Mercurial > hg > truffle
annotate src/share/vm/runtime/synchronizer.cpp @ 1145:e018e6884bd8
6631166: CMS: better heuristics when combatting fragmentation
Summary: Autonomic per-worker free block cache sizing, tunable coalition policies, fixes to per-size block statistics, retuned gain and bandwidth of some feedback loop filters to allow quicker reactivity to abrupt changes in ambient demand, and other heuristics to reduce fragmentation of the CMS old gen. Also tightened some assertions, including those related to locking.
Reviewed-by: jmasa
author | ysr |
---|---|
date | Wed, 23 Dec 2009 09:23:54 -0800 |
parents | b9fba36710f2 |
children | b30a2cd5e3a2 c18cbe5936b8 b96a3e44582f |
rev | line source |
---|---|
0 | 1 /* |
579 | 2 * Copyright 1998-2009 Sun Microsystems, Inc. All Rights Reserved. |
0 | 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
21 * have any questions. | |
22 * | |
23 */ | |
24 | |
25 # include "incls/_precompiled.incl" | |
26 # include "incls/_synchronizer.cpp.incl" | |
27 | |
28 #if defined(__GNUC__) && !defined(IA64) | |
29 // Need to inhibit inlining for older versions of GCC to avoid build-time failures | |
30 #define ATTR __attribute__((noinline)) | |
31 #else | |
32 #define ATTR | |
33 #endif | |
34 | |
35 // Native markword accessors for synchronization and hashCode(). | |
36 // | |
37 // The "core" versions of monitor enter and exit reside in this file. | |
38 // The interpreter and compilers contain specialized transliterated | |
39 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(), | |
40 // for instance. If you make changes here, make sure to modify the | |
41 // interpreter, and both C1 and C2 fast-path inline locking code emission. | |
42 // | |
43 // TODO: merge the objectMonitor and synchronizer classes. | |
44 // | |
45 // ----------------------------------------------------------------------------- | |
46 | |
47 #ifdef DTRACE_ENABLED | |
48 | |
49 // Only bother with this argument setup if dtrace is available | |
50 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. | |
51 | |
52 HS_DTRACE_PROBE_DECL5(hotspot, monitor__wait, | |
53 jlong, uintptr_t, char*, int, long); | |
54 HS_DTRACE_PROBE_DECL4(hotspot, monitor__waited, | |
55 jlong, uintptr_t, char*, int); | |
56 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, | |
57 jlong, uintptr_t, char*, int); | |
58 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, | |
59 jlong, uintptr_t, char*, int); | |
60 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, | |
61 jlong, uintptr_t, char*, int); | |
62 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, | |
63 jlong, uintptr_t, char*, int); | |
64 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, | |
65 jlong, uintptr_t, char*, int); | |
66 | |
67 #define DTRACE_MONITOR_PROBE_COMMON(klassOop, thread) \ | |
68 char* bytes = NULL; \ | |
69 int len = 0; \ | |
70 jlong jtid = SharedRuntime::get_java_tid(thread); \ | |
71 symbolOop klassname = ((oop)(klassOop))->klass()->klass_part()->name(); \ | |
72 if (klassname != NULL) { \ | |
73 bytes = (char*)klassname->bytes(); \ | |
74 len = klassname->utf8_length(); \ | |
75 } | |
76 | |
77 #define DTRACE_MONITOR_WAIT_PROBE(monitor, klassOop, thread, millis) \ | |
78 { \ | |
79 if (DTraceMonitorProbes) { \ | |
80 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ | |
81 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ | |
82 (monitor), bytes, len, (millis)); \ | |
83 } \ | |
84 } | |
85 | |
86 #define DTRACE_MONITOR_PROBE(probe, monitor, klassOop, thread) \ | |
87 { \ | |
88 if (DTraceMonitorProbes) { \ | |
89 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ | |
90 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ | |
91 (uintptr_t)(monitor), bytes, len); \ | |
92 } \ | |
93 } | |
94 | |
95 #else // ndef DTRACE_ENABLED | |
96 | |
97 #define DTRACE_MONITOR_WAIT_PROBE(klassOop, thread, millis, mon) {;} | |
98 #define DTRACE_MONITOR_PROBE(probe, klassOop, thread, mon) {;} | |
99 | |
100 #endif // ndef DTRACE_ENABLED | |
101 | |
102 // ObjectWaiter serves as a "proxy" or surrogate thread. | |
103 // TODO-FIXME: Eliminate ObjectWaiter and use the thread-specific | |
104 // ParkEvent instead. Beware, however, that the JVMTI code | |
105 // knows about ObjectWaiters, so we'll have to reconcile that code. | |
106 // See next_waiter(), first_waiter(), etc. | |
107 | |
108 class ObjectWaiter : public StackObj { | |
109 public: | |
110 enum TStates { TS_UNDEF, TS_READY, TS_RUN, TS_WAIT, TS_ENTER, TS_CXQ } ; | |
111 enum Sorted { PREPEND, APPEND, SORTED } ; | |
112 ObjectWaiter * volatile _next; | |
113 ObjectWaiter * volatile _prev; | |
114 Thread* _thread; | |
115 ParkEvent * _event; | |
116 volatile int _notified ; | |
117 volatile TStates TState ; | |
118 Sorted _Sorted ; // List placement disposition | |
119 bool _active ; // Contention monitoring is enabled | |
120 public: | |
121 ObjectWaiter(Thread* thread) { | |
122 _next = NULL; | |
123 _prev = NULL; | |
124 _notified = 0; | |
125 TState = TS_RUN ; | |
126 _thread = thread; | |
127 _event = thread->_ParkEvent ; | |
128 _active = false; | |
129 assert (_event != NULL, "invariant") ; | |
130 } | |
131 | |
132 void wait_reenter_begin(ObjectMonitor *mon) { | |
133 JavaThread *jt = (JavaThread *)this->_thread; | |
134 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); | |
135 } | |
136 | |
137 void wait_reenter_end(ObjectMonitor *mon) { | |
138 JavaThread *jt = (JavaThread *)this->_thread; | |
139 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); | |
140 } | |
141 }; | |
142 | |
143 enum ManifestConstants { | |
144 ClearResponsibleAtSTW = 0, | |
145 MaximumRecheckInterval = 1000 | |
146 } ; | |
147 | |
148 | |
149 #undef TEVENT | |
150 #define TEVENT(nom) {if (SyncVerbose) FEVENT(nom); } | |
151 | |
152 #define FEVENT(nom) { static volatile int ctr = 0 ; int v = ++ctr ; if ((v & (v-1)) == 0) { ::printf (#nom " : %d \n", v); ::fflush(stdout); }} | |
153 | |
154 #undef TEVENT | |
155 #define TEVENT(nom) {;} | |
156 | |
157 // Performance concern: | |
158 // OrderAccess::storestore() calls release() which STs 0 into the global volatile | |
159 // OrderAccess::Dummy variable. This store is unnecessary for correctness. | |
160 // Many threads STing into a common location causes considerable cache migration | |
161 // or "sloshing" on large SMP system. As such, I avoid using OrderAccess::storestore() | |
162 // until it's repaired. In some cases OrderAccess::fence() -- which incurs local | |
163 // latency on the executing processor -- is a better choice as it scales on SMP | |
164 // systems. See http://blogs.sun.com/dave/entry/biased_locking_in_hotspot for a | |
165 // discussion of coherency costs. Note that all our current reference platforms | |
166 // provide strong ST-ST order, so the issue is moot on IA32, x64, and SPARC. | |
167 // | |
168 // As a general policy we use "volatile" to control compiler-based reordering | |
169 // and explicit fences (barriers) to control for architectural reordering performed | |
170 // by the CPU(s) or platform. | |
171 | |
172 static int MBFence (int x) { OrderAccess::fence(); return x; } | |
173 | |
174 struct SharedGlobals { | |
175 // These are highly shared mostly-read variables. | |
176 // To avoid false-sharing they need to be the sole occupants of a $ line. | |
177 double padPrefix [8]; | |
178 volatile int stwRandom ; | |
179 volatile int stwCycle ; | |
180 | |
181 // Hot RW variables -- Sequester to avoid false-sharing | |
182 double padSuffix [16]; | |
183 volatile int hcSequence ; | |
184 double padFinal [8] ; | |
185 } ; | |
186 | |
187 static SharedGlobals GVars ; | |
188 | |
189 | |
190 // Tunables ... | |
191 // The knob* variables are effectively final. Once set they should | |
192 // never be modified hence. Consider using __read_mostly with GCC. | |
193 | |
194 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins | |
195 static int Knob_HandOff = 0 ; | |
196 static int Knob_Verbose = 0 ; | |
197 static int Knob_ReportSettings = 0 ; | |
198 | |
199 static int Knob_SpinLimit = 5000 ; // derived by an external tool - | |
200 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin | |
201 static int Knob_SpinBackOff = 0 ; // spin-loop backoff | |
202 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS | |
203 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change | |
204 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field | |
205 static int Knob_SpinEarly = 1 ; | |
206 static int Knob_SuccEnabled = 1 ; // futile wake throttling | |
207 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one | |
208 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs | |
209 static int Knob_Bonus = 100 ; // spin success bonus | |
210 static int Knob_BonusB = 100 ; // spin success bonus | |
211 static int Knob_Penalty = 200 ; // spin failure penalty | |
212 static int Knob_Poverty = 1000 ; | |
213 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() | |
214 static int Knob_FixedSpin = 0 ; | |
215 static int Knob_OState = 3 ; // Spinner checks thread state of _owner | |
216 static int Knob_UsePause = 1 ; | |
217 static int Knob_ExitPolicy = 0 ; | |
218 static int Knob_PreSpin = 10 ; // 20-100 likely better | |
219 static int Knob_ResetEvent = 0 ; | |
220 static int BackOffMask = 0 ; | |
221 | |
222 static int Knob_FastHSSEC = 0 ; | |
223 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee | |
224 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline | |
225 static volatile int InitDone = 0 ; | |
226 | |
227 | |
228 // hashCode() generation : | |
229 // | |
230 // Possibilities: | |
231 // * MD5Digest of {obj,stwRandom} | |
232 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. | |
233 // * A DES- or AES-style SBox[] mechanism | |
234 // * One of the Phi-based schemes, such as: | |
235 // 2654435761 = 2^32 * Phi (golden ratio) | |
236 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; | |
237 // * A variation of Marsaglia's shift-xor RNG scheme. | |
238 // * (obj ^ stwRandom) is appealing, but can result | |
239 // in undesirable regularity in the hashCode values of adjacent objects | |
240 // (objects allocated back-to-back, in particular). This could potentially | |
241 // result in hashtable collisions and reduced hashtable efficiency. | |
242 // There are simple ways to "diffuse" the middle address bits over the | |
243 // generated hashCode values: | |
244 // | |
245 | |
246 static inline intptr_t get_next_hash(Thread * Self, oop obj) { | |
247 intptr_t value = 0 ; | |
248 if (hashCode == 0) { | |
249 // This form uses an unguarded global Park-Miller RNG, | |
250 // so it's possible for two threads to race and generate the same RNG. | |
251 // On MP system we'll have lots of RW access to a global, so the | |
252 // mechanism induces lots of coherency traffic. | |
253 value = os::random() ; | |
254 } else | |
255 if (hashCode == 1) { | |
256 // This variation has the property of being stable (idempotent) | |
257 // between STW operations. This can be useful in some of the 1-0 | |
258 // synchronization schemes. | |
259 intptr_t addrBits = intptr_t(obj) >> 3 ; | |
260 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom ; | |
261 } else | |
262 if (hashCode == 2) { | |
263 value = 1 ; // for sensitivity testing | |
264 } else | |
265 if (hashCode == 3) { | |
266 value = ++GVars.hcSequence ; | |
267 } else | |
268 if (hashCode == 4) { | |
269 value = intptr_t(obj) ; | |
270 } else { | |
271 // Marsaglia's xor-shift scheme with thread-specific state | |
272 // This is probably the best overall implementation -- we'll | |
273 // likely make this the default in future releases. | |
274 unsigned t = Self->_hashStateX ; | |
275 t ^= (t << 11) ; | |
276 Self->_hashStateX = Self->_hashStateY ; | |
277 Self->_hashStateY = Self->_hashStateZ ; | |
278 Self->_hashStateZ = Self->_hashStateW ; | |
279 unsigned v = Self->_hashStateW ; | |
280 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)) ; | |
281 Self->_hashStateW = v ; | |
282 value = v ; | |
283 } | |
284 | |
285 value &= markOopDesc::hash_mask; | |
286 if (value == 0) value = 0xBAD ; | |
287 assert (value != markOopDesc::no_hash, "invariant") ; | |
288 TEVENT (hashCode: GENERATE) ; | |
289 return value; | |
290 } | |
291 | |
292 void BasicLock::print_on(outputStream* st) const { | |
293 st->print("monitor"); | |
294 } | |
295 | |
296 void BasicLock::move_to(oop obj, BasicLock* dest) { | |
297 // Check to see if we need to inflate the lock. This is only needed | |
298 // if an object is locked using "this" lightweight monitor. In that | |
299 // case, the displaced_header() is unlocked, because the | |
300 // displaced_header() contains the header for the originally unlocked | |
301 // object. However the object could have already been inflated. But it | |
302 // does not matter, the inflation will just a no-op. For other cases, | |
303 // the displaced header will be either 0x0 or 0x3, which are location | |
304 // independent, therefore the BasicLock is free to move. | |
305 // | |
306 // During OSR we may need to relocate a BasicLock (which contains a | |
307 // displaced word) from a location in an interpreter frame to a | |
308 // new location in a compiled frame. "this" refers to the source | |
309 // basiclock in the interpreter frame. "dest" refers to the destination | |
310 // basiclock in the new compiled frame. We *always* inflate in move_to(). | |
311 // The always-Inflate policy works properly, but in 1.5.0 it can sometimes | |
312 // cause performance problems in code that makes heavy use of a small # of | |
313 // uncontended locks. (We'd inflate during OSR, and then sync performance | |
314 // would subsequently plummet because the thread would be forced thru the slow-path). | |
315 // This problem has been made largely moot on IA32 by inlining the inflated fast-path | |
316 // operations in Fast_Lock and Fast_Unlock in i486.ad. | |
317 // | |
318 // Note that there is a way to safely swing the object's markword from | |
319 // one stack location to another. This avoids inflation. Obviously, | |
320 // we need to ensure that both locations refer to the current thread's stack. | |
321 // There are some subtle concurrency issues, however, and since the benefit is | |
322 // is small (given the support for inflated fast-path locking in the fast_lock, etc) | |
323 // we'll leave that optimization for another time. | |
324 | |
325 if (displaced_header()->is_neutral()) { | |
326 ObjectSynchronizer::inflate_helper(obj); | |
327 // WARNING: We can not put check here, because the inflation | |
328 // will not update the displaced header. Once BasicLock is inflated, | |
329 // no one should ever look at its content. | |
330 } else { | |
331 // Typically the displaced header will be 0 (recursive stack lock) or | |
332 // unused_mark. Naively we'd like to assert that the displaced mark | |
333 // value is either 0, neutral, or 3. But with the advent of the | |
334 // store-before-CAS avoidance in fast_lock/compiler_lock_object | |
335 // we can find any flavor mark in the displaced mark. | |
336 } | |
337 // [RGV] The next line appears to do nothing! | |
338 intptr_t dh = (intptr_t) displaced_header(); | |
339 dest->set_displaced_header(displaced_header()); | |
340 } | |
341 | |
342 // ----------------------------------------------------------------------------- | |
343 | |
344 // standard constructor, allows locking failures | |
345 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { | |
346 _dolock = doLock; | |
347 _thread = thread; | |
348 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) | |
349 _obj = obj; | |
350 | |
351 if (_dolock) { | |
352 TEVENT (ObjectLocker) ; | |
353 | |
354 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); | |
355 } | |
356 } | |
357 | |
358 ObjectLocker::~ObjectLocker() { | |
359 if (_dolock) { | |
360 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); | |
361 } | |
362 } | |
363 | |
364 // ----------------------------------------------------------------------------- | |
365 | |
366 | |
367 PerfCounter * ObjectSynchronizer::_sync_Inflations = NULL ; | |
368 PerfCounter * ObjectSynchronizer::_sync_Deflations = NULL ; | |
369 PerfCounter * ObjectSynchronizer::_sync_ContendedLockAttempts = NULL ; | |
370 PerfCounter * ObjectSynchronizer::_sync_FutileWakeups = NULL ; | |
371 PerfCounter * ObjectSynchronizer::_sync_Parks = NULL ; | |
372 PerfCounter * ObjectSynchronizer::_sync_EmptyNotifications = NULL ; | |
373 PerfCounter * ObjectSynchronizer::_sync_Notifications = NULL ; | |
374 PerfCounter * ObjectSynchronizer::_sync_PrivateA = NULL ; | |
375 PerfCounter * ObjectSynchronizer::_sync_PrivateB = NULL ; | |
376 PerfCounter * ObjectSynchronizer::_sync_SlowExit = NULL ; | |
377 PerfCounter * ObjectSynchronizer::_sync_SlowEnter = NULL ; | |
378 PerfCounter * ObjectSynchronizer::_sync_SlowNotify = NULL ; | |
379 PerfCounter * ObjectSynchronizer::_sync_SlowNotifyAll = NULL ; | |
380 PerfCounter * ObjectSynchronizer::_sync_FailedSpins = NULL ; | |
381 PerfCounter * ObjectSynchronizer::_sync_SuccessfulSpins = NULL ; | |
382 PerfCounter * ObjectSynchronizer::_sync_MonInCirculation = NULL ; | |
383 PerfCounter * ObjectSynchronizer::_sync_MonScavenged = NULL ; | |
384 PerfLongVariable * ObjectSynchronizer::_sync_MonExtant = NULL ; | |
385 | |
386 // One-shot global initialization for the sync subsystem. | |
387 // We could also defer initialization and initialize on-demand | |
388 // the first time we call inflate(). Initialization would | |
389 // be protected - like so many things - by the MonitorCache_lock. | |
390 | |
391 void ObjectSynchronizer::Initialize () { | |
392 static int InitializationCompleted = 0 ; | |
393 assert (InitializationCompleted == 0, "invariant") ; | |
394 InitializationCompleted = 1 ; | |
395 if (UsePerfData) { | |
396 EXCEPTION_MARK ; | |
397 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } | |
398 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } | |
399 NEWPERFCOUNTER(_sync_Inflations) ; | |
400 NEWPERFCOUNTER(_sync_Deflations) ; | |
401 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; | |
402 NEWPERFCOUNTER(_sync_FutileWakeups) ; | |
403 NEWPERFCOUNTER(_sync_Parks) ; | |
404 NEWPERFCOUNTER(_sync_EmptyNotifications) ; | |
405 NEWPERFCOUNTER(_sync_Notifications) ; | |
406 NEWPERFCOUNTER(_sync_SlowEnter) ; | |
407 NEWPERFCOUNTER(_sync_SlowExit) ; | |
408 NEWPERFCOUNTER(_sync_SlowNotify) ; | |
409 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; | |
410 NEWPERFCOUNTER(_sync_FailedSpins) ; | |
411 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; | |
412 NEWPERFCOUNTER(_sync_PrivateA) ; | |
413 NEWPERFCOUNTER(_sync_PrivateB) ; | |
414 NEWPERFCOUNTER(_sync_MonInCirculation) ; | |
415 NEWPERFCOUNTER(_sync_MonScavenged) ; | |
416 NEWPERFVARIABLE(_sync_MonExtant) ; | |
417 #undef NEWPERFCOUNTER | |
418 } | |
419 } | |
420 | |
421 // Compile-time asserts | |
422 // When possible, it's better to catch errors deterministically at | |
423 // compile-time than at runtime. The down-side to using compile-time | |
424 // asserts is that error message -- often something about negative array | |
425 // indices -- is opaque. | |
426 | |
513
2328d1d3f8cf
6781583: Hotspot build fails on linux 64 bit platform with gcc 4.3.2
xlu
parents:
478
diff
changeset
|
427 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } |
0 | 428 |
429 void ObjectMonitor::ctAsserts() { | |
430 CTASSERT(offset_of (ObjectMonitor, _header) == 0); | |
431 } | |
432 | |
433 static int Adjust (volatile int * adr, int dx) { | |
434 int v ; | |
435 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; | |
436 return v ; | |
437 } | |
438 | |
439 // Ad-hoc mutual exclusion primitives: SpinLock and Mux | |
440 // | |
441 // We employ SpinLocks _only for low-contention, fixed-length | |
442 // short-duration critical sections where we're concerned | |
443 // about native mutex_t or HotSpot Mutex:: latency. | |
444 // The mux construct provides a spin-then-block mutual exclusion | |
445 // mechanism. | |
446 // | |
447 // Testing has shown that contention on the ListLock guarding gFreeList | |
448 // is common. If we implement ListLock as a simple SpinLock it's common | |
449 // for the JVM to devolve to yielding with little progress. This is true | |
450 // despite the fact that the critical sections protected by ListLock are | |
451 // extremely short. | |
452 // | |
453 // TODO-FIXME: ListLock should be of type SpinLock. | |
454 // We should make this a 1st-class type, integrated into the lock | |
455 // hierarchy as leaf-locks. Critically, the SpinLock structure | |
456 // should have sufficient padding to avoid false-sharing and excessive | |
457 // cache-coherency traffic. | |
458 | |
459 | |
460 typedef volatile int SpinLockT ; | |
461 | |
462 void Thread::SpinAcquire (volatile int * adr, const char * LockName) { | |
463 if (Atomic::cmpxchg (1, adr, 0) == 0) { | |
464 return ; // normal fast-path return | |
465 } | |
466 | |
467 // Slow-path : We've encountered contention -- Spin/Yield/Block strategy. | |
468 TEVENT (SpinAcquire - ctx) ; | |
469 int ctr = 0 ; | |
470 int Yields = 0 ; | |
471 for (;;) { | |
472 while (*adr != 0) { | |
473 ++ctr ; | |
474 if ((ctr & 0xFFF) == 0 || !os::is_MP()) { | |
475 if (Yields > 5) { | |
476 // Consider using a simple NakedSleep() instead. | |
477 // Then SpinAcquire could be called by non-JVM threads | |
478 Thread::current()->_ParkEvent->park(1) ; | |
479 } else { | |
480 os::NakedYield() ; | |
481 ++Yields ; | |
482 } | |
483 } else { | |
484 SpinPause() ; | |
485 } | |
486 } | |
487 if (Atomic::cmpxchg (1, adr, 0) == 0) return ; | |
488 } | |
489 } | |
490 | |
491 void Thread::SpinRelease (volatile int * adr) { | |
492 assert (*adr != 0, "invariant") ; | |
493 OrderAccess::fence() ; // guarantee at least release consistency. | |
494 // Roach-motel semantics. | |
495 // It's safe if subsequent LDs and STs float "up" into the critical section, | |
496 // but prior LDs and STs within the critical section can't be allowed | |
497 // to reorder or float past the ST that releases the lock. | |
498 *adr = 0 ; | |
499 } | |
500 | |
501 // muxAcquire and muxRelease: | |
502 // | |
503 // * muxAcquire and muxRelease support a single-word lock-word construct. | |
504 // The LSB of the word is set IFF the lock is held. | |
505 // The remainder of the word points to the head of a singly-linked list | |
506 // of threads blocked on the lock. | |
507 // | |
508 // * The current implementation of muxAcquire-muxRelease uses its own | |
509 // dedicated Thread._MuxEvent instance. If we're interested in | |
510 // minimizing the peak number of extant ParkEvent instances then | |
511 // we could eliminate _MuxEvent and "borrow" _ParkEvent as long | |
512 // as certain invariants were satisfied. Specifically, care would need | |
513 // to be taken with regards to consuming unpark() "permits". | |
514 // A safe rule of thumb is that a thread would never call muxAcquire() | |
515 // if it's enqueued (cxq, EntryList, WaitList, etc) and will subsequently | |
516 // park(). Otherwise the _ParkEvent park() operation in muxAcquire() could | |
517 // consume an unpark() permit intended for monitorenter, for instance. | |
518 // One way around this would be to widen the restricted-range semaphore | |
519 // implemented in park(). Another alternative would be to provide | |
520 // multiple instances of the PlatformEvent() for each thread. One | |
521 // instance would be dedicated to muxAcquire-muxRelease, for instance. | |
522 // | |
523 // * Usage: | |
524 // -- Only as leaf locks | |
525 // -- for short-term locking only as muxAcquire does not perform | |
526 // thread state transitions. | |
527 // | |
528 // Alternatives: | |
529 // * We could implement muxAcquire and muxRelease with MCS or CLH locks | |
530 // but with parking or spin-then-park instead of pure spinning. | |
531 // * Use Taura-Oyama-Yonenzawa locks. | |
532 // * It's possible to construct a 1-0 lock if we encode the lockword as | |
533 // (List,LockByte). Acquire will CAS the full lockword while Release | |
534 // will STB 0 into the LockByte. The 1-0 scheme admits stranding, so | |
535 // acquiring threads use timers (ParkTimed) to detect and recover from | |
536 // the stranding window. Thread/Node structures must be aligned on 256-byte | |
537 // boundaries by using placement-new. | |
538 // * Augment MCS with advisory back-link fields maintained with CAS(). | |
539 // Pictorially: LockWord -> T1 <-> T2 <-> T3 <-> ... <-> Tn <-> Owner. | |
540 // The validity of the backlinks must be ratified before we trust the value. | |
541 // If the backlinks are invalid the exiting thread must back-track through the | |
542 // the forward links, which are always trustworthy. | |
543 // * Add a successor indication. The LockWord is currently encoded as | |
544 // (List, LOCKBIT:1). We could also add a SUCCBIT or an explicit _succ variable | |
545 // to provide the usual futile-wakeup optimization. | |
546 // See RTStt for details. | |
547 // * Consider schedctl.sc_nopreempt to cover the critical section. | |
548 // | |
549 | |
550 | |
551 typedef volatile intptr_t MutexT ; // Mux Lock-word | |
552 enum MuxBits { LOCKBIT = 1 } ; | |
553 | |
554 void Thread::muxAcquire (volatile intptr_t * Lock, const char * LockName) { | |
555 intptr_t w = Atomic::cmpxchg_ptr (LOCKBIT, Lock, 0) ; | |
556 if (w == 0) return ; | |
557 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { | |
558 return ; | |
559 } | |
560 | |
561 TEVENT (muxAcquire - Contention) ; | |
562 ParkEvent * const Self = Thread::current()->_MuxEvent ; | |
563 assert ((intptr_t(Self) & LOCKBIT) == 0, "invariant") ; | |
564 for (;;) { | |
565 int its = (os::is_MP() ? 100 : 0) + 1 ; | |
566 | |
567 // Optional spin phase: spin-then-park strategy | |
568 while (--its >= 0) { | |
569 w = *Lock ; | |
570 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { | |
571 return ; | |
572 } | |
573 } | |
574 | |
575 Self->reset() ; | |
576 Self->OnList = intptr_t(Lock) ; | |
577 // The following fence() isn't _strictly necessary as the subsequent | |
578 // CAS() both serializes execution and ratifies the fetched *Lock value. | |
579 OrderAccess::fence(); | |
580 for (;;) { | |
581 w = *Lock ; | |
