
517 bool CMSCollector::_foregroundGCShouldWait = false; | |
518 | |
519 CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen, | |
520 ConcurrentMarkSweepGeneration* permGen, | |
521 CardTableRS* ct, | |
522 ConcurrentMarkSweepPolicy* cp): | |
523 _cmsGen(cmsGen), | |
524 _permGen(permGen), | |
525 _ct(ct), | |
526 _ref_processor(NULL), // will be set later | |
527 _conc_workers(NULL), // may be set later | |
528 _abort_preclean(false), | |
529 _start_sampling(false), | |
530 _between_prologue_and_epilogue(false), | |
531 _markBitMap(0, Mutex::leaf + 1, "CMS_markBitMap_lock"), | |
532 _perm_gen_verify_bit_map(0, -1 /* no mutex */, "No_lock"), | |
533 _modUnionTable((CardTableModRefBS::card_shift - LogHeapWordSize), | |
534 -1 /* lock-free */, "No_lock" /* dummy */), | |
535 _modUnionClosure(&_modUnionTable), | |
536 _modUnionClosurePar(&_modUnionTable), | |
143
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
537 // Adjust my span to cover old (cms) gen and perm gen |
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
538 _span(cmsGen->reserved()._union(permGen->reserved())), |
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
539 // Construct the is_alive_closure with _span & markBitMap |
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
540 _is_alive_closure(_span, &_markBitMap), |
0 | 541 _restart_addr(NULL), |
542 _overflow_list(NULL), | |
543 _stats(cmsGen), | |
544 _eden_chunk_array(NULL), // may be set in ctor body | |
545 _eden_chunk_capacity(0), // -- ditto -- | |
546 _eden_chunk_index(0), // -- ditto -- | |
547 _survivor_plab_array(NULL), // -- ditto -- | |
548 _survivor_chunk_array(NULL), // -- ditto -- | |
549 _survivor_chunk_capacity(0), // -- ditto -- | |
550 _survivor_chunk_index(0), // -- ditto -- | |
551 _ser_pmc_preclean_ovflw(0), | |
452
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
552 _ser_kac_preclean_ovflw(0), |
0 | 553 _ser_pmc_remark_ovflw(0), |
554 _par_pmc_remark_ovflw(0), | |
555 _ser_kac_ovflw(0), | |
556 _par_kac_ovflw(0), | |
557 #ifndef PRODUCT | |
558 _num_par_pushes(0), | |
559 #endif | |
560 _collection_count_start(0), | |
561 _verifying(false), | |
562 _icms_start_limit(NULL), | |
563 _icms_stop_limit(NULL), | |
564 _verification_mark_bm(0, Mutex::leaf + 1, "CMS_verification_mark_bm_lock"), | |
565 _completed_initialization(false), | |
566 _collector_policy(cp), | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
567 _should_unload_classes(false), |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
568 _concurrent_cycles_since_last_unload(0), |
798
fe1574da39fc
6848641: CMSCollector::_roots_scanning_options should be initialized
ysr
parents:
679
diff
changeset
|
569 _roots_scanning_options(0), |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
570 _inter_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding), |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
571 _intra_sweep_estimate(CMS_SweepWeight, CMS_SweepPadding) |
0 | 572 { |
573 if (ExplicitGCInvokesConcurrentAndUnloadsClasses) { | |
574 ExplicitGCInvokesConcurrent = true; | |
575 } | |
576 // Now expand the span and allocate the collection support structures | |
577 // (MUT, marking bit map etc.) to cover both generations subject to | |
578 // collection. | |
579 | |
580 // First check that _permGen is adjacent to _cmsGen and above it. | |
581 assert( _cmsGen->reserved().word_size() > 0 | |
582 && _permGen->reserved().word_size() > 0, | |
583 "generations should not be of zero size"); | |
584 assert(_cmsGen->reserved().intersection(_permGen->reserved()).is_empty(), | |
585 "_cmsGen and _permGen should not overlap"); | |
586 assert(_cmsGen->reserved().end() == _permGen->reserved().start(), | |
587 "_cmsGen->end() different from _permGen->start()"); | |
588 | |
589 // For use by dirty card to oop closures. | |
590 _cmsGen->cmsSpace()->set_collector(this); | |
591 _permGen->cmsSpace()->set_collector(this); | |
592 | |
593 // Allocate MUT and marking bit map | |
594 { | |
595 MutexLockerEx x(_markBitMap.lock(), Mutex::_no_safepoint_check_flag); | |
596 if (!_markBitMap.allocate(_span)) { | |
597 warning("Failed to allocate CMS Bit Map"); | |
598 return; | |
599 } | |
600 assert(_markBitMap.covers(_span), "_markBitMap inconsistency?"); | |
601 } | |
602 { | |
603 _modUnionTable.allocate(_span); | |
604 assert(_modUnionTable.covers(_span), "_modUnionTable inconsistency?"); | |
605 } | |
606 | |
1284 | 607 if (!_markStack.allocate(MarkStackSize)) { |
0 | 608 warning("Failed to allocate CMS Marking Stack"); |
609 return; | |
610 } | |
611 if (!_revisitStack.allocate(CMSRevisitStackSize)) { | |
612 warning("Failed to allocate CMS Revisit Stack"); | |
613 return; | |
614 } | |
615 | |
616 // Support for multi-threaded concurrent phases | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
617 if (CollectedHeap::use_parallel_gc_threads() && CMSConcurrentMTEnabled) { |
1284 | 618 if (FLAG_IS_DEFAULT(ConcGCThreads)) { |
0 | 619 // just for now |
1284 | 620 FLAG_SET_DEFAULT(ConcGCThreads, (ParallelGCThreads + 3)/4); |
621 } | |
622 if (ConcGCThreads > 1) { | |
0 | 623 _conc_workers = new YieldingFlexibleWorkGang("Parallel CMS Threads", |
1284 | 624 ConcGCThreads, true); |
0 | 625 if (_conc_workers == NULL) { |
626 warning("GC/CMS: _conc_workers allocation failure: " | |
627 "forcing -CMSConcurrentMTEnabled"); | |
628 CMSConcurrentMTEnabled = false; | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
629 } else { |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
630 _conc_workers->initialize_workers(); |
0 | 631 } |
632 } else { | |
633 CMSConcurrentMTEnabled = false; | |
634 } | |
635 } | |
636 if (!CMSConcurrentMTEnabled) { | |
1284 | 637 ConcGCThreads = 0; |
0 | 638 } else { |
639 // Turn off CMSCleanOnEnter optimization temporarily for | |
640 // the MT case where it's not fixed yet; see 6178663. | |
641 CMSCleanOnEnter = false; | |
642 } | |
1284 | 643 assert((_conc_workers != NULL) == (ConcGCThreads > 1), |
0 | 644 "Inconsistency"); |
645 | |
646 // Parallel task queues; these are shared for the | |
647 // concurrent and stop-world phases of CMS, but | |
648 // are not shared with parallel scavenge (ParNew). | |
649 { | |
650 uint i; | |
1284 | 651 uint num_queues = (uint) MAX2(ParallelGCThreads, ConcGCThreads); |
0 | 652 |
653 if ((CMSParallelRemarkEnabled || CMSConcurrentMTEnabled | |
654 || ParallelRefProcEnabled) | |
655 && num_queues > 0) { | |
656 _task_queues = new OopTaskQueueSet(num_queues); | |
657 if (_task_queues == NULL) { | |
658 warning("task_queues allocation failure."); | |
659 return; | |
660 } | |
661 _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues); | |
662 if (_hash_seed == NULL) { | |
663 warning("_hash_seed array allocation failure"); | |
664 return; | |
665 } | |
666 | |
1665 | 667 typedef Padded<OopTaskQueue> PaddedOopTaskQueue; |
0 | 668 for (i = 0; i < num_queues; i++) { |
1665 | 669 PaddedOopTaskQueue *q = new PaddedOopTaskQueue(); |
670 if (q == NULL) { | |
0 | 671 warning("work_queue allocation failure."); |
672 return; | |
673 } | |
1665 | 674 _task_queues->register_queue(i, q); |
0 | 675 } |
676 for (i = 0; i < num_queues; i++) { | |
677 _task_queues->queue(i)->initialize(); | |
678 _hash_seed[i] = 17; // copied from ParNew | |
679 } | |
680 } | |
681 } | |
682 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
683 _cmsGen ->init_initiating_occupancy(CMSInitiatingOccupancyFraction, CMSTriggerRatio); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
684 _permGen->init_initiating_occupancy(CMSInitiatingPermOccupancyFraction, CMSTriggerPermRatio); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
685 |
0 | 686 // Clip CMSBootstrapOccupancy between 0 and 100. |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
687 _bootstrap_occupancy = ((double)MIN2((uintx)100, MAX2((uintx)0, CMSBootstrapOccupancy))) |
0 | 688 /(double)100; |
689 | |
690 _full_gcs_since_conc_gc = 0; | |
691 | |
692 // Now tell CMS generations the identity of their collector | |
693 ConcurrentMarkSweepGeneration::set_collector(this); | |
694 | |
695 // Create & start a CMS thread for this CMS collector | |
696 _cmsThread = ConcurrentMarkSweepThread::start(this); | |
697 assert(cmsThread() != NULL, "CMS Thread should have been created"); | |
698 assert(cmsThread()->collector() == this, | |
699 "CMS Thread should refer to this gen"); | |
700 assert(CGC_lock != NULL, "Where's the CGC_lock?"); | |
701 | |
702 // Support for parallelizing young gen rescan | |
703 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
704 _young_gen = gch->prev_gen(_cmsGen); | |
705 if (gch->supports_inline_contig_alloc()) { | |
706 _top_addr = gch->top_addr(); | |
707 _end_addr = gch->end_addr(); | |
708 assert(_young_gen != NULL, "no _young_gen"); | |
709 _eden_chunk_index = 0; | |
710 _eden_chunk_capacity = (_young_gen->max_capacity()+CMSSamplingGrain)/CMSSamplingGrain; | |
711 _eden_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, _eden_chunk_capacity); | |
712 if (_eden_chunk_array == NULL) { | |
713 _eden_chunk_capacity = 0; | |
714 warning("GC/CMS: _eden_chunk_array allocation failure"); | |
715 } | |
716 } | |
717 assert(_eden_chunk_array != NULL || _eden_chunk_capacity == 0, "Error"); | |
718 | |
719 // Support for parallelizing survivor space rescan | |
720 if (CMSParallelRemarkEnabled && CMSParallelSurvivorRemarkEnabled) { | |
1289
d47555d7aca8
6910182: CMS: assert(_cursor[j] == _survivor_plab_array[j].end(),"Ctl pt invariant")
jmasa
parents:
1284
diff
changeset
|
721 const size_t max_plab_samples = |
d47555d7aca8
6910182: CMS: assert(_cursor[j] == _survivor_plab_array[j].end(),"Ctl pt invariant")
jmasa
parents:
1284
diff
changeset
|
722 ((DefNewGeneration*)_young_gen)->max_survivor_size()/MinTLABSize; |
d47555d7aca8
6910182: CMS: assert(_cursor[j] == _survivor_plab_array[j].end(),"Ctl pt invariant")
jmasa
parents:
1284
diff
changeset
|
723 |
0 | 724 _survivor_plab_array = NEW_C_HEAP_ARRAY(ChunkArray, ParallelGCThreads); |
725 _survivor_chunk_array = NEW_C_HEAP_ARRAY(HeapWord*, 2*max_plab_samples); | |
726 _cursor = NEW_C_HEAP_ARRAY(size_t, ParallelGCThreads); | |
727 if (_survivor_plab_array == NULL || _survivor_chunk_array == NULL | |
728 || _cursor == NULL) { | |
729 warning("Failed to allocate survivor plab/chunk array"); | |
730 if (_survivor_plab_array != NULL) { | |
731 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); | |
732 _survivor_plab_array = NULL; | |
733 } | |
734 if (_survivor_chunk_array != NULL) { | |
735 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); | |
736 _survivor_chunk_array = NULL; | |
737 } | |
738 if (_cursor != NULL) { | |
739 FREE_C_HEAP_ARRAY(size_t, _cursor); | |
740 _cursor = NULL; | |
741 } | |
742 } else { | |
743 _survivor_chunk_capacity = 2*max_plab_samples; | |
744 for (uint i = 0; i < ParallelGCThreads; i++) { | |
745 HeapWord** vec = NEW_C_HEAP_ARRAY(HeapWord*, max_plab_samples); | |
746 if (vec == NULL) { | |
747 warning("Failed to allocate survivor plab array"); | |
748 for (int j = i; j > 0; j--) { | |
749 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_plab_array[j-1].array()); | |
750 } | |
751 FREE_C_HEAP_ARRAY(ChunkArray, _survivor_plab_array); | |
752 FREE_C_HEAP_ARRAY(HeapWord*, _survivor_chunk_array); | |
753 _survivor_plab_array = NULL; | |
754 _survivor_chunk_array = NULL; | |
755 _survivor_chunk_capacity = 0; | |
756 break; | |
757 } else { | |
758 ChunkArray* cur = | |
759 ::new (&_survivor_plab_array[i]) ChunkArray(vec, | |
760 max_plab_samples); | |
761 assert(cur->end() == 0, "Should be 0"); | |
762 assert(cur->array() == vec, "Should be vec"); | |
763 assert(cur->capacity() == max_plab_samples, "Error"); | |
764 } | |
765 } | |
766 } | |
767 } | |
768 assert( ( _survivor_plab_array != NULL | |
769 && _survivor_chunk_array != NULL) | |
770 || ( _survivor_chunk_capacity == 0 | |
771 && _survivor_chunk_index == 0), | |
772 "Error"); | |
773 | |
774 // Choose what strong roots should be scanned depending on verification options | |
775 // and perm gen collection mode. | |
776 if (!CMSClassUnloadingEnabled) { | |
777 // If class unloading is disabled we want to include all classes into the root set. | |
778 add_root_scanning_option(SharedHeap::SO_AllClasses); | |
779 } else { | |
780 add_root_scanning_option(SharedHeap::SO_SystemClasses); | |
781 } | |
782 | |
783 NOT_PRODUCT(_overflow_counter = CMSMarkStackOverflowInterval;) | |
784 _gc_counters = new CollectorCounters("CMS", 1); | |
785 _completed_initialization = true; | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
786 _inter_sweep_timer.start(); // start of time |
1518
3bfae429e2cf
6948537: CMS: BOT walkers observe out-of-thin-air zeros on sun4v sparc/CMT
ysr
parents:
1387
diff
changeset
|
787 #ifdef SPARC |
3bfae429e2cf
6948537: CMS: BOT walkers observe out-of-thin-air zeros on sun4v sparc/CMT
ysr
parents:
1387
diff
changeset
|
788 // Issue a stern warning, but allow use for experimentation and debugging. |
3bfae429e2cf
6948537: CMS: BOT walkers observe out-of-thin-air zeros on sun4v sparc/CMT
ysr
parents:
1387
diff
changeset
|
789 if (VM_Version::is_sun4v() && UseMemSetInBOT) { |
3bfae429e2cf
6948537: CMS: BOT walkers observe out-of-thin-air zeros on sun4v sparc/CMT
ysr
parents:
1387
diff
changeset
|
790 assert(!FLAG_IS_DEFAULT(UseMemSetInBOT), "Error"); |
3bfae429e2cf
6948537: CMS: BOT walkers observe out-of-thin-air zeros on sun4v sparc/CMT
ysr
parents:
1387
diff
changeset
|
791 warning("Experimental flag -XX:+UseMemSetInBOT is known to cause instability" |
3bfae429e2cf
6948537: CMS: BOT walkers observe out-of-thin-air zeros on sun4v sparc/CMT
ysr
parents:
1387
diff
changeset
|
792 " on sun4v; please understand that you are using at your own risk!"); |
3bfae429e2cf
6948537: CMS: BOT walkers observe out-of-thin-air zeros on sun4v sparc/CMT
ysr
parents:
1387
diff
changeset
|
793 } |
3bfae429e2cf
6948537: CMS: BOT walkers observe out-of-thin-air zeros on sun4v sparc/CMT
ysr
parents:
1387
diff
changeset
|
794 #endif |
0 | 795 } |
796 | |
797 const char* ConcurrentMarkSweepGeneration::name() const { | |
798 return "concurrent mark-sweep generation"; | |
799 } | |
800 void ConcurrentMarkSweepGeneration::update_counters() { | |
801 if (UsePerfData) { | |
802 _space_counters->update_all(); | |
803 _gen_counters->update_all(); | |
804 } | |
805 } | |
806 | |
807 // this is an optimized version of update_counters(). it takes the | |
808 // used value as a parameter rather than computing it. | |
809 // | |
810 void ConcurrentMarkSweepGeneration::update_counters(size_t used) { | |
811 if (UsePerfData) { | |
812 _space_counters->update_used(used); | |
813 _space_counters->update_capacity(); | |
814 _gen_counters->update_all(); | |
815 } | |
816 } | |
817 | |
818 void ConcurrentMarkSweepGeneration::print() const { | |
819 Generation::print(); | |
820 cmsSpace()->print(); | |
821 } | |
822 | |
823 #ifndef PRODUCT | |
824 void ConcurrentMarkSweepGeneration::print_statistics() { | |
825 cmsSpace()->printFLCensus(0); | |
826 } | |
827 #endif | |
828 | |
829 void ConcurrentMarkSweepGeneration::printOccupancy(const char *s) { | |
830 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
831 if (PrintGCDetails) { | |
832 if (Verbose) { | |
833 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"("SIZE_FORMAT")]", | |
834 level(), short_name(), s, used(), capacity()); | |
835 } else { | |
836 gclog_or_tty->print(" [%d %s-%s: "SIZE_FORMAT"K("SIZE_FORMAT"K)]", | |
837 level(), short_name(), s, used() / K, capacity() / K); | |
838 } | |
839 } | |
840 if (Verbose) { | |
841 gclog_or_tty->print(" "SIZE_FORMAT"("SIZE_FORMAT")", | |
842 gch->used(), gch->capacity()); | |
843 } else { | |
844 gclog_or_tty->print(" "SIZE_FORMAT"K("SIZE_FORMAT"K)", | |
845 gch->used() / K, gch->capacity() / K); | |
846 } | |
847 } | |
848 | |
849 size_t | |
850 ConcurrentMarkSweepGeneration::contiguous_available() const { | |
851 // dld proposes an improvement in precision here. If the committed | |
852 // part of the space ends in a free block we should add that to | |
853 // uncommitted size in the calculation below. Will make this | |
854 // change later, staying with the approximation below for the | |
855 // time being. -- ysr. | |
856 return MAX2(_virtual_space.uncommitted_size(), unsafe_max_alloc_nogc()); | |
857 } | |
858 | |
859 size_t | |
860 ConcurrentMarkSweepGeneration::unsafe_max_alloc_nogc() const { | |
861 return _cmsSpace->max_alloc_in_words() * HeapWordSize; | |
862 } | |
863 | |
864 size_t ConcurrentMarkSweepGeneration::max_available() const { | |
865 return free() + _virtual_space.uncommitted_size(); | |
866 } | |
867 | |
868 bool ConcurrentMarkSweepGeneration::promotion_attempt_is_safe( | |
869 size_t max_promotion_in_bytes, | |
870 bool younger_handles_promotion_failure) const { | |
871 | |
872 // This is the most conservative test. Full promotion is | |
873 // guaranteed if this is used. The multiplicative factor is to | |
874 // account for the worst case "dilatation". | |
875 double adjusted_max_promo_bytes = _dilatation_factor * max_promotion_in_bytes; | |
876 if (adjusted_max_promo_bytes > (double)max_uintx) { // larger than size_t | |
877 adjusted_max_promo_bytes = (double)max_uintx; | |
878 } | |
879 bool result = (max_contiguous_available() >= (size_t)adjusted_max_promo_bytes); | |
880 | |
881 if (younger_handles_promotion_failure && !result) { | |
882 // Full promotion is not guaranteed because fragmentation | |
883 // of the cms generation can prevent the full promotion. | |
884 result = (max_available() >= (size_t)adjusted_max_promo_bytes); | |
885 | |
886 if (!result) { | |
887 // With promotion failure handling the test for the ability | |
888 // to support the promotion does not have to be guaranteed. | |
889 // Use an average of the amount promoted. | |
890 result = max_available() >= (size_t) | |
891 gc_stats()->avg_promoted()->padded_average(); | |
892 if (PrintGC && Verbose && result) { | |
893 gclog_or_tty->print_cr( | |
894 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" | |
895 " max_available: " SIZE_FORMAT | |
896 " avg_promoted: " SIZE_FORMAT, | |
897 max_available(), (size_t) | |
898 gc_stats()->avg_promoted()->padded_average()); | |
899 } | |
900 } else { | |
901 if (PrintGC && Verbose) { | |
902 gclog_or_tty->print_cr( | |
903 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" | |
904 " max_available: " SIZE_FORMAT | |
905 " adj_max_promo_bytes: " SIZE_FORMAT, | |
906 max_available(), (size_t)adjusted_max_promo_bytes); | |
907 } | |
908 } | |
909 } else { | |
910 if (PrintGC && Verbose) { | |
911 gclog_or_tty->print_cr( | |
912 "\nConcurrentMarkSweepGeneration::promotion_attempt_is_safe" | |
913 " contiguous_available: " SIZE_FORMAT | |
914 " adj_max_promo_bytes: " SIZE_FORMAT, | |
915 max_contiguous_available(), (size_t)adjusted_max_promo_bytes); | |
916 } | |
917 } | |
918 return result; | |
919 } | |
920 | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
921 // At a promotion failure dump information on block layout in heap |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
922 // (cms old generation). |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
923 void ConcurrentMarkSweepGeneration::promotion_failure_occurred() { |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
924 if (CMSDumpAtPromotionFailure) { |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
925 cmsSpace()->dump_at_safepoint_with_locks(collector(), gclog_or_tty); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
926 } |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
927 } |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
928 |
0 | 929 CompactibleSpace* |
930 ConcurrentMarkSweepGeneration::first_compaction_space() const { | |
931 return _cmsSpace; | |
932 } | |
933 | |
934 void ConcurrentMarkSweepGeneration::reset_after_compaction() { | |
935 // Clear the promotion information. These pointers can be adjusted | |
936 // along with all the other pointers into the heap but | |
937 // compaction is expected to be a rare event with | |
938 // a heap using cms so don't do it without seeing the need. | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
939 if (CollectedHeap::use_parallel_gc_threads()) { |
0 | 940 for (uint i = 0; i < ParallelGCThreads; i++) { |
941 _par_gc_thread_states[i]->promo.reset(); | |
942 } | |
943 } | |
944 } | |
945 | |
946 void ConcurrentMarkSweepGeneration::space_iterate(SpaceClosure* blk, bool usedOnly) { | |
947 blk->do_space(_cmsSpace); | |
948 } | |
949 | |
950 void ConcurrentMarkSweepGeneration::compute_new_size() { | |
951 assert_locked_or_safepoint(Heap_lock); | |
952 | |
953 // If incremental collection failed, we just want to expand | |
954 // to the limit. | |
955 if (incremental_collection_failed()) { | |
956 clear_incremental_collection_failed(); | |
957 grow_to_reserved(); | |
958 return; | |
959 } | |
960 | |
961 size_t expand_bytes = 0; | |
962 double free_percentage = ((double) free()) / capacity(); | |
963 double desired_free_percentage = (double) MinHeapFreeRatio / 100; | |
964 double maximum_free_percentage = (double) MaxHeapFreeRatio / 100; | |
965 | |
966 // compute expansion delta needed for reaching desired free percentage | |
967 if (free_percentage < desired_free_percentage) { | |
968 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); | |
969 assert(desired_capacity >= capacity(), "invalid expansion size"); | |
970 expand_bytes = MAX2(desired_capacity - capacity(), MinHeapDeltaBytes); | |
971 } | |
972 if (expand_bytes > 0) { | |
973 if (PrintGCDetails && Verbose) { | |
974 size_t desired_capacity = (size_t)(used() / ((double) 1 - desired_free_percentage)); | |
975 gclog_or_tty->print_cr("\nFrom compute_new_size: "); | |
976 gclog_or_tty->print_cr(" Free fraction %f", free_percentage); | |
977 gclog_or_tty->print_cr(" Desired free fraction %f", | |
978 desired_free_percentage); | |
979 gclog_or_tty->print_cr(" Maximum free fraction %f", | |
980 maximum_free_percentage); | |
981 gclog_or_tty->print_cr(" Capactiy "SIZE_FORMAT, capacity()/1000); | |
982 gclog_or_tty->print_cr(" Desired capacity "SIZE_FORMAT, | |
983 desired_capacity/1000); | |
984 int prev_level = level() - 1; | |
985 if (prev_level >= 0) { | |
986 size_t prev_size = 0; | |
987 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
988 Generation* prev_gen = gch->_gens[prev_level]; | |
989 prev_size = prev_gen->capacity(); | |
990 gclog_or_tty->print_cr(" Younger gen size "SIZE_FORMAT, | |
991 prev_size/1000); | |
992 } | |
993 gclog_or_tty->print_cr(" unsafe_max_alloc_nogc "SIZE_FORMAT, | |
994 unsafe_max_alloc_nogc()/1000); | |
995 gclog_or_tty->print_cr(" contiguous available "SIZE_FORMAT, | |
996 contiguous_available()/1000); | |
997 gclog_or_tty->print_cr(" Expand by "SIZE_FORMAT" (bytes)", | |
998 expand_bytes); | |
999 } | |
1000 // safe if expansion fails | |
1001 expand(expand_bytes, 0, CMSExpansionCause::_satisfy_free_ratio); | |
1002 if (PrintGCDetails && Verbose) { | |
1003 gclog_or_tty->print_cr(" Expanded free fraction %f", | |
1004 ((double) free()) / capacity()); | |
1005 } | |
1006 } | |
1007 } | |
1008 | |
1009 Mutex* ConcurrentMarkSweepGeneration::freelistLock() const { | |
1010 return cmsSpace()->freelistLock(); | |
1011 } | |
1012 | |
1013 HeapWord* ConcurrentMarkSweepGeneration::allocate(size_t size, | |
1014 bool tlab) { | |
1015 CMSSynchronousYieldRequest yr; | |
1016 MutexLockerEx x(freelistLock(), | |
1017 Mutex::_no_safepoint_check_flag); | |
1018 return have_lock_and_allocate(size, tlab); | |
1019 } | |
1020 | |
1021 HeapWord* ConcurrentMarkSweepGeneration::have_lock_and_allocate(size_t size, | |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1022 bool tlab /* ignored */) { |
0 | 1023 assert_lock_strong(freelistLock()); |
1024 size_t adjustedSize = CompactibleFreeListSpace::adjustObjectSize(size); | |
1025 HeapWord* res = cmsSpace()->allocate(adjustedSize); | |
1026 // Allocate the object live (grey) if the background collector has | |
1027 // started marking. This is necessary because the marker may | |
1028 // have passed this address and consequently this object will | |
1029 // not otherwise be greyed and would be incorrectly swept up. | |
1030 // Note that if this object contains references, the writing | |
1031 // of those references will dirty the card containing this object | |
1032 // allowing the object to be blackened (and its references scanned) | |
1033 // either during a preclean phase or at the final checkpoint. | |
1034 if (res != NULL) { | |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1035 // We may block here with an uninitialized object with |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1036 // its mark-bit or P-bits not yet set. Such objects need |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1037 // to be safely navigable by block_start(). |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1038 assert(oop(res)->klass_or_null() == NULL, "Object should be uninitialized here."); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1039 assert(!((FreeChunk*)res)->isFree(), "Error, block will look free but show wrong size"); |
0 | 1040 collector()->direct_allocated(res, adjustedSize); |
1041 _direct_allocated_words += adjustedSize; | |
1042 // allocation counters | |
1043 NOT_PRODUCT( | |
1044 _numObjectsAllocated++; | |
1045 _numWordsAllocated += (int)adjustedSize; | |
1046 ) | |
1047 } | |
1048 return res; | |
1049 } | |
1050 | |
1051 // In the case of direct allocation by mutators in a generation that | |
1052 // is being concurrently collected, the object must be allocated | |
1053 // live (grey) if the background collector has started marking. | |
1054 // This is necessary because the marker may | |
1055 // have passed this address and consequently this object will | |
1056 // not otherwise be greyed and would be incorrectly swept up. | |
1057 // Note that if this object contains references, the writing | |
1058 // of those references will dirty the card containing this object | |
1059 // allowing the object to be blackened (and its references scanned) | |
1060 // either during a preclean phase or at the final checkpoint. | |
1061 void CMSCollector::direct_allocated(HeapWord* start, size_t size) { | |
1062 assert(_markBitMap.covers(start, size), "Out of bounds"); | |
1063 if (_collectorState >= Marking) { | |
1064 MutexLockerEx y(_markBitMap.lock(), | |
1065 Mutex::_no_safepoint_check_flag); | |
1066 // [see comments preceding SweepClosure::do_blk() below for details] | |
1067 // 1. need to mark the object as live so it isn't collected | |
1068 // 2. need to mark the 2nd bit to indicate the object may be uninitialized | |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1069 // 3. need to mark the end of the object so marking, precleaning or sweeping |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1070 // can skip over uninitialized or unparsable objects. An allocated |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1071 // object is considered uninitialized for our purposes as long as |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1072 // its klass word is NULL. (Unparsable objects are those which are |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1073 // initialized in the sense just described, but whose sizes can still |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1074 // not be correctly determined. Note that the class of unparsable objects |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1075 // can only occur in the perm gen. All old gen objects are parsable |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1076 // as soon as they are initialized.) |
0 | 1077 _markBitMap.mark(start); // object is live |
1078 _markBitMap.mark(start + 1); // object is potentially uninitialized? | |
1079 _markBitMap.mark(start + size - 1); | |
1080 // mark end of object | |
1081 } | |
1082 // check that oop looks uninitialized | |
187 | 1083 assert(oop(start)->klass_or_null() == NULL, "_klass should be NULL"); |
0 | 1084 } |
1085 | |
1086 void CMSCollector::promoted(bool par, HeapWord* start, | |
1087 bool is_obj_array, size_t obj_size) { | |
1088 assert(_markBitMap.covers(start), "Out of bounds"); | |
1089 // See comment in direct_allocated() about when objects should | |
1090 // be allocated live. | |
1091 if (_collectorState >= Marking) { | |
1092 // we already hold the marking bit map lock, taken in | |
1093 // the prologue | |
1094 if (par) { | |
1095 _markBitMap.par_mark(start); | |
1096 } else { | |
1097 _markBitMap.mark(start); | |
1098 } | |
1099 // We don't need to mark the object as uninitialized (as | |
1100 // in direct_allocated above) because this is being done with the | |
1101 // world stopped and the object will be initialized by the | |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1102 // time the marking, precleaning or sweeping get to look at it. |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1103 // But see the code for copying objects into the CMS generation, |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1104 // where we need to ensure that concurrent readers of the |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1105 // block offset table are able to safely navigate a block that |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1106 // is in flux from being free to being allocated (and in |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1107 // transition while being copied into) and subsequently |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1108 // becoming a bona-fide object when the copy/promotion is complete. |
0 | 1109 assert(SafepointSynchronize::is_at_safepoint(), |
1110 "expect promotion only at safepoints"); | |
1111 | |
1112 if (_collectorState < Sweeping) { | |
1113 // Mark the appropriate cards in the modUnionTable, so that | |
1114 // this object gets scanned before the sweep. If this is | |
1115 // not done, CMS generation references in the object might | |
1116 // not get marked. | |
1117 // For the case of arrays, which are otherwise precisely | |
1118 // marked, we need to dirty the entire array, not just its head. | |
1119 if (is_obj_array) { | |
1120 // The [par_]mark_range() method expects mr.end() below to | |
1121 // be aligned to the granularity of a bit's representation | |
1122 // in the heap. In the case of the MUT below, that's a | |
1123 // card size. | |
1124 MemRegion mr(start, | |
1125 (HeapWord*)round_to((intptr_t)(start + obj_size), | |
1126 CardTableModRefBS::card_size /* bytes */)); | |
1127 if (par) { | |
1128 _modUnionTable.par_mark_range(mr); | |
1129 } else { | |
1130 _modUnionTable.mark_range(mr); | |
1131 } | |
1132 } else { // not an obj array; we can just mark the head | |
1133 if (par) { | |
1134 _modUnionTable.par_mark(start); | |
1135 } else { | |
1136 _modUnionTable.mark(start); | |
1137 } | |
1138 } | |
1139 } | |
1140 } | |
1141 } | |
1142 | |
1143 static inline size_t percent_of_space(Space* space, HeapWord* addr) | |
1144 { | |
1145 size_t delta = pointer_delta(addr, space->bottom()); | |
1146 return (size_t)(delta * 100.0 / (space->capacity() / HeapWordSize)); | |
1147 } | |
1148 | |
1149 void CMSCollector::icms_update_allocation_limits() | |
1150 { | |
1151 Generation* gen0 = GenCollectedHeap::heap()->get_gen(0); | |
1152 EdenSpace* eden = gen0->as_DefNewGeneration()->eden(); | |
1153 | |
1154 const unsigned int duty_cycle = stats().icms_update_duty_cycle(); | |
1155 if (CMSTraceIncrementalPacing) { | |
1156 stats().print(); | |
1157 } | |
1158 | |
1159 assert(duty_cycle <= 100, "invalid duty cycle"); | |
1160 if (duty_cycle != 0) { | |
1161 // The duty_cycle is a percentage between 0 and 100; convert to words and | |
1162 // then compute the offset from the endpoints of the space. | |
1163 size_t free_words = eden->free() / HeapWordSize; | |
1164 double free_words_dbl = (double)free_words; | |
1165 size_t duty_cycle_words = (size_t)(free_words_dbl * duty_cycle / 100.0); | |
1166 size_t offset_words = (free_words - duty_cycle_words) / 2; | |
1167 | |
1168 _icms_start_limit = eden->top() + offset_words; | |
1169 _icms_stop_limit = eden->end() - offset_words; | |
1170 | |
1171 // The limits may be adjusted (shifted to the right) by | |
1172 // CMSIncrementalOffset, to allow the application more mutator time after a | |
1173 // young gen gc (when all mutators were stopped) and before CMS starts and | |
1174 // takes away one or more cpus. | |
1175 if (CMSIncrementalOffset != 0) { | |
1176 double adjustment_dbl = free_words_dbl * CMSIncrementalOffset / 100.0; | |
1177 size_t adjustment = (size_t)adjustment_dbl; | |
1178 HeapWord* tmp_stop = _icms_stop_limit + adjustment; | |
1179 if (tmp_stop > _icms_stop_limit && tmp_stop < eden->end()) { | |
1180 _icms_start_limit += adjustment; | |
1181 _icms_stop_limit = tmp_stop; | |
1182 } | |
1183 } | |
1184 } | |
1185 if (duty_cycle == 0 || (_icms_start_limit == _icms_stop_limit)) { | |
1186 _icms_start_limit = _icms_stop_limit = eden->end(); | |
1187 } | |
1188 | |
1189 // Install the new start limit. | |
1190 eden->set_soft_end(_icms_start_limit); | |
1191 | |
1192 if (CMSTraceIncrementalMode) { | |
1193 gclog_or_tty->print(" icms alloc limits: " | |
1194 PTR_FORMAT "," PTR_FORMAT | |
1195 " (" SIZE_FORMAT "%%," SIZE_FORMAT "%%) ", | |
1196 _icms_start_limit, _icms_stop_limit, | |
1197 percent_of_space(eden, _icms_start_limit), | |
1198 percent_of_space(eden, _icms_stop_limit)); | |
1199 if (Verbose) { | |
1200 gclog_or_tty->print("eden: "); | |
1201 eden->print_on(gclog_or_tty); | |
1202 } | |
1203 } | |
1204 } | |
1205 | |
1206 // Any changes here should try to maintain the invariant | |
1207 // that if this method is called with _icms_start_limit | |
1208 // and _icms_stop_limit both NULL, then it should return NULL | |
1209 // and not notify the icms thread. | |
1210 HeapWord* | |
1211 CMSCollector::allocation_limit_reached(Space* space, HeapWord* top, | |
1212 size_t word_size) | |
1213 { | |
1214 // A start_limit equal to end() means the duty cycle is 0, so treat that as a | |
1215 // nop. | |
1216 if (CMSIncrementalMode && _icms_start_limit != space->end()) { | |
1217 if (top <= _icms_start_limit) { | |
1218 if (CMSTraceIncrementalMode) { | |
1219 space->print_on(gclog_or_tty); | |
1220 gclog_or_tty->stamp(); | |
1221 gclog_or_tty->print_cr(" start limit top=" PTR_FORMAT | |
1222 ", new limit=" PTR_FORMAT | |
1223 " (" SIZE_FORMAT "%%)", | |
1224 top, _icms_stop_limit, | |
1225 percent_of_space(space, _icms_stop_limit)); | |
1226 } | |
1227 ConcurrentMarkSweepThread::start_icms(); | |
1228 assert(top < _icms_stop_limit, "Tautology"); | |
1229 if (word_size < pointer_delta(_icms_stop_limit, top)) { | |
1230 return _icms_stop_limit; | |
1231 } | |
1232 | |
1233 // The allocation will cross both the _start and _stop limits, so do the | |
1234 // stop notification also and return end(). | |
1235 if (CMSTraceIncrementalMode) { | |
1236 space->print_on(gclog_or_tty); | |
1237 gclog_or_tty->stamp(); | |
1238 gclog_or_tty->print_cr(" +stop limit top=" PTR_FORMAT | |
1239 ", new limit=" PTR_FORMAT | |
1240 " (" SIZE_FORMAT "%%)", | |
1241 top, space->end(), | |
1242 percent_of_space(space, space->end())); | |
1243 } | |
1244 ConcurrentMarkSweepThread::stop_icms(); | |
1245 return space->end(); | |
1246 } | |
1247 | |
1248 if (top <= _icms_stop_limit) { | |
1249 if (CMSTraceIncrementalMode) { | |
1250 space->print_on(gclog_or_tty); | |
1251 gclog_or_tty->stamp(); | |
1252 gclog_or_tty->print_cr(" stop limit top=" PTR_FORMAT | |
1253 ", new limit=" PTR_FORMAT | |
1254 " (" SIZE_FORMAT "%%)", | |
1255 top, space->end(), | |
1256 percent_of_space(space, space->end())); | |
1257 } | |
1258 ConcurrentMarkSweepThread::stop_icms(); | |
1259 return space->end(); | |
1260 } | |
1261 | |
1262 if (CMSTraceIncrementalMode) { | |
1263 space->print_on(gclog_or_tty); | |
1264 gclog_or_tty->stamp(); | |
1265 gclog_or_tty->print_cr(" end limit top=" PTR_FORMAT | |
1266 ", new limit=" PTR_FORMAT, | |
1267 top, NULL); | |
1268 } | |
1269 } | |
1270 | |
1271 return NULL; | |
1272 } | |
1273 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
1274 oop ConcurrentMarkSweepGeneration::promote(oop obj, size_t obj_size) { |
0 | 1275 assert(obj_size == (size_t)obj->size(), "bad obj_size passed in"); |
1276 // allocate, copy and if necessary update promoinfo -- | |
1277 // delegate to underlying space. | |
1278 assert_lock_strong(freelistLock()); | |
1279 | |
1280 #ifndef PRODUCT | |
1281 if (Universe::heap()->promotion_should_fail()) { | |
1282 return NULL; | |
1283 } | |
1284 #endif // #ifndef PRODUCT | |
1285 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
1286 oop res = _cmsSpace->promote(obj, obj_size); |
0 | 1287 if (res == NULL) { |
1288 // expand and retry | |
1289 size_t s = _cmsSpace->expansionSpaceRequired(obj_size); // HeapWords | |
1290 expand(s*HeapWordSize, MinHeapDeltaBytes, | |
1291 CMSExpansionCause::_satisfy_promotion); | |
1292 // Since there's currently no next generation, we don't try to promote | |
1293 // into a more senior generation. | |
1294 assert(next_gen() == NULL, "assumption, based upon which no attempt " | |
1295 "is made to pass on a possibly failing " | |
1296 "promotion to next generation"); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
1297 res = _cmsSpace->promote(obj, obj_size); |
0 | 1298 } |
1299 if (res != NULL) { | |
1300 // See comment in allocate() about when objects should | |
1301 // be allocated live. | |
1302 assert(obj->is_oop(), "Will dereference klass pointer below"); | |
1303 collector()->promoted(false, // Not parallel | |
1304 (HeapWord*)res, obj->is_objArray(), obj_size); | |
1305 // promotion counters | |
1306 NOT_PRODUCT( | |
1307 _numObjectsPromoted++; | |
1308 _numWordsPromoted += | |
1309 (int)(CompactibleFreeListSpace::adjustObjectSize(obj->size())); | |
1310 ) | |
1311 } | |
1312 return res; | |
1313 } | |
1314 | |
1315 | |
1316 HeapWord* | |
1317 ConcurrentMarkSweepGeneration::allocation_limit_reached(Space* space, | |
1318 HeapWord* top, | |
1319 size_t word_sz) | |
1320 { | |
1321 return collector()->allocation_limit_reached(space, top, word_sz); | |
1322 } | |
1323 | |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1324 // IMPORTANT: Notes on object size recognition in CMS. |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1325 // --------------------------------------------------- |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1326 // A block of storage in the CMS generation is always in |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1327 // one of three states. A free block (FREE), an allocated |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1328 // object (OBJECT) whose size() method reports the correct size, |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1329 // and an intermediate state (TRANSIENT) in which its size cannot |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1330 // be accurately determined. |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1331 // STATE IDENTIFICATION: (32 bit and 64 bit w/o COOPS) |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1332 // ----------------------------------------------------- |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1333 // FREE: klass_word & 1 == 1; mark_word holds block size |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1334 // |
1777
179464550c7d
6983930: CMS: Various small cleanups ca September 2010
ysr
parents:
1720
diff
changeset
|
1335 // OBJECT: klass_word installed; klass_word != 0 && klass_word & 1 == 0; |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1336 // obj->size() computes correct size |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1337 // [Perm Gen objects needs to be "parsable" before they can be navigated] |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1338 // |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1339 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1340 // |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1341 // STATE IDENTIFICATION: (64 bit+COOPS) |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1342 // ------------------------------------ |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1343 // FREE: mark_word & CMS_FREE_BIT == 1; mark_word & ~CMS_FREE_BIT gives block_size |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1344 // |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1345 // OBJECT: klass_word installed; klass_word != 0; |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1346 // obj->size() computes correct size |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1347 // [Perm Gen comment above continues to hold] |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1348 // |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1349 // TRANSIENT: klass_word == 0; size is indeterminate until we become an OBJECT |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1350 // |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1351 // |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1352 // STATE TRANSITION DIAGRAM |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1353 // |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1354 // mut / parnew mut / parnew |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1355 // FREE --------------------> TRANSIENT ---------------------> OBJECT --| |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1356 // ^ | |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1357 // |------------------------ DEAD <------------------------------------| |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1358 // sweep mut |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1359 // |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1360 // While a block is in TRANSIENT state its size cannot be determined |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1361 // so readers will either need to come back later or stall until |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1362 // the size can be determined. Note that for the case of direct |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1363 // allocation, P-bits, when available, may be used to determine the |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1364 // size of an object that may not yet have been initialized. |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1365 |
0 | 1366 // Things to support parallel young-gen collection. |
1367 oop | |
1368 ConcurrentMarkSweepGeneration::par_promote(int thread_num, | |
1369 oop old, markOop m, | |
1370 size_t word_sz) { | |
1371 #ifndef PRODUCT | |
1372 if (Universe::heap()->promotion_should_fail()) { | |
1373 return NULL; | |
1374 } | |
1375 #endif // #ifndef PRODUCT | |
1376 | |
1377 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; | |
1378 PromotionInfo* promoInfo = &ps->promo; | |
1379 // if we are tracking promotions, then first ensure space for | |
1380 // promotion (including spooling space for saving header if necessary). | |
1381 // then allocate and copy, then track promoted info if needed. | |
1382 // When tracking (see PromotionInfo::track()), the mark word may | |
1383 // be displaced and in this case restoration of the mark word | |
1384 // occurs in the (oop_since_save_marks_)iterate phase. | |
1385 if (promoInfo->tracking() && !promoInfo->ensure_spooling_space()) { | |
1386 // Out of space for allocating spooling buffers; | |
1387 // try expanding and allocating spooling buffers. | |
1388 if (!expand_and_ensure_spooling_space(promoInfo)) { | |
1389 return NULL; | |
1390 } | |
1391 } | |
1392 assert(promoInfo->has_spooling_space(), "Control point invariant"); | |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1393 const size_t alloc_sz = CompactibleFreeListSpace::adjustObjectSize(word_sz); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1394 HeapWord* obj_ptr = ps->lab.alloc(alloc_sz); |
0 | 1395 if (obj_ptr == NULL) { |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1396 obj_ptr = expand_and_par_lab_allocate(ps, alloc_sz); |
0 | 1397 if (obj_ptr == NULL) { |
1398 return NULL; | |
1399 } | |
1400 } | |
1401 oop obj = oop(obj_ptr); | |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1402 OrderAccess::storestore(); |
187 | 1403 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1404 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size"); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1405 // IMPORTANT: See note on object initialization for CMS above. |
0 | 1406 // Otherwise, copy the object. Here we must be careful to insert the |
1407 // klass pointer last, since this marks the block as an allocated object. | |
187 | 1408 // Except with compressed oops it's the mark word. |
0 | 1409 HeapWord* old_ptr = (HeapWord*)old; |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1410 // Restore the mark word copied above. |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1411 obj->set_mark(m); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1412 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1413 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size"); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1414 OrderAccess::storestore(); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1415 |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1416 if (UseCompressedOops) { |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1417 // Copy gap missed by (aligned) header size calculation below |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1418 obj->set_klass_gap(old->klass_gap()); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1419 } |
0 | 1420 if (word_sz > (size_t)oopDesc::header_size()) { |
1421 Copy::aligned_disjoint_words(old_ptr + oopDesc::header_size(), | |
1422 obj_ptr + oopDesc::header_size(), | |
1423 word_sz - oopDesc::header_size()); | |
1424 } | |
187 | 1425 |
0 | 1426 // Now we can track the promoted object, if necessary. We take care |
1521 | 1427 // to delay the transition from uninitialized to full object |
0 | 1428 // (i.e., insertion of klass pointer) until after, so that it |
1429 // atomically becomes a promoted object. | |
1430 if (promoInfo->tracking()) { | |
1431 promoInfo->track((PromotedObject*)obj, old->klass()); | |
1432 } | |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1433 assert(obj->klass_or_null() == NULL, "Object should be uninitialized here."); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1434 assert(!((FreeChunk*)obj_ptr)->isFree(), "Error, block will look free but show wrong size"); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1435 assert(old->is_oop(), "Will use and dereference old klass ptr below"); |
187 | 1436 |
1437 // Finally, install the klass pointer (this should be volatile). | |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1438 OrderAccess::storestore(); |
0 | 1439 obj->set_klass(old->klass()); |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1440 // We should now be able to calculate the right size for this object |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1441 assert(obj->is_oop() && obj->size() == (int)word_sz, "Error, incorrect size computed for promoted object"); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1442 |
0 | 1443 collector()->promoted(true, // parallel |
1444 obj_ptr, old->is_objArray(), word_sz); | |
1445 | |
1446 NOT_PRODUCT( | |
1716
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1447 Atomic::inc_ptr(&_numObjectsPromoted); |
be3f9c242c9d
6948538: CMS: BOT walkers can fall into object allocation and initialization cracks
ysr
parents:
1704
diff
changeset
|
1448 Atomic::add_ptr(alloc_sz, &_numWordsPromoted); |
0 | 1449 ) |
1450 | |
1451 return obj; | |
1452 } | |
1453 | |
1454 void | |
1455 ConcurrentMarkSweepGeneration:: | |
1456 par_promote_alloc_undo(int thread_num, | |
1457 HeapWord* obj, size_t word_sz) { | |
1458 // CMS does not support promotion undo. | |
1459 ShouldNotReachHere(); | |
1460 } | |
1461 | |
1462 void | |
1463 ConcurrentMarkSweepGeneration:: | |
1464 par_promote_alloc_done(int thread_num) { | |
1465 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
1466 ps->lab.retire(thread_num); |
0 | 1467 } |
1468 | |
1469 void | |
1470 ConcurrentMarkSweepGeneration:: | |
1471 par_oop_since_save_marks_iterate_done(int thread_num) { | |
1472 CMSParGCThreadState* ps = _par_gc_thread_states[thread_num]; | |
1473 ParScanWithoutBarrierClosure* dummy_cl = NULL; | |
1474 ps->promo.promoted_oops_iterate_nv(dummy_cl); | |
1475 } | |
1476 | |
1477 // XXXPERM | |
1478 bool ConcurrentMarkSweepGeneration::should_collect(bool full, | |
1479 size_t size, | |
1480 bool tlab) | |
1481 { | |
1482 // We allow a STW collection only if a full | |
1483 // collection was requested. | |
1484 return full || should_allocate(size, tlab); // FIX ME !!! | |
1485 // This and promotion failure handling are connected at the | |
1486 // hip and should be fixed by untying them. | |
1487 } | |
1488 | |
1489 bool CMSCollector::shouldConcurrentCollect() { | |
1490 if (_full_gc_requested) { | |
1491 if (Verbose && PrintGCDetails) { | |
1492 gclog_or_tty->print_cr("CMSCollector: collect because of explicit " | |
1520
bb843ebc7c55
6919638: CMS: ExplicitGCInvokesConcurrent misinteracts with gc locker
ysr
parents:
1518
diff
changeset
|
1493 " gc request (or gc_locker)"); |
0 | 1494 } |
1495 return true; | |
1496 } | |
1497 | |
1498 // For debugging purposes, change the type of collection. | |
1499 // If the rotation is not on the concurrent collection | |
1500 // type, don't start a concurrent collection. | |
1501 NOT_PRODUCT( | |
1502 if (RotateCMSCollectionTypes && | |
1503 (_cmsGen->debug_collection_type() != | |
1504 ConcurrentMarkSweepGeneration::Concurrent_collection_type)) { | |
1505 assert(_cmsGen->debug_collection_type() != | |
1506 ConcurrentMarkSweepGeneration::Unknown_collection_type, | |
1507 "Bad cms collection type"); | |
1508 return false; | |
1509 } | |
1510 ) | |
1511 | |
1512 FreelistLocker x(this); | |
1513 // ------------------------------------------------------------------ | |
1514 // Print out lots of information which affects the initiation of | |
1515 // a collection. | |
1516 if (PrintCMSInitiationStatistics && stats().valid()) { | |
1517 gclog_or_tty->print("CMSCollector shouldConcurrentCollect: "); | |
1518 gclog_or_tty->stamp(); | |
1519 gclog_or_tty->print_cr(""); | |
1520 stats().print_on(gclog_or_tty); | |
1521 gclog_or_tty->print_cr("time_until_cms_gen_full %3.7f", | |
1522 stats().time_until_cms_gen_full()); | |
1523 gclog_or_tty->print_cr("free="SIZE_FORMAT, _cmsGen->free()); | |
1524 gclog_or_tty->print_cr("contiguous_available="SIZE_FORMAT, | |
1525 _cmsGen->contiguous_available()); | |
1526 gclog_or_tty->print_cr("promotion_rate=%g", stats().promotion_rate()); | |
1527 gclog_or_tty->print_cr("cms_allocation_rate=%g", stats().cms_allocation_rate()); | |
1528 gclog_or_tty->print_cr("occupancy=%3.7f", _cmsGen->occupancy()); | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1529 gclog_or_tty->print_cr("initiatingOccupancy=%3.7f", _cmsGen->initiating_occupancy()); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1530 gclog_or_tty->print_cr("initiatingPermOccupancy=%3.7f", _permGen->initiating_occupancy()); |
0 | 1531 } |
1532 // ------------------------------------------------------------------ | |
1533 | |
1534 // If the estimated time to complete a cms collection (cms_duration()) | |
1535 // is less than the estimated time remaining until the cms generation | |
1536 // is full, start a collection. | |
1537 if (!UseCMSInitiatingOccupancyOnly) { | |
1538 if (stats().valid()) { | |
1539 if (stats().time_until_cms_start() == 0.0) { | |
1540 return true; | |
1541 } | |
1542 } else { | |
1543 // We want to conservatively collect somewhat early in order | |
1544 // to try and "bootstrap" our CMS/promotion statistics; | |
1545 // this branch will not fire after the first successful CMS | |
1546 // collection because the stats should then be valid. | |
1547 if (_cmsGen->occupancy() >= _bootstrap_occupancy) { | |
1548 if (Verbose && PrintGCDetails) { | |
1549 gclog_or_tty->print_cr( | |
1550 " CMSCollector: collect for bootstrapping statistics:" | |
1551 " occupancy = %f, boot occupancy = %f", _cmsGen->occupancy(), | |
1552 _bootstrap_occupancy); | |
1553 } | |
1554 return true; | |
1555 } | |
1556 } | |
1557 } | |
1558 | |
1559 // Otherwise, we start a collection cycle if either the perm gen or | |
1560 // old gen want a collection cycle started. Each may use | |
1561 // an appropriate criterion for making this decision. | |
1562 // XXX We need to make sure that the gen expansion | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1563 // criterion dovetails well with this. XXX NEED TO FIX THIS |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1564 if (_cmsGen->should_concurrent_collect()) { |
0 | 1565 if (Verbose && PrintGCDetails) { |
1566 gclog_or_tty->print_cr("CMS old gen initiated"); | |
1567 } | |
1568 return true; | |
1569 } | |
1570 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1571 // We start a collection if we believe an incremental collection may fail; |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1572 // this is not likely to be productive in practice because it's probably too |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1573 // late anyway. |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1574 GenCollectedHeap* gch = GenCollectedHeap::heap(); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1575 assert(gch->collector_policy()->is_two_generation_policy(), |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1576 "You may want to check the correctness of the following"); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1577 if (gch->incremental_collection_will_fail()) { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1578 if (PrintGCDetails && Verbose) { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1579 gclog_or_tty->print("CMSCollector: collect because incremental collection will fail "); |
0 | 1580 } |
1581 return true; | |
1582 } | |
1583 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1584 if (CMSClassUnloadingEnabled && _permGen->should_concurrent_collect()) { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1585 bool res = update_should_unload_classes(); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1586 if (res) { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1587 if (Verbose && PrintGCDetails) { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1588 gclog_or_tty->print_cr("CMS perm gen initiated"); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1589 } |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1590 return true; |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1591 } |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1592 } |
0 | 1593 return false; |
1594 } | |
1595 | |
1596 // Clear _expansion_cause fields of constituent generations | |
1597 void CMSCollector::clear_expansion_cause() { | |
1598 _cmsGen->clear_expansion_cause(); | |
1599 _permGen->clear_expansion_cause(); | |
1600 } | |
1601 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1602 // We should be conservative in starting a collection cycle. To |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1603 // start too eagerly runs the risk of collecting too often in the |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1604 // extreme. To collect too rarely falls back on full collections, |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1605 // which works, even if not optimum in terms of concurrent work. |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1606 // As a work around for too eagerly collecting, use the flag |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1607 // UseCMSInitiatingOccupancyOnly. This also has the advantage of |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1608 // giving the user an easily understandable way of controlling the |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1609 // collections. |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1610 // We want to start a new collection cycle if any of the following |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1611 // conditions hold: |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1612 // . our current occupancy exceeds the configured initiating occupancy |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1613 // for this generation, or |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1614 // . we recently needed to expand this space and have not, since that |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1615 // expansion, done a collection of this generation, or |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1616 // . the underlying space believes that it may be a good idea to initiate |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1617 // a concurrent collection (this may be based on criteria such as the |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1618 // following: the space uses linear allocation and linear allocation is |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1619 // going to fail, or there is believed to be excessive fragmentation in |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1620 // the generation, etc... or ... |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1621 // [.(currently done by CMSCollector::shouldConcurrentCollect() only for |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1622 // the case of the old generation, not the perm generation; see CR 6543076): |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1623 // we may be approaching a point at which allocation requests may fail because |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1624 // we will be out of sufficient free space given allocation rate estimates.] |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1625 bool ConcurrentMarkSweepGeneration::should_concurrent_collect() const { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1626 |
