1 /*
2 * Copyright (c) 1998, 2014, Oracle and/or its affiliates. All rights reserved.
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 Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/vmSymbols.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "oops/markOop.hpp"
29 #include "oops/oop.inline.hpp"
30 #include "runtime/atomic.inline.hpp"
31 #include "runtime/biasedLocking.hpp"
32 #include "runtime/handles.inline.hpp"
33 #include "runtime/interfaceSupport.hpp"
34 #include "runtime/mutexLocker.hpp"
35 #include "runtime/objectMonitor.hpp"
36 #include "runtime/objectMonitor.inline.hpp"
37 #include "runtime/osThread.hpp"
38 #include "runtime/stubRoutines.hpp"
39 #include "runtime/synchronizer.hpp"
40 #include "runtime/thread.inline.hpp"
41 #include "utilities/dtrace.hpp"
42 #include "utilities/events.hpp"
43 #include "utilities/preserveException.hpp"
44
45 #if defined(__GNUC__) && !defined(PPC64)
46 // Need to inhibit inlining for older versions of GCC to avoid build-time failures
47 #define ATTR __attribute__((noinline))
48 #else
49 #define ATTR
50 #endif
51
52 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
53
54 // The "core" versions of monitor enter and exit reside in this file.
55 // The interpreter and compilers contain specialized transliterated
56 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(),
57 // for instance. If you make changes here, make sure to modify the
58 // interpreter, and both C1 and C2 fast-path inline locking code emission.
59 //
60 //
61 // -----------------------------------------------------------------------------
62
63 #ifdef DTRACE_ENABLED
64
65 // Only bother with this argument setup if dtrace is available
66 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
67
68 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
69 char* bytes = NULL; \
70 int len = 0; \
71 jlong jtid = SharedRuntime::get_java_tid(thread); \
72 Symbol* klassname = ((oop)(obj))->klass()->name(); \
73 if (klassname != NULL) { \
74 bytes = (char*)klassname->bytes(); \
75 len = klassname->utf8_length(); \
76 }
77
78 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
79 { \
80 if (DTraceMonitorProbes) { \
81 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
82 HOTSPOT_MONITOR_WAIT(jtid, \
83 (uintptr_t)(monitor), bytes, len, (millis)); \
84 } \
85 }
86
87 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
88
89 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
90 { \
91 if (DTraceMonitorProbes) { \
92 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
93 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
94 (uintptr_t)(monitor), bytes, len); \
95 } \
96 }
97
98 #else // ndef DTRACE_ENABLED
99
100 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
101 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
102
103 #endif // ndef DTRACE_ENABLED
104
105 // This exists only as a workaround of dtrace bug 6254741
106 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
107 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
108 return 0;
109 }
110
111 #define NINFLATIONLOCKS 256
112 static volatile intptr_t InflationLocks[NINFLATIONLOCKS];
113
114 ObjectMonitor * ObjectSynchronizer::gBlockList = NULL;
115 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL;
116 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL;
117 int ObjectSynchronizer::gOmInUseCount = 0;
118 static volatile intptr_t ListLock = 0; // protects global monitor free-list cache
119 static volatile int MonitorFreeCount = 0; // # on gFreeList
120 static volatile int MonitorPopulation = 0; // # Extant -- in circulation
121 #define CHAINMARKER (cast_to_oop<intptr_t>(-1))
122
123 // -----------------------------------------------------------------------------
124 // Fast Monitor Enter/Exit
125 // This the fast monitor enter. The interpreter and compiler use
126 // some assembly copies of this code. Make sure update those code
127 // if the following function is changed. The implementation is
128 // extremely sensitive to race condition. Be careful.
129
130 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) {
131 if (UseBiasedLocking) {
132 if (!SafepointSynchronize::is_at_safepoint()) {
133 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD);
134 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) {
135 return;
136 }
137 } else {
138 assert(!attempt_rebias, "can not rebias toward VM thread");
139 BiasedLocking::revoke_at_safepoint(obj);
140 }
141 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
142 }
143
144 slow_enter(obj, lock, THREAD);
145 }
146
147 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) {
148 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here");
149 // if displaced header is null, the previous enter is recursive enter, no-op
150 markOop dhw = lock->displaced_header();
151 markOop mark;
152 if (dhw == NULL) {
153 // Recursive stack-lock.
154 // Diagnostics -- Could be: stack-locked, inflating, inflated.
155 mark = object->mark();
156 assert(!mark->is_neutral(), "invariant");
157 if (mark->has_locker() && mark != markOopDesc::INFLATING()) {
158 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant");
159 }
160 if (mark->has_monitor()) {
161 ObjectMonitor * m = mark->monitor();
162 assert(((oop)(m->object()))->mark() == mark, "invariant");
163 assert(m->is_entered(THREAD), "invariant");
164 }
165 return;
166 }
167
168 mark = object->mark();
169
170 // If the object is stack-locked by the current thread, try to
171 // swing the displaced header from the box back to the mark.
172 if (mark == (markOop) lock) {
173 assert(dhw->is_neutral(), "invariant");
174 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) {
175 TEVENT(fast_exit: release stacklock);
176 return;
177 }
178 }
179
180 ObjectSynchronizer::inflate(THREAD, object)->exit(true, THREAD);
181 }
182
183 // -----------------------------------------------------------------------------
184 // Interpreter/Compiler Slow Case
185 // This routine is used to handle interpreter/compiler slow case
186 // We don't need to use fast path here, because it must have been
187 // failed in the interpreter/compiler code.
188 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
189 markOop mark = obj->mark();
190 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
191
192 if (mark->is_neutral()) {
193 // Anticipate successful CAS -- the ST of the displaced mark must
194 // be visible <= the ST performed by the CAS.
195 lock->set_displaced_header(mark);
196 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) {
197 TEVENT(slow_enter: release stacklock);
198 return;
199 }
200 // Fall through to inflate() ...
201 } else
202 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
203 assert(lock != mark->locker(), "must not re-lock the same lock");
204 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
205 lock->set_displaced_header(NULL);
206 return;
207 }
208
209 #if 0
210 // The following optimization isn't particularly useful.
211 if (mark->has_monitor() && mark->monitor()->is_entered(THREAD)) {
212 lock->set_displaced_header(NULL);
213 return;
214 }
215 #endif
216
217 // The object header will never be displaced to this lock,
218 // so it does not matter what the value is, except that it
219 // must be non-zero to avoid looking like a re-entrant lock,
220 // and must not look locked either.
221 lock->set_displaced_header(markOopDesc::unused_mark());
222 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
223 }
224
225 // This routine is used to handle interpreter/compiler slow case
226 // We don't need to use fast path here, because it must have
227 // failed in the interpreter/compiler code. Simply use the heavy
228 // weight monitor should be ok, unless someone find otherwise.