582 if ((w & LOCKBIT) == 0) { | |
583 if (Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { | |
584 Self->OnList = 0 ; // hygiene - allows stronger asserts | |
585 return ; | |
586 } | |
587 continue ; // Interference -- *Lock changed -- Just retry | |
588 } | |
589 assert (w & LOCKBIT, "invariant") ; | |
590 Self->ListNext = (ParkEvent *) (w & ~LOCKBIT ); | |
591 if (Atomic::cmpxchg_ptr (intptr_t(Self)|LOCKBIT, Lock, w) == w) break ; | |
592 } | |
593 | |
594 while (Self->OnList != 0) { | |
595 Self->park() ; | |
596 } | |
597 } | |
598 } | |
599 | |
600 void Thread::muxAcquireW (volatile intptr_t * Lock, ParkEvent * ev) { | |
601 intptr_t w = Atomic::cmpxchg_ptr (LOCKBIT, Lock, 0) ; | |
602 if (w == 0) return ; | |
603 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { | |
604 return ; | |
605 } | |
606 | |
607 TEVENT (muxAcquire - Contention) ; | |
608 ParkEvent * ReleaseAfter = NULL ; | |
609 if (ev == NULL) { | |
610 ev = ReleaseAfter = ParkEvent::Allocate (NULL) ; | |
611 } | |
612 assert ((intptr_t(ev) & LOCKBIT) == 0, "invariant") ; | |
613 for (;;) { | |
614 guarantee (ev->OnList == 0, "invariant") ; | |
615 int its = (os::is_MP() ? 100 : 0) + 1 ; | |
616 | |
617 // Optional spin phase: spin-then-park strategy | |
618 while (--its >= 0) { | |
619 w = *Lock ; | |
620 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { | |
621 if (ReleaseAfter != NULL) { | |
622 ParkEvent::Release (ReleaseAfter) ; | |
623 } | |
624 return ; | |
625 } | |
626 } | |
627 | |
628 ev->reset() ; | |
629 ev->OnList = intptr_t(Lock) ; | |
630 // The following fence() isn't _strictly necessary as the subsequent | |
631 // CAS() both serializes execution and ratifies the fetched *Lock value. | |
632 OrderAccess::fence(); | |
633 for (;;) { | |
634 w = *Lock ; | |
635 if ((w & LOCKBIT) == 0) { | |
636 if (Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { | |
637 ev->OnList = 0 ; | |
638 // We call ::Release while holding the outer lock, thus | |
639 // artificially lengthening the critical section. | |
640 // Consider deferring the ::Release() until the subsequent unlock(), | |
641 // after we've dropped the outer lock. | |
642 if (ReleaseAfter != NULL) { | |
643 ParkEvent::Release (ReleaseAfter) ; | |
644 } | |
645 return ; | |
646 } | |
647 continue ; // Interference -- *Lock changed -- Just retry | |
648 } | |
649 assert (w & LOCKBIT, "invariant") ; | |
650 ev->ListNext = (ParkEvent *) (w & ~LOCKBIT ); | |
651 if (Atomic::cmpxchg_ptr (intptr_t(ev)|LOCKBIT, Lock, w) == w) break ; | |
652 } | |
653 | |
654 while (ev->OnList != 0) { | |
655 ev->park() ; | |
656 } | |
657 } | |
658 } | |
659 | |
660 // Release() must extract a successor from the list and then wake that thread. | |
661 // It can "pop" the front of the list or use a detach-modify-reattach (DMR) scheme | |
662 // similar to that used by ParkEvent::Allocate() and ::Release(). DMR-based | |
663 // Release() would : | |
664 // (A) CAS() or swap() null to *Lock, releasing the lock and detaching the list. | |
665 // (B) Extract a successor from the private list "in-hand" | |
666 // (C) attempt to CAS() the residual back into *Lock over null. | |
667 // If there were any newly arrived threads and the CAS() would fail. | |
668 // In that case Release() would detach the RATs, re-merge the list in-hand | |
669 // with the RATs and repeat as needed. Alternately, Release() might | |
670 // detach and extract a successor, but then pass the residual list to the wakee. | |
671 // The wakee would be responsible for reattaching and remerging before it | |
672 // competed for the lock. | |
673 // | |
674 // Both "pop" and DMR are immune from ABA corruption -- there can be | |
675 // multiple concurrent pushers, but only one popper or detacher. | |
676 // This implementation pops from the head of the list. This is unfair, | |
677 // but tends to provide excellent throughput as hot threads remain hot. | |
678 // (We wake recently run threads first). | |
679 | |
680 void Thread::muxRelease (volatile intptr_t * Lock) { | |
681 for (;;) { | |
682 const intptr_t w = Atomic::cmpxchg_ptr (0, Lock, LOCKBIT) ; | |
683 assert (w & LOCKBIT, "invariant") ; | |
684 if (w == LOCKBIT) return ; | |
685 ParkEvent * List = (ParkEvent *) (w & ~LOCKBIT) ; | |
686 assert (List != NULL, "invariant") ; | |
687 assert (List->OnList == intptr_t(Lock), "invariant") ; | |
688 ParkEvent * nxt = List->ListNext ; | |
689 | |
690 // The following CAS() releases the lock and pops the head element. | |
691 if (Atomic::cmpxchg_ptr (intptr_t(nxt), Lock, w) != w) { | |
692 continue ; | |
693 } | |
694 List->OnList = 0 ; | |
695 OrderAccess::fence() ; | |
696 List->unpark () ; | |
697 return ; | |
698 } | |
699 } | |
700 | |
701 // ObjectMonitor Lifecycle | |
702 // ----------------------- | |
703 // Inflation unlinks monitors from the global gFreeList and | |
704 // associates them with objects. Deflation -- which occurs at | |
705 // STW-time -- disassociates idle monitors from objects. Such | |
706 // scavenged monitors are returned to the gFreeList. | |
707 // | |
708 // The global list is protected by ListLock. All the critical sections | |
709 // are short and operate in constant-time. | |
710 // | |
711 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. | |
712 // | |
713 // Lifecycle: | |
714 // -- unassigned and on the global free list | |
715 // -- unassigned and on a thread's private omFreeList | |
716 // -- assigned to an object. The object is inflated and the mark refers | |
717 // to the objectmonitor. | |
718 // | |
719 // TODO-FIXME: | |
720 // | |
721 // * We currently protect the gFreeList with a simple lock. | |
722 // An alternate lock-free scheme would be to pop elements from the gFreeList | |
723 // with CAS. This would be safe from ABA corruption as long we only | |
724 // recycled previously appearing elements onto the list in deflate_idle_monitors() | |
725 // at STW-time. Completely new elements could always be pushed onto the gFreeList | |
726 // with CAS. Elements that appeared previously on the list could only | |
727 // be installed at STW-time. | |
728 // | |
729 // * For efficiency and to help reduce the store-before-CAS penalty | |
730 // the objectmonitors on gFreeList or local free lists should be ready to install | |
731 // with the exception of _header and _object. _object can be set after inflation. | |
732 // In particular, keep all objectMonitors on a thread's private list in ready-to-install | |
733 // state with m.Owner set properly. | |
734 // | |
735 // * We could all diffuse contention by using multiple global (FreeList, Lock) | |
736 // pairs -- threads could use trylock() and a cyclic-scan strategy to search for | |
737 // an unlocked free list. | |
738 // | |
739 // * Add lifecycle tags and assert()s. | |
740 // | |
741 // * Be more consistent about when we clear an objectmonitor's fields: | |
742 // A. After extracting the objectmonitor from a free list. | |
743 // B. After adding an objectmonitor to a free list. | |
744 // | |
745 | |
746 ObjectMonitor * ObjectSynchronizer::gBlockList = NULL ; | |
747 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL ; | |
748 static volatile intptr_t ListLock = 0 ; // protects global monitor free-list cache | |
749 #define CHAINMARKER ((oop)-1) | |
750 | |
751 ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) { | |
752 // A large MAXPRIVATE value reduces both list lock contention | |
753 // and list coherency traffic, but also tends to increase the | |
754 // number of objectMonitors in circulation as well as the STW | |
755 // scavenge costs. As usual, we lean toward time in space-time | |
756 // tradeoffs. | |
757 const int MAXPRIVATE = 1024 ; | |
758 for (;;) { | |
759 ObjectMonitor * m ; | |
760 | |
761 // 1: try to allocate from the thread's local omFreeList. | |
762 // Threads will attempt to allocate first from their local list, then | |
763 // from the global list, and only after those attempts fail will the thread | |
764 // attempt to instantiate new monitors. Thread-local free lists take | |
765 // heat off the ListLock and improve allocation latency, as well as reducing | |
766 // coherency traffic on the shared global list. | |
767 m = Self->omFreeList ; | |
768 if (m != NULL) { | |
769 Self->omFreeList = m->FreeNext ; | |
770 Self->omFreeCount -- ; | |
771 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene | |
772 guarantee (m->object() == NULL, "invariant") ; | |
773 return m ; | |
774 } | |
775 | |
776 // 2: try to allocate from the global gFreeList | |
777 // CONSIDER: use muxTry() instead of muxAcquire(). | |
778 // If the muxTry() fails then drop immediately into case 3. | |
779 // If we're using thread-local free lists then try | |
780 // to reprovision the caller's free list. | |
781 if (gFreeList != NULL) { | |
782 // Reprovision the thread's omFreeList. | |
783 // Use bulk transfers to reduce the allocation rate and heat | |
784 // on various locks. | |
785 Thread::muxAcquire (&ListLock, "omAlloc") ; | |
786 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL; ) { | |
787 ObjectMonitor * take = gFreeList ; | |
788 gFreeList = take->FreeNext ; | |
789 guarantee (take->object() == NULL, "invariant") ; | |
790 guarantee (!take->is_busy(), "invariant") ; | |
791 take->Recycle() ; | |
792 omRelease (Self, take) ; | |
793 } | |
794 Thread::muxRelease (&ListLock) ; | |
795 Self->omFreeProvision += 1 + (Self->omFreeProvision/2) ; | |
796 if (Self->omFreeProvision > MAXPRIVATE ) Self->omFreeProvision = MAXPRIVATE ; | |
797 TEVENT (omFirst - reprovision) ; | |
798 continue ; | |
799 } | |
800 | |
801 // 3: allocate a block of new ObjectMonitors | |
802 // Both the local and global free lists are empty -- resort to malloc(). | |
803 // In the current implementation objectMonitors are TSM - immortal. | |
804 assert (_BLOCKSIZE > 1, "invariant") ; | |
805 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE]; | |
806 | |
807 // NOTE: (almost) no way to recover if allocation failed. | |
808 // We might be able to induce a STW safepoint and scavenge enough | |
809 // objectMonitors to permit progress. | |
810 if (temp == NULL) { | |
811 vm_exit_out_of_memory (sizeof (ObjectMonitor[_BLOCKSIZE]), "Allocate ObjectMonitors") ; | |
812 } | |
813 | |
814 // Format the block. | |
815 // initialize the linked list, each monitor points to its next | |
816 // forming the single linked free list, the very first monitor | |
817 // will points to next block, which forms the block list. | |
818 // The trick of using the 1st element in the block as gBlockList | |
819 // linkage should be reconsidered. A better implementation would | |
820 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } | |
821 | |
822 for (int i = 1; i < _BLOCKSIZE ; i++) { | |
823 temp[i].FreeNext = &temp[i+1]; | |
824 } | |
825 | |
826 // terminate the last monitor as the end of list | |
827 temp[_BLOCKSIZE - 1].FreeNext = NULL ; | |
828 | |
829 // Element [0] is reserved for global list linkage | |
830 temp[0].set_object(CHAINMARKER); | |
831 | |
832 // Consider carving out this thread's current request from the | |
833 // block in hand. This avoids some lock traffic and redundant | |
834 // list activity. | |
835 | |
836 // Acquire the ListLock to manipulate BlockList and FreeList. | |
837 // An Oyama-Taura-Yonezawa scheme might be more efficient. | |
838 Thread::muxAcquire (&ListLock, "omAlloc [2]") ; | |
839 | |
840 // Add the new block to the list of extant blocks (gBlockList). | |
841 // The very first objectMonitor in a block is reserved and dedicated. | |
842 // It serves as blocklist "next" linkage. | |
843 temp[0].FreeNext = gBlockList; | |
844 gBlockList = temp; | |
845 | |
846 // Add the new string of objectMonitors to the global free list | |
847 temp[_BLOCKSIZE - 1].FreeNext = gFreeList ; | |
848 gFreeList = temp + 1; | |
849 Thread::muxRelease (&ListLock) ; | |
850 TEVENT (Allocate block of monitors) ; | |
851 } | |
852 } | |
853 | |
854 // Place "m" on the caller's private per-thread omFreeList. | |
855 // In practice there's no need to clamp or limit the number of | |
856 // monitors on a thread's omFreeList as the only time we'll call | |
857 // omRelease is to return a monitor to the free list after a CAS | |
858 // attempt failed. This doesn't allow unbounded #s of monitors to | |
859 // accumulate on a thread's free list. | |
860 // | |
861 // In the future the usage of omRelease() might change and monitors | |
862 // could migrate between free lists. In that case to avoid excessive | |
863 // accumulation we could limit omCount to (omProvision*2), otherwise return | |
864 // the objectMonitor to the global list. We should drain (return) in reasonable chunks. | |
865 // That is, *not* one-at-a-time. | |
866 | |
867 | |
868 void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m) { | |
869 guarantee (m->object() == NULL, "invariant") ; | |
870 m->FreeNext = Self->omFreeList ; | |
871 Self->omFreeList = m ; | |
872 Self->omFreeCount ++ ; | |
873 } | |
874 | |
875 // Return the monitors of a moribund thread's local free list to | |
876 // the global free list. Typically a thread calls omFlush() when | |
877 // it's dying. We could also consider having the VM thread steal | |
878 // monitors from threads that have not run java code over a few | |
879 // consecutive STW safepoints. Relatedly, we might decay | |
880 // omFreeProvision at STW safepoints. | |
881 // | |
882 // We currently call omFlush() from the Thread:: dtor _after the thread | |
883 // has been excised from the thread list and is no longer a mutator. | |
884 // That means that omFlush() can run concurrently with a safepoint and | |
885 // the scavenge operator. Calling omFlush() from JavaThread::exit() might | |
886 // be a better choice as we could safely reason that that the JVM is | |
887 // not at a safepoint at the time of the call, and thus there could | |
888 // be not inopportune interleavings between omFlush() and the scavenge | |
889 // operator. | |
890 | |
891 void ObjectSynchronizer::omFlush (Thread * Self) { | |
892 ObjectMonitor * List = Self->omFreeList ; // Null-terminated SLL | |
893 Self->omFreeList = NULL ; | |
894 if (List == NULL) return ; | |
895 ObjectMonitor * Tail = NULL ; | |
896 ObjectMonitor * s ; | |
897 for (s = List ; s != NULL ; s = s->FreeNext) { | |
898 Tail = s ; | |
899 guarantee (s->object() == NULL, "invariant") ; | |
900 guarantee (!s->is_busy(), "invariant") ; | |
901 s->set_owner (NULL) ; // redundant but good hygiene | |
902 TEVENT (omFlush - Move one) ; | |
903 } | |
904 | |
905 guarantee (Tail != NULL && List != NULL, "invariant") ; | |
906 Thread::muxAcquire (&ListLock, "omFlush") ; | |
907 Tail->FreeNext = gFreeList ; | |
908 gFreeList = List ; | |
909 Thread::muxRelease (&ListLock) ; | |
910 TEVENT (omFlush) ; | |
911 } | |
912 | |
913 | |
914 // Get the next block in the block list. | |
915 static inline ObjectMonitor* next(ObjectMonitor* block) { | |
916 assert(block->object() == CHAINMARKER, "must be a block header"); | |
917 block = block->FreeNext ; | |
918 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); | |
919 return block; | |
920 } | |
921 | |
922 // Fast path code shared by multiple functions | |
923 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { | |
924 markOop mark = obj->mark(); | |
925 if (mark->has_monitor()) { | |
926 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); | |
927 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); | |
928 return mark->monitor(); | |
929 } | |
930 return ObjectSynchronizer::inflate(Thread::current(), obj); | |
931 } | |
932 | |
933 // Note that we could encounter some performance loss through false-sharing as | |
934 // multiple locks occupy the same $ line. Padding might be appropriate. | |
935 | |
936 #define NINFLATIONLOCKS 256 | |
937 static volatile intptr_t InflationLocks [NINFLATIONLOCKS] ; | |
938 | |
939 static markOop ReadStableMark (oop obj) { | |
940 markOop mark = obj->mark() ; | |
941 if (!mark->is_being_inflated()) { | |
942 return mark ; // normal fast-path return | |
943 } | |
944 | |
945 int its = 0 ; | |
946 for (;;) { | |
947 markOop mark = obj->mark() ; | |
948 if (!mark->is_being_inflated()) { | |
949 return mark ; // normal fast-path return | |
950 } | |
951 | |
952 // The object is being inflated by some other thread. | |
953 // The caller of ReadStableMark() must wait for inflation to complete. | |
954 // Avoid live-lock | |
955 // TODO: consider calling SafepointSynchronize::do_call_back() while | |
956 // spinning to see if there's a safepoint pending. If so, immediately | |
957 // yielding or blocking would be appropriate. Avoid spinning while | |
958 // there is a safepoint pending. | |
959 // TODO: add inflation contention performance counters. | |
960 // TODO: restrict the aggregate number of spinners. | |
961 | |
962 ++its ; | |
963 if (its > 10000 || !os::is_MP()) { | |
964 if (its & 1) { | |
965 os::NakedYield() ; | |
966 TEVENT (Inflate: INFLATING - yield) ; | |
967 } else { | |
968 // Note that the following code attenuates the livelock problem but is not | |
969 // a complete remedy. A more complete solution would require that the inflating | |
970 // thread hold the associated inflation lock. The following code simply restricts | |
971 // the number of spinners to at most one. We'll have N-2 threads blocked | |
972 // on the inflationlock, 1 thread holding the inflation lock and using | |
973 // a yield/park strategy, and 1 thread in the midst of inflation. | |
974 // A more refined approach would be to change the encoding of INFLATING | |
975 // to allow encapsulation of a native thread pointer. Threads waiting for | |
976 // inflation to complete would use CAS to push themselves onto a singly linked | |
977 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag | |
978 // and calling park(). When inflation was complete the thread that accomplished inflation | |
979 // would detach the list and set the markword to inflated with a single CAS and | |
980 // then for each thread on the list, set the flag and unpark() the thread. | |
981 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease | |
982 // wakes at most one thread whereas we need to wake the entire list. | |
983 int ix = (intptr_t(obj) >> 5) & (NINFLATIONLOCKS-1) ; | |
984 int YieldThenBlock = 0 ; | |
985 assert (ix >= 0 && ix < NINFLATIONLOCKS, "invariant") ; | |
986 assert ((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant") ; | |
987 Thread::muxAcquire (InflationLocks + ix, "InflationLock") ; | |
988 while (obj->mark() == markOopDesc::INFLATING()) { | |
989 // Beware: NakedYield() is advisory and has almost no effect on some platforms | |
990 // so we periodically call Self->_ParkEvent->park(1). | |
991 // We use a mixed spin/yield/block mechanism. | |
992 if ((YieldThenBlock++) >= 16) { | |
993 Thread::current()->_ParkEvent->park(1) ; | |
994 } else { | |
995 os::NakedYield() ; | |
996 } | |
997 } | |
998 Thread::muxRelease (InflationLocks + ix ) ; | |
999 TEVENT (Inflate: INFLATING - yield/park) ; | |
1000 } | |
1001 } else { | |
1002 SpinPause() ; // SMP-polite spinning | |
1003 } | |
1004 } | |
1005 } | |
1006 | |
1007 ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) { | |
1008 // Inflate mutates the heap ... | |
1009 // Relaxing assertion for bug 6320749. | |
1010 assert (Universe::verify_in_progress() || | |
1011 !SafepointSynchronize::is_at_safepoint(), "invariant") ; | |
1012 | |
1013 for (;;) { | |
1014 const markOop mark = object->mark() ; | |
1015 assert (!mark->has_bias_pattern(), "invariant") ; | |
1016 | |
1017 // The mark can be in one of the following states: | |
1018 // * Inflated - just return | |
1019 // * Stack-locked - coerce it to inflated | |
1020 // * INFLATING - busy wait for conversion to complete | |
1021 // * Neutral - aggressively inflate the object. | |
1022 // * BIASED - Illegal. We should never see this | |
1023 | |
1024 // CASE: inflated | |
1025 if (mark->has_monitor()) { | |
1026 ObjectMonitor * inf = mark->monitor() ; | |
1027 assert (inf->header()->is_neutral(), "invariant"); | |
1028 assert (inf->object() == object, "invariant") ; | |
1029 assert (ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); | |
1030 return inf ; | |
1031 } | |
1032 | |
1033 // CASE: inflation in progress - inflating over a stack-lock. | |
1034 // Some other thread is converting from stack-locked to inflated. | |
1035 // Only that thread can complete inflation -- other threads must wait. | |
1036 // The INFLATING value is transient. | |
1037 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. | |
1038 // We could always eliminate polling by parking the thread on some auxiliary list. | |
1039 if (mark == markOopDesc::INFLATING()) { | |
1040 TEVENT (Inflate: spin while INFLATING) ; | |
1041 ReadStableMark(object) ; | |
1042 continue ; | |
1043 } | |
1044 | |
1045 // CASE: stack-locked | |
1046 // Could be stack-locked either by this thread or by some other thread. | |
1047 // | |
1048 // Note that we allocate the objectmonitor speculatively, _before_ attempting | |
1049 // to install INFLATING into the mark word. We originally installed INFLATING, | |
1050 // allocated the objectmonitor, and then finally STed the address of the | |
1051 // objectmonitor into the mark. This was correct, but artificially lengthened | |
1052 // the interval in which INFLATED appeared in the mark, thus increasing | |
1053 // the odds of inflation contention. | |
1054 // | |
1055 // We now use per-thread private objectmonitor free lists. | |
1056 // These list are reprovisioned from the global free list outside the | |
1057 // critical INFLATING...ST interval. A thread can transfer | |
1058 // multiple objectmonitors en-mass from the global free list to its local free list. | |
1059 // This reduces coherency traffic and lock contention on the global free list. | |
1060 // Using such local free lists, it doesn't matter if the omAlloc() call appears | |
1061 // before or after the CAS(INFLATING) operation. | |
1062 // See the comments in omAlloc(). | |
1063 | |
1064 if (mark->has_locker()) { | |
1065 ObjectMonitor * m = omAlloc (Self) ; | |
1066 // Optimistically prepare the objectmonitor - anticipate successful CAS | |
1067 // We do this before the CAS in order to minimize the length of time | |
1068 // in which INFLATING appears in the mark. | |
1069 m->Recycle(); | |
1070 m->FreeNext = NULL ; | |
1071 m->_Responsible = NULL ; | |
1072 m->OwnerIsThread = 0 ; | |
1073 m->_recursions = 0 ; | |
1074 m->_SpinDuration = Knob_SpinLimit ; // Consider: maintain by type/class | |
1075 | |
1076 markOop cmp = (markOop) Atomic::cmpxchg_ptr (markOopDesc::INFLATING(), object->mark_addr(), mark) ; | |
1077 if (cmp != mark) { | |
1078 omRelease (Self, m) ; | |
1079 continue ; // Interference -- just retry | |
1080 } | |
1081 | |
1082 // We've successfully installed INFLATING (0) into the mark-word. | |
1083 // This is the only case where 0 will appear in a mark-work. | |
1084 // Only the singular thread that successfully swings the mark-word | |
1085 // to 0 can perform (or more precisely, complete) inflation. | |
1086 // | |
1087 // Why do we CAS a 0 into the mark-word instead of just CASing the | |
1088 // mark-word from the stack-locked value directly to the new inflated state? | |
1089 // Consider what happens when a thread unlocks a stack-locked object. | |
1090 // It attempts to use CAS to swing the displaced header value from the | |
1091 // on-stack basiclock back into the object header. Recall also that the | |
1092 // header value (hashcode, etc) can reside in (a) the object header, or | |
1093 // (b) a displaced header associated with the stack-lock, or (c) a displaced | |
1094 // header in an objectMonitor. The inflate() routine must copy the header | |
1095 // value from the basiclock on the owner's stack to the objectMonitor, all | |
1096 // the while preserving the hashCode stability invariants. If the owner | |
1097 // decides to release the lock while the value is 0, the unlock will fail | |
1098 // and control will eventually pass from slow_exit() to inflate. The owner | |
1099 // will then spin, waiting for the 0 value to disappear. Put another way, | |
1100 // the 0 causes the owner to stall if the owner happens to try to | |
1101 // drop the lock (restoring the header from the basiclock to the object) | |
1102 // while inflation is in-progress. This protocol avoids races that might | |
1103 // would otherwise permit hashCode values to change or "flicker" for an object. | |
1104 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. | |
1105 // 0 serves as a "BUSY" inflate-in-progress indicator. | |
1106 | |
1107 | |
1108 // fetch the displaced mark from the owner's stack. | |
1109 // The owner can't die or unwind past the lock while our INFLATING | |
1110 // object is in the mark. Furthermore the owner can't complete | |
1111 // an unlock on the object, either. | |
1112 markOop dmw = mark->displaced_mark_helper() ; | |
1113 assert (dmw->is_neutral(), "invariant") ; | |
1114 | |
1115 // Setup monitor fields to proper values -- prepare the monitor | |
1116 m->set_header(dmw) ; | |
1117 | |
1118 // Optimization: if the mark->locker stack address is associated | |
1119 // with this thread we could simply set m->_owner = Self and | |
702
b9fba36710f2
6699669: Hotspot server leaves synchronized block with monitor in bad state
xlu
parents:
579
diff
changeset
|
1120 // m->OwnerIsThread = 1. Note that a thread can inflate an object |
0 | 1121 // that it has stack-locked -- as might happen in wait() -- directly |
1122 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. | |
702
b9fba36710f2
6699669: Hotspot server leaves synchronized block with monitor in bad state
xlu
parents:
579
diff
changeset
|
1123 m->set_owner(mark->locker()); |
0 | 1124 m->set_object(object); |
1125 // TODO-FIXME: assert BasicLock->dhw != 0. | |
1126 | |
1127 // Must preserve store ordering. The monitor state must | |
1128 // be stable at the time of publishing the monitor address. | |
1129 guarantee (object->mark() == markOopDesc::INFLATING(), "invariant") ; | |
1130 object->release_set_mark(markOopDesc::encode(m)); | |
1131 | |
1132 // Hopefully the performance counters are allocated on distinct cache lines | |