0 | 1627 assert_lock_strong(freelistLock()); |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1628 if (occupancy() > initiating_occupancy()) { |
0 | 1629 if (PrintGCDetails && Verbose) { |
1630 gclog_or_tty->print(" %s: collect because of occupancy %f / %f ", | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1631 short_name(), occupancy(), initiating_occupancy()); |
0 | 1632 } |
1633 return true; | |
1634 } | |
1635 if (UseCMSInitiatingOccupancyOnly) { | |
1636 return false; | |
1637 } | |
1638 if (expansion_cause() == CMSExpansionCause::_satisfy_allocation) { | |
1639 if (PrintGCDetails && Verbose) { | |
1640 gclog_or_tty->print(" %s: collect because expanded for allocation ", | |
1641 short_name()); | |
1642 } | |
1643 return true; | |
1644 } | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1645 if (_cmsSpace->should_concurrent_collect()) { |
0 | 1646 if (PrintGCDetails && Verbose) { |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
1647 gclog_or_tty->print(" %s: collect because cmsSpace says so ", |
0 | 1648 short_name()); |
1649 } | |
1650 return true; | |
1651 } | |
1652 return false; | |
1653 } | |
1654 | |
1655 void ConcurrentMarkSweepGeneration::collect(bool full, | |
1656 bool clear_all_soft_refs, | |
1657 size_t size, | |
1658 bool tlab) | |
1659 { | |
1660 collector()->collect(full, clear_all_soft_refs, size, tlab); | |
1661 } | |
1662 | |
1663 void CMSCollector::collect(bool full, | |
1664 bool clear_all_soft_refs, | |
1665 size_t size, | |
1666 bool tlab) | |
1667 { | |
1668 if (!UseCMSCollectionPassing && _collectorState > Idling) { | |
1669 // For debugging purposes skip the collection if the state | |
1670 // is not currently idle | |
1671 if (TraceCMSState) { | |
1672 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " skipped full:%d CMS state %d", | |
1673 Thread::current(), full, _collectorState); | |
1674 } | |
1675 return; | |
1676 } | |
1677 | |
1678 // The following "if" branch is present for defensive reasons. | |
1679 // In the current uses of this interface, it can be replaced with: | |
1680 // assert(!GC_locker.is_active(), "Can't be called otherwise"); | |
1681 // But I am not placing that assert here to allow future | |
1682 // generality in invoking this interface. | |
1683 if (GC_locker::is_active()) { | |
1684 // A consistency test for GC_locker | |
1685 assert(GC_locker::needs_gc(), "Should have been set already"); | |
1686 // Skip this foreground collection, instead | |
1687 // expanding the heap if necessary. | |
1688 // Need the free list locks for the call to free() in compute_new_size() | |
1689 compute_new_size(); | |
1690 return; | |
1691 } | |
1692 acquire_control_and_collect(full, clear_all_soft_refs); | |
1693 _full_gcs_since_conc_gc++; | |
1694 | |
1695 } | |
1696 | |
1697 void CMSCollector::request_full_gc(unsigned int full_gc_count) { | |
1698 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
1699 unsigned int gc_count = gch->total_full_collections(); | |
1700 if (gc_count == full_gc_count) { | |
1701 MutexLockerEx y(CGC_lock, Mutex::_no_safepoint_check_flag); | |
1702 _full_gc_requested = true; | |
1703 CGC_lock->notify(); // nudge CMS thread | |
1704 } | |
1705 } | |
1706 | |
1707 | |
1708 // The foreground and background collectors need to coordinate in order | |
1709 // to make sure that they do not mutually interfere with CMS collections. | |
1710 // When a background collection is active, | |
1711 // the foreground collector may need to take over (preempt) and | |
1712 // synchronously complete an ongoing collection. Depending on the | |
1713 // frequency of the background collections and the heap usage | |
1714 // of the application, this preemption can be seldom or frequent. | |
1715 // There are only certain | |
1716 // points in the background collection that the "collection-baton" | |
1717 // can be passed to the foreground collector. | |
1718 // | |
1719 // The foreground collector will wait for the baton before | |
1720 // starting any part of the collection. The foreground collector | |
1721 // will only wait at one location. | |
1722 // | |
1723 // The background collector will yield the baton before starting a new | |
1724 // phase of the collection (e.g., before initial marking, marking from roots, | |
1725 // precleaning, final re-mark, sweep etc.) This is normally done at the head | |
1726 // of the loop which switches the phases. The background collector does some | |
1727 // of the phases (initial mark, final re-mark) with the world stopped. | |
1728 // Because of locking involved in stopping the world, | |
1729 // the foreground collector should not block waiting for the background | |
1730 // collector when it is doing a stop-the-world phase. The background | |
1731 // collector will yield the baton at an additional point just before | |
1732 // it enters a stop-the-world phase. Once the world is stopped, the | |
1733 // background collector checks the phase of the collection. If the | |
1734 // phase has not changed, it proceeds with the collection. If the | |
1735 // phase has changed, it skips that phase of the collection. See | |
1736 // the comments on the use of the Heap_lock in collect_in_background(). | |
1737 // | |
1738 // Variable used in baton passing. | |
1739 // _foregroundGCIsActive - Set to true by the foreground collector when | |
1740 // it wants the baton. The foreground clears it when it has finished | |
1741 // the collection. | |
1742 // _foregroundGCShouldWait - Set to true by the background collector | |
1743 // when it is running. The foreground collector waits while | |
1744 // _foregroundGCShouldWait is true. | |
1745 // CGC_lock - monitor used to protect access to the above variables | |
1746 // and to notify the foreground and background collectors. | |
1747 // _collectorState - current state of the CMS collection. | |
1748 // | |
1749 // The foreground collector | |
1750 // acquires the CGC_lock | |
1751 // sets _foregroundGCIsActive | |
1752 // waits on the CGC_lock for _foregroundGCShouldWait to be false | |
1753 // various locks acquired in preparation for the collection | |
1754 // are released so as not to block the background collector | |
1755 // that is in the midst of a collection | |
1756 // proceeds with the collection | |
1757 // clears _foregroundGCIsActive | |
1758 // returns | |
1759 // | |
1760 // The background collector in a loop iterating on the phases of the | |
1761 // collection | |
1762 // acquires the CGC_lock | |
1763 // sets _foregroundGCShouldWait | |
1764 // if _foregroundGCIsActive is set | |
1765 // clears _foregroundGCShouldWait, notifies _CGC_lock | |
1766 // waits on _CGC_lock for _foregroundGCIsActive to become false | |
1767 // and exits the loop. | |
1768 // otherwise | |
1769 // proceed with that phase of the collection | |
1770 // if the phase is a stop-the-world phase, | |
1771 // yield the baton once more just before enqueueing | |
1772 // the stop-world CMS operation (executed by the VM thread). | |
1773 // returns after all phases of the collection are done | |
1774 // | |
1775 | |
1776 void CMSCollector::acquire_control_and_collect(bool full, | |
1777 bool clear_all_soft_refs) { | |
1778 assert(SafepointSynchronize::is_at_safepoint(), "should be at safepoint"); | |
1779 assert(!Thread::current()->is_ConcurrentGC_thread(), | |
1780 "shouldn't try to acquire control from self!"); | |
1781 | |
1782 // Start the protocol for acquiring control of the | |
1783 // collection from the background collector (aka CMS thread). | |
1784 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), | |
1785 "VM thread should have CMS token"); | |
1786 // Remember the possibly interrupted state of an ongoing | |
1787 // concurrent collection | |
1788 CollectorState first_state = _collectorState; | |
1789 | |
1790 // Signal to a possibly ongoing concurrent collection that | |
1791 // we want to do a foreground collection. | |
1792 _foregroundGCIsActive = true; | |
1793 | |
1794 // Disable incremental mode during a foreground collection. | |
1795 ICMSDisabler icms_disabler; | |
1796 | |
1797 // release locks and wait for a notify from the background collector | |
1798 // releasing the locks in only necessary for phases which | |
1799 // do yields to improve the granularity of the collection. | |
1800 assert_lock_strong(bitMapLock()); | |
1801 // We need to lock the Free list lock for the space that we are | |
1802 // currently collecting. | |
1803 assert(haveFreelistLocks(), "Must be holding free list locks"); | |
1804 bitMapLock()->unlock(); | |
1805 releaseFreelistLocks(); | |
1806 { | |
1807 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); | |
1808 if (_foregroundGCShouldWait) { | |
1809 // We are going to be waiting for action for the CMS thread; | |
1810 // it had better not be gone (for instance at shutdown)! | |
1811 assert(ConcurrentMarkSweepThread::cmst() != NULL, | |
1812 "CMS thread must be running"); | |
1813 // Wait here until the background collector gives us the go-ahead | |
1814 ConcurrentMarkSweepThread::clear_CMS_flag( | |
1815 ConcurrentMarkSweepThread::CMS_vm_has_token); // release token | |
1816 // Get a possibly blocked CMS thread going: | |
1817 // Note that we set _foregroundGCIsActive true above, | |
1818 // without protection of the CGC_lock. | |
1819 CGC_lock->notify(); | |
1820 assert(!ConcurrentMarkSweepThread::vm_thread_wants_cms_token(), | |
1821 "Possible deadlock"); | |
1822 while (_foregroundGCShouldWait) { | |
1823 // wait for notification | |
1824 CGC_lock->wait(Mutex::_no_safepoint_check_flag); | |
1825 // Possibility of delay/starvation here, since CMS token does | |
1826 // not know to give priority to VM thread? Actually, i think | |
1827 // there wouldn't be any delay/starvation, but the proof of | |
1828 // that "fact" (?) appears non-trivial. XXX 20011219YSR | |
1829 } | |
1830 ConcurrentMarkSweepThread::set_CMS_flag( | |
1831 ConcurrentMarkSweepThread::CMS_vm_has_token); | |
1832 } | |
1833 } | |
1834 // The CMS_token is already held. Get back the other locks. | |
1835 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), | |
1836 "VM thread should have CMS token"); | |
1837 getFreelistLocks(); | |
1838 bitMapLock()->lock_without_safepoint_check(); | |
1839 if (TraceCMSState) { | |
1840 gclog_or_tty->print_cr("CMS foreground collector has asked for control " | |
1841 INTPTR_FORMAT " with first state %d", Thread::current(), first_state); | |
1842 gclog_or_tty->print_cr(" gets control with state %d", _collectorState); | |
1843 } | |
1844 | |
1845 // Check if we need to do a compaction, or if not, whether | |
1846 // we need to start the mark-sweep from scratch. | |
1847 bool should_compact = false; | |
1848 bool should_start_over = false; | |
1849 decide_foreground_collection_type(clear_all_soft_refs, | |
1850 &should_compact, &should_start_over); | |
1851 | |
1852 NOT_PRODUCT( | |
1853 if (RotateCMSCollectionTypes) { | |
1854 if (_cmsGen->debug_collection_type() == | |
1855 ConcurrentMarkSweepGeneration::MSC_foreground_collection_type) { | |
1856 should_compact = true; | |
1857 } else if (_cmsGen->debug_collection_type() == | |
1858 ConcurrentMarkSweepGeneration::MS_foreground_collection_type) { | |
1859 should_compact = false; | |
1860 } | |
1861 } | |
1862 ) | |
1863 | |
1864 if (PrintGCDetails && first_state > Idling) { | |
1865 GCCause::Cause cause = GenCollectedHeap::heap()->gc_cause(); | |
1866 if (GCCause::is_user_requested_gc(cause) || | |
1867 GCCause::is_serviceability_requested_gc(cause)) { | |
1868 gclog_or_tty->print(" (concurrent mode interrupted)"); | |
1869 } else { | |
1870 gclog_or_tty->print(" (concurrent mode failure)"); | |
1871 } | |
1872 } | |
1873 | |
1874 if (should_compact) { | |
1875 // If the collection is being acquired from the background | |
1876 // collector, there may be references on the discovered | |
1877 // references lists that have NULL referents (being those | |
1878 // that were concurrently cleared by a mutator) or | |
1879 // that are no longer active (having been enqueued concurrently | |
1880 // by the mutator). | |
1881 // Scrub the list of those references because Mark-Sweep-Compact | |
1882 // code assumes referents are not NULL and that all discovered | |
1883 // Reference objects are active. | |
1884 ref_processor()->clean_up_discovered_references(); | |
1885 | |
1886 do_compaction_work(clear_all_soft_refs); | |
1887 | |
1888 // Has the GC time limit been exceeded? | |
1387
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1889 DefNewGeneration* young_gen = _young_gen->as_DefNewGeneration(); |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1890 size_t max_eden_size = young_gen->max_capacity() - |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1891 young_gen->to()->capacity() - |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1892 young_gen->from()->capacity(); |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1893 GenCollectedHeap* gch = GenCollectedHeap::heap(); |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1894 GCCause::Cause gc_cause = gch->gc_cause(); |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1895 size_policy()->check_gc_overhead_limit(_young_gen->used(), |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1896 young_gen->eden()->used(), |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1897 _cmsGen->max_capacity(), |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1898 max_eden_size, |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1899 full, |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1900 gc_cause, |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
1901 gch->collector_policy()); |
0 | 1902 } else { |
1903 do_mark_sweep_work(clear_all_soft_refs, first_state, | |
1904 should_start_over); | |
1905 } | |
1906 // Reset the expansion cause, now that we just completed | |
1907 // a collection cycle. | |
1908 clear_expansion_cause(); | |
1909 _foregroundGCIsActive = false; | |
1910 return; | |
1911 } | |
1912 | |
1913 // Resize the perm generation and the tenured generation | |
1914 // after obtaining the free list locks for the | |
1915 // two generations. | |
1916 void CMSCollector::compute_new_size() { | |
1917 assert_locked_or_safepoint(Heap_lock); | |
1918 FreelistLocker z(this); | |
1919 _permGen->compute_new_size(); | |
1920 _cmsGen->compute_new_size(); | |
1921 } | |
1922 | |
1923 // A work method used by foreground collection to determine | |
1924 // what type of collection (compacting or not, continuing or fresh) | |
1925 // it should do. | |
1926 // NOTE: the intent is to make UseCMSCompactAtFullCollection | |
1927 // and CMSCompactWhenClearAllSoftRefs the default in the future | |
1928 // and do away with the flags after a suitable period. | |
1929 void CMSCollector::decide_foreground_collection_type( | |
1930 bool clear_all_soft_refs, bool* should_compact, | |
1931 bool* should_start_over) { | |
1932 // Normally, we'll compact only if the UseCMSCompactAtFullCollection | |
1933 // flag is set, and we have either requested a System.gc() or | |
1934 // the number of full gc's since the last concurrent cycle | |
1935 // has exceeded the threshold set by CMSFullGCsBeforeCompaction, | |
1936 // or if an incremental collection has failed | |
1937 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
1938 assert(gch->collector_policy()->is_two_generation_policy(), | |
1939 "You may want to check the correctness of the following"); | |
1940 // Inform cms gen if this was due to partial collection failing. | |
1941 // The CMS gen may use this fact to determine its expansion policy. | |
1942 if (gch->incremental_collection_will_fail()) { | |
1943 assert(!_cmsGen->incremental_collection_failed(), | |
1944 "Should have been noticed, reacted to and cleared"); | |
1945 _cmsGen->set_incremental_collection_failed(); | |
1946 } | |
1947 *should_compact = | |
1948 UseCMSCompactAtFullCollection && | |
1949 ((_full_gcs_since_conc_gc >= CMSFullGCsBeforeCompaction) || | |
1950 GCCause::is_user_requested_gc(gch->gc_cause()) || | |
1951 gch->incremental_collection_will_fail()); | |
1952 *should_start_over = false; | |
1953 if (clear_all_soft_refs && !*should_compact) { | |
1954 // We are about to do a last ditch collection attempt | |
1955 // so it would normally make sense to do a compaction | |
1956 // to reclaim as much space as possible. | |
1957 if (CMSCompactWhenClearAllSoftRefs) { | |
1958 // Default: The rationale is that in this case either | |
1959 // we are past the final marking phase, in which case | |
1960 // we'd have to start over, or so little has been done | |
1961 // that there's little point in saving that work. Compaction | |
1962 // appears to be the sensible choice in either case. | |
1963 *should_compact = true; | |
1964 } else { | |
1965 // We have been asked to clear all soft refs, but not to | |
1966 // compact. Make sure that we aren't past the final checkpoint | |
1967 // phase, for that is where we process soft refs. If we are already | |
1968 // past that phase, we'll need to redo the refs discovery phase and | |
1969 // if necessary clear soft refs that weren't previously | |
1970 // cleared. We do so by remembering the phase in which | |
1971 // we came in, and if we are past the refs processing | |
1972 // phase, we'll choose to just redo the mark-sweep | |
1973 // collection from scratch. | |
1974 if (_collectorState > FinalMarking) { | |
1975 // We are past the refs processing phase; | |
1976 // start over and do a fresh synchronous CMS cycle | |
1977 _collectorState = Resetting; // skip to reset to start new cycle | |
1978 reset(false /* == !asynch */); | |
1979 *should_start_over = true; | |
1980 } // else we can continue a possibly ongoing current cycle | |
1981 } | |
1982 } | |
1983 } | |
1984 | |
1985 // A work method used by the foreground collector to do | |
1986 // a mark-sweep-compact. | |
1987 void CMSCollector::do_compaction_work(bool clear_all_soft_refs) { | |
1988 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
1989 TraceTime t("CMS:MSC ", PrintGCDetails && Verbose, true, gclog_or_tty); | |
1990 if (PrintGC && Verbose && !(GCCause::is_user_requested_gc(gch->gc_cause()))) { | |
1991 gclog_or_tty->print_cr("Compact ConcurrentMarkSweepGeneration after %d " | |
1992 "collections passed to foreground collector", _full_gcs_since_conc_gc); | |
1993 } | |
1994 | |
1995 // Sample collection interval time and reset for collection pause. | |
1996 if (UseAdaptiveSizePolicy) { | |
1997 size_policy()->msc_collection_begin(); | |
1998 } | |
1999 | |
2000 // Temporarily widen the span of the weak reference processing to | |
2001 // the entire heap. | |
2002 MemRegion new_span(GenCollectedHeap::heap()->reserved_region()); | |
2003 ReferenceProcessorSpanMutator x(ref_processor(), new_span); | |
2004 | |
2005 // Temporarily, clear the "is_alive_non_header" field of the | |
2006 // reference processor. | |
2007 ReferenceProcessorIsAliveMutator y(ref_processor(), NULL); | |
2008 | |
2009 // Temporarily make reference _processing_ single threaded (non-MT). | |
2010 ReferenceProcessorMTProcMutator z(ref_processor(), false); | |
2011 | |
2012 // Temporarily make refs discovery atomic | |
2013 ReferenceProcessorAtomicMutator w(ref_processor(), true); | |
2014 | |
2015 ref_processor()->set_enqueuing_is_done(false); | |
2016 ref_processor()->enable_discovery(); | |
457
27a80744a83b
6778647: snap(), snap_policy() should be renamed setup(), setup_policy()
ysr
parents:
453
diff
changeset
|
2017 ref_processor()->setup_policy(clear_all_soft_refs); |
0 | 2018 // If an asynchronous collection finishes, the _modUnionTable is |
2019 // all clear. If we are assuming the collection from an asynchronous | |
2020 // collection, clear the _modUnionTable. | |
2021 assert(_collectorState != Idling || _modUnionTable.isAllClear(), | |
2022 "_modUnionTable should be clear if the baton was not passed"); | |
2023 _modUnionTable.clear_all(); | |
2024 | |
2025 // We must adjust the allocation statistics being maintained | |
2026 // in the free list space. We do so by reading and clearing | |
2027 // the sweep timer and updating the block flux rate estimates below. | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2028 assert(!_intra_sweep_timer.is_active(), "_intra_sweep_timer should be inactive"); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2029 if (_inter_sweep_timer.is_active()) { |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2030 _inter_sweep_timer.stop(); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2031 // Note that we do not use this sample to update the _inter_sweep_estimate. |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2032 _cmsGen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2033 _inter_sweep_estimate.padded_average(), |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2034 _intra_sweep_estimate.padded_average()); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2035 } |
0 | 2036 |
1703
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
2037 { |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
2038 TraceCMSMemoryManagerStats(); |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
2039 } |
0 | 2040 GenMarkSweep::invoke_at_safepoint(_cmsGen->level(), |
2041 ref_processor(), clear_all_soft_refs); | |
2042 #ifdef ASSERT | |
2043 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); | |
2044 size_t free_size = cms_space->free(); | |
2045 assert(free_size == | |
2046 pointer_delta(cms_space->end(), cms_space->compaction_top()) | |
2047 * HeapWordSize, | |
2048 "All the free space should be compacted into one chunk at top"); | |
2049 assert(cms_space->dictionary()->totalChunkSize( | |
2050 debug_only(cms_space->freelistLock())) == 0 || | |
2051 cms_space->totalSizeInIndexedFreeLists() == 0, | |
2052 "All the free space should be in a single chunk"); | |
2053 size_t num = cms_space->totalCount(); | |
2054 assert((free_size == 0 && num == 0) || | |
2055 (free_size > 0 && (num == 1 || num == 2)), | |
2056 "There should be at most 2 free chunks after compaction"); | |
2057 #endif // ASSERT | |
2058 _collectorState = Resetting; | |
2059 assert(_restart_addr == NULL, | |
2060 "Should have been NULL'd before baton was passed"); | |
2061 reset(false /* == !asynch */); | |
2062 _cmsGen->reset_after_compaction(); | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
2063 _concurrent_cycles_since_last_unload = 0; |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
2064 |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
2065 if (verifying() && !should_unload_classes()) { |
0 | 2066 perm_gen_verify_bit_map()->clear_all(); |
2067 } | |
2068 | |
2069 // Clear any data recorded in the PLAB chunk arrays. | |
2070 if (_survivor_plab_array != NULL) { | |
2071 reset_survivor_plab_arrays(); | |
2072 } | |
2073 | |
2074 // Adjust the per-size allocation stats for the next epoch. | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2075 _cmsGen->cmsSpace()->endSweepFLCensus(sweep_count() /* fake */); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2076 // Restart the "inter sweep timer" for the next epoch. |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2077 _inter_sweep_timer.reset(); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2078 _inter_sweep_timer.start(); |
0 | 2079 |
2080 // Sample collection pause time and reset for collection interval. | |
2081 if (UseAdaptiveSizePolicy) { | |
2082 size_policy()->msc_collection_end(gch->gc_cause()); | |
2083 } | |
2084 | |
2085 // For a mark-sweep-compact, compute_new_size() will be called | |
2086 // in the heap's do_collection() method. | |
2087 } | |
2088 | |
2089 // A work method used by the foreground collector to do | |
2090 // a mark-sweep, after taking over from a possibly on-going | |
2091 // concurrent mark-sweep collection. | |
2092 void CMSCollector::do_mark_sweep_work(bool clear_all_soft_refs, | |
2093 CollectorState first_state, bool should_start_over) { | |
2094 if (PrintGC && Verbose) { | |
2095 gclog_or_tty->print_cr("Pass concurrent collection to foreground " | |
2096 "collector with count %d", | |
2097 _full_gcs_since_conc_gc); | |
2098 } | |
2099 switch (_collectorState) { | |
2100 case Idling: | |
2101 if (first_state == Idling || should_start_over) { | |
2102 // The background GC was not active, or should | |
2103 // restarted from scratch; start the cycle. | |
2104 _collectorState = InitialMarking; | |
2105 } | |
2106 // If first_state was not Idling, then a background GC | |
2107 // was in progress and has now finished. No need to do it | |
2108 // again. Leave the state as Idling. | |
2109 break; | |
2110 case Precleaning: | |
2111 // In the foreground case don't do the precleaning since | |
2112 // it is not done concurrently and there is extra work | |
2113 // required. | |
2114 _collectorState = FinalMarking; | |
2115 } | |
2116 if (PrintGCDetails && | |
2117 (_collectorState > Idling || | |
2118 !GCCause::is_user_requested_gc(GenCollectedHeap::heap()->gc_cause()))) { | |
2119 gclog_or_tty->print(" (concurrent mode failure)"); | |
2120 } | |
2121 collect_in_foreground(clear_all_soft_refs); | |
2122 | |
2123 // For a mark-sweep, compute_new_size() will be called | |
2124 // in the heap's do_collection() method. | |
2125 } | |
2126 | |
2127 | |
2128 void CMSCollector::getFreelistLocks() const { | |
2129 // Get locks for all free lists in all generations that this | |
2130 // collector is responsible for | |
2131 _cmsGen->freelistLock()->lock_without_safepoint_check(); | |
2132 _permGen->freelistLock()->lock_without_safepoint_check(); | |
2133 } | |
2134 | |
2135 void CMSCollector::releaseFreelistLocks() const { | |
2136 // Release locks for all free lists in all generations that this | |
2137 // collector is responsible for | |
2138 _cmsGen->freelistLock()->unlock(); | |
2139 _permGen->freelistLock()->unlock(); | |
2140 } | |
2141 | |
2142 bool CMSCollector::haveFreelistLocks() const { | |
2143 // Check locks for all free lists in all generations that this | |
2144 // collector is responsible for | |
2145 assert_lock_strong(_cmsGen->freelistLock()); | |
2146 assert_lock_strong(_permGen->freelistLock()); | |
2147 PRODUCT_ONLY(ShouldNotReachHere()); | |
2148 return true; | |
2149 } | |
2150 | |
2151 // A utility class that is used by the CMS collector to | |
2152 // temporarily "release" the foreground collector from its | |
2153 // usual obligation to wait for the background collector to | |
2154 // complete an ongoing phase before proceeding. | |
2155 class ReleaseForegroundGC: public StackObj { | |
2156 private: | |
2157 CMSCollector* _c; | |
2158 public: | |
2159 ReleaseForegroundGC(CMSCollector* c) : _c(c) { | |
2160 assert(_c->_foregroundGCShouldWait, "Else should not need to call"); | |
2161 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); | |
2162 // allow a potentially blocked foreground collector to proceed | |
2163 _c->_foregroundGCShouldWait = false; | |
2164 if (_c->_foregroundGCIsActive) { | |
2165 CGC_lock->notify(); | |
2166 } | |
2167 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
2168 "Possible deadlock"); | |
2169 } | |
2170 | |
2171 ~ReleaseForegroundGC() { | |
2172 assert(!_c->_foregroundGCShouldWait, "Usage protocol violation?"); | |
2173 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); | |
2174 _c->_foregroundGCShouldWait = true; | |
2175 } | |
2176 }; | |
2177 | |
2178 // There are separate collect_in_background and collect_in_foreground because of | |
2179 // the different locking requirements of the background collector and the | |
2180 // foreground collector. There was originally an attempt to share | |
2181 // one "collect" method between the background collector and the foreground | |
2182 // collector but the if-then-else required made it cleaner to have | |
2183 // separate methods. | |
2184 void CMSCollector::collect_in_background(bool clear_all_soft_refs) { | |
2185 assert(Thread::current()->is_ConcurrentGC_thread(), | |
2186 "A CMS asynchronous collection is only allowed on a CMS thread."); | |
2187 | |
2188 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
2189 { | |
2190 bool safepoint_check = Mutex::_no_safepoint_check_flag; | |
2191 MutexLockerEx hl(Heap_lock, safepoint_check); | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
2192 FreelistLocker fll(this); |
0 | 2193 MutexLockerEx x(CGC_lock, safepoint_check); |
2194 if (_foregroundGCIsActive || !UseAsyncConcMarkSweepGC) { | |
2195 // The foreground collector is active or we're | |
2196 // not using asynchronous collections. Skip this | |
2197 // background collection. | |
2198 assert(!_foregroundGCShouldWait, "Should be clear"); | |
2199 return; | |
2200 } else { | |
2201 assert(_collectorState == Idling, "Should be idling before start."); | |
2202 _collectorState = InitialMarking; | |
2203 // Reset the expansion cause, now that we are about to begin | |
2204 // a new cycle. | |
2205 clear_expansion_cause(); | |
2206 } | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
2207 // Decide if we want to enable class unloading as part of the |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
2208 // ensuing concurrent GC cycle. |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
2209 update_should_unload_classes(); |
0 | 2210 _full_gc_requested = false; // acks all outstanding full gc requests |
2211 // Signal that we are about to start a collection | |
2212 gch->increment_total_full_collections(); // ... starting a collection cycle | |
2213 _collection_count_start = gch->total_full_collections(); | |
2214 } | |
2215 | |
2216 // Used for PrintGC | |
2217 size_t prev_used; | |
2218 if (PrintGC && Verbose) { | |
2219 prev_used = _cmsGen->used(); // XXXPERM | |
2220 } | |
2221 | |
2222 // The change of the collection state is normally done at this level; | |
2223 // the exceptions are phases that are executed while the world is | |
2224 // stopped. For those phases the change of state is done while the | |
2225 // world is stopped. For baton passing purposes this allows the | |
2226 // background collector to finish the phase and change state atomically. | |
2227 // The foreground collector cannot wait on a phase that is done | |
2228 // while the world is stopped because the foreground collector already | |
2229 // has the world stopped and would deadlock. | |
2230 while (_collectorState != Idling) { | |
2231 if (TraceCMSState) { | |
2232 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", | |
2233 Thread::current(), _collectorState); | |
2234 } | |
2235 // The foreground collector | |
2236 // holds the Heap_lock throughout its collection. | |
2237 // holds the CMS token (but not the lock) | |
2238 // except while it is waiting for the background collector to yield. | |
2239 // | |
2240 // The foreground collector should be blocked (not for long) | |
2241 // if the background collector is about to start a phase | |
2242 // executed with world stopped. If the background | |
2243 // collector has already started such a phase, the | |
2244 // foreground collector is blocked waiting for the | |
2245 // Heap_lock. The stop-world phases (InitialMarking and FinalMarking) | |
2246 // are executed in the VM thread. | |
2247 // | |
2248 // The locking order is | |
2249 // PendingListLock (PLL) -- if applicable (FinalMarking) | |
2250 // Heap_lock (both this & PLL locked in VM_CMS_Operation::prologue()) | |
2251 // CMS token (claimed in | |
2252 // stop_world_and_do() --> | |
2253 // safepoint_synchronize() --> | |
2254 // CMSThread::synchronize()) | |
2255 | |
2256 { | |
2257 // Check if the FG collector wants us to yield. | |
2258 CMSTokenSync x(true); // is cms thread | |
2259 if (waitForForegroundGC()) { | |
2260 // We yielded to a foreground GC, nothing more to be | |
2261 // done this round. | |
2262 assert(_foregroundGCShouldWait == false, "We set it to false in " | |
2263 "waitForForegroundGC()"); | |
2264 if (TraceCMSState) { | |
2265 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT | |
2266 " exiting collection CMS state %d", | |
2267 Thread::current(), _collectorState); | |
2268 } | |
2269 return; | |
2270 } else { | |
2271 // The background collector can run but check to see if the | |
2272 // foreground collector has done a collection while the | |
2273 // background collector was waiting to get the CGC_lock | |
2274 // above. If yes, break so that _foregroundGCShouldWait | |
2275 // is cleared before returning. | |
2276 if (_collectorState == Idling) { | |
2277 break; | |
2278 } | |
2279 } | |
2280 } | |
2281 | |
2282 assert(_foregroundGCShouldWait, "Foreground collector, if active, " | |
2283 "should be waiting"); | |
2284 | |
2285 switch (_collectorState) { | |
2286 case InitialMarking: | |
2287 { | |
2288 ReleaseForegroundGC x(this); | |
2289 stats().record_cms_begin(); | |
2290 | |
2291 VM_CMS_Initial_Mark initial_mark_op(this); | |
2292 VMThread::execute(&initial_mark_op); | |
2293 } | |
2294 // The collector state may be any legal state at this point | |
2295 // since the background collector may have yielded to the | |
2296 // foreground collector. | |
2297 break; | |
2298 case Marking: | |
2299 // initial marking in checkpointRootsInitialWork has been completed | |
2300 if (markFromRoots(true)) { // we were successful | |
2301 assert(_collectorState == Precleaning, "Collector state should " | |
2302 "have changed"); | |
2303 } else { | |
2304 assert(_foregroundGCIsActive, "Internal state inconsistency"); | |
2305 } | |
2306 break; | |
2307 case Precleaning: | |
2308 if (UseAdaptiveSizePolicy) { | |
2309 size_policy()->concurrent_precleaning_begin(); | |
2310 } | |
2311 // marking from roots in markFromRoots has been completed | |
2312 preclean(); | |
2313 if (UseAdaptiveSizePolicy) { | |
2314 size_policy()->concurrent_precleaning_end(); | |
2315 } | |
2316 assert(_collectorState == AbortablePreclean || | |
2317 _collectorState == FinalMarking, | |
2318 "Collector state should have changed"); | |
2319 break; | |
2320 case AbortablePreclean: | |
2321 if (UseAdaptiveSizePolicy) { | |
2322 size_policy()->concurrent_phases_resume(); | |
2323 } | |
2324 abortable_preclean(); | |
2325 if (UseAdaptiveSizePolicy) { | |
2326 size_policy()->concurrent_precleaning_end(); | |
2327 } | |
2328 assert(_collectorState == FinalMarking, "Collector state should " | |
2329 "have changed"); | |
2330 break; | |
2331 case FinalMarking: | |
2332 { | |
2333 ReleaseForegroundGC x(this); | |
2334 | |
2335 VM_CMS_Final_Remark final_remark_op(this); | |
2336 VMThread::execute(&final_remark_op); | |
935 | 2337 } |
0 | 2338 assert(_foregroundGCShouldWait, "block post-condition"); |
2339 break; | |
2340 case Sweeping: | |
2341 if (UseAdaptiveSizePolicy) { | |
2342 size_policy()->concurrent_sweeping_begin(); | |
2343 } | |
2344 // final marking in checkpointRootsFinal has been completed | |
2345 sweep(true); | |
2346 assert(_collectorState == Resizing, "Collector state change " | |
2347 "to Resizing must be done under the free_list_lock"); | |
2348 _full_gcs_since_conc_gc = 0; | |
2349 | |
2350 // Stop the timers for adaptive size policy for the concurrent phases | |
2351 if (UseAdaptiveSizePolicy) { | |
2352 size_policy()->concurrent_sweeping_end(); | |
2353 size_policy()->concurrent_phases_end(gch->gc_cause(), | |
2354 gch->prev_gen(_cmsGen)->capacity(), | |
2355 _cmsGen->free()); | |
2356 } | |
2357 | |
2358 case Resizing: { | |
2359 // Sweeping has been completed... | |
2360 // At this point the background collection has completed. | |
2361 // Don't move the call to compute_new_size() down | |
2362 // into code that might be executed if the background | |
2363 // collection was preempted. | |
2364 { | |
2365 ReleaseForegroundGC x(this); // unblock FG collection | |
2366 MutexLockerEx y(Heap_lock, Mutex::_no_safepoint_check_flag); | |
2367 CMSTokenSync z(true); // not strictly needed. | |
2368 if (_collectorState == Resizing) { | |
2369 compute_new_size(); | |
2370 _collectorState = Resetting; | |
2371 } else { | |
2372 assert(_collectorState == Idling, "The state should only change" | |
2373 " because the foreground collector has finished the collection"); | |
2374 } | |
2375 } | |
2376 break; | |
2377 } | |
2378 case Resetting: | |
2379 // CMS heap resizing has been completed | |
2380 reset(true); | |
2381 assert(_collectorState == Idling, "Collector state should " | |
2382 "have changed"); | |
2383 stats().record_cms_end(); | |
2384 // Don't move the concurrent_phases_end() and compute_new_size() | |
2385 // calls to here because a preempted background collection | |
2386 // has it's state set to "Resetting". | |
2387 break; | |
2388 case Idling: | |
2389 default: | |
2390 ShouldNotReachHere(); | |
2391 break; | |
2392 } | |
2393 if (TraceCMSState) { | |
2394 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", | |
2395 Thread::current(), _collectorState); | |
2396 } | |
2397 assert(_foregroundGCShouldWait, "block post-condition"); | |
2398 } | |
2399 | |
2400 // Should this be in gc_epilogue? | |
2401 collector_policy()->counters()->update_counters(); | |
2402 | |
2403 { | |
2404 // Clear _foregroundGCShouldWait and, in the event that the | |
2405 // foreground collector is waiting, notify it, before | |
2406 // returning. | |
2407 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); | |
2408 _foregroundGCShouldWait = false; | |
2409 if (_foregroundGCIsActive) { | |
2410 CGC_lock->notify(); | |
2411 } | |
2412 assert(!ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
2413 "Possible deadlock"); | |
2414 } | |
2415 if (TraceCMSState) { | |
2416 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT | |
2417 " exiting collection CMS state %d", | |
2418 Thread::current(), _collectorState); | |
2419 } | |
2420 if (PrintGC && Verbose) { | |
2421 _cmsGen->print_heap_change(prev_used); | |
2422 } | |
2423 } | |
2424 | |
2425 void CMSCollector::collect_in_foreground(bool clear_all_soft_refs) { | |
2426 assert(_foregroundGCIsActive && !_foregroundGCShouldWait, | |
2427 "Foreground collector should be waiting, not executing"); | |
2428 assert(Thread::current()->is_VM_thread(), "A foreground collection" | |
2429 "may only be done by the VM Thread with the world stopped"); | |
2430 assert(ConcurrentMarkSweepThread::vm_thread_has_cms_token(), | |
2431 "VM thread should have CMS token"); | |
2432 | |
2433 NOT_PRODUCT(TraceTime t("CMS:MS (foreground) ", PrintGCDetails && Verbose, | |
2434 true, gclog_or_tty);) | |
2435 if (UseAdaptiveSizePolicy) { | |
2436 size_policy()->ms_collection_begin(); | |
2437 } | |
2438 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact); | |
2439 | |
2440 HandleMark hm; // Discard invalid handles created during verification | |
2441 | |
2442 if (VerifyBeforeGC && | |
2443 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { | |
2444 Universe::verify(true); | |
2445 } | |
2446 | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
2447 // Snapshot the soft reference policy to be used in this collection cycle. |
457
27a80744a83b
6778647: snap(), snap_policy() should be renamed setup(), setup_policy()
ysr
parents:
453
diff
changeset
|
2448 ref_processor()->setup_policy(clear_all_soft_refs); |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
2449 |
0 | 2450 bool init_mark_was_synchronous = false; // until proven otherwise |
2451 while (_collectorState != Idling) { | |
2452 if (TraceCMSState) { | |
2453 gclog_or_tty->print_cr("Thread " INTPTR_FORMAT " in CMS state %d", | |
2454 Thread::current(), _collectorState); | |
2455 } | |
2456 switch (_collectorState) { | |
2457 case InitialMarking: | |
2458 init_mark_was_synchronous = true; // fact to be exploited in re-mark | |
2459 checkpointRootsInitial(false); | |
2460 assert(_collectorState == Marking, "Collector state should have changed" | |
2461 " within checkpointRootsInitial()"); | |
2462 break; | |
2463 case Marking: | |
2464 // initial marking in checkpointRootsInitialWork has been completed | |
2465 if (VerifyDuringGC && | |
2466 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { | |
2467 gclog_or_tty->print("Verify before initial mark: "); | |
2468 Universe::verify(true); | |
2469 } | |
2470 { | |
2471 bool res = markFromRoots(false); | |
2472 assert(res && _collectorState == FinalMarking, "Collector state should " | |
2473 "have changed"); | |
2474 break; | |
2475 } | |
2476 case FinalMarking: | |
2477 if (VerifyDuringGC && | |
2478 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { | |
2479 gclog_or_tty->print("Verify before re-mark: "); | |
2480 Universe::verify(true); | |
2481 } | |
2482 checkpointRootsFinal(false, clear_all_soft_refs, | |
2483 init_mark_was_synchronous); | |
2484 assert(_collectorState == Sweeping, "Collector state should not " | |
2485 "have changed within checkpointRootsFinal()"); | |
2486 break; | |
2487 case Sweeping: | |
2488 // final marking in checkpointRootsFinal has been completed | |
2489 if (VerifyDuringGC && | |
2490 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { | |
2491 gclog_or_tty->print("Verify before sweep: "); | |
2492 Universe::verify(true); | |
2493 } | |
2494 sweep(false); | |
2495 assert(_collectorState == Resizing, "Incorrect state"); | |
2496 break; | |
2497 case Resizing: { | |
2498 // Sweeping has been completed; the actual resize in this case | |
2499 // is done separately; nothing to be done in this state. | |
2500 _collectorState = Resetting; | |
2501 break; | |
2502 } | |
2503 case Resetting: | |
2504 // The heap has been resized. | |
2505 if (VerifyDuringGC && | |
2506 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { | |
2507 gclog_or_tty->print("Verify before reset: "); | |
2508 Universe::verify(true); | |
2509 } | |
2510 reset(false); | |
2511 assert(_collectorState == Idling, "Collector state should " | |
2512 "have changed"); | |
2513 break; | |
2514 case Precleaning: | |
2515 case AbortablePreclean: | |
2516 // Elide the preclean phase | |
2517 _collectorState = FinalMarking; | |
2518 break; | |
2519 default: | |
2520 ShouldNotReachHere(); | |
2521 } | |
2522 if (TraceCMSState) { | |
2523 gclog_or_tty->print_cr(" Thread " INTPTR_FORMAT " done - next CMS state %d", | |
2524 Thread::current(), _collectorState); | |
2525 } | |
2526 } | |
2527 | |
2528 if (UseAdaptiveSizePolicy) { | |
2529 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
2530 size_policy()->ms_collection_end(gch->gc_cause()); | |
2531 } | |
2532 | |
2533 if (VerifyAfterGC && | |
2534 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { | |
2535 Universe::verify(true); | |
2536 } | |
2537 if (TraceCMSState) { | |
2538 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT | |
2539 " exiting collection CMS state %d", | |
2540 Thread::current(), _collectorState); | |
2541 } | |
2542 } | |
2543 | |
2544 bool CMSCollector::waitForForegroundGC() { | |
2545 bool res = false; | |
2546 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
2547 "CMS thread should have CMS token"); | |
2548 // Block the foreground collector until the | |
2549 // background collectors decides whether to | |
2550 // yield. | |
2551 MutexLockerEx x(CGC_lock, Mutex::_no_safepoint_check_flag); | |
2552 _foregroundGCShouldWait = true; | |
2553 if (_foregroundGCIsActive) { | |
2554 // The background collector yields to the | |
2555 // foreground collector and returns a value | |
2556 // indicating that it has yielded. The foreground | |
2557 // collector can proceed. | |
2558 res = true; | |
2559 _foregroundGCShouldWait = false; | |
2560 ConcurrentMarkSweepThread::clear_CMS_flag( | |
2561 ConcurrentMarkSweepThread::CMS_cms_has_token); | |
2562 ConcurrentMarkSweepThread::set_CMS_flag( | |
2563 ConcurrentMarkSweepThread::CMS_cms_wants_token); | |
2564 // Get a possibly blocked foreground thread going | |
2565 CGC_lock->notify(); | |
2566 if (TraceCMSState) { | |
2567 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " waiting at CMS state %d", | |
2568 Thread::current(), _collectorState); | |
2569 } | |
2570 while (_foregroundGCIsActive) { | |
2571 CGC_lock->wait(Mutex::_no_safepoint_check_flag); | |
2572 } | |
2573 ConcurrentMarkSweepThread::set_CMS_flag( | |
2574 ConcurrentMarkSweepThread::CMS_cms_has_token); | |
2575 ConcurrentMarkSweepThread::clear_CMS_flag( | |
2576 ConcurrentMarkSweepThread::CMS_cms_wants_token); | |
2577 } | |
2578 if (TraceCMSState) { | |
2579 gclog_or_tty->print_cr("CMS Thread " INTPTR_FORMAT " continuing at CMS state %d", | |
2580 Thread::current(), _collectorState); | |
2581 } | |
2582 return res; | |
2583 } | |
2584 | |
2585 // Because of the need to lock the free lists and other structures in | |
2586 // the collector, common to all the generations that the collector is | |
2587 // collecting, we need the gc_prologues of individual CMS generations | |
2588 // delegate to their collector. It may have been simpler had the | |
2589 // current infrastructure allowed one to call a prologue on a | |
2590 // collector. In the absence of that we have the generation's | |
2591 // prologue delegate to the collector, which delegates back | |
2592 // some "local" work to a worker method in the individual generations | |
2593 // that it's responsible for collecting, while itself doing any | |
2594 // work common to all generations it's responsible for. A similar | |
2595 // comment applies to the gc_epilogue()'s. | |
2596 // The role of the varaible _between_prologue_and_epilogue is to | |
2597 // enforce the invocation protocol. | |
2598 void CMSCollector::gc_prologue(bool full) { | |
2599 // Call gc_prologue_work() for each CMSGen and PermGen that | |
2600 // we are responsible for. | |
2601 | |
2602 // The following locking discipline assumes that we are only called | |
2603 // when the world is stopped. | |
2604 assert(SafepointSynchronize::is_at_safepoint(), "world is stopped assumption"); | |
2605 | |
2606 // The CMSCollector prologue must call the gc_prologues for the | |
2607 // "generations" (including PermGen if any) that it's responsible | |
2608 // for. | |
2609 | |
2610 assert( Thread::current()->is_VM_thread() | |
2611 || ( CMSScavengeBeforeRemark | |
2612 && Thread::current()->is_ConcurrentGC_thread()), | |
2613 "Incorrect thread type for prologue execution"); | |
2614 | |
2615 if (_between_prologue_and_epilogue) { | |
2616 // We have already been invoked; this is a gc_prologue delegation | |
2617 // from yet another CMS generation that we are responsible for, just | |
2618 // ignore it since all relevant work has already been done. | |
2619 return; | |
2620 } | |
2621 | |
2622 // set a bit saying prologue has been called; cleared in epilogue | |
2623 _between_prologue_and_epilogue = true; | |
2624 // Claim locks for common data structures, then call gc_prologue_work() | |
2625 // for each CMSGen and PermGen that we are responsible for. | |
2626 | |
2627 getFreelistLocks(); // gets free list locks on constituent spaces | |
2628 bitMapLock()->lock_without_safepoint_check(); | |
2629 | |
2630 // Should call gc_prologue_work() for all cms gens we are responsible for | |
2631 bool registerClosure = _collectorState >= Marking | |
2632 && _collectorState < Sweeping; | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
2633 ModUnionClosure* muc = CollectedHeap::use_parallel_gc_threads() ? |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
2634 &_modUnionClosurePar |
0 | 2635 : &_modUnionClosure; |
2636 _cmsGen->gc_prologue_work(full, registerClosure, muc); | |
2637 _permGen->gc_prologue_work(full, registerClosure, muc); | |
2638 | |
2639 if (!full) { | |
2640 stats().record_gc0_begin(); | |
2641 } | |
2642 } | |
2643 | |
2644 void ConcurrentMarkSweepGeneration::gc_prologue(bool full) { | |
2645 // Delegate to CMScollector which knows how to coordinate between | |
2646 // this and any other CMS generations that it is responsible for | |
2647 // collecting. | |
2648 collector()->gc_prologue(full); | |
2649 } | |
2650 | |
2651 // This is a "private" interface for use by this generation's CMSCollector. | |
2652 // Not to be called directly by any other entity (for instance, | |
2653 // GenCollectedHeap, which calls the "public" gc_prologue method above). | |
2654 void ConcurrentMarkSweepGeneration::gc_prologue_work(bool full, | |
2655 bool registerClosure, ModUnionClosure* modUnionClosure) { | |
2656 assert(!incremental_collection_failed(), "Shouldn't be set yet"); | |
2657 assert(cmsSpace()->preconsumptionDirtyCardClosure() == NULL, | |
2658 "Should be NULL"); | |
2659 if (registerClosure) { | |
2660 cmsSpace()->setPreconsumptionDirtyCardClosure(modUnionClosure); | |
2661 } | |
2662 cmsSpace()->gc_prologue(); | |
2663 // Clear stat counters | |
2664 NOT_PRODUCT( | |
2665 assert(_numObjectsPromoted == 0, "check"); | |
2666 assert(_numWordsPromoted == 0, "check"); | |
2667 if (Verbose && PrintGC) { | |
2668 gclog_or_tty->print("Allocated "SIZE_FORMAT" objects, " | |
2669 SIZE_FORMAT" bytes concurrently", | |
2670 _numObjectsAllocated, _numWordsAllocated*sizeof(HeapWord)); | |
2671 } | |
2672 _numObjectsAllocated = 0; | |
2673 _numWordsAllocated = 0; | |
2674 ) | |
2675 } | |
2676 | |
2677 void CMSCollector::gc_epilogue(bool full) { | |
2678 // The following locking discipline assumes that we are only called | |
2679 // when the world is stopped. | |
2680 assert(SafepointSynchronize::is_at_safepoint(), | |
2681 "world is stopped assumption"); | |
2682 | |
2683 // Currently the CMS epilogue (see CompactibleFreeListSpace) merely checks | |
2684 // if linear allocation blocks need to be appropriately marked to allow the | |
2685 // the blocks to be parsable. We also check here whether we need to nudge the | |
2686 // CMS collector thread to start a new cycle (if it's not already active). | |
2687 assert( Thread::current()->is_VM_thread() | |
2688 || ( CMSScavengeBeforeRemark | |
2689 && Thread::current()->is_ConcurrentGC_thread()), | |
2690 "Incorrect thread type for epilogue execution"); | |
2691 | |
2692 if (!_between_prologue_and_epilogue) { | |
2693 // We have already been invoked; this is a gc_epilogue delegation | |
2694 // from yet another CMS generation that we are responsible for, just | |
2695 // ignore it since all relevant work has already been done. | |
2696 return; | |
2697 } | |
2698 assert(haveFreelistLocks(), "must have freelist locks"); | |
2699 assert_lock_strong(bitMapLock()); | |
2700 | |
2701 _cmsGen->gc_epilogue_work(full); | |
2702 _permGen->gc_epilogue_work(full); | |
2703 | |
2704 if (_collectorState == AbortablePreclean || _collectorState == Precleaning) { | |
2705 // in case sampling was not already enabled, enable it | |
2706 _start_sampling = true; | |
2707 } | |
2708 // reset _eden_chunk_array so sampling starts afresh | |
2709 _eden_chunk_index = 0; | |
2710 | |
2711 size_t cms_used = _cmsGen->cmsSpace()->used(); | |
2712 size_t perm_used = _permGen->cmsSpace()->used(); | |
2713 | |
2714 // update performance counters - this uses a special version of | |
2715 // update_counters() that allows the utilization to be passed as a | |
2716 // parameter, avoiding multiple calls to used(). | |
2717 // | |
2718 _cmsGen->update_counters(cms_used); | |
2719 _permGen->update_counters(perm_used); | |
2720 | |
2721 if (CMSIncrementalMode) { | |
2722 icms_update_allocation_limits(); | |
2723 } | |
2724 | |
2725 bitMapLock()->unlock(); | |
2726 releaseFreelistLocks(); | |
2727 | |
2728 _between_prologue_and_epilogue = false; // ready for next cycle | |
2729 } | |
2730 | |
2731 void ConcurrentMarkSweepGeneration::gc_epilogue(bool full) { | |
2732 collector()->gc_epilogue(full); | |
2733 | |
2734 // Also reset promotion tracking in par gc thread states. | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
2735 if (CollectedHeap::use_parallel_gc_threads()) { |
0 | 2736 for (uint i = 0; i < ParallelGCThreads; i++) { |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2737 _par_gc_thread_states[i]->promo.stopTrackingPromotions(i); |
0 | 2738 } |
2739 } | |
2740 } | |
2741 | |
2742 void ConcurrentMarkSweepGeneration::gc_epilogue_work(bool full) { | |
2743 assert(!incremental_collection_failed(), "Should have been cleared"); | |
2744 cmsSpace()->setPreconsumptionDirtyCardClosure(NULL); | |
2745 cmsSpace()->gc_epilogue(); | |
2746 // Print stat counters | |
2747 NOT_PRODUCT( | |
2748 assert(_numObjectsAllocated == 0, "check"); | |
2749 assert(_numWordsAllocated == 0, "check"); | |
2750 if (Verbose && PrintGC) { | |
2751 gclog_or_tty->print("Promoted "SIZE_FORMAT" objects, " | |
2752 SIZE_FORMAT" bytes", | |
2753 _numObjectsPromoted, _numWordsPromoted*sizeof(HeapWord)); | |
2754 } | |
2755 _numObjectsPromoted = 0; | |
2756 _numWordsPromoted = 0; | |
2757 ) | |
2758 | |
2759 if (PrintGC && Verbose) { | |
2760 // Call down the chain in contiguous_available needs the freelistLock | |
2761 // so print this out before releasing the freeListLock. | |
2762 gclog_or_tty->print(" Contiguous available "SIZE_FORMAT" bytes ", | |
2763 contiguous_available()); | |
2764 } | |
2765 } | |
2766 | |
2767 #ifndef PRODUCT | |
2768 bool CMSCollector::have_cms_token() { | |
2769 Thread* thr = Thread::current(); | |
2770 if (thr->is_VM_thread()) { | |
2771 return ConcurrentMarkSweepThread::vm_thread_has_cms_token(); | |
2772 } else if (thr->is_ConcurrentGC_thread()) { | |
2773 return ConcurrentMarkSweepThread::cms_thread_has_cms_token(); | |
2774 } else if (thr->is_GC_task_thread()) { | |
2775 return ConcurrentMarkSweepThread::vm_thread_has_cms_token() && | |
2776 ParGCRareEvent_lock->owned_by_self(); | |
2777 } | |
2778 return false; | |
2779 } | |
2780 #endif | |
2781 | |
2782 // Check reachability of the given heap address in CMS generation, | |
2783 // treating all other generations as roots. | |
2784 bool CMSCollector::is_cms_reachable(HeapWord* addr) { | |
2785 // We could "guarantee" below, rather than assert, but i'll | |
2786 // leave these as "asserts" so that an adventurous debugger | |
2787 // could try this in the product build provided some subset of | |
2788 // the conditions were met, provided they were intersted in the | |
2789 // results and knew that the computation below wouldn't interfere | |
2790 // with other concurrent computations mutating the structures | |
2791 // being read or written. | |
2792 assert(SafepointSynchronize::is_at_safepoint(), | |
2793 "Else mutations in object graph will make answer suspect"); | |
2794 assert(have_cms_token(), "Should hold cms token"); | |
2795 assert(haveFreelistLocks(), "must hold free list locks"); | |
2796 assert_lock_strong(bitMapLock()); | |
2797 | |
2798 // Clear the marking bit map array before starting, but, just | |
2799 // for kicks, first report if the given address is already marked | |
2800 gclog_or_tty->print_cr("Start: Address 0x%x is%s marked", addr, | |
2801 _markBitMap.isMarked(addr) ? "" : " not"); | |
2802 | |
2803 if (verify_after_remark()) { | |
2804 MutexLockerEx x(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); | |
2805 bool result = verification_mark_bm()->isMarked(addr); | |
2806 gclog_or_tty->print_cr("TransitiveMark: Address 0x%x %s marked", addr, | |
2807 result ? "IS" : "is NOT"); | |
2808 return result; | |
2809 } else { | |
2810 gclog_or_tty->print_cr("Could not compute result"); | |
2811 return false; | |
2812 } | |
2813 } | |
2814 | |
2815 //////////////////////////////////////////////////////// | |
2816 // CMS Verification Support | |
2817 //////////////////////////////////////////////////////// | |
2818 // Following the remark phase, the following invariant | |
2819 // should hold -- each object in the CMS heap which is | |
2820 // marked in markBitMap() should be marked in the verification_mark_bm(). | |
2821 | |
2822 class VerifyMarkedClosure: public BitMapClosure { | |
2823 CMSBitMap* _marks; | |
2824 bool _failed; | |
2825 | |
2826 public: | |
2827 VerifyMarkedClosure(CMSBitMap* bm): _marks(bm), _failed(false) {} | |
2828 | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
2829 bool do_bit(size_t offset) { |
0 | 2830 HeapWord* addr = _marks->offsetToHeapWord(offset); |
2831 if (!_marks->isMarked(addr)) { | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2832 oop(addr)->print_on(gclog_or_tty); |
0 | 2833 gclog_or_tty->print_cr(" ("INTPTR_FORMAT" should have been marked)", addr); |
2834 _failed = true; | |
2835 } | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
2836 return true; |
0 | 2837 } |
2838 | |
2839 bool failed() { return _failed; } | |
2840 }; | |
2841 | |
2842 bool CMSCollector::verify_after_remark() { | |
2843 gclog_or_tty->print(" [Verifying CMS Marking... "); | |
2844 MutexLockerEx ml(verification_mark_bm()->lock(), Mutex::_no_safepoint_check_flag); | |
2845 static bool init = false; | |
2846 | |
2847 assert(SafepointSynchronize::is_at_safepoint(), | |
2848 "Else mutations in object graph will make answer suspect"); | |
2849 assert(have_cms_token(), | |
2850 "Else there may be mutual interference in use of " | |
2851 " verification data structures"); | |
2852 assert(_collectorState > Marking && _collectorState <= Sweeping, | |
2853 "Else marking info checked here may be obsolete"); | |
2854 assert(haveFreelistLocks(), "must hold free list locks"); | |
2855 assert_lock_strong(bitMapLock()); | |
2856 | |
2857 | |
2858 // Allocate marking bit map if not already allocated | |
2859 if (!init) { // first time | |
2860 if (!verification_mark_bm()->allocate(_span)) { | |
2861 return false; | |
2862 } | |
2863 init = true; | |
2864 } | |
2865 | |
2866 assert(verification_mark_stack()->isEmpty(), "Should be empty"); | |
2867 | |
2868 // Turn off refs discovery -- so we will be tracing through refs. | |
2869 // This is as intended, because by this time | |
2870 // GC must already have cleared any refs that need to be cleared, | |
2871 // and traced those that need to be marked; moreover, | |
2872 // the marking done here is not going to intefere in any | |
2873 // way with the marking information used by GC. | |
2874 NoRefDiscovery no_discovery(ref_processor()); | |
2875 | |