229 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
230 fast_exit(object, lock, THREAD);
231 }
232
233 // -----------------------------------------------------------------------------
234 // Class Loader support to workaround deadlocks on the class loader lock objects
235 // Also used by GC
236 // complete_exit()/reenter() are used to wait on a nested lock
237 // i.e. to give up an outer lock completely and then re-enter
238 // Used when holding nested locks - lock acquisition order: lock1 then lock2
239 // 1) complete_exit lock1 - saving recursion count
240 // 2) wait on lock2
241 // 3) when notified on lock2, unlock lock2
242 // 4) reenter lock1 with original recursion count
243 // 5) lock lock2
244 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
245 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
246 TEVENT(complete_exit);
247 if (UseBiasedLocking) {
248 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
249 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
250 }
251
252 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
253
254 return monitor->complete_exit(THREAD);
255 }
256
257 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
258 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
259 TEVENT(reenter);
260 if (UseBiasedLocking) {
261 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
262 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
263 }
264
265 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
266
267 monitor->reenter(recursion, THREAD);
268 }
269 // -----------------------------------------------------------------------------
270 // JNI locks on java objects
271 // NOTE: must use heavy weight monitor to handle jni monitor enter
272 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { // possible entry from jni enter
273 // the current locking is from JNI instead of Java code
274 TEVENT(jni_enter);
275 if (UseBiasedLocking) {
276 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
277 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
278 }
279 THREAD->set_current_pending_monitor_is_from_java(false);
280 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
281 THREAD->set_current_pending_monitor_is_from_java(true);
282 }
283
284 // NOTE: must use heavy weight monitor to handle jni monitor enter
285 bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) {
286 if (UseBiasedLocking) {
287 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
288 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
289 }
290
291 ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj());
292 return monitor->try_enter(THREAD);
293 }
294
295
296 // NOTE: must use heavy weight monitor to handle jni monitor exit
297 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
298 TEVENT(jni_exit);
299 if (UseBiasedLocking) {
300 Handle h_obj(THREAD, obj);
301 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD);
302 obj = h_obj();
303 }
304 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
305
306 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj);
307 // If this thread has locked the object, exit the monitor. Note: can't use
308 // monitor->check(CHECK); must exit even if an exception is pending.
309 if (monitor->check(THREAD)) {
310 monitor->exit(true, THREAD);
311 }
312 }
313
314 // -----------------------------------------------------------------------------
315 // Internal VM locks on java objects
316 // standard constructor, allows locking failures
317 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
318 _dolock = doLock;
319 _thread = thread;
320 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
321 _obj = obj;
322
323 if (_dolock) {
324 TEVENT(ObjectLocker);
325
326 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
327 }
328 }
329
330 ObjectLocker::~ObjectLocker() {
331 if (_dolock) {
332 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
333 }
334 }
335
336
337 // -----------------------------------------------------------------------------
338 // Wait/Notify/NotifyAll
339 // NOTE: must use heavy weight monitor to handle wait()
340 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
341 if (UseBiasedLocking) {
342 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
343 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
344 }
345 if (millis < 0) {
346 TEVENT(wait - throw IAX);
347 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
348 }
349 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
350 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
351 monitor->wait(millis, true, THREAD);
352
353 // This dummy call is in place to get around dtrace bug 6254741. Once
354 // that's fixed we can uncomment the following line, remove the call
355 // and change this function back into a "void" func.
356 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
357 return dtrace_waited_probe(monitor, obj, THREAD);
358 }
359
360 void ObjectSynchronizer::waitUninterruptibly (Handle obj, jlong millis, TRAPS) {
361 if (UseBiasedLocking) {
362 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
363 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
364 }
365 if (millis < 0) {
366 TEVENT(wait - throw IAX);
367 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
368 }
369 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD);
370 }
371
372 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
373 if (UseBiasedLocking) {
374 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
375 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
376 }
377
378 markOop mark = obj->mark();
379 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
380 return;
381 }
382 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD);
383 }
384
385 // NOTE: see comment of notify()
386 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
387 if (UseBiasedLocking) {
388 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
389 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
390 }
391
392 markOop mark = obj->mark();
393 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
394 return;
395 }
396 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD);
397 }
398
399 // -----------------------------------------------------------------------------
400 // Hash Code handling
401 //
402 // Performance concern:
403 // OrderAccess::storestore() calls release() which STs 0 into the global volatile
404 // OrderAccess::Dummy variable. This store is unnecessary for correctness.
405 // Many threads STing into a common location causes considerable cache migration
406 // or "sloshing" on large SMP system. As such, I avoid using OrderAccess::storestore()
407 // until it's repaired. In some cases OrderAccess::fence() -- which incurs local
408 // latency on the executing processor -- is a better choice as it scales on SMP
409 // systems. See http://blogs.sun.com/dave/entry/biased_locking_in_hotspot for a
410 // discussion of coherency costs. Note that all our current reference platforms
411 // provide strong ST-ST order, so the issue is moot on IA32, x64, and SPARC.
412 //
413 // As a general policy we use "volatile" to control compiler-based reordering
414 // and explicit fences (barriers) to control for architectural reordering performed
415 // by the CPU(s) or platform.
416
417 struct SharedGlobals {
418 // These are highly shared mostly-read variables.
419 // To avoid false-sharing they need to be the sole occupants of a $ line.
420 double padPrefix[8];
421 volatile int stwRandom;
422 volatile int stwCycle;
423
424 // Hot RW variables -- Sequester to avoid false-sharing
425 double padSuffix[16];
426 volatile int hcSequence;
427 double padFinal[8];
428 };
429
430 static SharedGlobals GVars;
431 static int MonitorScavengeThreshold = 1000000;
432 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending
433
434 static markOop ReadStableMark (oop obj) {
435 markOop mark = obj->mark();
436 if (!mark->is_being_inflated()) {
437 return mark; // normal fast-path return
438 }
439
440 int its = 0;
441 for (;;) {
442 markOop mark = obj->mark();
443 if (!mark->is_being_inflated()) {
444 return mark; // normal fast-path return
445 }
446
447 // The object is being inflated by some other thread.
448 // The caller of ReadStableMark() must wait for inflation to complete.
449 // Avoid live-lock
450 // TODO: consider calling SafepointSynchronize::do_call_back() while
451 // spinning to see if there's a safepoint pending. If so, immediately
452 // yielding or blocking would be appropriate. Avoid spinning while
453 // there is a safepoint pending.
454 // TODO: add inflation contention performance counters.
455 // TODO: restrict the aggregate number of spinners.
456
457 ++its;
458 if (its > 10000 || !os::is_MP()) {
459 if (its & 1) {
460 os::NakedYield();
461 TEVENT(Inflate: INFLATING - yield);
462 } else {
463 // Note that the following code attenuates the livelock problem but is not
464 // a complete remedy. A more complete solution would require that the inflating
465 // thread hold the associated inflation lock. The following code simply restricts
466 // the number of spinners to at most one. We'll have N-2 threads blocked
467 // on the inflationlock, 1 thread holding the inflation lock and using
468 // a yield/park strategy, and 1 thread in the midst of inflation.
469 // A more refined approach would be to change the encoding of INFLATING
470 // to allow encapsulation of a native thread pointer. Threads waiting for
471 // inflation to complete would use CAS to push themselves onto a singly linked
472 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
473 // and calling park(). When inflation was complete the thread that accomplished inflation
474 // would detach the list and set the markword to inflated with a single CAS and
475 // then for each thread on the list, set the flag and unpark() the thread.
476 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease
477 // wakes at most one thread whereas we need to wake the entire list.
478 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1);
479 int YieldThenBlock = 0;
480 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant");
481 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant");
482 Thread::muxAcquire(InflationLocks + ix, "InflationLock");
483 while (obj->mark() == markOopDesc::INFLATING()) {
484 // Beware: NakedYield() is advisory and has almost no effect on some platforms
485 // so we periodically call Self->_ParkEvent->park(1).
