1 /*
2 * Copyright (c) 1998, 2016, 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 "logging/log.hpp"
28 #include "memory/metaspaceShared.hpp"
29 #include "memory/padded.hpp"
30 #include "memory/resourceArea.hpp"
31 #include "oops/markOop.hpp"
32 #include "oops/oop.inline.hpp"
33 #include "runtime/atomic.hpp"
34 #include "runtime/biasedLocking.hpp"
35 #include "runtime/handles.inline.hpp"
36 #include "runtime/interfaceSupport.hpp"
37 #include "runtime/mutexLocker.hpp"
38 #include "runtime/objectMonitor.hpp"
39 #include "runtime/objectMonitor.inline.hpp"
40 #include "runtime/osThread.hpp"
41 #include "runtime/stubRoutines.hpp"
42 #include "runtime/synchronizer.hpp"
43 #include "runtime/thread.inline.hpp"
44 #include "runtime/vframe.hpp"
45 #include "trace/traceMacros.hpp"
46 #include "trace/tracing.hpp"
47 #include "utilities/dtrace.hpp"
48 #include "utilities/events.hpp"
49 #include "utilities/preserveException.hpp"
50
51 // The "core" versions of monitor enter and exit reside in this file.
52 // The interpreter and compilers contain specialized transliterated
53 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(),
54 // for instance. If you make changes here, make sure to modify the
55 // interpreter, and both C1 and C2 fast-path inline locking code emission.
56 //
57 // -----------------------------------------------------------------------------
58
59 #ifdef DTRACE_ENABLED
60
61 // Only bother with this argument setup if dtrace is available
62 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
63
64 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
65 char* bytes = NULL; \
66 int len = 0; \
67 jlong jtid = SharedRuntime::get_java_tid(thread); \
68 Symbol* klassname = ((oop)(obj))->klass()->name(); \
69 if (klassname != NULL) { \
70 bytes = (char*)klassname->bytes(); \
71 len = klassname->utf8_length(); \
72 }
73
74 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
75 { \
76 if (DTraceMonitorProbes) { \
77 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
78 HOTSPOT_MONITOR_WAIT(jtid, \
79 (uintptr_t)(monitor), bytes, len, (millis)); \
80 } \
81 }
82
83 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY
84 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL
85 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
86
87 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
88 { \
89 if (DTraceMonitorProbes) { \
90 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
91 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
92 (uintptr_t)(monitor), bytes, len); \
93 } \
94 }
95
96 #else // ndef DTRACE_ENABLED
97
98 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
99 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
100
101 #endif // ndef DTRACE_ENABLED
102
103 // This exists only as a workaround of dtrace bug 6254741
104 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
105 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
106 return 0;
107 }
108
109 #define NINFLATIONLOCKS 256
110 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS];
111
112 // global list of blocks of monitors
113 // gBlockList is really PaddedEnd<ObjectMonitor> *, but we don't
114 // want to expose the PaddedEnd template more than necessary.
115 ObjectMonitor * volatile ObjectSynchronizer::gBlockList = NULL;
116 // global monitor free list
117 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL;
118 // global monitor in-use list, for moribund threads,
119 // monitors they inflated need to be scanned for deflation
120 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL;
121 // count of entries in gOmInUseList
122 int ObjectSynchronizer::gOmInUseCount = 0;
123
124 static volatile intptr_t gListLock = 0; // protects global monitor lists
125 static volatile int gMonitorFreeCount = 0; // # on gFreeList
126 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation
127
128 static void post_monitor_inflate_event(EventJavaMonitorInflate&,
129 const oop,
130 const ObjectSynchronizer::InflateCause);
131
132 #define CHAINMARKER (cast_to_oop<intptr_t>(-1))
133 #define CLAIMEDMARKER (cast_to_oop<intptr_t>(-2))
134
135
136 // =====================> Quick functions
137
138 // The quick_* forms are special fast-path variants used to improve
139 // performance. In the simplest case, a "quick_*" implementation could
140 // simply return false, in which case the caller will perform the necessary
141 // state transitions and call the slow-path form.