2876 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) | |
2877 | |
2878 // Clear any marks from a previous round | |
2879 verification_mark_bm()->clear_all(); | |
2880 assert(verification_mark_stack()->isEmpty(), "markStack should be empty"); | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2881 verify_work_stacks_empty(); |
0 | 2882 |
2883 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
2884 gch->ensure_parsability(false); // fill TLABs, but no need to retire them | |
2885 // Update the saved marks which may affect the root scans. | |
2886 gch->save_marks(); | |
2887 | |
2888 if (CMSRemarkVerifyVariant == 1) { | |
2889 // In this first variant of verification, we complete | |
2890 // all marking, then check if the new marks-verctor is | |
2891 // a subset of the CMS marks-vector. | |
2892 verify_after_remark_work_1(); | |
2893 } else if (CMSRemarkVerifyVariant == 2) { | |
2894 // In this second variant of verification, we flag an error | |
2895 // (i.e. an object reachable in the new marks-vector not reachable | |
2896 // in the CMS marks-vector) immediately, also indicating the | |
2897 // identify of an object (A) that references the unmarked object (B) -- | |
2898 // presumably, a mutation to A failed to be picked up by preclean/remark? | |
2899 verify_after_remark_work_2(); | |
2900 } else { | |
2901 warning("Unrecognized value %d for CMSRemarkVerifyVariant", | |
2902 CMSRemarkVerifyVariant); | |
2903 } | |
2904 gclog_or_tty->print(" done] "); | |
2905 return true; | |
2906 } | |
2907 | |
2908 void CMSCollector::verify_after_remark_work_1() { | |
2909 ResourceMark rm; | |
2910 HandleMark hm; | |
2911 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
2912 | |
2913 // Mark from roots one level into CMS | |
994
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
2914 MarkRefsIntoClosure notOlder(_span, verification_mark_bm()); |
0 | 2915 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. |
2916 | |
2917 gch->gen_process_strong_roots(_cmsGen->level(), | |
2918 true, // younger gens are roots | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
2919 true, // activate StrongRootsScope |
0 | 2920 true, // collecting perm gen |
2921 SharedHeap::ScanningOption(roots_scanning_options()), | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
2922 ¬Older, |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
2923 true, // walk code active on stacks |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
2924 NULL); |
0 | 2925 |
2926 // Now mark from the roots | |
2927 assert(_revisitStack.isEmpty(), "Should be empty"); | |
2928 MarkFromRootsClosure markFromRootsClosure(this, _span, | |
2929 verification_mark_bm(), verification_mark_stack(), &_revisitStack, | |
2930 false /* don't yield */, true /* verifying */); | |
2931 assert(_restart_addr == NULL, "Expected pre-condition"); | |
2932 verification_mark_bm()->iterate(&markFromRootsClosure); | |
2933 while (_restart_addr != NULL) { | |
2934 // Deal with stack overflow: by restarting at the indicated | |
2935 // address. | |
2936 HeapWord* ra = _restart_addr; | |
2937 markFromRootsClosure.reset(ra); | |
2938 _restart_addr = NULL; | |
2939 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); | |
2940 } | |
2941 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); | |
2942 verify_work_stacks_empty(); | |
2943 // Should reset the revisit stack above, since no class tree | |
2944 // surgery is forthcoming. | |
2945 _revisitStack.reset(); // throwing away all contents | |
2946 | |
2947 // Marking completed -- now verify that each bit marked in | |
2948 // verification_mark_bm() is also marked in markBitMap(); flag all | |
2949 // errors by printing corresponding objects. | |
2950 VerifyMarkedClosure vcl(markBitMap()); | |
2951 verification_mark_bm()->iterate(&vcl); | |
2952 if (vcl.failed()) { | |
2953 gclog_or_tty->print("Verification failed"); | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2954 Universe::heap()->print_on(gclog_or_tty); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
2955 fatal("CMS: failed marking verification after remark"); |
0 | 2956 } |
2957 } | |
2958 | |
2959 void CMSCollector::verify_after_remark_work_2() { | |
2960 ResourceMark rm; | |
2961 HandleMark hm; | |
2962 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
2963 | |
2964 // Mark from roots one level into CMS | |
2965 MarkRefsIntoVerifyClosure notOlder(_span, verification_mark_bm(), | |
994
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
2966 markBitMap()); |
0 | 2967 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. |
2968 gch->gen_process_strong_roots(_cmsGen->level(), | |
2969 true, // younger gens are roots | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
2970 true, // activate StrongRootsScope |
0 | 2971 true, // collecting perm gen |
2972 SharedHeap::ScanningOption(roots_scanning_options()), | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
2973 ¬Older, |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
2974 true, // walk code active on stacks |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
2975 NULL); |
0 | 2976 |
2977 // Now mark from the roots | |
2978 assert(_revisitStack.isEmpty(), "Should be empty"); | |
2979 MarkFromRootsVerifyClosure markFromRootsClosure(this, _span, | |
2980 verification_mark_bm(), markBitMap(), verification_mark_stack()); | |
2981 assert(_restart_addr == NULL, "Expected pre-condition"); | |
2982 verification_mark_bm()->iterate(&markFromRootsClosure); | |
2983 while (_restart_addr != NULL) { | |
2984 // Deal with stack overflow: by restarting at the indicated | |
2985 // address. | |
2986 HeapWord* ra = _restart_addr; | |
2987 markFromRootsClosure.reset(ra); | |
2988 _restart_addr = NULL; | |
2989 verification_mark_bm()->iterate(&markFromRootsClosure, ra, _span.end()); | |
2990 } | |
2991 assert(verification_mark_stack()->isEmpty(), "Should have been drained"); | |
2992 verify_work_stacks_empty(); | |
2993 // Should reset the revisit stack above, since no class tree | |
2994 // surgery is forthcoming. | |
2995 _revisitStack.reset(); // throwing away all contents | |
2996 | |
2997 // Marking completed -- now verify that each bit marked in | |
2998 // verification_mark_bm() is also marked in markBitMap(); flag all | |
2999 // errors by printing corresponding objects. | |
3000 VerifyMarkedClosure vcl(markBitMap()); | |
3001 verification_mark_bm()->iterate(&vcl); | |
3002 assert(!vcl.failed(), "Else verification above should not have succeeded"); | |
3003 } | |
3004 | |
3005 void ConcurrentMarkSweepGeneration::save_marks() { | |
3006 // delegate to CMS space | |
3007 cmsSpace()->save_marks(); | |
3008 for (uint i = 0; i < ParallelGCThreads; i++) { | |
3009 _par_gc_thread_states[i]->promo.startTrackingPromotions(); | |
3010 } | |
3011 } | |
3012 | |
3013 bool ConcurrentMarkSweepGeneration::no_allocs_since_save_marks() { | |
3014 return cmsSpace()->no_allocs_since_save_marks(); | |
3015 } | |
3016 | |
3017 #define CMS_SINCE_SAVE_MARKS_DEFN(OopClosureType, nv_suffix) \ | |
3018 \ | |
3019 void ConcurrentMarkSweepGeneration:: \ | |
3020 oop_since_save_marks_iterate##nv_suffix(OopClosureType* cl) { \ | |
3021 cl->set_generation(this); \ | |
3022 cmsSpace()->oop_since_save_marks_iterate##nv_suffix(cl); \ | |
3023 cl->reset_generation(); \ | |
3024 save_marks(); \ | |
3025 } | |
3026 | |
3027 ALL_SINCE_SAVE_MARKS_CLOSURES(CMS_SINCE_SAVE_MARKS_DEFN) | |
3028 | |
3029 void | |
3030 ConcurrentMarkSweepGeneration::object_iterate_since_last_GC(ObjectClosure* blk) | |
3031 { | |
3032 // Not currently implemented; need to do the following. -- ysr. | |
3033 // dld -- I think that is used for some sort of allocation profiler. So it | |
3034 // really means the objects allocated by the mutator since the last | |
3035 // GC. We could potentially implement this cheaply by recording only | |
3036 // the direct allocations in a side data structure. | |
3037 // | |
3038 // I think we probably ought not to be required to support these | |
3039 // iterations at any arbitrary point; I think there ought to be some | |
3040 // call to enable/disable allocation profiling in a generation/space, | |
3041 // and the iterator ought to return the objects allocated in the | |
3042 // gen/space since the enable call, or the last iterator call (which | |
3043 // will probably be at a GC.) That way, for gens like CM&S that would | |
3044 // require some extra data structure to support this, we only pay the | |
3045 // cost when it's in use... | |
3046 cmsSpace()->object_iterate_since_last_GC(blk); | |
3047 } | |
3048 | |
3049 void | |
3050 ConcurrentMarkSweepGeneration::younger_refs_iterate(OopsInGenClosure* cl) { | |
3051 cl->set_generation(this); | |
3052 younger_refs_in_space_iterate(_cmsSpace, cl); | |
3053 cl->reset_generation(); | |
3054 } | |
3055 | |
3056 void | |
3057 ConcurrentMarkSweepGeneration::oop_iterate(MemRegion mr, OopClosure* cl) { | |
3058 if (freelistLock()->owned_by_self()) { | |
3059 Generation::oop_iterate(mr, cl); | |
3060 } else { | |
3061 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); | |
3062 Generation::oop_iterate(mr, cl); | |
3063 } | |
3064 } | |
3065 | |
3066 void | |
3067 ConcurrentMarkSweepGeneration::oop_iterate(OopClosure* cl) { | |
3068 if (freelistLock()->owned_by_self()) { | |
3069 Generation::oop_iterate(cl); | |
3070 } else { | |
3071 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); | |
3072 Generation::oop_iterate(cl); | |
3073 } | |
3074 } | |
3075 | |
3076 void | |
3077 ConcurrentMarkSweepGeneration::object_iterate(ObjectClosure* cl) { | |
3078 if (freelistLock()->owned_by_self()) { | |
3079 Generation::object_iterate(cl); | |
3080 } else { | |
3081 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); | |
3082 Generation::object_iterate(cl); | |
3083 } | |
3084 } | |
3085 | |
3086 void | |
517
e9be0e04635a
6689653: JMapPerm fails with UseConcMarkSweepIncGC and compressed oops off
jmasa
parents:
457
diff
changeset
|
3087 ConcurrentMarkSweepGeneration::safe_object_iterate(ObjectClosure* cl) { |
e9be0e04635a
6689653: JMapPerm fails with UseConcMarkSweepIncGC and compressed oops off
jmasa
parents:
457
diff
changeset
|
3088 if (freelistLock()->owned_by_self()) { |
e9be0e04635a
6689653: JMapPerm fails with UseConcMarkSweepIncGC and compressed oops off
jmasa
parents:
457
diff
changeset
|
3089 Generation::safe_object_iterate(cl); |
e9be0e04635a
6689653: JMapPerm fails with UseConcMarkSweepIncGC and compressed oops off
jmasa
parents:
457
diff
changeset
|
3090 } else { |
e9be0e04635a
6689653: JMapPerm fails with UseConcMarkSweepIncGC and compressed oops off
jmasa
parents:
457
diff
changeset
|
3091 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); |
e9be0e04635a
6689653: JMapPerm fails with UseConcMarkSweepIncGC and compressed oops off
jmasa
parents:
457
diff
changeset
|
3092 Generation::safe_object_iterate(cl); |
e9be0e04635a
6689653: JMapPerm fails with UseConcMarkSweepIncGC and compressed oops off
jmasa
parents:
457
diff
changeset
|
3093 } |
e9be0e04635a
6689653: JMapPerm fails with UseConcMarkSweepIncGC and compressed oops off
jmasa
parents:
457
diff
changeset
|
3094 } |
e9be0e04635a
6689653: JMapPerm fails with UseConcMarkSweepIncGC and compressed oops off
jmasa
parents:
457
diff
changeset
|
3095 |
e9be0e04635a
6689653: JMapPerm fails with UseConcMarkSweepIncGC and compressed oops off
jmasa
parents:
457
diff
changeset
|
3096 void |
0 | 3097 ConcurrentMarkSweepGeneration::pre_adjust_pointers() { |
3098 } | |
3099 | |
3100 void | |
3101 ConcurrentMarkSweepGeneration::post_compact() { | |
3102 } | |
3103 | |
3104 void | |
3105 ConcurrentMarkSweepGeneration::prepare_for_verify() { | |
3106 // Fix the linear allocation blocks to look like free blocks. | |
3107 | |
3108 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those | |
3109 // are not called when the heap is verified during universe initialization and | |
3110 // at vm shutdown. | |
3111 if (freelistLock()->owned_by_self()) { | |
3112 cmsSpace()->prepare_for_verify(); | |
3113 } else { | |
3114 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); | |
3115 cmsSpace()->prepare_for_verify(); | |
3116 } | |
3117 } | |
3118 | |
3119 void | |
3120 ConcurrentMarkSweepGeneration::verify(bool allow_dirty /* ignored */) { | |
3121 // Locks are normally acquired/released in gc_prologue/gc_epilogue, but those | |
3122 // are not called when the heap is verified during universe initialization and | |
3123 // at vm shutdown. | |
3124 if (freelistLock()->owned_by_self()) { | |
3125 cmsSpace()->verify(false /* ignored */); | |
3126 } else { | |
3127 MutexLockerEx fll(freelistLock(), Mutex::_no_safepoint_check_flag); | |
3128 cmsSpace()->verify(false /* ignored */); | |
3129 } | |
3130 } | |
3131 | |
3132 void CMSCollector::verify(bool allow_dirty /* ignored */) { | |
3133 _cmsGen->verify(allow_dirty); | |
3134 _permGen->verify(allow_dirty); | |
3135 } | |
3136 | |
3137 #ifndef PRODUCT | |
3138 bool CMSCollector::overflow_list_is_empty() const { | |
3139 assert(_num_par_pushes >= 0, "Inconsistency"); | |
3140 if (_overflow_list == NULL) { | |
3141 assert(_num_par_pushes == 0, "Inconsistency"); | |
3142 } | |
3143 return _overflow_list == NULL; | |
3144 } | |
3145 | |
3146 // The methods verify_work_stacks_empty() and verify_overflow_empty() | |
3147 // merely consolidate assertion checks that appear to occur together frequently. | |
3148 void CMSCollector::verify_work_stacks_empty() const { | |
3149 assert(_markStack.isEmpty(), "Marking stack should be empty"); | |
3150 assert(overflow_list_is_empty(), "Overflow list should be empty"); | |
3151 } | |
3152 | |
3153 void CMSCollector::verify_overflow_empty() const { | |
3154 assert(overflow_list_is_empty(), "Overflow list should be empty"); | |
3155 assert(no_preserved_marks(), "No preserved marks"); | |
3156 } | |
3157 #endif // PRODUCT | |
3158 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3159 // Decide if we want to enable class unloading as part of the |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3160 // ensuing concurrent GC cycle. We will collect the perm gen and |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3161 // unload classes if it's the case that: |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3162 // (1) an explicit gc request has been made and the flag |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3163 // ExplicitGCInvokesConcurrentAndUnloadsClasses is set, OR |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3164 // (2) (a) class unloading is enabled at the command line, and |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3165 // (b) (i) perm gen threshold has been crossed, or |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3166 // (ii) old gen is getting really full, or |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3167 // (iii) the previous N CMS collections did not collect the |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3168 // perm gen |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3169 // NOTE: Provided there is no change in the state of the heap between |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3170 // calls to this method, it should have idempotent results. Moreover, |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3171 // its results should be monotonically increasing (i.e. going from 0 to 1, |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3172 // but not 1 to 0) between successive calls between which the heap was |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3173 // not collected. For the implementation below, it must thus rely on |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3174 // the property that concurrent_cycles_since_last_unload() |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3175 // will not decrease unless a collection cycle happened and that |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3176 // _permGen->should_concurrent_collect() and _cmsGen->is_too_full() are |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3177 // themselves also monotonic in that sense. See check_monotonicity() |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3178 // below. |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3179 bool CMSCollector::update_should_unload_classes() { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3180 _should_unload_classes = false; |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3181 // Condition 1 above |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3182 if (_full_gc_requested && ExplicitGCInvokesConcurrentAndUnloadsClasses) { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3183 _should_unload_classes = true; |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3184 } else if (CMSClassUnloadingEnabled) { // Condition 2.a above |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3185 // Disjuncts 2.b.(i,ii,iii) above |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3186 _should_unload_classes = (concurrent_cycles_since_last_unload() >= |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3187 CMSClassUnloadingMaxInterval) |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3188 || _permGen->should_concurrent_collect() |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3189 || _cmsGen->is_too_full(); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3190 } |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3191 return _should_unload_classes; |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3192 } |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3193 |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3194 bool ConcurrentMarkSweepGeneration::is_too_full() const { |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3195 bool res = should_concurrent_collect(); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3196 res = res && (occupancy() > (double)CMSIsTooFullPercentage/100.0); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3197 return res; |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3198 } |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3199 |
0 | 3200 void CMSCollector::setup_cms_unloading_and_verification_state() { |
3201 const bool should_verify = VerifyBeforeGC || VerifyAfterGC || VerifyDuringGC | |
3202 || VerifyBeforeExit; | |
3203 const int rso = SharedHeap::SO_Symbols | SharedHeap::SO_Strings | |
3204 | SharedHeap::SO_CodeCache; | |
3205 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3206 if (should_unload_classes()) { // Should unload classes this cycle |
0 | 3207 remove_root_scanning_option(rso); // Shrink the root set appropriately |
3208 set_verifying(should_verify); // Set verification state for this cycle | |
3209 return; // Nothing else needs to be done at this time | |
3210 } | |
3211 | |
3212 // Not unloading classes this cycle | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3213 assert(!should_unload_classes(), "Inconsitency!"); |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
3214 if ((!verifying() || unloaded_classes_last_cycle()) && should_verify) { |
0 | 3215 // We were not verifying, or we _were_ unloading classes in the last cycle, |
3216 // AND some verification options are enabled this cycle; in this case, | |
3217 // we must make sure that the deadness map is allocated if not already so, | |
3218 // and cleared (if already allocated previously -- | |
3219 // CMSBitMap::sizeInBits() is used to determine if it's allocated). | |
3220 if (perm_gen_verify_bit_map()->sizeInBits() == 0) { | |
3221 if (!perm_gen_verify_bit_map()->allocate(_permGen->reserved())) { | |
3222 warning("Failed to allocate permanent generation verification CMS Bit Map;\n" | |
3223 "permanent generation verification disabled"); | |
3224 return; // Note that we leave verification disabled, so we'll retry this | |
3225 // allocation next cycle. We _could_ remember this failure | |
3226 // and skip further attempts and permanently disable verification | |
3227 // attempts if that is considered more desirable. | |
3228 } | |
3229 assert(perm_gen_verify_bit_map()->covers(_permGen->reserved()), | |
3230 "_perm_gen_ver_bit_map inconsistency?"); | |
3231 } else { | |
3232 perm_gen_verify_bit_map()->clear_all(); | |
3233 } | |
3234 // Include symbols, strings and code cache elements to prevent their resurrection. | |
3235 add_root_scanning_option(rso); | |
3236 set_verifying(true); | |
3237 } else if (verifying() && !should_verify) { | |
3238 // We were verifying, but some verification flags got disabled. | |
3239 set_verifying(false); | |
3240 // Exclude symbols, strings and code cache elements from root scanning to | |
3241 // reduce IM and RM pauses. | |
3242 remove_root_scanning_option(rso); | |
3243 } | |
3244 } | |
3245 | |
3246 | |
3247 #ifndef PRODUCT | |
3248 HeapWord* CMSCollector::block_start(const void* p) const { | |
3249 const HeapWord* addr = (HeapWord*)p; | |
3250 if (_span.contains(p)) { | |
3251 if (_cmsGen->cmsSpace()->is_in_reserved(addr)) { | |
3252 return _cmsGen->cmsSpace()->block_start(p); | |
3253 } else { | |
3254 assert(_permGen->cmsSpace()->is_in_reserved(addr), | |
3255 "Inconsistent _span?"); | |
3256 return _permGen->cmsSpace()->block_start(p); | |
3257 } | |
3258 } | |
3259 return NULL; | |
3260 } | |
3261 #endif | |
3262 | |
3263 HeapWord* | |
3264 ConcurrentMarkSweepGeneration::expand_and_allocate(size_t word_size, | |
3265 bool tlab, | |
3266 bool parallel) { | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3267 CMSSynchronousYieldRequest yr; |
0 | 3268 assert(!tlab, "Can't deal with TLAB allocation"); |
3269 MutexLockerEx x(freelistLock(), Mutex::_no_safepoint_check_flag); | |
3270 expand(word_size*HeapWordSize, MinHeapDeltaBytes, | |
3271 CMSExpansionCause::_satisfy_allocation); | |
3272 if (GCExpandToAllocateDelayMillis > 0) { | |
3273 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); | |
3274 } | |
9
173195ff483a
6642634: Test nsk/regression/b6186200 crashed with SIGSEGV
ysr
parents:
7
diff
changeset
|
3275 return have_lock_and_allocate(word_size, tlab); |
0 | 3276 } |
3277 | |
3278 // YSR: All of this generation expansion/shrinking stuff is an exact copy of | |
3279 // OneContigSpaceCardGeneration, which makes me wonder if we should move this | |
3280 // to CardGeneration and share it... | |
271
818a18cd69a8
6730514: assertion failure in mangling code when expanding by 0 bytes
jmasa
parents:
196
diff
changeset
|
3281 bool ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes) { |
818a18cd69a8
6730514: assertion failure in mangling code when expanding by 0 bytes
jmasa
parents:
196
diff
changeset
|
3282 return CardGeneration::expand(bytes, expand_bytes); |
818a18cd69a8
6730514: assertion failure in mangling code when expanding by 0 bytes
jmasa
parents:
196
diff
changeset
|
3283 } |
818a18cd69a8
6730514: assertion failure in mangling code when expanding by 0 bytes
jmasa
parents:
196
diff
changeset
|
3284 |
0 | 3285 void ConcurrentMarkSweepGeneration::expand(size_t bytes, size_t expand_bytes, |
3286 CMSExpansionCause::Cause cause) | |
3287 { | |
271
818a18cd69a8
6730514: assertion failure in mangling code when expanding by 0 bytes
jmasa
parents:
196
diff
changeset
|
3288 |
818a18cd69a8
6730514: assertion failure in mangling code when expanding by 0 bytes
jmasa
parents:
196
diff
changeset
|
3289 bool success = expand(bytes, expand_bytes); |
818a18cd69a8
6730514: assertion failure in mangling code when expanding by 0 bytes
jmasa
parents:
196
diff
changeset
|
3290 |
0 | 3291 // remember why we expanded; this information is used |
3292 // by shouldConcurrentCollect() when making decisions on whether to start | |
3293 // a new CMS cycle. | |
3294 if (success) { | |
3295 set_expansion_cause(cause); | |
3296 if (PrintGCDetails && Verbose) { | |
3297 gclog_or_tty->print_cr("Expanded CMS gen for %s", | |
3298 CMSExpansionCause::to_string(cause)); | |
3299 } | |
3300 } | |
3301 } | |
3302 | |
3303 HeapWord* ConcurrentMarkSweepGeneration::expand_and_par_lab_allocate(CMSParGCThreadState* ps, size_t word_sz) { | |
3304 HeapWord* res = NULL; | |
3305 MutexLocker x(ParGCRareEvent_lock); | |
3306 while (true) { | |
3307 // Expansion by some other thread might make alloc OK now: | |
3308 res = ps->lab.alloc(word_sz); | |
3309 if (res != NULL) return res; | |
3310 // If there's not enough expansion space available, give up. | |
3311 if (_virtual_space.uncommitted_size() < (word_sz * HeapWordSize)) { | |
3312 return NULL; | |
3313 } | |
3314 // Otherwise, we try expansion. | |
3315 expand(word_sz*HeapWordSize, MinHeapDeltaBytes, | |
3316 CMSExpansionCause::_allocate_par_lab); | |
3317 // Now go around the loop and try alloc again; | |
3318 // A competing par_promote might beat us to the expansion space, | |
3319 // so we may go around the loop again if promotion fails agaion. | |
3320 if (GCExpandToAllocateDelayMillis > 0) { | |
3321 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); | |
3322 } | |
3323 } | |
3324 } | |
3325 | |
3326 | |
3327 bool ConcurrentMarkSweepGeneration::expand_and_ensure_spooling_space( | |
3328 PromotionInfo* promo) { | |
3329 MutexLocker x(ParGCRareEvent_lock); | |
3330 size_t refill_size_bytes = promo->refillSize() * HeapWordSize; | |
3331 while (true) { | |
3332 // Expansion by some other thread might make alloc OK now: | |
3333 if (promo->ensure_spooling_space()) { | |
3334 assert(promo->has_spooling_space(), | |
3335 "Post-condition of successful ensure_spooling_space()"); | |
3336 return true; | |
3337 } | |
3338 // If there's not enough expansion space available, give up. | |
3339 if (_virtual_space.uncommitted_size() < refill_size_bytes) { | |
3340 return false; | |
3341 } | |
3342 // Otherwise, we try expansion. | |
3343 expand(refill_size_bytes, MinHeapDeltaBytes, | |
3344 CMSExpansionCause::_allocate_par_spooling_space); | |
3345 // Now go around the loop and try alloc again; | |
3346 // A competing allocation might beat us to the expansion space, | |
3347 // so we may go around the loop again if allocation fails again. | |
3348 if (GCExpandToAllocateDelayMillis > 0) { | |
3349 os::sleep(Thread::current(), GCExpandToAllocateDelayMillis, false); | |
3350 } | |
3351 } | |
3352 } | |
3353 | |
3354 | |
3355 | |
3356 void ConcurrentMarkSweepGeneration::shrink(size_t bytes) { | |
3357 assert_locked_or_safepoint(Heap_lock); | |
3358 size_t size = ReservedSpace::page_align_size_down(bytes); | |
3359 if (size > 0) { | |
3360 shrink_by(size); | |
3361 } | |
3362 } | |
3363 | |
3364 bool ConcurrentMarkSweepGeneration::grow_by(size_t bytes) { | |
3365 assert_locked_or_safepoint(Heap_lock); | |
3366 bool result = _virtual_space.expand_by(bytes); | |
3367 if (result) { | |
3368 HeapWord* old_end = _cmsSpace->end(); | |
3369 size_t new_word_size = | |
3370 heap_word_size(_virtual_space.committed_size()); | |
3371 MemRegion mr(_cmsSpace->bottom(), new_word_size); | |
3372 _bts->resize(new_word_size); // resize the block offset shared array | |
3373 Universe::heap()->barrier_set()->resize_covered_region(mr); | |
3374 // Hmmmm... why doesn't CFLS::set_end verify locking? | |
3375 // This is quite ugly; FIX ME XXX | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
3376 _cmsSpace->assert_locked(freelistLock()); |
0 | 3377 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); |
3378 | |
3379 // update the space and generation capacity counters | |
3380 if (UsePerfData) { | |
3381 _space_counters->update_capacity(); | |
3382 _gen_counters->update_all(); | |
3383 } | |
3384 | |
3385 if (Verbose && PrintGC) { | |
3386 size_t new_mem_size = _virtual_space.committed_size(); | |
3387 size_t old_mem_size = new_mem_size - bytes; | |
3388 gclog_or_tty->print_cr("Expanding %s from %ldK by %ldK to %ldK", | |
3389 name(), old_mem_size/K, bytes/K, new_mem_size/K); | |
3390 } | |
3391 } | |
3392 return result; | |
3393 } | |
3394 | |
3395 bool ConcurrentMarkSweepGeneration::grow_to_reserved() { | |
3396 assert_locked_or_safepoint(Heap_lock); | |
3397 bool success = true; | |
3398 const size_t remaining_bytes = _virtual_space.uncommitted_size(); | |
3399 if (remaining_bytes > 0) { | |
3400 success = grow_by(remaining_bytes); | |
3401 DEBUG_ONLY(if (!success) warning("grow to reserved failed");) | |
3402 } | |
3403 return success; | |
3404 } | |
3405 | |
3406 void ConcurrentMarkSweepGeneration::shrink_by(size_t bytes) { | |
3407 assert_locked_or_safepoint(Heap_lock); | |
3408 assert_lock_strong(freelistLock()); | |
3409 // XXX Fix when compaction is implemented. | |
3410 warning("Shrinking of CMS not yet implemented"); | |
3411 return; | |
3412 } | |
3413 | |
3414 | |
3415 // Simple ctor/dtor wrapper for accounting & timer chores around concurrent | |
3416 // phases. | |
3417 class CMSPhaseAccounting: public StackObj { | |
3418 public: | |
3419 CMSPhaseAccounting(CMSCollector *collector, | |
3420 const char *phase, | |
3421 bool print_cr = true); | |
3422 ~CMSPhaseAccounting(); | |
3423 | |
3424 private: | |
3425 CMSCollector *_collector; | |
3426 const char *_phase; | |
3427 elapsedTimer _wallclock; | |
3428 bool _print_cr; | |
3429 | |
3430 public: | |
3431 // Not MT-safe; so do not pass around these StackObj's | |
3432 // where they may be accessed by other threads. | |
3433 jlong wallclock_millis() { | |
3434 assert(_wallclock.is_active(), "Wall clock should not stop"); | |
3435 _wallclock.stop(); // to record time | |
3436 jlong ret = _wallclock.milliseconds(); | |
3437 _wallclock.start(); // restart | |
3438 return ret; | |
3439 } | |
3440 }; | |
3441 | |
3442 CMSPhaseAccounting::CMSPhaseAccounting(CMSCollector *collector, | |
3443 const char *phase, | |
3444 bool print_cr) : | |
3445 _collector(collector), _phase(phase), _print_cr(print_cr) { | |
3446 | |
3447 if (PrintCMSStatistics != 0) { | |
3448 _collector->resetYields(); | |
3449 } | |
3450 if (PrintGCDetails && PrintGCTimeStamps) { | |
3451 gclog_or_tty->date_stamp(PrintGCDateStamps); | |
3452 gclog_or_tty->stamp(); | |
3453 gclog_or_tty->print_cr(": [%s-concurrent-%s-start]", | |
3454 _collector->cmsGen()->short_name(), _phase); | |
3455 } | |
3456 _collector->resetTimer(); | |
3457 _wallclock.start(); | |
3458 _collector->startTimer(); | |
3459 } | |
3460 | |
3461 CMSPhaseAccounting::~CMSPhaseAccounting() { | |
3462 assert(_wallclock.is_active(), "Wall clock should not have stopped"); | |
3463 _collector->stopTimer(); | |
3464 _wallclock.stop(); | |
3465 if (PrintGCDetails) { | |
3466 gclog_or_tty->date_stamp(PrintGCDateStamps); | |
3467 if (PrintGCTimeStamps) { | |
3468 gclog_or_tty->stamp(); | |
3469 gclog_or_tty->print(": "); | |
3470 } | |
3471 gclog_or_tty->print("[%s-concurrent-%s: %3.3f/%3.3f secs]", | |
3472 _collector->cmsGen()->short_name(), | |
3473 _phase, _collector->timerValue(), _wallclock.seconds()); | |
3474 if (_print_cr) { | |
3475 gclog_or_tty->print_cr(""); | |
3476 } | |
3477 if (PrintCMSStatistics != 0) { | |
3478 gclog_or_tty->print_cr(" (CMS-concurrent-%s yielded %d times)", _phase, | |
3479 _collector->yields()); | |
3480 } | |
3481 } | |
3482 } | |
3483 | |
3484 // CMS work | |
3485 | |
3486 // Checkpoint the roots into this generation from outside | |
3487 // this generation. [Note this initial checkpoint need only | |
3488 // be approximate -- we'll do a catch up phase subsequently.] | |
3489 void CMSCollector::checkpointRootsInitial(bool asynch) { | |
3490 assert(_collectorState == InitialMarking, "Wrong collector state"); | |
3491 check_correct_thread_executing(); | |
1703
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
3492 TraceCMSMemoryManagerStats tms(_collectorState); |
0 | 3493 ReferenceProcessor* rp = ref_processor(); |
3494 SpecializationStats::clear(); | |
3495 assert(_restart_addr == NULL, "Control point invariant"); | |
3496 if (asynch) { | |
3497 // acquire locks for subsequent manipulations | |
3498 MutexLockerEx x(bitMapLock(), | |
3499 Mutex::_no_safepoint_check_flag); | |
3500 checkpointRootsInitialWork(asynch); | |
3501 rp->verify_no_references_recorded(); | |
3502 rp->enable_discovery(); // enable ("weak") refs discovery | |
3503 _collectorState = Marking; | |
3504 } else { | |
3505 // (Weak) Refs discovery: this is controlled from genCollectedHeap::do_collection | |
3506 // which recognizes if we are a CMS generation, and doesn't try to turn on | |
3507 // discovery; verify that they aren't meddling. | |
3508 assert(!rp->discovery_is_atomic(), | |
3509 "incorrect setting of discovery predicate"); | |
3510 assert(!rp->discovery_enabled(), "genCollectedHeap shouldn't control " | |
3511 "ref discovery for this generation kind"); | |
3512 // already have locks | |
3513 checkpointRootsInitialWork(asynch); | |
3514 rp->enable_discovery(); // now enable ("weak") refs discovery | |
3515 _collectorState = Marking; | |
3516 } | |
3517 SpecializationStats::print(); | |
3518 } | |
3519 | |
3520 void CMSCollector::checkpointRootsInitialWork(bool asynch) { | |
3521 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); | |
3522 assert(_collectorState == InitialMarking, "just checking"); | |
3523 | |
3524 // If there has not been a GC[n-1] since last GC[n] cycle completed, | |
3525 // precede our marking with a collection of all | |
3526 // younger generations to keep floating garbage to a minimum. | |
3527 // XXX: we won't do this for now -- it's an optimization to be done later. | |
3528 | |
3529 // already have locks | |
3530 assert_lock_strong(bitMapLock()); | |
3531 assert(_markBitMap.isAllClear(), "was reset at end of previous cycle"); | |
3532 | |
3533 // Setup the verification and class unloading state for this | |
3534 // CMS collection cycle. | |
3535 setup_cms_unloading_and_verification_state(); | |
3536 | |
3537 NOT_PRODUCT(TraceTime t("\ncheckpointRootsInitialWork", | |
3538 PrintGCDetails && Verbose, true, gclog_or_tty);) | |
3539 if (UseAdaptiveSizePolicy) { | |
3540 size_policy()->checkpoint_roots_initial_begin(); | |
3541 } | |
3542 | |
3543 // Reset all the PLAB chunk arrays if necessary. | |
3544 if (_survivor_plab_array != NULL && !CMSPLABRecordAlways) { | |
3545 reset_survivor_plab_arrays(); | |
3546 } | |
3547 | |
3548 ResourceMark rm; | |
3549 HandleMark hm; | |
3550 | |
3551 FalseClosure falseClosure; | |
3552 // In the case of a synchronous collection, we will elide the | |
3553 // remark step, so it's important to catch all the nmethod oops | |
994
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
3554 // in this step. |
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
3555 // The final 'true' flag to gen_process_strong_roots will ensure this. |
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
3556 // If 'async' is true, we can relax the nmethod tracing. |
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
3557 MarkRefsIntoClosure notOlder(_span, &_markBitMap); |
0 | 3558 GenCollectedHeap* gch = GenCollectedHeap::heap(); |
3559 | |
3560 verify_work_stacks_empty(); | |
3561 verify_overflow_empty(); | |
3562 | |
3563 gch->ensure_parsability(false); // fill TLABs, but no need to retire them | |
3564 // Update the saved marks which may affect the root scans. | |
3565 gch->save_marks(); | |
3566 | |
3567 // weak reference processing has not started yet. | |
3568 ref_processor()->set_enqueuing_is_done(false); | |
3569 | |
3570 { | |
935 | 3571 // This is not needed. DEBUG_ONLY(RememberKlassesChecker imx(true);) |
0 | 3572 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) |
3573 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. | |
3574 gch->gen_process_strong_roots(_cmsGen->level(), | |
3575 true, // younger gens are roots | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
3576 true, // activate StrongRootsScope |
0 | 3577 true, // collecting perm gen |
3578 SharedHeap::ScanningOption(roots_scanning_options()), | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
3579 ¬Older, |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
3580 true, // walk all of code cache if (so & SO_CodeCache) |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
3581 NULL); |
0 | 3582 } |
3583 | |
3584 // Clear mod-union table; it will be dirtied in the prologue of | |
3585 // CMS generation per each younger generation collection. | |
3586 | |
3587 assert(_modUnionTable.isAllClear(), | |
3588 "Was cleared in most recent final checkpoint phase" | |
3589 " or no bits are set in the gc_prologue before the start of the next " | |
3590 "subsequent marking phase."); | |
3591 | |
3592 // Temporarily disabled, since pre/post-consumption closures don't | |
3593 // care about precleaned cards | |
3594 #if 0 | |
3595 { | |
3596 MemRegion mr = MemRegion((HeapWord*)_virtual_space.low(), | |
3597 (HeapWord*)_virtual_space.high()); | |
3598 _ct->ct_bs()->preclean_dirty_cards(mr); | |
3599 } | |
3600 #endif | |
3601 | |
3602 // Save the end of the used_region of the constituent generations | |
3603 // to be used to limit the extent of sweep in each generation. | |
3604 save_sweep_limits(); | |
3605 if (UseAdaptiveSizePolicy) { | |
3606 size_policy()->checkpoint_roots_initial_end(gch->gc_cause()); | |
3607 } | |
3608 verify_overflow_empty(); | |
3609 } | |
3610 | |
3611 bool CMSCollector::markFromRoots(bool asynch) { | |
3612 // we might be tempted to assert that: | |
3613 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), | |
3614 // "inconsistent argument?"); | |
3615 // However that wouldn't be right, because it's possible that | |
3616 // a safepoint is indeed in progress as a younger generation | |
3617 // stop-the-world GC happens even as we mark in this generation. | |
3618 assert(_collectorState == Marking, "inconsistent state?"); | |
3619 check_correct_thread_executing(); | |
3620 verify_overflow_empty(); | |
3621 | |
3622 bool res; | |
3623 if (asynch) { | |
3624 | |
3625 // Start the timers for adaptive size policy for the concurrent phases | |
3626 // Do it here so that the foreground MS can use the concurrent | |
3627 // timer since a foreground MS might has the sweep done concurrently | |
3628 // or STW. | |
3629 if (UseAdaptiveSizePolicy) { | |
3630 size_policy()->concurrent_marking_begin(); | |
3631 } | |
3632 | |
3633 // Weak ref discovery note: We may be discovering weak | |
3634 // refs in this generation concurrent (but interleaved) with | |
3635 // weak ref discovery by a younger generation collector. | |
3636 | |
3637 CMSTokenSyncWithLocks ts(true, bitMapLock()); | |
3638 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); | |
3639 CMSPhaseAccounting pa(this, "mark", !PrintGCDetails); | |
3640 res = markFromRootsWork(asynch); | |
3641 if (res) { | |
3642 _collectorState = Precleaning; | |
3643 } else { // We failed and a foreground collection wants to take over | |
3644 assert(_foregroundGCIsActive, "internal state inconsistency"); | |
3645 assert(_restart_addr == NULL, "foreground will restart from scratch"); | |
3646 if (PrintGCDetails) { | |
3647 gclog_or_tty->print_cr("bailing out to foreground collection"); | |
3648 } | |
3649 } | |
3650 if (UseAdaptiveSizePolicy) { | |
3651 size_policy()->concurrent_marking_end(); | |
3652 } | |
3653 } else { | |
3654 assert(SafepointSynchronize::is_at_safepoint(), | |
3655 "inconsistent with asynch == false"); | |
3656 if (UseAdaptiveSizePolicy) { | |
3657 size_policy()->ms_collection_marking_begin(); | |
3658 } | |
3659 // already have locks | |
3660 res = markFromRootsWork(asynch); | |
3661 _collectorState = FinalMarking; | |
3662 if (UseAdaptiveSizePolicy) { | |
3663 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
3664 size_policy()->ms_collection_marking_end(gch->gc_cause()); | |
3665 } | |
3666 } | |
3667 verify_overflow_empty(); | |
3668 return res; | |
3669 } | |
3670 | |
3671 bool CMSCollector::markFromRootsWork(bool asynch) { | |
3672 // iterate over marked bits in bit map, doing a full scan and mark | |
3673 // from these roots using the following algorithm: | |
3674 // . if oop is to the right of the current scan pointer, | |
3675 // mark corresponding bit (we'll process it later) | |
3676 // . else (oop is to left of current scan pointer) | |
3677 // push oop on marking stack | |
3678 // . drain the marking stack | |
3679 | |
3680 // Note that when we do a marking step we need to hold the | |
3681 // bit map lock -- recall that direct allocation (by mutators) | |
3682 // and promotion (by younger generation collectors) is also | |
3683 // marking the bit map. [the so-called allocate live policy.] | |
3684 // Because the implementation of bit map marking is not | |
3685 // robust wrt simultaneous marking of bits in the same word, | |
3686 // we need to make sure that there is no such interference | |
3687 // between concurrent such updates. | |
3688 | |
3689 // already have locks | |
3690 assert_lock_strong(bitMapLock()); | |
3691 | |
3692 // Clear the revisit stack, just in case there are any | |
3693 // obsolete contents from a short-circuited previous CMS cycle. | |
3694 _revisitStack.reset(); | |
3695 verify_work_stacks_empty(); | |
3696 verify_overflow_empty(); | |
3697 assert(_revisitStack.isEmpty(), "tabula rasa"); | |
1190
4788266644c1
6895236: CMS: cmsOopClosures.inline.hpp:43 assert(..., "Should remember klasses in this context")
jmasa
parents:
1145
diff
changeset
|
3698 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) |
0 | 3699 bool result = false; |
1284 | 3700 if (CMSConcurrentMTEnabled && ConcGCThreads > 0) { |
0 | 3701 result = do_marking_mt(asynch); |
3702 } else { | |
3703 result = do_marking_st(asynch); | |
3704 } | |
3705 return result; | |
3706 } | |
3707 | |
3708 // Forward decl | |
3709 class CMSConcMarkingTask; | |
3710 | |
3711 class CMSConcMarkingTerminator: public ParallelTaskTerminator { | |
3712 CMSCollector* _collector; | |
3713 CMSConcMarkingTask* _task; | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3714 public: |
0 | 3715 virtual void yield(); |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3716 |
0 | 3717 // "n_threads" is the number of threads to be terminated. |
3718 // "queue_set" is a set of work queues of other threads. | |
3719 // "collector" is the CMS collector associated with this task terminator. | |
3720 // "yield" indicates whether we need the gang as a whole to yield. | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3721 CMSConcMarkingTerminator(int n_threads, TaskQueueSetSuper* queue_set, CMSCollector* collector) : |
0 | 3722 ParallelTaskTerminator(n_threads, queue_set), |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3723 _collector(collector) { } |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3724 |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3725 void set_task(CMSConcMarkingTask* task) { |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3726 _task = task; |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3727 } |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3728 }; |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3729 |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3730 class CMSConcMarkingTerminatorTerminator: public TerminatorTerminator { |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3731 CMSConcMarkingTask* _task; |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3732 public: |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3733 bool should_exit_termination(); |
0 | 3734 void set_task(CMSConcMarkingTask* task) { |
3735 _task = task; | |
3736 } | |
3737 }; | |
3738 | |
3739 // MT Concurrent Marking Task | |
3740 class CMSConcMarkingTask: public YieldingFlexibleGangTask { | |
3741 CMSCollector* _collector; | |
3742 int _n_workers; // requested/desired # workers | |
3743 bool _asynch; | |
3744 bool _result; | |
3745 CompactibleFreeListSpace* _cms_space; | |
3746 CompactibleFreeListSpace* _perm_space; | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3747 char _pad_front[64]; // padding to ... |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3748 HeapWord* _global_finger; // ... avoid sharing cache line |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3749 char _pad_back[64]; |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3750 HeapWord* _restart_addr; |
0 | 3751 |
3752 // Exposed here for yielding support | |
3753 Mutex* const _bit_map_lock; | |
3754 | |
3755 // The per thread work queues, available here for stealing | |
3756 OopTaskQueueSet* _task_queues; | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3757 |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3758 // Termination (and yielding) support |
0 | 3759 CMSConcMarkingTerminator _term; |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3760 CMSConcMarkingTerminatorTerminator _term_term; |
0 | 3761 |
3762 public: | |
3763 CMSConcMarkingTask(CMSCollector* collector, | |
3764 CompactibleFreeListSpace* cms_space, | |
3765 CompactibleFreeListSpace* perm_space, | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
3766 bool asynch, |
0 | 3767 YieldingFlexibleWorkGang* workers, |
3768 OopTaskQueueSet* task_queues): | |
3769 YieldingFlexibleGangTask("Concurrent marking done multi-threaded"), | |
3770 _collector(collector), | |
3771 _cms_space(cms_space), | |
3772 _perm_space(perm_space), | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
3773 _asynch(asynch), _n_workers(0), _result(true), |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
3774 _task_queues(task_queues), |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3775 _term(_n_workers, task_queues, _collector), |
0 | 3776 _bit_map_lock(collector->bitMapLock()) |
3777 { | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
3778 _requested_size = _n_workers; |
0 | 3779 _term.set_task(this); |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3780 _term_term.set_task(this); |
0 | 3781 assert(_cms_space->bottom() < _perm_space->bottom(), |
3782 "Finger incorrectly initialized below"); | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3783 _restart_addr = _global_finger = _cms_space->bottom(); |
0 | 3784 } |
3785 | |
3786 | |
3787 OopTaskQueueSet* task_queues() { return _task_queues; } | |
3788 | |
3789 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } | |
3790 | |
3791 HeapWord** global_finger_addr() { return &_global_finger; } | |
3792 | |
3793 CMSConcMarkingTerminator* terminator() { return &_term; } | |
3794 | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
3795 virtual void set_for_termination(int active_workers) { |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
3796 terminator()->reset_for_reuse(active_workers); |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
3797 } |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
3798 |
0 | 3799 void work(int i); |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3800 bool should_yield() { |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3801 return ConcurrentMarkSweepThread::should_yield() |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3802 && !_collector->foregroundGCIsActive() |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3803 && _asynch; |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3804 } |
0 | 3805 |
3806 virtual void coordinator_yield(); // stuff done by coordinator | |
3807 bool result() { return _result; } | |
3808 | |
3809 void reset(HeapWord* ra) { | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3810 assert(_global_finger >= _cms_space->end(), "Postcondition of ::work(i)"); |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3811 assert(_global_finger >= _perm_space->end(), "Postcondition of ::work(i)"); |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3812 assert(ra < _perm_space->end(), "ra too large"); |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3813 _restart_addr = _global_finger = ra; |
0 | 3814 _term.reset_for_reuse(); |
3815 } | |
3816 | |
3817 static bool get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, | |
3818 OopTaskQueue* work_q); | |
3819 | |
3820 private: | |
3821 void do_scan_and_mark(int i, CompactibleFreeListSpace* sp); | |
3822 void do_work_steal(int i); | |
3823 void bump_global_finger(HeapWord* f); | |
3824 }; | |
3825 | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3826 bool CMSConcMarkingTerminatorTerminator::should_exit_termination() { |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3827 assert(_task != NULL, "Error"); |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3828 return _task->yielding(); |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3829 // Note that we do not need the disjunct || _task->should_yield() above |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3830 // because we want terminating threads to yield only if the task |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3831 // is already in the midst of yielding, which happens only after at least one |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3832 // thread has yielded. |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3833 } |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3834 |
0 | 3835 void CMSConcMarkingTerminator::yield() { |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
3836 if (_task->should_yield()) { |
0 | 3837 _task->yield(); |
3838 } else { | |
3839 ParallelTaskTerminator::yield(); | |
3840 } | |
3841 } | |
3842 | |
3843 //////////////////////////////////////////////////////////////// | |
3844 // Concurrent Marking Algorithm Sketch | |
3845 //////////////////////////////////////////////////////////////// | |
3846 // Until all tasks exhausted (both spaces): | |
3847 // -- claim next available chunk | |
3848 // -- bump global finger via CAS | |
3849 // -- find first object that starts in this chunk | |
3850 // and start scanning bitmap from that position | |
3851 // -- scan marked objects for oops | |
3852 // -- CAS-mark target, and if successful: | |
3853 // . if target oop is above global finger (volatile read) | |
3854 // nothing to do | |
3855 // . if target oop is in chunk and above local finger | |
3856 // then nothing to do | |
3857 // . else push on work-queue | |
3858 // -- Deal with possible overflow issues: | |
3859 // . local work-queue overflow causes stuff to be pushed on | |
3860 // global (common) overflow queue | |
3861 // . always first empty local work queue | |
3862 // . then get a batch of oops from global work queue if any | |
3863 // . then do work stealing | |
3864 // -- When all tasks claimed (both spaces) | |
3865 // and local work queue empty, | |
3866 // then in a loop do: | |
3867 // . check global overflow stack; steal a batch of oops and trace | |
3868 // . try to steal from other threads oif GOS is empty | |
3869 // . if neither is available, offer termination | |
3870 // -- Terminate and return result | |
3871 // | |
3872 void CMSConcMarkingTask::work(int i) { | |
3873 elapsedTimer _timer; | |
3874 ResourceMark rm; | |
3875 HandleMark hm; | |
3876 | |
3877 DEBUG_ONLY(_collector->verify_overflow_empty();) | |
3878 | |
3879 // Before we begin work, our work queue should be empty | |
3880 assert(work_queue(i)->size() == 0, "Expected to be empty"); | |
3881 // Scan the bitmap covering _cms_space, tracing through grey objects. | |
3882 _timer.start(); | |
3883 do_scan_and_mark(i, _cms_space); | |
3884 _timer.stop(); | |
3885 if (PrintCMSStatistics != 0) { | |
3886 gclog_or_tty->print_cr("Finished cms space scanning in %dth thread: %3.3f sec", | |
3887 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers | |
3888 } | |
3889 | |
3890 // ... do the same for the _perm_space | |
3891 _timer.reset(); | |
3892 _timer.start(); | |
3893 do_scan_and_mark(i, _perm_space); | |
3894 _timer.stop(); | |
3895 if (PrintCMSStatistics != 0) { | |
3896 gclog_or_tty->print_cr("Finished perm space scanning in %dth thread: %3.3f sec", | |
3897 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers | |
3898 } | |
3899 | |
3900 // ... do work stealing | |
3901 _timer.reset(); | |
3902 _timer.start(); | |
3903 do_work_steal(i); | |
3904 _timer.stop(); | |
3905 if (PrintCMSStatistics != 0) { | |
3906 gclog_or_tty->print_cr("Finished work stealing in %dth thread: %3.3f sec", | |
3907 i, _timer.seconds()); // XXX: need xxx/xxx type of notation, two timers | |
3908 } | |
3909 assert(_collector->_markStack.isEmpty(), "Should have been emptied"); | |
3910 assert(work_queue(i)->size() == 0, "Should have been emptied"); | |
3911 // Note that under the current task protocol, the | |
3912 // following assertion is true even of the spaces | |
3913 // expanded since the completion of the concurrent | |
3914 // marking. XXX This will likely change under a strict | |
3915 // ABORT semantics. | |
3916 assert(_global_finger > _cms_space->end() && | |
3917 _global_finger >= _perm_space->end(), | |
3918 "All tasks have been completed"); | |
3919 DEBUG_ONLY(_collector->verify_overflow_empty();) | |
3920 } | |
3921 | |
3922 void CMSConcMarkingTask::bump_global_finger(HeapWord* f) { | |
3923 HeapWord* read = _global_finger; | |
3924 HeapWord* cur = read; | |
3925 while (f > read) { | |
3926 cur = read; | |
3927 read = (HeapWord*) Atomic::cmpxchg_ptr(f, &_global_finger, cur); | |
3928 if (cur == read) { | |
3929 // our cas succeeded | |
3930 assert(_global_finger >= f, "protocol consistency"); | |
3931 break; | |
3932 } | |
3933 } | |
3934 } | |
3935 | |
3936 // This is really inefficient, and should be redone by | |
3937 // using (not yet available) block-read and -write interfaces to the | |
3938 // stack and the work_queue. XXX FIX ME !!! | |
3939 bool CMSConcMarkingTask::get_work_from_overflow_stack(CMSMarkStack* ovflw_stk, | |
3940 OopTaskQueue* work_q) { | |
3941 // Fast lock-free check | |
3942 if (ovflw_stk->length() == 0) { | |
3943 return false; | |
3944 } | |
3945 assert(work_q->size() == 0, "Shouldn't steal"); | |
3946 MutexLockerEx ml(ovflw_stk->par_lock(), | |
3947 Mutex::_no_safepoint_check_flag); | |
3948 // Grab up to 1/4 the size of the work queue | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
628
diff
changeset
|
3949 size_t num = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, |
0 | 3950 (size_t)ParGCDesiredObjsFromOverflowList); |
3951 num = MIN2(num, ovflw_stk->length()); | |
3952 for (int i = (int) num; i > 0; i--) { | |
3953 oop cur = ovflw_stk->pop(); | |
3954 assert(cur != NULL, "Counted wrong?"); | |
3955 work_q->push(cur); | |
3956 } | |
3957 return num > 0; | |
3958 } | |
3959 | |
3960 void CMSConcMarkingTask::do_scan_and_mark(int i, CompactibleFreeListSpace* sp) { | |
3961 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); | |
3962 int n_tasks = pst->n_tasks(); | |
3963 // We allow that there may be no tasks to do here because | |
3964 // we are restarting after a stack overflow. | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3965 assert(pst->valid() || n_tasks == 0, "Uninitialized use?"); |
0 | 3966 int nth_task = 0; |
3967 | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3968 HeapWord* aligned_start = sp->bottom(); |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3969 if (sp->used_region().contains(_restart_addr)) { |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3970 // Align down to a card boundary for the start of 0th task |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3971 // for this space. |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3972 aligned_start = |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3973 (HeapWord*)align_size_down((uintptr_t)_restart_addr, |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3974 CardTableModRefBS::card_size); |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3975 } |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3976 |
0 | 3977 size_t chunk_size = sp->marking_task_size(); |
3978 while (!pst->is_task_claimed(/* reference */ nth_task)) { | |
3979 // Having claimed the nth task in this space, | |
3980 // compute the chunk that it corresponds to: | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3981 MemRegion span = MemRegion(aligned_start + nth_task*chunk_size, |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3982 aligned_start + (nth_task+1)*chunk_size); |
0 | 3983 // Try and bump the global finger via a CAS; |
3984 // note that we need to do the global finger bump | |
3985 // _before_ taking the intersection below, because | |
3986 // the task corresponding to that region will be | |
3987 // deemed done even if the used_region() expands | |
3988 // because of allocation -- as it almost certainly will | |
3989 // during start-up while the threads yield in the | |
3990 // closure below. | |
3991 HeapWord* finger = span.end(); | |
3992 bump_global_finger(finger); // atomically | |
3993 // There are null tasks here corresponding to chunks | |
3994 // beyond the "top" address of the space. | |
3995 span = span.intersection(sp->used_region()); | |
3996 if (!span.is_empty()) { // Non-null task | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3997 HeapWord* prev_obj; |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3998 assert(!span.contains(_restart_addr) || nth_task == 0, |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
3999 "Inconsistency"); |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4000 if (nth_task == 0) { |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4001 // For the 0th task, we'll not need to compute a block_start. |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4002 if (span.contains(_restart_addr)) { |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4003 // In the case of a restart because of stack overflow, |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4004 // we might additionally skip a chunk prefix. |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4005 prev_obj = _restart_addr; |
