Mercurial > hg > truffle
annotate src/share/vm/gc_implementation/g1/concurrentMark.cpp @ 1833:8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
Summary: Associate number of GC workers with the workgang as opposed to the task.
Reviewed-by: johnc, ysr
author | jmasa |
---|---|
date | Mon, 20 Sep 2010 14:38:38 -0700 |
parents | b63010841f78 |
children | 4805b9f4779e |
rev | line source |
---|---|
342 | 1 /* |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1547
diff
changeset
|
2 * Copyright (c) 2001, 2009, Oracle and/or its affiliates. All rights reserved. |
342 | 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:
1547
diff
changeset
|
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1547
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:
1547
diff
changeset
|
21 * questions. |
342 | 22 * |
23 */ | |
24 | |
25 #include "incls/_precompiled.incl" | |
26 #include "incls/_concurrentMark.cpp.incl" | |
27 | |
28 // | |
29 // CMS Bit Map Wrapper | |
30 | |
31 CMBitMapRO::CMBitMapRO(ReservedSpace rs, int shifter): | |
32 _bm((uintptr_t*)NULL,0), | |
33 _shifter(shifter) { | |
34 _bmStartWord = (HeapWord*)(rs.base()); | |
35 _bmWordSize = rs.size()/HeapWordSize; // rs.size() is in bytes | |
36 ReservedSpace brs(ReservedSpace::allocation_align_size_up( | |
37 (_bmWordSize >> (_shifter + LogBitsPerByte)) + 1)); | |
38 | |
39 guarantee(brs.is_reserved(), "couldn't allocate CMS bit map"); | |
40 // For now we'll just commit all of the bit map up fromt. | |
41 // Later on we'll try to be more parsimonious with swap. | |
42 guarantee(_virtual_space.initialize(brs, brs.size()), | |
43 "couldn't reseve backing store for CMS bit map"); | |
44 assert(_virtual_space.committed_size() == brs.size(), | |
45 "didn't reserve backing store for all of CMS bit map?"); | |
46 _bm.set_map((uintptr_t*)_virtual_space.low()); | |
47 assert(_virtual_space.committed_size() << (_shifter + LogBitsPerByte) >= | |
48 _bmWordSize, "inconsistency in bit map sizing"); | |
49 _bm.set_size(_bmWordSize >> _shifter); | |
50 } | |
51 | |
52 HeapWord* CMBitMapRO::getNextMarkedWordAddress(HeapWord* addr, | |
53 HeapWord* limit) const { | |
54 // First we must round addr *up* to a possible object boundary. | |
55 addr = (HeapWord*)align_size_up((intptr_t)addr, | |
56 HeapWordSize << _shifter); | |
57 size_t addrOffset = heapWordToOffset(addr); | |
58 if (limit == NULL) limit = _bmStartWord + _bmWordSize; | |
59 size_t limitOffset = heapWordToOffset(limit); | |
60 size_t nextOffset = _bm.get_next_one_offset(addrOffset, limitOffset); | |
61 HeapWord* nextAddr = offsetToHeapWord(nextOffset); | |
62 assert(nextAddr >= addr, "get_next_one postcondition"); | |
63 assert(nextAddr == limit || isMarked(nextAddr), | |
64 "get_next_one postcondition"); | |
65 return nextAddr; | |
66 } | |
67 | |
68 HeapWord* CMBitMapRO::getNextUnmarkedWordAddress(HeapWord* addr, | |
69 HeapWord* limit) const { | |
70 size_t addrOffset = heapWordToOffset(addr); | |
71 if (limit == NULL) limit = _bmStartWord + _bmWordSize; | |
72 size_t limitOffset = heapWordToOffset(limit); | |
73 size_t nextOffset = _bm.get_next_zero_offset(addrOffset, limitOffset); | |
74 HeapWord* nextAddr = offsetToHeapWord(nextOffset); | |
75 assert(nextAddr >= addr, "get_next_one postcondition"); | |
76 assert(nextAddr == limit || !isMarked(nextAddr), | |
77 "get_next_one postcondition"); | |
78 return nextAddr; | |
79 } | |
80 | |
81 int CMBitMapRO::heapWordDiffToOffsetDiff(size_t diff) const { | |
82 assert((diff & ((1 << _shifter) - 1)) == 0, "argument check"); | |
83 return (int) (diff >> _shifter); | |
84 } | |
85 | |
86 bool CMBitMapRO::iterate(BitMapClosure* cl, MemRegion mr) { | |
87 HeapWord* left = MAX2(_bmStartWord, mr.start()); | |
88 HeapWord* right = MIN2(_bmStartWord + _bmWordSize, mr.end()); | |
89 if (right > left) { | |
90 // Right-open interval [leftOffset, rightOffset). | |
91 return _bm.iterate(cl, heapWordToOffset(left), heapWordToOffset(right)); | |
92 } else { | |
93 return true; | |
94 } | |
95 } | |
96 | |
97 void CMBitMapRO::mostly_disjoint_range_union(BitMap* from_bitmap, | |
98 size_t from_start_index, | |
99 HeapWord* to_start_word, | |
100 size_t word_num) { | |
101 _bm.mostly_disjoint_range_union(from_bitmap, | |
102 from_start_index, | |
103 heapWordToOffset(to_start_word), | |
104 word_num); | |
105 } | |
106 | |
107 #ifndef PRODUCT | |
108 bool CMBitMapRO::covers(ReservedSpace rs) const { | |
109 // assert(_bm.map() == _virtual_space.low(), "map inconsistency"); | |
645
c3a720eefe82
6816308: Changes to allow builds with latest Windows SDK 6.1 on 64bit Windows 2003
kvn
parents:
619
diff
changeset
|
110 assert(((size_t)_bm.size() * (size_t)(1 << _shifter)) == _bmWordSize, |
342 | 111 "size inconsistency"); |
112 return _bmStartWord == (HeapWord*)(rs.base()) && | |
113 _bmWordSize == rs.size()>>LogHeapWordSize; | |
114 } | |
115 #endif | |
116 | |
117 void CMBitMap::clearAll() { | |
118 _bm.clear(); | |
119 return; | |
120 } | |
121 | |
122 void CMBitMap::markRange(MemRegion mr) { | |
123 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); | |
124 assert(!mr.is_empty(), "unexpected empty region"); | |
125 assert((offsetToHeapWord(heapWordToOffset(mr.end())) == | |
126 ((HeapWord *) mr.end())), | |
127 "markRange memory region end is not card aligned"); | |
128 // convert address range into offset range | |
129 _bm.at_put_range(heapWordToOffset(mr.start()), | |
130 heapWordToOffset(mr.end()), true); | |
131 } | |
132 | |
133 void CMBitMap::clearRange(MemRegion mr) { | |
134 mr.intersection(MemRegion(_bmStartWord, _bmWordSize)); | |
135 assert(!mr.is_empty(), "unexpected empty region"); | |
136 // convert address range into offset range | |
137 _bm.at_put_range(heapWordToOffset(mr.start()), | |
138 heapWordToOffset(mr.end()), false); | |
139 } | |
140 | |
141 MemRegion CMBitMap::getAndClearMarkedRegion(HeapWord* addr, | |
142 HeapWord* end_addr) { | |
143 HeapWord* start = getNextMarkedWordAddress(addr); | |
144 start = MIN2(start, end_addr); | |
145 HeapWord* end = getNextUnmarkedWordAddress(start); | |
146 end = MIN2(end, end_addr); | |
147 assert(start <= end, "Consistency check"); | |
148 MemRegion mr(start, end); | |
149 if (!mr.is_empty()) { | |
150 clearRange(mr); | |
151 } | |
152 return mr; | |
153 } | |
154 | |
155 CMMarkStack::CMMarkStack(ConcurrentMark* cm) : | |
156 _base(NULL), _cm(cm) | |
157 #ifdef ASSERT | |
158 , _drain_in_progress(false) | |
159 , _drain_in_progress_yields(false) | |
160 #endif | |
161 {} | |
162 | |
163 void CMMarkStack::allocate(size_t size) { | |
164 _base = NEW_C_HEAP_ARRAY(oop, size); | |
165 if (_base == NULL) | |
166 vm_exit_during_initialization("Failed to allocate " | |
167 "CM region mark stack"); | |
168 _index = 0; | |
169 // QQQQ cast ... | |
170 _capacity = (jint) size; | |
171 _oops_do_bound = -1; | |
172 NOT_PRODUCT(_max_depth = 0); | |
173 } | |
174 | |
175 CMMarkStack::~CMMarkStack() { | |
176 if (_base != NULL) FREE_C_HEAP_ARRAY(oop, _base); | |
177 } | |
178 | |
179 void CMMarkStack::par_push(oop ptr) { | |
180 while (true) { | |
181 if (isFull()) { | |
182 _overflow = true; | |
183 return; | |
184 } | |
185 // Otherwise... | |
186 jint index = _index; | |
187 jint next_index = index+1; | |
188 jint res = Atomic::cmpxchg(next_index, &_index, index); | |
189 if (res == index) { | |
190 _base[index] = ptr; | |
191 // Note that we don't maintain this atomically. We could, but it | |
192 // doesn't seem necessary. | |
193 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); | |
194 return; | |
195 } | |
196 // Otherwise, we need to try again. | |
197 } | |
198 } | |
199 | |
200 void CMMarkStack::par_adjoin_arr(oop* ptr_arr, int n) { | |
201 while (true) { | |
202 if (isFull()) { | |
203 _overflow = true; | |
204 return; | |
205 } | |
206 // Otherwise... | |
207 jint index = _index; | |
208 jint next_index = index + n; | |
209 if (next_index > _capacity) { | |
210 _overflow = true; | |
211 return; | |
212 } | |
213 jint res = Atomic::cmpxchg(next_index, &_index, index); | |
214 if (res == index) { | |
215 for (int i = 0; i < n; i++) { | |
216 int ind = index + i; | |
217 assert(ind < _capacity, "By overflow test above."); | |
218 _base[ind] = ptr_arr[i]; | |
219 } | |
220 NOT_PRODUCT(_max_depth = MAX2(_max_depth, next_index)); | |
221 return; | |
222 } | |
223 // Otherwise, we need to try again. | |
224 } | |
225 } | |
226 | |
227 | |
228 void CMMarkStack::par_push_arr(oop* ptr_arr, int n) { | |
229 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); | |
230 jint start = _index; | |
231 jint next_index = start + n; | |
232 if (next_index > _capacity) { | |
233 _overflow = true; | |
234 return; | |
235 } | |
236 // Otherwise. | |
237 _index = next_index; | |
238 for (int i = 0; i < n; i++) { | |
239 int ind = start + i; | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
240 assert(ind < _capacity, "By overflow test above."); |
342 | 241 _base[ind] = ptr_arr[i]; |
242 } | |
243 } | |
244 | |
245 | |
246 bool CMMarkStack::par_pop_arr(oop* ptr_arr, int max, int* n) { | |
247 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); | |
248 jint index = _index; | |
249 if (index == 0) { | |
250 *n = 0; | |
251 return false; | |
252 } else { | |
253 int k = MIN2(max, index); | |
254 jint new_ind = index - k; | |
255 for (int j = 0; j < k; j++) { | |
256 ptr_arr[j] = _base[new_ind + j]; | |
257 } | |
258 _index = new_ind; | |
259 *n = k; | |
260 return true; | |
261 } | |
262 } | |
263 | |
264 | |
265 CMRegionStack::CMRegionStack() : _base(NULL) {} | |
266 | |
267 void CMRegionStack::allocate(size_t size) { | |
268 _base = NEW_C_HEAP_ARRAY(MemRegion, size); | |
269 if (_base == NULL) | |
270 vm_exit_during_initialization("Failed to allocate " | |
271 "CM region mark stack"); | |
272 _index = 0; | |
273 // QQQQ cast ... | |
274 _capacity = (jint) size; | |
275 } | |
276 | |
277 CMRegionStack::~CMRegionStack() { | |
278 if (_base != NULL) FREE_C_HEAP_ARRAY(oop, _base); | |
279 } | |
280 | |
281 void CMRegionStack::push(MemRegion mr) { | |
282 assert(mr.word_size() > 0, "Precondition"); | |
283 while (true) { | |
284 if (isFull()) { | |
285 _overflow = true; | |
286 return; | |
287 } | |
288 // Otherwise... | |
289 jint index = _index; | |
290 jint next_index = index+1; | |
291 jint res = Atomic::cmpxchg(next_index, &_index, index); | |
292 if (res == index) { | |
293 _base[index] = mr; | |
294 return; | |
295 } | |
296 // Otherwise, we need to try again. | |
297 } | |
298 } | |
299 | |
1358
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
300 // Currently we do not call this at all. Normally we would call it |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
301 // during the concurrent marking / remark phases but we now call |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
302 // the lock-based version instead. But we might want to resurrect this |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
303 // code in the future. So, we'll leave it here commented out. |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
304 #if 0 |
342 | 305 MemRegion CMRegionStack::pop() { |
306 while (true) { | |
307 // Otherwise... | |
308 jint index = _index; | |
309 | |
310 if (index == 0) { | |
311 return MemRegion(); | |
312 } | |
313 jint next_index = index-1; | |
314 jint res = Atomic::cmpxchg(next_index, &_index, index); | |
315 if (res == index) { | |
316 MemRegion mr = _base[next_index]; | |
317 if (mr.start() != NULL) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
318 assert(mr.end() != NULL, "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
319 assert(mr.word_size() > 0, "invariant"); |
342 | 320 return mr; |
321 } else { | |
322 // that entry was invalidated... let's skip it | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
323 assert(mr.end() == NULL, "invariant"); |
342 | 324 } |
325 } | |
326 // Otherwise, we need to try again. | |
327 } | |
328 } | |
1358
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
329 #endif // 0 |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
330 |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
331 void CMRegionStack::push_with_lock(MemRegion mr) { |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
332 assert(mr.word_size() > 0, "Precondition"); |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
333 MutexLockerEx x(CMRegionStack_lock, Mutex::_no_safepoint_check_flag); |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
334 |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
335 if (isFull()) { |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
336 _overflow = true; |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
337 return; |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
338 } |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
339 |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
340 _base[_index] = mr; |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
341 _index += 1; |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
342 } |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
343 |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
344 MemRegion CMRegionStack::pop_with_lock() { |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
345 MutexLockerEx x(CMRegionStack_lock, Mutex::_no_safepoint_check_flag); |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
346 |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
347 while (true) { |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
348 if (_index == 0) { |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
349 return MemRegion(); |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
350 } |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
351 _index -= 1; |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
352 |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
353 MemRegion mr = _base[_index]; |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
354 if (mr.start() != NULL) { |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
355 assert(mr.end() != NULL, "invariant"); |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
356 assert(mr.word_size() > 0, "invariant"); |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
357 return mr; |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
358 } else { |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
359 // that entry was invalidated... let's skip it |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
360 assert(mr.end() == NULL, "invariant"); |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
361 } |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
362 } |
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
363 } |
342 | 364 |
365 bool CMRegionStack::invalidate_entries_into_cset() { | |
366 bool result = false; | |
367 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
368 for (int i = 0; i < _oops_do_bound; ++i) { | |
369 MemRegion mr = _base[i]; | |
370 if (mr.start() != NULL) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
371 assert(mr.end() != NULL, "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
372 assert(mr.word_size() > 0, "invariant"); |
342 | 373 HeapRegion* hr = g1h->heap_region_containing(mr.start()); |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
374 assert(hr != NULL, "invariant"); |
342 | 375 if (hr->in_collection_set()) { |
376 // The region points into the collection set | |
377 _base[i] = MemRegion(); | |
378 result = true; | |
379 } | |
380 } else { | |
381 // that entry was invalidated... let's skip it | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
382 assert(mr.end() == NULL, "invariant"); |
342 | 383 } |
384 } | |
385 return result; | |
386 } | |
387 | |
388 template<class OopClosureClass> | |
389 bool CMMarkStack::drain(OopClosureClass* cl, CMBitMap* bm, bool yield_after) { | |
390 assert(!_drain_in_progress || !_drain_in_progress_yields || yield_after | |
391 || SafepointSynchronize::is_at_safepoint(), | |
392 "Drain recursion must be yield-safe."); | |
393 bool res = true; | |
394 debug_only(_drain_in_progress = true); | |
395 debug_only(_drain_in_progress_yields = yield_after); | |
396 while (!isEmpty()) { | |
397 oop newOop = pop(); | |
398 assert(G1CollectedHeap::heap()->is_in_reserved(newOop), "Bad pop"); | |
399 assert(newOop->is_oop(), "Expected an oop"); | |
400 assert(bm == NULL || bm->isMarked((HeapWord*)newOop), | |
401 "only grey objects on this stack"); | |
402 // iterate over the oops in this oop, marking and pushing | |
403 // the ones in CMS generation. | |
404 newOop->oop_iterate(cl); | |
405 if (yield_after && _cm->do_yield_check()) { | |
406 res = false; break; | |
407 } | |
408 } | |
409 debug_only(_drain_in_progress = false); | |
410 return res; | |
411 } | |
412 | |
413 void CMMarkStack::oops_do(OopClosure* f) { | |
414 if (_index == 0) return; | |
415 assert(_oops_do_bound != -1 && _oops_do_bound <= _index, | |
416 "Bound must be set."); | |
417 for (int i = 0; i < _oops_do_bound; i++) { | |
418 f->do_oop(&_base[i]); | |
419 } | |
420 _oops_do_bound = -1; | |
421 } | |
422 | |
423 bool ConcurrentMark::not_yet_marked(oop obj) const { | |
424 return (_g1h->is_obj_ill(obj) | |
425 || (_g1h->is_in_permanent(obj) | |
426 && !nextMarkBitMap()->isMarked((HeapWord*)obj))); | |
427 } | |
428 | |
429 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away | |
430 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list | |
431 #endif // _MSC_VER | |
432 | |
433 ConcurrentMark::ConcurrentMark(ReservedSpace rs, | |
434 int max_regions) : | |
435 _markBitMap1(rs, MinObjAlignment - 1), | |
436 _markBitMap2(rs, MinObjAlignment - 1), | |
437 | |
438 _parallel_marking_threads(0), | |
439 _sleep_factor(0.0), | |
440 _marking_task_overhead(1.0), | |
441 _cleanup_sleep_factor(0.0), | |
442 _cleanup_task_overhead(1.0), | |
443 _region_bm(max_regions, false /* in_resource_area*/), | |
444 _card_bm((rs.size() + CardTableModRefBS::card_size - 1) >> | |
445 CardTableModRefBS::card_shift, | |
446 false /* in_resource_area*/), | |
447 _prevMarkBitMap(&_markBitMap1), | |
448 _nextMarkBitMap(&_markBitMap2), | |
449 _at_least_one_mark_complete(false), | |
450 | |
451 _markStack(this), | |
452 _regionStack(), | |
453 // _finger set in set_non_marking_state | |
454 | |
455 _max_task_num(MAX2(ParallelGCThreads, (size_t)1)), | |
456 // _active_tasks set in set_non_marking_state | |
457 // _tasks set inside the constructor | |
458 _task_queues(new CMTaskQueueSet((int) _max_task_num)), | |
459 _terminator(ParallelTaskTerminator((int) _max_task_num, _task_queues)), | |
460 | |
461 _has_overflown(false), | |
462 _concurrent(false), | |
619
7ea5ca260b28
6814467: G1: small fixes related to concurrent marking verboseness
tonyp
parents:
470
diff
changeset
|
463 _has_aborted(false), |
7ea5ca260b28
6814467: G1: small fixes related to concurrent marking verboseness
tonyp
parents:
470
diff
changeset
|
464 _restart_for_overflow(false), |
7ea5ca260b28
6814467: G1: small fixes related to concurrent marking verboseness
tonyp
parents:
470
diff
changeset
|
465 _concurrent_marking_in_progress(false), |
7ea5ca260b28
6814467: G1: small fixes related to concurrent marking verboseness
tonyp
parents:
470
diff
changeset
|
466 _should_gray_objects(false), |
342 | 467 |
468 // _verbose_level set below | |
469 | |
470 _init_times(), | |
471 _remark_times(), _remark_mark_times(), _remark_weak_ref_times(), | |
472 _cleanup_times(), | |
473 _total_counting_time(0.0), | |
474 _total_rs_scrub_time(0.0), | |
475 | |
936 | 476 _parallel_workers(NULL) |
342 | 477 { |
478 CMVerboseLevel verbose_level = | |
479 (CMVerboseLevel) G1MarkingVerboseLevel; | |
480 if (verbose_level < no_verbose) | |
481 verbose_level = no_verbose; | |
482 if (verbose_level > high_verbose) | |
483 verbose_level = high_verbose; | |
484 _verbose_level = verbose_level; | |
485 | |
486 if (verbose_low()) | |
487 gclog_or_tty->print_cr("[global] init, heap start = "PTR_FORMAT", " | |
488 "heap end = "PTR_FORMAT, _heap_start, _heap_end); | |
489 | |
1284 | 490 _markStack.allocate(MarkStackSize); |
751 | 491 _regionStack.allocate(G1MarkRegionStackSize); |
342 | 492 |
493 // Create & start a ConcurrentMark thread. | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
494 _cmThread = new ConcurrentMarkThread(this); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
495 assert(cmThread() != NULL, "CM Thread should have been created"); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
496 assert(cmThread()->cm() != NULL, "CM Thread should refer to this cm"); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
497 |
342 | 498 _g1h = G1CollectedHeap::heap(); |
499 assert(CGC_lock != NULL, "Where's the CGC_lock?"); | |
500 assert(_markBitMap1.covers(rs), "_markBitMap1 inconsistency"); | |
501 assert(_markBitMap2.covers(rs), "_markBitMap2 inconsistency"); | |
502 | |
503 SATBMarkQueueSet& satb_qs = JavaThread::satb_mark_queue_set(); | |
1282 | 504 satb_qs.set_buffer_size(G1SATBBufferSize); |
342 | 505 |
506 int size = (int) MAX2(ParallelGCThreads, (size_t)1); | |
507 _par_cleanup_thread_state = NEW_C_HEAP_ARRAY(ParCleanupThreadState*, size); | |
508 for (int i = 0 ; i < size; i++) { | |
509 _par_cleanup_thread_state[i] = new ParCleanupThreadState; | |
510 } | |
511 | |
512 _tasks = NEW_C_HEAP_ARRAY(CMTask*, _max_task_num); | |
513 _accum_task_vtime = NEW_C_HEAP_ARRAY(double, _max_task_num); | |
514 | |
515 // so that the assertion in MarkingTaskQueue::task_queue doesn't fail | |
516 _active_tasks = _max_task_num; | |
517 for (int i = 0; i < (int) _max_task_num; ++i) { | |
518 CMTaskQueue* task_queue = new CMTaskQueue(); | |
519 task_queue->initialize(); | |
520 _task_queues->register_queue(i, task_queue); | |
521 | |
522 _tasks[i] = new CMTask(i, this, task_queue, _task_queues); | |
523 _accum_task_vtime[i] = 0.0; | |
524 } | |
525 | |
1284 | 526 if (ConcGCThreads > ParallelGCThreads) { |
527 vm_exit_during_initialization("Can't have more ConcGCThreads " | |
342 | 528 "than ParallelGCThreads."); |
529 } | |
530 if (ParallelGCThreads == 0) { | |
531 // if we are not running with any parallel GC threads we will not | |
532 // spawn any marking threads either | |
533 _parallel_marking_threads = 0; | |
534 _sleep_factor = 0.0; | |
535 _marking_task_overhead = 1.0; | |
536 } else { | |
1284 | 537 if (ConcGCThreads > 0) { |
538 // notice that ConcGCThreads overwrites G1MarkingOverheadPercent | |
342 | 539 // if both are set |
540 | |
1284 | 541 _parallel_marking_threads = ConcGCThreads; |
342 | 542 _sleep_factor = 0.0; |