142 // The fast-path is designed to handle frequently arising cases in an efficient
143 // manner and is just a degenerate "optimistic" variant of the slow-path.
144 // returns true -- to indicate the call was satisfied.
145 // returns false -- to indicate the call needs the services of the slow-path.
146 // A no-loitering ordinance is in effect for code in the quick_* family
147 // operators: safepoints or indefinite blocking (blocking that might span a
148 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon
149 // entry.
150 //
151 // Consider: An interesting optimization is to have the JIT recognize the
152 // following common idiom:
153 // synchronized (someobj) { .... ; notify(); }
154 // That is, we find a notify() or notifyAll() call that immediately precedes
155 // the monitorexit operation. In that case the JIT could fuse the operations
156 // into a single notifyAndExit() runtime primitive.
157
158 bool ObjectSynchronizer::quick_notify(oopDesc * obj, Thread * self, bool all) {
159 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
160 assert(self->is_Java_thread(), "invariant");
161 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant");
162 NoSafepointVerifier nsv;
163 if (obj == NULL) return false; // slow-path for invalid obj
164 const markOop mark = obj->mark();
165
166 if (mark->has_locker() && self->is_lock_owned((address)mark->locker())) {
167 // Degenerate notify
168 // stack-locked by caller so by definition the implied waitset is empty.
169 return true;
170 }
171
172 if (mark->has_monitor()) {
173 ObjectMonitor * const mon = mark->monitor();
174 assert(oopDesc::equals((oop) mon->object(), obj), "invariant");
175 if (mon->owner() != self) return false; // slow-path for IMS exception
176
177 if (mon->first_waiter() != NULL) {
178 // We have one or more waiters. Since this is an inflated monitor
179 // that we own, we can transfer one or more threads from the waitset
180 // to the entrylist here and now, avoiding the slow-path.
181 if (all) {
182 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self);
183 } else {
184 DTRACE_MONITOR_PROBE(notify, mon, obj, self);
185 }
186 int tally = 0;
187 do {
188 mon->INotify(self);
189 ++tally;
190 } while (mon->first_waiter() != NULL && all);
191 OM_PERFDATA_OP(Notifications, inc(tally));
192 }
193 return true;
194 }
195
196 // biased locking and any other IMS exception states take the slow-path
197 return false;
198 }
199
200
201 // The LockNode emitted directly at the synchronization site would have
202 // been too big if it were to have included support for the cases of inflated
203 // recursive enter and exit, so they go here instead.
204 // Note that we can't safely call AsyncPrintJavaStack() from within
205 // quick_enter() as our thread state remains _in_Java.
206
207 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self,
208 BasicLock * lock) {
209 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
210 assert(Self->is_Java_thread(), "invariant");
211 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant");
212 NoSafepointVerifier nsv;
213 if (obj == NULL) return false; // Need to throw NPE
214 const markOop mark = obj->mark();
215
216 if (mark->has_monitor()) {
217 ObjectMonitor * const m = mark->monitor();
218 assert(oopDesc::equals((oop) m->object(), obj), "invariant");
219 Thread * const owner = (Thread *) m->_owner;
220
221 // Lock contention and Transactional Lock Elision (TLE) diagnostics
222 // and observability
223 // Case: light contention possibly amenable to TLE
224 // Case: TLE inimical operations such as nested/recursive synchronization
225
226 if (owner == Self) {
227 m->_recursions++;
228 return true;
229 }
230
231 // This Java Monitor is inflated so obj's header will never be
232 // displaced to this thread's BasicLock. Make the displaced header
233 // non-NULL so this BasicLock is not seen as recursive nor as
234 // being locked. We do this unconditionally so that this thread's
235 // BasicLock cannot be mis-interpreted by any stack walkers. For
236 // performance reasons, stack walkers generally first check for
237 // Biased Locking in the object's header, the second check is for
238 // stack-locking in the object's header, the third check is for
239 // recursive stack-locking in the displaced header in the BasicLock,
240 // and last are the inflated Java Monitor (ObjectMonitor) checks.