1133 // to avoid false sharing on MP systems ... | |
1134 if (_sync_Inflations != NULL) _sync_Inflations->inc() ; | |
1135 TEVENT(Inflate: overwrite stacklock) ; | |
1136 if (TraceMonitorInflation) { | |
1137 if (object->is_instance()) { | |
1138 ResourceMark rm; | |
1139 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", | |
1140 (intptr_t) object, (intptr_t) object->mark(), | |
1141 Klass::cast(object->klass())->external_name()); | |
1142 } | |
1143 } | |
1144 return m ; | |
1145 } | |
1146 | |
1147 // CASE: neutral | |
1148 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. | |
1149 // If we know we're inflating for entry it's better to inflate by swinging a | |
1150 // pre-locked objectMonitor pointer into the object header. A successful | |
1151 // CAS inflates the object *and* confers ownership to the inflating thread. | |
1152 // In the current implementation we use a 2-step mechanism where we CAS() | |
1153 // to inflate and then CAS() again to try to swing _owner from NULL to Self. | |
1154 // An inflateTry() method that we could call from fast_enter() and slow_enter() | |
1155 // would be useful. | |
1156 | |
1157 assert (mark->is_neutral(), "invariant"); | |
1158 ObjectMonitor * m = omAlloc (Self) ; | |
1159 // prepare m for installation - set monitor to initial state | |
1160 m->Recycle(); | |
1161 m->set_header(mark); | |
1162 m->set_owner(NULL); | |
1163 m->set_object(object); | |
1164 m->OwnerIsThread = 1 ; | |
1165 m->_recursions = 0 ; | |
1166 m->FreeNext = NULL ; | |
1167 m->_Responsible = NULL ; | |
1168 m->_SpinDuration = Knob_SpinLimit ; // consider: keep metastats by type/class | |
1169 | |
1170 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) { | |
1171 m->set_object (NULL) ; | |
1172 m->set_owner (NULL) ; | |
1173 m->OwnerIsThread = 0 ; | |
1174 m->Recycle() ; | |
1175 omRelease (Self, m) ; | |
1176 m = NULL ; | |
1177 continue ; | |
1178 // interference - the markword changed - just retry. | |
1179 // The state-transitions are one-way, so there's no chance of | |
1180 // live-lock -- "Inflated" is an absorbing state. | |
1181 } | |
1182 | |
1183 // Hopefully the performance counters are allocated on distinct | |
1184 // cache lines to avoid false sharing on MP systems ... | |
1185 if (_sync_Inflations != NULL) _sync_Inflations->inc() ; | |
1186 TEVENT(Inflate: overwrite neutral) ; | |
1187 if (TraceMonitorInflation) { | |
1188 if (object->is_instance()) { | |
1189 ResourceMark rm; | |
1190 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", | |
1191 (intptr_t) object, (intptr_t) object->mark(), | |
1192 Klass::cast(object->klass())->external_name()); | |
1193 } | |
1194 } | |
1195 return m ; | |
1196 } | |
1197 } | |
1198 | |
1199 | |
1200 // This the fast monitor enter. The interpreter and compiler use | |
1201 // some assembly copies of this code. Make sure update those code | |
1202 // if the following function is changed. The implementation is | |
1203 // extremely sensitive to race condition. Be careful. | |
1204 | |
1205 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) { | |
1206 if (UseBiasedLocking) { | |
1207 if (!SafepointSynchronize::is_at_safepoint()) { | |
1208 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); | |
1209 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { | |
1210 return; | |
1211 } | |
1212 } else { | |
1213 assert(!attempt_rebias, "can not rebias toward VM thread"); | |
1214 BiasedLocking::revoke_at_safepoint(obj); | |
1215 } | |
1216 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
702
b9fba36710f2
6699669: Hotspot server leaves synchronized block with monitor in bad state
xlu
parents:
579
diff
changeset
|
1217 } |
b9fba36710f2
6699669: Hotspot server leaves synchronized block with monitor in bad state
xlu
parents:
579
diff
changeset
|
1218 |
b9fba36710f2
6699669: Hotspot server leaves synchronized block with monitor in bad state
xlu
parents:
579
diff
changeset
|
1219 slow_enter (obj, lock, THREAD) ; |
0 | 1220 } |
1221 | |
1222 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { | |
1223 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here"); | |
1224 // if displaced header is null, the previous enter is recursive enter, no-op | |
1225 markOop dhw = lock->displaced_header(); | |
1226 markOop mark ; | |
1227 if (dhw == NULL) { | |
1228 // Recursive stack-lock. | |
1229 // Diagnostics -- Could be: stack-locked, inflating, inflated. | |
1230 mark = object->mark() ; | |
1231 assert (!mark->is_neutral(), "invariant") ; | |
1232 if (mark->has_locker() && mark != markOopDesc::INFLATING()) { | |
1233 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant") ; | |
1234 } | |
1235 if (mark->has_monitor()) { | |
1236 ObjectMonitor * m = mark->monitor() ; | |
1237 assert(((oop)(m->object()))->mark() == mark, "invariant") ; | |
1238 assert(m->is_entered(THREAD), "invariant") ; | |
1239 } | |
1240 return ; | |
1241 } | |
1242 | |
1243 mark = object->mark() ; | |
1244 | |
1245 // If the object is stack-locked by the current thread, try to | |
1246 // swing the displaced header from the box back to the mark. | |
1247 if (mark == (markOop) lock) { | |
1248 assert (dhw->is_neutral(), "invariant") ; | |
1249 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) { | |
1250 TEVENT (fast_exit: release stacklock) ; | |
1251 return; | |
1252 } | |
1253 } | |
1254 | |
1255 ObjectSynchronizer::inflate(THREAD, object)->exit (THREAD) ; | |
1256 } | |
1257 | |
1258 // This routine is used to handle interpreter/compiler slow case | |
1259 // We don't need to use fast path here, because it must have been | |
1260 // failed in the interpreter/compiler code. | |
1261 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { | |
1262 markOop mark = obj->mark(); | |
1263 assert(!mark->has_bias_pattern(), "should not see bias pattern here"); | |
1264 | |
1265 if (mark->is_neutral()) { | |
1266 // Anticipate successful CAS -- the ST of the displaced mark must | |
1267 // be visible <= the ST performed by the CAS. | |
1268 lock->set_displaced_header(mark); | |
1269 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) { | |
1270 TEVENT (slow_enter: release stacklock) ; | |
1271 return ; | |
1272 } | |
1273 // Fall through to inflate() ... | |
1274 } else | |
1275 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { | |
1276 assert(lock != mark->locker(), "must not re-lock the same lock"); | |
1277 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); | |
1278 lock->set_displaced_header(NULL); | |
1279 return; | |
1280 } | |
1281 | |
1282 #if 0 | |
1283 // The following optimization isn't particularly useful. | |
1284 if (mark->has_monitor() && mark->monitor()->is_entered(THREAD)) { | |
1285 lock->set_displaced_header (NULL) ; | |
1286 return ; | |
1287 } | |
1288 #endif | |
1289 | |
1290 // The object header will never be displaced to this lock, | |
1291 // so it does not matter what the value is, except that it | |
1292 // must be non-zero to avoid looking like a re-entrant lock, | |
1293 // and must not look locked either. | |
1294 lock->set_displaced_header(markOopDesc::unused_mark()); | |
1295 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); | |
1296 } | |
1297 | |
1298 // This routine is used to handle interpreter/compiler slow case | |
1299 // We don't need to use fast path here, because it must have | |
1300 // failed in the interpreter/compiler code. Simply use the heavy | |
1301 // weight monitor should be ok, unless someone find otherwise. | |
1302 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { | |
1303 fast_exit (object, lock, THREAD) ; | |
1304 } | |
1305 | |
1306 // NOTE: must use heavy weight monitor to handle jni monitor enter | |
1307 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { // possible entry from jni enter | |
1308 // the current locking is from JNI instead of Java code | |
1309 TEVENT (jni_enter) ; | |
1310 if (UseBiasedLocking) { | |
1311 BiasedLocking::revoke_and_rebias(obj, false, THREAD); | |
1312 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1313 } | |
1314 THREAD->set_current_pending_monitor_is_from_java(false); | |
1315 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); | |
1316 THREAD->set_current_pending_monitor_is_from_java(true); | |
1317 } | |
1318 | |
1319 // NOTE: must use heavy weight monitor to handle jni monitor enter | |
1320 bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) { | |
1321 if (UseBiasedLocking) { | |
1322 BiasedLocking::revoke_and_rebias(obj, false, THREAD); | |
1323 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1324 } | |
1325 | |
1326 ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj()); | |
1327 return monitor->try_enter(THREAD); | |
1328 } | |
1329 | |
1330 | |
1331 // NOTE: must use heavy weight monitor to handle jni monitor exit | |
1332 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { | |
1333 TEVENT (jni_exit) ; | |
1334 if (UseBiasedLocking) { | |
1335 BiasedLocking::revoke_and_rebias(obj, false, THREAD); | |
1336 } | |
1337 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1338 | |
1339 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj); | |
1340 // If this thread has locked the object, exit the monitor. Note: can't use | |
1341 // monitor->check(CHECK); must exit even if an exception is pending. | |
1342 if (monitor->check(THREAD)) { | |
1343 monitor->exit(THREAD); | |
1344 } | |
1345 } | |
1346 | |
1347 // complete_exit()/reenter() are used to wait on a nested lock | |
1348 // i.e. to give up an outer lock completely and then re-enter | |
1349 // Used when holding nested locks - lock acquisition order: lock1 then lock2 | |
1350 // 1) complete_exit lock1 - saving recursion count | |
1351 // 2) wait on lock2 | |
1352 // 3) when notified on lock2, unlock lock2 | |
1353 // 4) reenter lock1 with original recursion count | |
1354 // 5) lock lock2 | |
1355 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() | |
1356 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { | |
1357 TEVENT (complete_exit) ; | |
1358 if (UseBiasedLocking) { | |
1359 BiasedLocking::revoke_and_rebias(obj, false, THREAD); | |
1360 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1361 } | |
1362 | |
1363 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); | |
1364 | |
1365 return monitor->complete_exit(THREAD); | |
1366 } | |
1367 | |
1368 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() | |
1369 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { | |
1370 TEVENT (reenter) ; | |
1371 if (UseBiasedLocking) { | |
1372 BiasedLocking::revoke_and_rebias(obj, false, THREAD); | |
1373 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1374 } | |
1375 | |
1376 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); | |
1377 | |
1378 monitor->reenter(recursion, THREAD); | |
1379 } | |
1380 | |
1381 // This exists only as a workaround of dtrace bug 6254741 | |
1382 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { | |
1383 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); | |
1384 return 0; | |
1385 } | |
1386 | |
1387 // NOTE: must use heavy weight monitor to handle wait() | |
1388 void ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { | |
1389 if (UseBiasedLocking) { | |
1390 BiasedLocking::revoke_and_rebias(obj, false, THREAD); | |
1391 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1392 } | |
1393 if (millis < 0) { | |
1394 TEVENT (wait - throw IAX) ; | |
1395 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); | |
1396 } | |
1397 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); | |
1398 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); | |
1399 monitor->wait(millis, true, THREAD); | |
1400 | |
1401 /* This dummy call is in place to get around dtrace bug 6254741. Once | |
1402 that's fixed we can uncomment the following line and remove the call */ | |
1403 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); | |
1404 dtrace_waited_probe(monitor, obj, THREAD); | |
1405 } | |
1406 | |
1407 void ObjectSynchronizer::waitUninterruptibly (Handle obj, jlong millis, TRAPS) { | |
1408 if (UseBiasedLocking) { | |
1409 BiasedLocking::revoke_and_rebias(obj, false, THREAD); | |
1410 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1411 } | |
1412 if (millis < 0) { | |
1413 TEVENT (wait - throw IAX) ; | |
1414 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); | |
1415 } | |
1416 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD) ; | |
1417 } | |
1418 | |
1419 void ObjectSynchronizer::notify(Handle obj, TRAPS) { | |
1420 if (UseBiasedLocking) { | |
1421 BiasedLocking::revoke_and_rebias(obj, false, THREAD); | |
1422 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1423 } | |
1424 | |
1425 markOop mark = obj->mark(); | |
1426 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { | |
1427 return; | |
1428 } | |
1429 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD); | |
1430 } | |
1431 | |
1432 // NOTE: see comment of notify() | |
1433 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { | |
1434 if (UseBiasedLocking) { | |
1435 BiasedLocking::revoke_and_rebias(obj, false, THREAD); | |
1436 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1437 } | |
1438 | |
1439 markOop mark = obj->mark(); | |
1440 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { | |
1441 return; | |
1442 } | |
1443 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD); | |
1444 } | |
1445 | |
1446 intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) { | |
1447 if (UseBiasedLocking) { | |
1448 // NOTE: many places throughout the JVM do not expect a safepoint | |
1449 // to be taken here, in particular most operations on perm gen | |
1450 // objects. However, we only ever bias Java instances and all of | |
1451 // the call sites of identity_hash that might revoke biases have | |
1452 // been checked to make sure they can handle a safepoint. The | |
1453 // added check of the bias pattern is to avoid useless calls to | |
1454 // thread-local storage. | |
1455 if (obj->mark()->has_bias_pattern()) { | |
1456 // Box and unbox the raw reference just in case we cause a STW safepoint. | |
1457 Handle hobj (Self, obj) ; | |
1458 // Relaxing assertion for bug 6320749. | |
1459 assert (Universe::verify_in_progress() || | |
1460 !SafepointSynchronize::is_at_safepoint(), | |
1461 "biases should not be seen by VM thread here"); | |
1462 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); | |
1463 obj = hobj() ; | |
1464 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1465 } | |
1466 } | |
1467 | |
1468 // hashCode() is a heap mutator ... | |
1469 // Relaxing assertion for bug 6320749. | |
1470 assert (Universe::verify_in_progress() || | |
1471 !SafepointSynchronize::is_at_safepoint(), "invariant") ; | |
1472 assert (Universe::verify_in_progress() || | |
1473 Self->is_Java_thread() , "invariant") ; | |
1474 assert (Universe::verify_in_progress() || | |
1475 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; | |
1476 | |
1477 ObjectMonitor* monitor = NULL; | |
1478 markOop temp, test; | |
1479 intptr_t hash; | |
1480 markOop mark = ReadStableMark (obj); | |
1481 | |
1482 // object should remain ineligible for biased locking | |
1483 assert (!mark->has_bias_pattern(), "invariant") ; | |
1484 | |
1485 if (mark->is_neutral()) { | |
1486 hash = mark->hash(); // this is a normal header | |
1487 if (hash) { // if it has hash, just return it | |
1488 return hash; | |
1489 } | |
1490 hash = get_next_hash(Self, obj); // allocate a new hash code | |
1491 temp = mark->copy_set_hash(hash); // merge the hash code into header | |
1492 // use (machine word version) atomic operation to install the hash | |
1493 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); | |
1494 if (test == mark) { | |
1495 return hash; | |
1496 } | |
1497 // If atomic operation failed, we must inflate the header | |
1498 // into heavy weight monitor. We could add more code here | |
1499 // for fast path, but it does not worth the complexity. | |
1500 } else if (mark->has_monitor()) { | |
1501 monitor = mark->monitor(); | |
1502 temp = monitor->header(); | |
1503 assert (temp->is_neutral(), "invariant") ; | |
1504 hash = temp->hash(); | |
1505 if (hash) { | |
1506 return hash; | |
1507 } | |
1508 // Skip to the following code to reduce code size | |
1509 } else if (Self->is_lock_owned((address)mark->locker())) { | |
1510 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned | |
1511 assert (temp->is_neutral(), "invariant") ; | |
1512 hash = temp->hash(); // by current thread, check if the displaced | |
1513 if (hash) { // header contains hash code | |
1514 return hash; | |
1515 } | |
1516 // WARNING: | |
1517 // The displaced header is strictly immutable. | |
1518 // It can NOT be changed in ANY cases. So we have | |
1519 // to inflate the header into heavyweight monitor | |
1520 // even the current thread owns the lock. The reason | |
1521 // is the BasicLock (stack slot) will be asynchronously | |
1522 // read by other threads during the inflate() function. | |
1523 // Any change to stack may not propagate to other threads | |
1524 // correctly. | |
1525 } | |
1526 | |
1527 // Inflate the monitor to set hash code | |
1528 monitor = ObjectSynchronizer::inflate(Self, obj); | |
1529 // Load displaced header and check it has hash code | |
1530 mark = monitor->header(); | |
1531 assert (mark->is_neutral(), "invariant") ; | |
1532 hash = mark->hash(); | |
1533 if (hash == 0) { | |
1534 hash = get_next_hash(Self, obj); | |
1535 temp = mark->copy_set_hash(hash); // merge hash code into header | |
1536 assert (temp->is_neutral(), "invariant") ; | |
1537 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); | |
1538 if (test != mark) { | |
1539 // The only update to the header in the monitor (outside GC) | |
1540 // is install the hash code. If someone add new usage of | |
1541 // displaced header, please update this code | |
1542 hash = test->hash(); | |
1543 assert (test->is_neutral(), "invariant") ; | |
1544 assert (hash != 0, "Trivial unexpected object/monitor header usage."); | |
1545 } | |
1546 } | |
1547 // We finally get the hash | |
1548 return hash; | |
1549 } | |
1550 | |
1551 // Deprecated -- use FastHashCode() instead. | |
1552 | |
1553 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { | |
1554 return FastHashCode (Thread::current(), obj()) ; | |
1555 } | |
1556 | |
1557 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, | |
1558 Handle h_obj) { | |
1559 if (UseBiasedLocking) { | |
1560 BiasedLocking::revoke_and_rebias(h_obj, false, thread); | |
1561 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1562 } | |
1563 | |
1564 assert(thread == JavaThread::current(), "Can only be called on current thread"); | |
1565 oop obj = h_obj(); | |
1566 | |
1567 markOop mark = ReadStableMark (obj) ; | |
1568 | |
1569 // Uncontended case, header points to stack | |
1570 if (mark->has_locker()) { | |
1571 return thread->is_lock_owned((address)mark->locker()); | |
1572 } | |
1573 // Contended case, header points to ObjectMonitor (tagged pointer) | |
1574 if (mark->has_monitor()) { | |
1575 ObjectMonitor* monitor = mark->monitor(); | |
1576 return monitor->is_entered(thread) != 0 ; | |
1577 } | |
1578 // Unlocked case, header in place | |
1579 assert(mark->is_neutral(), "sanity check"); | |
1580 return false; | |
1581 } | |
1582 | |
1583 // Be aware of this method could revoke bias of the lock object. | |
1584 // This method querys the ownership of the lock handle specified by 'h_obj'. | |
1585 // If the current thread owns the lock, it returns owner_self. If no | |
1586 // thread owns the lock, it returns owner_none. Otherwise, it will return | |
1587 // ower_other. | |
1588 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership | |
1589 (JavaThread *self, Handle h_obj) { | |
1590 // The caller must beware this method can revoke bias, and | |
1591 // revocation can result in a safepoint. | |
1592 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; | |
1593 assert (self->thread_state() != _thread_blocked , "invariant") ; | |
1594 | |
1595 // Possible mark states: neutral, biased, stack-locked, inflated | |
1596 | |
1597 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { | |
1598 // CASE: biased | |
1599 BiasedLocking::revoke_and_rebias(h_obj, false, self); | |
1600 assert(!h_obj->mark()->has_bias_pattern(), | |
1601 "biases should be revoked by now"); | |
1602 } | |
1603 | |
1604 assert(self == JavaThread::current(), "Can only be called on current thread"); | |
1605 oop obj = h_obj(); | |
1606 markOop mark = ReadStableMark (obj) ; | |
1607 | |
1608 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. | |
1609 if (mark->has_locker()) { | |
1610 return self->is_lock_owned((address)mark->locker()) ? | |
1611 owner_self : owner_other; | |
1612 } | |
1613 | |
1614 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. | |
1615 // The Object:ObjectMonitor relationship is stable as long as we're | |
1616 // not at a safepoint. | |
1617 if (mark->has_monitor()) { | |
1618 void * owner = mark->monitor()->_owner ; | |
1619 if (owner == NULL) return owner_none ; | |
1620 return (owner == self || | |
1621 self->is_lock_owned((address)owner)) ? owner_self : owner_other; | |
1622 } | |
1623 | |
1624 // CASE: neutral | |
1625 assert(mark->is_neutral(), "sanity check"); | |
1626 return owner_none ; // it's unlocked | |
1627 } | |
1628 | |
1629 // FIXME: jvmti should call this | |
1630 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { | |
1631 if (UseBiasedLocking) { | |
1632 if (SafepointSynchronize::is_at_safepoint()) { | |
1633 BiasedLocking::revoke_at_safepoint(h_obj); | |
1634 } else { | |
1635 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); | |
1636 } | |
1637 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); | |
1638 } | |
1639 | |
1640 oop obj = h_obj(); | |
1641 address owner = NULL; | |
1642 | |
1643 markOop mark = ReadStableMark (obj) ; | |
1644 | |
1645 // Uncontended case, header points to stack | |
1646 if (mark->has_locker()) { | |
1647 owner = (address) mark->locker(); | |
1648 } | |
1649 | |
1650 // Contended case, header points to ObjectMonitor (tagged pointer) | |
1651 if (mark->has_monitor()) { | |
1652 ObjectMonitor* monitor = mark->monitor(); | |
1653 assert(monitor != NULL, "monitor should be non-null"); | |
1654 owner = (address) monitor->owner(); | |
1655 } | |
1656 | |
1657 if (owner != NULL) { | |
1658 return Threads::owning_thread_from_monitor_owner(owner, doLock); | |
1659 } | |
1660 | |
1661 // Unlocked case, header in place | |
1662 // Cannot have assertion since this object may have been | |
1663 // locked by another thread when reaching here. | |
1664 // assert(mark->is_neutral(), "sanity check"); | |
1665 | |
1666 return NULL; | |
1667 } | |
1668 | |
1669 // Iterate through monitor cache and attempt to release thread's monitors | |
1670 // Gives up on a particular monitor if an exception occurs, but continues | |
1671 // the overall iteration, swallowing the exception. | |
1672 class ReleaseJavaMonitorsClosure: public MonitorClosure { | |
1673 private: | |
1674 TRAPS; | |
1675 | |
1676 public: | |
1677 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} | |
1678 void do_monitor(ObjectMonitor* mid) { | |
1679 if (mid->owner() == THREAD) { | |
1680 (void)mid->complete_exit(CHECK); | |
1681 } | |
1682 } | |
1683 }; | |
1684 | |
1685 // Release all inflated monitors owned by THREAD. Lightweight monitors are | |
1686 // ignored. This is meant to be called during JNI thread detach which assumes | |
1687 // all remaining monitors are heavyweight. All exceptions are swallowed. | |
1688 // Scanning the extant monitor list can be time consuming. | |
1689 // A simple optimization is to add a per-thread flag that indicates a thread | |
1690 // called jni_monitorenter() during its lifetime. | |
1691 // | |
1692 // Instead of No_Savepoint_Verifier it might be cheaper to | |
1693 // use an idiom of the form: | |
1694 // auto int tmp = SafepointSynchronize::_safepoint_counter ; | |
1695 // <code that must not run at safepoint> | |
1696 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; | |
1697 // Since the tests are extremely cheap we could leave them enabled | |
1698 // for normal product builds. | |
1699 | |
1700 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { | |
1701 assert(THREAD == JavaThread::current(), "must be current Java thread"); | |
1702 No_Safepoint_Verifier nsv ; | |
1703 ReleaseJavaMonitorsClosure rjmc(THREAD); | |
1704 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread"); | |
1705 ObjectSynchronizer::monitors_iterate(&rjmc); | |
1706 Thread::muxRelease(&ListLock); | |
1707 THREAD->clear_pending_exception(); | |
1708 } | |
1709 | |
1710 // Visitors ... | |
1711 | |
1712 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { | |
1713 ObjectMonitor* block = gBlockList; | |
1714 ObjectMonitor* mid; | |
1715 while (block) { | |
1716 assert(block->object() == CHAINMARKER, "must be a block header"); | |
1717 for (int i = _BLOCKSIZE - 1; i > 0; i--) { | |
1718 mid = block + i; | |
1719 oop object = (oop) mid->object(); | |
1720 if (object != NULL) { | |
1721 closure->do_monitor(mid); | |
1722 } | |
1723 } | |
1724 block = (ObjectMonitor*) block->FreeNext; | |
1725 } | |
1726 } | |
1727 | |
1728 void ObjectSynchronizer::oops_do(OopClosure* f) { | |
1729 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); | |
1730 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { | |
1731 assert(block->object() == CHAINMARKER, "must be a block header"); | |
1732 for (int i = 1; i < _BLOCKSIZE; i++) { | |
1733 ObjectMonitor* mid = &block[i]; | |
1734 if (mid->object() != NULL) { | |
1735 f->do_oop((oop*)mid->object_addr()); | |
1736 } | |
1737 } | |
1738 } | |
1739 } | |
1740 | |
1741 // Deflate_idle_monitors() is called at all safepoints, immediately | |
1742 // after all mutators are stopped, but before any objects have moved. | |
1743 // It traverses the list of known monitors, deflating where possible. | |
1744 // The scavenged monitor are returned to the monitor free list. | |
1745 // | |
1746 // Beware that we scavenge at *every* stop-the-world point. | |
1747 // Having a large number of monitors in-circulation negatively | |
1748 // impacts the performance of some applications (e.g., PointBase). | |