486 // We use a mixed spin/yield/block mechanism.
487 if ((YieldThenBlock++) >= 16) {
488 Thread::current()->_ParkEvent->park(1);
489 } else {
490 os::NakedYield();
491 }
492 }
493 Thread::muxRelease(InflationLocks + ix);
494 TEVENT(Inflate: INFLATING - yield/park);
495 }
496 } else {
497 SpinPause(); // SMP-polite spinning
498 }
499 }
500 }
501
502 // hashCode() generation :
503 //
504 // Possibilities:
505 // * MD5Digest of {obj,stwRandom}
506 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function.
507 // * A DES- or AES-style SBox[] mechanism
508 // * One of the Phi-based schemes, such as:
509 // 2654435761 = 2^32 * Phi (golden ratio)
510 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ;
511 // * A variation of Marsaglia's shift-xor RNG scheme.
512 // * (obj ^ stwRandom) is appealing, but can result
513 // in undesirable regularity in the hashCode values of adjacent objects
514 // (objects allocated back-to-back, in particular). This could potentially
515 // result in hashtable collisions and reduced hashtable efficiency.
516 // There are simple ways to "diffuse" the middle address bits over the
517 // generated hashCode values:
518 //
519
520 static inline intptr_t get_next_hash(Thread * Self, oop obj) {
521 intptr_t value = 0;
522 if (hashCode == 0) {
523 // This form uses an unguarded global Park-Miller RNG,
524 // so it's possible for two threads to race and generate the same RNG.
525 // On MP system we'll have lots of RW access to a global, so the
526 // mechanism induces lots of coherency traffic.
527 value = os::random();
528 } else
529 if (hashCode == 1) {
530 // This variation has the property of being stable (idempotent)
531 // between STW operations. This can be useful in some of the 1-0
532 // synchronization schemes.
533 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3;
534 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom;
535 } else
536 if (hashCode == 2) {
537 value = 1; // for sensitivity testing
538 } else
539 if (hashCode == 3) {
540 value = ++GVars.hcSequence;
541 } else
542 if (hashCode == 4) {
543 value = cast_from_oop<intptr_t>(obj);
544 } else {
545 // Marsaglia's xor-shift scheme with thread-specific state
546 // This is probably the best overall implementation -- we'll
547 // likely make this the default in future releases.
548 unsigned t = Self->_hashStateX;
549 t ^= (t << 11);
550 Self->_hashStateX = Self->_hashStateY;
551 Self->_hashStateY = Self->_hashStateZ;
552 Self->_hashStateZ = Self->_hashStateW;
553 unsigned v = Self->_hashStateW;
554 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
555 Self->_hashStateW = v;
556 value = v;
557 }
558
559 value &= markOopDesc::hash_mask;
560 if (value == 0) value = 0xBAD;
561 assert(value != markOopDesc::no_hash, "invariant");
562 TEVENT(hashCode: GENERATE);
563 return value;
564 }
565 //
566 intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) {
567 if (UseBiasedLocking) {
568 // NOTE: many places throughout the JVM do not expect a safepoint
569 // to be taken here, in particular most operations on perm gen
570 // objects. However, we only ever bias Java instances and all of
571 // the call sites of identity_hash that might revoke biases have
572 // been checked to make sure they can handle a safepoint. The
573 // added check of the bias pattern is to avoid useless calls to
574 // thread-local storage.
575 if (obj->mark()->has_bias_pattern()) {
576 // Box and unbox the raw reference just in case we cause a STW safepoint.
577 Handle hobj(Self, obj);
578 // Relaxing assertion for bug 6320749.
579 assert(Universe::verify_in_progress() ||
580 !SafepointSynchronize::is_at_safepoint(),
581 "biases should not be seen by VM thread here");
582 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
583 obj = hobj();
584 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
585 }
586 }
587
588 // hashCode() is a heap mutator ...
589 // Relaxing assertion for bug 6320749.
590 assert(Universe::verify_in_progress() ||
591 !SafepointSynchronize::is_at_safepoint(), "invariant");
592 assert(Universe::verify_in_progress() ||
593 Self->is_Java_thread() , "invariant");
594 assert(Universe::verify_in_progress() ||
595 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
596
597 ObjectMonitor* monitor = NULL;
598 markOop temp, test;
599 intptr_t hash;
600 markOop mark = ReadStableMark (obj);
601
602 // object should remain ineligible for biased locking
603 assert(!mark->has_bias_pattern(), "invariant");
604
605 if (mark->is_neutral()) {
606 hash = mark->hash(); // this is a normal header
607 if (hash) { // if it has hash, just return it
608 return hash;
609 }
610 hash = get_next_hash(Self, obj); // allocate a new hash code
611 temp = mark->copy_set_hash(hash); // merge the hash code into header
612 // use (machine word version) atomic operation to install the hash
613 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark);
614 if (test == mark) {
615 return hash;
616 }
617 // If atomic operation failed, we must inflate the header
618 // into heavy weight monitor. We could add more code here
619 // for fast path, but it does not worth the complexity.
620 } else if (mark->has_monitor()) {
621 monitor = mark->monitor();
622 temp = monitor->header();
623 assert(temp->is_neutral(), "invariant");
624 hash = temp->hash();
625 if (hash) {
626 return hash;
627 }
628 // Skip to the following code to reduce code size
629 } else if (Self->is_lock_owned((address)mark->locker())) {
630 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
631 assert(temp->is_neutral(), "invariant");
632 hash = temp->hash(); // by current thread, check if the displaced
633 if (hash) { // header contains hash code
634 return hash;
635 }
636 // WARNING:
637 // The displaced header is strictly immutable.
638 // It can NOT be changed in ANY cases. So we have
639 // to inflate the header into heavyweight monitor
640 // even the current thread owns the lock. The reason
641 // is the BasicLock (stack slot) will be asynchronously
642 // read by other threads during the inflate() function.
643 // Any change to stack may not propagate to other threads
644 // correctly.
645 }
646
647 // Inflate the monitor to set hash code
648 monitor = ObjectSynchronizer::inflate(Self, obj);
649 // Load displaced header and check it has hash code
650 mark = monitor->header();
651 assert(mark->is_neutral(), "invariant");
652 hash = mark->hash();
653 if (hash == 0) {
654 hash = get_next_hash(Self, obj);
655 temp = mark->copy_set_hash(hash); // merge hash code into header
656 assert(temp->is_neutral(), "invariant");
657 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark);
658 if (test != mark) {
659 // The only update to the header in the monitor (outside GC)
660 // is install the hash code. If someone add new usage of
661 // displaced header, please update this code
662 hash = test->hash();
663 assert(test->is_neutral(), "invariant");
664 assert(hash != 0, "Trivial unexpected object/monitor header usage.");
665 }
666 }
667 // We finally get the hash
668 return hash;
669 }
670
671 // Deprecated -- use FastHashCode() instead.