241 lock->set_displaced_header(markOopDesc::unused_mark());
242
243 if (owner == NULL &&
244 Atomic::cmpxchg_ptr(Self, &(m->_owner), NULL) == NULL) {
245 assert(m->_recursions == 0, "invariant");
246 assert(m->_owner == Self, "invariant");
247 return true;
248 }
249 }
250
251 // Note that we could inflate in quick_enter.
252 // This is likely a useful optimization
253 // Critically, in quick_enter() we must not:
254 // -- perform bias revocation, or
255 // -- block indefinitely, or
256 // -- reach a safepoint
257
258 return false; // revert to slow-path
259 }
260
261 // -----------------------------------------------------------------------------
262 // Fast Monitor Enter/Exit
263 // This the fast monitor enter. The interpreter and compiler use
264 // some assembly copies of this code. Make sure update those code
265 // if the following function is changed. The implementation is
266 // extremely sensitive to race condition. Be careful.
267
268 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock,
269 bool attempt_rebias, TRAPS) {
270 if (UseBiasedLocking) {
271 if (!SafepointSynchronize::is_at_safepoint()) {
272 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD);
273 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) {
274 return;
275 }
276 } else {
277 assert(!attempt_rebias, "can not rebias toward VM thread");
278 BiasedLocking::revoke_at_safepoint(obj);
279 }
280 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
281 }
282
283 slow_enter(obj, lock, THREAD);
284 }
285
286 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) {
287 markOop mark = object->mark();
288 // We cannot check for Biased Locking if we are racing an inflation.
289 assert(mark == markOopDesc::INFLATING() ||
290 !mark->has_bias_pattern(), "should not see bias pattern here");
291
292 markOop dhw = lock->displaced_header();
293 if (dhw == NULL) {
294 // If the displaced header is NULL, then this exit matches up with
295 // a recursive enter. No real work to do here except for diagnostics.
296 #ifndef PRODUCT
297 if (mark != markOopDesc::INFLATING()) {
298 // Only do diagnostics if we are not racing an inflation. Simply
299 // exiting a recursive enter of a Java Monitor that is being
300 // inflated is safe; see the has_monitor() comment below.
301 assert(!mark->is_neutral(), "invariant");
302 assert(!mark->has_locker() ||
303 THREAD->is_lock_owned((address)mark->locker()), "invariant");
304 if (mark->has_monitor()) {
305 // The BasicLock's displaced_header is marked as a recursive
306 // enter and we have an inflated Java Monitor (ObjectMonitor).
307 // This is a special case where the Java Monitor was inflated
308 // after this thread entered the stack-lock recursively. When a
309 // Java Monitor is inflated, we cannot safely walk the Java
310 // Monitor owner's stack and update the BasicLocks because a
311 // Java Monitor can be asynchronously inflated by a thread that
312 // does not own the Java Monitor.
313 ObjectMonitor * m = mark->monitor();
314 assert(((oop)(m->object()))->mark() == mark, "invariant");
315 assert(m->is_entered(THREAD), "invariant");
316 }
317 }
318 #endif
319 return;
320 }
321
322 if (mark == (markOop) lock) {
323 // If the object is stack-locked by the current thread, try to
324 // swing the displaced header from the BasicLock back to the mark.
325 assert(dhw->is_neutral(), "invariant");
326 if (object->cas_set_mark(dhw, mark) == mark) {
327 TEVENT(fast_exit: release stack-lock);
328 return;
329 }
330 }
331
332 // We have to take the slow-path of possible inflation and then exit.
333 ObjectSynchronizer::inflate(THREAD,
334 object,
335 inflate_cause_vm_internal)->exit(true, THREAD);
336 }
337
338 // -----------------------------------------------------------------------------
339 // Interpreter/Compiler Slow Case
340 // This routine is used to handle interpreter/compiler slow case
341 // We don't need to use fast path here, because it must have been
342 // failed in the interpreter/compiler code.