1749 // Broadly, we want to minimize the # of monitors in circulation. | |
1750 // Alternately, we could partition the active monitors into sub-lists | |
1751 // of those that need scanning and those that do not. | |
1752 // Specifically, we would add a new sub-list of objectmonitors | |
1753 // that are in-circulation and potentially active. deflate_idle_monitors() | |
1754 // would scan only that list. Other monitors could reside on a quiescent | |
1755 // list. Such sequestered monitors wouldn't need to be scanned by | |
1756 // deflate_idle_monitors(). omAlloc() would first check the global free list, | |
1757 // then the quiescent list, and, failing those, would allocate a new block. | |
1758 // Deflate_idle_monitors() would scavenge and move monitors to the | |
1759 // quiescent list. | |
1760 // | |
1761 // Perversely, the heap size -- and thus the STW safepoint rate -- | |
1762 // typically drives the scavenge rate. Large heaps can mean infrequent GC, | |
1763 // which in turn can mean large(r) numbers of objectmonitors in circulation. | |
1764 // This is an unfortunate aspect of this design. | |
1765 // | |
1766 // Another refinement would be to refrain from calling deflate_idle_monitors() | |
1767 // except at stop-the-world points associated with garbage collections. | |
1768 // | |
1769 // An even better solution would be to deflate on-the-fly, aggressively, | |
1770 // at monitorexit-time as is done in EVM's metalock or Relaxed Locks. | |
1771 | |
1772 void ObjectSynchronizer::deflate_idle_monitors() { | |
1773 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); | |
1774 int nInuse = 0 ; // currently associated with objects | |
1775 int nInCirculation = 0 ; // extant | |
1776 int nScavenged = 0 ; // reclaimed | |
1777 | |
1778 ObjectMonitor * FreeHead = NULL ; // Local SLL of scavenged monitors | |
1779 ObjectMonitor * FreeTail = NULL ; | |
1780 | |
1781 // Iterate over all extant monitors - Scavenge all idle monitors. | |
1782 TEVENT (deflate_idle_monitors) ; | |
1783 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { | |
1784 assert(block->object() == CHAINMARKER, "must be a block header"); | |
1785 nInCirculation += _BLOCKSIZE ; | |
1786 for (int i = 1 ; i < _BLOCKSIZE; i++) { | |
1787 ObjectMonitor* mid = &block[i]; | |
1788 oop obj = (oop) mid->object(); | |
1789 | |
1790 if (obj == NULL) { | |
1791 // The monitor is not associated with an object. | |
1792 // The monitor should either be a thread-specific private | |
1793 // free list or the global free list. | |
1794 // obj == NULL IMPLIES mid->is_busy() == 0 | |
1795 guarantee (!mid->is_busy(), "invariant") ; | |
1796 continue ; | |
1797 } | |
1798 | |
1799 // Normal case ... The monitor is associated with obj. | |
1800 guarantee (obj->mark() == markOopDesc::encode(mid), "invariant") ; | |
1801 guarantee (mid == obj->mark()->monitor(), "invariant"); | |
1802 guarantee (mid->header()->is_neutral(), "invariant"); | |
1803 | |
1804 if (mid->is_busy()) { | |
1805 if (ClearResponsibleAtSTW) mid->_Responsible = NULL ; | |
1806 nInuse ++ ; | |
1807 } else { | |
1808 // Deflate the monitor if it is no longer being used | |
1809 // It's idle - scavenge and return to the global free list | |
1810 // plain old deflation ... | |
1811 TEVENT (deflate_idle_monitors - scavenge1) ; | |
1812 if (TraceMonitorInflation) { | |
1813 if (obj->is_instance()) { | |
1814 ResourceMark rm; | |
1815 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", | |
1816 (intptr_t) obj, (intptr_t) obj->mark(), Klass::cast(obj->klass())->external_name()); | |
1817 } | |
1818 } | |
1819 | |
1820 // Restore the header back to obj | |
1821 obj->release_set_mark(mid->header()); | |
1822 mid->clear(); | |
1823 | |
1824 assert (mid->object() == NULL, "invariant") ; | |
1825 | |
1826 // Move the object to the working free list defined by FreeHead,FreeTail. | |
1827 mid->FreeNext = NULL ; | |
1828 if (FreeHead == NULL) FreeHead = mid ; | |
1829 if (FreeTail != NULL) FreeTail->FreeNext = mid ; | |
1830 FreeTail = mid ; | |
1831 nScavenged ++ ; | |
1832 } | |
1833 } | |
1834 } | |
1835 | |
1836 // Move the scavenged monitors back to the global free list. | |
1837 // In theory we don't need the freelist lock as we're at a STW safepoint. | |
1838 // omAlloc() and omFree() can only be called while a thread is _not in safepoint state. | |
1839 // But it's remotely possible that omFlush() or release_monitors_owned_by_thread() | |
1840 // might be called while not at a global STW safepoint. In the interest of | |
1841 // safety we protect the following access with ListLock. | |
1842 // An even more conservative and prudent approach would be to guard | |
1843 // the main loop in scavenge_idle_monitors() with ListLock. | |
1844 if (FreeHead != NULL) { | |
1845 guarantee (FreeTail != NULL && nScavenged > 0, "invariant") ; | |
1846 assert (FreeTail->FreeNext == NULL, "invariant") ; | |
1847 // constant-time list splice - prepend scavenged segment to gFreeList | |
1848 Thread::muxAcquire (&ListLock, "scavenge - return") ; | |
1849 FreeTail->FreeNext = gFreeList ; | |
1850 gFreeList = FreeHead ; | |
1851 Thread::muxRelease (&ListLock) ; | |
1852 } | |
1853 | |
1854 if (_sync_Deflations != NULL) _sync_Deflations->inc(nScavenged) ; | |
1855 if (_sync_MonExtant != NULL) _sync_MonExtant ->set_value(nInCirculation); | |
1856 | |
1857 // TODO: Add objectMonitor leak detection. | |
1858 // Audit/inventory the objectMonitors -- make sure they're all accounted for. | |
1859 GVars.stwRandom = os::random() ; | |
1860 GVars.stwCycle ++ ; | |
1861 } | |
1862 | |
1863 // A macro is used below because there may already be a pending | |
1864 // exception which should not abort the execution of the routines | |
1865 // which use this (which is why we don't put this into check_slow and | |
1866 // call it with a CHECK argument). | |
1867 | |
1868 #define CHECK_OWNER() \ | |
1869 do { \ | |
1870 if (THREAD != _owner) { \ | |
1871 if (THREAD->is_lock_owned((address) _owner)) { \ | |
1872 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ | |
1873 _recursions = 0; \ | |
1874 OwnerIsThread = 1 ; \ | |
1875 } else { \ | |
1876 TEVENT (Throw IMSX) ; \ | |
1877 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ | |
1878 } \ | |
1879 } \ | |
1880 } while (false) | |
1881 | |
1882 // TODO-FIXME: eliminate ObjectWaiters. Replace this visitor/enumerator | |
1883 // interface with a simple FirstWaitingThread(), NextWaitingThread() interface. | |
1884 | |
1885 ObjectWaiter* ObjectMonitor::first_waiter() { | |
1886 return _WaitSet; | |
1887 } | |
1888 | |
1889 ObjectWaiter* ObjectMonitor::next_waiter(ObjectWaiter* o) { | |
1890 return o->_next; | |
1891 } | |
1892 | |
1893 Thread* ObjectMonitor::thread_of_waiter(ObjectWaiter* o) { | |
1894 return o->_thread; | |
1895 } | |
1896 | |
1897 // initialize the monitor, exception the semaphore, all other fields | |
1898 // are simple integers or pointers | |
1899 ObjectMonitor::ObjectMonitor() { | |
1900 _header = NULL; | |
1901 _count = 0; | |
1902 _waiters = 0, | |
1903 _recursions = 0; | |
1904 _object = NULL; | |
1905 _owner = NULL; | |
1906 _WaitSet = NULL; | |
1907 _WaitSetLock = 0 ; | |
1908 _Responsible = NULL ; | |
1909 _succ = NULL ; | |
1910 _cxq = NULL ; | |
1911 FreeNext = NULL ; | |
1912 _EntryList = NULL ; | |
1913 _SpinFreq = 0 ; | |
1914 _SpinClock = 0 ; | |
1915 OwnerIsThread = 0 ; | |
1916 } | |
1917 | |
1918 ObjectMonitor::~ObjectMonitor() { | |
1919 // TODO: Add asserts ... | |
1920 // _cxq == 0 _succ == NULL _owner == NULL _waiters == 0 | |
1921 // _count == 0 _EntryList == NULL etc | |
1922 } | |
1923 | |
1924 intptr_t ObjectMonitor::is_busy() const { | |
1925 // TODO-FIXME: merge _count and _waiters. | |
1926 // TODO-FIXME: assert _owner == null implies _recursions = 0 | |
1927 // TODO-FIXME: assert _WaitSet != null implies _count > 0 | |
1928 return _count|_waiters|intptr_t(_owner)|intptr_t(_cxq)|intptr_t(_EntryList ) ; | |
1929 } | |
1930 | |
1931 void ObjectMonitor::Recycle () { | |
1932 // TODO: add stronger asserts ... | |
1933 // _cxq == 0 _succ == NULL _owner == NULL _waiters == 0 | |
1934 // _count == 0 EntryList == NULL | |
1935 // _recursions == 0 _WaitSet == NULL | |
1936 // TODO: assert (is_busy()|_recursions) == 0 | |
1937 _succ = NULL ; | |
1938 _EntryList = NULL ; | |
1939 _cxq = NULL ; | |
1940 _WaitSet = NULL ; | |
1941 _recursions = 0 ; | |
1942 _SpinFreq = 0 ; | |
1943 _SpinClock = 0 ; | |
1944 OwnerIsThread = 0 ; | |
1945 } | |
1946 | |
1947 // WaitSet management ... | |
1948 | |
1949 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { | |
1950 assert(node != NULL, "should not dequeue NULL node"); | |
1951 assert(node->_prev == NULL, "node already in list"); | |
1952 assert(node->_next == NULL, "node already in list"); | |
1953 // put node at end of queue (circular doubly linked list) | |
1954 if (_WaitSet == NULL) { | |
1955 _WaitSet = node; | |
1956 node->_prev = node; | |
1957 node->_next = node; | |
1958 } else { | |
1959 ObjectWaiter* head = _WaitSet ; | |
1960 ObjectWaiter* tail = head->_prev; | |
1961 assert(tail->_next == head, "invariant check"); | |
1962 tail->_next = node; | |
1963 head->_prev = node; | |
1964 node->_next = head; | |
1965 node->_prev = tail; | |
1966 } | |
1967 } | |
1968 | |
1969 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { | |
1970 // dequeue the very first waiter | |
1971 ObjectWaiter* waiter = _WaitSet; | |
1972 if (waiter) { | |
1973 DequeueSpecificWaiter(waiter); | |
1974 } | |
1975 return waiter; | |
1976 } | |
1977 | |
1978 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { | |
1979 assert(node != NULL, "should not dequeue NULL node"); | |
1980 assert(node->_prev != NULL, "node already removed from list"); | |
1981 assert(node->_next != NULL, "node already removed from list"); | |
1982 // when the waiter has woken up because of interrupt, | |
1983 // timeout or other spurious wake-up, dequeue the | |
1984 // waiter from waiting list | |
1985 ObjectWaiter* next = node->_next; | |
1986 if (next == node) { | |
1987 assert(node->_prev == node, "invariant check"); | |
1988 _WaitSet = NULL; | |
1989 } else { | |
1990 ObjectWaiter* prev = node->_prev; | |
1991 assert(prev->_next == node, "invariant check"); | |
1992 assert(next->_prev == node, "invariant check"); | |
1993 next->_prev = prev; | |
1994 prev->_next = next; | |
1995 if (_WaitSet == node) { | |
1996 _WaitSet = next; | |
1997 } | |
1998 } | |
1999 node->_next = NULL; | |
2000 node->_prev = NULL; | |
2001 } | |
2002 | |
2003 static char * kvGet (char * kvList, const char * Key) { | |
2004 if (kvList == NULL) return NULL ; | |
2005 size_t n = strlen (Key) ; | |
2006 char * Search ; | |
2007 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { | |
2008 if (strncmp (Search, Key, n) == 0) { | |
2009 if (Search[n] == '=') return Search + n + 1 ; | |
2010 if (Search[n] == 0) return (char *) "1" ; | |
2011 } | |
2012 } | |
2013 return NULL ; | |
2014 } | |
2015 | |
2016 static int kvGetInt (char * kvList, const char * Key, int Default) { | |
2017 char * v = kvGet (kvList, Key) ; | |
2018 int rslt = v ? ::strtol (v, NULL, 0) : Default ; | |
2019 if (Knob_ReportSettings && v != NULL) { | |
2020 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; | |
2021 ::fflush (stdout) ; | |
2022 } | |
2023 return rslt ; | |
2024 } | |
2025 | |
2026 // By convention we unlink a contending thread from EntryList|cxq immediately | |
2027 // after the thread acquires the lock in ::enter(). Equally, we could defer | |
2028 // unlinking the thread until ::exit()-time. | |
2029 | |
2030 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) | |
2031 { | |
2032 assert (_owner == Self, "invariant") ; | |
2033 assert (SelfNode->_thread == Self, "invariant") ; | |
2034 | |
2035 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { | |
2036 // Normal case: remove Self from the DLL EntryList . | |
2037 // This is a constant-time operation. | |
2038 ObjectWaiter * nxt = SelfNode->_next ; | |
2039 ObjectWaiter * prv = SelfNode->_prev ; | |
2040 if (nxt != NULL) nxt->_prev = prv ; | |
2041 if (prv != NULL) prv->_next = nxt ; | |
2042 if (SelfNode == _EntryList ) _EntryList = nxt ; | |
2043 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
2044 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
2045 TEVENT (Unlink from EntryList) ; | |
2046 } else { | |
2047 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; | |
2048 // Inopportune interleaving -- Self is still on the cxq. | |
2049 // This usually means the enqueue of self raced an exiting thread. | |
2050 // Normally we'll find Self near the front of the cxq, so | |
2051 // dequeueing is typically fast. If needbe we can accelerate | |
2052 // this with some MCS/CHL-like bidirectional list hints and advisory | |
2053 // back-links so dequeueing from the interior will normally operate | |
2054 // in constant-time. | |
2055 // Dequeue Self from either the head (with CAS) or from the interior | |
2056 // with a linear-time scan and normal non-atomic memory operations. | |
2057 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList | |
2058 // and then unlink Self from EntryList. We have to drain eventually, | |
2059 // so it might as well be now. | |
2060 | |
2061 ObjectWaiter * v = _cxq ; | |
2062 assert (v != NULL, "invariant") ; | |
2063 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { | |
2064 // The CAS above can fail from interference IFF a "RAT" arrived. | |
2065 // In that case Self must be in the interior and can no longer be | |
2066 // at the head of cxq. | |
2067 if (v == SelfNode) { | |
2068 assert (_cxq != v, "invariant") ; | |
2069 v = _cxq ; // CAS above failed - start scan at head of list | |
2070 } | |
2071 ObjectWaiter * p ; | |
2072 ObjectWaiter * q = NULL ; | |
2073 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { | |
2074 q = p ; | |
2075 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; | |
2076 } | |
2077 assert (v != SelfNode, "invariant") ; | |
2078 assert (p == SelfNode, "Node not found on cxq") ; | |
2079 assert (p != _cxq, "invariant") ; | |
2080 assert (q != NULL, "invariant") ; | |
2081 assert (q->_next == p, "invariant") ; | |
2082 q->_next = p->_next ; | |
2083 } | |
2084 TEVENT (Unlink from cxq) ; | |
2085 } | |
2086 | |
2087 // Diagnostic hygiene ... | |
2088 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; | |
2089 SelfNode->_next = (ObjectWaiter *) 0xBAD ; | |
2090 SelfNode->TState = ObjectWaiter::TS_RUN ; | |
2091 } | |
2092 | |
2093 // Caveat: TryLock() is not necessarily serializing if it returns failure. | |
2094 // Callers must compensate as needed. | |
2095 | |
2096 int ObjectMonitor::TryLock (Thread * Self) { | |
2097 for (;;) { | |
2098 void * own = _owner ; | |
2099 if (own != NULL) return 0 ; | |
2100 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { | |
2101 // Either guarantee _recursions == 0 or set _recursions = 0. | |
2102 assert (_recursions == 0, "invariant") ; | |
2103 assert (_owner == Self, "invariant") ; | |
2104 // CONSIDER: set or assert that OwnerIsThread == 1 | |
2105 return 1 ; | |
2106 } | |
2107 // The lock had been free momentarily, but we lost the race to the lock. | |
2108 // Interference -- the CAS failed. | |
2109 // We can either return -1 or retry. | |
2110 // Retry doesn't make as much sense because the lock was just acquired. | |
2111 if (true) return -1 ; | |
2112 } | |
2113 } | |
2114 | |
2115 // NotRunnable() -- informed spinning | |
2116 // | |
2117 // Don't bother spinning if the owner is not eligible to drop the lock. | |
2118 // Peek at the owner's schedctl.sc_state and Thread._thread_values and | |
2119 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. | |
2120 // The thread must be runnable in order to drop the lock in timely fashion. | |
2121 // If the _owner is not runnable then spinning will not likely be | |
2122 // successful (profitable). | |
2123 // | |
2124 // Beware -- the thread referenced by _owner could have died | |
2125 // so a simply fetch from _owner->_thread_state might trap. | |
2126 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. | |
2127 // Because of the lifecycle issues the schedctl and _thread_state values | |
2128 // observed by NotRunnable() might be garbage. NotRunnable must | |
2129 // tolerate this and consider the observed _thread_state value | |
2130 // as advisory. | |
2131 // | |
2132 // Beware too, that _owner is sometimes a BasicLock address and sometimes | |
2133 // a thread pointer. We differentiate the two cases with OwnerIsThread. | |
2134 // Alternately, we might tag the type (thread pointer vs basiclock pointer) | |
2135 // with the LSB of _owner. Another option would be to probablistically probe | |
2136 // the putative _owner->TypeTag value. | |
2137 // | |
2138 // Checking _thread_state isn't perfect. Even if the thread is | |
2139 // in_java it might be blocked on a page-fault or have been preempted | |
2140 // and sitting on a ready/dispatch queue. _thread state in conjunction | |
2141 // with schedctl.sc_state gives us a good picture of what the | |
2142 // thread is doing, however. | |
2143 // | |
2144 // TODO: check schedctl.sc_state. | |
2145 // We'll need to use SafeFetch32() to read from the schedctl block. | |
2146 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ | |
2147 // | |
2148 // The return value from NotRunnable() is *advisory* -- the | |
2149 // result is based on sampling and is not necessarily coherent. | |
2150 // The caller must tolerate false-negative and false-positive errors. | |
2151 // Spinning, in general, is probabilistic anyway. | |
2152 | |
2153 | |
2154 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { | |
2155 // Check either OwnerIsThread or ox->TypeTag == 2BAD. | |
2156 if (!OwnerIsThread) return 0 ; | |
2157 | |
2158 if (ox == NULL) return 0 ; | |
2159 | |
2160 // Avoid transitive spinning ... | |
2161 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. | |
2162 // Immediately after T1 acquires L it's possible that T2, also | |
2163 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. | |
2164 // This occurs transiently after T1 acquired L but before | |
2165 // T1 managed to clear T1.Stalled. T2 does not need to abort | |
2166 // its spin in this circumstance. | |
2167 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; | |
2168 | |
2169 if (BlockedOn == 1) return 1 ; | |
2170 if (BlockedOn != 0) { | |
2171 return BlockedOn != intptr_t(this) && _owner == ox ; | |
2172 } | |
2173 | |
2174 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; | |
2175 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; | |
2176 // consider also: jst != _thread_in_Java -- but that's overspecific. | |
2177 return jst == _thread_blocked || jst == _thread_in_native ; | |
2178 } | |
2179 | |
2180 | |
2181 // Adaptive spin-then-block - rational spinning | |
2182 // | |
2183 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS | |
2184 // algorithm. On high order SMP systems it would be better to start with | |
2185 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, | |
2186 // a contending thread could enqueue itself on the cxq and then spin locally | |
2187 // on a thread-specific variable such as its ParkEvent._Event flag. | |
2188 // That's left as an exercise for the reader. Note that global spinning is | |
2189 // not problematic on Niagara, as the L2$ serves the interconnect and has both | |
2190 // low latency and massive bandwidth. | |
2191 // | |
2192 // Broadly, we can fix the spin frequency -- that is, the % of contended lock | |
2193 // acquisition attempts where we opt to spin -- at 100% and vary the spin count | |
2194 // (duration) or we can fix the count at approximately the duration of | |
2195 // a context switch and vary the frequency. Of course we could also | |
2196 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. | |
2197 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. | |
2198 // | |
2199 // This implementation varies the duration "D", where D varies with | |
2200 // the success rate of recent spin attempts. (D is capped at approximately | |
2201 // length of a round-trip context switch). The success rate for recent | |
2202 // spin attempts is a good predictor of the success rate of future spin | |
2203 // attempts. The mechanism adapts automatically to varying critical | |
2204 // section length (lock modality), system load and degree of parallelism. | |
2205 // D is maintained per-monitor in _SpinDuration and is initialized | |
2206 // optimistically. Spin frequency is fixed at 100%. | |
2207 // | |
2208 // Note that _SpinDuration is volatile, but we update it without locks | |
2209 // or atomics. The code is designed so that _SpinDuration stays within | |
2210 // a reasonable range even in the presence of races. The arithmetic | |
2211 // operations on _SpinDuration are closed over the domain of legal values, | |
2212 // so at worst a race will install and older but still legal value. | |
2213 // At the very worst this introduces some apparent non-determinism. | |
2214 // We might spin when we shouldn't or vice-versa, but since the spin | |
2215 // count are relatively short, even in the worst case, the effect is harmless. | |
2216 // | |
2217 // Care must be taken that a low "D" value does not become an | |
2218 // an absorbing state. Transient spinning failures -- when spinning | |
2219 // is overall profitable -- should not cause the system to converge | |
2220 // on low "D" values. We want spinning to be stable and predictable | |
2221 // and fairly responsive to change and at the same time we don't want | |
2222 // it to oscillate, become metastable, be "too" non-deterministic, | |
2223 // or converge on or enter undesirable stable absorbing states. | |
2224 // | |
2225 // We implement a feedback-based control system -- using past behavior | |
2226 // to predict future behavior. We face two issues: (a) if the | |
2227 // input signal is random then the spin predictor won't provide optimal | |
2228 // results, and (b) if the signal frequency is too high then the control | |
2229 // system, which has some natural response lag, will "chase" the signal. | |
2230 // (b) can arise from multimodal lock hold times. Transient preemption | |
2231 // can also result in apparent bimodal lock hold times. | |
2232 // Although sub-optimal, neither condition is particularly harmful, as | |
2233 // in the worst-case we'll spin when we shouldn't or vice-versa. | |
2234 // The maximum spin duration is rather short so the failure modes aren't bad. | |
2235 // To be conservative, I've tuned the gain in system to bias toward | |
2236 // _not spinning. Relatedly, the system can sometimes enter a mode where it | |
2237 // "rings" or oscillates between spinning and not spinning. This happens | |
2238 // when spinning is just on the cusp of profitability, however, so the | |
2239 // situation is not dire. The state is benign -- there's no need to add | |
2240 // hysteresis control to damp the transition rate between spinning and | |
2241 // not spinning. | |
2242 // | |
2243 // - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - | |
2244 // | |
2245 // Spin-then-block strategies ... | |
2246 // | |
2247 // Thoughts on ways to improve spinning : | |
2248 // | |
2249 // * Periodically call {psr_}getloadavg() while spinning, and | |
2250 // permit unbounded spinning if the load average is < | |
2251 // the number of processors. Beware, however, that getloadavg() | |
2252 // is exceptionally fast on solaris (about 1/10 the cost of a full | |
2253 // spin cycle, but quite expensive on linux. Beware also, that | |
2254 // multiple JVMs could "ring" or oscillate in a feedback loop. | |
2255 // Sufficient damping would solve that problem. | |
2256 // | |
2257 // * We currently use spin loops with iteration counters to approximate | |
2258 // spinning for some interval. Given the availability of high-precision | |
2259 // time sources such as gethrtime(), %TICK, %STICK, RDTSC, etc., we should | |
2260 // someday reimplement the spin loops to duration-based instead of iteration-based. | |
2261 // | |
2262 // * Don't spin if there are more than N = (CPUs/2) threads | |
2263 // currently spinning on the monitor (or globally). | |
2264 // That is, limit the number of concurrent spinners. | |
2265 // We might also limit the # of spinners in the JVM, globally. | |
2266 // | |
2267 // * If a spinning thread observes _owner change hands it should | |
2268 // abort the spin (and park immediately) or at least debit | |
2269 // the spin counter by a large "penalty". | |
2270 // | |
2271 // * Classically, the spin count is either K*(CPUs-1) or is a | |
2272 // simple constant that approximates the length of a context switch. | |
2273 // We currently use a value -- computed by a special utility -- that | |
2274 // approximates round-trip context switch times. | |
2275 // | |
2276 // * Normally schedctl_start()/_stop() is used to advise the kernel | |
2277 // to avoid preempting threads that are running in short, bounded | |
2278 // critical sections. We could use the schedctl hooks in an inverted | |
2279 // sense -- spinners would set the nopreempt flag, but poll the preempt | |
2280 // pending flag. If a spinner observed a pending preemption it'd immediately | |
2281 // abort the spin and park. As such, the schedctl service acts as | |
2282 // a preemption warning mechanism. | |
2283 // | |
2284 // * In lieu of spinning, if the system is running below saturation | |
2285 // (that is, loadavg() << #cpus), we can instead suppress futile | |
2286 // wakeup throttling, or even wake more than one successor at exit-time. | |
2287 // The net effect is largely equivalent to spinning. In both cases, | |
2288 // contending threads go ONPROC and opportunistically attempt to acquire | |
2289 // the lock, decreasing lock handover latency at the expense of wasted | |
2290 // cycles and context switching. | |
2291 // | |
2292 // * We might to spin less after we've parked as the thread will | |
2293 // have less $ and TLB affinity with the processor. | |
2294 // Likewise, we might spin less if we come ONPROC on a different | |
2295 // processor or after a long period (>> rechose_interval). | |
2296 // | |
2297 // * A table-driven state machine similar to Solaris' dispadmin scheduling | |