672
673 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
674 return FastHashCode(Thread::current(), obj());
675 }
676
677
678 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
679 Handle h_obj) {
680 if (UseBiasedLocking) {
681 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
682 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
683 }
684
685 assert(thread == JavaThread::current(), "Can only be called on current thread");
686 oop obj = h_obj();
687
688 markOop mark = ReadStableMark(obj);
689
690 // Uncontended case, header points to stack
691 if (mark->has_locker()) {
692 return thread->is_lock_owned((address)mark->locker());
693 }
694 // Contended case, header points to ObjectMonitor (tagged pointer)
695 if (mark->has_monitor()) {
696 ObjectMonitor* monitor = mark->monitor();
697 return monitor->is_entered(thread) != 0;
698 }
699 // Unlocked case, header in place
700 assert(mark->is_neutral(), "sanity check");
701 return false;
702 }
703
704 // Be aware of this method could revoke bias of the lock object.
705 // This method queries the ownership of the lock handle specified by 'h_obj'.
706 // If the current thread owns the lock, it returns owner_self. If no
707 // thread owns the lock, it returns owner_none. Otherwise, it will return
708 // owner_other.
709 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
710 (JavaThread *self, Handle h_obj) {
711 // The caller must beware this method can revoke bias, and
712 // revocation can result in a safepoint.
713 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
714 assert(self->thread_state() != _thread_blocked , "invariant");
715
716 // Possible mark states: neutral, biased, stack-locked, inflated
717
718 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
719 // CASE: biased
720 BiasedLocking::revoke_and_rebias(h_obj, false, self);
721 assert(!h_obj->mark()->has_bias_pattern(),
722 "biases should be revoked by now");
723 }
724
725 assert(self == JavaThread::current(), "Can only be called on current thread");
726 oop obj = h_obj();
727 markOop mark = ReadStableMark(obj);
728
729 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
730 if (mark->has_locker()) {
731 return self->is_lock_owned((address)mark->locker()) ?
732 owner_self : owner_other;
733 }
734
735 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor.
736 // The Object:ObjectMonitor relationship is stable as long as we're
737 // not at a safepoint.
738 if (mark->has_monitor()) {
739 void * owner = mark->monitor()->_owner;
740 if (owner == NULL) return owner_none;
741 return (owner == self ||
742 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
743 }
744
745 // CASE: neutral
746 assert(mark->is_neutral(), "sanity check");
747 return owner_none; // it's unlocked
748 }
749
750 // FIXME: jvmti should call this
751 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) {
752 if (UseBiasedLocking) {
753 if (SafepointSynchronize::is_at_safepoint()) {
754 BiasedLocking::revoke_at_safepoint(h_obj);
755 } else {
756 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
757 }
758 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
759 }
760
761 oop obj = h_obj();
762 address owner = NULL;
763
764 markOop mark = ReadStableMark(obj);
765
766 // Uncontended case, header points to stack
767 if (mark->has_locker()) {
768 owner = (address) mark->locker();
769 }
770
771 // Contended case, header points to ObjectMonitor (tagged pointer)
772 if (mark->has_monitor()) {
773 ObjectMonitor* monitor = mark->monitor();
774 assert(monitor != NULL, "monitor should be non-null");
775 owner = (address) monitor->owner();
776 }
777
778 if (owner != NULL) {
779 // owning_thread_from_monitor_owner() may also return NULL here
780 return Threads::owning_thread_from_monitor_owner(owner, doLock);
781 }
782
783 // Unlocked case, header in place
784 // Cannot have assertion since this object may have been
785 // locked by another thread when reaching here.
786 // assert(mark->is_neutral(), "sanity check");
787
788 return NULL;
789 }
790 // Visitors ...
791
792 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
793 ObjectMonitor* block = gBlockList;
794 ObjectMonitor* mid;
795 while (block) {
796 assert(block->object() == CHAINMARKER, "must be a block header");
797 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
798 mid = block + i;
799 oop object = (oop) mid->object();
800 if (object != NULL) {
801 closure->do_monitor(mid);
802 }
803 }
804 block = (ObjectMonitor*) block->FreeNext;
805 }
806 }
807
808 // Get the next block in the block list.
809 static inline ObjectMonitor* next(ObjectMonitor* block) {
810 assert(block->object() == CHAINMARKER, "must be a block header");
811 block = block->FreeNext;
812 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
813 return block;
814 }
815
816
817 void ObjectSynchronizer::oops_do(OopClosure* f) {
818 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
819 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) {
820 assert(block->object() == CHAINMARKER, "must be a block header");
821 for (int i = 1; i < _BLOCKSIZE; i++) {
822 ObjectMonitor* mid = &block[i];
823 if (mid->object() != NULL) {
824 f->do_oop((oop*)mid->object_addr());
825 }
826 }
827 }
828 }
829
830
831 // -----------------------------------------------------------------------------
832 // ObjectMonitor Lifecycle
833 // -----------------------
834 // Inflation unlinks monitors from the global gFreeList and
835 // associates them with objects. Deflation -- which occurs at
836 // STW-time -- disassociates idle monitors from objects. Such
837 // scavenged monitors are returned to the gFreeList.
838 //
839 // The global list is protected by ListLock. All the critical sections
840 // are short and operate in constant-time.
841 //
842 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
843 //
844 // Lifecycle:
845 // -- unassigned and on the global free list
846 // -- unassigned and on a thread's private omFreeList
847 // -- assigned to an object. The object is inflated and the mark refers
848 // to the objectmonitor.
849 //
850
851
852 // Constraining monitor pool growth via MonitorBound ...
853 //
854 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
855 // the rate of scavenging is driven primarily by GC. As such, we can find
856 // an inordinate number of monitors in circulation.
857 // To avoid that scenario we can artificially induce a STW safepoint
858 // if the pool appears to be growing past some reasonable bound.
859 // Generally we favor time in space-time tradeoffs, but as there's no
860 // natural back-pressure on the # of extant monitors we need to impose some
861 // type of limit. Beware that if MonitorBound is set to too low a value
862 // we could just loop. In addition, if MonitorBound is set to a low value
863 // we'll incur more safepoints, which are harmful to performance.
864 // See also: GuaranteedSafepointInterval
865 //
866 // The current implementation uses asynchronous VM operations.
867 //
868
869 static void InduceScavenge (Thread * Self, const char * Whence) {
870 // Induce STW safepoint to trim monitors
871 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
872 // More precisely, trigger an asynchronous STW safepoint as the number
873 // of active monitors passes the specified threshold.
874 // TODO: assert thread state is reasonable
875
876 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
877 if (ObjectMonitor::Knob_Verbose) {
878 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ;
879 ::fflush(stdout);
880 }
881 // Induce a 'null' safepoint to scavenge monitors
882 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
883 // to the VMthread and have a lifespan longer than that of this activation record.
884 // The VMThread will delete the op when completed.
885 VMThread::execute(new VM_ForceAsyncSafepoint());
886
887 if (ObjectMonitor::Knob_Verbose) {
888 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ;
889 ::fflush(stdout);
890 }
891 }
892 }
893 /* Too slow for general assert or debug
894 void ObjectSynchronizer::verifyInUse (Thread *Self) {
895 ObjectMonitor* mid;
896 int inusetally = 0;
897 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) {
898 inusetally ++;
899 }
900 assert(inusetally == Self->omInUseCount, "inuse count off");
901
902 int freetally = 0;
903 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) {
904 freetally ++;
905 }
906 assert(freetally == Self->omFreeCount, "free count off");
907 }
908 */
909 ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) {
910 // A large MAXPRIVATE value reduces both list lock contention
911 // and list coherency traffic, but also tends to increase the
912 // number of objectMonitors in circulation as well as the STW
913 // scavenge costs. As usual, we lean toward time in space-time
914 // tradeoffs.