343 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
344 markOop mark = obj->mark();
345 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
346
347 if (mark->is_neutral()) {
348 // Anticipate successful CAS -- the ST of the displaced mark must
349 // be visible <= the ST performed by the CAS.
350 lock->set_displaced_header(mark);
351 if (mark == obj()->cas_set_mark((markOop) lock, mark)) {
352 TEVENT(slow_enter: release stacklock);
353 return;
354 }
355 // Fall through to inflate() ...
356 } else if (mark->has_locker() &&
357 THREAD->is_lock_owned((address)mark->locker())) {
358 assert(lock != mark->locker(), "must not re-lock the same lock");
359 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
360 lock->set_displaced_header(NULL);
361 return;
362 }
363
364 // The object header will never be displaced to this lock,
365 // so it does not matter what the value is, except that it
366 // must be non-zero to avoid looking like a re-entrant lock,
367 // and must not look locked either.
368 lock->set_displaced_header(markOopDesc::unused_mark());
369 ObjectSynchronizer::inflate(THREAD,
370 obj(),
371 inflate_cause_monitor_enter)->enter(THREAD);
372 }
373
374 // This routine is used to handle interpreter/compiler slow case
375 // We don't need to use fast path here, because it must have
376 // failed in the interpreter/compiler code. Simply use the heavy
377 // weight monitor should be ok, unless someone find otherwise.
378 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
379 fast_exit(object, lock, THREAD);
380 }
381
382 // -----------------------------------------------------------------------------
383 // Class Loader support to workaround deadlocks on the class loader lock objects
384 // Also used by GC
385 // complete_exit()/reenter() are used to wait on a nested lock
386 // i.e. to give up an outer lock completely and then re-enter
387 // Used when holding nested locks - lock acquisition order: lock1 then lock2
388 // 1) complete_exit lock1 - saving recursion count
389 // 2) wait on lock2
390 // 3) when notified on lock2, unlock lock2
391 // 4) reenter lock1 with original recursion count
392 // 5) lock lock2
393 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
394 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
395 TEVENT(complete_exit);
396 if (UseBiasedLocking) {
397 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
398 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
399 }
400
401 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
402 obj(),
403 inflate_cause_vm_internal);
404
405 return monitor->complete_exit(THREAD);
406 }
407
408 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
409 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
410 TEVENT(reenter);
411 if (UseBiasedLocking) {
412 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
413 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
414 }
415
416 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
417 obj(),
418 inflate_cause_vm_internal);
419
420 monitor->reenter(recursion, THREAD);
421 }
422 // -----------------------------------------------------------------------------
423 // JNI locks on java objects
424 // NOTE: must use heavy weight monitor to handle jni monitor enter
425 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) {
426 // the current locking is from JNI instead of Java code
427 TEVENT(jni_enter);
428 if (UseBiasedLocking) {
429 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
430 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
431 }
432 THREAD->set_current_pending_monitor_is_from_java(false);
433 ObjectSynchronizer::inflate(THREAD, obj(), inflate_cause_jni_enter)->enter(THREAD);
434 THREAD->set_current_pending_monitor_is_from_java(true);
435 }
436
437 // NOTE: must use heavy weight monitor to handle jni monitor exit
438 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
439 TEVENT(jni_exit);
440 if (UseBiasedLocking) {
441 Handle h_obj(THREAD, obj);
442 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD);
443 obj = h_obj();
444 }
445 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
446
447 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
448 obj,
449 inflate_cause_jni_exit);
450 // If this thread has locked the object, exit the monitor. Note: can't use
451 // monitor->check(CHECK); must exit even if an exception is pending.