2298 // tables might be a better design. Instead of encoding information in | |
2299 // _SpinDuration, _SpinFreq and _SpinClock we'd just use explicit, | |
2300 // discrete states. Success or failure during a spin would drive | |
2301 // state transitions, and each state node would contain a spin count. | |
2302 // | |
2303 // * If the processor is operating in a mode intended to conserve power | |
2304 // (such as Intel's SpeedStep) or to reduce thermal output (thermal | |
2305 // step-down mode) then the Java synchronization subsystem should | |
2306 // forgo spinning. | |
2307 // | |
2308 // * The minimum spin duration should be approximately the worst-case | |
2309 // store propagation latency on the platform. That is, the time | |
2310 // it takes a store on CPU A to become visible on CPU B, where A and | |
2311 // B are "distant". | |
2312 // | |
2313 // * We might want to factor a thread's priority in the spin policy. | |
2314 // Threads with a higher priority might spin for slightly longer. | |
2315 // Similarly, if we use back-off in the TATAS loop, lower priority | |
2316 // threads might back-off longer. We don't currently use a | |
2317 // thread's priority when placing it on the entry queue. We may | |
2318 // want to consider doing so in future releases. | |
2319 // | |
2320 // * We might transiently drop a thread's scheduling priority while it spins. | |
2321 // SCHED_BATCH on linux and FX scheduling class at priority=0 on Solaris | |
2322 // would suffice. We could even consider letting the thread spin indefinitely at | |
2323 // a depressed or "idle" priority. This brings up fairness issues, however -- | |
2324 // in a saturated system a thread would with a reduced priority could languish | |
2325 // for extended periods on the ready queue. | |
2326 // | |
2327 // * While spinning try to use the otherwise wasted time to help the VM make | |
2328 // progress: | |
2329 // | |
2330 // -- YieldTo() the owner, if the owner is OFFPROC but ready | |
2331 // Done our remaining quantum directly to the ready thread. | |
2332 // This helps "push" the lock owner through the critical section. | |
2333 // It also tends to improve affinity/locality as the lock | |
2334 // "migrates" less frequently between CPUs. | |
2335 // -- Walk our own stack in anticipation of blocking. Memoize the roots. | |
2336 // -- Perform strand checking for other thread. Unpark potential strandees. | |
2337 // -- Help GC: trace or mark -- this would need to be a bounded unit of work. | |
2338 // Unfortunately this will pollute our $ and TLBs. Recall that we | |
2339 // spin to avoid context switching -- context switching has an | |
2340 // immediate cost in latency, a disruptive cost to other strands on a CMT | |
2341 // processor, and an amortized cost because of the D$ and TLB cache | |
2342 // reload transient when the thread comes back ONPROC and repopulates | |
2343 // $s and TLBs. | |
2344 // -- call getloadavg() to see if the system is saturated. It'd probably | |
2345 // make sense to call getloadavg() half way through the spin. | |
2346 // If the system isn't at full capacity the we'd simply reset | |
2347 // the spin counter to and extend the spin attempt. | |
2348 // -- Doug points out that we should use the same "helping" policy | |
2349 // in thread.yield(). | |
2350 // | |
2351 // * Try MONITOR-MWAIT on systems that support those instructions. | |
2352 // | |
2353 // * The spin statistics that drive spin decisions & frequency are | |
2354 // maintained in the objectmonitor structure so if we deflate and reinflate | |
2355 // we lose spin state. In practice this is not usually a concern | |
2356 // as the default spin state after inflation is aggressive (optimistic) | |
2357 // and tends toward spinning. So in the worst case for a lock where | |
2358 // spinning is not profitable we may spin unnecessarily for a brief | |
2359 // period. But then again, if a lock is contended it'll tend not to deflate | |
2360 // in the first place. | |
2361 | |
2362 | |
2363 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; | |
2364 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; | |
2365 | |
2366 // Spinning: Fixed frequency (100%), vary duration | |
2367 | |
2368 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { | |
2369 | |
2370 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. | |
2371 int ctr = Knob_FixedSpin ; | |
2372 if (ctr != 0) { | |
2373 while (--ctr >= 0) { | |
2374 if (TryLock (Self) > 0) return 1 ; | |
2375 SpinPause () ; | |
2376 } | |
2377 return 0 ; | |
2378 } | |
2379 | |
2380 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { | |
2381 if (TryLock(Self) > 0) { | |
2382 // Increase _SpinDuration ... | |
2383 // Note that we don't clamp SpinDuration precisely at SpinLimit. | |
2384 // Raising _SpurDuration to the poverty line is key. | |
2385 int x = _SpinDuration ; | |
2386 if (x < Knob_SpinLimit) { | |
2387 if (x < Knob_Poverty) x = Knob_Poverty ; | |
2388 _SpinDuration = x + Knob_BonusB ; | |
2389 } | |
2390 return 1 ; | |
2391 } | |
2392 SpinPause () ; | |
2393 } | |
2394 | |
2395 // Admission control - verify preconditions for spinning | |
2396 // | |
2397 // We always spin a little bit, just to prevent _SpinDuration == 0 from | |
2398 // becoming an absorbing state. Put another way, we spin briefly to | |
2399 // sample, just in case the system load, parallelism, contention, or lock | |
2400 // modality changed. | |
2401 // | |
2402 // Consider the following alternative: | |
2403 // Periodically set _SpinDuration = _SpinLimit and try a long/full | |
2404 // spin attempt. "Periodically" might mean after a tally of | |
2405 // the # of failed spin attempts (or iterations) reaches some threshold. | |
2406 // This takes us into the realm of 1-out-of-N spinning, where we | |
2407 // hold the duration constant but vary the frequency. | |
2408 | |
2409 ctr = _SpinDuration ; | |
2410 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; | |
2411 if (ctr <= 0) return 0 ; | |
2412 | |
2413 if (Knob_SuccRestrict && _succ != NULL) return 0 ; | |
2414 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { | |
2415 TEVENT (Spin abort - notrunnable [TOP]); | |
2416 return 0 ; | |
2417 } | |
2418 | |
2419 int MaxSpin = Knob_MaxSpinners ; | |
2420 if (MaxSpin >= 0) { | |
2421 if (_Spinner > MaxSpin) { | |
2422 TEVENT (Spin abort -- too many spinners) ; | |
2423 return 0 ; | |
2424 } | |
2425 // Slighty racy, but benign ... | |
2426 Adjust (&_Spinner, 1) ; | |
2427 } | |
2428 | |
2429 // We're good to spin ... spin ingress. | |
2430 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades | |
2431 // when preparing to LD...CAS _owner, etc and the CAS is likely | |
2432 // to succeed. | |
2433 int hits = 0 ; | |
2434 int msk = 0 ; | |
2435 int caspty = Knob_CASPenalty ; | |
2436 int oxpty = Knob_OXPenalty ; | |
2437 int sss = Knob_SpinSetSucc ; | |
2438 if (sss && _succ == NULL ) _succ = Self ; | |
2439 Thread * prv = NULL ; | |
2440 | |
2441 // There are three ways to exit the following loop: | |
2442 // 1. A successful spin where this thread has acquired the lock. | |
2443 // 2. Spin failure with prejudice | |
2444 // 3. Spin failure without prejudice | |
2445 | |
2446 while (--ctr >= 0) { | |
2447 | |
2448 // Periodic polling -- Check for pending GC | |
2449 // Threads may spin while they're unsafe. | |
2450 // We don't want spinning threads to delay the JVM from reaching | |
2451 // a stop-the-world safepoint or to steal cycles from GC. | |
2452 // If we detect a pending safepoint we abort in order that | |
2453 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) | |
2454 // this thread, if safe, doesn't steal cycles from GC. | |
2455 // This is in keeping with the "no loitering in runtime" rule. | |
2456 // We periodically check to see if there's a safepoint pending. | |
2457 if ((ctr & 0xFF) == 0) { | |
2458 if (SafepointSynchronize::do_call_back()) { | |
2459 TEVENT (Spin: safepoint) ; | |
2460 goto Abort ; // abrupt spin egress | |
2461 } | |
2462 if (Knob_UsePause & 1) SpinPause () ; | |
2463 | |
2464 int (*scb)(intptr_t,int) = SpinCallbackFunction ; | |
2465 if (hits > 50 && scb != NULL) { | |
2466 int abend = (*scb)(SpinCallbackArgument, 0) ; | |
2467 } | |
2468 } | |
2469 | |
2470 if (Knob_UsePause & 2) SpinPause() ; | |
2471 | |
2472 // Exponential back-off ... Stay off the bus to reduce coherency traffic. | |
2473 // This is useful on classic SMP systems, but is of less utility on | |
2474 // N1-style CMT platforms. | |
2475 // | |
2476 // Trade-off: lock acquisition latency vs coherency bandwidth. | |
2477 // Lock hold times are typically short. A histogram | |
2478 // of successful spin attempts shows that we usually acquire | |
2479 // the lock early in the spin. That suggests we want to | |
2480 // sample _owner frequently in the early phase of the spin, | |
2481 // but then back-off and sample less frequently as the spin | |
2482 // progresses. The back-off makes a good citizen on SMP big | |
2483 // SMP systems. Oversampling _owner can consume excessive | |
2484 // coherency bandwidth. Relatedly, if we _oversample _owner we | |
2485 // can inadvertently interfere with the the ST m->owner=null. | |
2486 // executed by the lock owner. | |
2487 if (ctr & msk) continue ; | |
2488 ++hits ; | |
2489 if ((hits & 0xF) == 0) { | |
2490 // The 0xF, above, corresponds to the exponent. | |
2491 // Consider: (msk+1)|msk | |
2492 msk = ((msk << 2)|3) & BackOffMask ; | |
2493 } | |
2494 | |
2495 // Probe _owner with TATAS | |
2496 // If this thread observes the monitor transition or flicker | |
2497 // from locked to unlocked to locked, then the odds that this | |
2498 // thread will acquire the lock in this spin attempt go down | |
2499 // considerably. The same argument applies if the CAS fails | |
2500 // or if we observe _owner change from one non-null value to | |
2501 // another non-null value. In such cases we might abort | |
2502 // the spin without prejudice or apply a "penalty" to the | |
2503 // spin count-down variable "ctr", reducing it by 100, say. | |
2504 | |
2505 Thread * ox = (Thread *) _owner ; | |
2506 if (ox == NULL) { | |
2507 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; | |
2508 if (ox == NULL) { | |
2509 // The CAS succeeded -- this thread acquired ownership | |
2510 // Take care of some bookkeeping to exit spin state. | |
2511 if (sss && _succ == Self) { | |
2512 _succ = NULL ; | |
2513 } | |
2514 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; | |
2515 | |
2516 // Increase _SpinDuration : | |
2517 // The spin was successful (profitable) so we tend toward | |
2518 // longer spin attempts in the future. | |
2519 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. | |
2520 // If we acquired the lock early in the spin cycle it | |
2521 // makes sense to increase _SpinDuration proportionally. | |
2522 // Note that we don't clamp SpinDuration precisely at SpinLimit. | |
2523 int x = _SpinDuration ; | |
2524 if (x < Knob_SpinLimit) { | |
2525 if (x < Knob_Poverty) x = Knob_Poverty ; | |
2526 _SpinDuration = x + Knob_Bonus ; | |
2527 } | |
2528 return 1 ; | |
2529 } | |
2530 | |
2531 // The CAS failed ... we can take any of the following actions: | |
2532 // * penalize: ctr -= Knob_CASPenalty | |
2533 // * exit spin with prejudice -- goto Abort; | |
2534 // * exit spin without prejudice. | |
2535 // * Since CAS is high-latency, retry again immediately. | |
2536 prv = ox ; | |
2537 TEVENT (Spin: cas failed) ; | |
2538 if (caspty == -2) break ; | |
2539 if (caspty == -1) goto Abort ; | |
2540 ctr -= caspty ; | |
2541 continue ; | |
2542 } | |
2543 | |
2544 // Did lock ownership change hands ? | |
2545 if (ox != prv && prv != NULL ) { | |
2546 TEVENT (spin: Owner changed) | |
2547 if (oxpty == -2) break ; | |
2548 if (oxpty == -1) goto Abort ; | |
2549 ctr -= oxpty ; | |
2550 } | |
2551 prv = ox ; | |
2552 | |
2553 // Abort the spin if the owner is not executing. | |
2554 // The owner must be executing in order to drop the lock. | |
2555 // Spinning while the owner is OFFPROC is idiocy. | |
2556 // Consider: ctr -= RunnablePenalty ; | |
2557 if (Knob_OState && NotRunnable (Self, ox)) { | |
2558 TEVENT (Spin abort - notrunnable); | |
2559 goto Abort ; | |
2560 } | |
2561 if (sss && _succ == NULL ) _succ = Self ; | |
2562 } | |
2563 | |
2564 // Spin failed with prejudice -- reduce _SpinDuration. | |
2565 // TODO: Use an AIMD-like policy to adjust _SpinDuration. | |
2566 // AIMD is globally stable. | |
2567 TEVENT (Spin failure) ; | |
2568 { | |
2569 int x = _SpinDuration ; | |
2570 if (x > 0) { | |
2571 // Consider an AIMD scheme like: x -= (x >> 3) + 100 | |
2572 // This is globally sample and tends to damp the response. | |
2573 x -= Knob_Penalty ; | |
2574 if (x < 0) x = 0 ; | |
2575 _SpinDuration = x ; | |
2576 } | |
2577 } | |
2578 | |
2579 Abort: | |
2580 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; | |
2581 if (sss && _succ == Self) { | |
2582 _succ = NULL ; | |
2583 // Invariant: after setting succ=null a contending thread | |
2584 // must recheck-retry _owner before parking. This usually happens | |
2585 // in the normal usage of TrySpin(), but it's safest | |
2586 // to make TrySpin() as foolproof as possible. | |
2587 OrderAccess::fence() ; | |
2588 if (TryLock(Self) > 0) return 1 ; | |
2589 } | |
2590 return 0 ; | |
2591 } | |
2592 | |
2593 #define TrySpin TrySpin_VaryDuration | |
2594 | |
2595 static void DeferredInitialize () { | |
2596 if (InitDone > 0) return ; | |
2597 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { | |
2598 while (InitDone != 1) ; | |
2599 return ; | |
2600 } | |
2601 | |
2602 // One-shot global initialization ... | |
2603 // The initialization is idempotent, so we don't need locks. | |
2604 // In the future consider doing this via os::init_2(). | |
2605 // SyncKnobs consist of <Key>=<Value> pairs in the style | |
2606 // of environment variables. Start by converting ':' to NUL. | |
2607 | |
2608 if (SyncKnobs == NULL) SyncKnobs = "" ; | |
2609 | |
2610 size_t sz = strlen (SyncKnobs) ; | |
2611 char * knobs = (char *) malloc (sz + 2) ; | |
2612 if (knobs == NULL) { | |
2613 vm_exit_out_of_memory (sz + 2, "Parse SyncKnobs") ; | |
2614 guarantee (0, "invariant") ; | |
2615 } | |
2616 strcpy (knobs, SyncKnobs) ; | |
2617 knobs[sz+1] = 0 ; | |
2618 for (char * p = knobs ; *p ; p++) { | |
2619 if (*p == ':') *p = 0 ; | |
2620 } | |
2621 | |
2622 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } | |
2623 SETKNOB(ReportSettings) ; | |
2624 SETKNOB(Verbose) ; | |
2625 SETKNOB(FixedSpin) ; | |
2626 SETKNOB(SpinLimit) ; | |
2627 SETKNOB(SpinBase) ; | |
2628 SETKNOB(SpinBackOff); | |
2629 SETKNOB(CASPenalty) ; | |
2630 SETKNOB(OXPenalty) ; | |
2631 SETKNOB(LogSpins) ; | |
2632 SETKNOB(SpinSetSucc) ; | |
2633 SETKNOB(SuccEnabled) ; | |
2634 SETKNOB(SuccRestrict) ; | |
2635 SETKNOB(Penalty) ; | |
2636 SETKNOB(Bonus) ; | |
2637 SETKNOB(BonusB) ; | |
2638 SETKNOB(Poverty) ; | |
2639 SETKNOB(SpinAfterFutile) ; | |
2640 SETKNOB(UsePause) ; | |
2641 SETKNOB(SpinEarly) ; | |
2642 SETKNOB(OState) ; | |
2643 SETKNOB(MaxSpinners) ; | |
2644 SETKNOB(PreSpin) ; | |
2645 SETKNOB(ExitPolicy) ; | |
2646 SETKNOB(QMode); | |
2647 SETKNOB(ResetEvent) ; | |
2648 SETKNOB(MoveNotifyee) ; | |
2649 SETKNOB(FastHSSEC) ; | |
2650 #undef SETKNOB | |
2651 | |
2652 if (os::is_MP()) { | |
2653 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; | |
2654 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; | |
2655 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) | |
2656 } else { | |
2657 Knob_SpinLimit = 0 ; | |
2658 Knob_SpinBase = 0 ; | |
2659 Knob_PreSpin = 0 ; | |
2660 Knob_FixedSpin = -1 ; | |
2661 } | |
2662 | |
2663 if (Knob_LogSpins == 0) { | |
2664 ObjectSynchronizer::_sync_FailedSpins = NULL ; | |
2665 } | |
2666 | |
2667 free (knobs) ; | |
2668 OrderAccess::fence() ; | |
2669 InitDone = 1 ; | |
2670 } | |
2671 | |
2672 // Theory of operations -- Monitors lists, thread residency, etc: | |
2673 // | |
2674 // * A thread acquires ownership of a monitor by successfully | |
2675 // CAS()ing the _owner field from null to non-null. | |
2676 // | |
2677 // * Invariant: A thread appears on at most one monitor list -- | |
2678 // cxq, EntryList or WaitSet -- at any one time. | |
2679 // | |
2680 // * Contending threads "push" themselves onto the cxq with CAS | |
2681 // and then spin/park. | |
2682 // | |
2683 // * After a contending thread eventually acquires the lock it must | |
2684 // dequeue itself from either the EntryList or the cxq. | |
2685 // | |
2686 // * The exiting thread identifies and unparks an "heir presumptive" | |
2687 // tentative successor thread on the EntryList. Critically, the | |
2688 // exiting thread doesn't unlink the successor thread from the EntryList. | |
2689 // After having been unparked, the wakee will recontend for ownership of | |
2690 // the monitor. The successor (wakee) will either acquire the lock or | |
2691 // re-park itself. | |
2692 // | |
2693 // Succession is provided for by a policy of competitive handoff. | |
2694 // The exiting thread does _not_ grant or pass ownership to the | |
2695 // successor thread. (This is also referred to as "handoff" succession"). | |
2696 // Instead the exiting thread releases ownership and possibly wakes | |
2697 // a successor, so the successor can (re)compete for ownership of the lock. | |
2698 // If the EntryList is empty but the cxq is populated the exiting | |
2699 // thread will drain the cxq into the EntryList. It does so by | |
2700 // by detaching the cxq (installing null with CAS) and folding | |
2701 // the threads from the cxq into the EntryList. The EntryList is | |
2702 // doubly linked, while the cxq is singly linked because of the | |
2703 // CAS-based "push" used to enqueue recently arrived threads (RATs). | |
2704 // | |
2705 // * Concurrency invariants: | |
2706 // | |
2707 // -- only the monitor owner may access or mutate the EntryList. | |
2708 // The mutex property of the monitor itself protects the EntryList | |
2709 // from concurrent interference. | |
2710 // -- Only the monitor owner may detach the cxq. | |
2711 // | |
2712 // * The monitor entry list operations avoid locks, but strictly speaking | |
2713 // they're not lock-free. Enter is lock-free, exit is not. | |
2714 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html | |
2715 // | |
2716 // * The cxq can have multiple concurrent "pushers" but only one concurrent | |
2717 // detaching thread. This mechanism is immune from the ABA corruption. | |
2718 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. | |
2719 // | |
2720 // * Taken together, the cxq and the EntryList constitute or form a | |
2721 // single logical queue of threads stalled trying to acquire the lock. | |
2722 // We use two distinct lists to improve the odds of a constant-time | |
2723 // dequeue operation after acquisition (in the ::enter() epilog) and | |
2724 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). | |
2725 // A key desideratum is to minimize queue & monitor metadata manipulation | |
2726 // that occurs while holding the monitor lock -- that is, we want to | |
2727 // minimize monitor lock holds times. Note that even a small amount of | |
2728 // fixed spinning will greatly reduce the # of enqueue-dequeue operations | |
2729 // on EntryList|cxq. That is, spinning relieves contention on the "inner" | |
2730 // locks and monitor metadata. | |
2731 // | |
2732 // Cxq points to the the set of Recently Arrived Threads attempting entry. | |
2733 // Because we push threads onto _cxq with CAS, the RATs must take the form of | |
2734 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when | |
2735 // the unlocking thread notices that EntryList is null but _cxq is != null. | |
2736 // | |
2737 // The EntryList is ordered by the prevailing queue discipline and | |
2738 // can be organized in any convenient fashion, such as a doubly-linked list or | |
2739 // a circular doubly-linked list. Critically, we want insert and delete operations | |
2740 // to operate in constant-time. If we need a priority queue then something akin | |
2741 // to Solaris' sleepq would work nicely. Viz., | |
2742 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. | |
2743 // Queue discipline is enforced at ::exit() time, when the unlocking thread | |
2744 // drains the cxq into the EntryList, and orders or reorders the threads on the | |
2745 // EntryList accordingly. | |
2746 // | |
2747 // Barring "lock barging", this mechanism provides fair cyclic ordering, | |
2748 // somewhat similar to an elevator-scan. | |
2749 // | |
2750 // * The monitor synchronization subsystem avoids the use of native | |
2751 // synchronization primitives except for the narrow platform-specific | |
2752 // park-unpark abstraction. See the comments in os_solaris.cpp regarding | |
2753 // the semantics of park-unpark. Put another way, this monitor implementation | |
2754 // depends only on atomic operations and park-unpark. The monitor subsystem | |
2755 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the | |
2756 // underlying OS manages the READY<->RUN transitions. | |
2757 // | |
2758 // * Waiting threads reside on the WaitSet list -- wait() puts | |
2759 // the caller onto the WaitSet. | |
2760 // | |
2761 // * notify() or notifyAll() simply transfers threads from the WaitSet to | |
2762 // either the EntryList or cxq. Subsequent exit() operations will | |
2763 // unpark the notifyee. Unparking a notifee in notify() is inefficient - | |
2764 // it's likely the notifyee would simply impale itself on the lock held | |
2765 // by the notifier. | |
2766 // | |
2767 // * An interesting alternative is to encode cxq as (List,LockByte) where | |
2768 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary | |
2769 // variable, like _recursions, in the scheme. The threads or Events that form | |
2770 // the list would have to be aligned in 256-byte addresses. A thread would | |
2771 // try to acquire the lock or enqueue itself with CAS, but exiting threads | |
2772 // could use a 1-0 protocol and simply STB to set the LockByte to 0. | |
2773 // Note that is is *not* word-tearing, but it does presume that full-word | |
2774 // CAS operations are coherent with intermix with STB operations. That's true | |
2775 // on most common processors. | |
2776 // | |
2777 // * See also http://blogs.sun.com/dave | |
2778 | |
2779 | |
2780 void ATTR ObjectMonitor::EnterI (TRAPS) { | |
2781 Thread * Self = THREAD ; | |
2782 assert (Self->is_Java_thread(), "invariant") ; | |
2783 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; | |
2784 | |
2785 // Try the lock - TATAS | |
2786 if (TryLock (Self) > 0) { | |
2787 assert (_succ != Self , "invariant") ; | |
2788 assert (_owner == Self , "invariant") ; | |
2789 assert (_Responsible != Self , "invariant") ; | |
2790 return ; | |
2791 } | |
2792 | |
2793 DeferredInitialize () ; | |
2794 | |
2795 // We try one round of spinning *before* enqueueing Self. | |
2796 // | |
2797 // If the _owner is ready but OFFPROC we could use a YieldTo() | |
2798 // operation to donate the remainder of this thread's quantum | |
2799 // to the owner. This has subtle but beneficial affinity | |
2800 // effects. | |
2801 | |
2802 if (TrySpin (Self) > 0) { | |
2803 assert (_owner == Self , "invariant") ; | |
2804 assert (_succ != Self , "invariant") ; | |
2805 assert (_Responsible != Self , "invariant") ; | |
2806 return ; | |
2807 } | |
2808 | |
2809 // The Spin failed -- Enqueue and park the thread ... | |
2810 assert (_succ != Self , "invariant") ; | |
2811 assert (_owner != Self , "invariant") ; | |
2812 assert (_Responsible != Self , "invariant") ; | |
2813 | |
2814 // Enqueue "Self" on ObjectMonitor's _cxq. | |
2815 // | |
2816 // Node acts as a proxy for Self. | |
2817 // As an aside, if were to ever rewrite the synchronization code mostly | |
2818 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class | |
2819 // Java objects. This would avoid awkward lifecycle and liveness issues, | |
2820 // as well as eliminate a subset of ABA issues. | |
2821 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. | |
2822 // | |
2823 | |
2824 ObjectWaiter node(Self) ; | |
2825 Self->_ParkEvent->reset() ; | |
2826 node._prev = (ObjectWaiter *) 0xBAD ; | |
2827 node.TState = ObjectWaiter::TS_CXQ ; | |
2828 | |
2829 // Push "Self" onto the front of the _cxq. | |
2830 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. | |
2831 // Note that spinning tends to reduce the rate at which threads | |
2832 // enqueue and dequeue on EntryList|cxq. | |
2833 ObjectWaiter * nxt ; | |
2834 for (;;) { | |
2835 node._next = nxt = _cxq ; | |
2836 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; | |
2837 | |
2838 // Interference - the CAS failed because _cxq changed. Just retry. | |
2839 // As an optional optimization we retry the lock. | |
2840 if (TryLock (Self) > 0) { | |
2841 assert (_succ != Self , "invariant") ; | |
2842 assert (_owner == Self , "invariant") ; | |
2843 assert (_Responsible != Self , "invariant") ; | |
2844 return ; | |
2845 } | |
2846 } | |
2847 | |
2848 // Check for cxq|EntryList edge transition to non-null. This indicates | |
2849 // the onset of contention. While contention persists exiting threads | |
2850 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit | |
2851 // operations revert to the faster 1-0 mode. This enter operation may interleave | |
2852 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we | |
2853 // arrange for one of the contending thread to use a timed park() operations | |
2854 // to detect and recover from the race. (Stranding is form of progress failure | |
2855 // where the monitor is unlocked but all the contending threads remain parked). | |
2856 // That is, at least one of the contended threads will periodically poll _owner. | |
2857 // One of the contending threads will become the designated "Responsible" thread. | |
2858 // The Responsible thread uses a timed park instead of a normal indefinite park | |
2859 // operation -- it periodically wakes and checks for and recovers from potential | |
2860 // strandings admitted by 1-0 exit operations. We need at most one Responsible | |