543 _marking_task_overhead = 1.0; | |
751 | 544 } else if (G1MarkingOverheadPercent > 0) { |
342 | 545 // we will calculate the number of parallel marking threads |
546 // based on a target overhead with respect to the soft real-time | |
547 // goal | |
548 | |
751 | 549 double marking_overhead = (double) G1MarkingOverheadPercent / 100.0; |
342 | 550 double overall_cm_overhead = |
751 | 551 (double) MaxGCPauseMillis * marking_overhead / |
552 (double) GCPauseIntervalMillis; | |
342 | 553 double cpu_ratio = 1.0 / (double) os::processor_count(); |
554 double marking_thread_num = ceil(overall_cm_overhead / cpu_ratio); | |
555 double marking_task_overhead = | |
556 overall_cm_overhead / marking_thread_num * | |
557 (double) os::processor_count(); | |
558 double sleep_factor = | |
559 (1.0 - marking_task_overhead) / marking_task_overhead; | |
560 | |
561 _parallel_marking_threads = (size_t) marking_thread_num; | |
562 _sleep_factor = sleep_factor; | |
563 _marking_task_overhead = marking_task_overhead; | |
564 } else { | |
565 _parallel_marking_threads = MAX2((ParallelGCThreads + 2) / 4, (size_t)1); | |
566 _sleep_factor = 0.0; | |
567 _marking_task_overhead = 1.0; | |
568 } | |
569 | |
570 if (parallel_marking_threads() > 1) | |
571 _cleanup_task_overhead = 1.0; | |
572 else | |
573 _cleanup_task_overhead = marking_task_overhead(); | |
574 _cleanup_sleep_factor = | |
575 (1.0 - cleanup_task_overhead()) / cleanup_task_overhead(); | |
576 | |
577 #if 0 | |
578 gclog_or_tty->print_cr("Marking Threads %d", parallel_marking_threads()); | |
579 gclog_or_tty->print_cr("CM Marking Task Overhead %1.4lf", marking_task_overhead()); | |
580 gclog_or_tty->print_cr("CM Sleep Factor %1.4lf", sleep_factor()); | |
581 gclog_or_tty->print_cr("CL Marking Task Overhead %1.4lf", cleanup_task_overhead()); | |
582 gclog_or_tty->print_cr("CL Sleep Factor %1.4lf", cleanup_sleep_factor()); | |
583 #endif | |
584 | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
585 guarantee(parallel_marking_threads() > 0, "peace of mind"); |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
586 _parallel_workers = new FlexibleWorkGang("G1 Parallel Marking Threads", |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
587 (int) _parallel_marking_threads, false, true); |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
588 if (_parallel_workers == NULL) { |
342 | 589 vm_exit_during_initialization("Failed necessary allocation."); |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
590 } else { |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
591 _parallel_workers->initialize_workers(); |
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
592 } |
342 | 593 } |
594 | |
595 // so that the call below can read a sensible value | |
596 _heap_start = (HeapWord*) rs.base(); | |
597 set_non_marking_state(); | |
598 } | |
599 | |
600 void ConcurrentMark::update_g1_committed(bool force) { | |
601 // If concurrent marking is not in progress, then we do not need to | |
602 // update _heap_end. This has a subtle and important | |
603 // side-effect. Imagine that two evacuation pauses happen between | |
604 // marking completion and remark. The first one can grow the | |
605 // heap (hence now the finger is below the heap end). Then, the | |
606 // second one could unnecessarily push regions on the region | |
607 // stack. This causes the invariant that the region stack is empty | |
608 // at the beginning of remark to be false. By ensuring that we do | |
609 // not observe heap expansions after marking is complete, then we do | |
610 // not have this problem. | |
611 if (!concurrent_marking_in_progress() && !force) | |
612 return; | |
613 | |
614 MemRegion committed = _g1h->g1_committed(); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
615 assert(committed.start() == _heap_start, "start shouldn't change"); |
342 | 616 HeapWord* new_end = committed.end(); |
617 if (new_end > _heap_end) { | |
618 // The heap has been expanded. | |
619 | |
620 _heap_end = new_end; | |
621 } | |
622 // Notice that the heap can also shrink. However, this only happens | |
623 // during a Full GC (at least currently) and the entire marking | |
624 // phase will bail out and the task will not be restarted. So, let's | |
625 // do nothing. | |
626 } | |
627 | |
628 void ConcurrentMark::reset() { | |
629 // Starting values for these two. This should be called in a STW | |
630 // phase. CM will be notified of any future g1_committed expansions | |
631 // will be at the end of evacuation pauses, when tasks are | |
632 // inactive. | |
633 MemRegion committed = _g1h->g1_committed(); | |
634 _heap_start = committed.start(); | |
635 _heap_end = committed.end(); | |
636 | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
637 // Separated the asserts so that we know which one fires. |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
638 assert(_heap_start != NULL, "heap bounds should look ok"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
639 assert(_heap_end != NULL, "heap bounds should look ok"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
640 assert(_heap_start < _heap_end, "heap bounds should look ok"); |
342 | 641 |
642 // reset all the marking data structures and any necessary flags | |
643 clear_marking_state(); | |
644 | |
645 if (verbose_low()) | |
646 gclog_or_tty->print_cr("[global] resetting"); | |
647 | |
648 // We do reset all of them, since different phases will use | |
649 // different number of active threads. So, it's easiest to have all | |
650 // of them ready. | |
651 for (int i = 0; i < (int) _max_task_num; ++i) | |
652 _tasks[i]->reset(_nextMarkBitMap); | |
653 | |
654 // we need this to make sure that the flag is on during the evac | |
655 // pause with initial mark piggy-backed | |
656 set_concurrent_marking_in_progress(); | |
657 } | |
658 | |
659 void ConcurrentMark::set_phase(size_t active_tasks, bool concurrent) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
660 assert(active_tasks <= _max_task_num, "we should not have more"); |
342 | 661 |
662 _active_tasks = active_tasks; | |
663 // Need to update the three data structures below according to the | |
664 // number of active threads for this phase. | |
665 _terminator = ParallelTaskTerminator((int) active_tasks, _task_queues); | |
666 _first_overflow_barrier_sync.set_n_workers((int) active_tasks); | |
667 _second_overflow_barrier_sync.set_n_workers((int) active_tasks); | |
668 | |
669 _concurrent = concurrent; | |
670 // We propagate this to all tasks, not just the active ones. | |
671 for (int i = 0; i < (int) _max_task_num; ++i) | |
672 _tasks[i]->set_concurrent(concurrent); | |
673 | |
674 if (concurrent) { | |
675 set_concurrent_marking_in_progress(); | |
676 } else { | |
677 // We currently assume that the concurrent flag has been set to | |
678 // false before we start remark. At this point we should also be | |
679 // in a STW phase. | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
680 assert(!concurrent_marking_in_progress(), "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
681 assert(_finger == _heap_end, "only way to get here"); |
342 | 682 update_g1_committed(true); |
683 } | |
684 } | |
685 | |
686 void ConcurrentMark::set_non_marking_state() { | |
687 // We set the global marking state to some default values when we're | |
688 // not doing marking. | |
689 clear_marking_state(); | |
690 _active_tasks = 0; | |
691 clear_concurrent_marking_in_progress(); | |
692 } | |
693 | |
694 ConcurrentMark::~ConcurrentMark() { | |
695 int size = (int) MAX2(ParallelGCThreads, (size_t)1); | |
696 for (int i = 0; i < size; i++) delete _par_cleanup_thread_state[i]; | |
697 FREE_C_HEAP_ARRAY(ParCleanupThreadState*, | |
698 _par_cleanup_thread_state); | |
699 | |
700 for (int i = 0; i < (int) _max_task_num; ++i) { | |
701 delete _task_queues->queue(i); | |
702 delete _tasks[i]; | |
703 } | |
704 delete _task_queues; | |
705 FREE_C_HEAP_ARRAY(CMTask*, _max_task_num); | |
706 } | |
707 | |
708 // This closure is used to mark refs into the g1 generation | |
709 // from external roots in the CMS bit map. | |
710 // Called at the first checkpoint. | |
711 // | |
712 | |
713 void ConcurrentMark::clearNextBitmap() { | |
1359
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
714 G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
715 G1CollectorPolicy* g1p = g1h->g1_policy(); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
716 |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
717 // Make sure that the concurrent mark thread looks to still be in |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
718 // the current cycle. |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
719 guarantee(cmThread()->during_cycle(), "invariant"); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
720 |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
721 // We are finishing up the current cycle by clearing the next |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
722 // marking bitmap and getting it ready for the next cycle. During |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
723 // this time no other cycle can start. So, let's make sure that this |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
724 // is the case. |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
725 guarantee(!g1h->mark_in_progress(), "invariant"); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
726 |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
727 // clear the mark bitmap (no grey objects to start with). |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
728 // We need to do this in chunks and offer to yield in between |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
729 // each chunk. |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
730 HeapWord* start = _nextMarkBitMap->startWord(); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
731 HeapWord* end = _nextMarkBitMap->endWord(); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
732 HeapWord* cur = start; |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
733 size_t chunkSize = M; |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
734 while (cur < end) { |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
735 HeapWord* next = cur + chunkSize; |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
736 if (next > end) |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
737 next = end; |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
738 MemRegion mr(cur,next); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
739 _nextMarkBitMap->clearRange(mr); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
740 cur = next; |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
741 do_yield_check(); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
742 |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
743 // Repeat the asserts from above. We'll do them as asserts here to |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
744 // minimize their overhead on the product. However, we'll have |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
745 // them as guarantees at the beginning / end of the bitmap |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
746 // clearing to get some checking in the product. |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
747 assert(cmThread()->during_cycle(), "invariant"); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
748 assert(!g1h->mark_in_progress(), "invariant"); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
749 } |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
750 |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
751 // Repeat the asserts from above. |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
752 guarantee(cmThread()->during_cycle(), "invariant"); |
23b1b27ac76c
6909756: G1: guarantee(G1CollectedHeap::heap()->mark_in_progress(),"Precondition.")
tonyp
parents:
1358
diff
changeset
|
753 guarantee(!g1h->mark_in_progress(), "invariant"); |
342 | 754 } |
755 | |
756 class NoteStartOfMarkHRClosure: public HeapRegionClosure { | |
757 public: | |
758 bool doHeapRegion(HeapRegion* r) { | |
759 if (!r->continuesHumongous()) { | |
760 r->note_start_of_marking(true); | |
761 } | |
762 return false; | |
763 } | |
764 }; | |
765 | |
766 void ConcurrentMark::checkpointRootsInitialPre() { | |
767 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
768 G1CollectorPolicy* g1p = g1h->g1_policy(); | |
769 | |
770 _has_aborted = false; | |
771 | |
1547
fb1a39993f69
6951319: enable solaris builds using Sun Studio 12 update 1
jcoomes
parents:
1388
diff
changeset
|
772 #ifndef PRODUCT |
1044 | 773 if (G1PrintReachableAtInitialMark) { |
1388 | 774 print_reachable("at-cycle-start", |
775 true /* use_prev_marking */, true /* all */); | |
1044 | 776 } |
1547
fb1a39993f69
6951319: enable solaris builds using Sun Studio 12 update 1
jcoomes
parents:
1388
diff
changeset
|
777 #endif |
342 | 778 |
779 // Initialise marking structures. This has to be done in a STW phase. | |
780 reset(); | |
781 } | |
782 | |
783 class CMMarkRootsClosure: public OopsInGenClosure { | |
784 private: | |
785 ConcurrentMark* _cm; | |
786 G1CollectedHeap* _g1h; | |
787 bool _do_barrier; | |
788 | |
789 public: | |
790 CMMarkRootsClosure(ConcurrentMark* cm, | |
791 G1CollectedHeap* g1h, | |
792 bool do_barrier) : _cm(cm), _g1h(g1h), | |
793 _do_barrier(do_barrier) { } | |
794 | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
795 virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
796 virtual void do_oop( oop* p) { do_oop_work(p); } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
797 |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
798 template <class T> void do_oop_work(T* p) { |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
799 T heap_oop = oopDesc::load_heap_oop(p); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
800 if (!oopDesc::is_null(heap_oop)) { |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
801 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
802 assert(obj->is_oop() || obj->mark() == NULL, |
342 | 803 "expected an oop, possibly with mark word displaced"); |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
804 HeapWord* addr = (HeapWord*)obj; |
342 | 805 if (_g1h->is_in_g1_reserved(addr)) { |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
806 _cm->grayRoot(obj); |
342 | 807 } |
808 } | |
809 if (_do_barrier) { | |
810 assert(!_g1h->is_in_g1_reserved(p), | |
811 "Should be called on external roots"); | |
812 do_barrier(p); | |
813 } | |
814 } | |
815 }; | |
816 | |
817 void ConcurrentMark::checkpointRootsInitialPost() { | |
818 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
819 | |
820 // For each region note start of marking. | |
821 NoteStartOfMarkHRClosure startcl; | |
822 g1h->heap_region_iterate(&startcl); | |
823 | |
824 // Start weak-reference discovery. | |
825 ReferenceProcessor* rp = g1h->ref_processor(); | |
826 rp->verify_no_references_recorded(); | |
827 rp->enable_discovery(); // enable ("weak") refs discovery | |
457
27a80744a83b
6778647: snap(), snap_policy() should be renamed setup(), setup_policy()
ysr
parents:
453
diff
changeset
|
828 rp->setup_policy(false); // snapshot the soft ref policy to be used in this cycle |
342 | 829 |
830 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
1317
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
831 // This is the start of the marking cycle, we're expected all |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
832 // threads to have SATB queues with active set to false. |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
833 satb_mq_set.set_active_all_threads(true, /* new active value */ |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
834 false /* expected_active */); |
342 | 835 |
836 // update_g1_committed() will be called at the end of an evac pause | |
837 // when marking is on. So, it's also called at the end of the | |
838 // initial-mark pause to update the heap end, if the heap expands | |
839 // during it. No need to call it here. | |
840 } | |
841 | |
842 // Checkpoint the roots into this generation from outside | |
843 // this generation. [Note this initial checkpoint need only | |
844 // be approximate -- we'll do a catch up phase subsequently.] | |
845 void ConcurrentMark::checkpointRootsInitial() { | |
846 assert(SafepointSynchronize::is_at_safepoint(), "world should be stopped"); | |
847 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
848 | |
849 double start = os::elapsedTime(); | |
850 | |
851 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy(); | |
852 g1p->record_concurrent_mark_init_start(); | |
853 checkpointRootsInitialPre(); | |
854 | |
855 // YSR: when concurrent precleaning is in place, we'll | |
856 // need to clear the cached card table here | |
857 | |
858 ResourceMark rm; | |
859 HandleMark hm; | |
860 | |
861 g1h->ensure_parsability(false); | |
862 g1h->perm_gen()->save_marks(); | |
863 | |
864 CMMarkRootsClosure notOlder(this, g1h, false); | |
865 CMMarkRootsClosure older(this, g1h, true); | |
866 | |
867 g1h->set_marking_started(); | |
868 g1h->rem_set()->prepare_for_younger_refs_iterate(false); | |
869 | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
870 g1h->process_strong_roots(true, // activate StrongRootsScope |
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
871 false, // fake perm gen collection |
342 | 872 SharedHeap::SO_AllClasses, |
873 ¬Older, // Regular roots | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
874 NULL, // do not visit active blobs |
342 | 875 &older // Perm Gen Roots |
876 ); | |
877 checkpointRootsInitialPost(); | |
878 | |
879 // Statistics. | |
880 double end = os::elapsedTime(); | |
881 _init_times.add((end - start) * 1000.0); | |
882 | |
883 g1p->record_concurrent_mark_init_end(); | |
884 } | |
885 | |
886 /* | |
887 Notice that in the next two methods, we actually leave the STS | |
888 during the barrier sync and join it immediately afterwards. If we | |
889 do not do this, this then the following deadlock can occur: one | |
890 thread could be in the barrier sync code, waiting for the other | |
891 thread to also sync up, whereas another one could be trying to | |
892 yield, while also waiting for the other threads to sync up too. | |
893 | |
894 Because the thread that does the sync barrier has left the STS, it | |
895 is possible to be suspended for a Full GC or an evacuation pause | |
896 could occur. This is actually safe, since the entering the sync | |
897 barrier is one of the last things do_marking_step() does, and it | |
898 doesn't manipulate any data structures afterwards. | |
899 */ | |
900 | |
901 void ConcurrentMark::enter_first_sync_barrier(int task_num) { | |
902 if (verbose_low()) | |
903 gclog_or_tty->print_cr("[%d] entering first barrier", task_num); | |
904 | |
905 ConcurrentGCThread::stsLeave(); | |
906 _first_overflow_barrier_sync.enter(); | |
907 ConcurrentGCThread::stsJoin(); | |
908 // at this point everyone should have synced up and not be doing any | |
909 // more work | |
910 | |
911 if (verbose_low()) | |
912 gclog_or_tty->print_cr("[%d] leaving first barrier", task_num); | |
913 | |
914 // let task 0 do this | |
915 if (task_num == 0) { | |
916 // task 0 is responsible for clearing the global data structures | |
917 clear_marking_state(); | |
918 | |
919 if (PrintGC) { | |
920 gclog_or_tty->date_stamp(PrintGCDateStamps); | |
921 gclog_or_tty->stamp(PrintGCTimeStamps); | |
922 gclog_or_tty->print_cr("[GC concurrent-mark-reset-for-overflow]"); | |
923 } | |
924 } | |
925 | |
926 // after this, each task should reset its own data structures then | |
927 // then go into the second barrier | |
928 } | |
929 | |
930 void ConcurrentMark::enter_second_sync_barrier(int task_num) { | |
931 if (verbose_low()) | |
932 gclog_or_tty->print_cr("[%d] entering second barrier", task_num); | |
933 | |
934 ConcurrentGCThread::stsLeave(); | |
935 _second_overflow_barrier_sync.enter(); | |
936 ConcurrentGCThread::stsJoin(); | |
937 // at this point everything should be re-initialised and ready to go | |
938 | |
939 if (verbose_low()) | |
940 gclog_or_tty->print_cr("[%d] leaving second barrier", task_num); | |
941 } | |
942 | |
943 void ConcurrentMark::grayRoot(oop p) { | |
944 HeapWord* addr = (HeapWord*) p; | |
945 // We can't really check against _heap_start and _heap_end, since it | |
946 // is possible during an evacuation pause with piggy-backed | |
947 // initial-mark that the committed space is expanded during the | |
948 // pause without CM observing this change. So the assertions below | |
949 // is a bit conservative; but better than nothing. | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
950 assert(_g1h->g1_committed().contains(addr), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
951 "address should be within the heap bounds"); |
342 | 952 |
953 if (!_nextMarkBitMap->isMarked(addr)) | |
954 _nextMarkBitMap->parMark(addr); | |
955 } | |
956 | |
957 void ConcurrentMark::grayRegionIfNecessary(MemRegion mr) { | |
958 // The objects on the region have already been marked "in bulk" by | |
959 // the caller. We only need to decide whether to push the region on | |
960 // the region stack or not. | |
961 | |
962 if (!concurrent_marking_in_progress() || !_should_gray_objects) | |
963 // We're done with marking and waiting for remark. We do not need to | |
964 // push anything else on the region stack. | |
965 return; | |
966 | |
967 HeapWord* finger = _finger; | |
968 | |
969 if (verbose_low()) | |
970 gclog_or_tty->print_cr("[global] attempting to push " | |
971 "region ["PTR_FORMAT", "PTR_FORMAT"), finger is at " | |
972 PTR_FORMAT, mr.start(), mr.end(), finger); | |
973 | |
974 if (mr.start() < finger) { | |
975 // The finger is always heap region aligned and it is not possible | |
976 // for mr to span heap regions. | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
977 assert(mr.end() <= finger, "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
978 |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
979 // Separated the asserts so that we know which one fires. |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
980 assert(mr.start() <= mr.end(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
981 "region boundaries should fall within the committed space"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
982 assert(_heap_start <= mr.start(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