452 if (monitor->check(THREAD)) {
453 monitor->exit(true, THREAD);
454 }
455 }
456
457 // -----------------------------------------------------------------------------
458 // Internal VM locks on java objects
459 // standard constructor, allows locking failures
460 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
461 _dolock = doLock;
462 _thread = thread;
463 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
464 _obj = obj;
465
466 if (_dolock) {
467 TEVENT(ObjectLocker);
468
469 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
470 }
471 }
472
473 ObjectLocker::~ObjectLocker() {
474 if (_dolock) {
475 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
476 }
477 }
478
479
480 // -----------------------------------------------------------------------------
481 // Wait/Notify/NotifyAll
482 // NOTE: must use heavy weight monitor to handle wait()
483 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
484 if (UseBiasedLocking) {
485 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
486 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
487 }
488 if (millis < 0) {
489 TEVENT(wait - throw IAX);
490 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
491 }
492 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
493 obj(),
494 inflate_cause_wait);
495
496 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
497 monitor->wait(millis, true, THREAD);
498
499 // This dummy call is in place to get around dtrace bug 6254741. Once
500 // that's fixed we can uncomment the following line, remove the call
501 // and change this function back into a "void" func.
502 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
503 return dtrace_waited_probe(monitor, obj, THREAD);
504 }
505
506 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
507 if (UseBiasedLocking) {
508 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
509 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
510 }
511 if (millis < 0) {
512 TEVENT(wait - throw IAX);
513 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
514 }
515 ObjectSynchronizer::inflate(THREAD,
516 obj(),
517 inflate_cause_wait)->wait(millis, false, THREAD);
518 }
519
520 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
521 if (UseBiasedLocking) {
522 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
523 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
524 }
525
526 markOop mark = obj->mark();
527 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
528 return;
529 }
530 ObjectSynchronizer::inflate(THREAD,
531 obj(),
532 inflate_cause_notify)->notify(THREAD);
533 }
534
535 // NOTE: see comment of notify()
536 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
537 if (UseBiasedLocking) {
538 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
539 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
540 }
541
542 markOop mark = obj->mark();
543 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
544 return;
545 }
546 ObjectSynchronizer::inflate(THREAD,
547 obj(),
548 inflate_cause_notify)->notifyAll(THREAD);
549 }
550
551 // -----------------------------------------------------------------------------
552 // Hash Code handling
553 //
554 // Performance concern:
555 // OrderAccess::storestore() calls release() which at one time stored 0
556 // into the global volatile OrderAccess::dummy variable. This store was
557 // unnecessary for correctness. Many threads storing into a common location
558 // causes considerable cache migration or "sloshing" on large SMP systems.
559 // As such, I avoided using OrderAccess::storestore(). In some cases
560 // OrderAccess::fence() -- which incurs local latency on the executing
561 // processor -- is a better choice as it scales on SMP systems.
562 //
563 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for
564 // a discussion of coherency costs. Note that all our current reference
565 // platforms provide strong ST-ST order, so the issue is moot on IA32,
566 // x64, and SPARC.
567 //
568 // As a general policy we use "volatile" to control compiler-based reordering
569 // and explicit fences (barriers) to control for architectural reordering
570 // performed by the CPU(s) or platform.
571
572 struct SharedGlobals {
573 char _pad_prefix[DEFAULT_CACHE_LINE_SIZE];
574 // These are highly shared mostly-read variables.
575 // To avoid false-sharing they need to be the sole occupants of a cache line.
576 volatile int stwRandom;
577 volatile int stwCycle;
578 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int) * 2);
579 // Hot RW variable -- Sequester to avoid false-sharing
580 volatile int hcSequence;
581 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int));
582 };
583
584 static SharedGlobals GVars;
585 static int MonitorScavengeThreshold = 1000000;
586 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending
587
588 static markOop ReadStableMark(oop obj) {
589 markOop mark = obj->mark();
590 if (!mark->is_being_inflated()) {
591 return mark; // normal fast-path return
592 }
593
594 int its = 0;
595 for (;;) {
596 markOop mark = obj->mark();
597 if (!mark->is_being_inflated()) {
598 return mark; // normal fast-path return
599 }
600
601 // The object is being inflated by some other thread.
602 // The caller of ReadStableMark() must wait for inflation to complete.