0 | 4006 } else { |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4007 prev_obj = span.start(); |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4008 } |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4009 } else { |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4010 // We want to skip the first object because |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4011 // the protocol is to scan any object in its entirety |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4012 // that _starts_ in this span; a fortiori, any |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4013 // object starting in an earlier span is scanned |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4014 // as part of an earlier claimed task. |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4015 // Below we use the "careful" version of block_start |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4016 // so we do not try to navigate uninitialized objects. |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4017 prev_obj = sp->block_start_careful(span.start()); |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4018 // Below we use a variant of block_size that uses the |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4019 // Printezis bits to avoid waiting for allocated |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4020 // objects to become initialized/parsable. |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4021 while (prev_obj < span.start()) { |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4022 size_t sz = sp->block_size_no_stall(prev_obj, _collector); |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4023 if (sz > 0) { |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4024 prev_obj += sz; |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4025 } else { |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4026 // In this case we may end up doing a bit of redundant |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4027 // scanning, but that appears unavoidable, short of |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4028 // locking the free list locks; see bug 6324141. |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4029 break; |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4030 } |
0 | 4031 } |
4032 } | |
4033 if (prev_obj < span.end()) { | |
4034 MemRegion my_span = MemRegion(prev_obj, span.end()); | |
4035 // Do the marking work within a non-empty span -- | |
4036 // the last argument to the constructor indicates whether the | |
4037 // iteration should be incremental with periodic yields. | |
4038 Par_MarkFromRootsClosure cl(this, _collector, my_span, | |
4039 &_collector->_markBitMap, | |
4040 work_queue(i), | |
4041 &_collector->_markStack, | |
4042 &_collector->_revisitStack, | |
4043 _asynch); | |
4044 _collector->_markBitMap.iterate(&cl, my_span.start(), my_span.end()); | |
4045 } // else nothing to do for this task | |
4046 } // else nothing to do for this task | |
4047 } | |
4048 // We'd be tempted to assert here that since there are no | |
4049 // more tasks left to claim in this space, the global_finger | |
4050 // must exceed space->top() and a fortiori space->end(). However, | |
4051 // that would not quite be correct because the bumping of | |
4052 // global_finger occurs strictly after the claiming of a task, | |
4053 // so by the time we reach here the global finger may not yet | |
4054 // have been bumped up by the thread that claimed the last | |
4055 // task. | |
4056 pst->all_tasks_completed(); | |
4057 } | |
4058 | |
935 | 4059 class Par_ConcMarkingClosure: public Par_KlassRememberingOopClosure { |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4060 private: |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4061 CMSConcMarkingTask* _task; |
0 | 4062 MemRegion _span; |
4063 CMSBitMap* _bit_map; | |
4064 CMSMarkStack* _overflow_stack; | |
4065 OopTaskQueue* _work_queue; | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4066 protected: |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4067 DO_OOP_WORK_DEFN |
0 | 4068 public: |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4069 Par_ConcMarkingClosure(CMSCollector* collector, CMSConcMarkingTask* task, OopTaskQueue* work_queue, |
935 | 4070 CMSBitMap* bit_map, CMSMarkStack* overflow_stack, |
4071 CMSMarkStack* revisit_stack): | |
4072 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4073 _task(task), |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4074 _span(collector->_span), |
0 | 4075 _work_queue(work_queue), |
4076 _bit_map(bit_map), | |
935 | 4077 _overflow_stack(overflow_stack) |
4078 { } | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4079 virtual void do_oop(oop* p); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4080 virtual void do_oop(narrowOop* p); |
0 | 4081 void trim_queue(size_t max); |
4082 void handle_stack_overflow(HeapWord* lost); | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4083 void do_yield_check() { |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4084 if (_task->should_yield()) { |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4085 _task->yield(); |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4086 } |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4087 } |
0 | 4088 }; |
4089 | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4090 // Grey object scanning during work stealing phase -- |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4091 // the salient assumption here is that any references |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4092 // that are in these stolen objects being scanned must |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4093 // already have been initialized (else they would not have |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4094 // been published), so we do not need to check for |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4095 // uninitialized objects before pushing here. |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4096 void Par_ConcMarkingClosure::do_oop(oop obj) { |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4097 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4098 HeapWord* addr = (HeapWord*)obj; |
0 | 4099 // Check if oop points into the CMS generation |
4100 // and is not marked | |
4101 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { | |
4102 // a white object ... | |
4103 // If we manage to "claim" the object, by being the | |
4104 // first thread to mark it, then we push it on our | |
4105 // marking stack | |
4106 if (_bit_map->par_mark(addr)) { // ... now grey | |
4107 // push on work queue (grey set) | |
4108 bool simulate_overflow = false; | |
4109 NOT_PRODUCT( | |
4110 if (CMSMarkStackOverflowALot && | |
4111 _collector->simulate_overflow()) { | |
4112 // simulate a stack overflow | |
4113 simulate_overflow = true; | |
4114 } | |
4115 ) | |
4116 if (simulate_overflow || | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4117 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { |
0 | 4118 // stack overflow |
4119 if (PrintCMSStatistics != 0) { | |
4120 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " | |
4121 SIZE_FORMAT, _overflow_stack->capacity()); | |
4122 } | |
4123 // We cannot assert that the overflow stack is full because | |
4124 // it may have been emptied since. | |
4125 assert(simulate_overflow || | |
4126 _work_queue->size() == _work_queue->max_elems(), | |
4127 "Else push should have succeeded"); | |
4128 handle_stack_overflow(addr); | |
4129 } | |
4130 } // Else, some other thread got there first | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4131 do_yield_check(); |
0 | 4132 } |
4133 } | |
4134 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4135 void Par_ConcMarkingClosure::do_oop(oop* p) { Par_ConcMarkingClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4136 void Par_ConcMarkingClosure::do_oop(narrowOop* p) { Par_ConcMarkingClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4137 |
0 | 4138 void Par_ConcMarkingClosure::trim_queue(size_t max) { |
4139 while (_work_queue->size() > max) { | |
4140 oop new_oop; | |
4141 if (_work_queue->pop_local(new_oop)) { | |
4142 assert(new_oop->is_oop(), "Should be an oop"); | |
4143 assert(_bit_map->isMarked((HeapWord*)new_oop), "Grey object"); | |
4144 assert(_span.contains((HeapWord*)new_oop), "Not in span"); | |
4145 assert(new_oop->is_parsable(), "Should be parsable"); | |
4146 new_oop->oop_iterate(this); // do_oop() above | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4147 do_yield_check(); |
0 | 4148 } |
4149 } | |
4150 } | |
4151 | |
4152 // Upon stack overflow, we discard (part of) the stack, | |
4153 // remembering the least address amongst those discarded | |
4154 // in CMSCollector's _restart_address. | |
4155 void Par_ConcMarkingClosure::handle_stack_overflow(HeapWord* lost) { | |
4156 // We need to do this under a mutex to prevent other | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
4157 // workers from interfering with the work done below. |
0 | 4158 MutexLockerEx ml(_overflow_stack->par_lock(), |
4159 Mutex::_no_safepoint_check_flag); | |
4160 // Remember the least grey address discarded | |
4161 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); | |
4162 _collector->lower_restart_addr(ra); | |
4163 _overflow_stack->reset(); // discard stack contents | |
4164 _overflow_stack->expand(); // expand the stack if possible | |
4165 } | |
4166 | |
4167 | |
4168 void CMSConcMarkingTask::do_work_steal(int i) { | |
4169 OopTaskQueue* work_q = work_queue(i); | |
4170 oop obj_to_scan; | |
4171 CMSBitMap* bm = &(_collector->_markBitMap); | |
4172 CMSMarkStack* ovflw = &(_collector->_markStack); | |
935 | 4173 CMSMarkStack* revisit = &(_collector->_revisitStack); |
0 | 4174 int* seed = _collector->hash_seed(i); |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4175 Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw, revisit); |
0 | 4176 while (true) { |
4177 cl.trim_queue(0); | |
4178 assert(work_q->size() == 0, "Should have been emptied above"); | |
4179 if (get_work_from_overflow_stack(ovflw, work_q)) { | |
4180 // Can't assert below because the work obtained from the | |
4181 // overflow stack may already have been stolen from us. | |
4182 // assert(work_q->size() > 0, "Work from overflow stack"); | |
4183 continue; | |
4184 } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { | |
4185 assert(obj_to_scan->is_oop(), "Should be an oop"); | |
4186 assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object"); | |
4187 obj_to_scan->oop_iterate(&cl); | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4188 } else if (terminator()->offer_termination(&_term_term)) { |
0 | 4189 assert(work_q->size() == 0, "Impossible!"); |
4190 break; | |
1837
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4191 } else if (yielding() || should_yield()) { |
c99c53f07c14
6692906: CMS: parallel concurrent marking may be prone to hanging or stalling mutators for periods of time
ysr
parents:
1836
diff
changeset
|
4192 yield(); |
0 | 4193 } |
4194 } | |
4195 } | |
4196 | |
4197 // This is run by the CMS (coordinator) thread. | |
4198 void CMSConcMarkingTask::coordinator_yield() { | |
4199 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
4200 "CMS thread should hold CMS token"); | |
935 | 4201 DEBUG_ONLY(RememberKlassesChecker mux(false);) |
0 | 4202 // First give up the locks, then yield, then re-lock |
4203 // We should probably use a constructor/destructor idiom to | |
4204 // do this unlock/lock or modify the MutexUnlocker class to | |
4205 // serve our purpose. XXX | |
4206 assert_lock_strong(_bit_map_lock); | |
4207 _bit_map_lock->unlock(); | |
4208 ConcurrentMarkSweepThread::desynchronize(true); | |
4209 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
4210 _collector->stopTimer(); | |
4211 if (PrintCMSStatistics != 0) { | |
4212 _collector->incrementYields(); | |
4213 } | |
4214 _collector->icms_wait(); | |
4215 | |
4216 // It is possible for whichever thread initiated the yield request | |
4217 // not to get a chance to wake up and take the bitmap lock between | |
4218 // this thread releasing it and reacquiring it. So, while the | |
4219 // should_yield() flag is on, let's sleep for a bit to give the | |
4220 // other thread a chance to wake up. The limit imposed on the number | |
4221 // of iterations is defensive, to avoid any unforseen circumstances | |
4222 // putting us into an infinite loop. Since it's always been this | |
4223 // (coordinator_yield()) method that was observed to cause the | |
4224 // problem, we are using a parameter (CMSCoordinatorYieldSleepCount) | |
4225 // which is by default non-zero. For the other seven methods that | |
4226 // also perform the yield operation, as are using a different | |
4227 // parameter (CMSYieldSleepCount) which is by default zero. This way we | |
4228 // can enable the sleeping for those methods too, if necessary. | |
4229 // See 6442774. | |
4230 // | |
4231 // We really need to reconsider the synchronization between the GC | |
4232 // thread and the yield-requesting threads in the future and we | |
4233 // should really use wait/notify, which is the recommended | |
4234 // way of doing this type of interaction. Additionally, we should | |
4235 // consolidate the eight methods that do the yield operation and they | |
4236 // are almost identical into one for better maintenability and | |
4237 // readability. See 6445193. | |
4238 // | |
4239 // Tony 2006.06.29 | |
4240 for (unsigned i = 0; i < CMSCoordinatorYieldSleepCount && | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4241 ConcurrentMarkSweepThread::should_yield() && |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
4242 !CMSCollector::foregroundGCIsActive(); ++i) { |
0 | 4243 os::sleep(Thread::current(), 1, false); |
4244 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
4245 } | |
4246 | |
4247 ConcurrentMarkSweepThread::synchronize(true); | |
4248 _bit_map_lock->lock_without_safepoint_check(); | |
4249 _collector->startTimer(); | |
4250 } | |
4251 | |
4252 bool CMSCollector::do_marking_mt(bool asynch) { | |
1284 | 4253 assert(ConcGCThreads > 0 && conc_workers() != NULL, "precondition"); |
0 | 4254 // In the future this would be determined ergonomically, based |
4255 // on #cpu's, # active mutator threads (and load), and mutation rate. | |
1284 | 4256 int num_workers = ConcGCThreads; |
0 | 4257 |
4258 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); | |
4259 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); | |
4260 | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4261 CMSConcMarkingTask tsk(this, |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4262 cms_space, |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4263 perm_space, |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4264 asynch, |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4265 conc_workers(), |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4266 task_queues()); |
0 | 4267 |
4268 // Since the actual number of workers we get may be different | |
4269 // from the number we requested above, do we need to do anything different | |
4270 // below? In particular, may be we need to subclass the SequantialSubTasksDone | |
4271 // class?? XXX | |
4272 cms_space ->initialize_sequential_subtasks_for_marking(num_workers); | |
4273 perm_space->initialize_sequential_subtasks_for_marking(num_workers); | |
4274 | |
4275 // Refs discovery is already non-atomic. | |
4276 assert(!ref_processor()->discovery_is_atomic(), "Should be non-atomic"); | |
4277 // Mutate the Refs discovery so it is MT during the | |
4278 // multi-threaded marking phase. | |
4279 ReferenceProcessorMTMutator mt(ref_processor(), num_workers > 1); | |
1190
4788266644c1
6895236: CMS: cmsOopClosures.inline.hpp:43 assert(..., "Should remember klasses in this context")
jmasa
parents:
1145
diff
changeset
|
4280 DEBUG_ONLY(RememberKlassesChecker cmx(should_unload_classes());) |
0 | 4281 conc_workers()->start_task(&tsk); |
4282 while (tsk.yielded()) { | |
4283 tsk.coordinator_yield(); | |
4284 conc_workers()->continue_task(&tsk); | |
4285 } | |
4286 // If the task was aborted, _restart_addr will be non-NULL | |
4287 assert(tsk.completed() || _restart_addr != NULL, "Inconsistency"); | |
4288 while (_restart_addr != NULL) { | |
4289 // XXX For now we do not make use of ABORTED state and have not | |
4290 // yet implemented the right abort semantics (even in the original | |
4291 // single-threaded CMS case). That needs some more investigation | |
4292 // and is deferred for now; see CR# TBF. 07252005YSR. XXX | |
4293 assert(!CMSAbortSemantics || tsk.aborted(), "Inconsistency"); | |
4294 // If _restart_addr is non-NULL, a marking stack overflow | |
605 | 4295 // occurred; we need to do a fresh marking iteration from the |
0 | 4296 // indicated restart address. |
4297 if (_foregroundGCIsActive && asynch) { | |
4298 // We may be running into repeated stack overflows, having | |
4299 // reached the limit of the stack size, while making very | |
4300 // slow forward progress. It may be best to bail out and | |
4301 // let the foreground collector do its job. | |
4302 // Clear _restart_addr, so that foreground GC | |
4303 // works from scratch. This avoids the headache of | |
4304 // a "rescan" which would otherwise be needed because | |
4305 // of the dirty mod union table & card table. | |
4306 _restart_addr = NULL; | |
4307 return false; | |
4308 } | |
4309 // Adjust the task to restart from _restart_addr | |
4310 tsk.reset(_restart_addr); | |
4311 cms_space ->initialize_sequential_subtasks_for_marking(num_workers, | |
4312 _restart_addr); | |
4313 perm_space->initialize_sequential_subtasks_for_marking(num_workers, | |
4314 _restart_addr); | |
4315 _restart_addr = NULL; | |
4316 // Get the workers going again | |
4317 conc_workers()->start_task(&tsk); | |
4318 while (tsk.yielded()) { | |
4319 tsk.coordinator_yield(); | |
4320 conc_workers()->continue_task(&tsk); | |
4321 } | |
4322 } | |
4323 assert(tsk.completed(), "Inconsistency"); | |
4324 assert(tsk.result() == true, "Inconsistency"); | |
4325 return true; | |
4326 } | |
4327 | |
4328 bool CMSCollector::do_marking_st(bool asynch) { | |
4329 ResourceMark rm; | |
4330 HandleMark hm; | |
4331 | |
4332 MarkFromRootsClosure markFromRootsClosure(this, _span, &_markBitMap, | |
4333 &_markStack, &_revisitStack, CMSYield && asynch); | |
4334 // the last argument to iterate indicates whether the iteration | |
4335 // should be incremental with periodic yields. | |
4336 _markBitMap.iterate(&markFromRootsClosure); | |
4337 // If _restart_addr is non-NULL, a marking stack overflow | |
605 | 4338 // occurred; we need to do a fresh iteration from the |
0 | 4339 // indicated restart address. |
4340 while (_restart_addr != NULL) { | |
4341 if (_foregroundGCIsActive && asynch) { | |
4342 // We may be running into repeated stack overflows, having | |
4343 // reached the limit of the stack size, while making very | |
4344 // slow forward progress. It may be best to bail out and | |
4345 // let the foreground collector do its job. | |
4346 // Clear _restart_addr, so that foreground GC | |
4347 // works from scratch. This avoids the headache of | |
4348 // a "rescan" which would otherwise be needed because | |
4349 // of the dirty mod union table & card table. | |
4350 _restart_addr = NULL; | |
4351 return false; // indicating failure to complete marking | |
4352 } | |
4353 // Deal with stack overflow: | |
4354 // we restart marking from _restart_addr | |
4355 HeapWord* ra = _restart_addr; | |
4356 markFromRootsClosure.reset(ra); | |
4357 _restart_addr = NULL; | |
4358 _markBitMap.iterate(&markFromRootsClosure, ra, _span.end()); | |
4359 } | |
4360 return true; | |
4361 } | |
4362 | |
4363 void CMSCollector::preclean() { | |
4364 check_correct_thread_executing(); | |
4365 assert(Thread::current()->is_ConcurrentGC_thread(), "Wrong thread"); | |
4366 verify_work_stacks_empty(); | |
4367 verify_overflow_empty(); | |
4368 _abort_preclean = false; | |
4369 if (CMSPrecleaningEnabled) { | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4370 // Precleaning is currently not MT but the reference processor |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4371 // may be set for MT. Disable it temporarily here. |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4372 ReferenceProcessor* rp = ref_processor(); |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4373 ReferenceProcessorMTProcMutator z(rp, false); |
0 | 4374 _eden_chunk_index = 0; |
4375 size_t used = get_eden_used(); | |
4376 size_t capacity = get_eden_capacity(); | |
4377 // Don't start sampling unless we will get sufficiently | |
4378 // many samples. | |
4379 if (used < (capacity/(CMSScheduleRemarkSamplingRatio * 100) | |
4380 * CMSScheduleRemarkEdenPenetration)) { | |
4381 _start_sampling = true; | |
4382 } else { | |
4383 _start_sampling = false; | |
4384 } | |
4385 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); | |
4386 CMSPhaseAccounting pa(this, "preclean", !PrintGCDetails); | |
4387 preclean_work(CMSPrecleanRefLists1, CMSPrecleanSurvivors1); | |
4388 } | |
4389 CMSTokenSync x(true); // is cms thread | |
4390 if (CMSPrecleaningEnabled) { | |
4391 sample_eden(); | |
4392 _collectorState = AbortablePreclean; | |
4393 } else { | |
4394 _collectorState = FinalMarking; | |
4395 } | |
4396 verify_work_stacks_empty(); | |
4397 verify_overflow_empty(); | |
4398 } | |
4399 | |
4400 // Try and schedule the remark such that young gen | |
4401 // occupancy is CMSScheduleRemarkEdenPenetration %. | |
4402 void CMSCollector::abortable_preclean() { | |
4403 check_correct_thread_executing(); | |
4404 assert(CMSPrecleaningEnabled, "Inconsistent control state"); | |
4405 assert(_collectorState == AbortablePreclean, "Inconsistent control state"); | |
4406 | |
4407 // If Eden's current occupancy is below this threshold, | |
4408 // immediately schedule the remark; else preclean | |
4409 // past the next scavenge in an effort to | |
4410 // schedule the pause as described avove. By choosing | |
4411 // CMSScheduleRemarkEdenSizeThreshold >= max eden size | |
4412 // we will never do an actual abortable preclean cycle. | |
4413 if (get_eden_used() > CMSScheduleRemarkEdenSizeThreshold) { | |
4414 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); | |
4415 CMSPhaseAccounting pa(this, "abortable-preclean", !PrintGCDetails); | |
4416 // We need more smarts in the abortable preclean | |
4417 // loop below to deal with cases where allocation | |
4418 // in young gen is very very slow, and our precleaning | |
4419 // is running a losing race against a horde of | |
4420 // mutators intent on flooding us with CMS updates | |
4421 // (dirty cards). | |
4422 // One, admittedly dumb, strategy is to give up | |
4423 // after a certain number of abortable precleaning loops | |
4424 // or after a certain maximum time. We want to make | |
4425 // this smarter in the next iteration. | |
4426 // XXX FIX ME!!! YSR | |
4427 size_t loops = 0, workdone = 0, cumworkdone = 0, waited = 0; | |
4428 while (!(should_abort_preclean() || | |
4429 ConcurrentMarkSweepThread::should_terminate())) { | |
4430 workdone = preclean_work(CMSPrecleanRefLists2, CMSPrecleanSurvivors2); | |
4431 cumworkdone += workdone; | |
4432 loops++; | |
4433 // Voluntarily terminate abortable preclean phase if we have | |
4434 // been at it for too long. | |
4435 if ((CMSMaxAbortablePrecleanLoops != 0) && | |
4436 loops >= CMSMaxAbortablePrecleanLoops) { | |
4437 if (PrintGCDetails) { | |
4438 gclog_or_tty->print(" CMS: abort preclean due to loops "); | |
4439 } | |
4440 break; | |
4441 } | |
4442 if (pa.wallclock_millis() > CMSMaxAbortablePrecleanTime) { | |
4443 if (PrintGCDetails) { | |
4444 gclog_or_tty->print(" CMS: abort preclean due to time "); | |
4445 } | |
4446 break; | |
4447 } | |
4448 // If we are doing little work each iteration, we should | |
4449 // take a short break. | |
4450 if (workdone < CMSAbortablePrecleanMinWorkPerIteration) { | |
4451 // Sleep for some time, waiting for work to accumulate | |
4452 stopTimer(); | |
4453 cmsThread()->wait_on_cms_lock(CMSAbortablePrecleanWaitMillis); | |
4454 startTimer(); | |
4455 waited++; | |
4456 } | |
4457 } | |
4458 if (PrintCMSStatistics > 0) { | |
4459 gclog_or_tty->print(" [%d iterations, %d waits, %d cards)] ", | |
4460 loops, waited, cumworkdone); | |
4461 } | |
4462 } | |
4463 CMSTokenSync x(true); // is cms thread | |
4464 if (_collectorState != Idling) { | |
4465 assert(_collectorState == AbortablePreclean, | |
4466 "Spontaneous state transition?"); | |
4467 _collectorState = FinalMarking; | |
4468 } // Else, a foreground collection completed this CMS cycle. | |
4469 return; | |
4470 } | |
4471 | |
4472 // Respond to an Eden sampling opportunity | |
4473 void CMSCollector::sample_eden() { | |
4474 // Make sure a young gc cannot sneak in between our | |
4475 // reading and recording of a sample. | |
4476 assert(Thread::current()->is_ConcurrentGC_thread(), | |
4477 "Only the cms thread may collect Eden samples"); | |
4478 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
4479 "Should collect samples while holding CMS token"); | |
4480 if (!_start_sampling) { | |
4481 return; | |
4482 } | |
4483 if (_eden_chunk_array) { | |
4484 if (_eden_chunk_index < _eden_chunk_capacity) { | |
4485 _eden_chunk_array[_eden_chunk_index] = *_top_addr; // take sample | |
4486 assert(_eden_chunk_array[_eden_chunk_index] <= *_end_addr, | |
4487 "Unexpected state of Eden"); | |
4488 // We'd like to check that what we just sampled is an oop-start address; | |
4489 // however, we cannot do that here since the object may not yet have been | |
4490 // initialized. So we'll instead do the check when we _use_ this sample | |
4491 // later. | |
4492 if (_eden_chunk_index == 0 || | |
4493 (pointer_delta(_eden_chunk_array[_eden_chunk_index], | |
4494 _eden_chunk_array[_eden_chunk_index-1]) | |
4495 >= CMSSamplingGrain)) { | |
4496 _eden_chunk_index++; // commit sample | |
4497 } | |
4498 } | |
4499 } | |
4500 if ((_collectorState == AbortablePreclean) && !_abort_preclean) { | |
4501 size_t used = get_eden_used(); | |
4502 size_t capacity = get_eden_capacity(); | |
4503 assert(used <= capacity, "Unexpected state of Eden"); | |
4504 if (used > (capacity/100 * CMSScheduleRemarkEdenPenetration)) { | |
4505 _abort_preclean = true; | |
4506 } | |
4507 } | |
4508 } | |
4509 | |
4510 | |
4511 size_t CMSCollector::preclean_work(bool clean_refs, bool clean_survivor) { | |
4512 assert(_collectorState == Precleaning || | |
4513 _collectorState == AbortablePreclean, "incorrect state"); | |
4514 ResourceMark rm; | |
4515 HandleMark hm; | |
4516 // Do one pass of scrubbing the discovered reference lists | |
4517 // to remove any reference objects with strongly-reachable | |
4518 // referents. | |
4519 if (clean_refs) { | |
4520 ReferenceProcessor* rp = ref_processor(); | |
4521 CMSPrecleanRefsYieldClosure yield_cl(this); | |
4522 assert(rp->span().equals(_span), "Spans should be equal"); | |
4523 CMSKeepAliveClosure keep_alive(this, _span, &_markBitMap, | |
935 | 4524 &_markStack, &_revisitStack, |
4525 true /* preclean */); | |
0 | 4526 CMSDrainMarkingStackClosure complete_trace(this, |
452
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
4527 _span, &_markBitMap, &_markStack, |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
4528 &keep_alive, true /* preclean */); |
0 | 4529 |
4530 // We don't want this step to interfere with a young | |
4531 // collection because we don't want to take CPU | |
4532 // or memory bandwidth away from the young GC threads | |
4533 // (which may be as many as there are CPUs). | |
4534 // Note that we don't need to protect ourselves from | |
4535 // interference with mutators because they can't | |
4536 // manipulate the discovered reference lists nor affect | |
4537 // the computed reachability of the referents, the | |
4538 // only properties manipulated by the precleaning | |
4539 // of these reference lists. | |
4540 stopTimer(); | |
4541 CMSTokenSyncWithLocks x(true /* is cms thread */, | |
4542 bitMapLock()); | |
4543 startTimer(); | |
4544 sample_eden(); | |
935 | 4545 |
0 | 4546 // The following will yield to allow foreground |
4547 // collection to proceed promptly. XXX YSR: | |
4548 // The code in this method may need further | |
4549 // tweaking for better performance and some restructuring | |
4550 // for cleaner interfaces. | |
4551 rp->preclean_discovered_references( | |
4552 rp->is_alive_non_header(), &keep_alive, &complete_trace, | |
1190
4788266644c1
6895236: CMS: cmsOopClosures.inline.hpp:43 assert(..., "Should remember klasses in this context")
jmasa
parents:
1145
diff
changeset
|
4553 &yield_cl, should_unload_classes()); |
0 | 4554 } |
4555 | |
4556 if (clean_survivor) { // preclean the active survivor space(s) | |
4557 assert(_young_gen->kind() == Generation::DefNew || | |
4558 _young_gen->kind() == Generation::ParNew || | |
4559 _young_gen->kind() == Generation::ASParNew, | |
4560 "incorrect type for cast"); | |
4561 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; | |
4562 PushAndMarkClosure pam_cl(this, _span, ref_processor(), | |
4563 &_markBitMap, &_modUnionTable, | |
4564 &_markStack, &_revisitStack, | |
4565 true /* precleaning phase */); | |
4566 stopTimer(); | |
4567 CMSTokenSyncWithLocks ts(true /* is cms thread */, | |
4568 bitMapLock()); | |
4569 startTimer(); | |
4570 unsigned int before_count = | |
4571 GenCollectedHeap::heap()->total_collections(); | |
4572 SurvivorSpacePrecleanClosure | |
4573 sss_cl(this, _span, &_markBitMap, &_markStack, | |
4574 &pam_cl, before_count, CMSYield); | |
1190
4788266644c1
6895236: CMS: cmsOopClosures.inline.hpp:43 assert(..., "Should remember klasses in this context")
jmasa
parents:
1145
diff
changeset
|
4575 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) |
0 | 4576 dng->from()->object_iterate_careful(&sss_cl); |
4577 dng->to()->object_iterate_careful(&sss_cl); | |
4578 } | |
4579 MarkRefsIntoAndScanClosure | |
4580 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, | |
4581 &_markStack, &_revisitStack, this, CMSYield, | |
4582 true /* precleaning phase */); | |
4583 // CAUTION: The following closure has persistent state that may need to | |
4584 // be reset upon a decrease in the sequence of addresses it | |
4585 // processes. | |
4586 ScanMarkedObjectsAgainCarefullyClosure | |
4587 smoac_cl(this, _span, | |
4588 &_markBitMap, &_markStack, &_revisitStack, &mrias_cl, CMSYield); | |
4589 | |
4590 // Preclean dirty cards in ModUnionTable and CardTable using | |
4591 // appropriate convergence criterion; | |
4592 // repeat CMSPrecleanIter times unless we find that | |
4593 // we are losing. | |
4594 assert(CMSPrecleanIter < 10, "CMSPrecleanIter is too large"); | |
4595 assert(CMSPrecleanNumerator < CMSPrecleanDenominator, | |
4596 "Bad convergence multiplier"); | |
4597 assert(CMSPrecleanThreshold >= 100, | |
4598 "Unreasonably low CMSPrecleanThreshold"); | |
4599 | |
4600 size_t numIter, cumNumCards, lastNumCards, curNumCards; | |
4601 for (numIter = 0, cumNumCards = lastNumCards = curNumCards = 0; | |
4602 numIter < CMSPrecleanIter; | |
4603 numIter++, lastNumCards = curNumCards, cumNumCards += curNumCards) { | |
4604 curNumCards = preclean_mod_union_table(_cmsGen, &smoac_cl); | |
4605 if (CMSPermGenPrecleaningEnabled) { | |
4606 curNumCards += preclean_mod_union_table(_permGen, &smoac_cl); | |
4607 } | |
4608 if (Verbose && PrintGCDetails) { | |
4609 gclog_or_tty->print(" (modUnionTable: %d cards)", curNumCards); | |
4610 } | |
4611 // Either there are very few dirty cards, so re-mark | |
4612 // pause will be small anyway, or our pre-cleaning isn't | |
4613 // that much faster than the rate at which cards are being | |
4614 // dirtied, so we might as well stop and re-mark since | |
4615 // precleaning won't improve our re-mark time by much. | |
4616 if (curNumCards <= CMSPrecleanThreshold || | |
4617 (numIter > 0 && | |
4618 (curNumCards * CMSPrecleanDenominator > | |
4619 lastNumCards * CMSPrecleanNumerator))) { | |
4620 numIter++; | |
4621 cumNumCards += curNumCards; | |
4622 break; | |
4623 } | |
4624 } | |
4625 curNumCards = preclean_card_table(_cmsGen, &smoac_cl); | |
4626 if (CMSPermGenPrecleaningEnabled) { | |
4627 curNumCards += preclean_card_table(_permGen, &smoac_cl); | |
4628 } | |
4629 cumNumCards += curNumCards; | |
4630 if (PrintGCDetails && PrintCMSStatistics != 0) { | |
4631 gclog_or_tty->print_cr(" (cardTable: %d cards, re-scanned %d cards, %d iterations)", | |
4632 curNumCards, cumNumCards, numIter); | |
4633 } | |
4634 return cumNumCards; // as a measure of useful work done | |
4635 } | |
4636 | |
4637 // PRECLEANING NOTES: | |
4638 // Precleaning involves: | |
4639 // . reading the bits of the modUnionTable and clearing the set bits. | |
4640 // . For the cards corresponding to the set bits, we scan the | |
4641 // objects on those cards. This means we need the free_list_lock | |
4642 // so that we can safely iterate over the CMS space when scanning | |
4643 // for oops. | |
4644 // . When we scan the objects, we'll be both reading and setting | |
4645 // marks in the marking bit map, so we'll need the marking bit map. | |
4646 // . For protecting _collector_state transitions, we take the CGC_lock. | |
4647 // Note that any races in the reading of of card table entries by the | |
4648 // CMS thread on the one hand and the clearing of those entries by the | |
4649 // VM thread or the setting of those entries by the mutator threads on the | |
4650 // other are quite benign. However, for efficiency it makes sense to keep | |
4651 // the VM thread from racing with the CMS thread while the latter is | |
4652 // dirty card info to the modUnionTable. We therefore also use the | |
4653 // CGC_lock to protect the reading of the card table and the mod union | |
4654 // table by the CM thread. | |
4655 // . We run concurrently with mutator updates, so scanning | |
4656 // needs to be done carefully -- we should not try to scan | |
4657 // potentially uninitialized objects. | |
4658 // | |
4659 // Locking strategy: While holding the CGC_lock, we scan over and | |
4660 // reset a maximal dirty range of the mod union / card tables, then lock | |
4661 // the free_list_lock and bitmap lock to do a full marking, then | |
4662 // release these locks; and repeat the cycle. This allows for a | |
4663 // certain amount of fairness in the sharing of these locks between | |
4664 // the CMS collector on the one hand, and the VM thread and the | |
4665 // mutators on the other. | |
4666 | |
4667 // NOTE: preclean_mod_union_table() and preclean_card_table() | |
4668 // further below are largely identical; if you need to modify | |
4669 // one of these methods, please check the other method too. | |
4670 | |
4671 size_t CMSCollector::preclean_mod_union_table( | |
4672 ConcurrentMarkSweepGeneration* gen, | |
4673 ScanMarkedObjectsAgainCarefullyClosure* cl) { | |
4674 verify_work_stacks_empty(); | |
4675 verify_overflow_empty(); | |
4676 | |
935 | 4677 // Turn off checking for this method but turn it back on |
4678 // selectively. There are yield points in this method | |
4679 // but it is difficult to turn the checking off just around | |
4680 // the yield points. It is simpler to selectively turn | |
4681 // it on. | |
4682 DEBUG_ONLY(RememberKlassesChecker mux(false);) | |
4683 | |
0 | 4684 // strategy: starting with the first card, accumulate contiguous |
4685 // ranges of dirty cards; clear these cards, then scan the region | |
4686 // covered by these cards. | |
4687 | |
4688 // Since all of the MUT is committed ahead, we can just use | |
4689 // that, in case the generations expand while we are precleaning. | |
4690 // It might also be fine to just use the committed part of the | |
4691 // generation, but we might potentially miss cards when the | |
4692 // generation is rapidly expanding while we are in the midst | |
4693 // of precleaning. | |
4694 HeapWord* startAddr = gen->reserved().start(); | |
4695 HeapWord* endAddr = gen->reserved().end(); | |
4696 | |
4697 cl->setFreelistLock(gen->freelistLock()); // needed for yielding | |
4698 | |
4699 size_t numDirtyCards, cumNumDirtyCards; | |
4700 HeapWord *nextAddr, *lastAddr; | |
4701 for (cumNumDirtyCards = numDirtyCards = 0, | |
4702 nextAddr = lastAddr = startAddr; | |
4703 nextAddr < endAddr; | |
4704 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { | |
4705 | |
4706 ResourceMark rm; | |
4707 HandleMark hm; | |
4708 | |
4709 MemRegion dirtyRegion; | |
4710 { | |
4711 stopTimer(); | |
935 | 4712 // Potential yield point |
0 | 4713 CMSTokenSync ts(true); |
4714 startTimer(); | |
4715 sample_eden(); | |
4716 // Get dirty region starting at nextOffset (inclusive), | |
4717 // simultaneously clearing it. | |
4718 dirtyRegion = | |
4719 _modUnionTable.getAndClearMarkedRegion(nextAddr, endAddr); | |
4720 assert(dirtyRegion.start() >= nextAddr, | |
4721 "returned region inconsistent?"); | |
4722 } | |
4723 // Remember where the next search should begin. | |
4724 // The returned region (if non-empty) is a right open interval, | |
4725 // so lastOffset is obtained from the right end of that | |
4726 // interval. | |
4727 lastAddr = dirtyRegion.end(); | |
4728 // Should do something more transparent and less hacky XXX | |
4729 numDirtyCards = | |
4730 _modUnionTable.heapWordDiffToOffsetDiff(dirtyRegion.word_size()); | |
4731 | |
4732 // We'll scan the cards in the dirty region (with periodic | |
4733 // yields for foreground GC as needed). | |
4734 if (!dirtyRegion.is_empty()) { | |
4735 assert(numDirtyCards > 0, "consistency check"); | |
4736 HeapWord* stop_point = NULL; | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
4737 stopTimer(); |
935 | 4738 // Potential yield point |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
4739 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
4740 bitMapLock()); |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
4741 startTimer(); |
0 | 4742 { |
4743 verify_work_stacks_empty(); | |
4744 verify_overflow_empty(); | |
4745 sample_eden(); | |
1190
4788266644c1
6895236: CMS: cmsOopClosures.inline.hpp:43 assert(..., "Should remember klasses in this context")
jmasa
parents:
1145
diff
changeset
|
4746 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) |
0 | 4747 stop_point = |
4748 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); | |
4749 } | |
4750 if (stop_point != NULL) { | |
4751 // The careful iteration stopped early either because it found an | |
4752 // uninitialized object, or because we were in the midst of an | |
4753 // "abortable preclean", which should now be aborted. Redirty | |
4754 // the bits corresponding to the partially-scanned or unscanned | |
4755 // cards. We'll either restart at the next block boundary or | |
4756 // abort the preclean. | |
4757 assert((CMSPermGenPrecleaningEnabled && (gen == _permGen)) || | |
4758 (_collectorState == AbortablePreclean && should_abort_preclean()), | |
4759 "Unparsable objects should only be in perm gen."); | |
4760 _modUnionTable.mark_range(MemRegion(stop_point, dirtyRegion.end())); | |
4761 if (should_abort_preclean()) { | |
4762 break; // out of preclean loop | |
4763 } else { | |
4764 // Compute the next address at which preclean should pick up; | |
4765 // might need bitMapLock in order to read P-bits. | |
4766 lastAddr = next_card_start_after_block(stop_point); | |
4767 } | |
4768 } | |
4769 } else { | |
4770 assert(lastAddr == endAddr, "consistency check"); | |
4771 assert(numDirtyCards == 0, "consistency check"); | |
4772 break; | |
4773 } | |
4774 } | |
4775 verify_work_stacks_empty(); | |
4776 verify_overflow_empty(); | |
4777 return cumNumDirtyCards; | |
4778 } | |
4779 | |
4780 // NOTE: preclean_mod_union_table() above and preclean_card_table() | |
4781 // below are largely identical; if you need to modify | |
4782 // one of these methods, please check the other method too. | |
4783 | |
4784 size_t CMSCollector::preclean_card_table(ConcurrentMarkSweepGeneration* gen, | |
4785 ScanMarkedObjectsAgainCarefullyClosure* cl) { | |
4786 // strategy: it's similar to precleamModUnionTable above, in that | |
4787 // we accumulate contiguous ranges of dirty cards, mark these cards | |
4788 // precleaned, then scan the region covered by these cards. | |
4789 HeapWord* endAddr = (HeapWord*)(gen->_virtual_space.high()); | |
4790 HeapWord* startAddr = (HeapWord*)(gen->_virtual_space.low()); | |
4791 | |
4792 cl->setFreelistLock(gen->freelistLock()); // needed for yielding | |
4793 | |
4794 size_t numDirtyCards, cumNumDirtyCards; | |
4795 HeapWord *lastAddr, *nextAddr; | |
4796 | |
4797 for (cumNumDirtyCards = numDirtyCards = 0, | |
4798 nextAddr = lastAddr = startAddr; | |
4799 nextAddr < endAddr; | |
4800 nextAddr = lastAddr, cumNumDirtyCards += numDirtyCards) { | |
4801 | |
4802 ResourceMark rm; | |
4803 HandleMark hm; | |
4804 | |
4805 MemRegion dirtyRegion; | |
4806 { | |
4807 // See comments in "Precleaning notes" above on why we | |
4808 // do this locking. XXX Could the locking overheads be | |
4809 // too high when dirty cards are sparse? [I don't think so.] | |
4810 stopTimer(); | |
4811 CMSTokenSync x(true); // is cms thread | |
4812 startTimer(); | |
4813 sample_eden(); | |
4814 // Get and clear dirty region from card table | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
4815 dirtyRegion = _ct->ct_bs()->dirty_card_range_after_reset( |
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
4816 MemRegion(nextAddr, endAddr), |
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
4817 true, |
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
4818 CardTableModRefBS::precleaned_card_val()); |
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
4819 |
0 | 4820 assert(dirtyRegion.start() >= nextAddr, |
4821 "returned region inconsistent?"); | |
4822 } | |
4823 lastAddr = dirtyRegion.end(); | |
4824 numDirtyCards = | |
4825 dirtyRegion.word_size()/CardTableModRefBS::card_size_in_words; | |
4826 | |
4827 if (!dirtyRegion.is_empty()) { | |
4828 stopTimer(); | |
4829 CMSTokenSyncWithLocks ts(true, gen->freelistLock(), bitMapLock()); | |
4830 startTimer(); | |
4831 sample_eden(); | |
4832 verify_work_stacks_empty(); | |
4833 verify_overflow_empty(); | |
1190
4788266644c1
6895236: CMS: cmsOopClosures.inline.hpp:43 assert(..., "Should remember klasses in this context")
jmasa
parents:
1145
diff
changeset
|
4834 DEBUG_ONLY(RememberKlassesChecker mx(should_unload_classes());) |
0 | 4835 HeapWord* stop_point = |
4836 gen->cmsSpace()->object_iterate_careful_m(dirtyRegion, cl); | |
4837 if (stop_point != NULL) { | |
4838 // The careful iteration stopped early because it found an | |
4839 // uninitialized object. Redirty the bits corresponding to the | |
4840 // partially-scanned or unscanned cards, and start again at the | |
4841 // next block boundary. | |
4842 assert(CMSPermGenPrecleaningEnabled || | |
4843 (_collectorState == AbortablePreclean && should_abort_preclean()), | |
4844 "Unparsable objects should only be in perm gen."); | |
4845 _ct->ct_bs()->invalidate(MemRegion(stop_point, dirtyRegion.end())); | |
4846 if (should_abort_preclean()) { | |
4847 break; // out of preclean loop | |
4848 } else { | |
4849 // Compute the next address at which preclean should pick up. | |
4850 lastAddr = next_card_start_after_block(stop_point); | |
4851 } | |
4852 } | |
4853 } else { | |
4854 break; | |
4855 } | |
4856 } | |
4857 verify_work_stacks_empty(); | |
4858 verify_overflow_empty(); | |
4859 return cumNumDirtyCards; | |
4860 } | |
4861 | |
4862 void CMSCollector::checkpointRootsFinal(bool asynch, | |
4863 bool clear_all_soft_refs, bool init_mark_was_synchronous) { | |
4864 assert(_collectorState == FinalMarking, "incorrect state transition?"); | |
4865 check_correct_thread_executing(); | |
4866 // world is stopped at this checkpoint | |
4867 assert(SafepointSynchronize::is_at_safepoint(), | |
4868 "world should be stopped"); | |
1703
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
4869 TraceCMSMemoryManagerStats tms(_collectorState); |
0 | 4870 verify_work_stacks_empty(); |
4871 verify_overflow_empty(); | |
4872 | |
4873 SpecializationStats::clear(); | |
4874 if (PrintGCDetails) { | |
4875 gclog_or_tty->print("[YG occupancy: "SIZE_FORMAT" K ("SIZE_FORMAT" K)]", | |
4876 _young_gen->used() / K, | |
4877 _young_gen->capacity() / K); | |
4878 } | |
4879 if (asynch) { | |
4880 if (CMSScavengeBeforeRemark) { | |
4881 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
4882 // Temporarily set flag to false, GCH->do_collection will | |
4883 // expect it to be false and set to true | |
4884 FlagSetting fl(gch->_is_gc_active, false); | |
4885 NOT_PRODUCT(TraceTime t("Scavenge-Before-Remark", | |
4886 PrintGCDetails && Verbose, true, gclog_or_tty);) | |
4887 int level = _cmsGen->level() - 1; | |
4888 if (level >= 0) { | |
4889 gch->do_collection(true, // full (i.e. force, see below) | |
4890 false, // !clear_all_soft_refs | |
4891 0, // size | |
4892 false, // is_tlab | |
4893 level // max_level | |
4894 ); | |
4895 } | |
4896 } | |
4897 FreelistLocker x(this); | |
4898 MutexLockerEx y(bitMapLock(), | |
4899 Mutex::_no_safepoint_check_flag); | |
4900 assert(!init_mark_was_synchronous, "but that's impossible!"); | |
4901 checkpointRootsFinalWork(asynch, clear_all_soft_refs, false); | |
4902 } else { | |
4903 // already have all the locks | |
4904 checkpointRootsFinalWork(asynch, clear_all_soft_refs, | |
4905 init_mark_was_synchronous); | |
4906 } | |
4907 verify_work_stacks_empty(); | |
4908 verify_overflow_empty(); | |
4909 SpecializationStats::print(); | |
4910 } | |
4911 | |
4912 void CMSCollector::checkpointRootsFinalWork(bool asynch, | |
4913 bool clear_all_soft_refs, bool init_mark_was_synchronous) { | |
4914 | |
4915 NOT_PRODUCT(TraceTime tr("checkpointRootsFinalWork", PrintGCDetails, false, gclog_or_tty);) | |
4916 | |
4917 assert(haveFreelistLocks(), "must have free list locks"); | |
4918 assert_lock_strong(bitMapLock()); | |
4919 | |
4920 if (UseAdaptiveSizePolicy) { | |
4921 size_policy()->checkpoint_roots_final_begin(); | |
4922 } | |
4923 | |
4924 ResourceMark rm; | |
4925 HandleMark hm; | |
4926 | |
4927 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
4928 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
4929 if (should_unload_classes()) { |
0 | 4930 CodeCache::gc_prologue(); |
4931 } | |
4932 assert(haveFreelistLocks(), "must have free list locks"); | |
4933 assert_lock_strong(bitMapLock()); | |
4934 | |
1190
4788266644c1
6895236: CMS: cmsOopClosures.inline.hpp:43 assert(..., "Should remember klasses in this context")
jmasa
parents:
1145
diff
changeset
|
4935 DEBUG_ONLY(RememberKlassesChecker fmx(should_unload_classes());) |
0 | 4936 if (!init_mark_was_synchronous) { |
4937 // We might assume that we need not fill TLAB's when | |
4938 // CMSScavengeBeforeRemark is set, because we may have just done | |
4939 // a scavenge which would have filled all TLAB's -- and besides | |
4940 // Eden would be empty. This however may not always be the case -- | |
4941 // for instance although we asked for a scavenge, it may not have | |
4942 // happened because of a JNI critical section. We probably need | |
4943 // a policy for deciding whether we can in that case wait until | |
4944 // the critical section releases and then do the remark following | |
4945 // the scavenge, and skip it here. In the absence of that policy, | |
4946 // or of an indication of whether the scavenge did indeed occur, | |
4947 // we cannot rely on TLAB's having been filled and must do | |
4948 // so here just in case a scavenge did not happen. | |
4949 gch->ensure_parsability(false); // fill TLAB's, but no need to retire them | |
4950 // Update the saved marks which may affect the root scans. | |
4951 gch->save_marks(); | |
4952 | |
4953 { | |
4954 COMPILER2_PRESENT(DerivedPointerTableDeactivate dpt_deact;) | |
4955 | |
4956 // Note on the role of the mod union table: | |
4957 // Since the marker in "markFromRoots" marks concurrently with | |
4958 // mutators, it is possible for some reachable objects not to have been | |
4959 // scanned. For instance, an only reference to an object A was | |
4960 // placed in object B after the marker scanned B. Unless B is rescanned, | |
4961 // A would be collected. Such updates to references in marked objects | |
4962 // are detected via the mod union table which is the set of all cards | |
4963 // dirtied since the first checkpoint in this GC cycle and prior to | |
4964 // the most recent young generation GC, minus those cleaned up by the | |
4965 // concurrent precleaning. | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
4966 if (CMSParallelRemarkEnabled && CollectedHeap::use_parallel_gc_threads()) { |
0 | 4967 TraceTime t("Rescan (parallel) ", PrintGCDetails, false, gclog_or_tty); |
4968 do_remark_parallel(); | |
4969 } else { | |
4970 TraceTime t("Rescan (non-parallel) ", PrintGCDetails, false, | |
4971 gclog_or_tty); | |
4972 do_remark_non_parallel(); | |
4973 } | |
4974 } | |
4975 } else { | |
4976 assert(!asynch, "Can't have init_mark_was_synchronous in asynch mode"); | |
4977 // The initial mark was stop-world, so there's no rescanning to | |
4978 // do; go straight on to the next step below. | |
4979 } | |
4980 verify_work_stacks_empty(); | |
4981 verify_overflow_empty(); | |
4982 | |
4983 { | |
4984 NOT_PRODUCT(TraceTime ts("refProcessingWork", PrintGCDetails, false, gclog_or_tty);) | |
4985 refProcessingWork(asynch, clear_all_soft_refs); | |
4986 } | |
4987 verify_work_stacks_empty(); | |
4988 verify_overflow_empty(); | |
4989 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
4990 if (should_unload_classes()) { |
0 | 4991 CodeCache::gc_epilogue(); |
4992 } | |
4993 | |
4994 // If we encountered any (marking stack / work queue) overflow | |
4995 // events during the current CMS cycle, take appropriate | |
4996 // remedial measures, where possible, so as to try and avoid | |
4997 // recurrence of that condition. | |
4998 assert(_markStack.isEmpty(), "No grey objects"); | |
4999 size_t ser_ovflw = _ser_pmc_remark_ovflw + _ser_pmc_preclean_ovflw + | |
452
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
5000 _ser_kac_ovflw + _ser_kac_preclean_ovflw; |
0 | 5001 if (ser_ovflw > 0) { |
5002 if (PrintCMSStatistics != 0) { | |
5003 gclog_or_tty->print_cr("Marking stack overflow (benign) " | |
452
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
5004 "(pmc_pc="SIZE_FORMAT", pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
5005 ", kac_preclean="SIZE_FORMAT")", |
0 | 5006 _ser_pmc_preclean_ovflw, _ser_pmc_remark_ovflw, |
452
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
5007 _ser_kac_ovflw, _ser_kac_preclean_ovflw); |
0 | 5008 } |
5009 _markStack.expand(); | |
5010 _ser_pmc_remark_ovflw = 0; | |
5011 _ser_pmc_preclean_ovflw = 0; | |
452
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
5012 _ser_kac_preclean_ovflw = 0; |
0 | 5013 _ser_kac_ovflw = 0; |
5014 } | |
5015 if (_par_pmc_remark_ovflw > 0 || _par_kac_ovflw > 0) { | |
5016 if (PrintCMSStatistics != 0) { | |
5017 gclog_or_tty->print_cr("Work queue overflow (benign) " | |
5018 "(pmc_rm="SIZE_FORMAT", kac="SIZE_FORMAT")", | |
5019 _par_pmc_remark_ovflw, _par_kac_ovflw); | |
5020 } | |
5021 _par_pmc_remark_ovflw = 0; | |
5022 _par_kac_ovflw = 0; | |
5023 } | |
5024 if (PrintCMSStatistics != 0) { | |
5025 if (_markStack._hit_limit > 0) { | |
5026 gclog_or_tty->print_cr(" (benign) Hit max stack size limit ("SIZE_FORMAT")", | |
5027 _markStack._hit_limit); | |
5028 } | |
5029 if (_markStack._failed_double > 0) { | |
5030 gclog_or_tty->print_cr(" (benign) Failed stack doubling ("SIZE_FORMAT")," | |
5031 " current capacity "SIZE_FORMAT, | |
5032 _markStack._failed_double, | |
5033 _markStack.capacity()); | |
5034 } | |
5035 } | |
5036 _markStack._hit_limit = 0; | |
5037 _markStack._failed_double = 0; | |
5038 | |
935 | 5039 // Check that all the klasses have been checked |
5040 assert(_revisitStack.isEmpty(), "Not all klasses revisited"); | |
5041 | |
0 | 5042 if ((VerifyAfterGC || VerifyDuringGC) && |
5043 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { | |
5044 verify_after_remark(); | |
5045 } | |
5046 | |
5047 // Change under the freelistLocks. | |
5048 _collectorState = Sweeping; | |
5049 // Call isAllClear() under bitMapLock | |
5050 assert(_modUnionTable.isAllClear(), "Should be clear by end of the" | |
5051 " final marking"); | |
5052 if (UseAdaptiveSizePolicy) { | |
5053 size_policy()->checkpoint_roots_final_end(gch->gc_cause()); | |
5054 } | |
5055 } | |
5056 | |
5057 // Parallel remark task | |
5058 class CMSParRemarkTask: public AbstractGangTask { | |
5059 CMSCollector* _collector; | |
5060 int _n_workers; | |
5061 CompactibleFreeListSpace* _cms_space; | |
5062 CompactibleFreeListSpace* _perm_space; | |
5063 | |
5064 // The per-thread work queues, available here for stealing. | |
5065 OopTaskQueueSet* _task_queues; | |
5066 ParallelTaskTerminator _term; | |
5067 | |
5068 public: | |
5069 CMSParRemarkTask(CMSCollector* collector, | |
5070 CompactibleFreeListSpace* cms_space, | |
5071 CompactibleFreeListSpace* perm_space, | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5072 int n_workers, FlexibleWorkGang* workers, |
0 | 5073 OopTaskQueueSet* task_queues): |
5074 AbstractGangTask("Rescan roots and grey objects in parallel"), | |
5075 _collector(collector), | |
5076 _cms_space(cms_space), _perm_space(perm_space), | |
5077 _n_workers(n_workers), | |
5078 _task_queues(task_queues), | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5079 _term(n_workers, task_queues) { } |
0 | 5080 |
5081 OopTaskQueueSet* task_queues() { return _task_queues; } | |
5082 | |
5083 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } | |
5084 | |
5085 ParallelTaskTerminator* terminator() { return &_term; } | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5086 int n_workers() { return _n_workers; } |
0 | 5087 |
5088 void work(int i); | |
5089 | |
5090 private: | |
5091 // Work method in support of parallel rescan ... of young gen spaces | |
5092 void do_young_space_rescan(int i, Par_MarkRefsIntoAndScanClosure* cl, | |
5093 ContiguousSpace* space, | |
5094 HeapWord** chunk_array, size_t chunk_top); | |
5095 | |
5096 // ... of dirty cards in old space | |
5097 void do_dirty_card_rescan_tasks(CompactibleFreeListSpace* sp, int i, | |
5098 Par_MarkRefsIntoAndScanClosure* cl); | |
5099 | |
5100 // ... work stealing for the above | |
5101 void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed); | |
5102 }; | |
5103 | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5104 // work_queue(i) is passed to the closure |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5105 // Par_MarkRefsIntoAndScanClosure. The "i" parameter |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5106 // also is passed to do_dirty_card_rescan_tasks() and to |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5107 // do_work_steal() to select the i-th task_queue. |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5108 |
0 | 5109 void CMSParRemarkTask::work(int i) { |
5110 elapsedTimer _timer; | |
5111 ResourceMark rm; | |
5112 HandleMark hm; | |
5113 | |
5114 // ---------- rescan from roots -------------- | |
5115 _timer.start(); | |
5116 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
5117 Par_MarkRefsIntoAndScanClosure par_mrias_cl(_collector, | |
5118 _collector->_span, _collector->ref_processor(), | |
5119 &(_collector->_markBitMap), | |
5120 work_queue(i), &(_collector->_revisitStack)); | |
5121 | |
5122 // Rescan young gen roots first since these are likely | |
5123 // coarsely partitioned and may, on that account, constitute | |
5124 // the critical path; thus, it's best to start off that | |
5125 // work first. | |
5126 // ---------- young gen roots -------------- | |
5127 { | |
5128 DefNewGeneration* dng = _collector->_young_gen->as_DefNewGeneration(); | |
5129 EdenSpace* eden_space = dng->eden(); | |
5130 ContiguousSpace* from_space = dng->from(); | |
5131 ContiguousSpace* to_space = dng->to(); | |
5132 | |
5133 HeapWord** eca = _collector->_eden_chunk_array; | |
5134 size_t ect = _collector->_eden_chunk_index; | |
5135 HeapWord** sca = _collector->_survivor_chunk_array; | |
5136 size_t sct = _collector->_survivor_chunk_index; | |
5137 | |
5138 assert(ect <= _collector->_eden_chunk_capacity, "out of bounds"); | |
5139 assert(sct <= _collector->_survivor_chunk_capacity, "out of bounds"); | |
5140 | |
5141 do_young_space_rescan(i, &par_mrias_cl, to_space, NULL, 0); | |
5142 do_young_space_rescan(i, &par_mrias_cl, from_space, sca, sct); | |
5143 do_young_space_rescan(i, &par_mrias_cl, eden_space, eca, ect); | |
5144 | |
5145 _timer.stop(); | |
5146 if (PrintCMSStatistics != 0) { | |
5147 gclog_or_tty->print_cr( | |
5148 "Finished young gen rescan work in %dth thread: %3.3f sec", | |
5149 i, _timer.seconds()); | |
5150 } | |
5151 } | |
5152 | |
5153 // ---------- remaining roots -------------- | |
5154 _timer.reset(); | |
5155 _timer.start(); | |
5156 gch->gen_process_strong_roots(_collector->_cmsGen->level(), | |
5157 false, // yg was scanned above | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5158 false, // this is parallel code |
0 | 5159 true, // collecting perm gen |
5160 SharedHeap::ScanningOption(_collector->CMSCollector::roots_scanning_options()), | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5161 &par_mrias_cl, |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5162 true, // walk all of code cache if (so & SO_CodeCache) |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5163 NULL); |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5164 assert(_collector->should_unload_classes() |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5165 || (_collector->CMSCollector::roots_scanning_options() & SharedHeap::SO_CodeCache), |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5166 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); |
0 | 5167 _timer.stop(); |
5168 if (PrintCMSStatistics != 0) { | |
5169 gclog_or_tty->print_cr( | |
5170 "Finished remaining root rescan work in %dth thread: %3.3f sec", | |
5171 i, _timer.seconds()); | |
5172 } | |
5173 | |
5174 // ---------- rescan dirty cards ------------ | |