983 "region boundaries should fall within the committed space"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
984 assert(mr.end() <= _heap_end, |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
985 "region boundaries should fall within the committed space"); |
342 | 986 if (verbose_low()) |
987 gclog_or_tty->print_cr("[global] region ["PTR_FORMAT", "PTR_FORMAT") " | |
988 "below the finger, pushing it", | |
989 mr.start(), mr.end()); | |
990 | |
991 if (!region_stack_push(mr)) { | |
992 if (verbose_low()) | |
993 gclog_or_tty->print_cr("[global] region stack has overflown."); | |
994 } | |
995 } | |
996 } | |
997 | |
998 void ConcurrentMark::markAndGrayObjectIfNecessary(oop p) { | |
999 // The object is not marked by the caller. We need to at least mark | |
1000 // it and maybe push in on the stack. | |
1001 | |
1002 HeapWord* addr = (HeapWord*)p; | |
1003 if (!_nextMarkBitMap->isMarked(addr)) { | |
1004 // We definitely need to mark it, irrespective whether we bail out | |
1005 // because we're done with marking. | |
1006 if (_nextMarkBitMap->parMark(addr)) { | |
1007 if (!concurrent_marking_in_progress() || !_should_gray_objects) | |
1008 // If we're done with concurrent marking and we're waiting for | |
1009 // remark, then we're not pushing anything on the stack. | |
1010 return; | |
1011 | |
1012 // No OrderAccess:store_load() is needed. It is implicit in the | |
1013 // CAS done in parMark(addr) above | |
1014 HeapWord* finger = _finger; | |
1015 | |
1016 if (addr < finger) { | |
1017 if (!mark_stack_push(oop(addr))) { | |
1018 if (verbose_low()) | |
1019 gclog_or_tty->print_cr("[global] global stack overflow " | |
1020 "during parMark"); | |
1021 } | |
1022 } | |
1023 } | |
1024 } | |
1025 } | |
1026 | |
1027 class CMConcurrentMarkingTask: public AbstractGangTask { | |
1028 private: | |
1029 ConcurrentMark* _cm; | |
1030 ConcurrentMarkThread* _cmt; | |
1031 | |
1032 public: | |
1033 void work(int worker_i) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
1034 assert(Thread::current()->is_ConcurrentGC_thread(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
1035 "this should only be done by a conc GC thread"); |
342 | 1036 |
1037 double start_vtime = os::elapsedVTime(); | |
1038 | |
1039 ConcurrentGCThread::stsJoin(); | |
1040 | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
1041 assert((size_t) worker_i < _cm->active_tasks(), "invariant"); |
342 | 1042 CMTask* the_task = _cm->task(worker_i); |
1043 the_task->record_start_time(); | |
1044 if (!_cm->has_aborted()) { | |
1045 do { | |
1046 double start_vtime_sec = os::elapsedVTime(); | |
1047 double start_time_sec = os::elapsedTime(); | |
1048 the_task->do_marking_step(10.0); | |
1049 double end_time_sec = os::elapsedTime(); | |
1050 double end_vtime_sec = os::elapsedVTime(); | |
1051 double elapsed_vtime_sec = end_vtime_sec - start_vtime_sec; | |
1052 double elapsed_time_sec = end_time_sec - start_time_sec; | |
1053 _cm->clear_has_overflown(); | |
1054 | |
1055 bool ret = _cm->do_yield_check(worker_i); | |
1056 | |
1057 jlong sleep_time_ms; | |
1058 if (!_cm->has_aborted() && the_task->has_aborted()) { | |
1059 sleep_time_ms = | |
1060 (jlong) (elapsed_vtime_sec * _cm->sleep_factor() * 1000.0); | |
1061 ConcurrentGCThread::stsLeave(); | |
1062 os::sleep(Thread::current(), sleep_time_ms, false); | |
1063 ConcurrentGCThread::stsJoin(); | |
1064 } | |
1065 double end_time2_sec = os::elapsedTime(); | |
1066 double elapsed_time2_sec = end_time2_sec - start_time_sec; | |
1067 | |
1068 #if 0 | |
1069 gclog_or_tty->print_cr("CM: elapsed %1.4lf ms, sleep %1.4lf ms, " | |
1070 "overhead %1.4lf", | |
1071 elapsed_vtime_sec * 1000.0, (double) sleep_time_ms, | |
1072 the_task->conc_overhead(os::elapsedTime()) * 8.0); | |
1073 gclog_or_tty->print_cr("elapsed time %1.4lf ms, time 2: %1.4lf ms", | |
1074 elapsed_time_sec * 1000.0, elapsed_time2_sec * 1000.0); | |
1075 #endif | |
1076 } while (!_cm->has_aborted() && the_task->has_aborted()); | |
1077 } | |
1078 the_task->record_end_time(); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
1079 guarantee(!the_task->has_aborted() || _cm->has_aborted(), "invariant"); |
342 | 1080 |
1081 ConcurrentGCThread::stsLeave(); | |
1082 | |
1083 double end_vtime = os::elapsedVTime(); | |
1084 _cm->update_accum_task_vtime(worker_i, end_vtime - start_vtime); | |
1085 } | |
1086 | |
1087 CMConcurrentMarkingTask(ConcurrentMark* cm, | |
1088 ConcurrentMarkThread* cmt) : | |
1089 AbstractGangTask("Concurrent Mark"), _cm(cm), _cmt(cmt) { } | |
1090 | |
1091 ~CMConcurrentMarkingTask() { } | |
1092 }; | |
1093 | |
1094 void ConcurrentMark::markFromRoots() { | |
1095 // we might be tempted to assert that: | |
1096 // assert(asynch == !SafepointSynchronize::is_at_safepoint(), | |
1097 // "inconsistent argument?"); | |
1098 // However that wouldn't be right, because it's possible that | |
1099 // a safepoint is indeed in progress as a younger generation | |
1100 // stop-the-world GC happens even as we mark in this generation. | |
1101 | |
1102 _restart_for_overflow = false; | |
1103 | |
1104 set_phase(MAX2((size_t) 1, parallel_marking_threads()), true); | |
1105 | |
1106 CMConcurrentMarkingTask markingTask(this, cmThread()); | |
1107 if (parallel_marking_threads() > 0) | |
1108 _parallel_workers->run_task(&markingTask); | |
1109 else | |
1110 markingTask.work(0); | |
1111 print_stats(); | |
1112 } | |
1113 | |
1114 void ConcurrentMark::checkpointRootsFinal(bool clear_all_soft_refs) { | |
1115 // world is stopped at this checkpoint | |
1116 assert(SafepointSynchronize::is_at_safepoint(), | |
1117 "world should be stopped"); | |
1118 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
1119 | |
1120 // If a full collection has happened, we shouldn't do this. | |
1121 if (has_aborted()) { | |
1122 g1h->set_marking_complete(); // So bitmap clearing isn't confused | |
1123 return; | |
1124 } | |
1125 | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1126 if (VerifyDuringGC) { |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1127 HandleMark hm; // handle scope |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1128 gclog_or_tty->print(" VerifyDuringGC:(before)"); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1129 Universe::heap()->prepare_for_verify(); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1130 Universe::verify(true, false, true); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1131 } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1132 |
342 | 1133 G1CollectorPolicy* g1p = g1h->g1_policy(); |
1134 g1p->record_concurrent_mark_remark_start(); | |
1135 | |
1136 double start = os::elapsedTime(); | |
1137 | |
1138 checkpointRootsFinalWork(); | |
1139 | |
1140 double mark_work_end = os::elapsedTime(); | |
1141 | |
1142 weakRefsWork(clear_all_soft_refs); | |
1143 | |
1144 if (has_overflown()) { | |
1145 // Oops. We overflowed. Restart concurrent marking. | |
1146 _restart_for_overflow = true; | |
1147 // Clear the flag. We do not need it any more. | |
1148 clear_has_overflown(); | |
1149 if (G1TraceMarkStackOverflow) | |
1150 gclog_or_tty->print_cr("\nRemark led to restart for overflow."); | |
1151 } else { | |
1152 // We're done with marking. | |
1317
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
1153 // This is the end of the marking cycle, we're expected all |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
1154 // threads to have SATB queues with active set to true. |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
1155 JavaThread::satb_mark_queue_set().set_active_all_threads( |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
1156 false, /* new active value */ |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
1157 true /* expected_active */); |
811 | 1158 |
1159 if (VerifyDuringGC) { | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1160 HandleMark hm; // handle scope |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1161 gclog_or_tty->print(" VerifyDuringGC:(after)"); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1162 Universe::heap()->prepare_for_verify(); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1163 Universe::heap()->verify(/* allow_dirty */ true, |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1164 /* silent */ false, |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1165 /* use_prev_marking */ false); |
811 | 1166 } |
342 | 1167 } |
1168 | |
1169 #if VERIFY_OBJS_PROCESSED | |
1170 _scan_obj_cl.objs_processed = 0; | |
1171 ThreadLocalObjQueue::objs_enqueued = 0; | |
1172 #endif | |
1173 | |
1174 // Statistics | |
1175 double now = os::elapsedTime(); | |
1176 _remark_mark_times.add((mark_work_end - start) * 1000.0); | |
1177 _remark_weak_ref_times.add((now - mark_work_end) * 1000.0); | |
1178 _remark_times.add((now - start) * 1000.0); | |
1179 | |
1180 g1p->record_concurrent_mark_remark_end(); | |
1181 } | |
1182 | |
1183 | |
1184 #define CARD_BM_TEST_MODE 0 | |
1185 | |
1186 class CalcLiveObjectsClosure: public HeapRegionClosure { | |
1187 | |
1188 CMBitMapRO* _bm; | |
1189 ConcurrentMark* _cm; | |
1190 bool _changed; | |
1191 bool _yield; | |
1192 size_t _words_done; | |
1193 size_t _tot_live; | |
1194 size_t _tot_used; | |
1195 size_t _regions_done; | |
1196 double _start_vtime_sec; | |
1197 | |
1198 BitMap* _region_bm; | |
1199 BitMap* _card_bm; | |
1200 intptr_t _bottom_card_num; | |
1201 bool _final; | |
1202 | |
1203 void mark_card_num_range(intptr_t start_card_num, intptr_t last_card_num) { | |
1204 for (intptr_t i = start_card_num; i <= last_card_num; i++) { | |
1205 #if CARD_BM_TEST_MODE | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
1206 guarantee(_card_bm->at(i - _bottom_card_num), "Should already be set."); |
342 | 1207 #else |
1208 _card_bm->par_at_put(i - _bottom_card_num, 1); | |
1209 #endif | |
1210 } | |
1211 } | |
1212 | |
1213 public: | |
1214 CalcLiveObjectsClosure(bool final, | |
1215 CMBitMapRO *bm, ConcurrentMark *cm, | |
936 | 1216 BitMap* region_bm, BitMap* card_bm) : |
342 | 1217 _bm(bm), _cm(cm), _changed(false), _yield(true), |
1218 _words_done(0), _tot_live(0), _tot_used(0), | |
936 | 1219 _region_bm(region_bm), _card_bm(card_bm),_final(final), |
342 | 1220 _regions_done(0), _start_vtime_sec(0.0) |
1221 { | |
1222 _bottom_card_num = | |
1223 intptr_t(uintptr_t(G1CollectedHeap::heap()->reserved_region().start()) >> | |
1224 CardTableModRefBS::card_shift); | |
1225 } | |
1226 | |
829 | 1227 // It takes a region that's not empty (i.e., it has at least one |
1228 // live object in it and sets its corresponding bit on the region | |
1229 // bitmap to 1. If the region is "starts humongous" it will also set | |
1230 // to 1 the bits on the region bitmap that correspond to its | |
1231 // associated "continues humongous" regions. | |
1232 void set_bit_for_region(HeapRegion* hr) { | |
1233 assert(!hr->continuesHumongous(), "should have filtered those out"); | |
1234 | |
1235 size_t index = hr->hrs_index(); | |
1236 if (!hr->startsHumongous()) { | |
1237 // Normal (non-humongous) case: just set the bit. | |
1238 _region_bm->par_at_put((BitMap::idx_t) index, true); | |
1239 } else { | |
1240 // Starts humongous case: calculate how many regions are part of | |
1241 // this humongous region and then set the bit range. It might | |
1242 // have been a bit more efficient to look at the object that | |
1243 // spans these humongous regions to calculate their number from | |
1244 // the object's size. However, it's a good idea to calculate | |
1245 // this based on the metadata itself, and not the region | |
1246 // contents, so that this code is not aware of what goes into | |
1247 // the humongous regions (in case this changes in the future). | |
1248 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
1249 size_t end_index = index + 1; | |
831 | 1250 while (end_index < g1h->n_regions()) { |
1251 HeapRegion* chr = g1h->region_at(end_index); | |
829 | 1252 if (!chr->continuesHumongous()) { |
1253 break; | |
1254 } | |
1255 end_index += 1; | |
1256 } | |
1257 _region_bm->par_at_put_range((BitMap::idx_t) index, | |
1258 (BitMap::idx_t) end_index, true); | |
1259 } | |
1260 } | |
1261 | |
342 | 1262 bool doHeapRegion(HeapRegion* hr) { |
1263 if (!_final && _regions_done == 0) | |
1264 _start_vtime_sec = os::elapsedVTime(); | |
1265 | |
639 | 1266 if (hr->continuesHumongous()) { |
829 | 1267 // We will ignore these here and process them when their |
1268 // associated "starts humongous" region is processed (see | |
1269 // set_bit_for_heap_region()). Note that we cannot rely on their | |
1270 // associated "starts humongous" region to have their bit set to | |
1271 // 1 since, due to the region chunking in the parallel region | |
1272 // iteration, a "continues humongous" region might be visited | |
1273 // before its associated "starts humongous". | |
639 | 1274 return false; |
1275 } | |
342 | 1276 |
1277 HeapWord* nextTop = hr->next_top_at_mark_start(); | |
1278 HeapWord* start = hr->top_at_conc_mark_count(); | |
1279 assert(hr->bottom() <= start && start <= hr->end() && | |
1280 hr->bottom() <= nextTop && nextTop <= hr->end() && | |
1281 start <= nextTop, | |
1282 "Preconditions."); | |
1283 // Otherwise, record the number of word's we'll examine. | |
1284 size_t words_done = (nextTop - start); | |
1285 // Find the first marked object at or after "start". | |
1286 start = _bm->getNextMarkedWordAddress(start, nextTop); | |
1287 size_t marked_bytes = 0; | |
1288 | |
1289 // Below, the term "card num" means the result of shifting an address | |
1290 // by the card shift -- address 0 corresponds to card number 0. One | |
1291 // must subtract the card num of the bottom of the heap to obtain a | |
1292 // card table index. | |
1293 // The first card num of the sequence of live cards currently being | |
1294 // constructed. -1 ==> no sequence. | |
1295 intptr_t start_card_num = -1; | |
1296 // The last card num of the sequence of live cards currently being | |
1297 // constructed. -1 ==> no sequence. | |
1298 intptr_t last_card_num = -1; | |
1299 | |
1300 while (start < nextTop) { | |
1301 if (_yield && _cm->do_yield_check()) { | |
1302 // We yielded. It might be for a full collection, in which case | |
1303 // all bets are off; terminate the traversal. | |
1304 if (_cm->has_aborted()) { | |
1305 _changed = false; | |
1306 return true; | |
1307 } else { | |
1308 // Otherwise, it might be a collection pause, and the region | |
1309 // we're looking at might be in the collection set. We'll | |
1310 // abandon this region. | |
1311 return false; | |
1312 } | |
1313 } | |
1314 oop obj = oop(start); | |
1315 int obj_sz = obj->size(); | |
1316 // The card num of the start of the current object. | |
1317 intptr_t obj_card_num = | |
1318 intptr_t(uintptr_t(start) >> CardTableModRefBS::card_shift); | |
1319 | |
1320 HeapWord* obj_last = start + obj_sz - 1; | |
1321 intptr_t obj_last_card_num = | |
1322 intptr_t(uintptr_t(obj_last) >> CardTableModRefBS::card_shift); | |
1323 | |
1324 if (obj_card_num != last_card_num) { | |
1325 if (start_card_num == -1) { | |
1326 assert(last_card_num == -1, "Both or neither."); | |
1327 start_card_num = obj_card_num; | |
1328 } else { | |
1329 assert(last_card_num != -1, "Both or neither."); | |
1330 assert(obj_card_num >= last_card_num, "Inv"); | |
1331 if ((obj_card_num - last_card_num) > 1) { | |
1332 // Mark the last run, and start a new one. | |
1333 mark_card_num_range(start_card_num, last_card_num); | |
1334 start_card_num = obj_card_num; | |
1335 } | |
1336 } | |
1337 #if CARD_BM_TEST_MODE | |
1338 /* | |
1339 gclog_or_tty->print_cr("Setting bits from %d/%d.", | |
1340 obj_card_num - _bottom_card_num, | |
1341 obj_last_card_num - _bottom_card_num); | |
1342 */ | |
1343 for (intptr_t j = obj_card_num; j <= obj_last_card_num; j++) { | |
1344 _card_bm->par_at_put(j - _bottom_card_num, 1); | |
1345 } | |
1346 #endif | |
1347 } | |
1348 // In any case, we set the last card num. | |
1349 last_card_num = obj_last_card_num; | |
1350 | |
1030
dfdaf65c3423
6858886: G1: guarantee(_next_marked_bytes <= used(),"invariant") at heapRegion.hpp:359
apetrusenko
parents:
1023
diff
changeset
|
1351 marked_bytes += (size_t)obj_sz * HeapWordSize; |
342 | 1352 // Find the next marked object after this one. |
1353 start = _bm->getNextMarkedWordAddress(start + 1, nextTop); | |
1354 _changed = true; | |
1355 } | |
1356 // Handle the last range, if any. | |
1357 if (start_card_num != -1) | |
1358 mark_card_num_range(start_card_num, last_card_num); | |
1359 if (_final) { | |
1360 // Mark the allocated-since-marking portion... | |
1361 HeapWord* tp = hr->top(); | |
1362 if (nextTop < tp) { | |
1363 start_card_num = | |
1364 intptr_t(uintptr_t(nextTop) >> CardTableModRefBS::card_shift); | |
1365 last_card_num = | |
1366 intptr_t(uintptr_t(tp) >> CardTableModRefBS::card_shift); | |
1367 mark_card_num_range(start_card_num, last_card_num); | |
1368 // This definitely means the region has live objects. | |
829 | 1369 set_bit_for_region(hr); |
342 | 1370 } |
1371 } | |
1372 | |
1373 hr->add_to_marked_bytes(marked_bytes); | |
1374 // Update the live region bitmap. | |
1375 if (marked_bytes > 0) { | |
829 | 1376 set_bit_for_region(hr); |
342 | 1377 } |
1378 hr->set_top_at_conc_mark_count(nextTop); | |
1379 _tot_live += hr->next_live_bytes(); | |
1380 _tot_used += hr->used(); | |
1381 _words_done = words_done; | |
1382 | |
1383 if (!_final) { | |
1384 ++_regions_done; | |
1385 if (_regions_done % 10 == 0) { | |
1386 double end_vtime_sec = os::elapsedVTime(); | |
1387 double elapsed_vtime_sec = end_vtime_sec - _start_vtime_sec; | |
1388 if (elapsed_vtime_sec > (10.0 / 1000.0)) { | |
1389 jlong sleep_time_ms = | |
1390 (jlong) (elapsed_vtime_sec * _cm->cleanup_sleep_factor() * 1000.0); | |
1391 os::sleep(Thread::current(), sleep_time_ms, false); | |
1392 _start_vtime_sec = end_vtime_sec; | |
1393 } | |
1394 } | |
1395 } | |
1396 | |
1397 return false; | |
1398 } | |
1399 | |
1400 bool changed() { return _changed; } | |
1401 void reset() { _changed = false; _words_done = 0; } | |
1402 void no_yield() { _yield = false; } | |
1403 size_t words_done() { return _words_done; } | |
1404 size_t tot_live() { return _tot_live; } | |
1405 size_t tot_used() { return _tot_used; } | |
1406 }; | |
1407 | |
1408 | |
1409 void ConcurrentMark::calcDesiredRegions() { | |
1410 _region_bm.clear(); | |
1411 _card_bm.clear(); | |
1412 CalcLiveObjectsClosure calccl(false /*final*/, | |
1413 nextMarkBitMap(), this, | |
936 | 1414 &_region_bm, &_card_bm); |
342 | 1415 G1CollectedHeap *g1h = G1CollectedHeap::heap(); |
1416 g1h->heap_region_iterate(&calccl); | |
1417 | |
1418 do { | |
1419 calccl.reset(); | |
1420 g1h->heap_region_iterate(&calccl); | |
1421 } while (calccl.changed()); | |
1422 } | |
1423 | |
1424 class G1ParFinalCountTask: public AbstractGangTask { | |
1425 protected: | |
1426 G1CollectedHeap* _g1h; | |
1427 CMBitMap* _bm; | |
1428 size_t _n_workers; | |
1429 size_t *_live_bytes; | |
1430 size_t *_used_bytes; | |
1431 BitMap* _region_bm; | |
1432 BitMap* _card_bm; | |
1433 public: | |
1434 G1ParFinalCountTask(G1CollectedHeap* g1h, CMBitMap* bm, | |
1435 BitMap* region_bm, BitMap* card_bm) : | |
1436 AbstractGangTask("G1 final counting"), _g1h(g1h), | |
1437 _bm(bm), _region_bm(region_bm), _card_bm(card_bm) | |
1438 { | |
1439 if (ParallelGCThreads > 0) | |
1440 _n_workers = _g1h->workers()->total_workers(); | |
1441 else | |
1442 _n_workers = 1; | |
1443 _live_bytes = NEW_C_HEAP_ARRAY(size_t, _n_workers); | |
1444 _used_bytes = NEW_C_HEAP_ARRAY(size_t, _n_workers); | |
1445 } | |
1446 | |
1447 ~G1ParFinalCountTask() { | |
1448 FREE_C_HEAP_ARRAY(size_t, _live_bytes); | |
1449 FREE_C_HEAP_ARRAY(size_t, _used_bytes); | |
1450 } | |
1451 | |
1452 void work(int i) { | |
1453 CalcLiveObjectsClosure calccl(true /*final*/, | |
1454 _bm, _g1h->concurrent_mark(), | |
936 | 1455 _region_bm, _card_bm); |
342 | 1456 calccl.no_yield(); |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
1457 if (G1CollectedHeap::use_parallel_gc_threads()) { |
355 | 1458 _g1h->heap_region_par_iterate_chunked(&calccl, i, |
1459 HeapRegion::FinalCountClaimValue); | |
342 | 1460 } else { |
1461 _g1h->heap_region_iterate(&calccl); | |
1462 } | |
1463 assert(calccl.complete(), "Shouldn't have yielded!"); | |
1464 | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
1465 assert((size_t) i < _n_workers, "invariant"); |
342 | 1466 _live_bytes[i] = calccl.tot_live(); |
1467 _used_bytes[i] = calccl.tot_used(); | |
1468 } | |
1469 size_t live_bytes() { | |
1470 size_t live_bytes = 0; | |
1471 for (size_t i = 0; i < _n_workers; ++i) | |
1472 live_bytes += _live_bytes[i]; | |
1473 return live_bytes; | |
1474 } | |
1475 size_t used_bytes() { | |
1476 size_t used_bytes = 0; | |
1477 for (size_t i = 0; i < _n_workers; ++i) | |
1478 used_bytes += _used_bytes[i]; | |
1479 return used_bytes; | |
1480 } | |
1481 }; | |
1482 | |
1483 class G1ParNoteEndTask; | |
1484 | |
1485 class G1NoteEndOfConcMarkClosure : public HeapRegionClosure { | |
1486 G1CollectedHeap* _g1; | |
1487 int _worker_num; | |
1488 size_t _max_live_bytes; | |
1489 size_t _regions_claimed; | |
1490 size_t _freed_bytes; | |
1491 size_t _cleared_h_regions; | |
1492 size_t _freed_regions; | |
1493 UncleanRegionList* _unclean_region_list; | |
1494 double _claimed_region_time; | |
1495 double _max_region_time; | |
1496 | |
1497 public: | |
1498 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, | |
1499 UncleanRegionList* list, | |
1500 int worker_num); | |
1501 size_t freed_bytes() { return _freed_bytes; } | |
1502 size_t cleared_h_regions() { return _cleared_h_regions; } | |
1503 size_t freed_regions() { return _freed_regions; } | |
1504 UncleanRegionList* unclean_region_list() { | |
1505 return _unclean_region_list; | |
1506 } | |
1507 | |
1508 bool doHeapRegion(HeapRegion *r); | |
1509 | |
1510 size_t max_live_bytes() { return _max_live_bytes; } | |
1511 size_t regions_claimed() { return _regions_claimed; } | |
1512 double claimed_region_time_sec() { return _claimed_region_time; } | |
1513 double max_region_time_sec() { return _max_region_time; } | |
1514 }; | |
1515 | |
1516 class G1ParNoteEndTask: public AbstractGangTask { | |
1517 friend class G1NoteEndOfConcMarkClosure; | |
1518 protected: | |
1519 G1CollectedHeap* _g1h; | |
1520 size_t _max_live_bytes; | |
1521 size_t _freed_bytes; | |
1522 ConcurrentMark::ParCleanupThreadState** _par_cleanup_thread_state; | |
1523 public: | |
1524 G1ParNoteEndTask(G1CollectedHeap* g1h, | |
1525 ConcurrentMark::ParCleanupThreadState** | |