2861 // thread per-monitor at any given moment. Only threads on cxq|EntryList may | |
2862 // be responsible for a monitor. | |
2863 // | |
2864 // Currently, one of the contended threads takes on the added role of "Responsible". | |
2865 // A viable alternative would be to use a dedicated "stranding checker" thread | |
2866 // that periodically iterated over all the threads (or active monitors) and unparked | |
2867 // successors where there was risk of stranding. This would help eliminate the | |
2868 // timer scalability issues we see on some platforms as we'd only have one thread | |
2869 // -- the checker -- parked on a timer. | |
2870 | |
2871 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { | |
2872 // Try to assume the role of responsible thread for the monitor. | |
2873 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } | |
2874 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; | |
2875 } | |
2876 | |
2877 // The lock have been released while this thread was occupied queueing | |
2878 // itself onto _cxq. To close the race and avoid "stranding" and | |
2879 // progress-liveness failure we must resample-retry _owner before parking. | |
2880 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. | |
2881 // In this case the ST-MEMBAR is accomplished with CAS(). | |
2882 // | |
2883 // TODO: Defer all thread state transitions until park-time. | |
2884 // Since state transitions are heavy and inefficient we'd like | |
2885 // to defer the state transitions until absolutely necessary, | |
2886 // and in doing so avoid some transitions ... | |
2887 | |
2888 TEVENT (Inflated enter - Contention) ; | |
2889 int nWakeups = 0 ; | |
2890 int RecheckInterval = 1 ; | |
2891 | |
2892 for (;;) { | |
2893 | |
2894 if (TryLock (Self) > 0) break ; | |
2895 assert (_owner != Self, "invariant") ; | |
2896 | |
2897 if ((SyncFlags & 2) && _Responsible == NULL) { | |
2898 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; | |
2899 } | |
2900 | |
2901 // park self | |
2902 if (_Responsible == Self || (SyncFlags & 1)) { | |
2903 TEVENT (Inflated enter - park TIMED) ; | |
2904 Self->_ParkEvent->park ((jlong) RecheckInterval) ; | |
2905 // Increase the RecheckInterval, but clamp the value. | |
2906 RecheckInterval *= 8 ; | |
2907 if (RecheckInterval > 1000) RecheckInterval = 1000 ; | |
2908 } else { | |
2909 TEVENT (Inflated enter - park UNTIMED) ; | |
2910 Self->_ParkEvent->park() ; | |
2911 } | |
2912 | |
2913 if (TryLock(Self) > 0) break ; | |
2914 | |
2915 // The lock is still contested. | |
2916 // Keep a tally of the # of futile wakeups. | |
2917 // Note that the counter is not protected by a lock or updated by atomics. | |
2918 // That is by design - we trade "lossy" counters which are exposed to | |
2919 // races during updates for a lower probe effect. | |
2920 TEVENT (Inflated enter - Futile wakeup) ; | |
2921 if (ObjectSynchronizer::_sync_FutileWakeups != NULL) { | |
2922 ObjectSynchronizer::_sync_FutileWakeups->inc() ; | |
2923 } | |
2924 ++ nWakeups ; | |
2925 | |
2926 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. | |
2927 // We can defer clearing _succ until after the spin completes | |
2928 // TrySpin() must tolerate being called with _succ == Self. | |
2929 // Try yet another round of adaptive spinning. | |
2930 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; | |
2931 | |
2932 // We can find that we were unpark()ed and redesignated _succ while | |
2933 // we were spinning. That's harmless. If we iterate and call park(), | |
2934 // park() will consume the event and return immediately and we'll | |
2935 // just spin again. This pattern can repeat, leaving _succ to simply | |
2936 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). | |
2937 // Alternately, we can sample fired() here, and if set, forgo spinning | |
2938 // in the next iteration. | |
2939 | |
2940 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { | |
2941 Self->_ParkEvent->reset() ; | |
2942 OrderAccess::fence() ; | |
2943 } | |
2944 if (_succ == Self) _succ = NULL ; | |
2945 | |
2946 // Invariant: after clearing _succ a thread *must* retry _owner before parking. | |
2947 OrderAccess::fence() ; | |
2948 } | |
2949 | |
2950 // Egress : | |
2951 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. | |
2952 // Normally we'll find Self on the EntryList . | |
2953 // From the perspective of the lock owner (this thread), the | |
2954 // EntryList is stable and cxq is prepend-only. | |
2955 // The head of cxq is volatile but the interior is stable. | |
2956 // In addition, Self.TState is stable. | |
2957 | |
2958 assert (_owner == Self , "invariant") ; | |
2959 assert (object() != NULL , "invariant") ; | |
2960 // I'd like to write: | |
2961 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; | |
2962 // but as we're at a safepoint that's not safe. | |
2963 | |
2964 UnlinkAfterAcquire (Self, &node) ; | |
2965 if (_succ == Self) _succ = NULL ; | |
2966 | |
2967 assert (_succ != Self, "invariant") ; | |
2968 if (_Responsible == Self) { | |
2969 _Responsible = NULL ; | |
2970 // Dekker pivot-point. | |
2971 // Consider OrderAccess::storeload() here | |
2972 | |
2973 // We may leave threads on cxq|EntryList without a designated | |
2974 // "Responsible" thread. This is benign. When this thread subsequently | |
2975 // exits the monitor it can "see" such preexisting "old" threads -- | |
2976 // threads that arrived on the cxq|EntryList before the fence, above -- | |
2977 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads | |
2978 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible | |
2979 // non-null and elect a new "Responsible" timer thread. | |
2980 // | |
2981 // This thread executes: | |
2982 // ST Responsible=null; MEMBAR (in enter epilog - here) | |
2983 // LD cxq|EntryList (in subsequent exit) | |
2984 // | |
2985 // Entering threads in the slow/contended path execute: | |
2986 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) | |
2987 // The (ST cxq; MEMBAR) is accomplished with CAS(). | |
2988 // | |
2989 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent | |
2990 // exit operation from floating above the ST Responsible=null. | |
2991 // | |
2992 // In *practice* however, EnterI() is always followed by some atomic | |
2993 // operation such as the decrement of _count in ::enter(). Those atomics | |
2994 // obviate the need for the explicit MEMBAR, above. | |
2995 } | |
2996 | |
2997 // We've acquired ownership with CAS(). | |
2998 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. | |
2999 // But since the CAS() this thread may have also stored into _succ, | |
3000 // EntryList, cxq or Responsible. These meta-data updates must be | |
3001 // visible __before this thread subsequently drops the lock. | |
3002 // Consider what could occur if we didn't enforce this constraint -- | |
3003 // STs to monitor meta-data and user-data could reorder with (become | |
3004 // visible after) the ST in exit that drops ownership of the lock. | |
3005 // Some other thread could then acquire the lock, but observe inconsistent | |
3006 // or old monitor meta-data and heap data. That violates the JMM. | |
3007 // To that end, the 1-0 exit() operation must have at least STST|LDST | |
3008 // "release" barrier semantics. Specifically, there must be at least a | |
3009 // STST|LDST barrier in exit() before the ST of null into _owner that drops | |
3010 // the lock. The barrier ensures that changes to monitor meta-data and data | |
3011 // protected by the lock will be visible before we release the lock, and | |
3012 // therefore before some other thread (CPU) has a chance to acquire the lock. | |
3013 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. | |
3014 // | |
3015 // Critically, any prior STs to _succ or EntryList must be visible before | |
3016 // the ST of null into _owner in the *subsequent* (following) corresponding | |
3017 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily | |
3018 // execute a serializing instruction. | |
3019 | |
3020 if (SyncFlags & 8) { | |
3021 OrderAccess::fence() ; | |
3022 } | |
3023 return ; | |
3024 } | |
3025 | |
3026 // ExitSuspendEquivalent: | |
3027 // A faster alternate to handle_special_suspend_equivalent_condition() | |
3028 // | |
3029 // handle_special_suspend_equivalent_condition() unconditionally | |
3030 // acquires the SR_lock. On some platforms uncontended MutexLocker() | |
3031 // operations have high latency. Note that in ::enter() we call HSSEC | |
3032 // while holding the monitor, so we effectively lengthen the critical sections. | |
3033 // | |
3034 // There are a number of possible solutions: | |
3035 // | |
3036 // A. To ameliorate the problem we might also defer state transitions | |
3037 // to as late as possible -- just prior to parking. | |
3038 // Given that, we'd call HSSEC after having returned from park(), | |
3039 // but before attempting to acquire the monitor. This is only a | |
3040 // partial solution. It avoids calling HSSEC while holding the | |
3041 // monitor (good), but it still increases successor reacquisition latency -- | |
3042 // the interval between unparking a successor and the time the successor | |
3043 // resumes and retries the lock. See ReenterI(), which defers state transitions. | |
3044 // If we use this technique we can also avoid EnterI()-exit() loop | |
3045 // in ::enter() where we iteratively drop the lock and then attempt | |
3046 // to reacquire it after suspending. | |
3047 // | |
3048 // B. In the future we might fold all the suspend bits into a | |
3049 // composite per-thread suspend flag and then update it with CAS(). | |
3050 // Alternately, a Dekker-like mechanism with multiple variables | |
3051 // would suffice: | |
3052 // ST Self->_suspend_equivalent = false | |
3053 // MEMBAR | |
3054 // LD Self_>_suspend_flags | |
3055 // | |
3056 | |
3057 | |
3058 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { | |
3059 int Mode = Knob_FastHSSEC ; | |
3060 if (Mode && !jSelf->is_external_suspend()) { | |
3061 assert (jSelf->is_suspend_equivalent(), "invariant") ; | |
3062 jSelf->clear_suspend_equivalent() ; | |
3063 if (2 == Mode) OrderAccess::storeload() ; | |
3064 if (!jSelf->is_external_suspend()) return false ; | |
3065 // We raced a suspension -- fall thru into the slow path | |
3066 TEVENT (ExitSuspendEquivalent - raced) ; | |
3067 jSelf->set_suspend_equivalent() ; | |
3068 } | |
3069 return jSelf->handle_special_suspend_equivalent_condition() ; | |
3070 } | |
3071 | |
3072 | |
3073 // ReenterI() is a specialized inline form of the latter half of the | |
3074 // contended slow-path from EnterI(). We use ReenterI() only for | |
3075 // monitor reentry in wait(). | |
3076 // | |
3077 // In the future we should reconcile EnterI() and ReenterI(), adding | |
3078 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the | |
3079 // loop accordingly. | |
3080 | |
3081 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { | |
3082 assert (Self != NULL , "invariant") ; | |
3083 assert (SelfNode != NULL , "invariant") ; | |
3084 assert (SelfNode->_thread == Self , "invariant") ; | |
3085 assert (_waiters > 0 , "invariant") ; | |
3086 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; | |
3087 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; | |
3088 JavaThread * jt = (JavaThread *) Self ; | |
3089 | |
3090 int nWakeups = 0 ; | |
3091 for (;;) { | |
3092 ObjectWaiter::TStates v = SelfNode->TState ; | |
3093 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; | |
3094 assert (_owner != Self, "invariant") ; | |
3095 | |
3096 if (TryLock (Self) > 0) break ; | |
3097 if (TrySpin (Self) > 0) break ; | |
3098 | |
3099 TEVENT (Wait Reentry - parking) ; | |
3100 | |
3101 // State transition wrappers around park() ... | |
3102 // ReenterI() wisely defers state transitions until | |
3103 // it's clear we must park the thread. | |
3104 { | |
3105 OSThreadContendState osts(Self->osthread()); | |
3106 ThreadBlockInVM tbivm(jt); | |
3107 | |
3108 // cleared by handle_special_suspend_equivalent_condition() | |
3109 // or java_suspend_self() | |
3110 jt->set_suspend_equivalent(); | |
3111 if (SyncFlags & 1) { | |
3112 Self->_ParkEvent->park ((jlong)1000) ; | |
3113 } else { | |
3114 Self->_ParkEvent->park () ; | |
3115 } | |
3116 | |
3117 // were we externally suspended while we were waiting? | |
3118 for (;;) { | |
3119 if (!ExitSuspendEquivalent (jt)) break ; | |
3120 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } | |
3121 jt->java_suspend_self(); | |
3122 jt->set_suspend_equivalent(); | |
3123 } | |
3124 } | |
3125 | |
3126 // Try again, but just so we distinguish between futile wakeups and | |
3127 // successful wakeups. The following test isn't algorithmically | |
3128 // necessary, but it helps us maintain sensible statistics. | |
3129 if (TryLock(Self) > 0) break ; | |
3130 | |
3131 // The lock is still contested. | |
3132 // Keep a tally of the # of futile wakeups. | |
3133 // Note that the counter is not protected by a lock or updated by atomics. | |
3134 // That is by design - we trade "lossy" counters which are exposed to | |
3135 // races during updates for a lower probe effect. | |
3136 TEVENT (Wait Reentry - futile wakeup) ; | |
3137 ++ nWakeups ; | |
3138 | |
3139 // Assuming this is not a spurious wakeup we'll normally | |
3140 // find that _succ == Self. | |
3141 if (_succ == Self) _succ = NULL ; | |
3142 | |
3143 // Invariant: after clearing _succ a contending thread | |
3144 // *must* retry _owner before parking. | |
3145 OrderAccess::fence() ; | |
3146 | |
3147 if (ObjectSynchronizer::_sync_FutileWakeups != NULL) { | |
3148 ObjectSynchronizer::_sync_FutileWakeups->inc() ; | |
3149 } | |
3150 } | |
3151 | |
3152 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . | |
3153 // Normally we'll find Self on the EntryList. | |
3154 // Unlinking from the EntryList is constant-time and atomic-free. | |
3155 // From the perspective of the lock owner (this thread), the | |
3156 // EntryList is stable and cxq is prepend-only. | |
3157 // The head of cxq is volatile but the interior is stable. | |
3158 // In addition, Self.TState is stable. | |
3159 | |
3160 assert (_owner == Self, "invariant") ; | |
3161 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; | |
3162 UnlinkAfterAcquire (Self, SelfNode) ; | |
3163 if (_succ == Self) _succ = NULL ; | |
3164 assert (_succ != Self, "invariant") ; | |
3165 SelfNode->TState = ObjectWaiter::TS_RUN ; | |
3166 OrderAccess::fence() ; // see comments at the end of EnterI() | |
3167 } | |
3168 | |
3169 bool ObjectMonitor::try_enter(Thread* THREAD) { | |
3170 if (THREAD != _owner) { | |
3171 if (THREAD->is_lock_owned ((address)_owner)) { | |
3172 assert(_recursions == 0, "internal state error"); | |
3173 _owner = THREAD ; | |
3174 _recursions = 1 ; | |
3175 OwnerIsThread = 1 ; | |
3176 return true; | |
3177 } | |
3178 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { | |
3179 return false; | |
3180 } | |
3181 return true; | |
3182 } else { | |
3183 _recursions++; | |
3184 return true; | |
3185 } | |
3186 } | |
3187 | |
3188 void ATTR ObjectMonitor::enter(TRAPS) { | |
3189 // The following code is ordered to check the most common cases first | |
3190 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. | |
3191 Thread * const Self = THREAD ; | |
3192 void * cur ; | |
3193 | |
3194 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; | |
3195 if (cur == NULL) { | |
3196 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. | |
3197 assert (_recursions == 0 , "invariant") ; | |
3198 assert (_owner == Self, "invariant") ; | |
3199 // CONSIDER: set or assert OwnerIsThread == 1 | |
3200 return ; | |
3201 } | |
3202 | |
3203 if (cur == Self) { | |
3204 // TODO-FIXME: check for integer overflow! BUGID 6557169. | |
3205 _recursions ++ ; | |
3206 return ; | |
3207 } | |
3208 | |
3209 if (Self->is_lock_owned ((address)cur)) { | |
3210 assert (_recursions == 0, "internal state error"); | |
3211 _recursions = 1 ; | |
3212 // Commute owner from a thread-specific on-stack BasicLockObject address to | |
3213 // a full-fledged "Thread *". | |
3214 _owner = Self ; | |
3215 OwnerIsThread = 1 ; | |
3216 return ; | |
3217 } | |
3218 | |
3219 // We've encountered genuine contention. | |
3220 assert (Self->_Stalled == 0, "invariant") ; | |
3221 Self->_Stalled = intptr_t(this) ; | |
3222 | |
3223 // Try one round of spinning *before* enqueueing Self | |
3224 // and before going through the awkward and expensive state | |
3225 // transitions. The following spin is strictly optional ... | |
3226 // Note that if we acquire the monitor from an initial spin | |
3227 // we forgo posting JVMTI events and firing DTRACE probes. | |
3228 if (Knob_SpinEarly && TrySpin (Self) > 0) { | |
3229 assert (_owner == Self , "invariant") ; | |
3230 assert (_recursions == 0 , "invariant") ; | |
3231 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; | |
3232 Self->_Stalled = 0 ; | |
3233 return ; | |
3234 } | |
3235 | |
3236 assert (_owner != Self , "invariant") ; | |
3237 assert (_succ != Self , "invariant") ; | |
3238 assert (Self->is_Java_thread() , "invariant") ; | |
3239 JavaThread * jt = (JavaThread *) Self ; | |
3240 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; | |
3241 assert (jt->thread_state() != _thread_blocked , "invariant") ; | |
3242 assert (this->object() != NULL , "invariant") ; | |
3243 assert (_count >= 0, "invariant") ; | |
3244 | |
3245 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). | |
3246 // Ensure the object-monitor relationship remains stable while there's contention. | |
3247 Atomic::inc_ptr(&_count); | |
3248 | |
3249 { // Change java thread status to indicate blocked on monitor enter. | |
3250 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); | |
3251 | |
3252 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); | |
3253 if (JvmtiExport::should_post_monitor_contended_enter()) { | |
3254 JvmtiExport::post_monitor_contended_enter(jt, this); | |
3255 } | |
3256 | |
3257 OSThreadContendState osts(Self->osthread()); | |
3258 ThreadBlockInVM tbivm(jt); | |
3259 | |
3260 Self->set_current_pending_monitor(this); | |
3261 | |
3262 // TODO-FIXME: change the following for(;;) loop to straight-line code. | |
3263 for (;;) { | |
3264 jt->set_suspend_equivalent(); | |
3265 // cleared by handle_special_suspend_equivalent_condition() | |
3266 // or java_suspend_self() | |
3267 | |
3268 EnterI (THREAD) ; | |
3269 | |
3270 if (!ExitSuspendEquivalent(jt)) break ; | |
3271 | |
3272 // | |
3273 // We have acquired the contended monitor, but while we were | |
3274 // waiting another thread suspended us. We don't want to enter | |
3275 // the monitor while suspended because that would surprise the | |
3276 // thread that suspended us. | |
3277 // | |
3278 _recursions = 0 ; | |
3279 _succ = NULL ; | |
3280 exit (Self) ; | |
3281 | |
3282 jt->java_suspend_self(); | |
3283 } | |
3284 Self->set_current_pending_monitor(NULL); | |
3285 } | |
3286 | |
3287 Atomic::dec_ptr(&_count); | |
3288 assert (_count >= 0, "invariant") ; | |
3289 Self->_Stalled = 0 ; | |
3290 | |
3291 // Must either set _recursions = 0 or ASSERT _recursions == 0. | |
3292 assert (_recursions == 0 , "invariant") ; | |
3293 assert (_owner == Self , "invariant") ; | |
3294 assert (_succ != Self , "invariant") ; | |
3295 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; | |
3296 | |
3297 // The thread -- now the owner -- is back in vm mode. | |
3298 // Report the glorious news via TI,DTrace and jvmstat. | |
3299 // The probe effect is non-trivial. All the reportage occurs | |
3300 // while we hold the monitor, increasing the length of the critical | |
3301 // section. Amdahl's parallel speedup law comes vividly into play. | |
3302 // | |
3303 // Another option might be to aggregate the events (thread local or | |
3304 // per-monitor aggregation) and defer reporting until a more opportune | |
3305 // time -- such as next time some thread encounters contention but has | |
3306 // yet to acquire the lock. While spinning that thread could | |
3307 // spinning we could increment JVMStat counters, etc. | |
3308 | |
3309 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); | |
3310 if (JvmtiExport::should_post_monitor_contended_entered()) { | |
3311 JvmtiExport::post_monitor_contended_entered(jt, this); | |
3312 } | |
3313 if (ObjectSynchronizer::_sync_ContendedLockAttempts != NULL) { | |
3314 ObjectSynchronizer::_sync_ContendedLockAttempts->inc() ; | |
3315 } | |
3316 } | |
3317 | |
3318 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { | |
3319 assert (_owner == Self, "invariant") ; | |
3320 | |
3321 // Exit protocol: | |
3322 // 1. ST _succ = wakee | |
3323 // 2. membar #loadstore|#storestore; | |
3324 // 2. ST _owner = NULL | |
3325 // 3. unpark(wakee) | |
3326 | |
3327 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; | |
3328 ParkEvent * Trigger = Wakee->_event ; | |
3329 | |
3330 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. | |
3331 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be | |
3332 // out-of-scope (non-extant). | |
3333 Wakee = NULL ; | |
3334 | |
3335 // Drop the lock | |
3336 OrderAccess::release_store_ptr (&_owner, NULL) ; | |
3337 OrderAccess::fence() ; // ST _owner vs LD in unpark() | |
3338 | |
3339 // TODO-FIXME: | |
3340 // If there's a safepoint pending the best policy would be to | |
3341 // get _this thread to a safepoint and only wake the successor | |
3342 // after the safepoint completed. monitorexit uses a "leaf" | |
3343 // state transition, however, so this thread can't become | |
3344 // safe at this point in time. (Its stack isn't walkable). | |
3345 // The next best thing is to defer waking the successor by | |
3346 // adding to a list of thread to be unparked after at the | |
3347 // end of the forthcoming STW). | |
3348 if (SafepointSynchronize::do_call_back()) { | |
3349 TEVENT (unpark before SAFEPOINT) ; | |
3350 } | |
3351 | |
3352 // Possible optimizations ... | |
3353 // | |
3354 // * Consider: set Wakee->UnparkTime = timeNow() | |
3355 // When the thread wakes up it'll compute (timeNow() - Self->UnparkTime()). | |
3356 // By measuring recent ONPROC latency we can approximate the | |
3357 // system load. In turn, we can feed that information back | |
3358 // into the spinning & succession policies. | |
3359 // (ONPROC latency correlates strongly with load). | |
3360 // | |
3361 // * Pull affinity: | |
3362 // If the wakee is cold then transiently setting it's affinity | |
3363 // to the current CPU is a good idea. | |
3364 // See http://j2se.east/~dice/PERSIST/050624-PullAffinity.txt | |
478
a7fac4381b50
6639341: sometimes contended-exit event comes after contended-entered on another thread
blacklion
parents:
0
diff
changeset
|
3365 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); |
0 | 3366 Trigger->unpark() ; |
3367 | |
3368 // Maintain stats and report events to JVMTI | |
3369 if (ObjectSynchronizer::_sync_Parks != NULL) { | |
3370 ObjectSynchronizer::_sync_Parks->inc() ; | |
3371 } | |
3372 } | |
3373 | |
3374 | |
3375 // exit() | |
3376 // ~~~~~~ | |
3377 // Note that the collector can't reclaim the objectMonitor or deflate | |
3378 // the object out from underneath the thread calling ::exit() as the | |
3379 // thread calling ::exit() never transitions to a stable state. | |
3380 // This inhibits GC, which in turn inhibits asynchronous (and | |
3381 // inopportune) reclamation of "this". | |
3382 // | |
3383 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; | |
3384 // There's one exception to the claim above, however. EnterI() can call | |
3385 // exit() to drop a lock if the acquirer has been externally suspended. | |
3386 // In that case exit() is called with _thread_state as _thread_blocked, | |
3387 // but the monitor's _count field is > 0, which inhibits reclamation. | |
3388 // | |
3389 // 1-0 exit | |
3390 // ~~~~~~~~ | |
3391 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of | |
3392 // the fast-path operators have been optimized so the common ::exit() | |
3393 // operation is 1-0. See i486.ad fast_unlock(), for instance. | |
3394 // The code emitted by fast_unlock() elides the usual MEMBAR. This | |
3395 // greatly improves latency -- MEMBAR and CAS having considerable local | |
3396 // latency on modern processors -- but at the cost of "stranding". Absent the | |
3397 // MEMBAR, a thread in fast_unlock() can race a thread in the slow | |
3398 // ::enter() path, resulting in the entering thread being stranding | |
3399 // and a progress-liveness failure. Stranding is extremely rare. | |
3400 // We use timers (timed park operations) & periodic polling to detect | |
3401 // and recover from stranding. Potentially stranded threads periodically | |
3402 // wake up and poll the lock. See the usage of the _Responsible variable. | |
3403 // | |
3404 // The CAS() in enter provides for safety and exclusion, while the CAS or | |
3405 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking | |
3406 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. | |
3407 // We detect and recover from stranding with timers. | |
3408 // | |
3409 // If a thread transiently strands it'll park until (a) another | |
3410 // thread acquires the lock and then drops the lock, at which time the | |
3411 // exiting thread will notice and unpark the stranded thread, or, (b) | |
3412 // the timer expires. If the lock is high traffic then the stranding latency | |
3413 // will be low due to (a). If the lock is low traffic then the odds of | |
3414 // stranding are lower, although the worst-case stranding latency | |
3415 // is longer. Critically, we don't want to put excessive load in the | |
3416 // platform's timer subsystem. We want to minimize both the timer injection | |
3417 // rate (timers created/sec) as well as the number of timers active at | |
3418 // any one time. (more precisely, we want to minimize timer-seconds, which is | |
3419 // the integral of the # of active timers at any instant over time). | |
3420 // Both impinge on OS scalability. Given that, at most one thread parked on | |
3421 // a monitor will use a timer. | |
3422 | |
3423 void ATTR ObjectMonitor::exit(TRAPS) { | |
3424 Thread * Self = THREAD ; | |
3425 if (THREAD != _owner) { | |