5175 _timer.reset(); | |
5176 _timer.start(); | |
5177 | |
5178 // Do the rescan tasks for each of the two spaces | |
5179 // (cms_space and perm_space) in turn. | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5180 // "i" is passed to select the "i-th" task_queue |
0 | 5181 do_dirty_card_rescan_tasks(_cms_space, i, &par_mrias_cl); |
5182 do_dirty_card_rescan_tasks(_perm_space, i, &par_mrias_cl); | |
5183 _timer.stop(); | |
5184 if (PrintCMSStatistics != 0) { | |
5185 gclog_or_tty->print_cr( | |
5186 "Finished dirty card rescan work in %dth thread: %3.3f sec", | |
5187 i, _timer.seconds()); | |
5188 } | |
5189 | |
5190 // ---------- steal work from other threads ... | |
5191 // ---------- ... and drain overflow list. | |
5192 _timer.reset(); | |
5193 _timer.start(); | |
5194 do_work_steal(i, &par_mrias_cl, _collector->hash_seed(i)); | |
5195 _timer.stop(); | |
5196 if (PrintCMSStatistics != 0) { | |
5197 gclog_or_tty->print_cr( | |
5198 "Finished work stealing in %dth thread: %3.3f sec", | |
5199 i, _timer.seconds()); | |
5200 } | |
5201 } | |
5202 | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5203 // Note that parameter "i" is not used. |
0 | 5204 void |
5205 CMSParRemarkTask::do_young_space_rescan(int i, | |
5206 Par_MarkRefsIntoAndScanClosure* cl, ContiguousSpace* space, | |
5207 HeapWord** chunk_array, size_t chunk_top) { | |
5208 // Until all tasks completed: | |
5209 // . claim an unclaimed task | |
5210 // . compute region boundaries corresponding to task claimed | |
5211 // using chunk_array | |
5212 // . par_oop_iterate(cl) over that region | |
5213 | |
5214 ResourceMark rm; | |
5215 HandleMark hm; | |
5216 | |
5217 SequentialSubTasksDone* pst = space->par_seq_tasks(); | |
5218 assert(pst->valid(), "Uninitialized use?"); | |
5219 | |
5220 int nth_task = 0; | |
5221 int n_tasks = pst->n_tasks(); | |
5222 | |
5223 HeapWord *start, *end; | |
5224 while (!pst->is_task_claimed(/* reference */ nth_task)) { | |
5225 // We claimed task # nth_task; compute its boundaries. | |
5226 if (chunk_top == 0) { // no samples were taken | |
5227 assert(nth_task == 0 && n_tasks == 1, "Can have only 1 EdenSpace task"); | |
5228 start = space->bottom(); | |
5229 end = space->top(); | |
5230 } else if (nth_task == 0) { | |
5231 start = space->bottom(); | |
5232 end = chunk_array[nth_task]; | |
5233 } else if (nth_task < (jint)chunk_top) { | |
5234 assert(nth_task >= 1, "Control point invariant"); | |
5235 start = chunk_array[nth_task - 1]; | |
5236 end = chunk_array[nth_task]; | |
5237 } else { | |
5238 assert(nth_task == (jint)chunk_top, "Control point invariant"); | |
5239 start = chunk_array[chunk_top - 1]; | |
5240 end = space->top(); | |
5241 } | |
5242 MemRegion mr(start, end); | |
5243 // Verify that mr is in space | |
5244 assert(mr.is_empty() || space->used_region().contains(mr), | |
5245 "Should be in space"); | |
5246 // Verify that "start" is an object boundary | |
5247 assert(mr.is_empty() || oop(mr.start())->is_oop(), | |
5248 "Should be an oop"); | |
5249 space->par_oop_iterate(mr, cl); | |
5250 } | |
5251 pst->all_tasks_completed(); | |
5252 } | |
5253 | |
5254 void | |
5255 CMSParRemarkTask::do_dirty_card_rescan_tasks( | |
5256 CompactibleFreeListSpace* sp, int i, | |
5257 Par_MarkRefsIntoAndScanClosure* cl) { | |
5258 // Until all tasks completed: | |
5259 // . claim an unclaimed task | |
5260 // . compute region boundaries corresponding to task claimed | |
5261 // . transfer dirty bits ct->mut for that region | |
5262 // . apply rescanclosure to dirty mut bits for that region | |
5263 | |
5264 ResourceMark rm; | |
5265 HandleMark hm; | |
5266 | |
5267 OopTaskQueue* work_q = work_queue(i); | |
5268 ModUnionClosure modUnionClosure(&(_collector->_modUnionTable)); | |
5269 // CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! CAUTION! | |
5270 // CAUTION: This closure has state that persists across calls to | |
5271 // the work method dirty_range_iterate_clear() in that it has | |
5272 // imbedded in it a (subtype of) UpwardsObjectClosure. The | |
5273 // use of that state in the imbedded UpwardsObjectClosure instance | |
5274 // assumes that the cards are always iterated (even if in parallel | |
5275 // by several threads) in monotonically increasing order per each | |
5276 // thread. This is true of the implementation below which picks | |
5277 // card ranges (chunks) in monotonically increasing order globally | |
5278 // and, a-fortiori, in monotonically increasing order per thread | |
5279 // (the latter order being a subsequence of the former). | |
5280 // If the work code below is ever reorganized into a more chaotic | |
5281 // work-partitioning form than the current "sequential tasks" | |
5282 // paradigm, the use of that persistent state will have to be | |
5283 // revisited and modified appropriately. See also related | |
5284 // bug 4756801 work on which should examine this code to make | |
5285 // sure that the changes there do not run counter to the | |
5286 // assumptions made here and necessary for correctness and | |
5287 // efficiency. Note also that this code might yield inefficient | |
5288 // behaviour in the case of very large objects that span one or | |
5289 // more work chunks. Such objects would potentially be scanned | |
5290 // several times redundantly. Work on 4756801 should try and | |
5291 // address that performance anomaly if at all possible. XXX | |
5292 MemRegion full_span = _collector->_span; | |
5293 CMSBitMap* bm = &(_collector->_markBitMap); // shared | |
5294 CMSMarkStack* rs = &(_collector->_revisitStack); // shared | |
5295 MarkFromDirtyCardsClosure | |
5296 greyRescanClosure(_collector, full_span, // entire span of interest | |
5297 sp, bm, work_q, rs, cl); | |
5298 | |
5299 SequentialSubTasksDone* pst = sp->conc_par_seq_tasks(); | |
5300 assert(pst->valid(), "Uninitialized use?"); | |
5301 int nth_task = 0; | |
5302 const int alignment = CardTableModRefBS::card_size * BitsPerWord; | |
5303 MemRegion span = sp->used_region(); | |
5304 HeapWord* start_addr = span.start(); | |
5305 HeapWord* end_addr = (HeapWord*)round_to((intptr_t)span.end(), | |
5306 alignment); | |
5307 const size_t chunk_size = sp->rescan_task_size(); // in HeapWord units | |
5308 assert((HeapWord*)round_to((intptr_t)start_addr, alignment) == | |
5309 start_addr, "Check alignment"); | |
5310 assert((size_t)round_to((intptr_t)chunk_size, alignment) == | |
5311 chunk_size, "Check alignment"); | |
5312 | |
5313 while (!pst->is_task_claimed(/* reference */ nth_task)) { | |
5314 // Having claimed the nth_task, compute corresponding mem-region, | |
5315 // which is a-fortiori aligned correctly (i.e. at a MUT bopundary). | |
5316 // The alignment restriction ensures that we do not need any | |
5317 // synchronization with other gang-workers while setting or | |
5318 // clearing bits in thus chunk of the MUT. | |
5319 MemRegion this_span = MemRegion(start_addr + nth_task*chunk_size, | |
5320 start_addr + (nth_task+1)*chunk_size); | |
5321 // The last chunk's end might be way beyond end of the | |
5322 // used region. In that case pull back appropriately. | |
5323 if (this_span.end() > end_addr) { | |
5324 this_span.set_end(end_addr); | |
5325 assert(!this_span.is_empty(), "Program logic (calculation of n_tasks)"); | |
5326 } | |
5327 // Iterate over the dirty cards covering this chunk, marking them | |
5328 // precleaned, and setting the corresponding bits in the mod union | |
5329 // table. Since we have been careful to partition at Card and MUT-word | |
5330 // boundaries no synchronization is needed between parallel threads. | |
5331 _collector->_ct->ct_bs()->dirty_card_iterate(this_span, | |
5332 &modUnionClosure); | |
5333 | |
5334 // Having transferred these marks into the modUnionTable, | |
5335 // rescan the marked objects on the dirty cards in the modUnionTable. | |
5336 // Even if this is at a synchronous collection, the initial marking | |
5337 // may have been done during an asynchronous collection so there | |
5338 // may be dirty bits in the mod-union table. | |
5339 _collector->_modUnionTable.dirty_range_iterate_clear( | |
5340 this_span, &greyRescanClosure); | |
5341 _collector->_modUnionTable.verifyNoOneBitsInRange( | |
5342 this_span.start(), | |
5343 this_span.end()); | |
5344 } | |
5345 pst->all_tasks_completed(); // declare that i am done | |
5346 } | |
5347 | |
5348 // . see if we can share work_queues with ParNew? XXX | |
5349 void | |
5350 CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, | |
5351 int* seed) { | |
5352 OopTaskQueue* work_q = work_queue(i); | |
5353 NOT_PRODUCT(int num_steals = 0;) | |
5354 oop obj_to_scan; | |
5355 CMSBitMap* bm = &(_collector->_markBitMap); | |
5356 | |
5357 while (true) { | |
5358 // Completely finish any left over work from (an) earlier round(s) | |
5359 cl->trim_queue(0); | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
628
diff
changeset
|
5360 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
628
diff
changeset
|
5361 (size_t)ParGCDesiredObjsFromOverflowList); |
0 | 5362 // Now check if there's any work in the overflow list |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5363 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5364 // only affects the number of attempts made to get work from the |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5365 // overflow list and does not affect the number of workers. Just |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5366 // pass ParallelGCThreads so this behavior is unchanged. |
0 | 5367 if (_collector->par_take_from_overflow_list(num_from_overflow_list, |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5368 work_q, |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5369 ParallelGCThreads)) { |
0 | 5370 // found something in global overflow list; |
5371 // not yet ready to go stealing work from others. | |
5372 // We'd like to assert(work_q->size() != 0, ...) | |
5373 // because we just took work from the overflow list, | |
5374 // but of course we can't since all of that could have | |
5375 // been already stolen from us. | |
5376 // "He giveth and He taketh away." | |
5377 continue; | |
5378 } | |
5379 // Verify that we have no work before we resort to stealing | |
5380 assert(work_q->size() == 0, "Have work, shouldn't steal"); | |
5381 // Try to steal from other queues that have work | |
5382 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { | |
5383 NOT_PRODUCT(num_steals++;) | |
5384 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); | |
5385 assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); | |
5386 // Do scanning work | |
5387 obj_to_scan->oop_iterate(cl); | |
5388 // Loop around, finish this work, and try to steal some more | |
5389 } else if (terminator()->offer_termination()) { | |
5390 break; // nirvana from the infinite cycle | |
5391 } | |
5392 } | |
5393 NOT_PRODUCT( | |
5394 if (PrintCMSStatistics != 0) { | |
5395 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); | |
5396 } | |
5397 ) | |
5398 assert(work_q->size() == 0 && _collector->overflow_list_is_empty(), | |
5399 "Else our work is not yet done"); | |
5400 } | |
5401 | |
5402 // Return a thread-local PLAB recording array, as appropriate. | |
5403 void* CMSCollector::get_data_recorder(int thr_num) { | |
5404 if (_survivor_plab_array != NULL && | |
5405 (CMSPLABRecordAlways || | |
5406 (_collectorState > Marking && _collectorState < FinalMarking))) { | |
5407 assert(thr_num < (int)ParallelGCThreads, "thr_num is out of bounds"); | |
5408 ChunkArray* ca = &_survivor_plab_array[thr_num]; | |
5409 ca->reset(); // clear it so that fresh data is recorded | |
5410 return (void*) ca; | |
5411 } else { | |
5412 return NULL; | |
5413 } | |
5414 } | |
5415 | |
5416 // Reset all the thread-local PLAB recording arrays | |
5417 void CMSCollector::reset_survivor_plab_arrays() { | |
5418 for (uint i = 0; i < ParallelGCThreads; i++) { | |
5419 _survivor_plab_array[i].reset(); | |
5420 } | |
5421 } | |
5422 | |
5423 // Merge the per-thread plab arrays into the global survivor chunk | |
5424 // array which will provide the partitioning of the survivor space | |
5425 // for CMS rescan. | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5426 void CMSCollector::merge_survivor_plab_arrays(ContiguousSpace* surv, |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5427 int no_of_gc_threads) { |
0 | 5428 assert(_survivor_plab_array != NULL, "Error"); |
5429 assert(_survivor_chunk_array != NULL, "Error"); | |
5430 assert(_collectorState == FinalMarking, "Error"); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5431 for (int j = 0; j < no_of_gc_threads; j++) { |
0 | 5432 _cursor[j] = 0; |
5433 } | |
5434 HeapWord* top = surv->top(); | |
5435 size_t i; | |
5436 for (i = 0; i < _survivor_chunk_capacity; i++) { // all sca entries | |
5437 HeapWord* min_val = top; // Higher than any PLAB address | |
5438 uint min_tid = 0; // position of min_val this round | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5439 for (int j = 0; j < no_of_gc_threads; j++) { |
0 | 5440 ChunkArray* cur_sca = &_survivor_plab_array[j]; |
5441 if (_cursor[j] == cur_sca->end()) { | |
5442 continue; | |
5443 } | |
5444 assert(_cursor[j] < cur_sca->end(), "ctl pt invariant"); | |
5445 HeapWord* cur_val = cur_sca->nth(_cursor[j]); | |
5446 assert(surv->used_region().contains(cur_val), "Out of bounds value"); | |
5447 if (cur_val < min_val) { | |
5448 min_tid = j; | |
5449 min_val = cur_val; | |
5450 } else { | |
5451 assert(cur_val < top, "All recorded addresses should be less"); | |
5452 } | |
5453 } | |
5454 // At this point min_val and min_tid are respectively | |
5455 // the least address in _survivor_plab_array[j]->nth(_cursor[j]) | |
5456 // and the thread (j) that witnesses that address. | |
5457 // We record this address in the _survivor_chunk_array[i] | |
5458 // and increment _cursor[min_tid] prior to the next round i. | |
5459 if (min_val == top) { | |
5460 break; | |
5461 } | |
5462 _survivor_chunk_array[i] = min_val; | |
5463 _cursor[min_tid]++; | |
5464 } | |
5465 // We are all done; record the size of the _survivor_chunk_array | |
5466 _survivor_chunk_index = i; // exclusive: [0, i) | |
5467 if (PrintCMSStatistics > 0) { | |
5468 gclog_or_tty->print(" (Survivor:" SIZE_FORMAT "chunks) ", i); | |
5469 } | |
5470 // Verify that we used up all the recorded entries | |
5471 #ifdef ASSERT | |
5472 size_t total = 0; | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5473 for (int j = 0; j < no_of_gc_threads; j++) { |
0 | 5474 assert(_cursor[j] == _survivor_plab_array[j].end(), "Ctl pt invariant"); |
5475 total += _cursor[j]; | |
5476 } | |
5477 assert(total == _survivor_chunk_index, "Ctl Pt Invariant"); | |
5478 // Check that the merged array is in sorted order | |
5479 if (total > 0) { | |
5480 for (size_t i = 0; i < total - 1; i++) { | |
5481 if (PrintCMSStatistics > 0) { | |
5482 gclog_or_tty->print(" (chunk" SIZE_FORMAT ":" INTPTR_FORMAT ") ", | |
5483 i, _survivor_chunk_array[i]); | |
5484 } | |
5485 assert(_survivor_chunk_array[i] < _survivor_chunk_array[i+1], | |
5486 "Not sorted"); | |
5487 } | |
5488 } | |
5489 #endif // ASSERT | |
5490 } | |
5491 | |
5492 // Set up the space's par_seq_tasks structure for work claiming | |
5493 // for parallel rescan of young gen. | |
5494 // See ParRescanTask where this is currently used. | |
5495 void | |
5496 CMSCollector:: | |
5497 initialize_sequential_subtasks_for_young_gen_rescan(int n_threads) { | |
5498 assert(n_threads > 0, "Unexpected n_threads argument"); | |
5499 DefNewGeneration* dng = (DefNewGeneration*)_young_gen; | |
5500 | |
5501 // Eden space | |
5502 { | |
5503 SequentialSubTasksDone* pst = dng->eden()->par_seq_tasks(); | |
5504 assert(!pst->valid(), "Clobbering existing data?"); | |
5505 // Each valid entry in [0, _eden_chunk_index) represents a task. | |
5506 size_t n_tasks = _eden_chunk_index + 1; | |
5507 assert(n_tasks == 1 || _eden_chunk_array != NULL, "Error"); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5508 // Sets the condition for completion of the subtask (how many threads |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5509 // need to finish in order to be done). |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5510 pst->set_n_threads(n_threads); |
0 | 5511 pst->set_n_tasks((int)n_tasks); |
5512 } | |
5513 | |
5514 // Merge the survivor plab arrays into _survivor_chunk_array | |
5515 if (_survivor_plab_array != NULL) { | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5516 merge_survivor_plab_arrays(dng->from(), n_threads); |
0 | 5517 } else { |
5518 assert(_survivor_chunk_index == 0, "Error"); | |
5519 } | |
5520 | |
5521 // To space | |
5522 { | |
5523 SequentialSubTasksDone* pst = dng->to()->par_seq_tasks(); | |
5524 assert(!pst->valid(), "Clobbering existing data?"); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5525 // Sets the condition for completion of the subtask (how many threads |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5526 // need to finish in order to be done). |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5527 pst->set_n_threads(n_threads); |
0 | 5528 pst->set_n_tasks(1); |
5529 assert(pst->valid(), "Error"); | |
5530 } | |
5531 | |
5532 // From space | |
5533 { | |
5534 SequentialSubTasksDone* pst = dng->from()->par_seq_tasks(); | |
5535 assert(!pst->valid(), "Clobbering existing data?"); | |
5536 size_t n_tasks = _survivor_chunk_index + 1; | |
5537 assert(n_tasks == 1 || _survivor_chunk_array != NULL, "Error"); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5538 // Sets the condition for completion of the subtask (how many threads |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5539 // need to finish in order to be done). |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5540 pst->set_n_threads(n_threads); |
0 | 5541 pst->set_n_tasks((int)n_tasks); |
5542 assert(pst->valid(), "Error"); | |
5543 } | |
5544 } | |
5545 | |
5546 // Parallel version of remark | |
5547 void CMSCollector::do_remark_parallel() { | |
5548 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5549 FlexibleWorkGang* workers = gch->workers(); |
0 | 5550 assert(workers != NULL, "Need parallel worker threads."); |
5551 int n_workers = workers->total_workers(); | |
5552 CompactibleFreeListSpace* cms_space = _cmsGen->cmsSpace(); | |
5553 CompactibleFreeListSpace* perm_space = _permGen->cmsSpace(); | |
5554 | |
5555 CMSParRemarkTask tsk(this, | |
5556 cms_space, perm_space, | |
5557 n_workers, workers, task_queues()); | |
5558 | |
5559 // Set up for parallel process_strong_roots work. | |
5560 gch->set_par_threads(n_workers); | |
5561 // We won't be iterating over the cards in the card table updating | |
5562 // the younger_gen cards, so we shouldn't call the following else | |
5563 // the verification code as well as subsequent younger_refs_iterate | |
5564 // code would get confused. XXX | |
5565 // gch->rem_set()->prepare_for_younger_refs_iterate(true); // parallel | |
5566 | |
5567 // The young gen rescan work will not be done as part of | |
5568 // process_strong_roots (which currently doesn't knw how to | |
5569 // parallelize such a scan), but rather will be broken up into | |
5570 // a set of parallel tasks (via the sampling that the [abortable] | |
5571 // preclean phase did of EdenSpace, plus the [two] tasks of | |
5572 // scanning the [two] survivor spaces. Further fine-grain | |
5573 // parallelization of the scanning of the survivor spaces | |
5574 // themselves, and of precleaning of the younger gen itself | |
5575 // is deferred to the future. | |
5576 initialize_sequential_subtasks_for_young_gen_rescan(n_workers); | |
5577 | |
5578 // The dirty card rescan work is broken up into a "sequence" | |
5579 // of parallel tasks (per constituent space) that are dynamically | |
5580 // claimed by the parallel threads. | |
5581 cms_space->initialize_sequential_subtasks_for_rescan(n_workers); | |
5582 perm_space->initialize_sequential_subtasks_for_rescan(n_workers); | |
5583 | |
5584 // It turns out that even when we're using 1 thread, doing the work in a | |
5585 // separate thread causes wide variance in run times. We can't help this | |
5586 // in the multi-threaded case, but we special-case n=1 here to get | |
5587 // repeatable measurements of the 1-thread overhead of the parallel code. | |
5588 if (n_workers > 1) { | |
5589 // Make refs discovery MT-safe | |
5590 ReferenceProcessorMTMutator mt(ref_processor(), true); | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5591 GenCollectedHeap::StrongRootsScope srs(gch); |
0 | 5592 workers->run_task(&tsk); |
5593 } else { | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5594 GenCollectedHeap::StrongRootsScope srs(gch); |
0 | 5595 tsk.work(0); |
5596 } | |
5597 gch->set_par_threads(0); // 0 ==> non-parallel. | |
5598 // restore, single-threaded for now, any preserved marks | |
5599 // as a result of work_q overflow | |
5600 restore_preserved_marks_if_any(); | |
5601 } | |
5602 | |
5603 // Non-parallel version of remark | |
5604 void CMSCollector::do_remark_non_parallel() { | |
5605 ResourceMark rm; | |
5606 HandleMark hm; | |
5607 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
5608 MarkRefsIntoAndScanClosure | |
5609 mrias_cl(_span, ref_processor(), &_markBitMap, &_modUnionTable, | |
5610 &_markStack, &_revisitStack, this, | |
5611 false /* should_yield */, false /* not precleaning */); | |
5612 MarkFromDirtyCardsClosure | |
5613 markFromDirtyCardsClosure(this, _span, | |
5614 NULL, // space is set further below | |
5615 &_markBitMap, &_markStack, &_revisitStack, | |
5616 &mrias_cl); | |
5617 { | |
5618 TraceTime t("grey object rescan", PrintGCDetails, false, gclog_or_tty); | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
5619 // Iterate over the dirty cards, setting the corresponding bits in the |
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
5620 // mod union table. |
0 | 5621 { |
5622 ModUnionClosure modUnionClosure(&_modUnionTable); | |
5623 _ct->ct_bs()->dirty_card_iterate( | |
5624 _cmsGen->used_region(), | |
5625 &modUnionClosure); | |
5626 _ct->ct_bs()->dirty_card_iterate( | |
5627 _permGen->used_region(), | |
5628 &modUnionClosure); | |
5629 } | |
5630 // Having transferred these marks into the modUnionTable, we just need | |
5631 // to rescan the marked objects on the dirty cards in the modUnionTable. | |
5632 // The initial marking may have been done during an asynchronous | |
5633 // collection so there may be dirty bits in the mod-union table. | |
5634 const int alignment = | |
5635 CardTableModRefBS::card_size * BitsPerWord; | |
5636 { | |
5637 // ... First handle dirty cards in CMS gen | |
5638 markFromDirtyCardsClosure.set_space(_cmsGen->cmsSpace()); | |
5639 MemRegion ur = _cmsGen->used_region(); | |
5640 HeapWord* lb = ur.start(); | |
5641 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); | |
5642 MemRegion cms_span(lb, ub); | |
5643 _modUnionTable.dirty_range_iterate_clear(cms_span, | |
5644 &markFromDirtyCardsClosure); | |
5645 verify_work_stacks_empty(); | |
5646 if (PrintCMSStatistics != 0) { | |
5647 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in cms gen) ", | |
5648 markFromDirtyCardsClosure.num_dirty_cards()); | |
5649 } | |
5650 } | |
5651 { | |
5652 // .. and then repeat for dirty cards in perm gen | |
5653 markFromDirtyCardsClosure.set_space(_permGen->cmsSpace()); | |
5654 MemRegion ur = _permGen->used_region(); | |
5655 HeapWord* lb = ur.start(); | |
5656 HeapWord* ub = (HeapWord*)round_to((intptr_t)ur.end(), alignment); | |
5657 MemRegion perm_span(lb, ub); | |
5658 _modUnionTable.dirty_range_iterate_clear(perm_span, | |
5659 &markFromDirtyCardsClosure); | |
5660 verify_work_stacks_empty(); | |
5661 if (PrintCMSStatistics != 0) { | |
5662 gclog_or_tty->print(" (re-scanned "SIZE_FORMAT" dirty cards in perm gen) ", | |
5663 markFromDirtyCardsClosure.num_dirty_cards()); | |
5664 } | |
5665 } | |
5666 } | |
5667 if (VerifyDuringGC && | |
5668 GenCollectedHeap::heap()->total_collections() >= VerifyGCStartAt) { | |
5669 HandleMark hm; // Discard invalid handles created during verification | |
5670 Universe::verify(true); | |
5671 } | |
5672 { | |
5673 TraceTime t("root rescan", PrintGCDetails, false, gclog_or_tty); | |
5674 | |
5675 verify_work_stacks_empty(); | |
5676 | |
5677 gch->rem_set()->prepare_for_younger_refs_iterate(false); // Not parallel. | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5678 GenCollectedHeap::StrongRootsScope srs(gch); |
0 | 5679 gch->gen_process_strong_roots(_cmsGen->level(), |
5680 true, // younger gens as roots | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5681 false, // use the local StrongRootsScope |
0 | 5682 true, // collecting perm gen |
5683 SharedHeap::ScanningOption(roots_scanning_options()), | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5684 &mrias_cl, |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5685 true, // walk code active on stacks |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5686 NULL); |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5687 assert(should_unload_classes() |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5688 || (roots_scanning_options() & SharedHeap::SO_CodeCache), |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
798
diff
changeset
|
5689 "if we didn't scan the code cache, we have to be ready to drop nmethods with expired weak oops"); |
0 | 5690 } |
5691 verify_work_stacks_empty(); | |
5692 // Restore evacuated mark words, if any, used for overflow list links | |
5693 if (!CMSOverflowEarlyRestoration) { | |
5694 restore_preserved_marks_if_any(); | |
5695 } | |
5696 verify_overflow_empty(); | |
5697 } | |
5698 | |
5699 //////////////////////////////////////////////////////// | |
5700 // Parallel Reference Processing Task Proxy Class | |
5701 //////////////////////////////////////////////////////// | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5702 class CMSRefProcTaskProxy: public AbstractGangTaskWOopQueues { |
0 | 5703 typedef AbstractRefProcTaskExecutor::ProcessTask ProcessTask; |
5704 CMSCollector* _collector; | |
5705 CMSBitMap* _mark_bit_map; | |
143
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
5706 const MemRegion _span; |
0 | 5707 ProcessTask& _task; |
5708 | |
5709 public: | |
5710 CMSRefProcTaskProxy(ProcessTask& task, | |
5711 CMSCollector* collector, | |
5712 const MemRegion& span, | |
5713 CMSBitMap* mark_bit_map, | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5714 AbstractWorkGang* workers, |
0 | 5715 OopTaskQueueSet* task_queues): |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5716 AbstractGangTaskWOopQueues("Process referents by policy in parallel", |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5717 task_queues), |
0 | 5718 _task(task), |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5719 _collector(collector), _span(span), _mark_bit_map(mark_bit_map) |
143
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
5720 { |
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
5721 assert(_collector->_span.equals(_span) && !_span.is_empty(), |
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
5722 "Inconsistency in _span"); |
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
5723 } |
0 | 5724 |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5725 OopTaskQueueSet* task_queues() { return queues(); } |
0 | 5726 |
5727 OopTaskQueue* work_queue(int i) { return task_queues()->queue(i); } | |
5728 | |
5729 void do_work_steal(int i, | |
5730 CMSParDrainMarkingStackClosure* drain, | |
5731 CMSParKeepAliveClosure* keep_alive, | |
5732 int* seed); | |
5733 | |
5734 virtual void work(int i); | |
5735 }; | |
5736 | |
5737 void CMSRefProcTaskProxy::work(int i) { | |
143
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
5738 assert(_collector->_span.equals(_span), "Inconsistency in _span"); |
0 | 5739 CMSParKeepAliveClosure par_keep_alive(_collector, _span, |
935 | 5740 _mark_bit_map, |
5741 &_collector->_revisitStack, | |
5742 work_queue(i)); | |
0 | 5743 CMSParDrainMarkingStackClosure par_drain_stack(_collector, _span, |
935 | 5744 _mark_bit_map, |
5745 &_collector->_revisitStack, | |
5746 work_queue(i)); | |
143
b5489bb705c9
6662086: 6u4+, 7b11+: CMS never clears referents when -XX:+ParallelRefProcEnabled
ysr
parents:
113
diff
changeset
|
5747 CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map); |
0 | 5748 _task.work(i, is_alive_closure, par_keep_alive, par_drain_stack); |
5749 if (_task.marks_oops_alive()) { | |
5750 do_work_steal(i, &par_drain_stack, &par_keep_alive, | |
5751 _collector->hash_seed(i)); | |
5752 } | |
5753 assert(work_queue(i)->size() == 0, "work_queue should be empty"); | |
5754 assert(_collector->_overflow_list == NULL, "non-empty _overflow_list"); | |
5755 } | |
5756 | |
5757 class CMSRefEnqueueTaskProxy: public AbstractGangTask { | |
5758 typedef AbstractRefProcTaskExecutor::EnqueueTask EnqueueTask; | |
5759 EnqueueTask& _task; | |
5760 | |
5761 public: | |
5762 CMSRefEnqueueTaskProxy(EnqueueTask& task) | |
5763 : AbstractGangTask("Enqueue reference objects in parallel"), | |
5764 _task(task) | |
5765 { } | |
5766 | |
5767 virtual void work(int i) | |
5768 { | |
5769 _task.work(i); | |
5770 } | |
5771 }; | |
5772 | |
5773 CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector, | |
935 | 5774 MemRegion span, CMSBitMap* bit_map, CMSMarkStack* revisit_stack, |
5775 OopTaskQueue* work_queue): | |
5776 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), | |
0 | 5777 _span(span), |
5778 _bit_map(bit_map), | |
5779 _work_queue(work_queue), | |
935 | 5780 _mark_and_push(collector, span, bit_map, revisit_stack, work_queue), |
0 | 5781 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), |
5782 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))) | |
5783 { } | |
5784 | |
5785 // . see if we can share work_queues with ParNew? XXX | |
5786 void CMSRefProcTaskProxy::do_work_steal(int i, | |
5787 CMSParDrainMarkingStackClosure* drain, | |
5788 CMSParKeepAliveClosure* keep_alive, | |
5789 int* seed) { | |
5790 OopTaskQueue* work_q = work_queue(i); | |
5791 NOT_PRODUCT(int num_steals = 0;) | |
5792 oop obj_to_scan; | |
5793 | |
5794 while (true) { | |
5795 // Completely finish any left over work from (an) earlier round(s) | |
5796 drain->trim_queue(0); | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
628
diff
changeset
|
5797 size_t num_from_overflow_list = MIN2((size_t)(work_q->max_elems() - work_q->size())/4, |
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
628
diff
changeset
|
5798 (size_t)ParGCDesiredObjsFromOverflowList); |
0 | 5799 // Now check if there's any work in the overflow list |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5800 // Passing ParallelGCThreads as the third parameter, no_of_gc_threads, |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5801 // only affects the number of attempts made to get work from the |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5802 // overflow list and does not affect the number of workers. Just |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5803 // pass ParallelGCThreads so this behavior is unchanged. |
0 | 5804 if (_collector->par_take_from_overflow_list(num_from_overflow_list, |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5805 work_q, |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5806 ParallelGCThreads)) { |
0 | 5807 // Found something in global overflow list; |
5808 // not yet ready to go stealing work from others. | |
5809 // We'd like to assert(work_q->size() != 0, ...) | |
5810 // because we just took work from the overflow list, | |
5811 // but of course we can't, since all of that might have | |
5812 // been already stolen from us. | |
5813 continue; | |
5814 } | |
5815 // Verify that we have no work before we resort to stealing | |
5816 assert(work_q->size() == 0, "Have work, shouldn't steal"); | |
5817 // Try to steal from other queues that have work | |
5818 if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) { | |
5819 NOT_PRODUCT(num_steals++;) | |
5820 assert(obj_to_scan->is_oop(), "Oops, not an oop!"); | |
5821 assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?"); | |
5822 // Do scanning work | |
5823 obj_to_scan->oop_iterate(keep_alive); | |
5824 // Loop around, finish this work, and try to steal some more | |
5825 } else if (terminator()->offer_termination()) { | |
5826 break; // nirvana from the infinite cycle | |
5827 } | |
5828 } | |
5829 NOT_PRODUCT( | |
5830 if (PrintCMSStatistics != 0) { | |
5831 gclog_or_tty->print("\n\t(%d: stole %d oops)", i, num_steals); | |
5832 } | |
5833 ) | |
5834 } | |
5835 | |
5836 void CMSRefProcTaskExecutor::execute(ProcessTask& task) | |
5837 { | |
5838 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5839 FlexibleWorkGang* workers = gch->workers(); |
0 | 5840 assert(workers != NULL, "Need parallel worker threads."); |
5841 CMSRefProcTaskProxy rp_task(task, &_collector, | |
5842 _collector.ref_processor()->span(), | |
5843 _collector.markBitMap(), | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5844 workers, _collector.task_queues()); |
0 | 5845 workers->run_task(&rp_task); |
5846 } | |
5847 | |
5848 void CMSRefProcTaskExecutor::execute(EnqueueTask& task) | |
5849 { | |
5850 | |
5851 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5852 FlexibleWorkGang* workers = gch->workers(); |
0 | 5853 assert(workers != NULL, "Need parallel worker threads."); |
5854 CMSRefEnqueueTaskProxy enq_task(task); | |
5855 workers->run_task(&enq_task); | |
5856 } | |
5857 | |
5858 void CMSCollector::refProcessingWork(bool asynch, bool clear_all_soft_refs) { | |
5859 | |
5860 ResourceMark rm; | |
5861 HandleMark hm; | |
5862 | |
5863 ReferenceProcessor* rp = ref_processor(); | |
5864 assert(rp->span().equals(_span), "Spans should be equal"); | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
5865 assert(!rp->enqueuing_is_done(), "Enqueuing should not be complete"); |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
5866 // Process weak references. |
457
27a80744a83b
6778647: snap(), snap_policy() should be renamed setup(), setup_policy()
ysr
parents:
453
diff
changeset
|
5867 rp->setup_policy(clear_all_soft_refs); |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
5868 verify_work_stacks_empty(); |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
5869 |
0 | 5870 CMSKeepAliveClosure cmsKeepAliveClosure(this, _span, &_markBitMap, |
935 | 5871 &_markStack, &_revisitStack, |
5872 false /* !preclean */); | |
0 | 5873 CMSDrainMarkingStackClosure cmsDrainMarkingStackClosure(this, |
5874 _span, &_markBitMap, &_markStack, | |
452
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
5875 &cmsKeepAliveClosure, false /* !preclean */); |
0 | 5876 { |
5877 TraceTime t("weak refs processing", PrintGCDetails, false, gclog_or_tty); | |
5878 if (rp->processing_is_mt()) { | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5879 // Set the degree of MT here. If the discovery is done MT, there |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5880 // may have been a different number of threads doing the discovery |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5881 // and a different number of discovered lists may have Ref objects. |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5882 // That is OK as long as the Reference lists are balanced (see |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5883 // balance_all_queues() and balance_queues()). |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5884 |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5885 |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5886 rp->set_mt_degree(ParallelGCThreads); |
0 | 5887 CMSRefProcTaskExecutor task_executor(*this); |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
5888 rp->process_discovered_references(&_is_alive_closure, |
0 | 5889 &cmsKeepAliveClosure, |
5890 &cmsDrainMarkingStackClosure, | |
5891 &task_executor); | |
5892 } else { | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
5893 rp->process_discovered_references(&_is_alive_closure, |
0 | 5894 &cmsKeepAliveClosure, |
5895 &cmsDrainMarkingStackClosure, | |
5896 NULL); | |
5897 } | |
5898 verify_work_stacks_empty(); | |
5899 } | |
5900 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
5901 if (should_unload_classes()) { |
0 | 5902 { |
5903 TraceTime t("class unloading", PrintGCDetails, false, gclog_or_tty); | |
5904 | |
5905 // Follow SystemDictionary roots and unload classes | |
5906 bool purged_class = SystemDictionary::do_unloading(&_is_alive_closure); | |
5907 | |
5908 // Follow CodeCache roots and unload any methods marked for unloading | |
5909 CodeCache::do_unloading(&_is_alive_closure, | |
5910 &cmsKeepAliveClosure, | |
5911 purged_class); | |
5912 | |
5913 cmsDrainMarkingStackClosure.do_void(); | |
5914 verify_work_stacks_empty(); | |
5915 | |
5916 // Update subklass/sibling/implementor links in KlassKlass descendants | |
5917 assert(!_revisitStack.isEmpty(), "revisit stack should not be empty"); | |
5918 oop k; | |
5919 while ((k = _revisitStack.pop()) != NULL) { | |
5920 ((Klass*)(oopDesc*)k)->follow_weak_klass_links( | |
5921 &_is_alive_closure, | |
5922 &cmsKeepAliveClosure); | |
5923 } | |
5924 assert(!ClassUnloading || | |
5925 (_markStack.isEmpty() && overflow_list_is_empty()), | |
5926 "Should not have found new reachable objects"); | |
5927 assert(_revisitStack.isEmpty(), "revisit stack should have been drained"); | |
5928 cmsDrainMarkingStackClosure.do_void(); | |
5929 verify_work_stacks_empty(); | |
5930 } | |
5931 | |
5932 { | |
5933 TraceTime t("scrub symbol & string tables", PrintGCDetails, false, gclog_or_tty); | |
5934 // Now clean up stale oops in SymbolTable and StringTable | |
5935 SymbolTable::unlink(&_is_alive_closure); | |
5936 StringTable::unlink(&_is_alive_closure); | |
5937 } | |
5938 } | |
5939 | |
5940 verify_work_stacks_empty(); | |
5941 // Restore any preserved marks as a result of mark stack or | |
5942 // work queue overflow | |
5943 restore_preserved_marks_if_any(); // done single-threaded for now | |
5944 | |
5945 rp->set_enqueuing_is_done(true); | |
5946 if (rp->processing_is_mt()) { | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
5947 rp->balance_all_queues(); |
0 | 5948 CMSRefProcTaskExecutor task_executor(*this); |
5949 rp->enqueue_discovered_references(&task_executor); | |
5950 } else { | |
5951 rp->enqueue_discovered_references(NULL); | |
5952 } | |
5953 rp->verify_no_references_recorded(); | |
5954 assert(!rp->discovery_enabled(), "should have been disabled"); | |
5955 | |
5956 // JVMTI object tagging is based on JNI weak refs. If any of these | |
5957 // refs were cleared then JVMTI needs to update its maps and | |
5958 // maybe post ObjectFrees to agents. | |
5959 JvmtiExport::cms_ref_processing_epilogue(); | |
5960 } | |
5961 | |
5962 #ifndef PRODUCT | |
5963 void CMSCollector::check_correct_thread_executing() { | |
5964 Thread* t = Thread::current(); | |
5965 // Only the VM thread or the CMS thread should be here. | |
5966 assert(t->is_ConcurrentGC_thread() || t->is_VM_thread(), | |
5967 "Unexpected thread type"); | |
5968 // If this is the vm thread, the foreground process | |
5969 // should not be waiting. Note that _foregroundGCIsActive is | |
5970 // true while the foreground collector is waiting. | |
5971 if (_foregroundGCShouldWait) { | |
5972 // We cannot be the VM thread | |
5973 assert(t->is_ConcurrentGC_thread(), | |
5974 "Should be CMS thread"); | |
5975 } else { | |
5976 // We can be the CMS thread only if we are in a stop-world | |
5977 // phase of CMS collection. | |
5978 if (t->is_ConcurrentGC_thread()) { | |
5979 assert(_collectorState == InitialMarking || | |
5980 _collectorState == FinalMarking, | |
5981 "Should be a stop-world phase"); | |
5982 // The CMS thread should be holding the CMS_token. | |
5983 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
5984 "Potential interference with concurrently " | |
5985 "executing VM thread"); | |
5986 } | |
5987 } | |
5988 } | |
5989 #endif | |
5990 | |
5991 void CMSCollector::sweep(bool asynch) { | |
5992 assert(_collectorState == Sweeping, "just checking"); | |
5993 check_correct_thread_executing(); | |
5994 verify_work_stacks_empty(); | |
5995 verify_overflow_empty(); | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
5996 increment_sweep_count(); |
1703
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
5997 TraceCMSMemoryManagerStats tms(_collectorState); |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
5998 |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
5999 _inter_sweep_timer.stop(); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6000 _inter_sweep_estimate.sample(_inter_sweep_timer.seconds()); |
0 | 6001 size_policy()->avg_cms_free_at_sweep()->sample(_cmsGen->free()); |
6002 | |
6003 // PermGen verification support: If perm gen sweeping is disabled in | |
6004 // this cycle, we preserve the perm gen object "deadness" information | |
6005 // in the perm_gen_verify_bit_map. In order to do that we traverse | |
6006 // all blocks in perm gen and mark all dead objects. | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
6007 if (verifying() && !should_unload_classes()) { |
0 | 6008 assert(perm_gen_verify_bit_map()->sizeInBits() != 0, |
6009 "Should have already been allocated"); | |
6010 MarkDeadObjectsClosure mdo(this, _permGen->cmsSpace(), | |
6011 markBitMap(), perm_gen_verify_bit_map()); | |
7
2faf283ce688
6621144: CMS: assertion failure "is_cms_thread == Thread::current()->is_ConcurrentGC_thread()"
ysr
parents:
0
diff
changeset
|
6012 if (asynch) { |
2faf283ce688
6621144: CMS: assertion failure "is_cms_thread == Thread::current()->is_ConcurrentGC_thread()"
ysr
parents:
0
diff
changeset
|
6013 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), |
2faf283ce688
6621144: CMS: assertion failure "is_cms_thread == Thread::current()->is_ConcurrentGC_thread()"
ysr
parents:
0
diff
changeset
|
6014 bitMapLock()); |
2faf283ce688
6621144: CMS: assertion failure "is_cms_thread == Thread::current()->is_ConcurrentGC_thread()"
ysr
parents:
0
diff
changeset
|
6015 _permGen->cmsSpace()->blk_iterate(&mdo); |
2faf283ce688
6621144: CMS: assertion failure "is_cms_thread == Thread::current()->is_ConcurrentGC_thread()"
ysr
parents:
0
diff
changeset
|
6016 } else { |
2faf283ce688
6621144: CMS: assertion failure "is_cms_thread == Thread::current()->is_ConcurrentGC_thread()"
ysr
parents:
0
diff
changeset
|
6017 // In the case of synchronous sweep, we already have |
2faf283ce688
6621144: CMS: assertion failure "is_cms_thread == Thread::current()->is_ConcurrentGC_thread()"
ysr
parents:
0
diff
changeset
|
6018 // the requisite locks/tokens. |
2faf283ce688
6621144: CMS: assertion failure "is_cms_thread == Thread::current()->is_ConcurrentGC_thread()"
ysr
parents:
0
diff
changeset
|
6019 _permGen->cmsSpace()->blk_iterate(&mdo); |
2faf283ce688
6621144: CMS: assertion failure "is_cms_thread == Thread::current()->is_ConcurrentGC_thread()"
ysr
parents:
0
diff
changeset
|
6020 } |
0 | 6021 } |
6022 | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6023 assert(!_intra_sweep_timer.is_active(), "Should not be active"); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6024 _intra_sweep_timer.reset(); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6025 _intra_sweep_timer.start(); |
0 | 6026 if (asynch) { |
6027 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); | |
6028 CMSPhaseAccounting pa(this, "sweep", !PrintGCDetails); | |
6029 // First sweep the old gen then the perm gen | |
6030 { | |
6031 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), | |
6032 bitMapLock()); | |
6033 sweepWork(_cmsGen, asynch); | |
6034 } | |
6035 | |
6036 // Now repeat for perm gen | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
6037 if (should_unload_classes()) { |
0 | 6038 CMSTokenSyncWithLocks ts(true, _permGen->freelistLock(), |
6039 bitMapLock()); | |
6040 sweepWork(_permGen, asynch); | |
6041 } | |
6042 | |
6043 // Update Universe::_heap_*_at_gc figures. | |
6044 // We need all the free list locks to make the abstract state | |
6045 // transition from Sweeping to Resetting. See detailed note | |
6046 // further below. | |
6047 { | |
6048 CMSTokenSyncWithLocks ts(true, _cmsGen->freelistLock(), | |
6049 _permGen->freelistLock()); | |
6050 // Update heap occupancy information which is used as | |
6051 // input to soft ref clearing policy at the next gc. | |
6052 Universe::update_heap_info_at_gc(); | |
6053 _collectorState = Resizing; | |
6054 } | |
6055 } else { | |
6056 // already have needed locks | |
6057 sweepWork(_cmsGen, asynch); | |
6058 | |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
6059 if (should_unload_classes()) { |
0 | 6060 sweepWork(_permGen, asynch); |
6061 } | |
6062 // Update heap occupancy information which is used as | |
6063 // input to soft ref clearing policy at the next gc. | |
6064 Universe::update_heap_info_at_gc(); | |
6065 _collectorState = Resizing; | |
6066 } | |
6067 verify_work_stacks_empty(); | |
6068 verify_overflow_empty(); | |
6069 | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6070 _intra_sweep_timer.stop(); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6071 _intra_sweep_estimate.sample(_intra_sweep_timer.seconds()); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6072 |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6073 _inter_sweep_timer.reset(); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6074 _inter_sweep_timer.start(); |
0 | 6075 |
6076 update_time_of_last_gc(os::javaTimeMillis()); | |
6077 | |
6078 // NOTE on abstract state transitions: | |
6079 // Mutators allocate-live and/or mark the mod-union table dirty | |
6080 // based on the state of the collection. The former is done in | |
6081 // the interval [Marking, Sweeping] and the latter in the interval | |
6082 // [Marking, Sweeping). Thus the transitions into the Marking state | |
6083 // and out of the Sweeping state must be synchronously visible | |
6084 // globally to the mutators. | |
6085 // The transition into the Marking state happens with the world | |
6086 // stopped so the mutators will globally see it. Sweeping is | |
6087 // done asynchronously by the background collector so the transition | |
6088 // from the Sweeping state to the Resizing state must be done | |
6089 // under the freelistLock (as is the check for whether to | |
6090 // allocate-live and whether to dirty the mod-union table). | |
6091 assert(_collectorState == Resizing, "Change of collector state to" | |
6092 " Resizing must be done under the freelistLocks (plural)"); | |
6093 | |
6094 // Now that sweeping has been completed, if the GCH's | |
6095 // incremental_collection_will_fail flag is set, clear it, | |
6096 // thus inviting a younger gen collection to promote into | |
6097 // this generation. If such a promotion may still fail, | |
6098 // the flag will be set again when a young collection is | |
6099 // attempted. | |
6100 // I think the incremental_collection_will_fail flag's use | |
6101 // is specific to a 2 generation collection policy, so i'll | |
6102 // assert that that's the configuration we are operating within. | |
6103 // The use of the flag can and should be generalized appropriately | |
6104 // in the future to deal with a general n-generation system. | |
6105 | |
6106 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
6107 assert(gch->collector_policy()->is_two_generation_policy(), | |
6108 "Resetting of incremental_collection_will_fail flag" | |
6109 " may be incorrect otherwise"); | |
6110 gch->clear_incremental_collection_will_fail(); | |
6111 gch->update_full_collections_completed(_collection_count_start); | |
6112 } | |
6113 | |
6114 // FIX ME!!! Looks like this belongs in CFLSpace, with | |
6115 // CMSGen merely delegating to it. | |
6116 void ConcurrentMarkSweepGeneration::setNearLargestChunk() { | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6117 double nearLargestPercent = FLSLargestBlockCoalesceProximity; |
0 | 6118 HeapWord* minAddr = _cmsSpace->bottom(); |
6119 HeapWord* largestAddr = | |
6120 (HeapWord*) _cmsSpace->dictionary()->findLargestDict(); | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6121 if (largestAddr == NULL) { |
0 | 6122 // The dictionary appears to be empty. In this case |
6123 // try to coalesce at the end of the heap. | |
6124 largestAddr = _cmsSpace->end(); | |
6125 } | |
6126 size_t largestOffset = pointer_delta(largestAddr, minAddr); | |
6127 size_t nearLargestOffset = | |
6128 (size_t)((double)largestOffset * nearLargestPercent) - MinChunkSize; | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6129 if (PrintFLSStatistics != 0) { |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6130 gclog_or_tty->print_cr( |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6131 "CMS: Large Block: " PTR_FORMAT ";" |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6132 " Proximity: " PTR_FORMAT " -> " PTR_FORMAT, |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6133 largestAddr, |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6134 _cmsSpace->nearLargestChunk(), minAddr + nearLargestOffset); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6135 } |
0 | 6136 _cmsSpace->set_nearLargestChunk(minAddr + nearLargestOffset); |
6137 } | |
6138 | |
6139 bool ConcurrentMarkSweepGeneration::isNearLargestChunk(HeapWord* addr) { | |
6140 return addr >= _cmsSpace->nearLargestChunk(); | |
6141 } | |
6142 | |
6143 FreeChunk* ConcurrentMarkSweepGeneration::find_chunk_at_end() { | |
6144 return _cmsSpace->find_chunk_at_end(); | |
6145 } | |
6146 | |
6147 void ConcurrentMarkSweepGeneration::update_gc_stats(int current_level, | |
6148 bool full) { | |
6149 // The next lower level has been collected. Gather any statistics | |
6150 // that are of interest at this point. | |
6151 if (!full && (current_level + 1) == level()) { | |
6152 // Gather statistics on the young generation collection. | |
6153 collector()->stats().record_gc0_end(used()); | |
6154 } | |
6155 } | |
6156 | |
6157 CMSAdaptiveSizePolicy* ConcurrentMarkSweepGeneration::size_policy() { | |
6158 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
6159 assert(gch->kind() == CollectedHeap::GenCollectedHeap, | |
6160 "Wrong type of heap"); | |
6161 CMSAdaptiveSizePolicy* sp = (CMSAdaptiveSizePolicy*) | |
6162 gch->gen_policy()->size_policy(); | |
6163 assert(sp->is_gc_cms_adaptive_size_policy(), | |
6164 "Wrong type of size policy"); | |
6165 return sp; | |
6166 } | |
6167 | |
6168 void ConcurrentMarkSweepGeneration::rotate_debug_collection_type() { | |
6169 if (PrintGCDetails && Verbose) { | |
6170 gclog_or_tty->print("Rotate from %d ", _debug_collection_type); | |
6171 } | |
6172 _debug_collection_type = (CollectionTypes) (_debug_collection_type + 1); | |
6173 _debug_collection_type = | |
6174 (CollectionTypes) (_debug_collection_type % Unknown_collection_type); | |
6175 if (PrintGCDetails && Verbose) { | |
6176 gclog_or_tty->print_cr("to %d ", _debug_collection_type); | |
6177 } | |
6178 } | |
6179 | |
6180 void CMSCollector::sweepWork(ConcurrentMarkSweepGeneration* gen, | |
6181 bool asynch) { | |
6182 // We iterate over the space(s) underlying this generation, | |
6183 // checking the mark bit map to see if the bits corresponding | |
6184 // to specific blocks are marked or not. Blocks that are | |
6185 // marked are live and are not swept up. All remaining blocks | |
6186 // are swept up, with coalescing on-the-fly as we sweep up | |
6187 // contiguous free and/or garbage blocks: | |
6188 // We need to ensure that the sweeper synchronizes with allocators | |
6189 // and stop-the-world collectors. In particular, the following | |
6190 // locks are used: | |
6191 // . CMS token: if this is held, a stop the world collection cannot occur | |
6192 // . freelistLock: if this is held no allocation can occur from this | |
6193 // generation by another thread | |
6194 // . bitMapLock: if this is held, no other thread can access or update | |
6195 // | |
6196 | |
6197 // Note that we need to hold the freelistLock if we use | |
6198 // block iterate below; else the iterator might go awry if | |
6199 // a mutator (or promotion) causes block contents to change | |
6200 // (for instance if the allocator divvies up a block). | |
6201 // If we hold the free list lock, for all practical purposes | |
6202 // young generation GC's can't occur (they'll usually need to | |
6203 // promote), so we might as well prevent all young generation | |
6204 // GC's while we do a sweeping step. For the same reason, we might | |
6205 // as well take the bit map lock for the entire duration | |
6206 | |
6207 // check that we hold the requisite locks | |
6208 assert(have_cms_token(), "Should hold cms token"); | |
6209 assert( (asynch && ConcurrentMarkSweepThread::cms_thread_has_cms_token()) | |
6210 || (!asynch && ConcurrentMarkSweepThread::vm_thread_has_cms_token()), | |
6211 "Should possess CMS token to sweep"); | |
6212 assert_lock_strong(gen->freelistLock()); | |
6213 assert_lock_strong(bitMapLock()); | |
6214 | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6215 assert(!_inter_sweep_timer.is_active(), "Was switched off in an outer context"); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6216 assert(_intra_sweep_timer.is_active(), "Was switched on in an outer context"); |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6217 gen->cmsSpace()->beginSweepFLCensus((float)(_inter_sweep_timer.seconds()), |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6218 _inter_sweep_estimate.padded_average(), |
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6219 _intra_sweep_estimate.padded_average()); |
0 | 6220 gen->setNearLargestChunk(); |
6221 | |
6222 { | |
6223 SweepClosure sweepClosure(this, gen, &_markBitMap, | |
6224 CMSYield && asynch); | |
6225 gen->cmsSpace()->blk_iterate_careful(&sweepClosure); | |
6226 // We need to free-up/coalesce garbage/blocks from a | |
6227 // co-terminal free run. This is done in the SweepClosure | |
6228 // destructor; so, do not remove this scope, else the | |
6229 // end-of-sweep-census below will be off by a little bit. | |
6230 } | |
6231 gen->cmsSpace()->sweep_completed(); | |
1145
e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
ysr
parents:
1085
diff
changeset
|
6232 gen->cmsSpace()->endSweepFLCensus(sweep_count()); |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
6233 if (should_unload_classes()) { // unloaded classes this cycle, |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
6234 _concurrent_cycles_since_last_unload = 0; // ... reset count |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
6235 } else { // did not unload classes, |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
6236 _concurrent_cycles_since_last_unload++; // ... increment count |
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
6237 } |
0 | 6238 } |
6239 | |
6240 // Reset CMS data structures (for now just the marking bit map) | |
6241 // preparatory for the next cycle. | |
6242 void CMSCollector::reset(bool asynch) { | |
6243 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
6244 CMSAdaptiveSizePolicy* sp = size_policy(); | |
6245 AdaptiveSizePolicyOutput(sp, gch->total_collections()); | |
6246 if (asynch) { | |
6247 CMSTokenSyncWithLocks ts(true, bitMapLock()); | |
6248 | |
6249 // If the state is not "Resetting", the foreground thread | |
6250 // has done a collection and the resetting. | |
6251 if (_collectorState != Resetting) { | |
6252 assert(_collectorState == Idling, "The state should only change" | |
6253 " because the foreground collector has finished the collection"); | |
6254 return; | |
6255 } | |
6256 | |
6257 // Clear the mark bitmap (no grey objects to start with) | |
6258 // for the next cycle. | |
6259 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); | |
6260 CMSPhaseAccounting cmspa(this, "reset", !PrintGCDetails); | |
6261 | |
6262 HeapWord* curAddr = _markBitMap.startWord(); | |
6263 while (curAddr < _markBitMap.endWord()) { | |
6264 size_t remaining = pointer_delta(_markBitMap.endWord(), curAddr); | |
6265 MemRegion chunk(curAddr, MIN2(CMSBitMapYieldQuantum, remaining)); | |
6266 _markBitMap.clear_large_range(chunk); | |
6267 if (ConcurrentMarkSweepThread::should_yield() && | |
6268 !foregroundGCIsActive() && | |
6269 CMSYield) { | |
6270 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
6271 "CMS thread should hold CMS token"); | |
6272 assert_lock_strong(bitMapLock()); | |
6273 bitMapLock()->unlock(); | |
6274 ConcurrentMarkSweepThread::desynchronize(true); | |
6275 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
6276 stopTimer(); | |
6277 if (PrintCMSStatistics != 0) { | |
6278 incrementYields(); | |
6279 } | |
6280 icms_wait(); | |
6281 | |
6282 // See the comment in coordinator_yield() | |
6283 for (unsigned i = 0; i < CMSYieldSleepCount && | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6284 ConcurrentMarkSweepThread::should_yield() && |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6285 !CMSCollector::foregroundGCIsActive(); ++i) { |
0 | 6286 os::sleep(Thread::current(), 1, false); |
6287 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
6288 } | |
6289 | |
6290 ConcurrentMarkSweepThread::synchronize(true); | |
6291 bitMapLock()->lock_without_safepoint_check(); | |
6292 startTimer(); | |
6293 } | |
6294 curAddr = chunk.end(); | |
6295 } | |
1387
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
6296 // A successful mostly concurrent collection has been done. |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
6297 // Because only the full (i.e., concurrent mode failure) collections |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
6298 // are being measured for gc overhead limits, clean the "near" flag |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
6299 // and count. |
0bfd3fb24150
6858496: Clear all SoftReferences before an out-of-memory due to GC overhead limit.