1526 par_cleanup_thread_state) : | |
1527 AbstractGangTask("G1 note end"), _g1h(g1h), | |
1528 _max_live_bytes(0), _freed_bytes(0), | |
1529 _par_cleanup_thread_state(par_cleanup_thread_state) | |
1530 {} | |
1531 | |
1532 void work(int i) { | |
1533 double start = os::elapsedTime(); | |
1534 G1NoteEndOfConcMarkClosure g1_note_end(_g1h, | |
1535 &_par_cleanup_thread_state[i]->list, | |
1536 i); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
1537 if (G1CollectedHeap::use_parallel_gc_threads()) { |
355 | 1538 _g1h->heap_region_par_iterate_chunked(&g1_note_end, i, |
1539 HeapRegion::NoteEndClaimValue); | |
342 | 1540 } else { |
1541 _g1h->heap_region_iterate(&g1_note_end); | |
1542 } | |
1543 assert(g1_note_end.complete(), "Shouldn't have yielded!"); | |
1544 | |
1545 // Now finish up freeing the current thread's regions. | |
1546 _g1h->finish_free_region_work(g1_note_end.freed_bytes(), | |
1547 g1_note_end.cleared_h_regions(), | |
1548 0, NULL); | |
1549 { | |
1550 MutexLockerEx x(ParGCRareEvent_lock, Mutex::_no_safepoint_check_flag); | |
1551 _max_live_bytes += g1_note_end.max_live_bytes(); | |
1552 _freed_bytes += g1_note_end.freed_bytes(); | |
1553 } | |
1554 double end = os::elapsedTime(); | |
1555 if (G1PrintParCleanupStats) { | |
1556 gclog_or_tty->print(" Worker thread %d [%8.3f..%8.3f = %8.3f ms] " | |
1557 "claimed %d regions (tot = %8.3f ms, max = %8.3f ms).\n", | |
1558 i, start, end, (end-start)*1000.0, | |
1559 g1_note_end.regions_claimed(), | |
1560 g1_note_end.claimed_region_time_sec()*1000.0, | |
1561 g1_note_end.max_region_time_sec()*1000.0); | |
1562 } | |
1563 } | |
1564 size_t max_live_bytes() { return _max_live_bytes; } | |
1565 size_t freed_bytes() { return _freed_bytes; } | |
1566 }; | |
1567 | |
1568 class G1ParScrubRemSetTask: public AbstractGangTask { | |
1569 protected: | |
1570 G1RemSet* _g1rs; | |
1571 BitMap* _region_bm; | |
1572 BitMap* _card_bm; | |
1573 public: | |
1574 G1ParScrubRemSetTask(G1CollectedHeap* g1h, | |
1575 BitMap* region_bm, BitMap* card_bm) : | |
1576 AbstractGangTask("G1 ScrubRS"), _g1rs(g1h->g1_rem_set()), | |
1577 _region_bm(region_bm), _card_bm(card_bm) | |
1578 {} | |
1579 | |
1580 void work(int i) { | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
1581 if (G1CollectedHeap::use_parallel_gc_threads()) { |
355 | 1582 _g1rs->scrub_par(_region_bm, _card_bm, i, |
1583 HeapRegion::ScrubRemSetClaimValue); | |
342 | 1584 } else { |
1585 _g1rs->scrub(_region_bm, _card_bm); | |
1586 } | |
1587 } | |
1588 | |
1589 }; | |
1590 | |
1591 G1NoteEndOfConcMarkClosure:: | |
1592 G1NoteEndOfConcMarkClosure(G1CollectedHeap* g1, | |
1593 UncleanRegionList* list, | |
1594 int worker_num) | |
1595 : _g1(g1), _worker_num(worker_num), | |
1596 _max_live_bytes(0), _regions_claimed(0), | |
1597 _freed_bytes(0), _cleared_h_regions(0), _freed_regions(0), | |
1598 _claimed_region_time(0.0), _max_region_time(0.0), | |
1599 _unclean_region_list(list) | |
1600 {} | |
1601 | |
1602 bool G1NoteEndOfConcMarkClosure::doHeapRegion(HeapRegion *r) { | |
1603 // We use a claim value of zero here because all regions | |
1604 // were claimed with value 1 in the FinalCount task. | |
1605 r->reset_gc_time_stamp(); | |
1606 if (!r->continuesHumongous()) { | |
1607 double start = os::elapsedTime(); | |
1608 _regions_claimed++; | |
1609 r->note_end_of_marking(); | |
1610 _max_live_bytes += r->max_live_bytes(); | |
1611 _g1->free_region_if_totally_empty_work(r, | |
1612 _freed_bytes, | |
1613 _cleared_h_regions, | |
1614 _freed_regions, | |
1615 _unclean_region_list, | |
1616 true /*par*/); | |
1617 double region_time = (os::elapsedTime() - start); | |
1618 _claimed_region_time += region_time; | |
1619 if (region_time > _max_region_time) _max_region_time = region_time; | |
1620 } | |
1621 return false; | |
1622 } | |
1623 | |
1624 void ConcurrentMark::cleanup() { | |
1625 // world is stopped at this checkpoint | |
1626 assert(SafepointSynchronize::is_at_safepoint(), | |
1627 "world should be stopped"); | |
1628 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
1629 | |
1630 // If a full collection has happened, we shouldn't do this. | |
1631 if (has_aborted()) { | |
1632 g1h->set_marking_complete(); // So bitmap clearing isn't confused | |
1633 return; | |
1634 } | |
1635 | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1636 if (VerifyDuringGC) { |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1637 HandleMark hm; // handle scope |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1638 gclog_or_tty->print(" VerifyDuringGC:(before)"); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1639 Universe::heap()->prepare_for_verify(); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1640 Universe::verify(/* allow dirty */ true, |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1641 /* silent */ false, |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1642 /* prev marking */ true); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1643 } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1644 |
342 | 1645 G1CollectorPolicy* g1p = G1CollectedHeap::heap()->g1_policy(); |
1646 g1p->record_concurrent_mark_cleanup_start(); | |
1647 | |
1648 double start = os::elapsedTime(); | |
1649 | |
1650 // Do counting once more with the world stopped for good measure. | |
1651 G1ParFinalCountTask g1_par_count_task(g1h, nextMarkBitMap(), | |
1652 &_region_bm, &_card_bm); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
1653 if (G1CollectedHeap::use_parallel_gc_threads()) { |
355 | 1654 assert(g1h->check_heap_region_claim_values( |
1655 HeapRegion::InitialClaimValue), | |
1656 "sanity check"); | |
1657 | |
342 | 1658 int n_workers = g1h->workers()->total_workers(); |
1659 g1h->set_par_threads(n_workers); | |
1660 g1h->workers()->run_task(&g1_par_count_task); | |
1661 g1h->set_par_threads(0); | |
355 | 1662 |
1663 assert(g1h->check_heap_region_claim_values( | |
1664 HeapRegion::FinalCountClaimValue), | |
1665 "sanity check"); | |
342 | 1666 } else { |
1667 g1_par_count_task.work(0); | |
1668 } | |
1669 | |
1670 size_t known_garbage_bytes = | |
1671 g1_par_count_task.used_bytes() - g1_par_count_task.live_bytes(); | |
1672 #if 0 | |
1673 gclog_or_tty->print_cr("used %1.2lf, live %1.2lf, garbage %1.2lf", | |
1674 (double) g1_par_count_task.used_bytes() / (double) (1024 * 1024), | |
1675 (double) g1_par_count_task.live_bytes() / (double) (1024 * 1024), | |
1676 (double) known_garbage_bytes / (double) (1024 * 1024)); | |
1677 #endif // 0 | |
1678 g1p->set_known_garbage_bytes(known_garbage_bytes); | |
1679 | |
1680 size_t start_used_bytes = g1h->used(); | |
1681 _at_least_one_mark_complete = true; | |
1682 g1h->set_marking_complete(); | |
1683 | |
1684 double count_end = os::elapsedTime(); | |
1685 double this_final_counting_time = (count_end - start); | |
1686 if (G1PrintParCleanupStats) { | |
1687 gclog_or_tty->print_cr("Cleanup:"); | |
1688 gclog_or_tty->print_cr(" Finalize counting: %8.3f ms", | |
1689 this_final_counting_time*1000.0); | |
1690 } | |
1691 _total_counting_time += this_final_counting_time; | |
1692 | |
1693 // Install newly created mark bitMap as "prev". | |
1694 swapMarkBitMaps(); | |
1695 | |
1696 g1h->reset_gc_time_stamp(); | |
1697 | |
1698 // Note end of marking in all heap regions. | |
1699 double note_end_start = os::elapsedTime(); | |
1700 G1ParNoteEndTask g1_par_note_end_task(g1h, _par_cleanup_thread_state); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
1701 if (G1CollectedHeap::use_parallel_gc_threads()) { |
342 | 1702 int n_workers = g1h->workers()->total_workers(); |
1703 g1h->set_par_threads(n_workers); | |
1704 g1h->workers()->run_task(&g1_par_note_end_task); | |
1705 g1h->set_par_threads(0); | |
355 | 1706 |
1707 assert(g1h->check_heap_region_claim_values(HeapRegion::NoteEndClaimValue), | |
1708 "sanity check"); | |
342 | 1709 } else { |
1710 g1_par_note_end_task.work(0); | |
1711 } | |
1712 g1h->set_unclean_regions_coming(true); | |
1713 double note_end_end = os::elapsedTime(); | |
1714 // Tell the mutators that there might be unclean regions coming... | |
1715 if (G1PrintParCleanupStats) { | |
1716 gclog_or_tty->print_cr(" note end of marking: %8.3f ms.", | |
1717 (note_end_end - note_end_start)*1000.0); | |
1718 } | |
1719 | |
355 | 1720 |
342 | 1721 // call below, since it affects the metric by which we sort the heap |
1722 // regions. | |
1723 if (G1ScrubRemSets) { | |
1724 double rs_scrub_start = os::elapsedTime(); | |
1725 G1ParScrubRemSetTask g1_par_scrub_rs_task(g1h, &_region_bm, &_card_bm); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
1726 if (G1CollectedHeap::use_parallel_gc_threads()) { |
342 | 1727 int n_workers = g1h->workers()->total_workers(); |
1728 g1h->set_par_threads(n_workers); | |
1729 g1h->workers()->run_task(&g1_par_scrub_rs_task); | |
1730 g1h->set_par_threads(0); | |
355 | 1731 |
1732 assert(g1h->check_heap_region_claim_values( | |
1733 HeapRegion::ScrubRemSetClaimValue), | |
1734 "sanity check"); | |
342 | 1735 } else { |
1736 g1_par_scrub_rs_task.work(0); | |
1737 } | |
1738 | |
1739 double rs_scrub_end = os::elapsedTime(); | |
1740 double this_rs_scrub_time = (rs_scrub_end - rs_scrub_start); | |
1741 _total_rs_scrub_time += this_rs_scrub_time; | |
1742 } | |
1743 | |
1744 // this will also free any regions totally full of garbage objects, | |
1745 // and sort the regions. | |
1746 g1h->g1_policy()->record_concurrent_mark_cleanup_end( | |
1747 g1_par_note_end_task.freed_bytes(), | |
1748 g1_par_note_end_task.max_live_bytes()); | |
1749 | |
1750 // Statistics. | |
1751 double end = os::elapsedTime(); | |
1752 _cleanup_times.add((end - start) * 1000.0); | |
1753 | |
1754 // G1CollectedHeap::heap()->print(); | |
1755 // gclog_or_tty->print_cr("HEAP GC TIME STAMP : %d", | |
1756 // G1CollectedHeap::heap()->get_gc_time_stamp()); | |
1757 | |
1758 if (PrintGC || PrintGCDetails) { | |
1759 g1h->print_size_transition(gclog_or_tty, | |
1760 start_used_bytes, | |
1761 g1h->used(), | |
1762 g1h->capacity()); | |
1763 } | |
1764 | |
1765 size_t cleaned_up_bytes = start_used_bytes - g1h->used(); | |
1766 g1p->decrease_known_garbage_bytes(cleaned_up_bytes); | |
1767 | |
1768 // We need to make this be a "collection" so any collection pause that | |
1769 // races with it goes around and waits for completeCleanup to finish. | |
1770 g1h->increment_total_collections(); | |
1771 | |
751 | 1772 if (VerifyDuringGC) { |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1773 HandleMark hm; // handle scope |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1774 gclog_or_tty->print(" VerifyDuringGC:(after)"); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1775 Universe::heap()->prepare_for_verify(); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1776 Universe::verify(/* allow dirty */ true, |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1777 /* silent */ false, |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1778 /* prev marking */ true); |
342 | 1779 } |
1780 } | |
1781 | |
1782 void ConcurrentMark::completeCleanup() { | |
1783 // A full collection intervened. | |
1784 if (has_aborted()) return; | |
1785 | |
1786 int first = 0; | |
1787 int last = (int)MAX2(ParallelGCThreads, (size_t)1); | |
1788 for (int t = 0; t < last; t++) { | |
1789 UncleanRegionList* list = &_par_cleanup_thread_state[t]->list; | |
1790 assert(list->well_formed(), "Inv"); | |
1791 HeapRegion* hd = list->hd(); | |
1792 while (hd != NULL) { | |
1793 // Now finish up the other stuff. | |
1794 hd->rem_set()->clear(); | |
1795 HeapRegion* next_hd = hd->next_from_unclean_list(); | |
1796 (void)list->pop(); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
1797 assert(list->hd() == next_hd, "how not?"); |
342 | 1798 _g1h->put_region_on_unclean_list(hd); |
1799 if (!hd->isHumongous()) { | |
1800 // Add this to the _free_regions count by 1. | |
1801 _g1h->finish_free_region_work(0, 0, 1, NULL); | |
1802 } | |
1803 hd = list->hd(); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
1804 assert(hd == next_hd, "how not?"); |
342 | 1805 } |
1806 } | |
1807 } | |
1808 | |
1809 | |
1810 class G1CMIsAliveClosure: public BoolObjectClosure { | |
1811 G1CollectedHeap* _g1; | |
1812 public: | |
1813 G1CMIsAliveClosure(G1CollectedHeap* g1) : | |
1814 _g1(g1) | |
1815 {} | |
1816 | |
1817 void do_object(oop obj) { | |
1818 assert(false, "not to be invoked"); | |
1819 } | |
1820 bool do_object_b(oop obj) { | |
1821 HeapWord* addr = (HeapWord*)obj; | |
1822 return addr != NULL && | |
1823 (!_g1->is_in_g1_reserved(addr) || !_g1->is_obj_ill(obj)); | |
1824 } | |
1825 }; | |
1826 | |
1827 class G1CMKeepAliveClosure: public OopClosure { | |
1828 G1CollectedHeap* _g1; | |
1829 ConcurrentMark* _cm; | |
1830 CMBitMap* _bitMap; | |
1831 public: | |
1832 G1CMKeepAliveClosure(G1CollectedHeap* g1, ConcurrentMark* cm, | |
1833 CMBitMap* bitMap) : | |
1834 _g1(g1), _cm(cm), | |
1835 _bitMap(bitMap) {} | |
1836 | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1837 virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1838 virtual void do_oop( oop* p) { do_oop_work(p); } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1839 |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1840 template <class T> void do_oop_work(T* p) { |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
1841 oop thisOop = oopDesc::load_decode_heap_oop(p); |
342 | 1842 HeapWord* addr = (HeapWord*)thisOop; |
1843 if (_g1->is_in_g1_reserved(addr) && _g1->is_obj_ill(thisOop)) { | |
1844 _bitMap->mark(addr); | |
1845 _cm->mark_stack_push(thisOop); | |
1846 } | |
1847 } | |
1848 }; | |
1849 | |
1850 class G1CMDrainMarkingStackClosure: public VoidClosure { | |
1851 CMMarkStack* _markStack; | |
1852 CMBitMap* _bitMap; | |
1853 G1CMKeepAliveClosure* _oopClosure; | |
1854 public: | |
1855 G1CMDrainMarkingStackClosure(CMBitMap* bitMap, CMMarkStack* markStack, | |
1856 G1CMKeepAliveClosure* oopClosure) : | |
1857 _bitMap(bitMap), | |
1858 _markStack(markStack), | |
1859 _oopClosure(oopClosure) | |
1860 {} | |
1861 | |
1862 void do_void() { | |
1863 _markStack->drain((OopClosure*)_oopClosure, _bitMap, false); | |
1864 } | |
1865 }; | |
1866 | |
1867 void ConcurrentMark::weakRefsWork(bool clear_all_soft_refs) { | |
1868 ResourceMark rm; | |
1869 HandleMark hm; | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
355
diff
changeset
|
1870 G1CollectedHeap* g1h = G1CollectedHeap::heap(); |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
355
diff
changeset
|
1871 ReferenceProcessor* rp = g1h->ref_processor(); |
342 | 1872 |
1873 // Process weak references. | |
457
27a80744a83b
6778647: snap(), snap_policy() should be renamed setup(), setup_policy()
ysr
parents:
453
diff
changeset
|
1874 rp->setup_policy(clear_all_soft_refs); |
342 | 1875 assert(_markStack.isEmpty(), "mark stack should be empty"); |
1876 | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
355
diff
changeset
|
1877 G1CMIsAliveClosure g1IsAliveClosure (g1h); |
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
355
diff
changeset
|
1878 G1CMKeepAliveClosure g1KeepAliveClosure(g1h, this, nextMarkBitMap()); |
342 | 1879 G1CMDrainMarkingStackClosure |
1880 g1DrainMarkingStackClosure(nextMarkBitMap(), &_markStack, | |
1881 &g1KeepAliveClosure); | |
1882 | |
1883 // XXXYYY Also: copy the parallel ref processing code from CMS. | |
453
c96030fff130
6684579: SoftReference processing can be made more efficient
ysr
parents:
355
diff
changeset
|
1884 rp->process_discovered_references(&g1IsAliveClosure, |
342 | 1885 &g1KeepAliveClosure, |
1886 &g1DrainMarkingStackClosure, | |
1887 NULL); | |
1888 assert(_markStack.overflow() || _markStack.isEmpty(), | |
1889 "mark stack should be empty (unless it overflowed)"); | |
1890 if (_markStack.overflow()) { | |
1891 set_has_overflown(); | |
1892 } | |
1893 | |
1894 rp->enqueue_discovered_references(); | |
1895 rp->verify_no_references_recorded(); | |
1896 assert(!rp->discovery_enabled(), "should have been disabled"); | |
1897 | |
1898 // Now clean up stale oops in SymbolTable and StringTable | |
1899 SymbolTable::unlink(&g1IsAliveClosure); | |
1900 StringTable::unlink(&g1IsAliveClosure); | |
1901 } | |
1902 | |
1903 void ConcurrentMark::swapMarkBitMaps() { | |
1904 CMBitMapRO* temp = _prevMarkBitMap; | |
1905 _prevMarkBitMap = (CMBitMapRO*)_nextMarkBitMap; | |
1906 _nextMarkBitMap = (CMBitMap*) temp; | |
1907 } | |
1908 | |
1909 class CMRemarkTask: public AbstractGangTask { | |
1910 private: | |
1911 ConcurrentMark *_cm; | |
1912 | |
1913 public: | |
1914 void work(int worker_i) { | |
1915 // Since all available tasks are actually started, we should | |
1916 // only proceed if we're supposed to be actived. | |
1917 if ((size_t)worker_i < _cm->active_tasks()) { | |
1918 CMTask* task = _cm->task(worker_i); | |
1919 task->record_start_time(); | |
1920 do { | |
1921 task->do_marking_step(1000000000.0 /* something very large */); | |
1922 } while (task->has_aborted() && !_cm->has_overflown()); | |
1923 // If we overflow, then we do not want to restart. We instead | |
1924 // want to abort remark and do concurrent marking again. | |
1925 task->record_end_time(); | |
1926 } | |
1927 } | |
1928 | |
1929 CMRemarkTask(ConcurrentMark* cm) : | |
1930 AbstractGangTask("Par Remark"), _cm(cm) { } | |
1931 }; | |
1932 | |
1933 void ConcurrentMark::checkpointRootsFinalWork() { | |
1934 ResourceMark rm; | |
1935 HandleMark hm; | |
1936 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
1937 | |
1938 g1h->ensure_parsability(false); | |
1939 | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
1940 if (G1CollectedHeap::use_parallel_gc_threads()) { |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
1941 G1CollectedHeap::StrongRootsScope srs(g1h); |
342 | 1942 // this is remark, so we'll use up all available threads |
1943 int active_workers = ParallelGCThreads; | |
1944 set_phase(active_workers, false); | |
1945 | |
1946 CMRemarkTask remarkTask(this); | |
1947 // We will start all available threads, even if we decide that the | |
1948 // active_workers will be fewer. The extra ones will just bail out | |
1949 // immediately. | |
1950 int n_workers = g1h->workers()->total_workers(); | |
1951 g1h->set_par_threads(n_workers); | |
1952 g1h->workers()->run_task(&remarkTask); | |
1953 g1h->set_par_threads(0); | |
1954 } else { | |
989
148e5441d916
6863023: need non-perm oops in code cache for JSR 292
jrose
parents:
912
diff
changeset
|
1955 G1CollectedHeap::StrongRootsScope srs(g1h); |
342 | 1956 // this is remark, so we'll use up all available threads |
1957 int active_workers = 1; | |
1958 set_phase(active_workers, false); | |
1959 | |
1960 CMRemarkTask remarkTask(this); | |
1961 // We will start all available threads, even if we decide that the | |
1962 // active_workers will be fewer. The extra ones will just bail out | |
1963 // immediately. | |
1964 remarkTask.work(0); | |
1965 } | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
1966 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
1967 guarantee(satb_mq_set.completed_buffers_num() == 0, "invariant"); |
342 | 1968 |
1969 print_stats(); | |
1970 | |
1971 if (!restart_for_overflow()) | |
1972 set_non_marking_state(); | |
1973 | |
1974 #if VERIFY_OBJS_PROCESSED | |
1975 if (_scan_obj_cl.objs_processed != ThreadLocalObjQueue::objs_enqueued) { | |
1976 gclog_or_tty->print_cr("Processed = %d, enqueued = %d.", | |
1977 _scan_obj_cl.objs_processed, | |
1978 ThreadLocalObjQueue::objs_enqueued); | |
1979 guarantee(_scan_obj_cl.objs_processed == | |
1980 ThreadLocalObjQueue::objs_enqueued, | |
1981 "Different number of objs processed and enqueued."); | |
1982 } | |
1983 #endif | |
1984 } | |
1985 | |
1044 | 1986 #ifndef PRODUCT |
1987 | |
1388 | 1988 class PrintReachableOopClosure: public OopClosure { |
342 | 1989 private: |
1990 G1CollectedHeap* _g1h; | |
1991 CMBitMapRO* _bitmap; | |
1992 outputStream* _out; | |
1044 | 1993 bool _use_prev_marking; |
1388 | 1994 bool _all; |
342 | 1995 |
1996 public: | |
1388 | 1997 PrintReachableOopClosure(CMBitMapRO* bitmap, |
1998 outputStream* out, | |
1999 bool use_prev_marking, | |
2000 bool all) : | |
1044 | 2001 _g1h(G1CollectedHeap::heap()), |
1388 | 2002 _bitmap(bitmap), _out(out), _use_prev_marking(use_prev_marking), _all(all) { } |
342 | 2003 |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2004 void do_oop(narrowOop* p) { do_oop_work(p); } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2005 void do_oop( oop* p) { do_oop_work(p); } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2006 |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2007 template <class T> void do_oop_work(T* p) { |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2008 oop obj = oopDesc::load_decode_heap_oop(p); |
342 | 2009 const char* str = NULL; |
2010 const char* str2 = ""; | |
2011 | |
1388 | 2012 if (obj == NULL) { |
2013 str = ""; | |
2014 } else if (!_g1h->is_in_g1_reserved(obj)) { | |
2015 str = " O"; | |
2016 } else { | |
342 | 2017 HeapRegion* hr = _g1h->heap_region_containing(obj); |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2018 guarantee(hr != NULL, "invariant"); |
1044 | 2019 bool over_tams = false; |
2020 if (_use_prev_marking) { | |
2021 over_tams = hr->obj_allocated_since_prev_marking(obj); | |
2022 } else { | |
2023 over_tams = hr->obj_allocated_since_next_marking(obj); | |
2024 } | |
1388 | 2025 bool marked = _bitmap->isMarked((HeapWord*) obj); |
1044 | 2026 |
2027 if (over_tams) { | |
1388 | 2028 str = " >"; |
2029 if (marked) { | |
342 | 2030 str2 = " AND MARKED"; |
1044 | 2031 } |
1388 | 2032 } else if (marked) { |
2033 str = " M"; | |
1044 | 2034 } else { |
1388 | 2035 str = " NOT"; |
1044 | 2036 } |
342 | 2037 } |
2038 | |
1388 | 2039 _out->print_cr(" "PTR_FORMAT": "PTR_FORMAT"%s%s", |
342 | 2040 p, (void*) obj, str, str2); |
2041 } | |
2042 }; | |
2043 | |
1388 | 2044 class PrintReachableObjectClosure : public ObjectClosure { |
342 | 2045 private: |
1044 | 2046 CMBitMapRO* _bitmap; |
342 | 2047 outputStream* _out; |
1044 | 2048 bool _use_prev_marking; |
1388 | 2049 bool _all; |
2050 HeapRegion* _hr; | |
342 | 2051 |
2052 public: | |
1388 | 2053 PrintReachableObjectClosure(CMBitMapRO* bitmap, |
2054 outputStream* out, | |
2055 bool use_prev_marking, | |
2056 bool all, | |
2057 HeapRegion* hr) : | |
2058 _bitmap(bitmap), _out(out), | |
2059 _use_prev_marking(use_prev_marking), _all(all), _hr(hr) { } | |
2060 | |
2061 void do_object(oop o) { | |
2062 bool over_tams; | |
2063 if (_use_prev_marking) { | |
2064 over_tams = _hr->obj_allocated_since_prev_marking(o); | |