915 const int MAXPRIVATE = 1024;
916 for (;;) {
917 ObjectMonitor * m;
918
919 // 1: try to allocate from the thread's local omFreeList.
920 // Threads will attempt to allocate first from their local list, then
921 // from the global list, and only after those attempts fail will the thread
922 // attempt to instantiate new monitors. Thread-local free lists take
923 // heat off the ListLock and improve allocation latency, as well as reducing
924 // coherency traffic on the shared global list.
925 m = Self->omFreeList;
926 if (m != NULL) {
927 Self->omFreeList = m->FreeNext;
928 Self->omFreeCount--;
929 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene
930 guarantee(m->object() == NULL, "invariant");
931 if (MonitorInUseLists) {
932 m->FreeNext = Self->omInUseList;
933 Self->omInUseList = m;
934 Self->omInUseCount++;
935 // verifyInUse(Self);
936 } else {
937 m->FreeNext = NULL;
938 }
939 return m;
940 }
941
942 // 2: try to allocate from the global gFreeList
943 // CONSIDER: use muxTry() instead of muxAcquire().
944 // If the muxTry() fails then drop immediately into case 3.
945 // If we're using thread-local free lists then try
946 // to reprovision the caller's free list.
947 if (gFreeList != NULL) {
948 // Reprovision the thread's omFreeList.
949 // Use bulk transfers to reduce the allocation rate and heat
950 // on various locks.
951 Thread::muxAcquire(&ListLock, "omAlloc");
952 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) {
953 MonitorFreeCount--;
954 ObjectMonitor * take = gFreeList;
955 gFreeList = take->FreeNext;
956 guarantee(take->object() == NULL, "invariant");
957 guarantee(!take->is_busy(), "invariant");
958 take->Recycle();
959 omRelease(Self, take, false);
960 }
961 Thread::muxRelease(&ListLock);
962 Self->omFreeProvision += 1 + (Self->omFreeProvision/2);
963 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE;
964 TEVENT(omFirst - reprovision);
965
966 const int mx = MonitorBound;
967 if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) {
968 // We can't safely induce a STW safepoint from omAlloc() as our thread
969 // state may not be appropriate for such activities and callers may hold
970 // naked oops, so instead we defer the action.
971 InduceScavenge(Self, "omAlloc");
972 }
973 continue;
974 }
975
976 // 3: allocate a block of new ObjectMonitors
977 // Both the local and global free lists are empty -- resort to malloc().
978 // In the current implementation objectMonitors are TSM - immortal.
979 assert(_BLOCKSIZE > 1, "invariant");
980 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE];
981
982 // NOTE: (almost) no way to recover if allocation failed.
983 // We might be able to induce a STW safepoint and scavenge enough
984 // objectMonitors to permit progress.
985 if (temp == NULL) {
986 vm_exit_out_of_memory(sizeof (ObjectMonitor[_BLOCKSIZE]), OOM_MALLOC_ERROR,
987 "Allocate ObjectMonitors");
988 }
989
990 // Format the block.
991 // initialize the linked list, each monitor points to its next
992 // forming the single linked free list, the very first monitor
993 // will points to next block, which forms the block list.
994 // The trick of using the 1st element in the block as gBlockList
995 // linkage should be reconsidered. A better implementation would
996 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
997
998 for (int i = 1; i < _BLOCKSIZE; i++) {
999 temp[i].FreeNext = &temp[i+1];
1000 }
1001
1002 // terminate the last monitor as the end of list
1003 temp[_BLOCKSIZE - 1].FreeNext = NULL;
1004
1005 // Element [0] is reserved for global list linkage
1006 temp[0].set_object(CHAINMARKER);
1007
1008 // Consider carving out this thread's current request from the
1009 // block in hand. This avoids some lock traffic and redundant
1010 // list activity.
1011
1012 // Acquire the ListLock to manipulate BlockList and FreeList.
1013 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1014 Thread::muxAcquire(&ListLock, "omAlloc [2]");
1015 MonitorPopulation += _BLOCKSIZE-1;
1016 MonitorFreeCount += _BLOCKSIZE-1;
1017
1018 // Add the new block to the list of extant blocks (gBlockList).
1019 // The very first objectMonitor in a block is reserved and dedicated.
1020 // It serves as blocklist "next" linkage.
1021 temp[0].FreeNext = gBlockList;
1022 gBlockList = temp;
1023
1024 // Add the new string of objectMonitors to the global free list
1025 temp[_BLOCKSIZE - 1].FreeNext = gFreeList;
1026 gFreeList = temp + 1;
1027 Thread::muxRelease(&ListLock);
1028 TEVENT(Allocate block of monitors);
1029 }
1030 }
1031
1032 // Place "m" on the caller's private per-thread omFreeList.
1033 // In practice there's no need to clamp or limit the number of
1034 // monitors on a thread's omFreeList as the only time we'll call
1035 // omRelease is to return a monitor to the free list after a CAS
1036 // attempt failed. This doesn't allow unbounded #s of monitors to
1037 // accumulate on a thread's free list.
1038 //
1039
1040 void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m, bool fromPerThreadAlloc) {
1041 guarantee(m->object() == NULL, "invariant");
1042
1043 // Remove from omInUseList
1044 if (MonitorInUseLists && fromPerThreadAlloc) {
1045 ObjectMonitor* curmidinuse = NULL;
1046 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL;) {
1047 if (m == mid) {
1048 // extract from per-thread in-use-list
1049 if (mid == Self->omInUseList) {
1050 Self->omInUseList = mid->FreeNext;
1051 } else if (curmidinuse != NULL) {
1052 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist
1053 }
1054 Self->omInUseCount--;
1055 // verifyInUse(Self);
1056 break;
1057 } else {
1058 curmidinuse = mid;
1059 mid = mid->FreeNext;
1060 }
1061 }
1062 }
1063
1064 // FreeNext is used for both onInUseList and omFreeList, so clear old before setting new
1065 m->FreeNext = Self->omFreeList;
1066 Self->omFreeList = m;
1067 Self->omFreeCount++;
1068 }
1069
1070 // Return the monitors of a moribund thread's local free list to
1071 // the global free list. Typically a thread calls omFlush() when
1072 // it's dying. We could also consider having the VM thread steal
1073 // monitors from threads that have not run java code over a few
1074 // consecutive STW safepoints. Relatedly, we might decay
1075 // omFreeProvision at STW safepoints.
1076 //
1077 // Also return the monitors of a moribund thread"s omInUseList to
1078 // a global gOmInUseList under the global list lock so these
1079 // will continue to be scanned.
1080 //
1081 // We currently call omFlush() from the Thread:: dtor _after the thread
1082 // has been excised from the thread list and is no longer a mutator.
1083 // That means that omFlush() can run concurrently with a safepoint and
1084 // the scavenge operator. Calling omFlush() from JavaThread::exit() might
1085 // be a better choice as we could safely reason that that the JVM is
1086 // not at a safepoint at the time of the call, and thus there could
1087 // be not inopportune interleavings between omFlush() and the scavenge
1088 // operator.