603 // Avoid live-lock
604 // TODO: consider calling SafepointSynchronize::do_call_back() while
605 // spinning to see if there's a safepoint pending. If so, immediately
606 // yielding or blocking would be appropriate. Avoid spinning while
607 // there is a safepoint pending.
608 // TODO: add inflation contention performance counters.
609 // TODO: restrict the aggregate number of spinners.
610
611 ++its;
612 if (its > 10000 || !os::is_MP()) {
613 if (its & 1) {
614 os::naked_yield();
615 TEVENT(Inflate: INFLATING - yield);
616 } else {
617 // Note that the following code attenuates the livelock problem but is not
618 // a complete remedy. A more complete solution would require that the inflating
619 // thread hold the associated inflation lock. The following code simply restricts
620 // the number of spinners to at most one. We'll have N-2 threads blocked
621 // on the inflationlock, 1 thread holding the inflation lock and using
622 // a yield/park strategy, and 1 thread in the midst of inflation.
623 // A more refined approach would be to change the encoding of INFLATING
624 // to allow encapsulation of a native thread pointer. Threads waiting for
625 // inflation to complete would use CAS to push themselves onto a singly linked
626 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
627 // and calling park(). When inflation was complete the thread that accomplished inflation
628 // would detach the list and set the markword to inflated with a single CAS and
629 // then for each thread on the list, set the flag and unpark() the thread.
630 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease
631 // wakes at most one thread whereas we need to wake the entire list.
632 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1);
633 int YieldThenBlock = 0;
634 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant");
635 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant");
636 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock");
637 while (obj->mark() == markOopDesc::INFLATING()) {
638 // Beware: NakedYield() is advisory and has almost no effect on some platforms
639 // so we periodically call Self->_ParkEvent->park(1).
640 // We use a mixed spin/yield/block mechanism.
641 if ((YieldThenBlock++) >= 16) {
642 Thread::current()->_ParkEvent->park(1);
643 } else {
644 os::naked_yield();
645 }
646 }
647 Thread::muxRelease(gInflationLocks + ix);
648 TEVENT(Inflate: INFLATING - yield/park);
649 }
650 } else {
651 SpinPause(); // SMP-polite spinning
652 }
653 }
654 }
655
656 // hashCode() generation :
657 //
658 // Possibilities:
659 // * MD5Digest of {obj,stwRandom}
660 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function.
661 // * A DES- or AES-style SBox[] mechanism
662 // * One of the Phi-based schemes, such as:
663 // 2654435761 = 2^32 * Phi (golden ratio)
664 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ;
665 // * A variation of Marsaglia's shift-xor RNG scheme.
666 // * (obj ^ stwRandom) is appealing, but can result
667 // in undesirable regularity in the hashCode values of adjacent objects
668 // (objects allocated back-to-back, in particular). This could potentially
669 // result in hashtable collisions and reduced hashtable efficiency.
670 // There are simple ways to "diffuse" the middle address bits over the
671 // generated hashCode values:
672
673 static inline intptr_t get_next_hash(Thread * Self, oop obj) {
674 intptr_t value = 0;
675 if (hashCode == 0) {
676 // This form uses an unguarded global Park-Miller RNG,
677 // so it's possible for two threads to race and generate the same RNG.
678 // On MP system we'll have lots of RW access to a global, so the
679 // mechanism induces lots of coherency traffic.
680 value = os::random();
681 } else if (hashCode == 1) {
682 // This variation has the property of being stable (idempotent)
683 // between STW operations. This can be useful in some of the 1-0
684 // synchronization schemes.
685 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3;
686 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom;
687 } else if (hashCode == 2) {
688 value = 1; // for sensitivity testing
689 } else if (hashCode == 3) {
690 value = ++GVars.hcSequence;
691 } else if (hashCode == 4) {
692 value = cast_from_oop<intptr_t>(obj);
693 } else {
694 // Marsaglia's xor-shift scheme with thread-specific state
695 // This is probably the best overall implementation -- we'll
696 // likely make this the default in future releases.