2065 } else { | |
2066 over_tams = _hr->obj_allocated_since_next_marking(o); | |
2067 } | |
2068 bool marked = _bitmap->isMarked((HeapWord*) o); | |
2069 bool print_it = _all || over_tams || marked; | |
2070 | |
2071 if (print_it) { | |
2072 _out->print_cr(" "PTR_FORMAT"%s", | |
2073 o, (over_tams) ? " >" : (marked) ? " M" : ""); | |
2074 PrintReachableOopClosure oopCl(_bitmap, _out, _use_prev_marking, _all); | |
2075 o->oop_iterate(&oopCl); | |
2076 } | |
342 | 2077 } |
2078 }; | |
2079 | |
1388 | 2080 class PrintReachableRegionClosure : public HeapRegionClosure { |
342 | 2081 private: |
1044 | 2082 CMBitMapRO* _bitmap; |
342 | 2083 outputStream* _out; |
1044 | 2084 bool _use_prev_marking; |
1388 | 2085 bool _all; |
342 | 2086 |
2087 public: | |
2088 bool doHeapRegion(HeapRegion* hr) { | |
2089 HeapWord* b = hr->bottom(); | |
2090 HeapWord* e = hr->end(); | |
2091 HeapWord* t = hr->top(); | |
1044 | 2092 HeapWord* p = NULL; |
2093 if (_use_prev_marking) { | |
2094 p = hr->prev_top_at_mark_start(); | |
2095 } else { | |
2096 p = hr->next_top_at_mark_start(); | |
2097 } | |
342 | 2098 _out->print_cr("** ["PTR_FORMAT", "PTR_FORMAT"] top: "PTR_FORMAT" " |
1044 | 2099 "TAMS: "PTR_FORMAT, b, e, t, p); |
1388 | 2100 _out->cr(); |
2101 | |
2102 HeapWord* from = b; | |
2103 HeapWord* to = t; | |
2104 | |
2105 if (to > from) { | |
2106 _out->print_cr("Objects in ["PTR_FORMAT", "PTR_FORMAT"]", from, to); | |
2107 _out->cr(); | |
2108 PrintReachableObjectClosure ocl(_bitmap, _out, | |
2109 _use_prev_marking, _all, hr); | |
2110 hr->object_iterate_mem_careful(MemRegion(from, to), &ocl); | |
2111 _out->cr(); | |
2112 } | |
342 | 2113 |
2114 return false; | |
2115 } | |
2116 | |
1388 | 2117 PrintReachableRegionClosure(CMBitMapRO* bitmap, |
2118 outputStream* out, | |
2119 bool use_prev_marking, | |
2120 bool all) : | |
2121 _bitmap(bitmap), _out(out), _use_prev_marking(use_prev_marking), _all(all) { } | |
342 | 2122 }; |
2123 | |
1388 | 2124 void ConcurrentMark::print_reachable(const char* str, |
2125 bool use_prev_marking, | |
2126 bool all) { | |
2127 gclog_or_tty->cr(); | |
2128 gclog_or_tty->print_cr("== Doing heap dump... "); | |
1044 | 2129 |
2130 if (G1PrintReachableBaseFile == NULL) { | |
2131 gclog_or_tty->print_cr(" #### error: no base file defined"); | |
2132 return; | |
2133 } | |
2134 | |
2135 if (strlen(G1PrintReachableBaseFile) + 1 + strlen(str) > | |
2136 (JVM_MAXPATHLEN - 1)) { | |
2137 gclog_or_tty->print_cr(" #### error: file name too long"); | |
2138 return; | |
2139 } | |
2140 | |
2141 char file_name[JVM_MAXPATHLEN]; | |
2142 sprintf(file_name, "%s.%s", G1PrintReachableBaseFile, str); | |
2143 gclog_or_tty->print_cr(" dumping to file %s", file_name); | |
2144 | |
2145 fileStream fout(file_name); | |
2146 if (!fout.is_open()) { | |
2147 gclog_or_tty->print_cr(" #### error: could not open file"); | |
2148 return; | |
2149 } | |
2150 | |
2151 outputStream* out = &fout; | |
2152 | |
2153 CMBitMapRO* bitmap = NULL; | |
2154 if (use_prev_marking) { | |
2155 bitmap = _prevMarkBitMap; | |
2156 } else { | |
2157 bitmap = _nextMarkBitMap; | |
2158 } | |
2159 | |
2160 out->print_cr("-- USING %s", (use_prev_marking) ? "PTAMS" : "NTAMS"); | |
2161 out->cr(); | |
2162 | |
1388 | 2163 out->print_cr("--- ITERATING OVER REGIONS"); |
1044 | 2164 out->cr(); |
1388 | 2165 PrintReachableRegionClosure rcl(bitmap, out, use_prev_marking, all); |
2166 _g1h->heap_region_iterate(&rcl); | |
1044 | 2167 out->cr(); |
2168 | |
2169 gclog_or_tty->print_cr(" done"); | |
1388 | 2170 gclog_or_tty->flush(); |
342 | 2171 } |
2172 | |
1044 | 2173 #endif // PRODUCT |
2174 | |
342 | 2175 // This note is for drainAllSATBBuffers and the code in between. |
2176 // In the future we could reuse a task to do this work during an | |
2177 // evacuation pause (since now tasks are not active and can be claimed | |
2178 // during an evacuation pause). This was a late change to the code and | |
2179 // is currently not being taken advantage of. | |
2180 | |
2181 class CMGlobalObjectClosure : public ObjectClosure { | |
2182 private: | |
2183 ConcurrentMark* _cm; | |
2184 | |
2185 public: | |
2186 void do_object(oop obj) { | |
2187 _cm->deal_with_reference(obj); | |
2188 } | |
2189 | |
2190 CMGlobalObjectClosure(ConcurrentMark* cm) : _cm(cm) { } | |
2191 }; | |
2192 | |
2193 void ConcurrentMark::deal_with_reference(oop obj) { | |
2194 if (verbose_high()) | |
2195 gclog_or_tty->print_cr("[global] we're dealing with reference "PTR_FORMAT, | |
2196 (void*) obj); | |
2197 | |
2198 | |
2199 HeapWord* objAddr = (HeapWord*) obj; | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2200 assert(obj->is_oop_or_null(true /* ignore mark word */), "Error"); |
342 | 2201 if (_g1h->is_in_g1_reserved(objAddr)) { |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2202 assert(obj != NULL, "is_in_g1_reserved should ensure this"); |
342 | 2203 HeapRegion* hr = _g1h->heap_region_containing(obj); |
2204 if (_g1h->is_obj_ill(obj, hr)) { | |
2205 if (verbose_high()) | |
2206 gclog_or_tty->print_cr("[global] "PTR_FORMAT" is not considered " | |
2207 "marked", (void*) obj); | |
2208 | |
2209 // we need to mark it first | |
2210 if (_nextMarkBitMap->parMark(objAddr)) { | |
2211 // No OrderAccess:store_load() is needed. It is implicit in the | |
2212 // CAS done in parMark(objAddr) above | |
2213 HeapWord* finger = _finger; | |
2214 if (objAddr < finger) { | |
2215 if (verbose_high()) | |
2216 gclog_or_tty->print_cr("[global] below the global finger " | |
2217 "("PTR_FORMAT"), pushing it", finger); | |
2218 if (!mark_stack_push(obj)) { | |
2219 if (verbose_low()) | |
2220 gclog_or_tty->print_cr("[global] global stack overflow during " | |
2221 "deal_with_reference"); | |
2222 } | |
2223 } | |
2224 } | |
2225 } | |
2226 } | |
2227 } | |
2228 | |
2229 void ConcurrentMark::drainAllSATBBuffers() { | |
2230 CMGlobalObjectClosure oc(this); | |
2231 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
2232 satb_mq_set.set_closure(&oc); | |
2233 | |
2234 while (satb_mq_set.apply_closure_to_completed_buffer()) { | |
2235 if (verbose_medium()) | |
2236 gclog_or_tty->print_cr("[global] processed an SATB buffer"); | |
2237 } | |
2238 | |
2239 // no need to check whether we should do this, as this is only | |
2240 // called during an evacuation pause | |
2241 satb_mq_set.iterate_closure_all_threads(); | |
2242 | |
2243 satb_mq_set.set_closure(NULL); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2244 assert(satb_mq_set.completed_buffers_num() == 0, "invariant"); |
342 | 2245 } |
2246 | |
2247 void ConcurrentMark::markPrev(oop p) { | |
2248 // Note we are overriding the read-only view of the prev map here, via | |
2249 // the cast. | |
2250 ((CMBitMap*)_prevMarkBitMap)->mark((HeapWord*)p); | |
2251 } | |
2252 | |
2253 void ConcurrentMark::clear(oop p) { | |
2254 assert(p != NULL && p->is_oop(), "expected an oop"); | |
2255 HeapWord* addr = (HeapWord*)p; | |
2256 assert(addr >= _nextMarkBitMap->startWord() || | |
2257 addr < _nextMarkBitMap->endWord(), "in a region"); | |
2258 | |
2259 _nextMarkBitMap->clear(addr); | |
2260 } | |
2261 | |
2262 void ConcurrentMark::clearRangeBothMaps(MemRegion mr) { | |
2263 // Note we are overriding the read-only view of the prev map here, via | |
2264 // the cast. | |
2265 ((CMBitMap*)_prevMarkBitMap)->clearRange(mr); | |
2266 _nextMarkBitMap->clearRange(mr); | |
2267 } | |
2268 | |
2269 HeapRegion* | |
2270 ConcurrentMark::claim_region(int task_num) { | |
2271 // "checkpoint" the finger | |
2272 HeapWord* finger = _finger; | |
2273 | |
2274 // _heap_end will not change underneath our feet; it only changes at | |
2275 // yield points. | |
2276 while (finger < _heap_end) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2277 assert(_g1h->is_in_g1_reserved(finger), "invariant"); |
342 | 2278 |
2279 // is the gap between reading the finger and doing the CAS too long? | |
2280 | |
2281 HeapRegion* curr_region = _g1h->heap_region_containing(finger); | |
2282 HeapWord* bottom = curr_region->bottom(); | |
2283 HeapWord* end = curr_region->end(); | |
2284 HeapWord* limit = curr_region->next_top_at_mark_start(); | |
2285 | |
2286 if (verbose_low()) | |
2287 gclog_or_tty->print_cr("[%d] curr_region = "PTR_FORMAT" " | |
2288 "["PTR_FORMAT", "PTR_FORMAT"), " | |
2289 "limit = "PTR_FORMAT, | |
2290 task_num, curr_region, bottom, end, limit); | |
2291 | |
2292 HeapWord* res = | |
2293 (HeapWord*) Atomic::cmpxchg_ptr(end, &_finger, finger); | |
2294 if (res == finger) { | |
2295 // we succeeded | |
2296 | |
2297 // notice that _finger == end cannot be guaranteed here since, | |
2298 // someone else might have moved the finger even further | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2299 assert(_finger >= end, "the finger should have moved forward"); |
342 | 2300 |
2301 if (verbose_low()) | |
2302 gclog_or_tty->print_cr("[%d] we were successful with region = " | |
2303 PTR_FORMAT, task_num, curr_region); | |
2304 | |
2305 if (limit > bottom) { | |
2306 if (verbose_low()) | |
2307 gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is not empty, " | |
2308 "returning it ", task_num, curr_region); | |
2309 return curr_region; | |
2310 } else { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2311 assert(limit == bottom, |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2312 "the region limit should be at bottom"); |
342 | 2313 if (verbose_low()) |
2314 gclog_or_tty->print_cr("[%d] region "PTR_FORMAT" is empty, " | |
2315 "returning NULL", task_num, curr_region); | |
2316 // we return NULL and the caller should try calling | |
2317 // claim_region() again. | |
2318 return NULL; | |
2319 } | |
2320 } else { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2321 assert(_finger > finger, "the finger should have moved forward"); |
342 | 2322 if (verbose_low()) |
2323 gclog_or_tty->print_cr("[%d] somebody else moved the finger, " | |
2324 "global finger = "PTR_FORMAT", " | |
2325 "our finger = "PTR_FORMAT, | |
2326 task_num, _finger, finger); | |
2327 | |
2328 // read it again | |
2329 finger = _finger; | |
2330 } | |
2331 } | |
2332 | |
2333 return NULL; | |
2334 } | |
2335 | |
2336 void ConcurrentMark::oops_do(OopClosure* cl) { | |
2337 if (_markStack.size() > 0 && verbose_low()) | |
2338 gclog_or_tty->print_cr("[global] scanning the global marking stack, " | |
2339 "size = %d", _markStack.size()); | |
2340 // we first iterate over the contents of the mark stack... | |
2341 _markStack.oops_do(cl); | |
2342 | |
2343 for (int i = 0; i < (int)_max_task_num; ++i) { | |
2344 OopTaskQueue* queue = _task_queues->queue((int)i); | |
2345 | |
2346 if (queue->size() > 0 && verbose_low()) | |
2347 gclog_or_tty->print_cr("[global] scanning task queue of task %d, " | |
2348 "size = %d", i, queue->size()); | |
2349 | |
2350 // ...then over the contents of the all the task queues. | |
2351 queue->oops_do(cl); | |
2352 } | |
2353 | |
2354 // finally, invalidate any entries that in the region stack that | |
2355 // point into the collection set | |
2356 if (_regionStack.invalidate_entries_into_cset()) { | |
2357 // otherwise, any gray objects copied during the evacuation pause | |
2358 // might not be visited. | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2359 assert(_should_gray_objects, "invariant"); |
342 | 2360 } |
2361 } | |
2362 | |
2363 void ConcurrentMark::clear_marking_state() { | |
2364 _markStack.setEmpty(); | |
2365 _markStack.clear_overflow(); | |
2366 _regionStack.setEmpty(); | |
2367 _regionStack.clear_overflow(); | |
2368 clear_has_overflown(); | |
2369 _finger = _heap_start; | |
2370 | |
2371 for (int i = 0; i < (int)_max_task_num; ++i) { | |
2372 OopTaskQueue* queue = _task_queues->queue(i); | |
2373 queue->set_empty(); | |
2374 } | |
2375 } | |
2376 | |
2377 void ConcurrentMark::print_stats() { | |
2378 if (verbose_stats()) { | |
2379 gclog_or_tty->print_cr("---------------------------------------------------------------------"); | |
2380 for (size_t i = 0; i < _active_tasks; ++i) { | |
2381 _tasks[i]->print_stats(); | |
2382 gclog_or_tty->print_cr("---------------------------------------------------------------------"); | |
2383 } | |
2384 } | |
2385 } | |
2386 | |
2387 class CSMarkOopClosure: public OopClosure { | |
2388 friend class CSMarkBitMapClosure; | |
2389 | |
2390 G1CollectedHeap* _g1h; | |
2391 CMBitMap* _bm; | |
2392 ConcurrentMark* _cm; | |
2393 oop* _ms; | |
2394 jint* _array_ind_stack; | |
2395 int _ms_size; | |
2396 int _ms_ind; | |
2397 int _array_increment; | |
2398 | |
2399 bool push(oop obj, int arr_ind = 0) { | |
2400 if (_ms_ind == _ms_size) { | |
2401 gclog_or_tty->print_cr("Mark stack is full."); | |
2402 return false; | |
2403 } | |
2404 _ms[_ms_ind] = obj; | |
2405 if (obj->is_objArray()) _array_ind_stack[_ms_ind] = arr_ind; | |
2406 _ms_ind++; | |
2407 return true; | |
2408 } | |
2409 | |
2410 oop pop() { | |
2411 if (_ms_ind == 0) return NULL; | |
2412 else { | |
2413 _ms_ind--; | |
2414 return _ms[_ms_ind]; | |
2415 } | |
2416 } | |
2417 | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2418 template <class T> bool drain() { |
342 | 2419 while (_ms_ind > 0) { |
2420 oop obj = pop(); | |
2421 assert(obj != NULL, "Since index was non-zero."); | |
2422 if (obj->is_objArray()) { | |
2423 jint arr_ind = _array_ind_stack[_ms_ind]; | |
2424 objArrayOop aobj = objArrayOop(obj); | |
2425 jint len = aobj->length(); | |
2426 jint next_arr_ind = arr_ind + _array_increment; | |
2427 if (next_arr_ind < len) { | |
2428 push(obj, next_arr_ind); | |
2429 } | |
2430 // Now process this portion of this one. | |
2431 int lim = MIN2(next_arr_ind, len); | |
2432 for (int j = arr_ind; j < lim; j++) { | |
912
308762b2bf14
6872000: G1: compilation fails on linux/older gcc
apetrusenko
parents:
866
diff
changeset
|
2433 do_oop(aobj->objArrayOopDesc::obj_at_addr<T>(j)); |
342 | 2434 } |
2435 | |
2436 } else { | |
2437 obj->oop_iterate(this); | |
2438 } | |
2439 if (abort()) return false; | |
2440 } | |
2441 return true; | |
2442 } | |
2443 | |
2444 public: | |
2445 CSMarkOopClosure(ConcurrentMark* cm, int ms_size) : | |
2446 _g1h(G1CollectedHeap::heap()), | |
2447 _cm(cm), | |
2448 _bm(cm->nextMarkBitMap()), | |
2449 _ms_size(ms_size), _ms_ind(0), | |
2450 _ms(NEW_C_HEAP_ARRAY(oop, ms_size)), | |
2451 _array_ind_stack(NEW_C_HEAP_ARRAY(jint, ms_size)), | |
2452 _array_increment(MAX2(ms_size/8, 16)) | |
2453 {} | |
2454 | |
2455 ~CSMarkOopClosure() { | |
2456 FREE_C_HEAP_ARRAY(oop, _ms); | |
2457 FREE_C_HEAP_ARRAY(jint, _array_ind_stack); | |
2458 } | |
2459 | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2460 virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2461 virtual void do_oop( oop* p) { do_oop_work(p); } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2462 |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2463 template <class T> void do_oop_work(T* p) { |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2464 T heap_oop = oopDesc::load_heap_oop(p); |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2465 if (oopDesc::is_null(heap_oop)) return; |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2466 oop obj = oopDesc::decode_heap_oop_not_null(heap_oop); |
342 | 2467 if (obj->is_forwarded()) { |
2468 // If the object has already been forwarded, we have to make sure | |
2469 // that it's marked. So follow the forwarding pointer. Note that | |
2470 // this does the right thing for self-forwarding pointers in the | |
2471 // evacuation failure case. | |
2472 obj = obj->forwardee(); | |
2473 } | |
2474 HeapRegion* hr = _g1h->heap_region_containing(obj); | |
2475 if (hr != NULL) { | |
2476 if (hr->in_collection_set()) { | |
2477 if (_g1h->is_obj_ill(obj)) { | |
2478 _bm->mark((HeapWord*)obj); | |
2479 if (!push(obj)) { | |
2480 gclog_or_tty->print_cr("Setting abort in CSMarkOopClosure because push failed."); | |
2481 set_abort(); | |
2482 } | |
2483 } | |
2484 } else { | |
2485 // Outside the collection set; we need to gray it | |
2486 _cm->deal_with_reference(obj); | |
2487 } | |
2488 } | |
2489 } | |
2490 }; | |
2491 | |
2492 class CSMarkBitMapClosure: public BitMapClosure { | |
2493 G1CollectedHeap* _g1h; | |
2494 CMBitMap* _bitMap; | |
2495 ConcurrentMark* _cm; | |
2496 CSMarkOopClosure _oop_cl; | |
2497 public: | |
2498 CSMarkBitMapClosure(ConcurrentMark* cm, int ms_size) : | |
2499 _g1h(G1CollectedHeap::heap()), | |
2500 _bitMap(cm->nextMarkBitMap()), | |
2501 _oop_cl(cm, ms_size) | |
2502 {} | |
2503 | |
2504 ~CSMarkBitMapClosure() {} | |
2505 | |
2506 bool do_bit(size_t offset) { | |
2507 // convert offset into a HeapWord* | |
2508 HeapWord* addr = _bitMap->offsetToHeapWord(offset); | |
2509 assert(_bitMap->endWord() && addr < _bitMap->endWord(), | |
2510 "address out of range"); | |
2511 assert(_bitMap->isMarked(addr), "tautology"); | |
2512 oop obj = oop(addr); | |
2513 if (!obj->is_forwarded()) { | |
2514 if (!_oop_cl.push(obj)) return false; | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2515 if (UseCompressedOops) { |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2516 if (!_oop_cl.drain<narrowOop>()) return false; |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2517 } else { |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2518 if (!_oop_cl.drain<oop>()) return false; |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2519 } |
342 | 2520 } |
2521 // Otherwise... | |
2522 return true; | |
2523 } | |
2524 }; | |
2525 | |
2526 | |
2527 class CompleteMarkingInCSHRClosure: public HeapRegionClosure { | |
2528 CMBitMap* _bm; | |
2529 CSMarkBitMapClosure _bit_cl; | |
2530 enum SomePrivateConstants { | |
2531 MSSize = 1000 | |
2532 }; | |
2533 bool _completed; | |
2534 public: | |
2535 CompleteMarkingInCSHRClosure(ConcurrentMark* cm) : | |
2536 _bm(cm->nextMarkBitMap()), | |
2537 _bit_cl(cm, MSSize), | |
2538 _completed(true) | |
2539 {} | |
2540 | |
2541 ~CompleteMarkingInCSHRClosure() {} | |
2542 | |
2543 bool doHeapRegion(HeapRegion* r) { | |
2544 if (!r->evacuation_failed()) { | |
2545 MemRegion mr = MemRegion(r->bottom(), r->next_top_at_mark_start()); | |
2546 if (!mr.is_empty()) { | |
2547 if (!_bm->iterate(&_bit_cl, mr)) { | |
2548 _completed = false; | |
2549 return true; | |
2550 } | |
2551 } | |
2552 } | |
2553 return false; | |
2554 } | |
2555 | |
2556 bool completed() { return _completed; } | |
2557 }; | |
2558 | |
2559 class ClearMarksInHRClosure: public HeapRegionClosure { | |
2560 CMBitMap* _bm; | |
2561 public: | |
2562 ClearMarksInHRClosure(CMBitMap* bm): _bm(bm) { } | |
2563 | |
2564 bool doHeapRegion(HeapRegion* r) { | |
2565 if (!r->used_region().is_empty() && !r->evacuation_failed()) { | |
2566 MemRegion usedMR = r->used_region(); | |
2567 _bm->clearRange(r->used_region()); | |
2568 } | |
2569 return false; | |
2570 } | |
2571 }; | |
2572 | |
2573 void ConcurrentMark::complete_marking_in_collection_set() { | |
2574 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
2575 | |
2576 if (!g1h->mark_in_progress()) { | |
2577 g1h->g1_policy()->record_mark_closure_time(0.0); | |
2578 return; | |
2579 } | |
2580 | |
2581 int i = 1; | |
2582 double start = os::elapsedTime(); | |
2583 while (true) { | |
2584 i++; | |
2585 CompleteMarkingInCSHRClosure cmplt(this); | |
2586 g1h->collection_set_iterate(&cmplt); | |
2587 if (cmplt.completed()) break; | |
2588 } | |
2589 double end_time = os::elapsedTime(); | |
2590 double elapsed_time_ms = (end_time - start) * 1000.0; | |
2591 g1h->g1_policy()->record_mark_closure_time(elapsed_time_ms); | |
2592 | |
2593 ClearMarksInHRClosure clr(nextMarkBitMap()); | |
2594 g1h->collection_set_iterate(&clr); | |
2595 } | |
2596 | |
2597 // The next two methods deal with the following optimisation. Some | |
2598 // objects are gray by being marked and located above the finger. If | |
2599 // they are copied, during an evacuation pause, below the finger then | |
2600 // the need to be pushed on the stack. The observation is that, if | |
2601 // there are no regions in the collection set located above the | |
2602 // finger, then the above cannot happen, hence we do not need to | |
2603 // explicitly gray any objects when copying them to below the | |
2604 // finger. The global stack will be scanned to ensure that, if it | |
2605 // points to objects being copied, it will update their | |
2606 // location. There is a tricky situation with the gray objects in | |
2607 // region stack that are being coped, however. See the comment in | |
2608 // newCSet(). | |
2609 | |
2610 void ConcurrentMark::newCSet() { | |
2611 if (!concurrent_marking_in_progress()) | |
2612 // nothing to do if marking is not in progress | |
2613 return; | |
2614 | |
2615 // find what the lowest finger is among the global and local fingers | |
2616 _min_finger = _finger; | |
2617 for (int i = 0; i < (int)_max_task_num; ++i) { | |
2618 CMTask* task = _tasks[i]; | |
2619 HeapWord* task_finger = task->finger(); | |
2620 if (task_finger != NULL && task_finger < _min_finger) | |
2621 _min_finger = task_finger; | |
2622 } | |
2623 | |
2624 _should_gray_objects = false; | |
2625 | |
2626 // This fixes a very subtle and fustrating bug. It might be the case | |
2627 // that, during en evacuation pause, heap regions that contain | |
2628 // objects that are gray (by being in regions contained in the | |
2629 // region stack) are included in the collection set. Since such gray | |
2630 // objects will be moved, and because it's not easy to redirect | |
2631 // region stack entries to point to a new location (because objects | |
2632 // in one region might be scattered to multiple regions after they | |
2633 // are copied), one option is to ensure that all marked objects | |
2634 // copied during a pause are pushed on the stack. Notice, however, | |
2635 // that this problem can only happen when the region stack is not | |
2636 // empty during an evacuation pause. So, we make the fix a bit less | |
2637 // conservative and ensure that regions are pushed on the stack, | |
2638 // irrespective whether all collection set regions are below the | |
2639 // finger, if the region stack is not empty. This is expected to be | |
2640 // a rare case, so I don't think it's necessary to be smarted about it. | |
2641 if (!region_stack_empty()) | |
2642 _should_gray_objects = true; | |
2643 } | |
2644 | |
2645 void ConcurrentMark::registerCSetRegion(HeapRegion* hr) { | |
2646 if (!concurrent_marking_in_progress()) | |
2647 return; | |
2648 | |
2649 HeapWord* region_end = hr->end(); | |
2650 if (region_end > _min_finger) | |
2651 _should_gray_objects = true; | |