1089
1090 void ObjectSynchronizer::omFlush (Thread * Self) {
1091 ObjectMonitor * List = Self->omFreeList; // Null-terminated SLL
1092 Self->omFreeList = NULL;
1093 ObjectMonitor * Tail = NULL;
1094 int Tally = 0;
1095 if (List != NULL) {
1096 ObjectMonitor * s;
1097 for (s = List; s != NULL; s = s->FreeNext) {
1098 Tally++;
1099 Tail = s;
1100 guarantee(s->object() == NULL, "invariant");
1101 guarantee(!s->is_busy(), "invariant");
1102 s->set_owner(NULL); // redundant but good hygiene
1103 TEVENT(omFlush - Move one);
1104 }
1105 guarantee(Tail != NULL && List != NULL, "invariant");
1106 }
1107
1108 ObjectMonitor * InUseList = Self->omInUseList;
1109 ObjectMonitor * InUseTail = NULL;
1110 int InUseTally = 0;
1111 if (InUseList != NULL) {
1112 Self->omInUseList = NULL;
1113 ObjectMonitor *curom;
1114 for (curom = InUseList; curom != NULL; curom = curom->FreeNext) {
1115 InUseTail = curom;
1116 InUseTally++;
1117 }
1118 // TODO debug
1119 assert(Self->omInUseCount == InUseTally, "inuse count off");
1120 Self->omInUseCount = 0;
1121 guarantee(InUseTail != NULL && InUseList != NULL, "invariant");
1122 }
1123
1124 Thread::muxAcquire(&ListLock, "omFlush");
1125 if (Tail != NULL) {
1126 Tail->FreeNext = gFreeList;
1127 gFreeList = List;
1128 MonitorFreeCount += Tally;
1129 }
1130
1131 if (InUseTail != NULL) {
1132 InUseTail->FreeNext = gOmInUseList;
1133 gOmInUseList = InUseList;
1134 gOmInUseCount += InUseTally;
1135 }
1136
1137 Thread::muxRelease(&ListLock);
1138 TEVENT(omFlush);
1139 }
1140
1141 // Fast path code shared by multiple functions
1142 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) {
1143 markOop mark = obj->mark();
1144 if (mark->has_monitor()) {
1145 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid");
1146 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header");
1147 return mark->monitor();
1148 }
1149 return ObjectSynchronizer::inflate(Thread::current(), obj);
1150 }
1151
1152
1153 // Note that we could encounter some performance loss through false-sharing as
1154 // multiple locks occupy the same $ line. Padding might be appropriate.
1155
1156
1157 ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) {
1158 // Inflate mutates the heap ...
1159 // Relaxing assertion for bug 6320749.
1160 assert(Universe::verify_in_progress() ||
1161 !SafepointSynchronize::is_at_safepoint(), "invariant");
1162
1163 for (;;) {
1164 const markOop mark = object->mark();
1165 assert(!mark->has_bias_pattern(), "invariant");
1166
1167 // The mark can be in one of the following states:
1168 // * Inflated - just return
1169 // * Stack-locked - coerce it to inflated
1170 // * INFLATING - busy wait for conversion to complete
1171 // * Neutral - aggressively inflate the object.
1172 // * BIASED - Illegal. We should never see this
1173
1174 // CASE: inflated
1175 if (mark->has_monitor()) {
1176 ObjectMonitor * inf = mark->monitor();
1177 assert(inf->header()->is_neutral(), "invariant");
1178 assert(inf->object() == object, "invariant");
1179 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1180 return inf;
1181 }
1182
1183 // CASE: inflation in progress - inflating over a stack-lock.
1184 // Some other thread is converting from stack-locked to inflated.
1185 // Only that thread can complete inflation -- other threads must wait.
1186 // The INFLATING value is transient.
1187 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1188 // We could always eliminate polling by parking the thread on some auxiliary list.
1189 if (mark == markOopDesc::INFLATING()) {
1190 TEVENT(Inflate: spin while INFLATING);
1191 ReadStableMark(object);
1192 continue;
1193 }
1194
1195 // CASE: stack-locked
1196 // Could be stack-locked either by this thread or by some other thread.
1197 //
1198 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1199 // to install INFLATING into the mark word. We originally installed INFLATING,
1200 // allocated the objectmonitor, and then finally STed the address of the
1201 // objectmonitor into the mark. This was correct, but artificially lengthened
1202 // the interval in which INFLATED appeared in the mark, thus increasing
1203 // the odds of inflation contention.
1204 //
1205 // We now use per-thread private objectmonitor free lists.
1206 // These list are reprovisioned from the global free list outside the
1207 // critical INFLATING...ST interval. A thread can transfer
1208 // multiple objectmonitors en-mass from the global free list to its local free list.
1209 // This reduces coherency traffic and lock contention on the global free list.
1210 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1211 // before or after the CAS(INFLATING) operation.
1212 // See the comments in omAlloc().
1213
1214 if (mark->has_locker()) {
1215 ObjectMonitor * m = omAlloc(Self);
1216 // Optimistically prepare the objectmonitor - anticipate successful CAS
1217 // We do this before the CAS in order to minimize the length of time
1218 // in which INFLATING appears in the mark.
1219 m->Recycle();
1220 m->_Responsible = NULL;
1221 m->OwnerIsThread = 0;
1222 m->_recursions = 0;
1223 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1224
1225 markOop cmp = (markOop) Atomic::cmpxchg_ptr(markOopDesc::INFLATING(), object->mark_addr(), mark);
1226 if (cmp != mark) {
1227 omRelease(Self, m, true);
1228 continue; // Interference -- just retry
1229 }
1230
1231 // We've successfully installed INFLATING (0) into the mark-word.
1232 // This is the only case where 0 will appear in a mark-work.
1233 // Only the singular thread that successfully swings the mark-word
1234 // to 0 can perform (or more precisely, complete) inflation.
1235 //
1236 // Why do we CAS a 0 into the mark-word instead of just CASing the
1237 // mark-word from the stack-locked value directly to the new inflated state?
1238 // Consider what happens when a thread unlocks a stack-locked object.
1239 // It attempts to use CAS to swing the displaced header value from the
1240 // on-stack basiclock back into the object header. Recall also that the
1241 // header value (hashcode, etc) can reside in (a) the object header, or
1242 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1243 // header in an objectMonitor. The inflate() routine must copy the header
1244 // value from the basiclock on the owner's stack to the objectMonitor, all
1245 // the while preserving the hashCode stability invariants. If the owner
1246 // decides to release the lock while the value is 0, the unlock will fail
1247 // and control will eventually pass from slow_exit() to inflate. The owner
1248 // will then spin, waiting for the 0 value to disappear. Put another way,
1249 // the 0 causes the owner to stall if the owner happens to try to
1250 // drop the lock (restoring the header from the basiclock to the object)
1251 // while inflation is in-progress. This protocol avoids races that might
1252 // would otherwise permit hashCode values to change or "flicker" for an object.
1253 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1254 // 0 serves as a "BUSY" inflate-in-progress indicator.
1255
1256
1257 // fetch the displaced mark from the owner's stack.
1258 // The owner can't die or unwind past the lock while our INFLATING
1259 // object is in the mark. Furthermore the owner can't complete
1260 // an unlock on the object, either.