697 unsigned t = Self->_hashStateX;
698 t ^= (t << 11);
699 Self->_hashStateX = Self->_hashStateY;
700 Self->_hashStateY = Self->_hashStateZ;
701 Self->_hashStateZ = Self->_hashStateW;
702 unsigned v = Self->_hashStateW;
703 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
704 Self->_hashStateW = v;
705 value = v;
706 }
707
708 value &= markOopDesc::hash_mask;
709 if (value == 0) value = 0xBAD;
710 assert(value != markOopDesc::no_hash, "invariant");
711 TEVENT(hashCode: GENERATE);
712 return value;
713 }
714
715 intptr_t ObjectSynchronizer::FastHashCode(Thread * Self, oop obj) {
716 if (UseBiasedLocking) {
717 // NOTE: many places throughout the JVM do not expect a safepoint
718 // to be taken here, in particular most operations on perm gen
719 // objects. However, we only ever bias Java instances and all of
720 // the call sites of identity_hash that might revoke biases have
721 // been checked to make sure they can handle a safepoint. The
722 // added check of the bias pattern is to avoid useless calls to
723 // thread-local storage.
724 if (obj->mark()->has_bias_pattern()) {
725 // Handle for oop obj in case of STW safepoint
726 Handle hobj(Self, obj);
727 // Relaxing assertion for bug 6320749.
728 assert(Universe::verify_in_progress() ||
729 !SafepointSynchronize::is_at_safepoint(),
730 "biases should not be seen by VM thread here");
731 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
732 obj = hobj();
733 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
734 }
735 }
736
737 // hashCode() is a heap mutator ...
738 // Relaxing assertion for bug 6320749.
739 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
740 !SafepointSynchronize::is_at_safepoint(), "invariant");
741 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
742 Self->is_Java_thread() , "invariant");
743 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
744 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
745
746 ObjectMonitor* monitor = NULL;
747 markOop temp, test;
748 intptr_t hash;
749 markOop mark = ReadStableMark(obj);
750
751 // object should remain ineligible for biased locking
752 assert(!mark->has_bias_pattern(), "invariant");
753
754 if (mark->is_neutral()) {
755 hash = mark->hash(); // this is a normal header
756 if (hash) { // if it has hash, just return it
757 return hash;
758 }
759 hash = get_next_hash(Self, obj); // allocate a new hash code
760 temp = mark->copy_set_hash(hash); // merge the hash code into header
761 // use (machine word version) atomic operation to install the hash
762 test = obj->cas_set_mark(temp, mark);
763 if (test == mark) {
764 return hash;
765 }
766 // If atomic operation failed, we must inflate the header
767 // into heavy weight monitor. We could add more code here
768 // for fast path, but it does not worth the complexity.
769 } else if (mark->has_monitor()) {
770 monitor = mark->monitor();
771 temp = monitor->header();
772 assert(temp->is_neutral(), "invariant");
773 hash = temp->hash();
774 if (hash) {
775 return hash;
776 }
777 // Skip to the following code to reduce code size
778 } else if (Self->is_lock_owned((address)mark->locker())) {
779 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
780 assert(temp->is_neutral(), "invariant");
781 hash = temp->hash(); // by current thread, check if the displaced
782 if (hash) { // header contains hash code
783 return hash;
784 }
785 // WARNING:
786 // The displaced header is strictly immutable.
787 // It can NOT be changed in ANY cases. So we have
788 // to inflate the header into heavyweight monitor
789 // even the current thread owns the lock. The reason
790 // is the BasicLock (stack slot) will be asynchronously
791 // read by other threads during the inflate() function.
792 // Any change to stack may not propagate to other threads
793 // correctly.