2652 } | |
2653 | |
2654 // abandon current marking iteration due to a Full GC | |
2655 void ConcurrentMark::abort() { | |
2656 // Clear all marks to force marking thread to do nothing | |
2657 _nextMarkBitMap->clearAll(); | |
2658 // Empty mark stack | |
2659 clear_marking_state(); | |
2660 for (int i = 0; i < (int)_max_task_num; ++i) | |
2661 _tasks[i]->clear_region_fields(); | |
2662 _has_aborted = true; | |
2663 | |
2664 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
2665 satb_mq_set.abandon_partial_marking(); | |
1317
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
2666 // This can be called either during or outside marking, we'll read |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
2667 // the expected_active value from the SATB queue set. |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
2668 satb_mq_set.set_active_all_threads( |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
2669 false, /* new active value */ |
d4197f8d516a
6935821: G1: threads created during marking do not active their SATB queues
tonyp
parents:
1314
diff
changeset
|
2670 satb_mq_set.is_active() /* expected_active */); |
342 | 2671 } |
2672 | |
2673 static void print_ms_time_info(const char* prefix, const char* name, | |
2674 NumberSeq& ns) { | |
2675 gclog_or_tty->print_cr("%s%5d %12s: total time = %8.2f s (avg = %8.2f ms).", | |
2676 prefix, ns.num(), name, ns.sum()/1000.0, ns.avg()); | |
2677 if (ns.num() > 0) { | |
2678 gclog_or_tty->print_cr("%s [std. dev = %8.2f ms, max = %8.2f ms]", | |
2679 prefix, ns.sd(), ns.maximum()); | |
2680 } | |
2681 } | |
2682 | |
2683 void ConcurrentMark::print_summary_info() { | |
2684 gclog_or_tty->print_cr(" Concurrent marking:"); | |
2685 print_ms_time_info(" ", "init marks", _init_times); | |
2686 print_ms_time_info(" ", "remarks", _remark_times); | |
2687 { | |
2688 print_ms_time_info(" ", "final marks", _remark_mark_times); | |
2689 print_ms_time_info(" ", "weak refs", _remark_weak_ref_times); | |
2690 | |
2691 } | |
2692 print_ms_time_info(" ", "cleanups", _cleanup_times); | |
2693 gclog_or_tty->print_cr(" Final counting total time = %8.2f s (avg = %8.2f ms).", | |
2694 _total_counting_time, | |
2695 (_cleanup_times.num() > 0 ? _total_counting_time * 1000.0 / | |
2696 (double)_cleanup_times.num() | |
2697 : 0.0)); | |
2698 if (G1ScrubRemSets) { | |
2699 gclog_or_tty->print_cr(" RS scrub total time = %8.2f s (avg = %8.2f ms).", | |
2700 _total_rs_scrub_time, | |
2701 (_cleanup_times.num() > 0 ? _total_rs_scrub_time * 1000.0 / | |
2702 (double)_cleanup_times.num() | |
2703 : 0.0)); | |
2704 } | |
2705 gclog_or_tty->print_cr(" Total stop_world time = %8.2f s.", | |
2706 (_init_times.sum() + _remark_times.sum() + | |
2707 _cleanup_times.sum())/1000.0); | |
2708 gclog_or_tty->print_cr(" Total concurrent time = %8.2f s " | |
2709 "(%8.2f s marking, %8.2f s counting).", | |
2710 cmThread()->vtime_accum(), | |
2711 cmThread()->vtime_mark_accum(), | |
2712 cmThread()->vtime_count_accum()); | |
2713 } | |
2714 | |
1019 | 2715 void ConcurrentMark::print_worker_threads_on(outputStream* st) const { |
2716 _parallel_workers->print_worker_threads_on(st); | |
2717 } | |
2718 | |
342 | 2719 // Closures |
2720 // XXX: there seems to be a lot of code duplication here; | |
2721 // should refactor and consolidate the shared code. | |
2722 | |
2723 // This closure is used to mark refs into the CMS generation in | |
2724 // the CMS bit map. Called at the first checkpoint. | |
2725 | |
2726 // We take a break if someone is trying to stop the world. | |
2727 bool ConcurrentMark::do_yield_check(int worker_i) { | |
2728 if (should_yield()) { | |
2729 if (worker_i == 0) | |
2730 _g1h->g1_policy()->record_concurrent_pause(); | |
2731 cmThread()->yield(); | |
2732 if (worker_i == 0) | |
2733 _g1h->g1_policy()->record_concurrent_pause_end(); | |
2734 return true; | |
2735 } else { | |
2736 return false; | |
2737 } | |
2738 } | |
2739 | |
2740 bool ConcurrentMark::should_yield() { | |
2741 return cmThread()->should_yield(); | |
2742 } | |
2743 | |
2744 bool ConcurrentMark::containing_card_is_marked(void* p) { | |
2745 size_t offset = pointer_delta(p, _g1h->reserved_region().start(), 1); | |
2746 return _card_bm.at(offset >> CardTableModRefBS::card_shift); | |
2747 } | |
2748 | |
2749 bool ConcurrentMark::containing_cards_are_marked(void* start, | |
2750 void* last) { | |
2751 return | |
2752 containing_card_is_marked(start) && | |
2753 containing_card_is_marked(last); | |
2754 } | |
2755 | |
2756 #ifndef PRODUCT | |
2757 // for debugging purposes | |
2758 void ConcurrentMark::print_finger() { | |
2759 gclog_or_tty->print_cr("heap ["PTR_FORMAT", "PTR_FORMAT"), global finger = "PTR_FORMAT, | |
2760 _heap_start, _heap_end, _finger); | |
2761 for (int i = 0; i < (int) _max_task_num; ++i) { | |
2762 gclog_or_tty->print(" %d: "PTR_FORMAT, i, _tasks[i]->finger()); | |
2763 } | |
2764 gclog_or_tty->print_cr(""); | |
2765 } | |
2766 #endif | |
2767 | |
2768 // Closure for iteration over bitmaps | |
2769 class CMBitMapClosure : public BitMapClosure { | |
2770 private: | |
2771 // the bitmap that is being iterated over | |
2772 CMBitMap* _nextMarkBitMap; | |
2773 ConcurrentMark* _cm; | |
2774 CMTask* _task; | |
2775 // true if we're scanning a heap region claimed by the task (so that | |
2776 // we move the finger along), false if we're not, i.e. currently when | |
2777 // scanning a heap region popped from the region stack (so that we | |
2778 // do not move the task finger along; it'd be a mistake if we did so). | |
2779 bool _scanning_heap_region; | |
2780 | |
2781 public: | |
2782 CMBitMapClosure(CMTask *task, | |
2783 ConcurrentMark* cm, | |
2784 CMBitMap* nextMarkBitMap) | |
2785 : _task(task), _cm(cm), _nextMarkBitMap(nextMarkBitMap) { } | |
2786 | |
2787 void set_scanning_heap_region(bool scanning_heap_region) { | |
2788 _scanning_heap_region = scanning_heap_region; | |
2789 } | |
2790 | |
2791 bool do_bit(size_t offset) { | |
2792 HeapWord* addr = _nextMarkBitMap->offsetToHeapWord(offset); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2793 assert(_nextMarkBitMap->isMarked(addr), "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2794 assert( addr < _cm->finger(), "invariant"); |
342 | 2795 |
2796 if (_scanning_heap_region) { | |
2797 statsOnly( _task->increase_objs_found_on_bitmap() ); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2798 assert(addr >= _task->finger(), "invariant"); |
342 | 2799 // We move that task's local finger along. |
2800 _task->move_finger_to(addr); | |
2801 } else { | |
2802 // We move the task's region finger along. | |
2803 _task->move_region_finger_to(addr); | |
2804 } | |
2805 | |
2806 _task->scan_object(oop(addr)); | |
2807 // we only partially drain the local queue and global stack | |
2808 _task->drain_local_queue(true); | |
2809 _task->drain_global_stack(true); | |
2810 | |
2811 // if the has_aborted flag has been raised, we need to bail out of | |
2812 // the iteration | |
2813 return !_task->has_aborted(); | |
2814 } | |
2815 }; | |
2816 | |
2817 // Closure for iterating over objects, currently only used for | |
2818 // processing SATB buffers. | |
2819 class CMObjectClosure : public ObjectClosure { | |
2820 private: | |
2821 CMTask* _task; | |
2822 | |
2823 public: | |
2824 void do_object(oop obj) { | |
2825 _task->deal_with_reference(obj); | |
2826 } | |
2827 | |
2828 CMObjectClosure(CMTask* task) : _task(task) { } | |
2829 }; | |
2830 | |
2831 // Closure for iterating over object fields | |
2832 class CMOopClosure : public OopClosure { | |
2833 private: | |
2834 G1CollectedHeap* _g1h; | |
2835 ConcurrentMark* _cm; | |
2836 CMTask* _task; | |
2837 | |
2838 public: | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2839 virtual void do_oop(narrowOop* p) { do_oop_work(p); } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2840 virtual void do_oop( oop* p) { do_oop_work(p); } |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2841 |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2842 template <class T> void do_oop_work(T* p) { |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2843 assert(_g1h->is_in_g1_reserved((HeapWord*) p), "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2844 assert(!_g1h->heap_region_containing((HeapWord*) p)->is_on_free_list(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2845 "invariant"); |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2846 |
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2847 oop obj = oopDesc::load_decode_heap_oop(p); |
342 | 2848 if (_cm->verbose_high()) |
2849 gclog_or_tty->print_cr("[%d] we're looking at location " | |
2850 "*"PTR_FORMAT" = "PTR_FORMAT, | |
2851 _task->task_id(), p, (void*) obj); | |
2852 _task->deal_with_reference(obj); | |
2853 } | |
2854 | |
2855 CMOopClosure(G1CollectedHeap* g1h, | |
2856 ConcurrentMark* cm, | |
2857 CMTask* task) | |
2858 : _g1h(g1h), _cm(cm), _task(task) { } | |
2859 }; | |
2860 | |
2861 void CMTask::setup_for_region(HeapRegion* hr) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2862 // Separated the asserts so that we know which one fires. |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2863 assert(hr != NULL, |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2864 "claim_region() should have filtered out continues humongous regions"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2865 assert(!hr->continuesHumongous(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2866 "claim_region() should have filtered out continues humongous regions"); |
342 | 2867 |
2868 if (_cm->verbose_low()) | |
2869 gclog_or_tty->print_cr("[%d] setting up for region "PTR_FORMAT, | |
2870 _task_id, hr); | |
2871 | |
2872 _curr_region = hr; | |
2873 _finger = hr->bottom(); | |
2874 update_region_limit(); | |
2875 } | |
2876 | |
2877 void CMTask::update_region_limit() { | |
2878 HeapRegion* hr = _curr_region; | |
2879 HeapWord* bottom = hr->bottom(); | |
2880 HeapWord* limit = hr->next_top_at_mark_start(); | |
2881 | |
2882 if (limit == bottom) { | |
2883 if (_cm->verbose_low()) | |
2884 gclog_or_tty->print_cr("[%d] found an empty region " | |
2885 "["PTR_FORMAT", "PTR_FORMAT")", | |
2886 _task_id, bottom, limit); | |
2887 // The region was collected underneath our feet. | |
2888 // We set the finger to bottom to ensure that the bitmap | |
2889 // iteration that will follow this will not do anything. | |
2890 // (this is not a condition that holds when we set the region up, | |
2891 // as the region is not supposed to be empty in the first place) | |
2892 _finger = bottom; | |
2893 } else if (limit >= _region_limit) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2894 assert(limit >= _finger, "peace of mind"); |
342 | 2895 } else { |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2896 assert(limit < _region_limit, "only way to get here"); |
342 | 2897 // This can happen under some pretty unusual circumstances. An |
2898 // evacuation pause empties the region underneath our feet (NTAMS | |
2899 // at bottom). We then do some allocation in the region (NTAMS | |
2900 // stays at bottom), followed by the region being used as a GC | |
2901 // alloc region (NTAMS will move to top() and the objects | |
2902 // originally below it will be grayed). All objects now marked in | |
2903 // the region are explicitly grayed, if below the global finger, | |
2904 // and we do not need in fact to scan anything else. So, we simply | |
2905 // set _finger to be limit to ensure that the bitmap iteration | |
2906 // doesn't do anything. | |
2907 _finger = limit; | |
2908 } | |
2909 | |
2910 _region_limit = limit; | |
2911 } | |
2912 | |
2913 void CMTask::giveup_current_region() { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2914 assert(_curr_region != NULL, "invariant"); |
342 | 2915 if (_cm->verbose_low()) |
2916 gclog_or_tty->print_cr("[%d] giving up region "PTR_FORMAT, | |
2917 _task_id, _curr_region); | |
2918 clear_region_fields(); | |
2919 } | |
2920 | |
2921 void CMTask::clear_region_fields() { | |
2922 // Values for these three fields that indicate that we're not | |
2923 // holding on to a region. | |
2924 _curr_region = NULL; | |
2925 _finger = NULL; | |
2926 _region_limit = NULL; | |
2927 | |
2928 _region_finger = NULL; | |
2929 } | |
2930 | |
2931 void CMTask::reset(CMBitMap* nextMarkBitMap) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2932 guarantee(nextMarkBitMap != NULL, "invariant"); |
342 | 2933 |
2934 if (_cm->verbose_low()) | |
2935 gclog_or_tty->print_cr("[%d] resetting", _task_id); | |
2936 | |
2937 _nextMarkBitMap = nextMarkBitMap; | |
2938 clear_region_fields(); | |
2939 | |
2940 _calls = 0; | |
2941 _elapsed_time_ms = 0.0; | |
2942 _termination_time_ms = 0.0; | |
2943 _termination_start_time_ms = 0.0; | |
2944 | |
2945 #if _MARKING_STATS_ | |
2946 _local_pushes = 0; | |
2947 _local_pops = 0; | |
2948 _local_max_size = 0; | |
2949 _objs_scanned = 0; | |
2950 _global_pushes = 0; | |
2951 _global_pops = 0; | |
2952 _global_max_size = 0; | |
2953 _global_transfers_to = 0; | |
2954 _global_transfers_from = 0; | |
2955 _region_stack_pops = 0; | |
2956 _regions_claimed = 0; | |
2957 _objs_found_on_bitmap = 0; | |
2958 _satb_buffers_processed = 0; | |
2959 _steal_attempts = 0; | |
2960 _steals = 0; | |
2961 _aborted = 0; | |
2962 _aborted_overflow = 0; | |
2963 _aborted_cm_aborted = 0; | |
2964 _aborted_yield = 0; | |
2965 _aborted_timed_out = 0; | |
2966 _aborted_satb = 0; | |
2967 _aborted_termination = 0; | |
2968 #endif // _MARKING_STATS_ | |
2969 } | |
2970 | |
2971 bool CMTask::should_exit_termination() { | |
2972 regular_clock_call(); | |
2973 // This is called when we are in the termination protocol. We should | |
2974 // quit if, for some reason, this task wants to abort or the global | |
2975 // stack is not empty (this means that we can get work from it). | |
2976 return !_cm->mark_stack_empty() || has_aborted(); | |
2977 } | |
2978 | |
2979 // This determines whether the method below will check both the local | |
2980 // and global fingers when determining whether to push on the stack a | |
2981 // gray object (value 1) or whether it will only check the global one | |
2982 // (value 0). The tradeoffs are that the former will be a bit more | |
2983 // accurate and possibly push less on the stack, but it might also be | |
2984 // a little bit slower. | |
2985 | |
2986 #define _CHECK_BOTH_FINGERS_ 1 | |
2987 | |
2988 void CMTask::deal_with_reference(oop obj) { | |
2989 if (_cm->verbose_high()) | |
2990 gclog_or_tty->print_cr("[%d] we're dealing with reference = "PTR_FORMAT, | |
2991 _task_id, (void*) obj); | |
2992 | |
2993 ++_refs_reached; | |
2994 | |
2995 HeapWord* objAddr = (HeapWord*) obj; | |
845
df6caf649ff7
6700789: G1: Enable use of compressed oops with G1 heaps
ysr
parents:
831
diff
changeset
|
2996 assert(obj->is_oop_or_null(true /* ignore mark word */), "Error"); |
342 | 2997 if (_g1h->is_in_g1_reserved(objAddr)) { |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
2998 assert(obj != NULL, "is_in_g1_reserved should ensure this"); |
342 | 2999 HeapRegion* hr = _g1h->heap_region_containing(obj); |
3000 if (_g1h->is_obj_ill(obj, hr)) { | |
3001 if (_cm->verbose_high()) | |
3002 gclog_or_tty->print_cr("[%d] "PTR_FORMAT" is not considered marked", | |
3003 _task_id, (void*) obj); | |
3004 | |
3005 // we need to mark it first | |
3006 if (_nextMarkBitMap->parMark(objAddr)) { | |
3007 // No OrderAccess:store_load() is needed. It is implicit in the | |
3008 // CAS done in parMark(objAddr) above | |
3009 HeapWord* global_finger = _cm->finger(); | |
3010 | |
3011 #if _CHECK_BOTH_FINGERS_ | |
3012 // we will check both the local and global fingers | |
3013 | |
3014 if (_finger != NULL && objAddr < _finger) { | |
3015 if (_cm->verbose_high()) | |
3016 gclog_or_tty->print_cr("[%d] below the local finger ("PTR_FORMAT"), " | |
3017 "pushing it", _task_id, _finger); | |
3018 push(obj); | |
3019 } else if (_curr_region != NULL && objAddr < _region_limit) { | |
3020 // do nothing | |
3021 } else if (objAddr < global_finger) { | |
3022 // Notice that the global finger might be moving forward | |
3023 // concurrently. This is not a problem. In the worst case, we | |
3024 // mark the object while it is above the global finger and, by | |
3025 // the time we read the global finger, it has moved forward | |
3026 // passed this object. In this case, the object will probably | |
3027 // be visited when a task is scanning the region and will also | |
3028 // be pushed on the stack. So, some duplicate work, but no | |
3029 // correctness problems. | |
3030 | |
3031 if (_cm->verbose_high()) | |
3032 gclog_or_tty->print_cr("[%d] below the global finger " | |
3033 "("PTR_FORMAT"), pushing it", | |
3034 _task_id, global_finger); | |
3035 push(obj); | |
3036 } else { | |
3037 // do nothing | |
3038 } | |
3039 #else // _CHECK_BOTH_FINGERS_ | |
3040 // we will only check the global finger | |
3041 | |
3042 if (objAddr < global_finger) { | |
3043 // see long comment above | |
3044 | |
3045 if (_cm->verbose_high()) | |
3046 gclog_or_tty->print_cr("[%d] below the global finger " | |
3047 "("PTR_FORMAT"), pushing it", | |
3048 _task_id, global_finger); | |
3049 push(obj); | |
3050 } | |
3051 #endif // _CHECK_BOTH_FINGERS_ | |
3052 } | |
3053 } | |
3054 } | |
3055 } | |
3056 | |
3057 void CMTask::push(oop obj) { | |
3058 HeapWord* objAddr = (HeapWord*) obj; | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3059 assert(_g1h->is_in_g1_reserved(objAddr), "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3060 assert(!_g1h->heap_region_containing(objAddr)->is_on_free_list(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3061 "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3062 assert(!_g1h->is_obj_ill(obj), "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3063 assert(_nextMarkBitMap->isMarked(objAddr), "invariant"); |
342 | 3064 |
3065 if (_cm->verbose_high()) | |
3066 gclog_or_tty->print_cr("[%d] pushing "PTR_FORMAT, _task_id, (void*) obj); | |
3067 | |
3068 if (!_task_queue->push(obj)) { | |
3069 // The local task queue looks full. We need to push some entries | |
3070 // to the global stack. | |
3071 | |
3072 if (_cm->verbose_medium()) | |
3073 gclog_or_tty->print_cr("[%d] task queue overflow, " | |
3074 "moving entries to the global stack", | |
3075 _task_id); | |
3076 move_entries_to_global_stack(); | |
3077 | |
3078 // this should succeed since, even if we overflow the global | |
3079 // stack, we should have definitely removed some entries from the | |
3080 // local queue. So, there must be space on it. | |
3081 bool success = _task_queue->push(obj); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3082 assert(success, "invariant"); |
342 | 3083 } |
3084 | |
3085 statsOnly( int tmp_size = _task_queue->size(); | |
3086 if (tmp_size > _local_max_size) | |
3087 _local_max_size = tmp_size; | |
3088 ++_local_pushes ); | |
3089 } | |
3090 | |
3091 void CMTask::reached_limit() { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3092 assert(_words_scanned >= _words_scanned_limit || |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3093 _refs_reached >= _refs_reached_limit , |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3094 "shouldn't have been called otherwise"); |
342 | 3095 regular_clock_call(); |
3096 } | |
3097 | |
3098 void CMTask::regular_clock_call() { | |
3099 if (has_aborted()) | |
3100 return; | |
3101 | |
3102 // First, we need to recalculate the words scanned and refs reached | |
3103 // limits for the next clock call. | |
3104 recalculate_limits(); | |
3105 | |
3106 // During the regular clock call we do the following | |
3107 | |
3108 // (1) If an overflow has been flagged, then we abort. | |
3109 if (_cm->has_overflown()) { | |
3110 set_has_aborted(); | |
3111 return; | |
3112 } | |
3113 | |
3114 // If we are not concurrent (i.e. we're doing remark) we don't need | |
3115 // to check anything else. The other steps are only needed during | |
3116 // the concurrent marking phase. | |
3117 if (!concurrent()) | |
3118 return; | |
3119 | |
3120 // (2) If marking has been aborted for Full GC, then we also abort. | |
3121 if (_cm->has_aborted()) { | |
3122 set_has_aborted(); | |
3123 statsOnly( ++_aborted_cm_aborted ); | |
3124 return; | |
3125 } | |
3126 | |
3127 double curr_time_ms = os::elapsedVTime() * 1000.0; | |
3128 | |
3129 // (3) If marking stats are enabled, then we update the step history. | |
3130 #if _MARKING_STATS_ | |
3131 if (_words_scanned >= _words_scanned_limit) | |
3132 ++_clock_due_to_scanning; | |
3133 if (_refs_reached >= _refs_reached_limit) | |
3134 ++_clock_due_to_marking; | |
3135 | |
3136 double last_interval_ms = curr_time_ms - _interval_start_time_ms; | |
3137 _interval_start_time_ms = curr_time_ms; | |
3138 _all_clock_intervals_ms.add(last_interval_ms); | |
3139 | |
3140 if (_cm->verbose_medium()) { | |
3141 gclog_or_tty->print_cr("[%d] regular clock, interval = %1.2lfms, " | |
3142 "scanned = %d%s, refs reached = %d%s", | |
3143 _task_id, last_interval_ms, | |
3144 _words_scanned, | |
3145 (_words_scanned >= _words_scanned_limit) ? " (*)" : "", | |
3146 _refs_reached, | |
3147 (_refs_reached >= _refs_reached_limit) ? " (*)" : ""); | |
3148 } | |
3149 #endif // _MARKING_STATS_ | |
3150 | |
3151 // (4) We check whether we should yield. If we have to, then we abort. | |
3152 if (_cm->should_yield()) { | |
3153 // We should yield. To do this we abort the task. The caller is | |
3154 // responsible for yielding. | |
3155 set_has_aborted(); | |
3156 statsOnly( ++_aborted_yield ); | |
3157 return; | |
3158 } | |
3159 | |
3160 // (5) We check whether we've reached our time quota. If we have, | |
3161 // then we abort. | |
3162 double elapsed_time_ms = curr_time_ms - _start_time_ms; | |
3163 if (elapsed_time_ms > _time_target_ms) { | |
3164 set_has_aborted(); | |
3165 _has_aborted_timed_out = true; | |
3166 statsOnly( ++_aborted_timed_out ); | |
3167 return; | |
3168 } | |
3169 | |
3170 // (6) Finally, we check whether there are enough completed STAB | |
3171 // buffers available for processing. If there are, we abort. | |
3172 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
3173 if (!_draining_satb_buffers && satb_mq_set.process_completed_buffers()) { | |
3174 if (_cm->verbose_low()) | |
3175 gclog_or_tty->print_cr("[%d] aborting to deal with pending SATB buffers", | |
3176 _task_id); | |
3177 // we do need to process SATB buffers, we'll abort and restart | |
3178 // the marking task to do so | |
3179 set_has_aborted(); | |