jmasa
parents:
1289
diff
changeset
|
6300 sp->reset_gc_overhead_limit_count(); |
0 | 6301 _collectorState = Idling; |
6302 } else { | |
6303 // already have the lock | |
6304 assert(_collectorState == Resetting, "just checking"); | |
6305 assert_lock_strong(bitMapLock()); | |
6306 _markBitMap.clear_all(); | |
6307 _collectorState = Idling; | |
6308 } | |
6309 | |
6310 // Stop incremental mode after a cycle completes, so that any future cycles | |
6311 // are triggered by allocation. | |
6312 stop_icms(); | |
6313 | |
6314 NOT_PRODUCT( | |
6315 if (RotateCMSCollectionTypes) { | |
6316 _cmsGen->rotate_debug_collection_type(); | |
6317 } | |
6318 ) | |
6319 } | |
6320 | |
6321 void CMSCollector::do_CMS_operation(CMS_op_type op) { | |
6322 gclog_or_tty->date_stamp(PrintGC && PrintGCDateStamps); | |
6323 TraceCPUTime tcpu(PrintGCDetails, true, gclog_or_tty); | |
6324 TraceTime t("GC", PrintGC, !PrintGCDetails, gclog_or_tty); | |
6325 TraceCollectorStats tcs(counters()); | |
6326 | |
6327 switch (op) { | |
6328 case CMS_op_checkpointRootsInitial: { | |
6329 checkpointRootsInitial(true); // asynch | |
6330 if (PrintGC) { | |
6331 _cmsGen->printOccupancy("initial-mark"); | |
6332 } | |
6333 break; | |
6334 } | |
6335 case CMS_op_checkpointRootsFinal: { | |
6336 checkpointRootsFinal(true, // asynch | |
6337 false, // !clear_all_soft_refs | |
6338 false); // !init_mark_was_synchronous | |
6339 if (PrintGC) { | |
6340 _cmsGen->printOccupancy("remark"); | |
6341 } | |
6342 break; | |
6343 } | |
6344 default: | |
6345 fatal("No such CMS_op"); | |
6346 } | |
6347 } | |
6348 | |
6349 #ifndef PRODUCT | |
6350 size_t const CMSCollector::skip_header_HeapWords() { | |
6351 return FreeChunk::header_size(); | |
6352 } | |
6353 | |
6354 // Try and collect here conditions that should hold when | |
6355 // CMS thread is exiting. The idea is that the foreground GC | |
6356 // thread should not be blocked if it wants to terminate | |
6357 // the CMS thread and yet continue to run the VM for a while | |
6358 // after that. | |
6359 void CMSCollector::verify_ok_to_terminate() const { | |
6360 assert(Thread::current()->is_ConcurrentGC_thread(), | |
6361 "should be called by CMS thread"); | |
6362 assert(!_foregroundGCShouldWait, "should be false"); | |
6363 // We could check here that all the various low-level locks | |
6364 // are not held by the CMS thread, but that is overkill; see | |
6365 // also CMSThread::verify_ok_to_terminate() where the CGC_lock | |
6366 // is checked. | |
6367 } | |
6368 #endif | |
6369 | |
6370 size_t CMSCollector::block_size_using_printezis_bits(HeapWord* addr) const { | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
6371 assert(_markBitMap.isMarked(addr) && _markBitMap.isMarked(addr + 1), |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
452
diff
changeset
|
6372 "missing Printezis mark?"); |
0 | 6373 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); |
6374 size_t size = pointer_delta(nextOneAddr + 1, addr); | |
6375 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), | |
6376 "alignment problem"); | |
6377 assert(size >= 3, "Necessary for Printezis marks to work"); | |
6378 return size; | |
6379 } | |
6380 | |
6381 // A variant of the above (block_size_using_printezis_bits()) except | |
6382 // that we return 0 if the P-bits are not yet set. | |
6383 size_t CMSCollector::block_size_if_printezis_bits(HeapWord* addr) const { | |
6384 if (_markBitMap.isMarked(addr)) { | |
6385 assert(_markBitMap.isMarked(addr + 1), "Missing Printezis bit?"); | |
6386 HeapWord* nextOneAddr = _markBitMap.getNextMarkedWordAddress(addr + 2); | |
6387 size_t size = pointer_delta(nextOneAddr + 1, addr); | |
6388 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), | |
6389 "alignment problem"); | |
6390 assert(size >= 3, "Necessary for Printezis marks to work"); | |
6391 return size; | |
6392 } else { | |
6393 assert(!_markBitMap.isMarked(addr + 1), "Bit map inconsistency?"); | |
6394 return 0; | |
6395 } | |
6396 } | |
6397 | |
6398 HeapWord* CMSCollector::next_card_start_after_block(HeapWord* addr) const { | |
6399 size_t sz = 0; | |
6400 oop p = (oop)addr; | |
187 | 6401 if (p->klass_or_null() != NULL && p->is_parsable()) { |
0 | 6402 sz = CompactibleFreeListSpace::adjustObjectSize(p->size()); |
6403 } else { | |
6404 sz = block_size_using_printezis_bits(addr); | |
6405 } | |
6406 assert(sz > 0, "size must be nonzero"); | |
6407 HeapWord* next_block = addr + sz; | |
6408 HeapWord* next_card = (HeapWord*)round_to((uintptr_t)next_block, | |
6409 CardTableModRefBS::card_size); | |
6410 assert(round_down((uintptr_t)addr, CardTableModRefBS::card_size) < | |
6411 round_down((uintptr_t)next_card, CardTableModRefBS::card_size), | |
6412 "must be different cards"); | |
6413 return next_card; | |
6414 } | |
6415 | |
6416 | |
6417 // CMS Bit Map Wrapper ///////////////////////////////////////// | |
6418 | |
6419 // Construct a CMS bit map infrastructure, but don't create the | |
6420 // bit vector itself. That is done by a separate call CMSBitMap::allocate() | |
6421 // further below. | |
6422 CMSBitMap::CMSBitMap(int shifter, int mutex_rank, const char* mutex_name): | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
6423 _bm(), |
0 | 6424 _shifter(shifter), |
6425 _lock(mutex_rank >= 0 ? new Mutex(mutex_rank, mutex_name, true) : NULL) | |
6426 { | |
6427 _bmStartWord = 0; | |
6428 _bmWordSize = 0; | |
6429 } | |
6430 | |
6431 bool CMSBitMap::allocate(MemRegion mr) { | |
6432 _bmStartWord = mr.start(); | |
6433 _bmWordSize = mr.word_size(); | |
6434 ReservedSpace brs(ReservedSpace::allocation_align_size_up( | |
6435 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); | |
6436 if (!brs.is_reserved()) { | |
6437 warning("CMS bit map allocation failure"); | |
6438 return false; | |
6439 } | |
6440 // For now we'll just commit all of the bit map up fromt. | |
6441 // Later on we'll try to be more parsimonious with swap. | |
6442 if (!_virtual_space.initialize(brs, brs.size())) { | |
6443 warning("CMS bit map backing store failure"); | |
6444 return false; | |
6445 } | |
6446 assert(_virtual_space.committed_size() == brs.size(), | |
6447 "didn't reserve backing store for all of CMS bit map?"); | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
6448 _bm.set_map((BitMap::bm_word_t*)_virtual_space.low()); |
0 | 6449 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= |
6450 _bmWordSize, "inconsistency in bit map sizing"); | |
6451 _bm.set_size(_bmWordSize >> _shifter); | |
6452 | |
6453 // bm.clear(); // can we rely on getting zero'd memory? verify below | |
6454 assert(isAllClear(), | |
6455 "Expected zero'd memory from ReservedSpace constructor"); | |
6456 assert(_bm.size() == heapWordDiffToOffsetDiff(sizeInWords()), | |
6457 "consistency check"); | |
6458 return true; | |
6459 } | |
6460 | |
6461 void CMSBitMap::dirty_range_iterate_clear(MemRegion mr, MemRegionClosure* cl) { | |
6462 HeapWord *next_addr, *end_addr, *last_addr; | |
6463 assert_locked(); | |
6464 assert(covers(mr), "out-of-range error"); | |
6465 // XXX assert that start and end are appropriately aligned | |
6466 for (next_addr = mr.start(), end_addr = mr.end(); | |
6467 next_addr < end_addr; next_addr = last_addr) { | |
6468 MemRegion dirty_region = getAndClearMarkedRegion(next_addr, end_addr); | |
6469 last_addr = dirty_region.end(); | |
6470 if (!dirty_region.is_empty()) { | |
6471 cl->do_MemRegion(dirty_region); | |
6472 } else { | |
6473 assert(last_addr == end_addr, "program logic"); | |
6474 return; | |
6475 } | |
6476 } | |
6477 } | |
6478 | |
6479 #ifndef PRODUCT | |
6480 void CMSBitMap::assert_locked() const { | |
6481 CMSLockVerifier::assert_locked(lock()); | |
6482 } | |
6483 | |
6484 bool CMSBitMap::covers(MemRegion mr) const { | |
6485 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); | |
6486 assert((size_t)_bm.size() == (_bmWordSize >> _shifter), | |
6487 "size inconsistency"); | |
6488 return (mr.start() >= _bmStartWord) && | |
6489 (mr.end() <= endWord()); | |
6490 } | |
6491 | |
6492 bool CMSBitMap::covers(HeapWord* start, size_t size) const { | |
6493 return (start >= _bmStartWord && (start + size) <= endWord()); | |
6494 } | |
6495 | |
6496 void CMSBitMap::verifyNoOneBitsInRange(HeapWord* left, HeapWord* right) { | |
6497 // verify that there are no 1 bits in the interval [left, right) | |
6498 FalseBitMapClosure falseBitMapClosure; | |
6499 iterate(&falseBitMapClosure, left, right); | |
6500 } | |
6501 | |
6502 void CMSBitMap::region_invariant(MemRegion mr) | |
6503 { | |
6504 assert_locked(); | |
6505 // mr = mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); | |
6506 assert(!mr.is_empty(), "unexpected empty region"); | |
6507 assert(covers(mr), "mr should be covered by bit map"); | |
6508 // convert address range into offset range | |
6509 size_t start_ofs = heapWordToOffset(mr.start()); | |
6510 // Make sure that end() is appropriately aligned | |
6511 assert(mr.end() == (HeapWord*)round_to((intptr_t)mr.end(), | |
6512 (1 << (_shifter+LogHeapWordSize))), | |
6513 "Misaligned mr.end()"); | |
6514 size_t end_ofs = heapWordToOffset(mr.end()); | |
6515 assert(end_ofs > start_ofs, "Should mark at least one bit"); | |
6516 } | |
6517 | |
6518 #endif | |
6519 | |
6520 bool CMSMarkStack::allocate(size_t size) { | |
6521 // allocate a stack of the requisite depth | |
6522 ReservedSpace rs(ReservedSpace::allocation_align_size_up( | |
6523 size * sizeof(oop))); | |
6524 if (!rs.is_reserved()) { | |
6525 warning("CMSMarkStack allocation failure"); | |
6526 return false; | |
6527 } | |
6528 if (!_virtual_space.initialize(rs, rs.size())) { | |
6529 warning("CMSMarkStack backing store failure"); | |
6530 return false; | |
6531 } | |
6532 assert(_virtual_space.committed_size() == rs.size(), | |
6533 "didn't reserve backing store for all of CMS stack?"); | |
6534 _base = (oop*)(_virtual_space.low()); | |
6535 _index = 0; | |
6536 _capacity = size; | |
6537 NOT_PRODUCT(_max_depth = 0); | |
6538 return true; | |
6539 } | |
6540 | |
6541 // XXX FIX ME !!! In the MT case we come in here holding a | |
6542 // leaf lock. For printing we need to take a further lock | |
6543 // which has lower rank. We need to recallibrate the two | |
6544 // lock-ranks involved in order to be able to rpint the | |
6545 // messages below. (Or defer the printing to the caller. | |
6546 // For now we take the expedient path of just disabling the | |
6547 // messages for the problematic case.) | |
6548 void CMSMarkStack::expand() { | |
1284 | 6549 assert(_capacity <= MarkStackSizeMax, "stack bigger than permitted"); |
6550 if (_capacity == MarkStackSizeMax) { | |
0 | 6551 if (_hit_limit++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { |
6552 // We print a warning message only once per CMS cycle. | |
6553 gclog_or_tty->print_cr(" (benign) Hit CMSMarkStack max size limit"); | |
6554 } | |
6555 return; | |
6556 } | |
6557 // Double capacity if possible | |
1284 | 6558 size_t new_capacity = MIN2(_capacity*2, MarkStackSizeMax); |
0 | 6559 // Do not give up existing stack until we have managed to |
6560 // get the double capacity that we desired. | |
6561 ReservedSpace rs(ReservedSpace::allocation_align_size_up( | |
6562 new_capacity * sizeof(oop))); | |
6563 if (rs.is_reserved()) { | |
6564 // Release the backing store associated with old stack | |
6565 _virtual_space.release(); | |
6566 // Reinitialize virtual space for new stack | |
6567 if (!_virtual_space.initialize(rs, rs.size())) { | |
6568 fatal("Not enough swap for expanded marking stack"); | |
6569 } | |
6570 _base = (oop*)(_virtual_space.low()); | |
6571 _index = 0; | |
6572 _capacity = new_capacity; | |
6573 } else if (_failed_double++ == 0 && !CMSConcurrentMTEnabled && PrintGCDetails) { | |
6574 // Failed to double capacity, continue; | |
6575 // we print a detail message only once per CMS cycle. | |
6576 gclog_or_tty->print(" (benign) Failed to expand marking stack from "SIZE_FORMAT"K to " | |
6577 SIZE_FORMAT"K", | |
6578 _capacity / K, new_capacity / K); | |
6579 } | |
6580 } | |
6581 | |
6582 | |
6583 // Closures | |
6584 // XXX: there seems to be a lot of code duplication here; | |
6585 // should refactor and consolidate common code. | |
6586 | |
6587 // This closure is used to mark refs into the CMS generation in | |
6588 // the CMS bit map. Called at the first checkpoint. This closure | |
6589 // assumes that we do not need to re-mark dirty cards; if the CMS | |
6590 // generation on which this is used is not an oldest (modulo perm gen) | |
6591 // generation then this will lose younger_gen cards! | |
6592 | |
6593 MarkRefsIntoClosure::MarkRefsIntoClosure( | |
994
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
6594 MemRegion span, CMSBitMap* bitMap): |
0 | 6595 _span(span), |
994
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
6596 _bitMap(bitMap) |
0 | 6597 { |
6598 assert(_ref_processor == NULL, "deliberately left NULL"); | |
6599 assert(_bitMap->covers(_span), "_bitMap/_span mismatch"); | |
6600 } | |
6601 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6602 void MarkRefsIntoClosure::do_oop(oop obj) { |
0 | 6603 // if p points into _span, then mark corresponding bit in _markBitMap |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6604 assert(obj->is_oop(), "expected an oop"); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6605 HeapWord* addr = (HeapWord*)obj; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6606 if (_span.contains(addr)) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6607 // this should be made more efficient |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6608 _bitMap->mark(addr); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6609 } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6610 } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6611 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6612 void MarkRefsIntoClosure::do_oop(oop* p) { MarkRefsIntoClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6613 void MarkRefsIntoClosure::do_oop(narrowOop* p) { MarkRefsIntoClosure::do_oop_work(p); } |
0 | 6614 |
6615 // A variant of the above, used for CMS marking verification. | |
6616 MarkRefsIntoVerifyClosure::MarkRefsIntoVerifyClosure( | |
994
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
6617 MemRegion span, CMSBitMap* verification_bm, CMSBitMap* cms_bm): |
0 | 6618 _span(span), |
6619 _verification_bm(verification_bm), | |
994
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
6620 _cms_bm(cms_bm) |
753cf9794df9
6885169: merge of 4957990 and 6863023 causes conflict on do_nmethods
jrose
parents:
993
diff
changeset
|
6621 { |
0 | 6622 assert(_ref_processor == NULL, "deliberately left NULL"); |
6623 assert(_verification_bm->covers(_span), "_verification_bm/_span mismatch"); | |
6624 } | |
6625 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6626 void MarkRefsIntoVerifyClosure::do_oop(oop obj) { |
0 | 6627 // if p points into _span, then mark corresponding bit in _markBitMap |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6628 assert(obj->is_oop(), "expected an oop"); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6629 HeapWord* addr = (HeapWord*)obj; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6630 if (_span.contains(addr)) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6631 _verification_bm->mark(addr); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6632 if (!_cms_bm->isMarked(addr)) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6633 oop(addr)->print(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6634 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", addr); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6635 fatal("... aborting"); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6636 } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6637 } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6638 } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6639 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6640 void MarkRefsIntoVerifyClosure::do_oop(oop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6641 void MarkRefsIntoVerifyClosure::do_oop(narrowOop* p) { MarkRefsIntoVerifyClosure::do_oop_work(p); } |
0 | 6642 |
6643 ////////////////////////////////////////////////// | |
6644 // MarkRefsIntoAndScanClosure | |
6645 ////////////////////////////////////////////////// | |
6646 | |
6647 MarkRefsIntoAndScanClosure::MarkRefsIntoAndScanClosure(MemRegion span, | |
6648 ReferenceProcessor* rp, | |
6649 CMSBitMap* bit_map, | |
6650 CMSBitMap* mod_union_table, | |
6651 CMSMarkStack* mark_stack, | |
6652 CMSMarkStack* revisit_stack, | |
6653 CMSCollector* collector, | |
6654 bool should_yield, | |
6655 bool concurrent_precleaning): | |
6656 _collector(collector), | |
6657 _span(span), | |
6658 _bit_map(bit_map), | |
6659 _mark_stack(mark_stack), | |
6660 _pushAndMarkClosure(collector, span, rp, bit_map, mod_union_table, | |
6661 mark_stack, revisit_stack, concurrent_precleaning), | |
6662 _yield(should_yield), | |
6663 _concurrent_precleaning(concurrent_precleaning), | |
6664 _freelistLock(NULL) | |
6665 { | |
6666 _ref_processor = rp; | |
6667 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); | |
6668 } | |
6669 | |
6670 // This closure is used to mark refs into the CMS generation at the | |
6671 // second (final) checkpoint, and to scan and transitively follow | |
6672 // the unmarked oops. It is also used during the concurrent precleaning | |
6673 // phase while scanning objects on dirty cards in the CMS generation. | |
6674 // The marks are made in the marking bit map and the marking stack is | |
6675 // used for keeping the (newly) grey objects during the scan. | |
6676 // The parallel version (Par_...) appears further below. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6677 void MarkRefsIntoAndScanClosure::do_oop(oop obj) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6678 if (obj != NULL) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6679 assert(obj->is_oop(), "expected an oop"); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6680 HeapWord* addr = (HeapWord*)obj; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6681 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6682 assert(_collector->overflow_list_is_empty(), |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6683 "overflow list should be empty"); |
0 | 6684 if (_span.contains(addr) && |
6685 !_bit_map->isMarked(addr)) { | |
6686 // mark bit map (object is now grey) | |
6687 _bit_map->mark(addr); | |
6688 // push on marking stack (stack should be empty), and drain the | |
6689 // stack by applying this closure to the oops in the oops popped | |
6690 // from the stack (i.e. blacken the grey objects) | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6691 bool res = _mark_stack->push(obj); |
0 | 6692 assert(res, "Should have space to push on empty stack"); |
6693 do { | |
6694 oop new_oop = _mark_stack->pop(); | |
6695 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); | |
6696 assert(new_oop->is_parsable(), "Found unparsable oop"); | |
6697 assert(_bit_map->isMarked((HeapWord*)new_oop), | |
6698 "only grey objects on this stack"); | |
6699 // iterate over the oops in this oop, marking and pushing | |
6700 // the ones in CMS heap (i.e. in _span). | |
6701 new_oop->oop_iterate(&_pushAndMarkClosure); | |
6702 // check if it's time to yield | |
6703 do_yield_check(); | |
6704 } while (!_mark_stack->isEmpty() || | |
6705 (!_concurrent_precleaning && take_from_overflow_list())); | |
6706 // if marking stack is empty, and we are not doing this | |
6707 // during precleaning, then check the overflow list | |
6708 } | |
6709 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); | |
6710 assert(_collector->overflow_list_is_empty(), | |
6711 "overflow list was drained above"); | |
6712 // We could restore evacuated mark words, if any, used for | |
6713 // overflow list links here because the overflow list is | |
6714 // provably empty here. That would reduce the maximum | |
6715 // size requirements for preserved_{oop,mark}_stack. | |
6716 // But we'll just postpone it until we are all done | |
6717 // so we can just stream through. | |
6718 if (!_concurrent_precleaning && CMSOverflowEarlyRestoration) { | |
6719 _collector->restore_preserved_marks_if_any(); | |
6720 assert(_collector->no_preserved_marks(), "No preserved marks"); | |
6721 } | |
6722 assert(!CMSOverflowEarlyRestoration || _collector->no_preserved_marks(), | |
6723 "All preserved marks should have been restored above"); | |
6724 } | |
6725 } | |
6726 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6727 void MarkRefsIntoAndScanClosure::do_oop(oop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6728 void MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { MarkRefsIntoAndScanClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6729 |
0 | 6730 void MarkRefsIntoAndScanClosure::do_yield_work() { |
6731 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
6732 "CMS thread should hold CMS token"); | |
6733 assert_lock_strong(_freelistLock); | |
6734 assert_lock_strong(_bit_map->lock()); | |
6735 // relinquish the free_list_lock and bitMaplock() | |
935 | 6736 DEBUG_ONLY(RememberKlassesChecker mux(false);) |
0 | 6737 _bit_map->lock()->unlock(); |
6738 _freelistLock->unlock(); | |
6739 ConcurrentMarkSweepThread::desynchronize(true); | |
6740 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
6741 _collector->stopTimer(); | |
6742 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); | |
6743 if (PrintCMSStatistics != 0) { | |
6744 _collector->incrementYields(); | |
6745 } | |
6746 _collector->icms_wait(); | |
6747 | |
6748 // See the comment in coordinator_yield() | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6749 for (unsigned i = 0; |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6750 i < CMSYieldSleepCount && |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6751 ConcurrentMarkSweepThread::should_yield() && |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6752 !CMSCollector::foregroundGCIsActive(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6753 ++i) { |
0 | 6754 os::sleep(Thread::current(), 1, false); |
6755 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
6756 } | |
6757 | |
6758 ConcurrentMarkSweepThread::synchronize(true); | |
6759 _freelistLock->lock_without_safepoint_check(); | |
6760 _bit_map->lock()->lock_without_safepoint_check(); | |
6761 _collector->startTimer(); | |
6762 } | |
6763 | |
6764 /////////////////////////////////////////////////////////// | |
6765 // Par_MarkRefsIntoAndScanClosure: a parallel version of | |
6766 // MarkRefsIntoAndScanClosure | |
6767 /////////////////////////////////////////////////////////// | |
6768 Par_MarkRefsIntoAndScanClosure::Par_MarkRefsIntoAndScanClosure( | |
6769 CMSCollector* collector, MemRegion span, ReferenceProcessor* rp, | |
6770 CMSBitMap* bit_map, OopTaskQueue* work_queue, CMSMarkStack* revisit_stack): | |
6771 _span(span), | |
6772 _bit_map(bit_map), | |
6773 _work_queue(work_queue), | |
6774 _low_water_mark(MIN2((uint)(work_queue->max_elems()/4), | |
6775 (uint)(CMSWorkQueueDrainThreshold * ParallelGCThreads))), | |
6776 _par_pushAndMarkClosure(collector, span, rp, bit_map, work_queue, | |
6777 revisit_stack) | |
6778 { | |
6779 _ref_processor = rp; | |
6780 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); | |
6781 } | |
6782 | |
6783 // This closure is used to mark refs into the CMS generation at the | |
6784 // second (final) checkpoint, and to scan and transitively follow | |
6785 // the unmarked oops. The marks are made in the marking bit map and | |
6786 // the work_queue is used for keeping the (newly) grey objects during | |
6787 // the scan phase whence they are also available for stealing by parallel | |
6788 // threads. Since the marking bit map is shared, updates are | |
6789 // synchronized (via CAS). | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6790 void Par_MarkRefsIntoAndScanClosure::do_oop(oop obj) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6791 if (obj != NULL) { |
0 | 6792 // Ignore mark word because this could be an already marked oop |
6793 // that may be chained at the end of the overflow list. | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
6794 assert(obj->is_oop(true), "expected an oop"); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6795 HeapWord* addr = (HeapWord*)obj; |
0 | 6796 if (_span.contains(addr) && |
6797 !_bit_map->isMarked(addr)) { | |
6798 // mark bit map (object will become grey): | |
6799 // It is possible for several threads to be | |
6800 // trying to "claim" this object concurrently; | |
6801 // the unique thread that succeeds in marking the | |
6802 // object first will do the subsequent push on | |
6803 // to the work queue (or overflow list). | |
6804 if (_bit_map->par_mark(addr)) { | |
6805 // push on work_queue (which may not be empty), and trim the | |
6806 // queue to an appropriate length by applying this closure to | |
6807 // the oops in the oops popped from the stack (i.e. blacken the | |
6808 // grey objects) | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6809 bool res = _work_queue->push(obj); |
0 | 6810 assert(res, "Low water mark should be less than capacity?"); |
6811 trim_queue(_low_water_mark); | |
6812 } // Else, another thread claimed the object | |
6813 } | |
6814 } | |
6815 } | |
6816 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6817 void Par_MarkRefsIntoAndScanClosure::do_oop(oop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6818 void Par_MarkRefsIntoAndScanClosure::do_oop(narrowOop* p) { Par_MarkRefsIntoAndScanClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6819 |
0 | 6820 // This closure is used to rescan the marked objects on the dirty cards |
6821 // in the mod union table and the card table proper. | |
6822 size_t ScanMarkedObjectsAgainCarefullyClosure::do_object_careful_m( | |
6823 oop p, MemRegion mr) { | |
6824 | |
6825 size_t size = 0; | |
6826 HeapWord* addr = (HeapWord*)p; | |
6827 DEBUG_ONLY(_collector->verify_work_stacks_empty();) | |
6828 assert(_span.contains(addr), "we are scanning the CMS generation"); | |
6829 // check if it's time to yield | |
6830 if (do_yield_check()) { | |
6831 // We yielded for some foreground stop-world work, | |
6832 // and we have been asked to abort this ongoing preclean cycle. | |
6833 return 0; | |
6834 } | |
6835 if (_bitMap->isMarked(addr)) { | |
6836 // it's marked; is it potentially uninitialized? | |
187 | 6837 if (p->klass_or_null() != NULL) { |
518
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
6838 // If is_conc_safe is false, the object may be undergoing |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
6839 // change by the VM outside a safepoint. Don't try to |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
6840 // scan it, but rather leave it for the remark phase. |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
6841 if (CMSPermGenPrecleaningEnabled && |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
6842 (!p->is_conc_safe() || !p->is_parsable())) { |
0 | 6843 // Signal precleaning to redirty the card since |
6844 // the klass pointer is already installed. | |
6845 assert(size == 0, "Initial value"); | |
6846 } else { | |
6847 assert(p->is_parsable(), "must be parsable."); | |
6848 // an initialized object; ignore mark word in verification below | |
6849 // since we are running concurrent with mutators | |
6850 assert(p->is_oop(true), "should be an oop"); | |
6851 if (p->is_objArray()) { | |
6852 // objArrays are precisely marked; restrict scanning | |
6853 // to dirty cards only. | |
187 | 6854 size = CompactibleFreeListSpace::adjustObjectSize( |
6855 p->oop_iterate(_scanningClosure, mr)); | |
0 | 6856 } else { |
6857 // A non-array may have been imprecisely marked; we need | |
6858 // to scan object in its entirety. | |
6859 size = CompactibleFreeListSpace::adjustObjectSize( | |
6860 p->oop_iterate(_scanningClosure)); | |
6861 } | |
6862 #ifdef DEBUG | |
6863 size_t direct_size = | |
6864 CompactibleFreeListSpace::adjustObjectSize(p->size()); | |
6865 assert(size == direct_size, "Inconsistency in size"); | |
6866 assert(size >= 3, "Necessary for Printezis marks to work"); | |
6867 if (!_bitMap->isMarked(addr+1)) { | |
6868 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size); | |
6869 } else { | |
6870 _bitMap->verifyNoOneBitsInRange(addr+2, addr+size-1); | |
6871 assert(_bitMap->isMarked(addr+size-1), | |
6872 "inconsistent Printezis mark"); | |
6873 } | |
6874 #endif // DEBUG | |
6875 } | |
6876 } else { | |
6877 // an unitialized object | |
6878 assert(_bitMap->isMarked(addr+1), "missing Printezis mark?"); | |
6879 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); | |
6880 size = pointer_delta(nextOneAddr + 1, addr); | |
6881 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), | |
6882 "alignment problem"); | |
6883 // Note that pre-cleaning needn't redirty the card. OopDesc::set_klass() | |
6884 // will dirty the card when the klass pointer is installed in the | |
6885 // object (signalling the completion of initialization). | |
6886 } | |
6887 } else { | |
6888 // Either a not yet marked object or an uninitialized object | |
187 | 6889 if (p->klass_or_null() == NULL || !p->is_parsable()) { |
0 | 6890 // An uninitialized object, skip to the next card, since |
6891 // we may not be able to read its P-bits yet. | |
6892 assert(size == 0, "Initial value"); | |
6893 } else { | |
6894 // An object not (yet) reached by marking: we merely need to | |
6895 // compute its size so as to go look at the next block. | |
6896 assert(p->is_oop(true), "should be an oop"); | |
6897 size = CompactibleFreeListSpace::adjustObjectSize(p->size()); | |
6898 } | |
6899 } | |
6900 DEBUG_ONLY(_collector->verify_work_stacks_empty();) | |
6901 return size; | |
6902 } | |
6903 | |
6904 void ScanMarkedObjectsAgainCarefullyClosure::do_yield_work() { | |
6905 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
6906 "CMS thread should hold CMS token"); | |
6907 assert_lock_strong(_freelistLock); | |
6908 assert_lock_strong(_bitMap->lock()); | |
935 | 6909 DEBUG_ONLY(RememberKlassesChecker mux(false);) |
0 | 6910 // relinquish the free_list_lock and bitMaplock() |
6911 _bitMap->lock()->unlock(); | |
6912 _freelistLock->unlock(); | |
6913 ConcurrentMarkSweepThread::desynchronize(true); | |
6914 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
6915 _collector->stopTimer(); | |
6916 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); | |
6917 if (PrintCMSStatistics != 0) { | |
6918 _collector->incrementYields(); | |
6919 } | |
6920 _collector->icms_wait(); | |
6921 | |
6922 // See the comment in coordinator_yield() | |
6923 for (unsigned i = 0; i < CMSYieldSleepCount && | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6924 ConcurrentMarkSweepThread::should_yield() && |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
6925 !CMSCollector::foregroundGCIsActive(); ++i) { |
0 | 6926 os::sleep(Thread::current(), 1, false); |
6927 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
6928 } | |
6929 | |
6930 ConcurrentMarkSweepThread::synchronize(true); | |
6931 _freelistLock->lock_without_safepoint_check(); | |
6932 _bitMap->lock()->lock_without_safepoint_check(); | |
6933 _collector->startTimer(); | |
6934 } | |
6935 | |
6936 | |
6937 ////////////////////////////////////////////////////////////////// | |
6938 // SurvivorSpacePrecleanClosure | |
6939 ////////////////////////////////////////////////////////////////// | |
6940 // This (single-threaded) closure is used to preclean the oops in | |
6941 // the survivor spaces. | |
6942 size_t SurvivorSpacePrecleanClosure::do_object_careful(oop p) { | |
6943 | |
6944 HeapWord* addr = (HeapWord*)p; | |
6945 DEBUG_ONLY(_collector->verify_work_stacks_empty();) | |
6946 assert(!_span.contains(addr), "we are scanning the survivor spaces"); | |
187 | 6947 assert(p->klass_or_null() != NULL, "object should be initializd"); |
0 | 6948 assert(p->is_parsable(), "must be parsable."); |
6949 // an initialized object; ignore mark word in verification below | |
6950 // since we are running concurrent with mutators | |
6951 assert(p->is_oop(true), "should be an oop"); | |
6952 // Note that we do not yield while we iterate over | |
6953 // the interior oops of p, pushing the relevant ones | |
6954 // on our marking stack. | |
6955 size_t size = p->oop_iterate(_scanning_closure); | |
6956 do_yield_check(); | |
6957 // Observe that below, we do not abandon the preclean | |
6958 // phase as soon as we should; rather we empty the | |
6959 // marking stack before returning. This is to satisfy | |
6960 // some existing assertions. In general, it may be a | |
6961 // good idea to abort immediately and complete the marking | |
6962 // from the grey objects at a later time. | |
6963 while (!_mark_stack->isEmpty()) { | |
6964 oop new_oop = _mark_stack->pop(); | |
6965 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); | |
6966 assert(new_oop->is_parsable(), "Found unparsable oop"); | |
6967 assert(_bit_map->isMarked((HeapWord*)new_oop), | |
6968 "only grey objects on this stack"); | |
6969 // iterate over the oops in this oop, marking and pushing | |
6970 // the ones in CMS heap (i.e. in _span). | |
6971 new_oop->oop_iterate(_scanning_closure); | |
6972 // check if it's time to yield | |
6973 do_yield_check(); | |
6974 } | |
6975 unsigned int after_count = | |
6976 GenCollectedHeap::heap()->total_collections(); | |
6977 bool abort = (_before_count != after_count) || | |
6978 _collector->should_abort_preclean(); | |
6979 return abort ? 0 : size; | |
6980 } | |
6981 | |
6982 void SurvivorSpacePrecleanClosure::do_yield_work() { | |
6983 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
6984 "CMS thread should hold CMS token"); | |
6985 assert_lock_strong(_bit_map->lock()); | |
935 | 6986 DEBUG_ONLY(RememberKlassesChecker smx(false);) |
0 | 6987 // Relinquish the bit map lock |
6988 _bit_map->lock()->unlock(); | |
6989 ConcurrentMarkSweepThread::desynchronize(true); | |
6990 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
6991 _collector->stopTimer(); | |
6992 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); | |
6993 if (PrintCMSStatistics != 0) { | |
6994 _collector->incrementYields(); | |
6995 } | |
6996 _collector->icms_wait(); | |
6997 | |
6998 // See the comment in coordinator_yield() | |
6999 for (unsigned i = 0; i < CMSYieldSleepCount && | |
7000 ConcurrentMarkSweepThread::should_yield() && | |
7001 !CMSCollector::foregroundGCIsActive(); ++i) { | |
7002 os::sleep(Thread::current(), 1, false); | |
7003 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
7004 } | |
7005 | |
7006 ConcurrentMarkSweepThread::synchronize(true); | |
7007 _bit_map->lock()->lock_without_safepoint_check(); | |
7008 _collector->startTimer(); | |
7009 } | |
7010 | |
7011 // This closure is used to rescan the marked objects on the dirty cards | |
7012 // in the mod union table and the card table proper. In the parallel | |
7013 // case, although the bitMap is shared, we do a single read so the | |
7014 // isMarked() query is "safe". | |
7015 bool ScanMarkedObjectsAgainClosure::do_object_bm(oop p, MemRegion mr) { | |
7016 // Ignore mark word because we are running concurrent with mutators | |
7017 assert(p->is_oop_or_null(true), "expected an oop or null"); | |
7018 HeapWord* addr = (HeapWord*)p; | |
7019 assert(_span.contains(addr), "we are scanning the CMS generation"); | |
7020 bool is_obj_array = false; | |
7021 #ifdef DEBUG | |
7022 if (!_parallel) { | |
7023 assert(_mark_stack->isEmpty(), "pre-condition (eager drainage)"); | |
7024 assert(_collector->overflow_list_is_empty(), | |
7025 "overflow list should be empty"); | |
7026 | |
7027 } | |
7028 #endif // DEBUG | |
7029 if (_bit_map->isMarked(addr)) { | |
7030 // Obj arrays are precisely marked, non-arrays are not; | |
7031 // so we scan objArrays precisely and non-arrays in their | |
7032 // entirety. | |
7033 if (p->is_objArray()) { | |
7034 is_obj_array = true; | |
7035 if (_parallel) { | |
7036 p->oop_iterate(_par_scan_closure, mr); | |
7037 } else { | |
7038 p->oop_iterate(_scan_closure, mr); | |
7039 } | |
7040 } else { | |
7041 if (_parallel) { | |
7042 p->oop_iterate(_par_scan_closure); | |
7043 } else { | |
7044 p->oop_iterate(_scan_closure); | |
7045 } | |
7046 } | |
7047 } | |
7048 #ifdef DEBUG | |
7049 if (!_parallel) { | |
7050 assert(_mark_stack->isEmpty(), "post-condition (eager drainage)"); | |
7051 assert(_collector->overflow_list_is_empty(), | |
7052 "overflow list should be empty"); | |
7053 | |
7054 } | |
7055 #endif // DEBUG | |
7056 return is_obj_array; | |
7057 } | |
7058 | |
7059 MarkFromRootsClosure::MarkFromRootsClosure(CMSCollector* collector, | |
7060 MemRegion span, | |
7061 CMSBitMap* bitMap, CMSMarkStack* markStack, | |
7062 CMSMarkStack* revisitStack, | |
7063 bool should_yield, bool verifying): | |
7064 _collector(collector), | |
7065 _span(span), | |
7066 _bitMap(bitMap), | |
7067 _mut(&collector->_modUnionTable), | |
7068 _markStack(markStack), | |
7069 _revisitStack(revisitStack), | |
7070 _yield(should_yield), | |
7071 _skipBits(0) | |
7072 { | |
7073 assert(_markStack->isEmpty(), "stack should be empty"); | |
7074 _finger = _bitMap->startWord(); | |
7075 _threshold = _finger; | |
7076 assert(_collector->_restart_addr == NULL, "Sanity check"); | |
7077 assert(_span.contains(_finger), "Out of bounds _finger?"); | |
7078 DEBUG_ONLY(_verifying = verifying;) | |
7079 } | |
7080 | |
7081 void MarkFromRootsClosure::reset(HeapWord* addr) { | |
7082 assert(_markStack->isEmpty(), "would cause duplicates on stack"); | |
7083 assert(_span.contains(addr), "Out of bounds _finger?"); | |
7084 _finger = addr; | |
7085 _threshold = (HeapWord*)round_to( | |
7086 (intptr_t)_finger, CardTableModRefBS::card_size); | |
7087 } | |
7088 | |
7089 // Should revisit to see if this should be restructured for | |
7090 // greater efficiency. | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7091 bool MarkFromRootsClosure::do_bit(size_t offset) { |
0 | 7092 if (_skipBits > 0) { |
7093 _skipBits--; | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7094 return true; |
0 | 7095 } |
7096 // convert offset into a HeapWord* | |
7097 HeapWord* addr = _bitMap->startWord() + offset; | |
7098 assert(_bitMap->endWord() && addr < _bitMap->endWord(), | |
7099 "address out of range"); | |
7100 assert(_bitMap->isMarked(addr), "tautology"); | |
7101 if (_bitMap->isMarked(addr+1)) { | |
7102 // this is an allocated but not yet initialized object | |
7103 assert(_skipBits == 0, "tautology"); | |
7104 _skipBits = 2; // skip next two marked bits ("Printezis-marks") | |
7105 oop p = oop(addr); | |
187 | 7106 if (p->klass_or_null() == NULL || !p->is_parsable()) { |
0 | 7107 DEBUG_ONLY(if (!_verifying) {) |
7108 // We re-dirty the cards on which this object lies and increase | |
7109 // the _threshold so that we'll come back to scan this object | |
7110 // during the preclean or remark phase. (CMSCleanOnEnter) | |
7111 if (CMSCleanOnEnter) { | |
7112 size_t sz = _collector->block_size_using_printezis_bits(addr); | |
7113 HeapWord* end_card_addr = (HeapWord*)round_to( | |
7114 (intptr_t)(addr+sz), CardTableModRefBS::card_size); | |
283
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7115 MemRegion redirty_range = MemRegion(addr, end_card_addr); |
0 | 7116 assert(!redirty_range.is_empty(), "Arithmetical tautology"); |
7117 // Bump _threshold to end_card_addr; note that | |
7118 // _threshold cannot possibly exceed end_card_addr, anyhow. | |
7119 // This prevents future clearing of the card as the scan proceeds | |
7120 // to the right. | |
7121 assert(_threshold <= end_card_addr, | |
7122 "Because we are just scanning into this object"); | |
7123 if (_threshold < end_card_addr) { | |
7124 _threshold = end_card_addr; | |
7125 } | |
187 | 7126 if (p->klass_or_null() != NULL) { |
0 | 7127 // Redirty the range of cards... |
7128 _mut->mark_range(redirty_range); | |
7129 } // ...else the setting of klass will dirty the card anyway. | |
7130 } | |
7131 DEBUG_ONLY(}) | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7132 return true; |
0 | 7133 } |
7134 } | |
7135 scanOopsInOop(addr); | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7136 return true; |
0 | 7137 } |
7138 | |
7139 // We take a break if we've been at this for a while, | |
7140 // so as to avoid monopolizing the locks involved. | |
7141 void MarkFromRootsClosure::do_yield_work() { | |
7142 // First give up the locks, then yield, then re-lock | |
7143 // We should probably use a constructor/destructor idiom to | |
7144 // do this unlock/lock or modify the MutexUnlocker class to | |
7145 // serve our purpose. XXX | |
7146 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
7147 "CMS thread should hold CMS token"); | |
7148 assert_lock_strong(_bitMap->lock()); | |
935 | 7149 DEBUG_ONLY(RememberKlassesChecker mux(false);) |
0 | 7150 _bitMap->lock()->unlock(); |
7151 ConcurrentMarkSweepThread::desynchronize(true); | |
7152 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
7153 _collector->stopTimer(); | |
7154 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); | |
7155 if (PrintCMSStatistics != 0) { | |
7156 _collector->incrementYields(); | |
7157 } | |
7158 _collector->icms_wait(); | |
7159 | |
7160 // See the comment in coordinator_yield() | |
7161 for (unsigned i = 0; i < CMSYieldSleepCount && | |
7162 ConcurrentMarkSweepThread::should_yield() && | |
7163 !CMSCollector::foregroundGCIsActive(); ++i) { | |
7164 os::sleep(Thread::current(), 1, false); | |
7165 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
7166 } | |
7167 | |
7168 ConcurrentMarkSweepThread::synchronize(true); | |
7169 _bitMap->lock()->lock_without_safepoint_check(); | |
7170 _collector->startTimer(); | |
7171 } | |
7172 | |
7173 void MarkFromRootsClosure::scanOopsInOop(HeapWord* ptr) { | |
7174 assert(_bitMap->isMarked(ptr), "expected bit to be set"); | |
7175 assert(_markStack->isEmpty(), | |
7176 "should drain stack to limit stack usage"); | |
7177 // convert ptr to an oop preparatory to scanning | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7178 oop obj = oop(ptr); |
0 | 7179 // Ignore mark word in verification below, since we |
7180 // may be running concurrent with mutators. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7181 assert(obj->is_oop(true), "should be an oop"); |
0 | 7182 assert(_finger <= ptr, "_finger runneth ahead"); |
7183 // advance the finger to right end of this object | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7184 _finger = ptr + obj->size(); |
0 | 7185 assert(_finger > ptr, "we just incremented it above"); |
7186 // On large heaps, it may take us some time to get through | |
7187 // the marking phase (especially if running iCMS). During | |
7188 // this time it's possible that a lot of mutations have | |
7189 // accumulated in the card table and the mod union table -- | |
7190 // these mutation records are redundant until we have | |
7191 // actually traced into the corresponding card. | |
7192 // Here, we check whether advancing the finger would make | |
7193 // us cross into a new card, and if so clear corresponding | |
7194 // cards in the MUT (preclean them in the card-table in the | |
7195 // future). | |
7196 | |
7197 DEBUG_ONLY(if (!_verifying) {) | |
7198 // The clean-on-enter optimization is disabled by default, | |
7199 // until we fix 6178663. | |
7200 if (CMSCleanOnEnter && (_finger > _threshold)) { | |
7201 // [_threshold, _finger) represents the interval | |
7202 // of cards to be cleared in MUT (or precleaned in card table). | |
7203 // The set of cards to be cleared is all those that overlap | |
7204 // with the interval [_threshold, _finger); note that | |
7205 // _threshold is always kept card-aligned but _finger isn't | |
7206 // always card-aligned. | |
7207 HeapWord* old_threshold = _threshold; | |
7208 assert(old_threshold == (HeapWord*)round_to( | |
7209 (intptr_t)old_threshold, CardTableModRefBS::card_size), | |
7210 "_threshold should always be card-aligned"); | |
7211 _threshold = (HeapWord*)round_to( | |
7212 (intptr_t)_finger, CardTableModRefBS::card_size); | |
7213 MemRegion mr(old_threshold, _threshold); | |
7214 assert(!mr.is_empty(), "Control point invariant"); | |
7215 assert(_span.contains(mr), "Should clear within span"); | |
7216 // XXX When _finger crosses from old gen into perm gen | |
7217 // we may be doing unnecessary cleaning; do better in the | |
7218 // future by detecting that condition and clearing fewer | |
7219 // MUT/CT entries. | |
7220 _mut->clear_range(mr); | |
7221 } | |
7222 DEBUG_ONLY(}) | |
7223 // Note: the finger doesn't advance while we drain | |
7224 // the stack below. | |
7225 PushOrMarkClosure pushOrMarkClosure(_collector, | |
7226 _span, _bitMap, _markStack, | |
7227 _revisitStack, | |
7228 _finger, this); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7229 bool res = _markStack->push(obj); |
0 | 7230 assert(res, "Empty non-zero size stack should have space for single push"); |
7231 while (!_markStack->isEmpty()) { | |
7232 oop new_oop = _markStack->pop(); | |
7233 // Skip verifying header mark word below because we are | |
7234 // running concurrent with mutators. | |
7235 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); | |
7236 // now scan this oop's oops | |
7237 new_oop->oop_iterate(&pushOrMarkClosure); | |
7238 do_yield_check(); | |
7239 } | |
7240 assert(_markStack->isEmpty(), "tautology, emphasizing post-condition"); | |
7241 } | |
7242 | |
7243 Par_MarkFromRootsClosure::Par_MarkFromRootsClosure(CMSConcMarkingTask* task, | |
7244 CMSCollector* collector, MemRegion span, | |
7245 CMSBitMap* bit_map, | |
7246 OopTaskQueue* work_queue, | |
7247 CMSMarkStack* overflow_stack, | |
7248 CMSMarkStack* revisit_stack, | |
7249 bool should_yield): | |
7250 _collector(collector), | |
7251 _whole_span(collector->_span), | |
7252 _span(span), | |
7253 _bit_map(bit_map), | |
7254 _mut(&collector->_modUnionTable), | |
7255 _work_queue(work_queue), | |
7256 _overflow_stack(overflow_stack), | |
7257 _revisit_stack(revisit_stack), | |
7258 _yield(should_yield), | |
7259 _skip_bits(0), | |
7260 _task(task) | |
7261 { | |
7262 assert(_work_queue->size() == 0, "work_queue should be empty"); | |
7263 _finger = span.start(); | |
7264 _threshold = _finger; // XXX Defer clear-on-enter optimization for now | |
7265 assert(_span.contains(_finger), "Out of bounds _finger?"); | |
7266 } | |
7267 | |
7268 // Should revisit to see if this should be restructured for | |
7269 // greater efficiency. | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7270 bool Par_MarkFromRootsClosure::do_bit(size_t offset) { |
0 | 7271 if (_skip_bits > 0) { |
7272 _skip_bits--; | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7273 return true; |
0 | 7274 } |
7275 // convert offset into a HeapWord* | |
7276 HeapWord* addr = _bit_map->startWord() + offset; | |
7277 assert(_bit_map->endWord() && addr < _bit_map->endWord(), | |
7278 "address out of range"); | |
7279 assert(_bit_map->isMarked(addr), "tautology"); | |
7280 if (_bit_map->isMarked(addr+1)) { | |
7281 // this is an allocated object that might not yet be initialized | |
7282 assert(_skip_bits == 0, "tautology"); | |
7283 _skip_bits = 2; // skip next two marked bits ("Printezis-marks") | |
7284 oop p = oop(addr); | |
187 | 7285 if (p->klass_or_null() == NULL || !p->is_parsable()) { |
0 | 7286 // in the case of Clean-on-Enter optimization, redirty card |
7287 // and avoid clearing card by increasing the threshold. | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7288 return true; |
0 | 7289 } |
7290 } | |
7291 scan_oops_in_oop(addr); | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7292 return true; |
0 | 7293 } |
7294 | |
7295 void Par_MarkFromRootsClosure::scan_oops_in_oop(HeapWord* ptr) { | |
7296 assert(_bit_map->isMarked(ptr), "expected bit to be set"); | |
7297 // Should we assert that our work queue is empty or | |
7298 // below some drain limit? | |
7299 assert(_work_queue->size() == 0, | |
7300 "should drain stack to limit stack usage"); | |
7301 // convert ptr to an oop preparatory to scanning | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7302 oop obj = oop(ptr); |
0 | 7303 // Ignore mark word in verification below, since we |
7304 // may be running concurrent with mutators. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7305 assert(obj->is_oop(true), "should be an oop"); |
0 | 7306 assert(_finger <= ptr, "_finger runneth ahead"); |
7307 // advance the finger to right end of this object | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7308 _finger = ptr + obj->size(); |
0 | 7309 assert(_finger > ptr, "we just incremented it above"); |
7310 // On large heaps, it may take us some time to get through | |
7311 // the marking phase (especially if running iCMS). During | |
7312 // this time it's possible that a lot of mutations have | |
7313 // accumulated in the card table and the mod union table -- | |
7314 // these mutation records are redundant until we have | |
7315 // actually traced into the corresponding card. | |
7316 // Here, we check whether advancing the finger would make | |
7317 // us cross into a new card, and if so clear corresponding | |
7318 // cards in the MUT (preclean them in the card-table in the | |
7319 // future). | |
7320 | |
7321 // The clean-on-enter optimization is disabled by default, | |