1261 markOop dmw = mark->displaced_mark_helper();
1262 assert(dmw->is_neutral(), "invariant");
1263
1264 // Setup monitor fields to proper values -- prepare the monitor
1265 m->set_header(dmw);
1266
1267 // Optimization: if the mark->locker stack address is associated
1268 // with this thread we could simply set m->_owner = Self and
1269 // m->OwnerIsThread = 1. Note that a thread can inflate an object
1270 // that it has stack-locked -- as might happen in wait() -- directly
1271 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1272 m->set_owner(mark->locker());
1273 m->set_object(object);
1274 // TODO-FIXME: assert BasicLock->dhw != 0.
1275
1276 // Must preserve store ordering. The monitor state must
1277 // be stable at the time of publishing the monitor address.
1278 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant");
1279 object->release_set_mark(markOopDesc::encode(m));
1280
1281 // Hopefully the performance counters are allocated on distinct cache lines
1282 // to avoid false sharing on MP systems ...
1283 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc();
1284 TEVENT(Inflate: overwrite stacklock);
1285 if (TraceMonitorInflation) {
1286 if (object->is_instance()) {
1287 ResourceMark rm;
1288 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1289 (void *) object, (intptr_t) object->mark(),
1290 object->klass()->external_name());
1291 }
1292 }
1293 return m;
1294 }
1295
1296 // CASE: neutral
1297 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1298 // If we know we're inflating for entry it's better to inflate by swinging a
1299 // pre-locked objectMonitor pointer into the object header. A successful
1300 // CAS inflates the object *and* confers ownership to the inflating thread.
1301 // In the current implementation we use a 2-step mechanism where we CAS()
1302 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1303 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1304 // would be useful.
1305
1306 assert(mark->is_neutral(), "invariant");
1307 ObjectMonitor * m = omAlloc(Self);
1308 // prepare m for installation - set monitor to initial state
1309 m->Recycle();
1310 m->set_header(mark);
1311 m->set_owner(NULL);
1312 m->set_object(object);
1313 m->OwnerIsThread = 1;
1314 m->_recursions = 0;
1315 m->_Responsible = NULL;
1316 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1317
1318 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) {
1319 m->set_object(NULL);
1320 m->set_owner(NULL);
1321 m->OwnerIsThread = 0;
1322 m->Recycle();
1323 omRelease(Self, m, true);
1324 m = NULL;
1325 continue;
1326 // interference - the markword changed - just retry.
1327 // The state-transitions are one-way, so there's no chance of
1328 // live-lock -- "Inflated" is an absorbing state.
1329 }
1330
1331 // Hopefully the performance counters are allocated on distinct
1332 // cache lines to avoid false sharing on MP systems ...
1333 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc();
1334 TEVENT(Inflate: overwrite neutral);
1335 if (TraceMonitorInflation) {
1336 if (object->is_instance()) {
1337 ResourceMark rm;
1338 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1339 (void *) object, (intptr_t) object->mark(),
1340 object->klass()->external_name());
1341 }
1342 }
1343 return m;
1344 }
1345 }
1346
1347 // Note that we could encounter some performance loss through false-sharing as
1348 // multiple locks occupy the same $ line. Padding might be appropriate.
1349
1350
1351 // Deflate_idle_monitors() is called at all safepoints, immediately
1352 // after all mutators are stopped, but before any objects have moved.
1353 // It traverses the list of known monitors, deflating where possible.
1354 // The scavenged monitor are returned to the monitor free list.
1355 //
1356 // Beware that we scavenge at *every* stop-the-world point.
1357 // Having a large number of monitors in-circulation negatively
1358 // impacts the performance of some applications (e.g., PointBase).
1359 // Broadly, we want to minimize the # of monitors in circulation.
1360 //
1361 // We have added a flag, MonitorInUseLists, which creates a list
1362 // of active monitors for each thread. deflate_idle_monitors()
1363 // only scans the per-thread inuse lists. omAlloc() puts all
1364 // assigned monitors on the per-thread list. deflate_idle_monitors()
1365 // returns the non-busy monitors to the global free list.
1366 // When a thread dies, omFlush() adds the list of active monitors for
1367 // that thread to a global gOmInUseList acquiring the
1368 // global list lock. deflate_idle_monitors() acquires the global
1369 // list lock to scan for non-busy monitors to the global free list.
1370 // An alternative could have used a single global inuse list. The
1371 // downside would have been the additional cost of acquiring the global list lock
1372 // for every omAlloc().
1373 //
1374 // Perversely, the heap size -- and thus the STW safepoint rate --
1375 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1376 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1377 // This is an unfortunate aspect of this design.
1378 //
1379
1380 enum ManifestConstants {
1381 ClearResponsibleAtSTW = 0,
1382 MaximumRecheckInterval = 1000
1383 };
1384
1385 // Deflate a single monitor if not in use
1386 // Return true if deflated, false if in use
1387 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1388 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) {
1389 bool deflated;
1390 // Normal case ... The monitor is associated with obj.
1391 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant");
1392 guarantee(mid == obj->mark()->monitor(), "invariant");
1393 guarantee(mid->header()->is_neutral(), "invariant");
1394
1395 if (mid->is_busy()) {
1396 if (ClearResponsibleAtSTW) mid->_Responsible = NULL;
1397 deflated = false;
1398 } else {
1399 // Deflate the monitor if it is no longer being used
1400 // It's idle - scavenge and return to the global free list
1401 // plain old deflation ...
1402 TEVENT(deflate_idle_monitors - scavenge1);
1403 if (TraceMonitorInflation) {
1404 if (obj->is_instance()) {
1405 ResourceMark rm;
1406 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1407 (void *) obj, (intptr_t) obj->mark(), obj->klass()->external_name());
1408 }
1409 }
1410
1411 // Restore the header back to obj
1412 obj->release_set_mark(mid->header());
1413 mid->clear();
1414
1415 assert(mid->object() == NULL, "invariant");
1416
1417 // Move the object to the working free list defined by FreeHead,FreeTail.