3426 if (THREAD->is_lock_owned((address) _owner)) { | |
3427 // Transmute _owner from a BasicLock pointer to a Thread address. | |
3428 // We don't need to hold _mutex for this transition. | |
3429 // Non-null to Non-null is safe as long as all readers can | |
3430 // tolerate either flavor. | |
3431 assert (_recursions == 0, "invariant") ; | |
3432 _owner = THREAD ; | |
3433 _recursions = 0 ; | |
3434 OwnerIsThread = 1 ; | |
3435 } else { | |
3436 // NOTE: we need to handle unbalanced monitor enter/exit | |
3437 // in native code by throwing an exception. | |
3438 // TODO: Throw an IllegalMonitorStateException ? | |
3439 TEVENT (Exit - Throw IMSX) ; | |
3440 assert(false, "Non-balanced monitor enter/exit!"); | |
3441 if (false) { | |
3442 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); | |
3443 } | |
3444 return; | |
3445 } | |
3446 } | |
3447 | |
3448 if (_recursions != 0) { | |
3449 _recursions--; // this is simple recursive enter | |
3450 TEVENT (Inflated exit - recursive) ; | |
3451 return ; | |
3452 } | |
3453 | |
3454 // Invariant: after setting Responsible=null an thread must execute | |
3455 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. | |
3456 if ((SyncFlags & 4) == 0) { | |
3457 _Responsible = NULL ; | |
3458 } | |
3459 | |
3460 for (;;) { | |
3461 assert (THREAD == _owner, "invariant") ; | |
3462 | |
3463 // Fast-path monitor exit: | |
3464 // | |
3465 // Observe the Dekker/Lamport duality: | |
3466 // A thread in ::exit() executes: | |
3467 // ST Owner=null; MEMBAR; LD EntryList|cxq. | |
3468 // A thread in the contended ::enter() path executes the complementary: | |
3469 // ST EntryList|cxq = nonnull; MEMBAR; LD Owner. | |
3470 // | |
3471 // Note that there's a benign race in the exit path. We can drop the | |
3472 // lock, another thread can reacquire the lock immediately, and we can | |
3473 // then wake a thread unnecessarily (yet another flavor of futile wakeup). | |
3474 // This is benign, and we've structured the code so the windows are short | |
3475 // and the frequency of such futile wakeups is low. | |
3476 // | |
3477 // We could eliminate the race by encoding both the "LOCKED" state and | |
3478 // the queue head in a single word. Exit would then use either CAS to | |
3479 // clear the LOCKED bit/byte. This precludes the desirable 1-0 optimization, | |
3480 // however. | |
3481 // | |
3482 // Possible fast-path ::exit() optimization: | |
3483 // The current fast-path exit implementation fetches both cxq and EntryList. | |
3484 // See also i486.ad fast_unlock(). Testing has shown that two LDs | |
3485 // isn't measurably slower than a single LD on any platforms. | |
3486 // Still, we could reduce the 2 LDs to one or zero by one of the following: | |
3487 // | |
3488 // - Use _count instead of cxq|EntryList | |
3489 // We intend to eliminate _count, however, when we switch | |
3490 // to on-the-fly deflation in ::exit() as is used in | |
3491 // Metalocks and RelaxedLocks. | |
3492 // | |
3493 // - Establish the invariant that cxq == null implies EntryList == null. | |
3494 // set cxq == EMPTY (1) to encode the state where cxq is empty | |
3495 // by EntryList != null. EMPTY is a distinguished value. | |
3496 // The fast-path exit() would fetch cxq but not EntryList. | |
3497 // | |
3498 // - Encode succ as follows: | |
3499 // succ = t : Thread t is the successor -- t is ready or is spinning. | |
3500 // Exiting thread does not need to wake a successor. | |
3501 // succ = 0 : No successor required -> (EntryList|cxq) == null | |
3502 // Exiting thread does not need to wake a successor | |
3503 // succ = 1 : Successor required -> (EntryList|cxq) != null and | |
3504 // logically succ == null. | |
3505 // Exiting thread must wake a successor. | |
3506 // | |
3507 // The 1-1 fast-exit path would appear as : | |
3508 // _owner = null ; membar ; | |
3509 // if (_succ == 1 && CAS (&_owner, null, Self) == null) goto SlowPath | |
3510 // goto FastPathDone ; | |
3511 // | |
3512 // and the 1-0 fast-exit path would appear as: | |
3513 // if (_succ == 1) goto SlowPath | |
3514 // Owner = null ; | |
3515 // goto FastPathDone | |
3516 // | |
3517 // - Encode the LSB of _owner as 1 to indicate that exit() | |
3518 // must use the slow-path and make a successor ready. | |
3519 // (_owner & 1) == 0 IFF succ != null || (EntryList|cxq) == null | |
3520 // (_owner & 1) == 0 IFF succ == null && (EntryList|cxq) != null (obviously) | |
3521 // The 1-0 fast exit path would read: | |
3522 // if (_owner != Self) goto SlowPath | |
3523 // _owner = null | |
3524 // goto FastPathDone | |
3525 | |
3526 if (Knob_ExitPolicy == 0) { | |
3527 // release semantics: prior loads and stores from within the critical section | |
3528 // must not float (reorder) past the following store that drops the lock. | |
3529 // On SPARC that requires MEMBAR #loadstore|#storestore. | |
3530 // But of course in TSO #loadstore|#storestore is not required. | |
3531 // I'd like to write one of the following: | |
3532 // A. OrderAccess::release() ; _owner = NULL | |
3533 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; | |
3534 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both | |
3535 // store into a _dummy variable. That store is not needed, but can result | |
3536 // in massive wasteful coherency traffic on classic SMP systems. | |
3537 // Instead, I use release_store(), which is implemented as just a simple | |
3538 // ST on x64, x86 and SPARC. | |
3539 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock | |
3540 OrderAccess::storeload() ; // See if we need to wake a successor | |
3541 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { | |
3542 TEVENT (Inflated exit - simple egress) ; | |
3543 return ; | |
3544 } | |
3545 TEVENT (Inflated exit - complex egress) ; | |
3546 | |
3547 // Normally the exiting thread is responsible for ensuring succession, | |
3548 // but if other successors are ready or other entering threads are spinning | |
3549 // then this thread can simply store NULL into _owner and exit without | |
3550 // waking a successor. The existence of spinners or ready successors | |
3551 // guarantees proper succession (liveness). Responsibility passes to the | |
3552 // ready or running successors. The exiting thread delegates the duty. | |
3553 // More precisely, if a successor already exists this thread is absolved | |
3554 // of the responsibility of waking (unparking) one. | |
3555 // | |
3556 // The _succ variable is critical to reducing futile wakeup frequency. | |
3557 // _succ identifies the "heir presumptive" thread that has been made | |
3558 // ready (unparked) but that has not yet run. We need only one such | |
3559 // successor thread to guarantee progress. | |
3560 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf | |
3561 // section 3.3 "Futile Wakeup Throttling" for details. | |
3562 // | |
3563 // Note that spinners in Enter() also set _succ non-null. | |
3564 // In the current implementation spinners opportunistically set | |
3565 // _succ so that exiting threads might avoid waking a successor. | |
3566 // Another less appealing alternative would be for the exiting thread | |
3567 // to drop the lock and then spin briefly to see if a spinner managed | |
3568 // to acquire the lock. If so, the exiting thread could exit | |
3569 // immediately without waking a successor, otherwise the exiting | |
3570 // thread would need to dequeue and wake a successor. | |
3571 // (Note that we'd need to make the post-drop spin short, but no | |
3572 // shorter than the worst-case round-trip cache-line migration time. | |
3573 // The dropped lock needs to become visible to the spinner, and then | |
3574 // the acquisition of the lock by the spinner must become visible to | |
3575 // the exiting thread). | |
3576 // | |
3577 | |
3578 // It appears that an heir-presumptive (successor) must be made ready. | |
3579 // Only the current lock owner can manipulate the EntryList or | |
3580 // drain _cxq, so we need to reacquire the lock. If we fail | |
3581 // to reacquire the lock the responsibility for ensuring succession | |
3582 // falls to the new owner. | |
3583 // | |
3584 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { | |
3585 return ; | |
3586 } | |
3587 TEVENT (Exit - Reacquired) ; | |
3588 } else { | |
3589 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { | |
3590 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock | |
3591 OrderAccess::storeload() ; | |
3592 // Ratify the previously observed values. | |
3593 if (_cxq == NULL || _succ != NULL) { | |
3594 TEVENT (Inflated exit - simple egress) ; | |
3595 return ; | |
3596 } | |
3597 | |
3598 // inopportune interleaving -- the exiting thread (this thread) | |
3599 // in the fast-exit path raced an entering thread in the slow-enter | |
3600 // path. | |
3601 // We have two choices: | |
3602 // A. Try to reacquire the lock. | |
3603 // If the CAS() fails return immediately, otherwise | |
3604 // we either restart/rerun the exit operation, or simply | |
3605 // fall-through into the code below which wakes a successor. | |
3606 // B. If the elements forming the EntryList|cxq are TSM | |
3607 // we could simply unpark() the lead thread and return | |
3608 // without having set _succ. | |
3609 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { | |
3610 TEVENT (Inflated exit - reacquired succeeded) ; | |
3611 return ; | |
3612 } | |
3613 TEVENT (Inflated exit - reacquired failed) ; | |
3614 } else { | |
3615 TEVENT (Inflated exit - complex egress) ; | |
3616 } | |
3617 } | |
3618 | |
3619 guarantee (_owner == THREAD, "invariant") ; | |
3620 | |
3621 // Select an appropriate successor ("heir presumptive") from the EntryList | |
3622 // and make it ready. Generally we just wake the head of EntryList . | |
3623 // There's no algorithmic constraint that we use the head - it's just | |
3624 // a policy decision. Note that the thread at head of the EntryList | |
3625 // remains at the head until it acquires the lock. This means we'll | |
3626 // repeatedly wake the same thread until it manages to grab the lock. | |
3627 // This is generally a good policy - if we're seeing lots of futile wakeups | |
3628 // at least we're waking/rewaking a thread that's like to be hot or warm | |
3629 // (have residual D$ and TLB affinity). | |
3630 // | |
3631 // "Wakeup locality" optimization: | |
3632 // http://j2se.east/~dice/PERSIST/040825-WakeLocality.txt | |
3633 // In the future we'll try to bias the selection mechanism | |
3634 // to preferentially pick a thread that recently ran on | |
3635 // a processor element that shares cache with the CPU on which | |
3636 // the exiting thread is running. We need access to Solaris' | |
3637 // schedctl.sc_cpu to make that work. | |
3638 // | |
3639 ObjectWaiter * w = NULL ; | |
3640 int QMode = Knob_QMode ; | |
3641 | |
3642 if (QMode == 2 && _cxq != NULL) { | |
3643 // QMode == 2 : cxq has precedence over EntryList. | |
3644 // Try to directly wake a successor from the cxq. | |
3645 // If successful, the successor will need to unlink itself from cxq. | |
3646 w = _cxq ; | |
3647 assert (w != NULL, "invariant") ; | |
3648 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; | |
3649 ExitEpilog (Self, w) ; | |
3650 return ; | |
3651 } | |
3652 | |
3653 if (QMode == 3 && _cxq != NULL) { | |
3654 // Aggressively drain cxq into EntryList at the first opportunity. | |
3655 // This policy ensure that recently-run threads live at the head of EntryList. | |
3656 // Drain _cxq into EntryList - bulk transfer. | |
3657 // First, detach _cxq. | |
3658 // The following loop is tantamount to: w = swap (&cxq, NULL) | |
3659 w = _cxq ; | |
3660 for (;;) { | |
3661 assert (w != NULL, "Invariant") ; | |
3662 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; | |
3663 if (u == w) break ; | |
3664 w = u ; | |
3665 } | |
3666 assert (w != NULL , "invariant") ; | |
3667 | |
3668 ObjectWaiter * q = NULL ; | |
3669 ObjectWaiter * p ; | |
3670 for (p = w ; p != NULL ; p = p->_next) { | |
3671 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; | |
3672 p->TState = ObjectWaiter::TS_ENTER ; | |
3673 p->_prev = q ; | |
3674 q = p ; | |
3675 } | |
3676 | |
3677 // Append the RATs to the EntryList | |
3678 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. | |
3679 ObjectWaiter * Tail ; | |
3680 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; | |
3681 if (Tail == NULL) { | |
3682 _EntryList = w ; | |
3683 } else { | |
3684 Tail->_next = w ; | |
3685 w->_prev = Tail ; | |
3686 } | |
3687 | |
3688 // Fall thru into code that tries to wake a successor from EntryList | |
3689 } | |
3690 | |
3691 if (QMode == 4 && _cxq != NULL) { | |
3692 // Aggressively drain cxq into EntryList at the first opportunity. | |
3693 // This policy ensure that recently-run threads live at the head of EntryList. | |
3694 | |
3695 // Drain _cxq into EntryList - bulk transfer. | |
3696 // First, detach _cxq. | |
3697 // The following loop is tantamount to: w = swap (&cxq, NULL) | |
3698 w = _cxq ; | |
3699 for (;;) { | |
3700 assert (w != NULL, "Invariant") ; | |
3701 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; | |
3702 if (u == w) break ; | |
3703 w = u ; | |
3704 } | |
3705 assert (w != NULL , "invariant") ; | |
3706 | |
3707 ObjectWaiter * q = NULL ; | |
3708 ObjectWaiter * p ; | |
3709 for (p = w ; p != NULL ; p = p->_next) { | |
3710 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; | |
3711 p->TState = ObjectWaiter::TS_ENTER ; | |
3712 p->_prev = q ; | |
3713 q = p ; | |
3714 } | |
3715 | |
3716 // Prepend the RATs to the EntryList | |
3717 if (_EntryList != NULL) { | |
3718 q->_next = _EntryList ; | |
3719 _EntryList->_prev = q ; | |
3720 } | |
3721 _EntryList = w ; | |
3722 | |
3723 // Fall thru into code that tries to wake a successor from EntryList | |
3724 } | |
3725 | |
3726 w = _EntryList ; | |
3727 if (w != NULL) { | |
3728 // I'd like to write: guarantee (w->_thread != Self). | |
3729 // But in practice an exiting thread may find itself on the EntryList. | |
3730 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and | |
3731 // then calls exit(). Exit release the lock by setting O._owner to NULL. | |
3732 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The | |
3733 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then | |
3734 // release the lock "O". T2 resumes immediately after the ST of null into | |
3735 // _owner, above. T2 notices that the EntryList is populated, so it | |
3736 // reacquires the lock and then finds itself on the EntryList. | |
3737 // Given all that, we have to tolerate the circumstance where "w" is | |
3738 // associated with Self. | |
3739 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
3740 ExitEpilog (Self, w) ; | |
3741 return ; | |
3742 } | |
3743 | |
3744 // If we find that both _cxq and EntryList are null then just | |
3745 // re-run the exit protocol from the top. | |
3746 w = _cxq ; | |
3747 if (w == NULL) continue ; | |
3748 | |
3749 // Drain _cxq into EntryList - bulk transfer. | |
3750 // First, detach _cxq. | |
3751 // The following loop is tantamount to: w = swap (&cxq, NULL) | |
3752 for (;;) { | |
3753 assert (w != NULL, "Invariant") ; | |
3754 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; | |
3755 if (u == w) break ; | |
3756 w = u ; | |
3757 } | |
3758 TEVENT (Inflated exit - drain cxq into EntryList) ; | |
3759 | |
3760 assert (w != NULL , "invariant") ; | |
3761 assert (_EntryList == NULL , "invariant") ; | |
3762 | |
3763 // Convert the LIFO SLL anchored by _cxq into a DLL. | |
3764 // The list reorganization step operates in O(LENGTH(w)) time. | |
3765 // It's critical that this step operate quickly as | |
3766 // "Self" still holds the outer-lock, restricting parallelism | |
3767 // and effectively lengthening the critical section. | |
3768 // Invariant: s chases t chases u. | |
3769 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so | |
3770 // we have faster access to the tail. | |
3771 | |
3772 if (QMode == 1) { | |
3773 // QMode == 1 : drain cxq to EntryList, reversing order | |
3774 // We also reverse the order of the list. | |
3775 ObjectWaiter * s = NULL ; | |
3776 ObjectWaiter * t = w ; | |
3777 ObjectWaiter * u = NULL ; | |
3778 while (t != NULL) { | |
3779 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; | |
3780 t->TState = ObjectWaiter::TS_ENTER ; | |
3781 u = t->_next ; | |
3782 t->_prev = u ; | |
3783 t->_next = s ; | |
3784 s = t; | |
3785 t = u ; | |
3786 } | |
3787 _EntryList = s ; | |
3788 assert (s != NULL, "invariant") ; | |
3789 } else { | |
3790 // QMode == 0 or QMode == 2 | |
3791 _EntryList = w ; | |
3792 ObjectWaiter * q = NULL ; | |
3793 ObjectWaiter * p ; | |
3794 for (p = w ; p != NULL ; p = p->_next) { | |
3795 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; | |
3796 p->TState = ObjectWaiter::TS_ENTER ; | |
3797 p->_prev = q ; | |
3798 q = p ; | |
3799 } | |
3800 } | |
3801 | |
3802 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL | |
3803 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). | |
3804 | |
3805 // See if we can abdicate to a spinner instead of waking a thread. | |
3806 // A primary goal of the implementation is to reduce the | |
3807 // context-switch rate. | |
3808 if (_succ != NULL) continue; | |
3809 | |
3810 w = _EntryList ; | |
3811 if (w != NULL) { | |
3812 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
3813 ExitEpilog (Self, w) ; | |
3814 return ; | |
3815 } | |
3816 } | |
3817 } | |
3818 // complete_exit exits a lock returning recursion count | |
3819 // complete_exit/reenter operate as a wait without waiting | |
3820 // complete_exit requires an inflated monitor | |
3821 // The _owner field is not always the Thread addr even with an | |
3822 // inflated monitor, e.g. the monitor can be inflated by a non-owning | |
3823 // thread due to contention. | |
3824 intptr_t ObjectMonitor::complete_exit(TRAPS) { | |
3825 Thread * const Self = THREAD; | |
3826 assert(Self->is_Java_thread(), "Must be Java thread!"); | |
3827 JavaThread *jt = (JavaThread *)THREAD; | |
3828 | |
3829 DeferredInitialize(); | |
3830 | |
3831 if (THREAD != _owner) { | |
3832 if (THREAD->is_lock_owned ((address)_owner)) { | |
3833 assert(_recursions == 0, "internal state error"); | |
3834 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ | |
3835 _recursions = 0 ; | |
3836 OwnerIsThread = 1 ; | |
3837 } | |
3838 } | |
3839 | |
3840 guarantee(Self == _owner, "complete_exit not owner"); | |
3841 intptr_t save = _recursions; // record the old recursion count | |
3842 _recursions = 0; // set the recursion level to be 0 | |
3843 exit (Self) ; // exit the monitor | |
3844 guarantee (_owner != Self, "invariant"); | |
3845 return save; | |
3846 } | |
3847 | |
3848 // reenter() enters a lock and sets recursion count | |
3849 // complete_exit/reenter operate as a wait without waiting | |
3850 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { | |
3851 Thread * const Self = THREAD; | |
3852 assert(Self->is_Java_thread(), "Must be Java thread!"); | |
3853 JavaThread *jt = (JavaThread *)THREAD; | |
3854 | |
3855 guarantee(_owner != Self, "reenter already owner"); | |
3856 enter (THREAD); // enter the monitor | |
3857 guarantee (_recursions == 0, "reenter recursion"); | |
3858 _recursions = recursions; | |
3859 return; | |
3860 } | |
3861 | |
3862 // Note: a subset of changes to ObjectMonitor::wait() | |
3863 // will need to be replicated in complete_exit above | |
3864 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { | |
3865 Thread * const Self = THREAD ; | |
3866 assert(Self->is_Java_thread(), "Must be Java thread!"); | |
3867 JavaThread *jt = (JavaThread *)THREAD; | |
3868 | |
3869 DeferredInitialize () ; | |
3870 | |
3871 // Throw IMSX or IEX. | |
3872 CHECK_OWNER(); | |
3873 | |
3874 // check for a pending interrupt | |
3875 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { | |
3876 // post monitor waited event. Note that this is past-tense, we are done waiting. | |
3877 if (JvmtiExport::should_post_monitor_waited()) { | |
3878 // Note: 'false' parameter is passed here because the | |
3879 // wait was not timed out due to thread interrupt. | |
3880 JvmtiExport::post_monitor_waited(jt, this, false); | |
3881 } | |
3882 TEVENT (Wait - Throw IEX) ; | |
3883 THROW(vmSymbols::java_lang_InterruptedException()); | |
3884 return ; | |
3885 } | |
3886 TEVENT (Wait) ; | |
3887 | |
3888 assert (Self->_Stalled == 0, "invariant") ; | |
3889 Self->_Stalled = intptr_t(this) ; | |
3890 jt->set_current_waiting_monitor(this); | |
3891 | |
3892 // create a node to be put into the queue | |
3893 // Critically, after we reset() the event but prior to park(), we must check | |
3894 // for a pending interrupt. | |
3895 ObjectWaiter node(Self); | |
3896 node.TState = ObjectWaiter::TS_WAIT ; | |
3897 Self->_ParkEvent->reset() ; | |
3898 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag | |
3899 | |
3900 // Enter the waiting queue, which is a circular doubly linked list in this case | |
3901 // but it could be a priority queue or any data structure. | |
3902 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only | |
3903 // by the the owner of the monitor *except* in the case where park() | |
3904 // returns because of a timeout of interrupt. Contention is exceptionally rare | |
3905 // so we use a simple spin-lock instead of a heavier-weight blocking lock. | |
3906 | |
3907 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; | |
3908 AddWaiter (&node) ; | |
3909 Thread::SpinRelease (&_WaitSetLock) ; | |
3910 | |
3911 if ((SyncFlags & 4) == 0) { | |
3912 _Responsible = NULL ; | |
3913 } | |
3914 intptr_t save = _recursions; // record the old recursion count | |
3915 _waiters++; // increment the number of waiters | |
3916 _recursions = 0; // set the recursion level to be 1 | |
3917 exit (Self) ; // exit the monitor | |
3918 guarantee (_owner != Self, "invariant") ; | |
3919 | |
3920 // As soon as the ObjectMonitor's ownership is dropped in the exit() | |
3921 // call above, another thread can enter() the ObjectMonitor, do the | |
3922 // notify(), and exit() the ObjectMonitor. If the other thread's | |
3923 // exit() call chooses this thread as the successor and the unpark() | |
3924 // call happens to occur while this thread is posting a | |
3925 // MONITOR_CONTENDED_EXIT event, then we run the risk of the event | |
3926 // handler using RawMonitors and consuming the unpark(). | |
3927 // | |
3928 // To avoid the problem, we re-post the event. This does no harm | |
3929 // even if the original unpark() was not consumed because we are the | |
3930 // chosen successor for this monitor. | |
3931 if (node._notified != 0 && _succ == Self) { | |
3932 node._event->unpark(); | |
3933 } | |
3934 | |
3935 // The thread is on the WaitSet list - now park() it. | |
3936 // On MP systems it's conceivable that a brief spin before we park | |
3937 // could be profitable. | |
3938 // | |
3939 // TODO-FIXME: change the following logic to a loop of the form | |
3940 // while (!timeout && !interrupted && _notified == 0) park() | |
3941 | |
3942 int ret = OS_OK ; | |
3943 int WasNotified = 0 ; | |
3944 { // State transition wrappers | |
3945 OSThread* osthread = Self->osthread(); | |
3946 OSThreadWaitState osts(osthread, true); | |
3947 { | |
3948 ThreadBlockInVM tbivm(jt); | |
3949 // Thread is in thread_blocked state and oop access is unsafe. | |
3950 jt->set_suspend_equivalent(); | |
3951 | |
3952 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { | |
3953 // Intentionally empty | |
3954 } else | |
3955 if (node._notified == 0) { | |
3956 if (millis <= 0) { | |
3957 Self->_ParkEvent->park () ; | |
3958 } else { | |
3959 ret = Self->_ParkEvent->park (millis) ; | |
3960 } | |
3961 } | |
3962 | |
3963 // were we externally suspended while we were waiting? | |
3964 if (ExitSuspendEquivalent (jt)) { | |
3965 // TODO-FIXME: add -- if succ == Self then succ = null. | |
3966 jt->java_suspend_self(); | |
3967 } | |
3968 | |
3969 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm | |
3970 | |
3971 | |
3972 // Node may be on the WaitSet, the EntryList (or cxq), or in transition | |
3973 // from the WaitSet to the EntryList. | |
3974 // See if we need to remove Node from the WaitSet. | |
3975 // We use double-checked locking to avoid grabbing _WaitSetLock | |
3976 // if the thread is not on the wait queue. | |
3977 // | |
3978 // Note that we don't need a fence before the fetch of TState. | |
3979 // In the worst case we'll fetch a old-stale value of TS_WAIT previously | |
3980 // written by the is thread. (perhaps the fetch might even be satisfied | |
3981 // by a look-aside into the processor's own store buffer, although given | |
3982 // the length of the code path between the prior ST and this load that's | |
3983 // highly unlikely). If the following LD fetches a stale TS_WAIT value | |
3984 // then we'll acquire the lock and then re-fetch a fresh TState value. | |
3985 // That is, we fail toward safety. | |
3986 | |
3987 if (node.TState == ObjectWaiter::TS_WAIT) { | |
3988 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; | |
3989 if (node.TState == ObjectWaiter::TS_WAIT) { | |
3990 DequeueSpecificWaiter (&node) ; // unlink from WaitSet | |
3991 assert(node._notified == 0, "invariant"); | |
3992 node.TState = ObjectWaiter::TS_RUN ; | |
3993 } | |
3994 Thread::SpinRelease (&_WaitSetLock) ; | |
3995 } | |
3996 | |
3997 // The thread is now either on off-list (TS_RUN), | |
3998 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). | |
3999 // The Node's TState variable is stable from the perspective of this thread. | |
4000 // No other threads will asynchronously modify TState. | |
4001 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; | |
4002 OrderAccess::loadload() ; | |
4003 if (_succ == Self) _succ = NULL ; | |
4004 WasNotified = node._notified ; | |
4005 | |
4006 // Reentry phase -- reacquire the monitor. | |
4007 // re-enter contended monitor after object.wait(). | |
4008 // retain OBJECT_WAIT state until re-enter successfully completes | |
4009 // Thread state is thread_in_vm and oop access is again safe, | |
4010 // although the raw address of the object may have changed. | |
4011 // (Don't cache naked oops over safepoints, of course). | |
4012 | |