7322 // until we fix 6178663. | |
7323 if (CMSCleanOnEnter && (_finger > _threshold)) { | |
7324 // [_threshold, _finger) represents the interval | |
7325 // of cards to be cleared in MUT (or precleaned in card table). | |
7326 // The set of cards to be cleared is all those that overlap | |
7327 // with the interval [_threshold, _finger); note that | |
7328 // _threshold is always kept card-aligned but _finger isn't | |
7329 // always card-aligned. | |
7330 HeapWord* old_threshold = _threshold; | |
7331 assert(old_threshold == (HeapWord*)round_to( | |
7332 (intptr_t)old_threshold, CardTableModRefBS::card_size), | |
7333 "_threshold should always be card-aligned"); | |
7334 _threshold = (HeapWord*)round_to( | |
7335 (intptr_t)_finger, CardTableModRefBS::card_size); | |
7336 MemRegion mr(old_threshold, _threshold); | |
7337 assert(!mr.is_empty(), "Control point invariant"); | |
7338 assert(_span.contains(mr), "Should clear within span"); // _whole_span ?? | |
7339 // XXX When _finger crosses from old gen into perm gen | |
7340 // we may be doing unnecessary cleaning; do better in the | |
7341 // future by detecting that condition and clearing fewer | |
7342 // MUT/CT entries. | |
7343 _mut->clear_range(mr); | |
7344 } | |
7345 | |
7346 // Note: the local finger doesn't advance while we drain | |
7347 // the stack below, but the global finger sure can and will. | |
7348 HeapWord** gfa = _task->global_finger_addr(); | |
7349 Par_PushOrMarkClosure pushOrMarkClosure(_collector, | |
7350 _span, _bit_map, | |
7351 _work_queue, | |
7352 _overflow_stack, | |
7353 _revisit_stack, | |
7354 _finger, | |
7355 gfa, this); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7356 bool res = _work_queue->push(obj); // overflow could occur here |
0 | 7357 assert(res, "Will hold once we use workqueues"); |
7358 while (true) { | |
7359 oop new_oop; | |
7360 if (!_work_queue->pop_local(new_oop)) { | |
7361 // We emptied our work_queue; check if there's stuff that can | |
7362 // be gotten from the overflow stack. | |
7363 if (CMSConcMarkingTask::get_work_from_overflow_stack( | |
7364 _overflow_stack, _work_queue)) { | |
7365 do_yield_check(); | |
7366 continue; | |
7367 } else { // done | |
7368 break; | |
7369 } | |
7370 } | |
7371 // Skip verifying header mark word below because we are | |
7372 // running concurrent with mutators. | |
7373 assert(new_oop->is_oop(true), "Oops! expected to pop an oop"); | |
7374 // now scan this oop's oops | |
7375 new_oop->oop_iterate(&pushOrMarkClosure); | |
7376 do_yield_check(); | |
7377 } | |
7378 assert(_work_queue->size() == 0, "tautology, emphasizing post-condition"); | |
7379 } | |
7380 | |
7381 // Yield in response to a request from VM Thread or | |
7382 // from mutators. | |
7383 void Par_MarkFromRootsClosure::do_yield_work() { | |
7384 assert(_task != NULL, "sanity"); | |
7385 _task->yield(); | |
7386 } | |
7387 | |
7388 // A variant of the above used for verifying CMS marking work. | |
7389 MarkFromRootsVerifyClosure::MarkFromRootsVerifyClosure(CMSCollector* collector, | |
7390 MemRegion span, | |
7391 CMSBitMap* verification_bm, CMSBitMap* cms_bm, | |
7392 CMSMarkStack* mark_stack): | |
7393 _collector(collector), | |
7394 _span(span), | |
7395 _verification_bm(verification_bm), | |
7396 _cms_bm(cms_bm), | |
7397 _mark_stack(mark_stack), | |
7398 _pam_verify_closure(collector, span, verification_bm, cms_bm, | |
7399 mark_stack) | |
7400 { | |
7401 assert(_mark_stack->isEmpty(), "stack should be empty"); | |
7402 _finger = _verification_bm->startWord(); | |
7403 assert(_collector->_restart_addr == NULL, "Sanity check"); | |
7404 assert(_span.contains(_finger), "Out of bounds _finger?"); | |
7405 } | |
7406 | |
7407 void MarkFromRootsVerifyClosure::reset(HeapWord* addr) { | |
7408 assert(_mark_stack->isEmpty(), "would cause duplicates on stack"); | |
7409 assert(_span.contains(addr), "Out of bounds _finger?"); | |
7410 _finger = addr; | |
7411 } | |
7412 | |
7413 // Should revisit to see if this should be restructured for | |
7414 // greater efficiency. | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7415 bool MarkFromRootsVerifyClosure::do_bit(size_t offset) { |
0 | 7416 // convert offset into a HeapWord* |
7417 HeapWord* addr = _verification_bm->startWord() + offset; | |
7418 assert(_verification_bm->endWord() && addr < _verification_bm->endWord(), | |
7419 "address out of range"); | |
7420 assert(_verification_bm->isMarked(addr), "tautology"); | |
7421 assert(_cms_bm->isMarked(addr), "tautology"); | |
7422 | |
7423 assert(_mark_stack->isEmpty(), | |
7424 "should drain stack to limit stack usage"); | |
7425 // convert addr to an oop preparatory to scanning | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7426 oop obj = oop(addr); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7427 assert(obj->is_oop(), "should be an oop"); |
0 | 7428 assert(_finger <= addr, "_finger runneth ahead"); |
7429 // advance the finger to right end of this object | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7430 _finger = addr + obj->size(); |
0 | 7431 assert(_finger > addr, "we just incremented it above"); |
7432 // Note: the finger doesn't advance while we drain | |
7433 // the stack below. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7434 bool res = _mark_stack->push(obj); |
0 | 7435 assert(res, "Empty non-zero size stack should have space for single push"); |
7436 while (!_mark_stack->isEmpty()) { | |
7437 oop new_oop = _mark_stack->pop(); | |
7438 assert(new_oop->is_oop(), "Oops! expected to pop an oop"); | |
7439 // now scan this oop's oops | |
7440 new_oop->oop_iterate(&_pam_verify_closure); | |
7441 } | |
7442 assert(_mark_stack->isEmpty(), "tautology, emphasizing post-condition"); | |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7443 return true; |
0 | 7444 } |
7445 | |
7446 PushAndMarkVerifyClosure::PushAndMarkVerifyClosure( | |
7447 CMSCollector* collector, MemRegion span, | |
7448 CMSBitMap* verification_bm, CMSBitMap* cms_bm, | |
7449 CMSMarkStack* mark_stack): | |
7450 OopClosure(collector->ref_processor()), | |
7451 _collector(collector), | |
7452 _span(span), | |
7453 _verification_bm(verification_bm), | |
7454 _cms_bm(cms_bm), | |
7455 _mark_stack(mark_stack) | |
7456 { } | |
7457 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7458 void PushAndMarkVerifyClosure::do_oop(oop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7459 void PushAndMarkVerifyClosure::do_oop(narrowOop* p) { PushAndMarkVerifyClosure::do_oop_work(p); } |
0 | 7460 |
7461 // Upon stack overflow, we discard (part of) the stack, | |
7462 // remembering the least address amongst those discarded | |
7463 // in CMSCollector's _restart_address. | |
7464 void PushAndMarkVerifyClosure::handle_stack_overflow(HeapWord* lost) { | |
7465 // Remember the least grey address discarded | |
7466 HeapWord* ra = (HeapWord*)_mark_stack->least_value(lost); | |
7467 _collector->lower_restart_addr(ra); | |
7468 _mark_stack->reset(); // discard stack contents | |
7469 _mark_stack->expand(); // expand the stack if possible | |
7470 } | |
7471 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7472 void PushAndMarkVerifyClosure::do_oop(oop obj) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7473 assert(obj->is_oop_or_null(), "expected an oop or NULL"); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7474 HeapWord* addr = (HeapWord*)obj; |
0 | 7475 if (_span.contains(addr) && !_verification_bm->isMarked(addr)) { |
7476 // Oop lies in _span and isn't yet grey or black | |
7477 _verification_bm->mark(addr); // now grey | |
7478 if (!_cms_bm->isMarked(addr)) { | |
7479 oop(addr)->print(); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7480 gclog_or_tty->print_cr(" (" INTPTR_FORMAT " should have been marked)", |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7481 addr); |
0 | 7482 fatal("... aborting"); |
7483 } | |
7484 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7485 if (!_mark_stack->push(obj)) { // stack overflow |
0 | 7486 if (PrintCMSStatistics != 0) { |
7487 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " | |
7488 SIZE_FORMAT, _mark_stack->capacity()); | |
7489 } | |
7490 assert(_mark_stack->isFull(), "Else push should have succeeded"); | |
7491 handle_stack_overflow(addr); | |
7492 } | |
7493 // anything including and to the right of _finger | |
7494 // will be scanned as we iterate over the remainder of the | |
7495 // bit map | |
7496 } | |
7497 } | |
7498 | |
7499 PushOrMarkClosure::PushOrMarkClosure(CMSCollector* collector, | |
7500 MemRegion span, | |
7501 CMSBitMap* bitMap, CMSMarkStack* markStack, | |
7502 CMSMarkStack* revisitStack, | |
7503 HeapWord* finger, MarkFromRootsClosure* parent) : | |
935 | 7504 KlassRememberingOopClosure(collector, collector->ref_processor(), revisitStack), |
0 | 7505 _span(span), |
7506 _bitMap(bitMap), | |
7507 _markStack(markStack), | |
7508 _finger(finger), | |
935 | 7509 _parent(parent) |
0 | 7510 { } |
7511 | |
7512 Par_PushOrMarkClosure::Par_PushOrMarkClosure(CMSCollector* collector, | |
7513 MemRegion span, | |
7514 CMSBitMap* bit_map, | |
7515 OopTaskQueue* work_queue, | |
7516 CMSMarkStack* overflow_stack, | |
7517 CMSMarkStack* revisit_stack, | |
7518 HeapWord* finger, | |
7519 HeapWord** global_finger_addr, | |
7520 Par_MarkFromRootsClosure* parent) : | |
935 | 7521 Par_KlassRememberingOopClosure(collector, |
7522 collector->ref_processor(), | |
7523 revisit_stack), | |
0 | 7524 _whole_span(collector->_span), |
7525 _span(span), | |
7526 _bit_map(bit_map), | |
7527 _work_queue(work_queue), | |
7528 _overflow_stack(overflow_stack), | |
7529 _finger(finger), | |
7530 _global_finger_addr(global_finger_addr), | |
935 | 7531 _parent(parent) |
0 | 7532 { } |
7533 | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
7534 // Assumes thread-safe access by callers, who are |
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
7535 // responsible for mutual exclusion. |
0 | 7536 void CMSCollector::lower_restart_addr(HeapWord* low) { |
7537 assert(_span.contains(low), "Out of bounds addr"); | |
7538 if (_restart_addr == NULL) { | |
7539 _restart_addr = low; | |
7540 } else { | |
7541 _restart_addr = MIN2(_restart_addr, low); | |
7542 } | |
7543 } | |
7544 | |
7545 // Upon stack overflow, we discard (part of) the stack, | |
7546 // remembering the least address amongst those discarded | |
7547 // in CMSCollector's _restart_address. | |
7548 void PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { | |
7549 // Remember the least grey address discarded | |
7550 HeapWord* ra = (HeapWord*)_markStack->least_value(lost); | |
7551 _collector->lower_restart_addr(ra); | |
7552 _markStack->reset(); // discard stack contents | |
7553 _markStack->expand(); // expand the stack if possible | |
7554 } | |
7555 | |
7556 // Upon stack overflow, we discard (part of) the stack, | |
7557 // remembering the least address amongst those discarded | |
7558 // in CMSCollector's _restart_address. | |
7559 void Par_PushOrMarkClosure::handle_stack_overflow(HeapWord* lost) { | |
7560 // We need to do this under a mutex to prevent other | |
340
ebeb6490b814
6722116: CMS: Incorrect overflow handling when using parallel concurrent marking
ysr
parents:
283
diff
changeset
|
7561 // workers from interfering with the work done below. |
0 | 7562 MutexLockerEx ml(_overflow_stack->par_lock(), |
7563 Mutex::_no_safepoint_check_flag); | |
7564 // Remember the least grey address discarded | |
7565 HeapWord* ra = (HeapWord*)_overflow_stack->least_value(lost); | |
7566 _collector->lower_restart_addr(ra); | |
7567 _overflow_stack->reset(); // discard stack contents | |
7568 _overflow_stack->expand(); // expand the stack if possible | |
7569 } | |
7570 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7571 void PushOrMarkClosure::do_oop(oop obj) { |
0 | 7572 // Ignore mark word because we are running concurrent with mutators. |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7573 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7574 HeapWord* addr = (HeapWord*)obj; |
0 | 7575 if (_span.contains(addr) && !_bitMap->isMarked(addr)) { |
7576 // Oop lies in _span and isn't yet grey or black | |
7577 _bitMap->mark(addr); // now grey | |
7578 if (addr < _finger) { | |
7579 // the bit map iteration has already either passed, or | |
7580 // sampled, this bit in the bit map; we'll need to | |
7581 // use the marking stack to scan this oop's oops. | |
7582 bool simulate_overflow = false; | |
7583 NOT_PRODUCT( | |
7584 if (CMSMarkStackOverflowALot && | |
7585 _collector->simulate_overflow()) { | |
7586 // simulate a stack overflow | |
7587 simulate_overflow = true; | |
7588 } | |
7589 ) | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7590 if (simulate_overflow || !_markStack->push(obj)) { // stack overflow |
0 | 7591 if (PrintCMSStatistics != 0) { |
7592 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " | |
7593 SIZE_FORMAT, _markStack->capacity()); | |
7594 } | |
7595 assert(simulate_overflow || _markStack->isFull(), "Else push should have succeeded"); | |
7596 handle_stack_overflow(addr); | |
7597 } | |
7598 } | |
7599 // anything including and to the right of _finger | |
7600 // will be scanned as we iterate over the remainder of the | |
7601 // bit map | |
7602 do_yield_check(); | |
7603 } | |
7604 } | |
7605 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7606 void PushOrMarkClosure::do_oop(oop* p) { PushOrMarkClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7607 void PushOrMarkClosure::do_oop(narrowOop* p) { PushOrMarkClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7608 |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7609 void Par_PushOrMarkClosure::do_oop(oop obj) { |
0 | 7610 // Ignore mark word because we are running concurrent with mutators. |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7611 assert(obj->is_oop_or_null(true), "expected an oop or NULL"); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7612 HeapWord* addr = (HeapWord*)obj; |
0 | 7613 if (_whole_span.contains(addr) && !_bit_map->isMarked(addr)) { |
7614 // Oop lies in _span and isn't yet grey or black | |
7615 // We read the global_finger (volatile read) strictly after marking oop | |
7616 bool res = _bit_map->par_mark(addr); // now grey | |
7617 volatile HeapWord** gfa = (volatile HeapWord**)_global_finger_addr; | |
7618 // Should we push this marked oop on our stack? | |
7619 // -- if someone else marked it, nothing to do | |
7620 // -- if target oop is above global finger nothing to do | |
7621 // -- if target oop is in chunk and above local finger | |
7622 // then nothing to do | |
7623 // -- else push on work queue | |
7624 if ( !res // someone else marked it, they will deal with it | |
7625 || (addr >= *gfa) // will be scanned in a later task | |
7626 || (_span.contains(addr) && addr >= _finger)) { // later in this chunk | |
7627 return; | |
7628 } | |
7629 // the bit map iteration has already either passed, or | |
7630 // sampled, this bit in the bit map; we'll need to | |
7631 // use the marking stack to scan this oop's oops. | |
7632 bool simulate_overflow = false; | |
7633 NOT_PRODUCT( | |
7634 if (CMSMarkStackOverflowALot && | |
7635 _collector->simulate_overflow()) { | |
7636 // simulate a stack overflow | |
7637 simulate_overflow = true; | |
7638 } | |
7639 ) | |
7640 if (simulate_overflow || | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7641 !(_work_queue->push(obj) || _overflow_stack->par_push(obj))) { |
0 | 7642 // stack overflow |
7643 if (PrintCMSStatistics != 0) { | |
7644 gclog_or_tty->print_cr("CMS marking stack overflow (benign) at " | |
7645 SIZE_FORMAT, _overflow_stack->capacity()); | |
7646 } | |
7647 // We cannot assert that the overflow stack is full because | |
7648 // it may have been emptied since. | |
7649 assert(simulate_overflow || | |
7650 _work_queue->size() == _work_queue->max_elems(), | |
7651 "Else push should have succeeded"); | |
7652 handle_stack_overflow(addr); | |
7653 } | |
7654 do_yield_check(); | |
7655 } | |
7656 } | |
7657 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7658 void Par_PushOrMarkClosure::do_oop(oop* p) { Par_PushOrMarkClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7659 void Par_PushOrMarkClosure::do_oop(narrowOop* p) { Par_PushOrMarkClosure::do_oop_work(p); } |
0 | 7660 |
935 | 7661 KlassRememberingOopClosure::KlassRememberingOopClosure(CMSCollector* collector, |
7662 ReferenceProcessor* rp, | |
7663 CMSMarkStack* revisit_stack) : | |
7664 OopClosure(rp), | |
7665 _collector(collector), | |
7666 _revisit_stack(revisit_stack), | |
7667 _should_remember_klasses(collector->should_unload_classes()) {} | |
7668 | |
0 | 7669 PushAndMarkClosure::PushAndMarkClosure(CMSCollector* collector, |
7670 MemRegion span, | |
7671 ReferenceProcessor* rp, | |
7672 CMSBitMap* bit_map, | |
7673 CMSBitMap* mod_union_table, | |
7674 CMSMarkStack* mark_stack, | |
7675 CMSMarkStack* revisit_stack, | |
7676 bool concurrent_precleaning): | |
935 | 7677 KlassRememberingOopClosure(collector, rp, revisit_stack), |
0 | 7678 _span(span), |
7679 _bit_map(bit_map), | |
7680 _mod_union_table(mod_union_table), | |
7681 _mark_stack(mark_stack), | |
935 | 7682 _concurrent_precleaning(concurrent_precleaning) |
0 | 7683 { |
7684 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); | |
7685 } | |
7686 | |
7687 // Grey object rescan during pre-cleaning and second checkpoint phases -- | |
7688 // the non-parallel version (the parallel version appears further below.) | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7689 void PushAndMarkClosure::do_oop(oop obj) { |
342
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7690 // Ignore mark word verification. If during concurrent precleaning, |
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7691 // the object monitor may be locked. If during the checkpoint |
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7692 // phases, the object may already have been reached by a different |
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7693 // path and may be at the end of the global overflow list (so |
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7694 // the mark word may be NULL). |
37f87013dfd8
6711316: Open source the Garbage-First garbage collector
ysr
parents:
143
diff
changeset
|
7695 assert(obj->is_oop_or_null(true /* ignore mark word */), |
0 | 7696 "expected an oop or NULL"); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7697 HeapWord* addr = (HeapWord*)obj; |
0 | 7698 // Check if oop points into the CMS generation |
7699 // and is not marked | |
7700 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { | |
7701 // a white object ... | |
7702 _bit_map->mark(addr); // ... now grey | |
7703 // push on the marking stack (grey set) | |
7704 bool simulate_overflow = false; | |
7705 NOT_PRODUCT( | |
7706 if (CMSMarkStackOverflowALot && | |
7707 _collector->simulate_overflow()) { | |
7708 // simulate a stack overflow | |
7709 simulate_overflow = true; | |
7710 } | |
7711 ) | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7712 if (simulate_overflow || !_mark_stack->push(obj)) { |
0 | 7713 if (_concurrent_precleaning) { |
283
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7714 // During precleaning we can just dirty the appropriate card(s) |
0 | 7715 // in the mod union table, thus ensuring that the object remains |
283
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7716 // in the grey set and continue. In the case of object arrays |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7717 // we need to dirty all of the cards that the object spans, |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7718 // since the rescan of object arrays will be limited to the |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7719 // dirty cards. |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7720 // Note that no one can be intefering with us in this action |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7721 // of dirtying the mod union table, so no locking or atomics |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7722 // are required. |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7723 if (obj->is_objArray()) { |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7724 size_t sz = obj->size(); |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7725 HeapWord* end_card_addr = (HeapWord*)round_to( |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7726 (intptr_t)(addr+sz), CardTableModRefBS::card_size); |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7727 MemRegion redirty_range = MemRegion(addr, end_card_addr); |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7728 assert(!redirty_range.is_empty(), "Arithmetical tautology"); |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7729 _mod_union_table->mark_range(redirty_range); |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7730 } else { |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7731 _mod_union_table->mark(addr); |
9199f248b0ee
6722112: CMS: Incorrect encoding of overflown object arrays during concurrent precleaning
ysr
parents:
271
diff
changeset
|
7732 } |
0 | 7733 _collector->_ser_pmc_preclean_ovflw++; |
7734 } else { | |
7735 // During the remark phase, we need to remember this oop | |
7736 // in the overflow list. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7737 _collector->push_on_overflow_list(obj); |
0 | 7738 _collector->_ser_pmc_remark_ovflw++; |
7739 } | |
7740 } | |
7741 } | |
7742 } | |
7743 | |
7744 Par_PushAndMarkClosure::Par_PushAndMarkClosure(CMSCollector* collector, | |
7745 MemRegion span, | |
7746 ReferenceProcessor* rp, | |
7747 CMSBitMap* bit_map, | |
7748 OopTaskQueue* work_queue, | |
7749 CMSMarkStack* revisit_stack): | |
935 | 7750 Par_KlassRememberingOopClosure(collector, rp, revisit_stack), |
0 | 7751 _span(span), |
7752 _bit_map(bit_map), | |
935 | 7753 _work_queue(work_queue) |
0 | 7754 { |
7755 assert(_ref_processor != NULL, "_ref_processor shouldn't be NULL"); | |
7756 } | |
7757 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7758 void PushAndMarkClosure::do_oop(oop* p) { PushAndMarkClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7759 void PushAndMarkClosure::do_oop(narrowOop* p) { PushAndMarkClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7760 |
0 | 7761 // Grey object rescan during second checkpoint phase -- |
7762 // the parallel version. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7763 void Par_PushAndMarkClosure::do_oop(oop obj) { |
0 | 7764 // In the assert below, we ignore the mark word because |
7765 // this oop may point to an already visited object that is | |
7766 // on the overflow stack (in which case the mark word has | |
7767 // been hijacked for chaining into the overflow stack -- | |
7768 // if this is the last object in the overflow stack then | |
7769 // its mark word will be NULL). Because this object may | |
7770 // have been subsequently popped off the global overflow | |
7771 // stack, and the mark word possibly restored to the prototypical | |
7772 // value, by the time we get to examined this failing assert in | |
7773 // the debugger, is_oop_or_null(false) may subsequently start | |
7774 // to hold. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7775 assert(obj->is_oop_or_null(true), |
0 | 7776 "expected an oop or NULL"); |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7777 HeapWord* addr = (HeapWord*)obj; |
0 | 7778 // Check if oop points into the CMS generation |
7779 // and is not marked | |
7780 if (_span.contains(addr) && !_bit_map->isMarked(addr)) { | |
7781 // a white object ... | |
7782 // If we manage to "claim" the object, by being the | |
7783 // first thread to mark it, then we push it on our | |
7784 // marking stack | |
7785 if (_bit_map->par_mark(addr)) { // ... now grey | |
7786 // push on work queue (grey set) | |
7787 bool simulate_overflow = false; | |
7788 NOT_PRODUCT( | |
7789 if (CMSMarkStackOverflowALot && | |
7790 _collector->par_simulate_overflow()) { | |
7791 // simulate a stack overflow | |
7792 simulate_overflow = true; | |
7793 } | |
7794 ) | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7795 if (simulate_overflow || !_work_queue->push(obj)) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7796 _collector->par_push_on_overflow_list(obj); |
0 | 7797 _collector->_par_pmc_remark_ovflw++; // imprecise OK: no need to CAS |
7798 } | |
7799 } // Else, some other thread got there first | |
7800 } | |
7801 } | |
7802 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7803 void Par_PushAndMarkClosure::do_oop(oop* p) { Par_PushAndMarkClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7804 void Par_PushAndMarkClosure::do_oop(narrowOop* p) { Par_PushAndMarkClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
7805 |
941 | 7806 void PushAndMarkClosure::remember_mdo(DataLayout* v) { |
7807 // TBD | |
7808 } | |
7809 | |
7810 void Par_PushAndMarkClosure::remember_mdo(DataLayout* v) { | |
7811 // TBD | |
7812 } | |
7813 | |
0 | 7814 void CMSPrecleanRefsYieldClosure::do_yield_work() { |
935 | 7815 DEBUG_ONLY(RememberKlassesChecker mux(false);) |
0 | 7816 Mutex* bml = _collector->bitMapLock(); |
7817 assert_lock_strong(bml); | |
7818 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
7819 "CMS thread should hold CMS token"); | |
7820 | |
7821 bml->unlock(); | |
7822 ConcurrentMarkSweepThread::desynchronize(true); | |
7823 | |
7824 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
7825 | |
7826 _collector->stopTimer(); | |
7827 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); | |
7828 if (PrintCMSStatistics != 0) { | |
7829 _collector->incrementYields(); | |
7830 } | |
7831 _collector->icms_wait(); | |
7832 | |
7833 // See the comment in coordinator_yield() | |
7834 for (unsigned i = 0; i < CMSYieldSleepCount && | |
7835 ConcurrentMarkSweepThread::should_yield() && | |
7836 !CMSCollector::foregroundGCIsActive(); ++i) { | |
7837 os::sleep(Thread::current(), 1, false); | |
7838 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
7839 } | |
7840 | |
7841 ConcurrentMarkSweepThread::synchronize(true); | |
7842 bml->lock(); | |
7843 | |
7844 _collector->startTimer(); | |
7845 } | |
7846 | |
7847 bool CMSPrecleanRefsYieldClosure::should_return() { | |
7848 if (ConcurrentMarkSweepThread::should_yield()) { | |
7849 do_yield_work(); | |
7850 } | |
7851 return _collector->foregroundGCIsActive(); | |
7852 } | |
7853 | |
7854 void MarkFromDirtyCardsClosure::do_MemRegion(MemRegion mr) { | |
7855 assert(((size_t)mr.start())%CardTableModRefBS::card_size_in_words == 0, | |
7856 "mr should be aligned to start at a card boundary"); | |
7857 // We'd like to assert: | |
7858 // assert(mr.word_size()%CardTableModRefBS::card_size_in_words == 0, | |
7859 // "mr should be a range of cards"); | |
7860 // However, that would be too strong in one case -- the last | |
7861 // partition ends at _unallocated_block which, in general, can be | |
7862 // an arbitrary boundary, not necessarily card aligned. | |
7863 if (PrintCMSStatistics != 0) { | |
7864 _num_dirty_cards += | |
7865 mr.word_size()/CardTableModRefBS::card_size_in_words; | |
7866 } | |
7867 _space->object_iterate_mem(mr, &_scan_cl); | |
7868 } | |
7869 | |
7870 SweepClosure::SweepClosure(CMSCollector* collector, | |
7871 ConcurrentMarkSweepGeneration* g, | |
7872 CMSBitMap* bitMap, bool should_yield) : | |
7873 _collector(collector), | |
7874 _g(g), | |
7875 _sp(g->cmsSpace()), | |
7876 _limit(_sp->sweep_limit()), | |
7877 _freelistLock(_sp->freelistLock()), | |
7878 _bitMap(bitMap), | |
7879 _yield(should_yield), | |
7880 _inFreeRange(false), // No free range at beginning of sweep | |
7881 _freeRangeInFreeLists(false), // No free range at beginning of sweep | |
7882 _lastFreeRangeCoalesced(false), | |
7883 _freeFinger(g->used_region().start()) | |
7884 { | |
7885 NOT_PRODUCT( | |
7886 _numObjectsFreed = 0; | |
7887 _numWordsFreed = 0; | |
7888 _numObjectsLive = 0; | |
7889 _numWordsLive = 0; | |
7890 _numObjectsAlreadyFree = 0; | |
7891 _numWordsAlreadyFree = 0; | |
7892 _last_fc = NULL; | |
7893 | |
7894 _sp->initializeIndexedFreeListArrayReturnedBytes(); | |
7895 _sp->dictionary()->initializeDictReturnedBytes(); | |
7896 ) | |
7897 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), | |
7898 "sweep _limit out of bounds"); | |
7899 if (CMSTraceSweeper) { | |
7900 gclog_or_tty->print("\n====================\nStarting new sweep\n"); | |
7901 } | |
7902 } | |
7903 | |
7904 // We need this destructor to reclaim any space at the end | |
7905 // of the space, which do_blk below may not have added back to | |
7906 // the free lists. [basically dealing with the "fringe effect"] | |
7907 SweepClosure::~SweepClosure() { | |
7908 assert_lock_strong(_freelistLock); | |
7909 // this should be treated as the end of a free run if any | |
7910 // The current free range should be returned to the free lists | |
7911 // as one coalesced chunk. | |
7912 if (inFreeRange()) { | |
7913 flushCurFreeChunk(freeFinger(), | |
7914 pointer_delta(_limit, freeFinger())); | |
7915 assert(freeFinger() < _limit, "the finger pointeth off base"); | |
7916 if (CMSTraceSweeper) { | |
7917 gclog_or_tty->print("destructor:"); | |
7918 gclog_or_tty->print("Sweep:put_free_blk 0x%x ("SIZE_FORMAT") " | |
7919 "[coalesced:"SIZE_FORMAT"]\n", | |
7920 freeFinger(), pointer_delta(_limit, freeFinger()), | |
7921 lastFreeRangeCoalesced()); | |
7922 } | |
7923 } | |
7924 NOT_PRODUCT( | |
7925 if (Verbose && PrintGC) { | |
7926 gclog_or_tty->print("Collected "SIZE_FORMAT" objects, " | |
7927 SIZE_FORMAT " bytes", | |
7928 _numObjectsFreed, _numWordsFreed*sizeof(HeapWord)); | |
7929 gclog_or_tty->print_cr("\nLive "SIZE_FORMAT" objects, " | |
7930 SIZE_FORMAT" bytes " | |
7931 "Already free "SIZE_FORMAT" objects, "SIZE_FORMAT" bytes", | |
7932 _numObjectsLive, _numWordsLive*sizeof(HeapWord), | |
7933 _numObjectsAlreadyFree, _numWordsAlreadyFree*sizeof(HeapWord)); | |
7934 size_t totalBytes = (_numWordsFreed + _numWordsLive + _numWordsAlreadyFree) * | |
7935 sizeof(HeapWord); | |
7936 gclog_or_tty->print_cr("Total sweep: "SIZE_FORMAT" bytes", totalBytes); | |
7937 | |
7938 if (PrintCMSStatistics && CMSVerifyReturnedBytes) { | |
7939 size_t indexListReturnedBytes = _sp->sumIndexedFreeListArrayReturnedBytes(); | |
7940 size_t dictReturnedBytes = _sp->dictionary()->sumDictReturnedBytes(); | |
7941 size_t returnedBytes = indexListReturnedBytes + dictReturnedBytes; | |
7942 gclog_or_tty->print("Returned "SIZE_FORMAT" bytes", returnedBytes); | |
7943 gclog_or_tty->print(" Indexed List Returned "SIZE_FORMAT" bytes", | |
7944 indexListReturnedBytes); | |
7945 gclog_or_tty->print_cr(" Dictionary Returned "SIZE_FORMAT" bytes", | |
7946 dictReturnedBytes); | |
7947 } | |
7948 } | |
7949 ) | |
7950 // Now, in debug mode, just null out the sweep_limit | |
7951 NOT_PRODUCT(_sp->clear_sweep_limit();) | |
7952 if (CMSTraceSweeper) { | |
7953 gclog_or_tty->print("end of sweep\n================\n"); | |
7954 } | |
7955 } | |
7956 | |
7957 void SweepClosure::initialize_free_range(HeapWord* freeFinger, | |
7958 bool freeRangeInFreeLists) { | |
7959 if (CMSTraceSweeper) { | |
7960 gclog_or_tty->print("---- Start free range 0x%x with free block [%d] (%d)\n", | |
7961 freeFinger, _sp->block_size(freeFinger), | |
7962 freeRangeInFreeLists); | |
7963 } | |
7964 assert(!inFreeRange(), "Trampling existing free range"); | |
7965 set_inFreeRange(true); | |
7966 set_lastFreeRangeCoalesced(false); | |
7967 | |
7968 set_freeFinger(freeFinger); | |
7969 set_freeRangeInFreeLists(freeRangeInFreeLists); | |
7970 if (CMSTestInFreeList) { | |
7971 if (freeRangeInFreeLists) { | |
7972 FreeChunk* fc = (FreeChunk*) freeFinger; | |
7973 assert(fc->isFree(), "A chunk on the free list should be free."); | |
7974 assert(fc->size() > 0, "Free range should have a size"); | |
7975 assert(_sp->verifyChunkInFreeLists(fc), "Chunk is not in free lists"); | |
7976 } | |
7977 } | |
7978 } | |
7979 | |
7980 // Note that the sweeper runs concurrently with mutators. Thus, | |
7981 // it is possible for direct allocation in this generation to happen | |
7982 // in the middle of the sweep. Note that the sweeper also coalesces | |
7983 // contiguous free blocks. Thus, unless the sweeper and the allocator | |
7984 // synchronize appropriately freshly allocated blocks may get swept up. | |
7985 // This is accomplished by the sweeper locking the free lists while | |
7986 // it is sweeping. Thus blocks that are determined to be free are | |
7987 // indeed free. There is however one additional complication: | |
7988 // blocks that have been allocated since the final checkpoint and | |
7989 // mark, will not have been marked and so would be treated as | |
7990 // unreachable and swept up. To prevent this, the allocator marks | |
7991 // the bit map when allocating during the sweep phase. This leads, | |
7992 // however, to a further complication -- objects may have been allocated | |
7993 // but not yet initialized -- in the sense that the header isn't yet | |
7994 // installed. The sweeper can not then determine the size of the block | |
7995 // in order to skip over it. To deal with this case, we use a technique | |
7996 // (due to Printezis) to encode such uninitialized block sizes in the | |
7997 // bit map. Since the bit map uses a bit per every HeapWord, but the | |
7998 // CMS generation has a minimum object size of 3 HeapWords, it follows | |
7999 // that "normal marks" won't be adjacent in the bit map (there will | |
8000 // always be at least two 0 bits between successive 1 bits). We make use | |
8001 // of these "unused" bits to represent uninitialized blocks -- the bit | |
8002 // corresponding to the start of the uninitialized object and the next | |
8003 // bit are both set. Finally, a 1 bit marks the end of the object that | |
8004 // started with the two consecutive 1 bits to indicate its potentially | |
8005 // uninitialized state. | |
8006 | |
8007 size_t SweepClosure::do_blk_careful(HeapWord* addr) { | |
8008 FreeChunk* fc = (FreeChunk*)addr; | |
8009 size_t res; | |
8010 | |
1720
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8011 // Check if we are done sweeping. Below we check "addr >= _limit" rather |
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8012 // than "addr == _limit" because although _limit was a block boundary when |
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8013 // we started the sweep, it may no longer be one because heap expansion |
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8014 // may have caused us to coalesce the block ending at the address _limit |
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8015 // with a newly expanded chunk (this happens when _limit was set to the |
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8016 // previous _end of the space), so we may have stepped past _limit; see CR 6977970. |
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8017 if (addr >= _limit) { // we have swept up to or past the limit, do nothing more |
0 | 8018 assert(_limit >= _sp->bottom() && _limit <= _sp->end(), |
8019 "sweep _limit out of bounds"); | |
1720
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8020 assert(addr < _sp->end(), "addr out of bounds"); |
0 | 8021 // help the closure application finish |
1720
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8022 return pointer_delta(_sp->end(), addr); |
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8023 } |
5ed703250bff
6977970: CMS: concurrentMarkSweepGeneration.cpp:7947 assert(addr <= _limit) failed: sweep invariant
ysr
parents:
1716
diff
changeset
|
8024 assert(addr < _limit, "sweep invariant"); |
0 | 8025 |
8026 // check if we should yield | |
8027 do_yield_check(addr); | |
8028 if (fc->isFree()) { | |
8029 // Chunk that is already free | |
8030 res = fc->size(); | |
8031 doAlreadyFreeChunk(fc); | |
8032 debug_only(_sp->verifyFreeLists()); | |
8033 assert(res == fc->size(), "Don't expect the size to change"); | |
8034 NOT_PRODUCT( | |
8035 _numObjectsAlreadyFree++; | |
8036 _numWordsAlreadyFree += res; | |
8037 ) | |
8038 NOT_PRODUCT(_last_fc = fc;) | |
8039 } else if (!_bitMap->isMarked(addr)) { | |
8040 // Chunk is fresh garbage | |
8041 res = doGarbageChunk(fc); | |
8042 debug_only(_sp->verifyFreeLists()); | |
8043 NOT_PRODUCT( | |
8044 _numObjectsFreed++; | |
8045 _numWordsFreed += res; | |
8046 ) | |
8047 } else { | |
8048 // Chunk that is alive. | |
8049 res = doLiveChunk(fc); | |
8050 debug_only(_sp->verifyFreeLists()); | |
8051 NOT_PRODUCT( | |
8052 _numObjectsLive++; | |
8053 _numWordsLive += res; | |
8054 ) | |
8055 } | |
8056 return res; | |
8057 } | |
8058 | |
8059 // For the smart allocation, record following | |
8060 // split deaths - a free chunk is removed from its free list because | |
8061 // it is being split into two or more chunks. | |
8062 // split birth - a free chunk is being added to its free list because | |
8063 // a larger free chunk has been split and resulted in this free chunk. | |
8064 // coal death - a free chunk is being removed from its free list because | |
8065 // it is being coalesced into a large free chunk. | |
8066 // coal birth - a free chunk is being added to its free list because | |
8067 // it was created when two or more free chunks where coalesced into | |
8068 // this free chunk. | |
8069 // | |
8070 // These statistics are used to determine the desired number of free | |
8071 // chunks of a given size. The desired number is chosen to be relative | |
8072 // to the end of a CMS sweep. The desired number at the end of a sweep | |
8073 // is the | |
8074 // count-at-end-of-previous-sweep (an amount that was enough) | |
8075 // - count-at-beginning-of-current-sweep (the excess) | |
8076 // + split-births (gains in this size during interval) | |
8077 // - split-deaths (demands on this size during interval) | |
8078 // where the interval is from the end of one sweep to the end of the | |
8079 // next. | |
8080 // | |
8081 // When sweeping the sweeper maintains an accumulated chunk which is | |
8082 // the chunk that is made up of chunks that have been coalesced. That | |
8083 // will be termed the left-hand chunk. A new chunk of garbage that | |
8084 // is being considered for coalescing will be referred to as the | |
8085 // right-hand chunk. | |
8086 // | |
8087 // When making a decision on whether to coalesce a right-hand chunk with | |
8088 // the current left-hand chunk, the current count vs. the desired count | |
8089 // of the left-hand chunk is considered. Also if the right-hand chunk | |
8090 // is near the large chunk at the end of the heap (see | |
8091 // ConcurrentMarkSweepGeneration::isNearLargestChunk()), then the | |
8092 // left-hand chunk is coalesced. | |
8093 // | |
8094 // When making a decision about whether to split a chunk, the desired count | |
8095 // vs. the current count of the candidate to be split is also considered. | |
8096 // If the candidate is underpopulated (currently fewer chunks than desired) | |
8097 // a chunk of an overpopulated (currently more chunks than desired) size may | |
8098 // be chosen. The "hint" associated with a free list, if non-null, points | |
8099 // to a free list which may be overpopulated. | |
8100 // | |
8101 | |
8102 void SweepClosure::doAlreadyFreeChunk(FreeChunk* fc) { | |
8103 size_t size = fc->size(); | |
8104 // Chunks that cannot be coalesced are not in the | |
8105 // free lists. | |
8106 if (CMSTestInFreeList && !fc->cantCoalesce()) { | |
8107 assert(_sp->verifyChunkInFreeLists(fc), | |
8108 "free chunk should be in free lists"); | |
8109 } | |
8110 // a chunk that is already free, should not have been | |
8111 // marked in the bit map | |
8112 HeapWord* addr = (HeapWord*) fc; | |
8113 assert(!_bitMap->isMarked(addr), "free chunk should be unmarked"); | |
8114 // Verify that the bit map has no bits marked between | |
8115 // addr and purported end of this block. | |
8116 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); | |
8117 | |
8118 // Some chunks cannot be coalesced in under any circumstances. | |
8119 // See the definition of cantCoalesce(). | |
8120 if (!fc->cantCoalesce()) { | |
8121 // This chunk can potentially be coalesced. | |
8122 if (_sp->adaptive_freelists()) { | |
8123 // All the work is done in | |
8124 doPostIsFreeOrGarbageChunk(fc, size); | |
8125 } else { // Not adaptive free lists | |
8126 // this is a free chunk that can potentially be coalesced by the sweeper; | |
8127 if (!inFreeRange()) { | |
8128 // if the next chunk is a free block that can't be coalesced | |
8129 // it doesn't make sense to remove this chunk from the free lists | |
8130 FreeChunk* nextChunk = (FreeChunk*)(addr + size); | |
8131 assert((HeapWord*)nextChunk <= _limit, "sweep invariant"); | |
8132 if ((HeapWord*)nextChunk < _limit && // there's a next chunk... | |
8133 nextChunk->isFree() && // which is free... | |
8134 nextChunk->cantCoalesce()) { // ... but cant be coalesced | |
8135 // nothing to do | |
8136 } else { | |
8137 // Potentially the start of a new free range: | |
8138 // Don't eagerly remove it from the free lists. | |
8139 // No need to remove it if it will just be put | |
8140 // back again. (Also from a pragmatic point of view | |
8141 // if it is a free block in a region that is beyond | |
8142 // any allocated blocks, an assertion will fail) | |
8143 // Remember the start of a free run. | |
8144 initialize_free_range(addr, true); | |
8145 // end - can coalesce with next chunk | |
8146 } | |
8147 } else { | |
8148 // the midst of a free range, we are coalescing | |
8149 debug_only(record_free_block_coalesced(fc);) | |
8150 if (CMSTraceSweeper) { | |
8151 gclog_or_tty->print(" -- pick up free block 0x%x (%d)\n", fc, size); | |
8152 } | |
8153 // remove it from the free lists | |
8154 _sp->removeFreeChunkFromFreeLists(fc); | |
8155 set_lastFreeRangeCoalesced(true); | |
8156 // If the chunk is being coalesced and the current free range is | |
8157 // in the free lists, remove the current free range so that it | |
8158 // will be returned to the free lists in its entirety - all | |
8159 // the coalesced pieces included. | |
8160 if (freeRangeInFreeLists()) { | |
8161 FreeChunk* ffc = (FreeChunk*) freeFinger(); | |
8162 assert(ffc->size() == pointer_delta(addr, freeFinger()), | |
8163 "Size of free range is inconsistent with chunk size."); | |
8164 if (CMSTestInFreeList) { | |
8165 assert(_sp->verifyChunkInFreeLists(ffc), | |
8166 "free range is not in free lists"); | |
8167 } | |
8168 _sp->removeFreeChunkFromFreeLists(ffc); | |
8169 set_freeRangeInFreeLists(false); | |
8170 } | |
8171 } | |
8172 } | |
8173 } else { | |
8174 // Code path common to both original and adaptive free lists. | |
8175 | |
8176 // cant coalesce with previous block; this should be treated | |
8177 // as the end of a free run if any | |
8178 if (inFreeRange()) { | |
8179 // we kicked some butt; time to pick up the garbage | |
8180 assert(freeFinger() < addr, "the finger pointeth off base"); | |
8181 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger())); | |
8182 } | |
8183 // else, nothing to do, just continue | |
8184 } | |
8185 } | |
8186 | |
8187 size_t SweepClosure::doGarbageChunk(FreeChunk* fc) { | |
8188 // This is a chunk of garbage. It is not in any free list. | |
8189 // Add it to a free list or let it possibly be coalesced into | |
8190 // a larger chunk. | |
8191 HeapWord* addr = (HeapWord*) fc; | |
8192 size_t size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); | |
8193 | |
8194 if (_sp->adaptive_freelists()) { | |
8195 // Verify that the bit map has no bits marked between | |
8196 // addr and purported end of just dead object. | |
8197 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); | |
8198 | |
8199 doPostIsFreeOrGarbageChunk(fc, size); | |
8200 } else { | |
8201 if (!inFreeRange()) { | |
8202 // start of a new free range | |
8203 assert(size > 0, "A free range should have a size"); | |
8204 initialize_free_range(addr, false); | |
8205 | |
8206 } else { | |
8207 // this will be swept up when we hit the end of the | |
8208 // free range | |
8209 if (CMSTraceSweeper) { | |
8210 gclog_or_tty->print(" -- pick up garbage 0x%x (%d) \n", fc, size); | |
8211 } | |
8212 // If the chunk is being coalesced and the current free range is | |
8213 // in the free lists, remove the current free range so that it | |
8214 // will be returned to the free lists in its entirety - all | |
8215 // the coalesced pieces included. | |
8216 if (freeRangeInFreeLists()) { | |
8217 FreeChunk* ffc = (FreeChunk*)freeFinger(); | |
8218 assert(ffc->size() == pointer_delta(addr, freeFinger()), | |
8219 "Size of free range is inconsistent with chunk size."); | |
8220 if (CMSTestInFreeList) { | |
8221 assert(_sp->verifyChunkInFreeLists(ffc), | |
8222 "free range is not in free lists"); | |
8223 } | |
8224 _sp->removeFreeChunkFromFreeLists(ffc); | |
8225 set_freeRangeInFreeLists(false); | |
8226 } | |
8227 set_lastFreeRangeCoalesced(true); | |
8228 } | |
8229 // this will be swept up when we hit the end of the free range | |
8230 | |
8231 // Verify that the bit map has no bits marked between | |
8232 // addr and purported end of just dead object. | |
8233 _bitMap->verifyNoOneBitsInRange(addr + 1, addr + size); | |
8234 } | |
8235 return size; | |
8236 } | |
8237 | |
8238 size_t SweepClosure::doLiveChunk(FreeChunk* fc) { | |
8239 HeapWord* addr = (HeapWord*) fc; | |
8240 // The sweeper has just found a live object. Return any accumulated | |
8241 // left hand chunk to the free lists. | |
8242 if (inFreeRange()) { | |
8243 if (_sp->adaptive_freelists()) { | |
8244 flushCurFreeChunk(freeFinger(), | |
8245 pointer_delta(addr, freeFinger())); | |
8246 } else { // not adaptive freelists | |
8247 set_inFreeRange(false); | |
8248 // Add the free range back to the free list if it is not already | |
8249 // there. | |
8250 if (!freeRangeInFreeLists()) { | |
8251 assert(freeFinger() < addr, "the finger pointeth off base"); | |
8252 if (CMSTraceSweeper) { | |
8253 gclog_or_tty->print("Sweep:put_free_blk 0x%x (%d) " | |
8254 "[coalesced:%d]\n", | |
8255 freeFinger(), pointer_delta(addr, freeFinger()), | |
8256 lastFreeRangeCoalesced()); | |
8257 } | |
8258 _sp->addChunkAndRepairOffsetTable(freeFinger(), | |
8259 pointer_delta(addr, freeFinger()), lastFreeRangeCoalesced()); | |
8260 } | |
8261 } | |
8262 } | |
8263 | |
8264 // Common code path for original and adaptive free lists. | |
8265 | |
8266 // this object is live: we'd normally expect this to be | |
8267 // an oop, and like to assert the following: | |
8268 // assert(oop(addr)->is_oop(), "live block should be an oop"); | |
8269 // However, as we commented above, this may be an object whose | |
8270 // header hasn't yet been initialized. | |
8271 size_t size; | |
8272 assert(_bitMap->isMarked(addr), "Tautology for this control point"); | |
8273 if (_bitMap->isMarked(addr + 1)) { | |
8274 // Determine the size from the bit map, rather than trying to | |
8275 // compute it from the object header. | |
8276 HeapWord* nextOneAddr = _bitMap->getNextMarkedWordAddress(addr + 2); | |
8277 size = pointer_delta(nextOneAddr + 1, addr); | |
8278 assert(size == CompactibleFreeListSpace::adjustObjectSize(size), | |
8279 "alignment problem"); | |
8280 | |
8281 #ifdef DEBUG | |
187 | 8282 if (oop(addr)->klass_or_null() != NULL && |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
8283 ( !_collector->should_unload_classes() |
518
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8284 || (oop(addr)->is_parsable()) && |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8285 oop(addr)->is_conc_safe())) { |
0 | 8286 // Ignore mark word because we are running concurrent with mutators |
8287 assert(oop(addr)->is_oop(true), "live block should be an oop"); | |
518
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8288 // is_conc_safe is checked before performing this assertion |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8289 // because an object that is not is_conc_safe may yet have |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8290 // the return from size() correct. |
0 | 8291 assert(size == |
8292 CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()), | |
8293 "P-mark and computed size do not agree"); | |
8294 } | |
8295 #endif | |
8296 | |
8297 } else { | |
8298 // This should be an initialized object that's alive. | |
187 | 8299 assert(oop(addr)->klass_or_null() != NULL && |
94
0834225a7916
6634032: CMS: Need CMSInitiatingPermOccupancyFraction for perm, divorcing from CMSInitiatingOccupancyFraction
ysr
parents:
9
diff
changeset
|
8300 (!_collector->should_unload_classes() |
0 | 8301 || oop(addr)->is_parsable()), |
8302 "Should be an initialized object"); | |
518
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8303 // Note that there are objects used during class redefinition |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8304 // (e.g., merge_cp in VM_RedefineClasses::merge_cp_and_rewrite() |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8305 // which are discarded with their is_conc_safe state still |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8306 // false. These object may be floating garbage so may be |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8307 // seen here. If they are floating garbage their size |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8308 // should be attainable from their klass. Do not that |
0af8b0718fc9