3180 statsOnly( ++_aborted_satb ); | |
3181 return; | |
3182 } | |
3183 } | |
3184 | |
3185 void CMTask::recalculate_limits() { | |
3186 _real_words_scanned_limit = _words_scanned + words_scanned_period; | |
3187 _words_scanned_limit = _real_words_scanned_limit; | |
3188 | |
3189 _real_refs_reached_limit = _refs_reached + refs_reached_period; | |
3190 _refs_reached_limit = _real_refs_reached_limit; | |
3191 } | |
3192 | |
3193 void CMTask::decrease_limits() { | |
3194 // This is called when we believe that we're going to do an infrequent | |
3195 // operation which will increase the per byte scanned cost (i.e. move | |
3196 // entries to/from the global stack). It basically tries to decrease the | |
3197 // scanning limit so that the clock is called earlier. | |
3198 | |
3199 if (_cm->verbose_medium()) | |
3200 gclog_or_tty->print_cr("[%d] decreasing limits", _task_id); | |
3201 | |
3202 _words_scanned_limit = _real_words_scanned_limit - | |
3203 3 * words_scanned_period / 4; | |
3204 _refs_reached_limit = _real_refs_reached_limit - | |
3205 3 * refs_reached_period / 4; | |
3206 } | |
3207 | |
3208 void CMTask::move_entries_to_global_stack() { | |
3209 // local array where we'll store the entries that will be popped | |
3210 // from the local queue | |
3211 oop buffer[global_stack_transfer_size]; | |
3212 | |
3213 int n = 0; | |
3214 oop obj; | |
3215 while (n < global_stack_transfer_size && _task_queue->pop_local(obj)) { | |
3216 buffer[n] = obj; | |
3217 ++n; | |
3218 } | |
3219 | |
3220 if (n > 0) { | |
3221 // we popped at least one entry from the local queue | |
3222 | |
3223 statsOnly( ++_global_transfers_to; _local_pops += n ); | |
3224 | |
3225 if (!_cm->mark_stack_push(buffer, n)) { | |
3226 if (_cm->verbose_low()) | |
3227 gclog_or_tty->print_cr("[%d] aborting due to global stack overflow", _task_id); | |
3228 set_has_aborted(); | |
3229 } else { | |
3230 // the transfer was successful | |
3231 | |
3232 if (_cm->verbose_medium()) | |
3233 gclog_or_tty->print_cr("[%d] pushed %d entries to the global stack", | |
3234 _task_id, n); | |
3235 statsOnly( int tmp_size = _cm->mark_stack_size(); | |
3236 if (tmp_size > _global_max_size) | |
3237 _global_max_size = tmp_size; | |
3238 _global_pushes += n ); | |
3239 } | |
3240 } | |
3241 | |
3242 // this operation was quite expensive, so decrease the limits | |
3243 decrease_limits(); | |
3244 } | |
3245 | |
3246 void CMTask::get_entries_from_global_stack() { | |
3247 // local array where we'll store the entries that will be popped | |
3248 // from the global stack. | |
3249 oop buffer[global_stack_transfer_size]; | |
3250 int n; | |
3251 _cm->mark_stack_pop(buffer, global_stack_transfer_size, &n); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3252 assert(n <= global_stack_transfer_size, |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3253 "we should not pop more than the given limit"); |
342 | 3254 if (n > 0) { |
3255 // yes, we did actually pop at least one entry | |
3256 | |
3257 statsOnly( ++_global_transfers_from; _global_pops += n ); | |
3258 if (_cm->verbose_medium()) | |
3259 gclog_or_tty->print_cr("[%d] popped %d entries from the global stack", | |
3260 _task_id, n); | |
3261 for (int i = 0; i < n; ++i) { | |
3262 bool success = _task_queue->push(buffer[i]); | |
3263 // We only call this when the local queue is empty or under a | |
3264 // given target limit. So, we do not expect this push to fail. | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3265 assert(success, "invariant"); |
342 | 3266 } |
3267 | |
3268 statsOnly( int tmp_size = _task_queue->size(); | |
3269 if (tmp_size > _local_max_size) | |
3270 _local_max_size = tmp_size; | |
3271 _local_pushes += n ); | |
3272 } | |
3273 | |
3274 // this operation was quite expensive, so decrease the limits | |
3275 decrease_limits(); | |
3276 } | |
3277 | |
3278 void CMTask::drain_local_queue(bool partially) { | |
3279 if (has_aborted()) | |
3280 return; | |
3281 | |
3282 // Decide what the target size is, depending whether we're going to | |
3283 // drain it partially (so that other tasks can steal if they run out | |
3284 // of things to do) or totally (at the very end). | |
3285 size_t target_size; | |
3286 if (partially) | |
3287 target_size = MIN2((size_t)_task_queue->max_elems()/3, GCDrainStackTargetSize); | |
3288 else | |
3289 target_size = 0; | |
3290 | |
3291 if (_task_queue->size() > target_size) { | |
3292 if (_cm->verbose_high()) | |
3293 gclog_or_tty->print_cr("[%d] draining local queue, target size = %d", | |
3294 _task_id, target_size); | |
3295 | |
3296 oop obj; | |
3297 bool ret = _task_queue->pop_local(obj); | |
3298 while (ret) { | |
3299 statsOnly( ++_local_pops ); | |
3300 | |
3301 if (_cm->verbose_high()) | |
3302 gclog_or_tty->print_cr("[%d] popped "PTR_FORMAT, _task_id, | |
3303 (void*) obj); | |
3304 | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3305 assert(_g1h->is_in_g1_reserved((HeapWord*) obj), "invariant" ); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3306 assert(!_g1h->heap_region_containing(obj)->is_on_free_list(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3307 "invariant"); |
342 | 3308 |
3309 scan_object(obj); | |
3310 | |
3311 if (_task_queue->size() <= target_size || has_aborted()) | |
3312 ret = false; | |
3313 else | |
3314 ret = _task_queue->pop_local(obj); | |
3315 } | |
3316 | |
3317 if (_cm->verbose_high()) | |
3318 gclog_or_tty->print_cr("[%d] drained local queue, size = %d", | |
3319 _task_id, _task_queue->size()); | |
3320 } | |
3321 } | |
3322 | |
3323 void CMTask::drain_global_stack(bool partially) { | |
3324 if (has_aborted()) | |
3325 return; | |
3326 | |
3327 // We have a policy to drain the local queue before we attempt to | |
3328 // drain the global stack. | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3329 assert(partially || _task_queue->size() == 0, "invariant"); |
342 | 3330 |
3331 // Decide what the target size is, depending whether we're going to | |
3332 // drain it partially (so that other tasks can steal if they run out | |
3333 // of things to do) or totally (at the very end). Notice that, | |
3334 // because we move entries from the global stack in chunks or | |
3335 // because another task might be doing the same, we might in fact | |
3336 // drop below the target. But, this is not a problem. | |
3337 size_t target_size; | |
3338 if (partially) | |
3339 target_size = _cm->partial_mark_stack_size_target(); | |
3340 else | |
3341 target_size = 0; | |
3342 | |
3343 if (_cm->mark_stack_size() > target_size) { | |
3344 if (_cm->verbose_low()) | |
3345 gclog_or_tty->print_cr("[%d] draining global_stack, target size %d", | |
3346 _task_id, target_size); | |
3347 | |
3348 while (!has_aborted() && _cm->mark_stack_size() > target_size) { | |
3349 get_entries_from_global_stack(); | |
3350 drain_local_queue(partially); | |
3351 } | |
3352 | |
3353 if (_cm->verbose_low()) | |
3354 gclog_or_tty->print_cr("[%d] drained global stack, size = %d", | |
3355 _task_id, _cm->mark_stack_size()); | |
3356 } | |
3357 } | |
3358 | |
3359 // SATB Queue has several assumptions on whether to call the par or | |
3360 // non-par versions of the methods. this is why some of the code is | |
3361 // replicated. We should really get rid of the single-threaded version | |
3362 // of the code to simplify things. | |
3363 void CMTask::drain_satb_buffers() { | |
3364 if (has_aborted()) | |
3365 return; | |
3366 | |
3367 // We set this so that the regular clock knows that we're in the | |
3368 // middle of draining buffers and doesn't set the abort flag when it | |
3369 // notices that SATB buffers are available for draining. It'd be | |
3370 // very counter productive if it did that. :-) | |
3371 _draining_satb_buffers = true; | |
3372 | |
3373 CMObjectClosure oc(this); | |
3374 SATBMarkQueueSet& satb_mq_set = JavaThread::satb_mark_queue_set(); | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
3375 if (G1CollectedHeap::use_parallel_gc_threads()) |
342 | 3376 satb_mq_set.set_par_closure(_task_id, &oc); |
3377 else | |
3378 satb_mq_set.set_closure(&oc); | |
3379 | |
3380 // This keeps claiming and applying the closure to completed buffers | |
3381 // until we run out of buffers or we need to abort. | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
3382 if (G1CollectedHeap::use_parallel_gc_threads()) { |
342 | 3383 while (!has_aborted() && |
3384 satb_mq_set.par_apply_closure_to_completed_buffer(_task_id)) { | |
3385 if (_cm->verbose_medium()) | |
3386 gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id); | |
3387 statsOnly( ++_satb_buffers_processed ); | |
3388 regular_clock_call(); | |
3389 } | |
3390 } else { | |
3391 while (!has_aborted() && | |
3392 satb_mq_set.apply_closure_to_completed_buffer()) { | |
3393 if (_cm->verbose_medium()) | |
3394 gclog_or_tty->print_cr("[%d] processed an SATB buffer", _task_id); | |
3395 statsOnly( ++_satb_buffers_processed ); | |
3396 regular_clock_call(); | |
3397 } | |
3398 } | |
3399 | |
3400 if (!concurrent() && !has_aborted()) { | |
3401 // We should only do this during remark. | |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
3402 if (G1CollectedHeap::use_parallel_gc_threads()) |
342 | 3403 satb_mq_set.par_iterate_closure_all_threads(_task_id); |
3404 else | |
3405 satb_mq_set.iterate_closure_all_threads(); | |
3406 } | |
3407 | |
3408 _draining_satb_buffers = false; | |
3409 | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3410 assert(has_aborted() || |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3411 concurrent() || |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3412 satb_mq_set.completed_buffers_num() == 0, "invariant"); |
342 | 3413 |
1833
8b10f48633dc
6984287: Regularize how GC parallel workers are specified.
jmasa
parents:
1719
diff
changeset
|
3414 if (G1CollectedHeap::use_parallel_gc_threads()) |
342 | 3415 satb_mq_set.set_par_closure(_task_id, NULL); |
3416 else | |
3417 satb_mq_set.set_closure(NULL); | |
3418 | |
3419 // again, this was a potentially expensive operation, decrease the | |
3420 // limits to get the regular clock call early | |
3421 decrease_limits(); | |
3422 } | |
3423 | |
3424 void CMTask::drain_region_stack(BitMapClosure* bc) { | |
3425 if (has_aborted()) | |
3426 return; | |
3427 | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3428 assert(_region_finger == NULL, |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3429 "it should be NULL when we're not scanning a region"); |
342 | 3430 |
3431 if (!_cm->region_stack_empty()) { | |
3432 if (_cm->verbose_low()) | |
3433 gclog_or_tty->print_cr("[%d] draining region stack, size = %d", | |
3434 _task_id, _cm->region_stack_size()); | |
3435 | |
1358
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
3436 MemRegion mr = _cm->region_stack_pop_with_lock(); |
342 | 3437 // it returns MemRegion() if the pop fails |
3438 statsOnly(if (mr.start() != NULL) ++_region_stack_pops ); | |
3439 | |
3440 while (mr.start() != NULL) { | |
3441 if (_cm->verbose_medium()) | |
3442 gclog_or_tty->print_cr("[%d] we are scanning region " | |
3443 "["PTR_FORMAT", "PTR_FORMAT")", | |
3444 _task_id, mr.start(), mr.end()); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3445 assert(mr.end() <= _cm->finger(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3446 "otherwise the region shouldn't be on the stack"); |
342 | 3447 assert(!mr.is_empty(), "Only non-empty regions live on the region stack"); |
3448 if (_nextMarkBitMap->iterate(bc, mr)) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3449 assert(!has_aborted(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3450 "cannot abort the task without aborting the bitmap iteration"); |
342 | 3451 |
3452 // We finished iterating over the region without aborting. | |
3453 regular_clock_call(); | |
3454 if (has_aborted()) | |
3455 mr = MemRegion(); | |
3456 else { | |
1358
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
3457 mr = _cm->region_stack_pop_with_lock(); |
342 | 3458 // it returns MemRegion() if the pop fails |
3459 statsOnly(if (mr.start() != NULL) ++_region_stack_pops ); | |
3460 } | |
3461 } else { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3462 assert(has_aborted(), "currently the only way to do so"); |
342 | 3463 |
3464 // The only way to abort the bitmap iteration is to return | |
3465 // false from the do_bit() method. However, inside the | |
3466 // do_bit() method we move the _region_finger to point to the | |
3467 // object currently being looked at. So, if we bail out, we | |
3468 // have definitely set _region_finger to something non-null. | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3469 assert(_region_finger != NULL, "invariant"); |
342 | 3470 |
3471 // The iteration was actually aborted. So now _region_finger | |
3472 // points to the address of the object we last scanned. If we | |
3473 // leave it there, when we restart this task, we will rescan | |
3474 // the object. It is easy to avoid this. We move the finger by | |
3475 // enough to point to the next possible object header (the | |
3476 // bitmap knows by how much we need to move it as it knows its | |
3477 // granularity). | |
3478 MemRegion newRegion = | |
3479 MemRegion(_nextMarkBitMap->nextWord(_region_finger), mr.end()); | |
3480 | |
3481 if (!newRegion.is_empty()) { | |
3482 if (_cm->verbose_low()) { | |
3483 gclog_or_tty->print_cr("[%d] pushing unscanned region" | |
3484 "[" PTR_FORMAT "," PTR_FORMAT ") on region stack", | |
3485 _task_id, | |
3486 newRegion.start(), newRegion.end()); | |
3487 } | |
3488 // Now push the part of the region we didn't scan on the | |
3489 // region stack to make sure a task scans it later. | |
1358
72f725c5a7be
6940310: G1: MT-unsafe calls to CM::region_stack_push() / CM::region_stack_pop()
tonyp
parents:
1317
diff
changeset
|
3490 _cm->region_stack_push_with_lock(newRegion); |
342 | 3491 } |
3492 // break from while | |
3493 mr = MemRegion(); | |
3494 } | |
3495 _region_finger = NULL; | |
3496 } | |
3497 | |
3498 if (_cm->verbose_low()) | |
3499 gclog_or_tty->print_cr("[%d] drained region stack, size = %d", | |
3500 _task_id, _cm->region_stack_size()); | |
3501 } | |
3502 } | |
3503 | |
3504 void CMTask::print_stats() { | |
3505 gclog_or_tty->print_cr("Marking Stats, task = %d, calls = %d", | |
3506 _task_id, _calls); | |
3507 gclog_or_tty->print_cr(" Elapsed time = %1.2lfms, Termination time = %1.2lfms", | |
3508 _elapsed_time_ms, _termination_time_ms); | |
3509 gclog_or_tty->print_cr(" Step Times (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", | |
3510 _step_times_ms.num(), _step_times_ms.avg(), | |
3511 _step_times_ms.sd()); | |
3512 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", | |
3513 _step_times_ms.maximum(), _step_times_ms.sum()); | |
3514 | |
3515 #if _MARKING_STATS_ | |
3516 gclog_or_tty->print_cr(" Clock Intervals (cum): num = %d, avg = %1.2lfms, sd = %1.2lfms", | |
3517 _all_clock_intervals_ms.num(), _all_clock_intervals_ms.avg(), | |
3518 _all_clock_intervals_ms.sd()); | |
3519 gclog_or_tty->print_cr(" max = %1.2lfms, total = %1.2lfms", | |
3520 _all_clock_intervals_ms.maximum(), | |
3521 _all_clock_intervals_ms.sum()); | |
3522 gclog_or_tty->print_cr(" Clock Causes (cum): scanning = %d, marking = %d", | |
3523 _clock_due_to_scanning, _clock_due_to_marking); | |
3524 gclog_or_tty->print_cr(" Objects: scanned = %d, found on the bitmap = %d", | |
3525 _objs_scanned, _objs_found_on_bitmap); | |
3526 gclog_or_tty->print_cr(" Local Queue: pushes = %d, pops = %d, max size = %d", | |
3527 _local_pushes, _local_pops, _local_max_size); | |
3528 gclog_or_tty->print_cr(" Global Stack: pushes = %d, pops = %d, max size = %d", | |
3529 _global_pushes, _global_pops, _global_max_size); | |
3530 gclog_or_tty->print_cr(" transfers to = %d, transfers from = %d", | |
3531 _global_transfers_to,_global_transfers_from); | |
3532 gclog_or_tty->print_cr(" Regions: claimed = %d, Region Stack: pops = %d", | |
3533 _regions_claimed, _region_stack_pops); | |
3534 gclog_or_tty->print_cr(" SATB buffers: processed = %d", _satb_buffers_processed); | |
3535 gclog_or_tty->print_cr(" Steals: attempts = %d, successes = %d", | |
3536 _steal_attempts, _steals); | |
3537 gclog_or_tty->print_cr(" Aborted: %d, due to", _aborted); | |
3538 gclog_or_tty->print_cr(" overflow: %d, global abort: %d, yield: %d", | |
3539 _aborted_overflow, _aborted_cm_aborted, _aborted_yield); | |
3540 gclog_or_tty->print_cr(" time out: %d, SATB: %d, termination: %d", | |
3541 _aborted_timed_out, _aborted_satb, _aborted_termination); | |
3542 #endif // _MARKING_STATS_ | |
3543 } | |
3544 | |
3545 /***************************************************************************** | |
3546 | |
3547 The do_marking_step(time_target_ms) method is the building block | |
3548 of the parallel marking framework. It can be called in parallel | |
3549 with other invocations of do_marking_step() on different tasks | |
3550 (but only one per task, obviously) and concurrently with the | |
3551 mutator threads, or during remark, hence it eliminates the need | |
3552 for two versions of the code. When called during remark, it will | |
3553 pick up from where the task left off during the concurrent marking | |
3554 phase. Interestingly, tasks are also claimable during evacuation | |
3555 pauses too, since do_marking_step() ensures that it aborts before | |
3556 it needs to yield. | |
3557 | |
3558 The data structures that is uses to do marking work are the | |
3559 following: | |
3560 | |
3561 (1) Marking Bitmap. If there are gray objects that appear only | |
3562 on the bitmap (this happens either when dealing with an overflow | |
3563 or when the initial marking phase has simply marked the roots | |
3564 and didn't push them on the stack), then tasks claim heap | |
3565 regions whose bitmap they then scan to find gray objects. A | |
3566 global finger indicates where the end of the last claimed region | |
3567 is. A local finger indicates how far into the region a task has | |
3568 scanned. The two fingers are used to determine how to gray an | |
3569 object (i.e. whether simply marking it is OK, as it will be | |
3570 visited by a task in the future, or whether it needs to be also | |
3571 pushed on a stack). | |
3572 | |
3573 (2) Local Queue. The local queue of the task which is accessed | |
3574 reasonably efficiently by the task. Other tasks can steal from | |
3575 it when they run out of work. Throughout the marking phase, a | |
3576 task attempts to keep its local queue short but not totally | |
3577 empty, so that entries are available for stealing by other | |
3578 tasks. Only when there is no more work, a task will totally | |
3579 drain its local queue. | |
3580 | |
3581 (3) Global Mark Stack. This handles local queue overflow. During | |
3582 marking only sets of entries are moved between it and the local | |
3583 queues, as access to it requires a mutex and more fine-grain | |
3584 interaction with it which might cause contention. If it | |
3585 overflows, then the marking phase should restart and iterate | |
3586 over the bitmap to identify gray objects. Throughout the marking | |
3587 phase, tasks attempt to keep the global mark stack at a small | |
3588 length but not totally empty, so that entries are available for | |
3589 popping by other tasks. Only when there is no more work, tasks | |
3590 will totally drain the global mark stack. | |
3591 | |
3592 (4) Global Region Stack. Entries on it correspond to areas of | |
3593 the bitmap that need to be scanned since they contain gray | |
3594 objects. Pushes on the region stack only happen during | |
3595 evacuation pauses and typically correspond to areas covered by | |
3596 GC LABS. If it overflows, then the marking phase should restart | |
3597 and iterate over the bitmap to identify gray objects. Tasks will | |
3598 try to totally drain the region stack as soon as possible. | |
3599 | |
3600 (5) SATB Buffer Queue. This is where completed SATB buffers are | |
3601 made available. Buffers are regularly removed from this queue | |
3602 and scanned for roots, so that the queue doesn't get too | |
3603 long. During remark, all completed buffers are processed, as | |
3604 well as the filled in parts of any uncompleted buffers. | |
3605 | |
3606 The do_marking_step() method tries to abort when the time target | |
3607 has been reached. There are a few other cases when the | |
3608 do_marking_step() method also aborts: | |
3609 | |
3610 (1) When the marking phase has been aborted (after a Full GC). | |
3611 | |
3612 (2) When a global overflow (either on the global stack or the | |
3613 region stack) has been triggered. Before the task aborts, it | |
3614 will actually sync up with the other tasks to ensure that all | |
3615 the marking data structures (local queues, stacks, fingers etc.) | |
3616 are re-initialised so that when do_marking_step() completes, | |
3617 the marking phase can immediately restart. | |
3618 | |
3619 (3) When enough completed SATB buffers are available. The | |
3620 do_marking_step() method only tries to drain SATB buffers right | |
3621 at the beginning. So, if enough buffers are available, the | |
3622 marking step aborts and the SATB buffers are processed at | |
3623 the beginning of the next invocation. | |
3624 | |
3625 (4) To yield. when we have to yield then we abort and yield | |
3626 right at the end of do_marking_step(). This saves us from a lot | |
3627 of hassle as, by yielding we might allow a Full GC. If this | |
3628 happens then objects will be compacted underneath our feet, the | |
3629 heap might shrink, etc. We save checking for this by just | |
3630 aborting and doing the yield right at the end. | |
3631 | |
3632 From the above it follows that the do_marking_step() method should | |
3633 be called in a loop (or, otherwise, regularly) until it completes. | |
3634 | |
3635 If a marking step completes without its has_aborted() flag being | |
3636 true, it means it has completed the current marking phase (and | |
3637 also all other marking tasks have done so and have all synced up). | |
3638 | |
3639 A method called regular_clock_call() is invoked "regularly" (in | |
3640 sub ms intervals) throughout marking. It is this clock method that | |
3641 checks all the abort conditions which were mentioned above and | |
3642 decides when the task should abort. A work-based scheme is used to | |
3643 trigger this clock method: when the number of object words the | |
3644 marking phase has scanned or the number of references the marking | |