1418 if (*FreeHeadp == NULL) *FreeHeadp = mid;
1419 if (*FreeTailp != NULL) {
1420 ObjectMonitor * prevtail = *FreeTailp;
1421 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); // TODO KK
1422 prevtail->FreeNext = mid;
1423 }
1424 *FreeTailp = mid;
1425 deflated = true;
1426 }
1427 return deflated;
1428 }
1429
1430 // Caller acquires ListLock
1431 int ObjectSynchronizer::walk_monitor_list(ObjectMonitor** listheadp,
1432 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) {
1433 ObjectMonitor* mid;
1434 ObjectMonitor* next;
1435 ObjectMonitor* curmidinuse = NULL;
1436 int deflatedcount = 0;
1437
1438 for (mid = *listheadp; mid != NULL;) {
1439 oop obj = (oop) mid->object();
1440 bool deflated = false;
1441 if (obj != NULL) {
1442 deflated = deflate_monitor(mid, obj, FreeHeadp, FreeTailp);
1443 }
1444 if (deflated) {
1445 // extract from per-thread in-use-list
1446 if (mid == *listheadp) {
1447 *listheadp = mid->FreeNext;
1448 } else if (curmidinuse != NULL) {
1449 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist
1450 }
1451 next = mid->FreeNext;
1452 mid->FreeNext = NULL; // This mid is current tail in the FreeHead list
1453 mid = next;
1454 deflatedcount++;
1455 } else {
1456 curmidinuse = mid;
1457 mid = mid->FreeNext;
1458 }
1459 }
1460 return deflatedcount;
1461 }
1462
1463 void ObjectSynchronizer::deflate_idle_monitors() {
1464 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1465 int nInuse = 0; // currently associated with objects
1466 int nInCirculation = 0; // extant
1467 int nScavenged = 0; // reclaimed
1468 bool deflated = false;
1469
1470 ObjectMonitor * FreeHead = NULL; // Local SLL of scavenged monitors
1471 ObjectMonitor * FreeTail = NULL;
1472
1473 TEVENT(deflate_idle_monitors);
1474 // Prevent omFlush from changing mids in Thread dtor's during deflation
1475 // And in case the vm thread is acquiring a lock during a safepoint
1476 // See e.g. 6320749
1477 Thread::muxAcquire(&ListLock, "scavenge - return");
1478
1479 if (MonitorInUseLists) {
1480 int inUse = 0;
1481 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) {
1482 nInCirculation+= cur->omInUseCount;
1483 int deflatedcount = walk_monitor_list(cur->omInUseList_addr(), &FreeHead, &FreeTail);
1484 cur->omInUseCount-= deflatedcount;
1485 // verifyInUse(cur);
1486 nScavenged += deflatedcount;
1487 nInuse += cur->omInUseCount;
1488 }
1489
1490 // For moribund threads, scan gOmInUseList
1491 if (gOmInUseList) {
1492 nInCirculation += gOmInUseCount;
1493 int deflatedcount = walk_monitor_list((ObjectMonitor **)&gOmInUseList, &FreeHead, &FreeTail);
1494 gOmInUseCount-= deflatedcount;
1495 nScavenged += deflatedcount;
1496 nInuse += gOmInUseCount;
1497 }
1498
1499 } else for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) {
1500 // Iterate over all extant monitors - Scavenge all idle monitors.
1501 assert(block->object() == CHAINMARKER, "must be a block header");
1502 nInCirculation += _BLOCKSIZE;
1503 for (int i = 1; i < _BLOCKSIZE; i++) {
1504 ObjectMonitor* mid = &block[i];
1505 oop obj = (oop) mid->object();
1506
1507 if (obj == NULL) {
1508 // The monitor is not associated with an object.
1509 // The monitor should either be a thread-specific private
1510 // free list or the global free list.
1511 // obj == NULL IMPLIES mid->is_busy() == 0
1512 guarantee(!mid->is_busy(), "invariant");
1513 continue;
1514 }
1515 deflated = deflate_monitor(mid, obj, &FreeHead, &FreeTail);
1516
1517 if (deflated) {
1518 mid->FreeNext = NULL;
1519 nScavenged++;
1520 } else {
1521 nInuse++;
1522 }
1523 }
1524 }
1525
1526 MonitorFreeCount += nScavenged;
1527
1528 // Consider: audit gFreeList to ensure that MonitorFreeCount and list agree.
1529
1530 if (ObjectMonitor::Knob_Verbose) {
1531 ::printf("Deflate: InCirc=%d InUse=%d Scavenged=%d ForceMonitorScavenge=%d : pop=%d free=%d\n",
1532 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge,
1533 MonitorPopulation, MonitorFreeCount);
1534 ::fflush(stdout);
1535 }
1536
1537 ForceMonitorScavenge = 0; // Reset
1538
1539 // Move the scavenged monitors back to the global free list.
1540 if (FreeHead != NULL) {
1541 guarantee(FreeTail != NULL && nScavenged > 0, "invariant");
1542 assert(FreeTail->FreeNext == NULL, "invariant");
1543 // constant-time list splice - prepend scavenged segment to gFreeList
1544 FreeTail->FreeNext = gFreeList;
1545 gFreeList = FreeHead;
1546 }
1547 Thread::muxRelease(&ListLock);
1548
1549 if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged);
1550 if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation);
1551
1552 // TODO: Add objectMonitor leak detection.
1553 // Audit/inventory the objectMonitors -- make sure they're all accounted for.
1554 GVars.stwRandom = os::random();
1555 GVars.stwCycle++;
1556 }
1557
1558 // Monitor cleanup on JavaThread::exit
1559
1560 // Iterate through monitor cache and attempt to release thread's monitors
1561 // Gives up on a particular monitor if an exception occurs, but continues
1562 // the overall iteration, swallowing the exception.
1563 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1564 private:
1565 TRAPS;
1566
1567 public:
1568 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {}
1569 void do_monitor(ObjectMonitor* mid) {
1570 if (mid->owner() == THREAD) {
1571 (void)mid->complete_exit(CHECK);
1572 }
1573 }
1574 };
1575
1576 // Release all inflated monitors owned by THREAD. Lightweight monitors are
1577 // ignored. This is meant to be called during JNI thread detach which assumes
1578 // all remaining monitors are heavyweight. All exceptions are swallowed.
1579 // Scanning the extant monitor list can be time consuming.
1580 // A simple optimization is to add a per-thread flag that indicates a thread
1581 // called jni_monitorenter() during its lifetime.
1582 //
1583 // Instead of No_Savepoint_Verifier it might be cheaper to
1584 // use an idiom of the form:
1585 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1586 // <code that must not run at safepoint>
1587 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1588 // Since the tests are extremely cheap we could leave them enabled
1589 // for normal product builds.
1590
1591 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) {
1592 assert(THREAD == JavaThread::current(), "must be current Java thread");
1593 No_Safepoint_Verifier nsv;
1594 ReleaseJavaMonitorsClosure rjmc(THREAD);
1595 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread");
1596 ObjectSynchronizer::monitors_iterate(&rjmc);
1597 Thread::muxRelease(&ListLock);
1598 THREAD->clear_pending_exception();
1599 }
1600
1601 //------------------------------------------------------------------------------
1602 // Non-product code
1603
1604 #ifndef PRODUCT
1605
1606 // Verify all monitors in the monitor cache, the verification is weak.
1607 void ObjectSynchronizer::verify() {
1608 ObjectMonitor* block = gBlockList;
1609 ObjectMonitor* mid;
1610 while (block) {
1611 assert(block->object() == CHAINMARKER, "must be a block header");
1612 for (int i = 1; i < _BLOCKSIZE; i++) {
1613 mid = block + i;
1614 oop object = (oop) mid->object();
1615 if (object != NULL) {
1616 mid->verify();
1617 }
1618 }
1619 block = (ObjectMonitor*) block->FreeNext;
1620 }
1621 }
1622
1623 // Check if monitor belongs to the monitor cache
1624 // The list is grow-only so it's *relatively* safe to traverse
1625 // the list of extant blocks without taking a lock.
1626
1627 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) {
1628 ObjectMonitor* block = gBlockList;
1629
1630 while (block) {
1631 assert(block->object() == CHAINMARKER, "must be a block header");
1632 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) {
1633 address mon = (address) monitor;
1634 address blk = (address) block;
1635 size_t diff = mon - blk;
1636 assert((diff % sizeof(ObjectMonitor)) == 0, "check");
1637 return 1;
1638 }
1639 block = (ObjectMonitor*) block->FreeNext;
1640 }
1641 return 0;
1642 }
1643
1644 #endif