6692899: CMS: many vm.parallel_class_loading tests fail with assert "missing Printezis mark"
jmasa
parents:
517
diff
changeset
|
8309 // is_conc_safe() is true for oop(addr). |
0 | 8310 // Ignore mark word because we are running concurrent with mutators |
8311 assert(oop(addr)->is_oop(true), "live block should be an oop"); | |
8312 // Verify that the bit map has no bits marked between | |
8313 // addr and purported end of this block. | |
8314 size = CompactibleFreeListSpace::adjustObjectSize(oop(addr)->size()); | |
8315 assert(size >= 3, "Necessary for Printezis marks to work"); | |
8316 assert(!_bitMap->isMarked(addr+1), "Tautology for this control point"); | |
8317 DEBUG_ONLY(_bitMap->verifyNoOneBitsInRange(addr+2, addr+size);) | |
8318 } | |
8319 return size; | |
8320 } | |
8321 | |
8322 void SweepClosure::doPostIsFreeOrGarbageChunk(FreeChunk* fc, | |
8323 size_t chunkSize) { | |
8324 // doPostIsFreeOrGarbageChunk() should only be called in the smart allocation | |
8325 // scheme. | |
8326 bool fcInFreeLists = fc->isFree(); | |
8327 assert(_sp->adaptive_freelists(), "Should only be used in this case."); | |
8328 assert((HeapWord*)fc <= _limit, "sweep invariant"); | |
8329 if (CMSTestInFreeList && fcInFreeLists) { | |
8330 assert(_sp->verifyChunkInFreeLists(fc), | |
8331 "free chunk is not in free lists"); | |
8332 } | |
8333 | |
8334 | |
8335 if (CMSTraceSweeper) { | |
8336 gclog_or_tty->print_cr(" -- pick up another chunk at 0x%x (%d)", fc, chunkSize); | |
8337 } | |
8338 | |
8339 HeapWord* addr = (HeapWord*) fc; | |
8340 | |
8341 bool coalesce; | |
8342 size_t left = pointer_delta(addr, freeFinger()); | |
8343 size_t right = chunkSize; | |
8344 switch (FLSCoalescePolicy) { | |
8345 // numeric value forms a coalition aggressiveness metric | |
8346 case 0: { // never coalesce | |
8347 coalesce = false; | |
8348 break; | |
8349 } | |
8350 case 1: { // coalesce if left & right chunks on overpopulated lists | |
8351 coalesce = _sp->coalOverPopulated(left) && | |
8352 _sp->coalOverPopulated(right); | |
8353 break; | |
8354 } | |
8355 case 2: { // coalesce if left chunk on overpopulated list (default) | |
8356 coalesce = _sp->coalOverPopulated(left); | |
8357 break; | |
8358 } | |
8359 case 3: { // coalesce if left OR right chunk on overpopulated list | |
8360 coalesce = _sp->coalOverPopulated(left) || | |
8361 _sp->coalOverPopulated(right); | |
8362 break; | |
8363 } | |
8364 case 4: { // always coalesce | |
8365 coalesce = true; | |
8366 break; | |
8367 } | |
8368 default: | |
8369 ShouldNotReachHere(); | |
8370 } | |
8371 | |
8372 // Should the current free range be coalesced? | |
8373 // If the chunk is in a free range and either we decided to coalesce above | |
8374 // or the chunk is near the large block at the end of the heap | |
8375 // (isNearLargestChunk() returns true), then coalesce this chunk. | |
8376 bool doCoalesce = inFreeRange() && | |
8377 (coalesce || _g->isNearLargestChunk((HeapWord*)fc)); | |
8378 if (doCoalesce) { | |
8379 // Coalesce the current free range on the left with the new | |
8380 // chunk on the right. If either is on a free list, | |
8381 // it must be removed from the list and stashed in the closure. | |
8382 if (freeRangeInFreeLists()) { | |
8383 FreeChunk* ffc = (FreeChunk*)freeFinger(); | |
8384 assert(ffc->size() == pointer_delta(addr, freeFinger()), | |
8385 "Size of free range is inconsistent with chunk size."); | |
8386 if (CMSTestInFreeList) { | |
8387 assert(_sp->verifyChunkInFreeLists(ffc), | |
8388 "Chunk is not in free lists"); | |
8389 } | |
8390 _sp->coalDeath(ffc->size()); | |
8391 _sp->removeFreeChunkFromFreeLists(ffc); | |
8392 set_freeRangeInFreeLists(false); | |
8393 } | |
8394 if (fcInFreeLists) { | |
8395 _sp->coalDeath(chunkSize); | |
8396 assert(fc->size() == chunkSize, | |
8397 "The chunk has the wrong size or is not in the free lists"); | |
8398 _sp->removeFreeChunkFromFreeLists(fc); | |
8399 } | |
8400 set_lastFreeRangeCoalesced(true); | |
8401 } else { // not in a free range and/or should not coalesce | |
8402 // Return the current free range and start a new one. | |
8403 if (inFreeRange()) { | |
8404 // In a free range but cannot coalesce with the right hand chunk. | |
8405 // Put the current free range into the free lists. | |
8406 flushCurFreeChunk(freeFinger(), | |
8407 pointer_delta(addr, freeFinger())); | |
8408 } | |
8409 // Set up for new free range. Pass along whether the right hand | |
8410 // chunk is in the free lists. | |
8411 initialize_free_range((HeapWord*)fc, fcInFreeLists); | |
8412 } | |
8413 } | |
8414 void SweepClosure::flushCurFreeChunk(HeapWord* chunk, size_t size) { | |
8415 assert(inFreeRange(), "Should only be called if currently in a free range."); | |
8416 assert(size > 0, | |
8417 "A zero sized chunk cannot be added to the free lists."); | |
8418 if (!freeRangeInFreeLists()) { | |
8419 if(CMSTestInFreeList) { | |
8420 FreeChunk* fc = (FreeChunk*) chunk; | |
8421 fc->setSize(size); | |
8422 assert(!_sp->verifyChunkInFreeLists(fc), | |
8423 "chunk should not be in free lists yet"); | |
8424 } | |
8425 if (CMSTraceSweeper) { | |
8426 gclog_or_tty->print_cr(" -- add free block 0x%x (%d) to free lists", | |
8427 chunk, size); | |
8428 } | |
8429 // A new free range is going to be starting. The current | |
8430 // free range has not been added to the free lists yet or | |
8431 // was removed so add it back. | |
8432 // If the current free range was coalesced, then the death | |
8433 // of the free range was recorded. Record a birth now. | |
8434 if (lastFreeRangeCoalesced()) { | |
8435 _sp->coalBirth(size); | |
8436 } | |
8437 _sp->addChunkAndRepairOffsetTable(chunk, size, | |
8438 lastFreeRangeCoalesced()); | |
8439 } | |
8440 set_inFreeRange(false); | |
8441 set_freeRangeInFreeLists(false); | |
8442 } | |
8443 | |
8444 // We take a break if we've been at this for a while, | |
8445 // so as to avoid monopolizing the locks involved. | |
8446 void SweepClosure::do_yield_work(HeapWord* addr) { | |
8447 // Return current free chunk being used for coalescing (if any) | |
8448 // to the appropriate freelist. After yielding, the next | |
8449 // free block encountered will start a coalescing range of | |
8450 // free blocks. If the next free block is adjacent to the | |
8451 // chunk just flushed, they will need to wait for the next | |
8452 // sweep to be coalesced. | |
8453 if (inFreeRange()) { | |
8454 flushCurFreeChunk(freeFinger(), pointer_delta(addr, freeFinger())); | |
8455 } | |
8456 | |
8457 // First give up the locks, then yield, then re-lock. | |
8458 // We should probably use a constructor/destructor idiom to | |
8459 // do this unlock/lock or modify the MutexUnlocker class to | |
8460 // serve our purpose. XXX | |
8461 assert_lock_strong(_bitMap->lock()); | |
8462 assert_lock_strong(_freelistLock); | |
8463 assert(ConcurrentMarkSweepThread::cms_thread_has_cms_token(), | |
8464 "CMS thread should hold CMS token"); | |
8465 _bitMap->lock()->unlock(); | |
8466 _freelistLock->unlock(); | |
8467 ConcurrentMarkSweepThread::desynchronize(true); | |
8468 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
8469 _collector->stopTimer(); | |
8470 GCPauseTimer p(_collector->size_policy()->concurrent_timer_ptr()); | |
8471 if (PrintCMSStatistics != 0) { | |
8472 _collector->incrementYields(); | |
8473 } | |
8474 _collector->icms_wait(); | |
8475 | |
8476 // See the comment in coordinator_yield() | |
8477 for (unsigned i = 0; i < CMSYieldSleepCount && | |
8478 ConcurrentMarkSweepThread::should_yield() && | |
8479 !CMSCollector::foregroundGCIsActive(); ++i) { | |
8480 os::sleep(Thread::current(), 1, false); | |
8481 ConcurrentMarkSweepThread::acknowledge_yield_request(); | |
8482 } | |
8483 | |
8484 ConcurrentMarkSweepThread::synchronize(true); | |
8485 _freelistLock->lock(); | |
8486 _bitMap->lock()->lock_without_safepoint_check(); | |
8487 _collector->startTimer(); | |
8488 } | |
8489 | |
8490 #ifndef PRODUCT | |
8491 // This is actually very useful in a product build if it can | |
8492 // be called from the debugger. Compile it into the product | |
8493 // as needed. | |
8494 bool debug_verifyChunkInFreeLists(FreeChunk* fc) { | |
8495 return debug_cms_space->verifyChunkInFreeLists(fc); | |
8496 } | |
8497 | |
8498 void SweepClosure::record_free_block_coalesced(FreeChunk* fc) const { | |
8499 if (CMSTraceSweeper) { | |
8500 gclog_or_tty->print("Sweep:coal_free_blk 0x%x (%d)\n", fc, fc->size()); | |
8501 } | |
8502 } | |
8503 #endif | |
8504 | |
8505 // CMSIsAliveClosure | |
8506 bool CMSIsAliveClosure::do_object_b(oop obj) { | |
8507 HeapWord* addr = (HeapWord*)obj; | |
8508 return addr != NULL && | |
8509 (!_span.contains(addr) || _bit_map->isMarked(addr)); | |
8510 } | |
8511 | |
935 | 8512 CMSKeepAliveClosure::CMSKeepAliveClosure( CMSCollector* collector, |
8513 MemRegion span, | |
8514 CMSBitMap* bit_map, CMSMarkStack* mark_stack, | |
8515 CMSMarkStack* revisit_stack, bool cpc): | |
8516 KlassRememberingOopClosure(collector, NULL, revisit_stack), | |
8517 _span(span), | |
8518 _bit_map(bit_map), | |
8519 _mark_stack(mark_stack), | |
8520 _concurrent_precleaning(cpc) { | |
8521 assert(!_span.is_empty(), "Empty span could spell trouble"); | |
8522 } | |
8523 | |
8524 | |
0 | 8525 // CMSKeepAliveClosure: the serial version |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8526 void CMSKeepAliveClosure::do_oop(oop obj) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8527 HeapWord* addr = (HeapWord*)obj; |
0 | 8528 if (_span.contains(addr) && |
8529 !_bit_map->isMarked(addr)) { | |
8530 _bit_map->mark(addr); | |
8531 bool simulate_overflow = false; | |
8532 NOT_PRODUCT( | |
8533 if (CMSMarkStackOverflowALot && | |
8534 _collector->simulate_overflow()) { | |
8535 // simulate a stack overflow | |
8536 simulate_overflow = true; | |
8537 } | |
8538 ) | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8539 if (simulate_overflow || !_mark_stack->push(obj)) { |
452
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8540 if (_concurrent_precleaning) { |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8541 // We dirty the overflown object and let the remark |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8542 // phase deal with it. |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8543 assert(_collector->overflow_list_is_empty(), "Error"); |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8544 // In the case of object arrays, we need to dirty all of |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8545 // the cards that the object spans. No locking or atomics |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8546 // are needed since no one else can be mutating the mod union |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8547 // table. |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8548 if (obj->is_objArray()) { |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8549 size_t sz = obj->size(); |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8550 HeapWord* end_card_addr = |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8551 (HeapWord*)round_to((intptr_t)(addr+sz), CardTableModRefBS::card_size); |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8552 MemRegion redirty_range = MemRegion(addr, end_card_addr); |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8553 assert(!redirty_range.is_empty(), "Arithmetical tautology"); |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8554 _collector->_modUnionTable.mark_range(redirty_range); |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8555 } else { |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8556 _collector->_modUnionTable.mark(addr); |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8557 } |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8558 _collector->_ser_kac_preclean_ovflw++; |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8559 } else { |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8560 _collector->push_on_overflow_list(obj); |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8561 _collector->_ser_kac_ovflw++; |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8562 } |
0 | 8563 } |
8564 } | |
8565 } | |
8566 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8567 void CMSKeepAliveClosure::do_oop(oop* p) { CMSKeepAliveClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8568 void CMSKeepAliveClosure::do_oop(narrowOop* p) { CMSKeepAliveClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8569 |
0 | 8570 // CMSParKeepAliveClosure: a parallel version of the above. |
8571 // The work queues are private to each closure (thread), | |
8572 // but (may be) available for stealing by other threads. | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8573 void CMSParKeepAliveClosure::do_oop(oop obj) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8574 HeapWord* addr = (HeapWord*)obj; |
0 | 8575 if (_span.contains(addr) && |
8576 !_bit_map->isMarked(addr)) { | |
8577 // In general, during recursive tracing, several threads | |
8578 // may be concurrently getting here; the first one to | |
8579 // "tag" it, claims it. | |
8580 if (_bit_map->par_mark(addr)) { | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8581 bool res = _work_queue->push(obj); |
0 | 8582 assert(res, "Low water mark should be much less than capacity"); |
8583 // Do a recursive trim in the hope that this will keep | |
8584 // stack usage lower, but leave some oops for potential stealers | |
8585 trim_queue(_low_water_mark); | |
8586 } // Else, another thread got there first | |
8587 } | |
8588 } | |
8589 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8590 void CMSParKeepAliveClosure::do_oop(oop* p) { CMSParKeepAliveClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8591 void CMSParKeepAliveClosure::do_oop(narrowOop* p) { CMSParKeepAliveClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8592 |
0 | 8593 void CMSParKeepAliveClosure::trim_queue(uint max) { |
8594 while (_work_queue->size() > max) { | |
8595 oop new_oop; | |
8596 if (_work_queue->pop_local(new_oop)) { | |
8597 assert(new_oop != NULL && new_oop->is_oop(), "Expected an oop"); | |
8598 assert(_bit_map->isMarked((HeapWord*)new_oop), | |
8599 "no white objects on this stack!"); | |
8600 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); | |
8601 // iterate over the oops in this oop, marking and pushing | |
8602 // the ones in CMS heap (i.e. in _span). | |
8603 new_oop->oop_iterate(&_mark_and_push); | |
8604 } | |
8605 } | |
8606 } | |
8607 | |
935 | 8608 CMSInnerParMarkAndPushClosure::CMSInnerParMarkAndPushClosure( |
8609 CMSCollector* collector, | |
8610 MemRegion span, CMSBitMap* bit_map, | |
8611 CMSMarkStack* revisit_stack, | |
8612 OopTaskQueue* work_queue): | |
8613 Par_KlassRememberingOopClosure(collector, NULL, revisit_stack), | |
8614 _span(span), | |
8615 _bit_map(bit_map), | |
8616 _work_queue(work_queue) { } | |
8617 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8618 void CMSInnerParMarkAndPushClosure::do_oop(oop obj) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8619 HeapWord* addr = (HeapWord*)obj; |
0 | 8620 if (_span.contains(addr) && |
8621 !_bit_map->isMarked(addr)) { | |
8622 if (_bit_map->par_mark(addr)) { | |
8623 bool simulate_overflow = false; | |
8624 NOT_PRODUCT( | |
8625 if (CMSMarkStackOverflowALot && | |
8626 _collector->par_simulate_overflow()) { | |
8627 // simulate a stack overflow | |
8628 simulate_overflow = true; | |
8629 } | |
8630 ) | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8631 if (simulate_overflow || !_work_queue->push(obj)) { |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8632 _collector->par_push_on_overflow_list(obj); |
0 | 8633 _collector->_par_kac_ovflw++; |
8634 } | |
8635 } // Else another thread got there already | |
8636 } | |
8637 } | |
8638 | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8639 void CMSInnerParMarkAndPushClosure::do_oop(oop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8640 void CMSInnerParMarkAndPushClosure::do_oop(narrowOop* p) { CMSInnerParMarkAndPushClosure::do_oop_work(p); } |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8641 |
0 | 8642 ////////////////////////////////////////////////////////////////// |
8643 // CMSExpansionCause ///////////////////////////// | |
8644 ////////////////////////////////////////////////////////////////// | |
8645 const char* CMSExpansionCause::to_string(CMSExpansionCause::Cause cause) { | |
8646 switch (cause) { | |
8647 case _no_expansion: | |
8648 return "No expansion"; | |
8649 case _satisfy_free_ratio: | |
8650 return "Free ratio"; | |
8651 case _satisfy_promotion: | |
8652 return "Satisfy promotion"; | |
8653 case _satisfy_allocation: | |
8654 return "allocation"; | |
8655 case _allocate_par_lab: | |
8656 return "Par LAB"; | |
8657 case _allocate_par_spooling_space: | |
8658 return "Par Spooling Space"; | |
8659 case _adaptive_size_policy: | |
8660 return "Ergonomics"; | |
8661 default: | |
8662 return "unknown"; | |
8663 } | |
8664 } | |
8665 | |
8666 void CMSDrainMarkingStackClosure::do_void() { | |
8667 // the max number to take from overflow list at a time | |
8668 const size_t num = _mark_stack->capacity()/4; | |
452
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8669 assert(!_concurrent_precleaning || _collector->overflow_list_is_empty(), |
00b023ae2d78
6722113: CMS: Incorrect overflow handling during precleaning of Reference lists
ysr
parents:
360
diff
changeset
|
8670 "Overflow list should be NULL during concurrent phases"); |
0 | 8671 while (!_mark_stack->isEmpty() || |
8672 // if stack is empty, check the overflow list | |
8673 _collector->take_from_overflow_list(num, _mark_stack)) { | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8674 oop obj = _mark_stack->pop(); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8675 HeapWord* addr = (HeapWord*)obj; |
0 | 8676 assert(_span.contains(addr), "Should be within span"); |
8677 assert(_bit_map->isMarked(addr), "Should be marked"); | |
113
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8678 assert(obj->is_oop(), "Should be an oop"); |
ba764ed4b6f2
6420645: Create a vm that uses compressed oops for up to 32gb heapsizes
coleenp
parents:
94
diff
changeset
|
8679 obj->oop_iterate(_keep_alive); |
0 | 8680 } |
8681 } | |
8682 | |
8683 void CMSParDrainMarkingStackClosure::do_void() { | |
8684 // drain queue | |
8685 trim_queue(0); | |
8686 } | |
8687 | |
8688 // Trim our work_queue so its length is below max at return | |
8689 void CMSParDrainMarkingStackClosure::trim_queue(uint max) { | |
8690 while (_work_queue->size() > max) { | |
8691 oop new_oop; | |
8692 if (_work_queue->pop_local(new_oop)) { | |
8693 assert(new_oop->is_oop(), "Expected an oop"); | |
8694 assert(_bit_map->isMarked((HeapWord*)new_oop), | |
8695 "no white objects on this stack!"); | |
8696 assert(_span.contains((HeapWord*)new_oop), "Out of bounds oop"); | |
8697 // iterate over the oops in this oop, marking and pushing | |
8698 // the ones in CMS heap (i.e. in _span). | |
8699 new_oop->oop_iterate(&_mark_and_push); | |
8700 } | |
8701 } | |
8702 } | |
8703 | |
8704 //////////////////////////////////////////////////////////////////// | |
8705 // Support for Marking Stack Overflow list handling and related code | |
8706 //////////////////////////////////////////////////////////////////// | |
8707 // Much of the following code is similar in shape and spirit to the | |
8708 // code used in ParNewGC. We should try and share that code | |
8709 // as much as possible in the future. | |
8710 | |
8711 #ifndef PRODUCT | |
8712 // Debugging support for CMSStackOverflowALot | |
8713 | |
8714 // It's OK to call this multi-threaded; the worst thing | |
8715 // that can happen is that we'll get a bunch of closely | |
8716 // spaced simulated oveflows, but that's OK, in fact | |
8717 // probably good as it would exercise the overflow code | |
8718 // under contention. | |
8719 bool CMSCollector::simulate_overflow() { | |
8720 if (_overflow_counter-- <= 0) { // just being defensive | |
8721 _overflow_counter = CMSMarkStackOverflowInterval; | |
8722 return true; | |
8723 } else { | |
8724 return false; | |
8725 } | |
8726 } | |
8727 | |
8728 bool CMSCollector::par_simulate_overflow() { | |
8729 return simulate_overflow(); | |
8730 } | |
8731 #endif | |
8732 | |
8733 // Single-threaded | |
8734 bool CMSCollector::take_from_overflow_list(size_t num, CMSMarkStack* stack) { | |
8735 assert(stack->isEmpty(), "Expected precondition"); | |
8736 assert(stack->capacity() > num, "Shouldn't bite more than can chew"); | |
8737 size_t i = num; | |
8738 oop cur = _overflow_list; | |
8739 const markOop proto = markOopDesc::prototype(); | |
534 | 8740 NOT_PRODUCT(ssize_t n = 0;) |
0 | 8741 for (oop next; i > 0 && cur != NULL; cur = next, i--) { |
8742 next = oop(cur->mark()); | |
8743 cur->set_mark(proto); // until proven otherwise | |
8744 assert(cur->is_oop(), "Should be an oop"); | |
8745 bool res = stack->push(cur); | |
8746 assert(res, "Bit off more than can chew?"); | |
8747 NOT_PRODUCT(n++;) | |
8748 } | |
8749 _overflow_list = cur; | |
8750 #ifndef PRODUCT | |
8751 assert(_num_par_pushes >= n, "Too many pops?"); | |
8752 _num_par_pushes -=n; | |
8753 #endif | |
8754 return !stack->isEmpty(); | |
8755 } | |
8756 | |
534 | 8757 #define BUSY (oop(0x1aff1aff)) |
8758 // (MT-safe) Get a prefix of at most "num" from the list. | |
8759 // The overflow list is chained through the mark word of | |
8760 // each object in the list. We fetch the entire list, | |
8761 // break off a prefix of the right size and return the | |
8762 // remainder. If other threads try to take objects from | |
8763 // the overflow list at that time, they will wait for | |
8764 // some time to see if data becomes available. If (and | |
8765 // only if) another thread places one or more object(s) | |
8766 // on the global list before we have returned the suffix | |
8767 // to the global list, we will walk down our local list | |
8768 // to find its end and append the global list to | |
8769 // our suffix before returning it. This suffix walk can | |
8770 // prove to be expensive (quadratic in the amount of traffic) | |
8771 // when there are many objects in the overflow list and | |
8772 // there is much producer-consumer contention on the list. | |
8773 // *NOTE*: The overflow list manipulation code here and | |
8774 // in ParNewGeneration:: are very similar in shape, | |
8775 // except that in the ParNew case we use the old (from/eden) | |
8776 // copy of the object to thread the list via its klass word. | |
8777 // Because of the common code, if you make any changes in | |
8778 // the code below, please check the ParNew version to see if | |
8779 // similar changes might be needed. | |
8780 // CR 6797058 has been filed to consolidate the common code. | |
0 | 8781 bool CMSCollector::par_take_from_overflow_list(size_t num, |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
8782 OopTaskQueue* work_q, |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
8783 int no_of_gc_threads) { |
534 | 8784 assert(work_q->size() == 0, "First empty local work queue"); |
0 | 8785 assert(num < work_q->max_elems(), "Can't bite more than we can chew"); |
8786 if (_overflow_list == NULL) { | |
8787 return false; | |
8788 } | |
8789 // Grab the entire list; we'll put back a suffix | |
534 | 8790 oop prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); |
8791 Thread* tid = Thread::current(); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
8792 // Before "no_of_gc_threads" was introduced CMSOverflowSpinCount was |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
8793 // set to ParallelGCThreads. |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1777
diff
changeset
|
8794 size_t CMSOverflowSpinCount = (size_t) no_of_gc_threads; // was ParallelGCThreads; |
534 | 8795 size_t sleep_time_millis = MAX2((size_t)1, num/100); |
8796 // If the list is busy, we spin for a short while, | |
8797 // sleeping between attempts to get the list. | |
8798 for (size_t spin = 0; prefix == BUSY && spin < CMSOverflowSpinCount; spin++) { | |
8799 os::sleep(tid, sleep_time_millis, false); | |
8800 if (_overflow_list == NULL) { | |
8801 // Nothing left to take | |
8802 return false; | |
8803 } else if (_overflow_list != BUSY) { | |
8804 // Try and grab the prefix | |
8805 prefix = (oop)Atomic::xchg_ptr(BUSY, &_overflow_list); | |
8806 } | |
8807 } | |
8808 // If the list was found to be empty, or we spun long | |
8809 // enough, we give up and return empty-handed. If we leave | |
8810 // the list in the BUSY state below, it must be the case that | |
8811 // some other thread holds the overflow list and will set it | |
8812 // to a non-BUSY state in the future. | |
8813 if (prefix == NULL || prefix == BUSY) { | |
8814 // Nothing to take or waited long enough | |
8815 if (prefix == NULL) { | |
8816 // Write back the NULL in case we overwrote it with BUSY above | |
8817 // and it is still the same value. | |
8818 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); | |
8819 } | |
8820 return false; | |
8821 } | |
8822 assert(prefix != NULL && prefix != BUSY, "Error"); | |
0 | 8823 size_t i = num; |
8824 oop cur = prefix; | |
534 | 8825 // Walk down the first "num" objects, unless we reach the end. |
0 | 8826 for (; i > 1 && cur->mark() != NULL; cur = oop(cur->mark()), i--); |
534 | 8827 if (cur->mark() == NULL) { |
8828 // We have "num" or fewer elements in the list, so there | |
8829 // is nothing to return to the global list. | |
8830 // Write back the NULL in lieu of the BUSY we wrote | |
8831 // above, if it is still the same value. | |
8832 if (_overflow_list == BUSY) { | |
8833 (void) Atomic::cmpxchg_ptr(NULL, &_overflow_list, BUSY); | |
8834 } | |
8835 } else { | |
8836 // Chop off the suffix and rerturn it to the global list. | |
8837 assert(cur->mark() != BUSY, "Error"); | |
0 | 8838 oop suffix_head = cur->mark(); // suffix will be put back on global list |
8839 cur->set_mark(NULL); // break off suffix | |
534 | 8840 // It's possible that the list is still in the empty(busy) state |
8841 // we left it in a short while ago; in that case we may be | |
8842 // able to place back the suffix without incurring the cost | |
8843 // of a walk down the list. | |
0 | 8844 oop observed_overflow_list = _overflow_list; |
534 | 8845 oop cur_overflow_list = observed_overflow_list; |
8846 bool attached = false; | |
8847 while (observed_overflow_list == BUSY || observed_overflow_list == NULL) { | |
0 | 8848 observed_overflow_list = |
534 | 8849 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); |
8850 if (cur_overflow_list == observed_overflow_list) { | |
8851 attached = true; | |
8852 break; | |
8853 } else cur_overflow_list = observed_overflow_list; | |
8854 } | |
8855 if (!attached) { | |
8856 // Too bad, someone else sneaked in (at least) an element; we'll need | |
8857 // to do a splice. Find tail of suffix so we can prepend suffix to global | |
8858 // list. | |
8859 for (cur = suffix_head; cur->mark() != NULL; cur = (oop)(cur->mark())); | |
8860 oop suffix_tail = cur; | |
8861 assert(suffix_tail != NULL && suffix_tail->mark() == NULL, | |
8862 "Tautology"); | |
8863 observed_overflow_list = _overflow_list; | |
8864 do { | |
8865 cur_overflow_list = observed_overflow_list; | |
8866 if (cur_overflow_list != BUSY) { | |
8867 // Do the splice ... | |
8868 suffix_tail->set_mark(markOop(cur_overflow_list)); | |
8869 } else { // cur_overflow_list == BUSY | |
8870 suffix_tail->set_mark(NULL); | |
8871 } | |
8872 // ... and try to place spliced list back on overflow_list ... | |
8873 observed_overflow_list = | |
8874 (oop) Atomic::cmpxchg_ptr(suffix_head, &_overflow_list, cur_overflow_list); | |
8875 } while (cur_overflow_list != observed_overflow_list); | |
8876 // ... until we have succeeded in doing so. | |
8877 } | |
0 | 8878 } |
8879 | |
8880 // Push the prefix elements on work_q | |
8881 assert(prefix != NULL, "control point invariant"); | |
8882 const markOop proto = markOopDesc::prototype(); | |
8883 oop next; | |
534 | 8884 NOT_PRODUCT(ssize_t n = 0;) |
0 | 8885 for (cur = prefix; cur != NULL; cur = next) { |
8886 next = oop(cur->mark()); | |
8887 cur->set_mark(proto); // until proven otherwise | |
8888 assert(cur->is_oop(), "Should be an oop"); | |
8889 bool res = work_q->push(cur); | |
8890 assert(res, "Bit off more than we can chew?"); | |
8891 NOT_PRODUCT(n++;) | |
8892 } | |
8893 #ifndef PRODUCT | |
8894 assert(_num_par_pushes >= n, "Too many pops?"); | |
8895 Atomic::add_ptr(-(intptr_t)n, &_num_par_pushes); | |
8896 #endif | |
8897 return true; | |
8898 } | |
8899 | |
8900 // Single-threaded | |
8901 void CMSCollector::push_on_overflow_list(oop p) { | |
8902 NOT_PRODUCT(_num_par_pushes++;) | |
8903 assert(p->is_oop(), "Not an oop"); | |
8904 preserve_mark_if_necessary(p); | |
8905 p->set_mark((markOop)_overflow_list); | |
8906 _overflow_list = p; | |
8907 } | |
8908 | |
8909 // Multi-threaded; use CAS to prepend to overflow list | |
8910 void CMSCollector::par_push_on_overflow_list(oop p) { | |
8911 NOT_PRODUCT(Atomic::inc_ptr(&_num_par_pushes);) | |
8912 assert(p->is_oop(), "Not an oop"); | |
8913 par_preserve_mark_if_necessary(p); | |
8914 oop observed_overflow_list = _overflow_list; | |
8915 oop cur_overflow_list; | |
8916 do { | |
8917 cur_overflow_list = observed_overflow_list; | |
534 | 8918 if (cur_overflow_list != BUSY) { |
8919 p->set_mark(markOop(cur_overflow_list)); | |
8920 } else { | |
8921 p->set_mark(NULL); | |
8922 } | |
0 | 8923 observed_overflow_list = |
8924 (oop) Atomic::cmpxchg_ptr(p, &_overflow_list, cur_overflow_list); | |
8925 } while (cur_overflow_list != observed_overflow_list); | |
8926 } | |
534 | 8927 #undef BUSY |
0 | 8928 |
8929 // Single threaded | |
8930 // General Note on GrowableArray: pushes may silently fail | |
8931 // because we are (temporarily) out of C-heap for expanding | |
8932 // the stack. The problem is quite ubiquitous and affects | |
8933 // a lot of code in the JVM. The prudent thing for GrowableArray | |
8934 // to do (for now) is to exit with an error. However, that may | |
8935 // be too draconian in some cases because the caller may be | |
534 | 8936 // able to recover without much harm. For such cases, we |
0 | 8937 // should probably introduce a "soft_push" method which returns |
8938 // an indication of success or failure with the assumption that | |
8939 // the caller may be able to recover from a failure; code in | |
8940 // the VM can then be changed, incrementally, to deal with such | |
8941 // failures where possible, thus, incrementally hardening the VM | |
8942 // in such low resource situations. | |
8943 void CMSCollector::preserve_mark_work(oop p, markOop m) { | |
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
8944 _preserved_oop_stack.push(p); |
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
8945 _preserved_mark_stack.push(m); |
0 | 8946 assert(m == p->mark(), "Mark word changed"); |
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
8947 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), |
0 | 8948 "bijection"); |
8949 } | |
8950 | |
8951 // Single threaded | |
8952 void CMSCollector::preserve_mark_if_necessary(oop p) { | |
8953 markOop m = p->mark(); | |
8954 if (m->must_be_preserved(p)) { | |
8955 preserve_mark_work(p, m); | |
8956 } | |
8957 } | |
8958 | |
8959 void CMSCollector::par_preserve_mark_if_necessary(oop p) { | |
8960 markOop m = p->mark(); | |
8961 if (m->must_be_preserved(p)) { | |
8962 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); | |
8963 // Even though we read the mark word without holding | |
8964 // the lock, we are assured that it will not change | |
8965 // because we "own" this oop, so no other thread can | |
8966 // be trying to push it on the overflow list; see | |
8967 // the assertion in preserve_mark_work() that checks | |
8968 // that m == p->mark(). | |
8969 preserve_mark_work(p, m); | |
8970 } | |
8971 } | |
8972 | |
8973 // We should be able to do this multi-threaded, | |
8974 // a chunk of stack being a task (this is | |
8975 // correct because each oop only ever appears | |
8976 // once in the overflow list. However, it's | |
8977 // not very easy to completely overlap this with | |
8978 // other operations, so will generally not be done | |
8979 // until all work's been completed. Because we | |
8980 // expect the preserved oop stack (set) to be small, | |
8981 // it's probably fine to do this single-threaded. | |
8982 // We can explore cleverer concurrent/overlapped/parallel | |
8983 // processing of preserved marks if we feel the | |
8984 // need for this in the future. Stack overflow should | |
8985 // be so rare in practice and, when it happens, its | |
8986 // effect on performance so great that this will | |
8987 // likely just be in the noise anyway. | |
8988 void CMSCollector::restore_preserved_marks_if_any() { | |
8989 assert(SafepointSynchronize::is_at_safepoint(), | |
8990 "world should be stopped"); | |
8991 assert(Thread::current()->is_ConcurrentGC_thread() || | |
8992 Thread::current()->is_VM_thread(), | |
8993 "should be single-threaded"); | |
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
8994 assert(_preserved_oop_stack.size() == _preserved_mark_stack.size(), |
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
8995 "bijection"); |
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
8996 |
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
8997 while (!_preserved_oop_stack.is_empty()) { |
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
8998 oop p = _preserved_oop_stack.pop(); |
0 | 8999 assert(p->is_oop(), "Should be an oop"); |
9000 assert(_span.contains(p), "oop should be in _span"); | |
9001 assert(p->mark() == markOopDesc::prototype(), | |
9002 "Set when taken from overflow list"); | |
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
9003 markOop m = _preserved_mark_stack.pop(); |
0 | 9004 p->set_mark(m); |
9005 } | |
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
9006 assert(_preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(), |
0 | 9007 "stacks were cleared above"); |
9008 } | |
9009 | |
9010 #ifndef PRODUCT | |
9011 bool CMSCollector::no_preserved_marks() const { | |
1836
894b1d7c7e01
6423256: GC stacks should use a better data structure
jcoomes
parents:
1833
diff
changeset
|
9012 return _preserved_mark_stack.is_empty() && _preserved_oop_stack.is_empty(); |
0 | 9013 } |
9014 #endif | |
9015 | |
9016 CMSAdaptiveSizePolicy* ASConcurrentMarkSweepGeneration::cms_size_policy() const | |
9017 { | |
9018 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); | |
9019 CMSAdaptiveSizePolicy* size_policy = | |
9020 (CMSAdaptiveSizePolicy*) gch->gen_policy()->size_policy(); | |
9021 assert(size_policy->is_gc_cms_adaptive_size_policy(), | |
9022 "Wrong type for size policy"); | |
9023 return size_policy; | |
9024 } | |
9025 | |
9026 void ASConcurrentMarkSweepGeneration::resize(size_t cur_promo_size, | |
9027 size_t desired_promo_size) { | |
9028 if (cur_promo_size < desired_promo_size) { | |
9029 size_t expand_bytes = desired_promo_size - cur_promo_size; | |
9030 if (PrintAdaptiveSizePolicy && Verbose) { | |
9031 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " | |
9032 "Expanding tenured generation by " SIZE_FORMAT " (bytes)", | |
9033 expand_bytes); | |
9034 } | |
9035 expand(expand_bytes, | |
9036 MinHeapDeltaBytes, | |
9037 CMSExpansionCause::_adaptive_size_policy); | |
9038 } else if (desired_promo_size < cur_promo_size) { | |
9039 size_t shrink_bytes = cur_promo_size - desired_promo_size; | |
9040 if (PrintAdaptiveSizePolicy && Verbose) { | |
9041 gclog_or_tty->print_cr(" ASConcurrentMarkSweepGeneration::resize " | |
9042 "Shrinking tenured generation by " SIZE_FORMAT " (bytes)", | |
9043 shrink_bytes); | |
9044 } | |
9045 shrink(shrink_bytes); | |
9046 } | |
9047 } | |
9048 | |
9049 CMSGCAdaptivePolicyCounters* ASConcurrentMarkSweepGeneration::gc_adaptive_policy_counters() { | |
9050 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
9051 CMSGCAdaptivePolicyCounters* counters = | |
9052 (CMSGCAdaptivePolicyCounters*) gch->collector_policy()->counters(); | |
9053 assert(counters->kind() == GCPolicyCounters::CMSGCAdaptivePolicyCountersKind, | |
9054 "Wrong kind of counters"); | |
9055 return counters; | |
9056 } | |
9057 | |
9058 | |
9059 void ASConcurrentMarkSweepGeneration::update_counters() { | |
9060 if (UsePerfData) { | |
9061 _space_counters->update_all(); | |
9062 _gen_counters->update_all(); | |
9063 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); | |
9064 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
9065 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); | |
9066 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, | |
9067 "Wrong gc statistics type"); | |
9068 counters->update_counters(gc_stats_l); | |
9069 } | |
9070 } | |
9071 | |
9072 void ASConcurrentMarkSweepGeneration::update_counters(size_t used) { | |
9073 if (UsePerfData) { | |
9074 _space_counters->update_used(used); | |
9075 _space_counters->update_capacity(); | |
9076 _gen_counters->update_all(); | |
9077 | |
9078 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); | |
9079 GenCollectedHeap* gch = GenCollectedHeap::heap(); | |
9080 CMSGCStats* gc_stats_l = (CMSGCStats*) gc_stats(); | |
9081 assert(gc_stats_l->kind() == GCStats::CMSGCStatsKind, | |
9082 "Wrong gc statistics type"); | |
9083 counters->update_counters(gc_stats_l); | |
9084 } | |
9085 } | |
9086 | |
9087 // The desired expansion delta is computed so that: | |
9088 // . desired free percentage or greater is used | |
9089 void ASConcurrentMarkSweepGeneration::compute_new_size() { | |
9090 assert_locked_or_safepoint(Heap_lock); | |
9091 | |
9092 GenCollectedHeap* gch = (GenCollectedHeap*) GenCollectedHeap::heap(); | |
9093 | |
9094 // If incremental collection failed, we just want to expand | |
9095 // to the limit. | |
9096 if (incremental_collection_failed()) { | |
9097 clear_incremental_collection_failed(); | |
9098 grow_to_reserved(); | |
9099 return; | |
9100 } | |
9101 | |
9102 assert(UseAdaptiveSizePolicy, "Should be using adaptive sizing"); | |
9103 | |
9104 assert(gch->kind() == CollectedHeap::GenCollectedHeap, | |
9105 "Wrong type of heap"); | |
9106 int prev_level = level() - 1; | |
9107 assert(prev_level >= 0, "The cms generation is the lowest generation"); | |
9108 Generation* prev_gen = gch->get_gen(prev_level); | |
9109 assert(prev_gen->kind() == Generation::ASParNew, | |
9110 "Wrong type of young generation"); | |
9111 ParNewGeneration* younger_gen = (ParNewGeneration*) prev_gen; | |
9112 size_t cur_eden = younger_gen->eden()->capacity(); | |
9113 CMSAdaptiveSizePolicy* size_policy = cms_size_policy(); | |
9114 size_t cur_promo = free(); | |
9115 size_policy->compute_tenured_generation_free_space(cur_promo, | |
9116 max_available(), | |
9117 cur_eden); | |
9118 resize(cur_promo, size_policy->promo_size()); | |
9119 | |
9120 // Record the new size of the space in the cms generation | |
9121 // that is available for promotions. This is temporary. | |
9122 // It should be the desired promo size. | |
9123 size_policy->avg_cms_promo()->sample(free()); | |
9124 size_policy->avg_old_live()->sample(used()); | |
9125 | |
9126 if (UsePerfData) { | |
9127 CMSGCAdaptivePolicyCounters* counters = gc_adaptive_policy_counters(); | |
9128 counters->update_cms_capacity_counter(capacity()); | |
9129 } | |
9130 } | |
9131 | |
9132 void ASConcurrentMarkSweepGeneration::shrink_by(size_t desired_bytes) { | |
9133 assert_locked_or_safepoint(Heap_lock); | |
9134 assert_lock_strong(freelistLock()); | |
9135 HeapWord* old_end = _cmsSpace->end(); | |
9136 HeapWord* unallocated_start = _cmsSpace->unallocated_block(); | |
9137 assert(old_end >= unallocated_start, "Miscalculation of unallocated_start"); | |
9138 FreeChunk* chunk_at_end = find_chunk_at_end(); | |
9139 if (chunk_at_end == NULL) { | |
9140 // No room to shrink | |
9141 if (PrintGCDetails && Verbose) { | |
9142 gclog_or_tty->print_cr("No room to shrink: old_end " | |
9143 PTR_FORMAT " unallocated_start " PTR_FORMAT | |
9144 " chunk_at_end " PTR_FORMAT, | |
9145 old_end, unallocated_start, chunk_at_end); | |
9146 } | |
9147 return; | |
9148 } else { | |
9149 | |
9150 // Find the chunk at the end of the space and determine | |
9151 // how much it can be shrunk. | |
9152 size_t shrinkable_size_in_bytes = chunk_at_end->size(); | |
9153 size_t aligned_shrinkable_size_in_bytes = | |
9154 align_size_down(shrinkable_size_in_bytes, os::vm_page_size()); | |
9155 assert(unallocated_start <= chunk_at_end->end(), | |
9156 "Inconsistent chunk at end of space"); | |
9157 size_t bytes = MIN2(desired_bytes, aligned_shrinkable_size_in_bytes); | |
9158 size_t word_size_before = heap_word_size(_virtual_space.committed_size()); | |
9159 | |
9160 // Shrink the underlying space | |
9161 _virtual_space.shrink_by(bytes); | |
9162 if (PrintGCDetails && Verbose) { | |
9163 gclog_or_tty->print_cr("ConcurrentMarkSweepGeneration::shrink_by:" | |
9164 " desired_bytes " SIZE_FORMAT | |
9165 " shrinkable_size_in_bytes " SIZE_FORMAT | |
9166 " aligned_shrinkable_size_in_bytes " SIZE_FORMAT | |
9167 " bytes " SIZE_FORMAT, | |
9168 desired_bytes, shrinkable_size_in_bytes, | |
9169 aligned_shrinkable_size_in_bytes, bytes); | |
9170 gclog_or_tty->print_cr(" old_end " SIZE_FORMAT | |
9171 " unallocated_start " SIZE_FORMAT, | |
9172 old_end, unallocated_start); | |
9173 } | |
9174 | |
9175 // If the space did shrink (shrinking is not guaranteed), | |
9176 // shrink the chunk at the end by the appropriate amount. | |
9177 if (((HeapWord*)_virtual_space.high()) < old_end) { | |
9178 size_t new_word_size = | |
9179 heap_word_size(_virtual_space.committed_size()); | |
9180 | |
9181 // Have to remove the chunk from the dictionary because it is changing | |
9182 // size and might be someplace elsewhere in the dictionary. | |
9183 | |
9184 // Get the chunk at end, shrink it, and put it | |
9185 // back. | |
9186 _cmsSpace->removeChunkFromDictionary(chunk_at_end); | |
9187 size_t word_size_change = word_size_before - new_word_size; | |
9188 size_t chunk_at_end_old_size = chunk_at_end->size(); | |
9189 assert(chunk_at_end_old_size >= word_size_change, | |
9190 "Shrink is too large"); | |
9191 chunk_at_end->setSize(chunk_at_end_old_size - | |
9192 word_size_change); | |
9193 _cmsSpace->freed((HeapWord*) chunk_at_end->end(), | |
9194 word_size_change); | |
9195 | |
9196 _cmsSpace->returnChunkToDictionary(chunk_at_end); | |
9197 | |
9198 MemRegion mr(_cmsSpace->bottom(), new_word_size); | |
9199 _bts->resize(new_word_size); // resize the block offset shared array | |
9200 Universe::heap()->barrier_set()->resize_covered_region(mr); | |
9201 _cmsSpace->assert_locked(); | |
9202 _cmsSpace->set_end((HeapWord*)_virtual_space.high()); | |
9203 | |
9204 NOT_PRODUCT(_cmsSpace->dictionary()->verify()); | |
9205 | |
9206 // update the space and generation capacity counters | |
9207 if (UsePerfData) { | |
9208 _space_counters->update_capacity(); | |
9209 _gen_counters->update_all(); | |
9210 } | |
9211 | |
9212 if (Verbose && PrintGCDetails) { | |
9213 size_t new_mem_size = _virtual_space.committed_size(); | |
9214 size_t old_mem_size = new_mem_size + bytes; | |
9215 gclog_or_tty->print_cr("Shrinking %s from %ldK by %ldK to %ldK", | |
9216 name(), old_mem_size/K, bytes/K, new_mem_size/K); | |
9217 } | |
9218 } | |
9219 | |
9220 assert(_cmsSpace->unallocated_block() <= _cmsSpace->end(), | |
9221 "Inconsistency at end of space"); | |
9222 assert(chunk_at_end->end() == _cmsSpace->end(), | |
9223 "Shrinking is inconsistent"); | |
9224 return; | |
9225 } | |
9226 } | |
9227 | |
9228 // Transfer some number of overflown objects to usual marking | |
9229 // stack. Return true if some objects were transferred. | |
9230 bool MarkRefsIntoAndScanClosure::take_from_overflow_list() { | |
679
cea947c8a988
6819891: ParNew: Fix work queue overflow code to deal correctly with +UseCompressedOops
ysr
parents:
628
diff
changeset
|
9231 size_t num = MIN2((size_t)(_mark_stack->capacity() - _mark_stack->length())/4, |
0 | 9232 (size_t)ParGCDesiredObjsFromOverflowList); |
9233 | |
9234 bool res = _collector->take_from_overflow_list(num, _mark_stack); | |
9235 assert(_collector->overflow_list_is_empty() || res, | |
9236 "If list is not empty, we should have taken something"); | |
9237 assert(!res || !_mark_stack->isEmpty(), | |
9238 "If we took something, it should now be on our stack"); | |
9239 return res; | |
9240 } | |
9241 | |
9242 size_t MarkDeadObjectsClosure::do_blk(HeapWord* addr) { | |
9243 size_t res = _sp->block_size_no_stall(addr, _collector); | |
9244 assert(res != 0, "Should always be able to compute a size"); | |
9245 if (_sp->block_is_obj(addr)) { | |
9246 if (_live_bit_map->isMarked(addr)) { | |
9247 // It can't have been dead in a previous cycle | |
9248 guarantee(!_dead_bit_map->isMarked(addr), "No resurrection!"); | |
9249 } else { | |
9250 _dead_bit_map->mark(addr); // mark the dead object | |
9251 } | |
9252 } | |
9253 return res; | |
9254 } | |
1703
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9255 |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9256 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(CMSCollector::CollectorState phase): TraceMemoryManagerStats() { |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9257 |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9258 switch (phase) { |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9259 case CMSCollector::InitialMarking: |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9260 initialize(true /* fullGC */ , |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9261 true /* recordGCBeginTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9262 true /* recordPreGCUsage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9263 false /* recordPeakUsage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9264 false /* recordPostGCusage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9265 true /* recordAccumulatedGCTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9266 false /* recordGCEndTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9267 false /* countCollection */ ); |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9268 break; |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9269 |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9270 case CMSCollector::FinalMarking: |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9271 initialize(true /* fullGC */ , |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9272 false /* recordGCBeginTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9273 false /* recordPreGCUsage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9274 false /* recordPeakUsage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9275 false /* recordPostGCusage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9276 true /* recordAccumulatedGCTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9277 false /* recordGCEndTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9278 false /* countCollection */ ); |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9279 break; |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9280 |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9281 case CMSCollector::Sweeping: |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9282 initialize(true /* fullGC */ , |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9283 false /* recordGCBeginTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9284 false /* recordPreGCUsage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9285 true /* recordPeakUsage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9286 true /* recordPostGCusage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9287 false /* recordAccumulatedGCTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9288 true /* recordGCEndTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9289 true /* countCollection */ ); |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9290 break; |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9291 |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9292 default: |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9293 ShouldNotReachHere(); |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9294 } |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9295 } |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9296 |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9297 // when bailing out of cms in concurrent mode failure |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9298 TraceCMSMemoryManagerStats::TraceCMSMemoryManagerStats(): TraceMemoryManagerStats() { |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9299 initialize(true /* fullGC */ , |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9300 true /* recordGCBeginTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9301 true /* recordPreGCUsage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9302 true /* recordPeakUsage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9303 true /* recordPostGCusage */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9304 true /* recordAccumulatedGCTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9305 true /* recordGCEndTime */, |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9306 true /* countCollection */ ); |
f6f3eef8a521
6581734: CMS Old Gen's collection usage is zero after GC which is incorrect
kevinw
parents:
1579
diff
changeset
|
9307 } |