3645 phase has visited reach a given limit. Additional invocations to | |
3646 the method clock have been planted in a few other strategic places | |
3647 too. The initial reason for the clock method was to avoid calling | |
3648 vtime too regularly, as it is quite expensive. So, once it was in | |
3649 place, it was natural to piggy-back all the other conditions on it | |
3650 too and not constantly check them throughout the code. | |
3651 | |
3652 *****************************************************************************/ | |
3653 | |
3654 void CMTask::do_marking_step(double time_target_ms) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3655 assert(time_target_ms >= 1.0, "minimum granularity is 1ms"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3656 assert(concurrent() == _cm->concurrent(), "they should be the same"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3657 |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3658 assert(concurrent() || _cm->region_stack_empty(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3659 "the region stack should have been cleared before remark"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3660 assert(_region_finger == NULL, |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3661 "this should be non-null only when a region is being scanned"); |
342 | 3662 |
3663 G1CollectorPolicy* g1_policy = _g1h->g1_policy(); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3664 assert(_task_queues != NULL, "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3665 assert(_task_queue != NULL, "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3666 assert(_task_queues->queue(_task_id) == _task_queue, "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3667 |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3668 assert(!_claimed, |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3669 "only one thread should claim this task at any one time"); |
342 | 3670 |
3671 // OK, this doesn't safeguard again all possible scenarios, as it is | |
3672 // possible for two threads to set the _claimed flag at the same | |
3673 // time. But it is only for debugging purposes anyway and it will | |
3674 // catch most problems. | |
3675 _claimed = true; | |
3676 | |
3677 _start_time_ms = os::elapsedVTime() * 1000.0; | |
3678 statsOnly( _interval_start_time_ms = _start_time_ms ); | |
3679 | |
3680 double diff_prediction_ms = | |
3681 g1_policy->get_new_prediction(&_marking_step_diffs_ms); | |
3682 _time_target_ms = time_target_ms - diff_prediction_ms; | |
3683 | |
3684 // set up the variables that are used in the work-based scheme to | |
3685 // call the regular clock method | |
3686 _words_scanned = 0; | |
3687 _refs_reached = 0; | |
3688 recalculate_limits(); | |
3689 | |
3690 // clear all flags | |
3691 clear_has_aborted(); | |
3692 _has_aborted_timed_out = false; | |
3693 _draining_satb_buffers = false; | |
3694 | |
3695 ++_calls; | |
3696 | |
3697 if (_cm->verbose_low()) | |
3698 gclog_or_tty->print_cr("[%d] >>>>>>>>>> START, call = %d, " | |
3699 "target = %1.2lfms >>>>>>>>>>", | |
3700 _task_id, _calls, _time_target_ms); | |
3701 | |
3702 // Set up the bitmap and oop closures. Anything that uses them is | |
3703 // eventually called from this method, so it is OK to allocate these | |
3704 // statically. | |
3705 CMBitMapClosure bitmap_closure(this, _cm, _nextMarkBitMap); | |
3706 CMOopClosure oop_closure(_g1h, _cm, this); | |
3707 set_oop_closure(&oop_closure); | |
3708 | |
3709 if (_cm->has_overflown()) { | |
3710 // This can happen if the region stack or the mark stack overflows | |
3711 // during a GC pause and this task, after a yield point, | |
3712 // restarts. We have to abort as we need to get into the overflow | |
3713 // protocol which happens right at the end of this task. | |
3714 set_has_aborted(); | |
3715 } | |
3716 | |
3717 // First drain any available SATB buffers. After this, we will not | |
3718 // look at SATB buffers before the next invocation of this method. | |
3719 // If enough completed SATB buffers are queued up, the regular clock | |
3720 // will abort this task so that it restarts. | |
3721 drain_satb_buffers(); | |
3722 // ...then partially drain the local queue and the global stack | |
3723 drain_local_queue(true); | |
3724 drain_global_stack(true); | |
3725 | |
3726 // Then totally drain the region stack. We will not look at | |
3727 // it again before the next invocation of this method. Entries on | |
3728 // the region stack are only added during evacuation pauses, for | |
3729 // which we have to yield. When we do, we abort the task anyway so | |
3730 // it will look at the region stack again when it restarts. | |
3731 bitmap_closure.set_scanning_heap_region(false); | |
3732 drain_region_stack(&bitmap_closure); | |
3733 // ...then partially drain the local queue and the global stack | |
3734 drain_local_queue(true); | |
3735 drain_global_stack(true); | |
3736 | |
3737 do { | |
3738 if (!has_aborted() && _curr_region != NULL) { | |
3739 // This means that we're already holding on to a region. | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3740 assert(_finger != NULL, "if region is not NULL, then the finger " |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3741 "should not be NULL either"); |
342 | 3742 |
3743 // We might have restarted this task after an evacuation pause | |
3744 // which might have evacuated the region we're holding on to | |
3745 // underneath our feet. Let's read its limit again to make sure | |
3746 // that we do not iterate over a region of the heap that | |
3747 // contains garbage (update_region_limit() will also move | |
3748 // _finger to the start of the region if it is found empty). | |
3749 update_region_limit(); | |
3750 // We will start from _finger not from the start of the region, | |
3751 // as we might be restarting this task after aborting half-way | |
3752 // through scanning this region. In this case, _finger points to | |
3753 // the address where we last found a marked object. If this is a | |
3754 // fresh region, _finger points to start(). | |
3755 MemRegion mr = MemRegion(_finger, _region_limit); | |
3756 | |
3757 if (_cm->verbose_low()) | |
3758 gclog_or_tty->print_cr("[%d] we're scanning part " | |
3759 "["PTR_FORMAT", "PTR_FORMAT") " | |
3760 "of region "PTR_FORMAT, | |
3761 _task_id, _finger, _region_limit, _curr_region); | |
3762 | |
3763 // Let's iterate over the bitmap of the part of the | |
3764 // region that is left. | |
3765 bitmap_closure.set_scanning_heap_region(true); | |
3766 if (mr.is_empty() || | |
3767 _nextMarkBitMap->iterate(&bitmap_closure, mr)) { | |
3768 // We successfully completed iterating over the region. Now, | |
3769 // let's give up the region. | |
3770 giveup_current_region(); | |
3771 regular_clock_call(); | |
3772 } else { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3773 assert(has_aborted(), "currently the only way to do so"); |
342 | 3774 // The only way to abort the bitmap iteration is to return |
3775 // false from the do_bit() method. However, inside the | |
3776 // do_bit() method we move the _finger to point to the | |
3777 // object currently being looked at. So, if we bail out, we | |
3778 // have definitely set _finger to something non-null. | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3779 assert(_finger != NULL, "invariant"); |
342 | 3780 |
3781 // Region iteration was actually aborted. So now _finger | |
3782 // points to the address of the object we last scanned. If we | |
3783 // leave it there, when we restart this task, we will rescan | |
3784 // the object. It is easy to avoid this. We move the finger by | |
3785 // enough to point to the next possible object header (the | |
3786 // bitmap knows by how much we need to move it as it knows its | |
3787 // granularity). | |
1314
3f0549ed0c98
6921710: G1: assert(new_finger >= _finger && new_finger < _region_limit,"invariant")
apetrusenko
parents:
1284
diff
changeset
|
3788 assert(_finger < _region_limit, "invariant"); |
3f0549ed0c98
6921710: G1: assert(new_finger >= _finger && new_finger < _region_limit,"invariant")
apetrusenko
parents:
1284
diff
changeset
|
3789 HeapWord* new_finger = _nextMarkBitMap->nextWord(_finger); |
3f0549ed0c98
6921710: G1: assert(new_finger >= _finger && new_finger < _region_limit,"invariant")
apetrusenko
parents:
1284
diff
changeset
|
3790 // Check if bitmap iteration was aborted while scanning the last object |
3f0549ed0c98
6921710: G1: assert(new_finger >= _finger && new_finger < _region_limit,"invariant")
apetrusenko
parents:
1284
diff
changeset
|
3791 if (new_finger >= _region_limit) { |
3f0549ed0c98
6921710: G1: assert(new_finger >= _finger && new_finger < _region_limit,"invariant")
apetrusenko
parents:
1284
diff
changeset
|
3792 giveup_current_region(); |
3f0549ed0c98
6921710: G1: assert(new_finger >= _finger && new_finger < _region_limit,"invariant")
apetrusenko
parents:
1284
diff
changeset
|
3793 } else { |
3f0549ed0c98
6921710: G1: assert(new_finger >= _finger && new_finger < _region_limit,"invariant")
apetrusenko
parents:
1284
diff
changeset
|
3794 move_finger_to(new_finger); |
3f0549ed0c98
6921710: G1: assert(new_finger >= _finger && new_finger < _region_limit,"invariant")
apetrusenko
parents:
1284
diff
changeset
|
3795 } |
342 | 3796 } |
3797 } | |
3798 // At this point we have either completed iterating over the | |
3799 // region we were holding on to, or we have aborted. | |
3800 | |
3801 // We then partially drain the local queue and the global stack. | |
3802 // (Do we really need this?) | |
3803 drain_local_queue(true); | |
3804 drain_global_stack(true); | |
3805 | |
3806 // Read the note on the claim_region() method on why it might | |
3807 // return NULL with potentially more regions available for | |
3808 // claiming and why we have to check out_of_regions() to determine | |
3809 // whether we're done or not. | |
3810 while (!has_aborted() && _curr_region == NULL && !_cm->out_of_regions()) { | |
3811 // We are going to try to claim a new region. We should have | |
3812 // given up on the previous one. | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3813 // Separated the asserts so that we know which one fires. |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3814 assert(_curr_region == NULL, "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3815 assert(_finger == NULL, "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3816 assert(_region_limit == NULL, "invariant"); |
342 | 3817 if (_cm->verbose_low()) |
3818 gclog_or_tty->print_cr("[%d] trying to claim a new region", _task_id); | |
3819 HeapRegion* claimed_region = _cm->claim_region(_task_id); | |
3820 if (claimed_region != NULL) { | |
3821 // Yes, we managed to claim one | |
3822 statsOnly( ++_regions_claimed ); | |
3823 | |
3824 if (_cm->verbose_low()) | |
3825 gclog_or_tty->print_cr("[%d] we successfully claimed " | |
3826 "region "PTR_FORMAT, | |
3827 _task_id, claimed_region); | |
3828 | |
3829 setup_for_region(claimed_region); | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3830 assert(_curr_region == claimed_region, "invariant"); |
342 | 3831 } |
3832 // It is important to call the regular clock here. It might take | |
3833 // a while to claim a region if, for example, we hit a large | |
3834 // block of empty regions. So we need to call the regular clock | |
3835 // method once round the loop to make sure it's called | |
3836 // frequently enough. | |
3837 regular_clock_call(); | |
3838 } | |
3839 | |
3840 if (!has_aborted() && _curr_region == NULL) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3841 assert(_cm->out_of_regions(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3842 "at this point we should be out of regions"); |
342 | 3843 } |
3844 } while ( _curr_region != NULL && !has_aborted()); | |
3845 | |
3846 if (!has_aborted()) { | |
3847 // We cannot check whether the global stack is empty, since other | |
343
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3848 // tasks might be pushing objects to it concurrently. We also cannot |
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3849 // check if the region stack is empty because if a thread is aborting |
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3850 // it can push a partially done region back. |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3851 assert(_cm->out_of_regions(), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3852 "at this point we should be out of regions"); |
342 | 3853 |
3854 if (_cm->verbose_low()) | |
3855 gclog_or_tty->print_cr("[%d] all regions claimed", _task_id); | |
3856 | |
3857 // Try to reduce the number of available SATB buffers so that | |
3858 // remark has less work to do. | |
3859 drain_satb_buffers(); | |
3860 } | |
3861 | |
3862 // Since we've done everything else, we can now totally drain the | |
3863 // local queue and global stack. | |
3864 drain_local_queue(false); | |
3865 drain_global_stack(false); | |
3866 | |
3867 // Attempt at work stealing from other task's queues. | |
3868 if (!has_aborted()) { | |
3869 // We have not aborted. This means that we have finished all that | |
3870 // we could. Let's try to do some stealing... | |
3871 | |
3872 // We cannot check whether the global stack is empty, since other | |
343
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3873 // tasks might be pushing objects to it concurrently. We also cannot |
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3874 // check if the region stack is empty because if a thread is aborting |
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3875 // it can push a partially done region back. |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3876 assert(_cm->out_of_regions() && _task_queue->size() == 0, |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3877 "only way to reach here"); |
342 | 3878 |
3879 if (_cm->verbose_low()) | |
3880 gclog_or_tty->print_cr("[%d] starting to steal", _task_id); | |
3881 | |
3882 while (!has_aborted()) { | |
3883 oop obj; | |
3884 statsOnly( ++_steal_attempts ); | |
3885 | |
3886 if (_cm->try_stealing(_task_id, &_hash_seed, obj)) { | |
3887 if (_cm->verbose_medium()) | |
3888 gclog_or_tty->print_cr("[%d] stolen "PTR_FORMAT" successfully", | |
3889 _task_id, (void*) obj); | |
3890 | |
3891 statsOnly( ++_steals ); | |
3892 | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3893 assert(_nextMarkBitMap->isMarked((HeapWord*) obj), |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3894 "any stolen object should be marked"); |
342 | 3895 scan_object(obj); |
3896 | |
3897 // And since we're towards the end, let's totally drain the | |
3898 // local queue and global stack. | |
3899 drain_local_queue(false); | |
3900 drain_global_stack(false); | |
3901 } else { | |
3902 break; | |
3903 } | |
3904 } | |
3905 } | |
3906 | |
3907 // We still haven't aborted. Now, let's try to get into the | |
3908 // termination protocol. | |
3909 if (!has_aborted()) { | |
3910 // We cannot check whether the global stack is empty, since other | |
343
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3911 // tasks might be concurrently pushing objects on it. We also cannot |
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3912 // check if the region stack is empty because if a thread is aborting |
afc1ce1efe66
6710665: G1: guarantee(_cm->out_of_regions() && _cm->region_stack_empty() && _task_queue->size() == 0, ...)
iveresov
parents:
342
diff
changeset
|
3913 // it can push a partially done region back. |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3914 // Separated the asserts so that we know which one fires. |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3915 assert(_cm->out_of_regions(), "only way to reach here"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3916 assert(_task_queue->size() == 0, "only way to reach here"); |
342 | 3917 |
3918 if (_cm->verbose_low()) | |
3919 gclog_or_tty->print_cr("[%d] starting termination protocol", _task_id); | |
3920 | |
3921 _termination_start_time_ms = os::elapsedVTime() * 1000.0; | |
3922 // The CMTask class also extends the TerminatorTerminator class, | |
3923 // hence its should_exit_termination() method will also decide | |
3924 // whether to exit the termination protocol or not. | |
3925 bool finished = _cm->terminator()->offer_termination(this); | |
3926 double termination_end_time_ms = os::elapsedVTime() * 1000.0; | |
3927 _termination_time_ms += | |
3928 termination_end_time_ms - _termination_start_time_ms; | |
3929 | |
3930 if (finished) { | |
3931 // We're all done. | |
3932 | |
3933 if (_task_id == 0) { | |
3934 // let's allow task 0 to do this | |
3935 if (concurrent()) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3936 assert(_cm->concurrent_marking_in_progress(), "invariant"); |
342 | 3937 // we need to set this to false before the next |
3938 // safepoint. This way we ensure that the marking phase | |
3939 // doesn't observe any more heap expansions. | |
3940 _cm->clear_concurrent_marking_in_progress(); | |
3941 } | |
3942 } | |
3943 | |
3944 // We can now guarantee that the global stack is empty, since | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3945 // all other tasks have finished. We separated the guarantees so |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3946 // that, if a condition is false, we can immediately find out |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3947 // which one. |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3948 guarantee(_cm->out_of_regions(), "only way to reach here"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3949 guarantee(_cm->region_stack_empty(), "only way to reach here"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3950 guarantee(_cm->mark_stack_empty(), "only way to reach here"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3951 guarantee(_task_queue->size() == 0, "only way to reach here"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3952 guarantee(!_cm->has_overflown(), "only way to reach here"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3953 guarantee(!_cm->mark_stack_overflow(), "only way to reach here"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
3954 guarantee(!_cm->region_stack_overflow(), "only way to reach here"); |
342 | 3955 |
3956 if (_cm->verbose_low()) | |
3957 gclog_or_tty->print_cr("[%d] all tasks terminated", _task_id); | |
3958 } else { | |
3959 // Apparently there's more work to do. Let's abort this task. It | |
3960 // will restart it and we can hopefully find more things to do. | |
3961 | |
3962 if (_cm->verbose_low()) | |
3963 gclog_or_tty->print_cr("[%d] apparently there is more work to do", _task_id); | |
3964 | |
3965 set_has_aborted(); | |
3966 statsOnly( ++_aborted_termination ); | |
3967 } | |
3968 } | |
3969 | |
3970 // Mainly for debugging purposes to make sure that a pointer to the | |
3971 // closure which was statically allocated in this frame doesn't | |
3972 // escape it by accident. | |
3973 set_oop_closure(NULL); | |
3974 double end_time_ms = os::elapsedVTime() * 1000.0; | |
3975 double elapsed_time_ms = end_time_ms - _start_time_ms; | |
3976 // Update the step history. | |
3977 _step_times_ms.add(elapsed_time_ms); | |
3978 | |
3979 if (has_aborted()) { | |
3980 // The task was aborted for some reason. | |
3981 | |
3982 statsOnly( ++_aborted ); | |
3983 | |
3984 if (_has_aborted_timed_out) { | |
3985 double diff_ms = elapsed_time_ms - _time_target_ms; | |
3986 // Keep statistics of how well we did with respect to hitting | |
3987 // our target only if we actually timed out (if we aborted for | |
3988 // other reasons, then the results might get skewed). | |
3989 _marking_step_diffs_ms.add(diff_ms); | |
3990 } | |
3991 | |
3992 if (_cm->has_overflown()) { | |
3993 // This is the interesting one. We aborted because a global | |
3994 // overflow was raised. This means we have to restart the | |
3995 // marking phase and start iterating over regions. However, in | |
3996 // order to do this we have to make sure that all tasks stop | |
3997 // what they are doing and re-initialise in a safe manner. We | |
3998 // will achieve this with the use of two barrier sync points. | |
3999 | |
4000 if (_cm->verbose_low()) | |
4001 gclog_or_tty->print_cr("[%d] detected overflow", _task_id); | |
4002 | |
4003 _cm->enter_first_sync_barrier(_task_id); | |
4004 // When we exit this sync barrier we know that all tasks have | |
4005 // stopped doing marking work. So, it's now safe to | |
4006 // re-initialise our data structures. At the end of this method, | |
4007 // task 0 will clear the global data structures. | |
4008 | |
4009 statsOnly( ++_aborted_overflow ); | |
4010 | |
4011 // We clear the local state of this task... | |
4012 clear_region_fields(); | |
4013 | |
4014 // ...and enter the second barrier. | |
4015 _cm->enter_second_sync_barrier(_task_id); | |
4016 // At this point everything has bee re-initialised and we're | |
4017 // ready to restart. | |
4018 } | |
4019 | |
4020 if (_cm->verbose_low()) { | |
4021 gclog_or_tty->print_cr("[%d] <<<<<<<<<< ABORTING, target = %1.2lfms, " | |
4022 "elapsed = %1.2lfms <<<<<<<<<<", | |
4023 _task_id, _time_target_ms, elapsed_time_ms); | |
4024 if (_cm->has_aborted()) | |
4025 gclog_or_tty->print_cr("[%d] ========== MARKING ABORTED ==========", | |
4026 _task_id); | |
4027 } | |
4028 } else { | |
4029 if (_cm->verbose_low()) | |
4030 gclog_or_tty->print_cr("[%d] <<<<<<<<<< FINISHED, target = %1.2lfms, " | |
4031 "elapsed = %1.2lfms <<<<<<<<<<", | |
4032 _task_id, _time_target_ms, elapsed_time_ms); | |
4033 } | |
4034 | |
4035 _claimed = false; | |
4036 } | |
4037 | |
4038 CMTask::CMTask(int task_id, | |
4039 ConcurrentMark* cm, | |
4040 CMTaskQueue* task_queue, | |
4041 CMTaskQueueSet* task_queues) | |
4042 : _g1h(G1CollectedHeap::heap()), | |
4043 _task_id(task_id), _cm(cm), | |
4044 _claimed(false), | |
4045 _nextMarkBitMap(NULL), _hash_seed(17), | |
4046 _task_queue(task_queue), | |
4047 _task_queues(task_queues), | |
4048 _oop_closure(NULL) { | |
1023
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
4049 guarantee(task_queue != NULL, "invariant"); |
11d4857fe5e1
6888619: G1: too many guarantees in concurrent marking
tonyp
parents:
1022
diff
changeset
|
4050 guarantee(task_queues != NULL, "invariant"); |
342 | 4051 |
4052 statsOnly( _clock_due_to_scanning = 0; | |
4053 _clock_due_to_marking = 0 ); | |
4054 | |
4055 _marking_step_diffs_ms.add(0.5); | |
4056 } |