4013 // post monitor waited event. Note that this is past-tense, we are done waiting. | |
4014 if (JvmtiExport::should_post_monitor_waited()) { | |
4015 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); | |
4016 } | |
4017 OrderAccess::fence() ; | |
4018 | |
4019 assert (Self->_Stalled != 0, "invariant") ; | |
4020 Self->_Stalled = 0 ; | |
4021 | |
4022 assert (_owner != Self, "invariant") ; | |
4023 ObjectWaiter::TStates v = node.TState ; | |
4024 if (v == ObjectWaiter::TS_RUN) { | |
4025 enter (Self) ; | |
4026 } else { | |
4027 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; | |
4028 ReenterI (Self, &node) ; | |
4029 node.wait_reenter_end(this); | |
4030 } | |
4031 | |
4032 // Self has reacquired the lock. | |
4033 // Lifecycle - the node representing Self must not appear on any queues. | |
4034 // Node is about to go out-of-scope, but even if it were immortal we wouldn't | |
4035 // want residual elements associated with this thread left on any lists. | |
4036 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; | |
4037 assert (_owner == Self, "invariant") ; | |
4038 assert (_succ != Self , "invariant") ; | |
4039 } // OSThreadWaitState() | |
4040 | |
4041 jt->set_current_waiting_monitor(NULL); | |
4042 | |
4043 guarantee (_recursions == 0, "invariant") ; | |
4044 _recursions = save; // restore the old recursion count | |
4045 _waiters--; // decrement the number of waiters | |
4046 | |
4047 // Verify a few postconditions | |
4048 assert (_owner == Self , "invariant") ; | |
4049 assert (_succ != Self , "invariant") ; | |
4050 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; | |
4051 | |
4052 if (SyncFlags & 32) { | |
4053 OrderAccess::fence() ; | |
4054 } | |
4055 | |
4056 // check if the notification happened | |
4057 if (!WasNotified) { | |
4058 // no, it could be timeout or Thread.interrupt() or both | |
4059 // check for interrupt event, otherwise it is timeout | |
4060 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { | |
4061 TEVENT (Wait - throw IEX from epilog) ; | |
4062 THROW(vmSymbols::java_lang_InterruptedException()); | |
4063 } | |
4064 } | |
4065 | |
4066 // NOTE: Spurious wake up will be consider as timeout. | |
4067 // Monitor notify has precedence over thread interrupt. | |
4068 } | |
4069 | |
4070 | |
4071 // Consider: | |
4072 // If the lock is cool (cxq == null && succ == null) and we're on an MP system | |
4073 // then instead of transferring a thread from the WaitSet to the EntryList | |
4074 // we might just dequeue a thread from the WaitSet and directly unpark() it. | |
4075 | |
4076 void ObjectMonitor::notify(TRAPS) { | |
4077 CHECK_OWNER(); | |
4078 if (_WaitSet == NULL) { | |
4079 TEVENT (Empty-Notify) ; | |
4080 return ; | |
4081 } | |
4082 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); | |
4083 | |
4084 int Policy = Knob_MoveNotifyee ; | |
4085 | |
4086 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; | |
4087 ObjectWaiter * iterator = DequeueWaiter() ; | |
4088 if (iterator != NULL) { | |
4089 TEVENT (Notify1 - Transfer) ; | |
4090 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; | |
4091 guarantee (iterator->_notified == 0, "invariant") ; | |
4092 // Disposition - what might we do with iterator ? | |
4093 // a. add it directly to the EntryList - either tail or head. | |
4094 // b. push it onto the front of the _cxq. | |
4095 // For now we use (a). | |
4096 if (Policy != 4) { | |
4097 iterator->TState = ObjectWaiter::TS_ENTER ; | |
4098 } | |
4099 iterator->_notified = 1 ; | |
4100 | |
4101 ObjectWaiter * List = _EntryList ; | |
4102 if (List != NULL) { | |
4103 assert (List->_prev == NULL, "invariant") ; | |
4104 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
4105 assert (List != iterator, "invariant") ; | |
4106 } | |
4107 | |
4108 if (Policy == 0) { // prepend to EntryList | |
4109 if (List == NULL) { | |
4110 iterator->_next = iterator->_prev = NULL ; | |
4111 _EntryList = iterator ; | |
4112 } else { | |
4113 List->_prev = iterator ; | |
4114 iterator->_next = List ; | |
4115 iterator->_prev = NULL ; | |
4116 _EntryList = iterator ; | |
4117 } | |
4118 } else | |
4119 if (Policy == 1) { // append to EntryList | |
4120 if (List == NULL) { | |
4121 iterator->_next = iterator->_prev = NULL ; | |
4122 _EntryList = iterator ; | |
4123 } else { | |
4124 // CONSIDER: finding the tail currently requires a linear-time walk of | |
4125 // the EntryList. We can make tail access constant-time by converting to | |
4126 // a CDLL instead of using our current DLL. | |
4127 ObjectWaiter * Tail ; | |
4128 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; | |
4129 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; | |
4130 Tail->_next = iterator ; | |
4131 iterator->_prev = Tail ; | |
4132 iterator->_next = NULL ; | |
4133 } | |
4134 } else | |
4135 if (Policy == 2) { // prepend to cxq | |
4136 // prepend to cxq | |
4137 if (List == NULL) { | |
4138 iterator->_next = iterator->_prev = NULL ; | |
4139 _EntryList = iterator ; | |
4140 } else { | |
4141 iterator->TState = ObjectWaiter::TS_CXQ ; | |
4142 for (;;) { | |
4143 ObjectWaiter * Front = _cxq ; | |
4144 iterator->_next = Front ; | |
4145 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { | |
4146 break ; | |
4147 } | |
4148 } | |
4149 } | |
4150 } else | |
4151 if (Policy == 3) { // append to cxq | |
4152 iterator->TState = ObjectWaiter::TS_CXQ ; | |
4153 for (;;) { | |
4154 ObjectWaiter * Tail ; | |
4155 Tail = _cxq ; | |
4156 if (Tail == NULL) { | |
4157 iterator->_next = NULL ; | |
4158 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { | |
4159 break ; | |
4160 } | |
4161 } else { | |
4162 while (Tail->_next != NULL) Tail = Tail->_next ; | |
4163 Tail->_next = iterator ; | |
4164 iterator->_prev = Tail ; | |
4165 iterator->_next = NULL ; | |
4166 break ; | |
4167 } | |
4168 } | |
4169 } else { | |
4170 ParkEvent * ev = iterator->_event ; | |
4171 iterator->TState = ObjectWaiter::TS_RUN ; | |
4172 OrderAccess::fence() ; | |
4173 ev->unpark() ; | |
4174 } | |
4175 | |
4176 if (Policy < 4) { | |
4177 iterator->wait_reenter_begin(this); | |
4178 } | |
4179 | |
4180 // _WaitSetLock protects the wait queue, not the EntryList. We could | |
4181 // move the add-to-EntryList operation, above, outside the critical section | |
4182 // protected by _WaitSetLock. In practice that's not useful. With the | |
4183 // exception of wait() timeouts and interrupts the monitor owner | |
4184 // is the only thread that grabs _WaitSetLock. There's almost no contention | |
4185 // on _WaitSetLock so it's not profitable to reduce the length of the | |
4186 // critical section. | |
4187 } | |
4188 | |
4189 Thread::SpinRelease (&_WaitSetLock) ; | |
4190 | |
4191 if (iterator != NULL && ObjectSynchronizer::_sync_Notifications != NULL) { | |
4192 ObjectSynchronizer::_sync_Notifications->inc() ; | |
4193 } | |
4194 } | |
4195 | |
4196 | |
4197 void ObjectMonitor::notifyAll(TRAPS) { | |
4198 CHECK_OWNER(); | |
4199 ObjectWaiter* iterator; | |
4200 if (_WaitSet == NULL) { | |
4201 TEVENT (Empty-NotifyAll) ; | |
4202 return ; | |
4203 } | |
4204 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); | |
4205 | |
4206 int Policy = Knob_MoveNotifyee ; | |
4207 int Tally = 0 ; | |
4208 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; | |
4209 | |
4210 for (;;) { | |
4211 iterator = DequeueWaiter () ; | |
4212 if (iterator == NULL) break ; | |
4213 TEVENT (NotifyAll - Transfer1) ; | |
4214 ++Tally ; | |
4215 | |
4216 // Disposition - what might we do with iterator ? | |
4217 // a. add it directly to the EntryList - either tail or head. | |
4218 // b. push it onto the front of the _cxq. | |
4219 // For now we use (a). | |
4220 // | |
4221 // TODO-FIXME: currently notifyAll() transfers the waiters one-at-a-time from the waitset | |
4222 // to the EntryList. This could be done more efficiently with a single bulk transfer, | |
4223 // but in practice it's not time-critical. Beware too, that in prepend-mode we invert the | |
4224 // order of the waiters. Lets say that the waitset is "ABCD" and the EntryList is "XYZ". | |
4225 // After a notifyAll() in prepend mode the waitset will be empty and the EntryList will | |
4226 // be "DCBAXYZ". | |
4227 | |
4228 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; | |
4229 guarantee (iterator->_notified == 0, "invariant") ; | |
4230 iterator->_notified = 1 ; | |
4231 if (Policy != 4) { | |
4232 iterator->TState = ObjectWaiter::TS_ENTER ; | |
4233 } | |
4234 | |
4235 ObjectWaiter * List = _EntryList ; | |
4236 if (List != NULL) { | |
4237 assert (List->_prev == NULL, "invariant") ; | |
4238 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
4239 assert (List != iterator, "invariant") ; | |
4240 } | |
4241 | |
4242 if (Policy == 0) { // prepend to EntryList | |
4243 if (List == NULL) { | |
4244 iterator->_next = iterator->_prev = NULL ; | |
4245 _EntryList = iterator ; | |
4246 } else { | |
4247 List->_prev = iterator ; | |
4248 iterator->_next = List ; | |
4249 iterator->_prev = NULL ; | |
4250 _EntryList = iterator ; | |
4251 } | |
4252 } else | |
4253 if (Policy == 1) { // append to EntryList | |
4254 if (List == NULL) { | |
4255 iterator->_next = iterator->_prev = NULL ; | |
4256 _EntryList = iterator ; | |
4257 } else { | |
4258 // CONSIDER: finding the tail currently requires a linear-time walk of | |
4259 // the EntryList. We can make tail access constant-time by converting to | |
4260 // a CDLL instead of using our current DLL. | |
4261 ObjectWaiter * Tail ; | |
4262 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; | |
4263 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; | |
4264 Tail->_next = iterator ; | |
4265 iterator->_prev = Tail ; | |
4266 iterator->_next = NULL ; | |
4267 } | |
4268 } else | |
4269 if (Policy == 2) { // prepend to cxq | |
4270 // prepend to cxq | |
4271 iterator->TState = ObjectWaiter::TS_CXQ ; | |
4272 for (;;) { | |
4273 ObjectWaiter * Front = _cxq ; | |
4274 iterator->_next = Front ; | |
4275 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { | |
4276 break ; | |
4277 } | |
4278 } | |
4279 } else | |
4280 if (Policy == 3) { // append to cxq | |
4281 iterator->TState = ObjectWaiter::TS_CXQ ; | |
4282 for (;;) { | |
4283 ObjectWaiter * Tail ; | |
4284 Tail = _cxq ; | |
4285 if (Tail == NULL) { | |
4286 iterator->_next = NULL ; | |
4287 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { | |
4288 break ; | |
4289 } | |
4290 } else { | |
4291 while (Tail->_next != NULL) Tail = Tail->_next ; | |
4292 Tail->_next = iterator ; | |
4293 iterator->_prev = Tail ; | |
4294 iterator->_next = NULL ; | |
4295 break ; | |
4296 } | |
4297 } | |
4298 } else { | |
4299 ParkEvent * ev = iterator->_event ; | |
4300 iterator->TState = ObjectWaiter::TS_RUN ; | |
4301 OrderAccess::fence() ; | |
4302 ev->unpark() ; | |
4303 } | |
4304 | |
4305 if (Policy < 4) { | |
4306 iterator->wait_reenter_begin(this); | |
4307 } | |
4308 | |
4309 // _WaitSetLock protects the wait queue, not the EntryList. We could | |
4310 // move the add-to-EntryList operation, above, outside the critical section | |
4311 // protected by _WaitSetLock. In practice that's not useful. With the | |
4312 // exception of wait() timeouts and interrupts the monitor owner | |
4313 // is the only thread that grabs _WaitSetLock. There's almost no contention | |
4314 // on _WaitSetLock so it's not profitable to reduce the length of the | |
4315 // critical section. | |
4316 } | |
4317 | |
4318 Thread::SpinRelease (&_WaitSetLock) ; | |
4319 | |
4320 if (Tally != 0 && ObjectSynchronizer::_sync_Notifications != NULL) { | |
4321 ObjectSynchronizer::_sync_Notifications->inc(Tally) ; | |
4322 } | |
4323 } | |
4324 | |
4325 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. | |
4326 // TODO-FIXME: remove check_slow() -- it's likely dead. | |
4327 | |
4328 void ObjectMonitor::check_slow(TRAPS) { | |
4329 TEVENT (check_slow - throw IMSX) ; | |
4330 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); | |
4331 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); | |
4332 } | |
4333 | |
4334 | |
4335 // ------------------------------------------------------------------------- | |
4336 // The raw monitor subsystem is entirely distinct from normal | |
4337 // java-synchronization or jni-synchronization. raw monitors are not | |
4338 // associated with objects. They can be implemented in any manner | |
4339 // that makes sense. The original implementors decided to piggy-back | |
4340 // the raw-monitor implementation on the existing Java objectMonitor mechanism. | |
4341 // This flaw needs to fixed. We should reimplement raw monitors as sui-generis. | |
4342 // Specifically, we should not implement raw monitors via java monitors. | |
4343 // Time permitting, we should disentangle and deconvolve the two implementations | |
4344 // and move the resulting raw monitor implementation over to the JVMTI directories. | |
4345 // Ideally, the raw monitor implementation would be built on top of | |
4346 // park-unpark and nothing else. | |
4347 // | |
4348 // raw monitors are used mainly by JVMTI | |
4349 // The raw monitor implementation borrows the ObjectMonitor structure, | |
4350 // but the operators are degenerate and extremely simple. | |
4351 // | |
4352 // Mixed use of a single objectMonitor instance -- as both a raw monitor | |
4353 // and a normal java monitor -- is not permissible. | |
4354 // | |
4355 // Note that we use the single RawMonitor_lock to protect queue operations for | |
4356 // _all_ raw monitors. This is a scalability impediment, but since raw monitor usage | |
4357 // is deprecated and rare, this is not of concern. The RawMonitor_lock can not | |
4358 // be held indefinitely. The critical sections must be short and bounded. | |
4359 // | |
4360 // ------------------------------------------------------------------------- | |
4361 | |
4362 int ObjectMonitor::SimpleEnter (Thread * Self) { | |
4363 for (;;) { | |
4364 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { | |
4365 return OS_OK ; | |
4366 } | |
4367 | |
4368 ObjectWaiter Node (Self) ; | |
4369 Self->_ParkEvent->reset() ; // strictly optional | |
4370 Node.TState = ObjectWaiter::TS_ENTER ; | |
4371 | |
4372 RawMonitor_lock->lock_without_safepoint_check() ; | |
4373 Node._next = _EntryList ; | |
4374 _EntryList = &Node ; | |
4375 OrderAccess::fence() ; | |
4376 if (_owner == NULL && Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { | |
4377 _EntryList = Node._next ; | |
4378 RawMonitor_lock->unlock() ; | |
4379 return OS_OK ; | |
4380 } | |
4381 RawMonitor_lock->unlock() ; | |
4382 while (Node.TState == ObjectWaiter::TS_ENTER) { | |
4383 Self->_ParkEvent->park() ; | |
4384 } | |
4385 } | |
4386 } | |
4387 | |
4388 int ObjectMonitor::SimpleExit (Thread * Self) { | |
4389 guarantee (_owner == Self, "invariant") ; | |
4390 OrderAccess::release_store_ptr (&_owner, NULL) ; | |
4391 OrderAccess::fence() ; | |
4392 if (_EntryList == NULL) return OS_OK ; | |
4393 ObjectWaiter * w ; | |
4394 | |
4395 RawMonitor_lock->lock_without_safepoint_check() ; | |
4396 w = _EntryList ; | |
4397 if (w != NULL) { | |
4398 _EntryList = w->_next ; | |
4399 } | |
4400 RawMonitor_lock->unlock() ; | |
4401 if (w != NULL) { | |
4402 guarantee (w ->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
4403 ParkEvent * ev = w->_event ; | |
4404 w->TState = ObjectWaiter::TS_RUN ; | |
4405 OrderAccess::fence() ; | |
4406 ev->unpark() ; | |
4407 } | |
4408 return OS_OK ; | |
4409 } | |
4410 | |
4411 int ObjectMonitor::SimpleWait (Thread * Self, jlong millis) { | |
4412 guarantee (_owner == Self , "invariant") ; | |
4413 guarantee (_recursions == 0, "invariant") ; | |
4414 | |
4415 ObjectWaiter Node (Self) ; | |
4416 Node._notified = 0 ; | |
4417 Node.TState = ObjectWaiter::TS_WAIT ; | |
4418 | |
4419 RawMonitor_lock->lock_without_safepoint_check() ; | |
4420 Node._next = _WaitSet ; | |
4421 _WaitSet = &Node ; | |
4422 RawMonitor_lock->unlock() ; | |
4423 | |
4424 SimpleExit (Self) ; | |
4425 guarantee (_owner != Self, "invariant") ; | |
4426 | |
4427 int ret = OS_OK ; | |
4428 if (millis <= 0) { | |
4429 Self->_ParkEvent->park(); | |
4430 } else { | |
4431 ret = Self->_ParkEvent->park(millis); | |
4432 } | |
4433 | |
4434 // If thread still resides on the waitset then unlink it. | |
4435 // Double-checked locking -- the usage is safe in this context | |
4436 // as we TState is volatile and the lock-unlock operators are | |
4437 // serializing (barrier-equivalent). | |
4438 | |
4439 if (Node.TState == ObjectWaiter::TS_WAIT) { | |
4440 RawMonitor_lock->lock_without_safepoint_check() ; | |
4441 if (Node.TState == ObjectWaiter::TS_WAIT) { | |
4442 // Simple O(n) unlink, but performance isn't critical here. | |
4443 ObjectWaiter * p ; | |
4444 ObjectWaiter * q = NULL ; | |
4445 for (p = _WaitSet ; p != &Node; p = p->_next) { | |
4446 q = p ; | |
4447 } | |
4448 guarantee (p == &Node, "invariant") ; | |
4449 if (q == NULL) { | |
4450 guarantee (p == _WaitSet, "invariant") ; | |
4451 _WaitSet = p->_next ; | |
4452 } else { | |
4453 guarantee (p == q->_next, "invariant") ; | |
4454 q->_next = p->_next ; | |
4455 } | |
4456 Node.TState = ObjectWaiter::TS_RUN ; | |
4457 } | |
4458 RawMonitor_lock->unlock() ; | |
4459 } | |
4460 | |
4461 guarantee (Node.TState == ObjectWaiter::TS_RUN, "invariant") ; | |
4462 SimpleEnter (Self) ; | |
4463 | |
4464 guarantee (_owner == Self, "invariant") ; | |
4465 guarantee (_recursions == 0, "invariant") ; | |
4466 return ret ; | |
4467 } | |
4468 | |
4469 int ObjectMonitor::SimpleNotify (Thread * Self, bool All) { | |
4470 guarantee (_owner == Self, "invariant") ; | |
4471 if (_WaitSet == NULL) return OS_OK ; | |
4472 | |
4473 // We have two options: | |
4474 // A. Transfer the threads from the WaitSet to the EntryList | |
4475 // B. Remove the thread from the WaitSet and unpark() it. | |
4476 // | |
4477 // We use (B), which is crude and results in lots of futile | |
4478 // context switching. In particular (B) induces lots of contention. | |
4479 | |
4480 ParkEvent * ev = NULL ; // consider using a small auto array ... | |
4481 RawMonitor_lock->lock_without_safepoint_check() ; | |
4482 for (;;) { | |
4483 ObjectWaiter * w = _WaitSet ; | |
4484 if (w == NULL) break ; | |
4485 _WaitSet = w->_next ; | |
4486 if (ev != NULL) { ev->unpark(); ev = NULL; } | |
4487 ev = w->_event ; | |
4488 OrderAccess::loadstore() ; | |
4489 w->TState = ObjectWaiter::TS_RUN ; | |
4490 OrderAccess::storeload(); | |
4491 if (!All) break ; | |
4492 } | |
4493 RawMonitor_lock->unlock() ; | |
4494 if (ev != NULL) ev->unpark(); | |
4495 return OS_OK ; | |
4496 } | |
4497 | |
4498 // Any JavaThread will enter here with state _thread_blocked | |
4499 int ObjectMonitor::raw_enter(TRAPS) { | |
4500 TEVENT (raw_enter) ; | |
4501 void * Contended ; | |
4502 | |
4503 // don't enter raw monitor if thread is being externally suspended, it will | |
4504 // surprise the suspender if a "suspended" thread can still enter monitor | |
4505 JavaThread * jt = (JavaThread *)THREAD; | |
4506 if (THREAD->is_Java_thread()) { | |
4507 jt->SR_lock()->lock_without_safepoint_check(); | |
4508 while (jt->is_external_suspend()) { | |
4509 jt->SR_lock()->unlock(); | |
4510 jt->java_suspend_self(); | |
4511 jt->SR_lock()->lock_without_safepoint_check(); | |
4512 } | |
4513 // guarded by SR_lock to avoid racing with new external suspend requests. | |
4514 Contended = Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) ; | |
4515 jt->SR_lock()->unlock(); | |
4516 } else { | |
4517 Contended = Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) ; | |
4518 } | |
4519 | |
4520 if (Contended == THREAD) { | |
4521 _recursions ++ ; | |
4522 return OM_OK ; | |
4523 } | |
4524 | |
4525 if (Contended == NULL) { | |
4526 guarantee (_owner == THREAD, "invariant") ; | |
4527 guarantee (_recursions == 0, "invariant") ; | |
4528 return OM_OK ; | |
4529 } | |
4530 | |
4531 THREAD->set_current_pending_monitor(this); | |
4532 | |
4533 if (!THREAD->is_Java_thread()) { | |
4534 // No other non-Java threads besides VM thread would acquire | |
4535 // a raw monitor. | |
4536 assert(THREAD->is_VM_thread(), "must be VM thread"); | |
4537 SimpleEnter (THREAD) ; | |
4538 } else { | |
4539 guarantee (jt->thread_state() == _thread_blocked, "invariant") ; | |
4540 for (;;) { | |
4541 jt->set_suspend_equivalent(); | |
4542 // cleared by handle_special_suspend_equivalent_condition() or | |
4543 // java_suspend_self() | |
4544 SimpleEnter (THREAD) ; | |
4545 | |
4546 // were we externally suspended while we were waiting? | |
4547 if (!jt->handle_special_suspend_equivalent_condition()) break ; | |
4548 | |
4549 // This thread was externally suspended | |
4550 // | |
4551 // This logic isn't needed for JVMTI raw monitors, | |
4552 // but doesn't hurt just in case the suspend rules change. This | |
4553 // logic is needed for the ObjectMonitor.wait() reentry phase. | |
4554 // We have reentered the contended monitor, but while we were | |
4555 // waiting another thread suspended us. We don't want to reenter | |
4556 // the monitor while suspended because that would surprise the | |
4557 // thread that suspended us. | |
4558 // | |
4559 // Drop the lock - | |
4560 SimpleExit (THREAD) ; | |
4561 | |
4562 jt->java_suspend_self(); | |
4563 } | |
4564 | |
4565 assert(_owner == THREAD, "Fatal error with monitor owner!"); | |
4566 assert(_recursions == 0, "Fatal error with monitor recursions!"); | |
4567 } | |
4568 | |
4569 THREAD->set_current_pending_monitor(NULL); | |
4570 guarantee (_recursions == 0, "invariant") ; | |
4571 return OM_OK; | |
4572 } | |
4573 | |
4574 // Used mainly for JVMTI raw monitor implementation | |
4575 // Also used for ObjectMonitor::wait(). | |
4576 int ObjectMonitor::raw_exit(TRAPS) { | |
4577 TEVENT (raw_exit) ; | |
4578 if (THREAD != _owner) { | |
4579 return OM_ILLEGAL_MONITOR_STATE; | |
4580 } | |
4581 if (_recursions > 0) { | |
4582 --_recursions ; | |
4583 return OM_OK ; | |
4584 } | |
4585 | |
4586 void * List = _EntryList ; | |
4587 SimpleExit (THREAD) ; | |
4588 | |
4589 return OM_OK; | |
4590 } | |
4591 | |
4592 // Used for JVMTI raw monitor implementation. | |
4593 // All JavaThreads will enter here with state _thread_blocked | |
4594 | |
4595 int ObjectMonitor::raw_wait(jlong millis, bool interruptible, TRAPS) { | |
4596 TEVENT (raw_wait) ; | |
4597 if (THREAD != _owner) { | |
4598 return OM_ILLEGAL_MONITOR_STATE; | |
4599 } | |
4600 | |
4601 // To avoid spurious wakeups we reset the parkevent -- This is strictly optional. | |
4602 // The caller must be able to tolerate spurious returns from raw_wait(). | |
4603 THREAD->_ParkEvent->reset() ; | |
4604 OrderAccess::fence() ; | |
4605 | |
4606 // check interrupt event | |
4607 if (interruptible && Thread::is_interrupted(THREAD, true)) { | |
4608 return OM_INTERRUPTED; | |
4609 } | |
4610 | |
4611 intptr_t save = _recursions ; | |
4612 _recursions = 0 ; | |
4613 _waiters ++ ; | |
4614 if (THREAD->is_Java_thread()) { | |
4615 guarantee (((JavaThread *) THREAD)->thread_state() == _thread_blocked, "invariant") ; | |
4616 ((JavaThread *)THREAD)->set_suspend_equivalent(); | |
4617 } | |
4618 int rv = SimpleWait (THREAD, millis) ; | |
4619 _recursions = save ; | |
4620 _waiters -- ; | |
4621 | |
4622 guarantee (THREAD == _owner, "invariant") ; | |
4623 if (THREAD->is_Java_thread()) { | |
4624 JavaThread * jSelf = (JavaThread *) THREAD ; | |
4625 for (;;) { | |
4626 if (!jSelf->handle_special_suspend_equivalent_condition()) break ; | |
4627 SimpleExit (THREAD) ; | |
4628 jSelf->java_suspend_self(); | |
4629 SimpleEnter (THREAD) ; | |
4630 jSelf->set_suspend_equivalent() ; | |
4631 } | |
4632 } | |
4633 guarantee (THREAD == _owner, "invariant") ; | |
4634 | |
4635 if (interruptible && Thread::is_interrupted(THREAD, true)) { | |
4636 return OM_INTERRUPTED; | |
4637 } | |
4638 return OM_OK ; | |
4639 } | |
4640 | |
4641 int ObjectMonitor::raw_notify(TRAPS) { | |
4642 TEVENT (raw_notify) ; | |
4643 if (THREAD != _owner) { | |
4644 return OM_ILLEGAL_MONITOR_STATE; | |
4645 } | |
4646 SimpleNotify (THREAD, false) ; | |
4647 return OM_OK; | |
4648 } | |
4649 | |
4650 int ObjectMonitor::raw_notifyAll(TRAPS) { | |
4651 TEVENT (raw_notifyAll) ; | |
4652 if (THREAD != _owner) { | |
4653 return OM_ILLEGAL_MONITOR_STATE; | |
4654 } | |
4655 SimpleNotify (THREAD, true) ; | |
4656 return OM_OK; | |
4657 } | |
4658 | |
4659 #ifndef PRODUCT | |
4660 void ObjectMonitor::verify() { | |
4661 } | |
4662 | |
4663 void ObjectMonitor::print() { | |
4664 } | |
4665 #endif | |
4666 | |
4667 //------------------------------------------------------------------------------ | |
4668 // Non-product code | |
4669 | |
4670 #ifndef PRODUCT | |
4671 | |
4672 void ObjectSynchronizer::trace_locking(Handle locking_obj, bool is_compiled, | |
4673 bool is_method, bool is_locking) { | |
4674 // Don't know what to do here | |
4675 } | |
4676 | |
4677 // Verify all monitors in the monitor cache, the verification is weak. | |
4678 void ObjectSynchronizer::verify() { | |
4679 ObjectMonitor* block = gBlockList; | |
4680 ObjectMonitor* mid; | |
4681 while (block) { | |
4682 assert(block->object() == CHAINMARKER, "must be a block header"); | |
4683 for (int i = 1; i < _BLOCKSIZE; i++) { | |
4684 mid = block + i; | |
4685 oop object = (oop) mid->object(); | |
4686 if (object != NULL) { | |
4687 mid->verify(); | |
4688 } | |
4689 } | |
4690 block = (ObjectMonitor*) block->FreeNext; | |
4691 } | |
4692 } | |
4693 | |
4694 // Check if monitor belongs to the monitor cache | |
4695 // The list is grow-only so it's *relatively* safe to traverse | |
4696 // the list of extant blocks without taking a lock. | |
4697 | |
4698 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { | |
4699 ObjectMonitor* block = gBlockList; | |
4700 | |
4701 while (block) { | |
4702 assert(block->object() == CHAINMARKER, "must be a block header"); | |
4703 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { | |
4704 address mon = (address) monitor; | |
4705 address blk = (address) block; | |
4706 size_t diff = mon - blk; | |
4707 assert((diff % sizeof(ObjectMonitor)) == 0, "check"); | |
4708 return 1; | |
4709 } | |
4710 block = (ObjectMonitor*) block->FreeNext; | |
4711 } | |
4712 return 0; | |
4713 } | |
4714 | |
4715 #endif |