794 }
795
796 // Inflate the monitor to set hash code
797 monitor = ObjectSynchronizer::inflate(Self, obj, inflate_cause_hash_code);
798 // Load displaced header and check it has hash code
799 mark = monitor->header();
800 assert(mark->is_neutral(), "invariant");
801 hash = mark->hash();
802 if (hash == 0) {
803 hash = get_next_hash(Self, obj);
804 temp = mark->copy_set_hash(hash); // merge hash code into header
805 assert(temp->is_neutral(), "invariant");
806 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark);
807 if (test != mark) {
808 // The only update to the header in the monitor (outside GC)
809 // is install the hash code. If someone add new usage of
810 // displaced header, please update this code
811 hash = test->hash();
812 assert(test->is_neutral(), "invariant");
813 assert(hash != 0, "Trivial unexpected object/monitor header usage.");
814 }
815 }
816 // We finally get the hash
817 return hash;
818 }
819
820 // Deprecated -- use FastHashCode() instead.
821
822 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
823 return FastHashCode(Thread::current(), obj());
824 }
825
826
827 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
828 Handle h_obj) {
829 if (UseBiasedLocking) {
830 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
831 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
832 }
833
834 assert(thread == JavaThread::current(), "Can only be called on current thread");
835 oop obj = h_obj();
836
837 markOop mark = ReadStableMark(obj);
838
839 // Uncontended case, header points to stack
840 if (mark->has_locker()) {
841 return thread->is_lock_owned((address)mark->locker());
842 }
843 // Contended case, header points to ObjectMonitor (tagged pointer)
844 if (mark->has_monitor()) {
845 ObjectMonitor* monitor = mark->monitor();
846 return monitor->is_entered(thread) != 0;
847 }
848 // Unlocked case, header in place
849 assert(mark->is_neutral(), "sanity check");
850 return false;
851 }
852
853 // Be aware of this method could revoke bias of the lock object.
854 // This method queries the ownership of the lock handle specified by 'h_obj'.
855 // If the current thread owns the lock, it returns owner_self. If no
856 // thread owns the lock, it returns owner_none. Otherwise, it will return
857 // owner_other.
858 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
859 (JavaThread *self, Handle h_obj) {
860 // The caller must beware this method can revoke bias, and
861 // revocation can result in a safepoint.
862 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
863 assert(self->thread_state() != _thread_blocked, "invariant");
864
865 // Possible mark states: neutral, biased, stack-locked, inflated
866
867 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
868 // CASE: biased
869 BiasedLocking::revoke_and_rebias(h_obj, false, self);
870 assert(!h_obj->mark()->has_bias_pattern(),
871 "biases should be revoked by now");
872 }
873
874 assert(self == JavaThread::current(), "Can only be called on current thread");
875 oop obj = h_obj();
876 markOop mark = ReadStableMark(obj);
877
878 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
879 if (mark->has_locker()) {
880 return self->is_lock_owned((address)mark->locker()) ?
881 owner_self : owner_other;
882 }
883
884 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor.
885 // The Object:ObjectMonitor relationship is stable as long as we're
886 // not at a safepoint.
887 if (mark->has_monitor()) {
888 void * owner = mark->monitor()->_owner;
889 if (owner == NULL) return owner_none;
890 return (owner == self ||
891 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
892 }
893
894 // CASE: neutral
895 assert(mark->is_neutral(), "sanity check");
896 return owner_none; // it's unlocked
897 }
898
899 // FIXME: jvmti should call this
900 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) {
901 if (UseBiasedLocking) {
902 if (SafepointSynchronize::is_at_safepoint()) {
903 BiasedLocking::revoke_at_safepoint(h_obj);
904 } else {
905 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
906 }
907 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
908 }
909
910 oop obj = h_obj();
911 address owner = NULL;
912
913 markOop mark = ReadStableMark(obj);
914
915 // Uncontended case, header points to stack
916 if (mark->has_locker()) {
917 owner = (address) mark->locker();
918 }
919
920 // Contended case, header points to ObjectMonitor (tagged pointer)
921 if (mark->has_monitor()) {
922 ObjectMonitor* monitor = mark->monitor();
923 assert(monitor != NULL, "monitor should be non-null");
924 owner = (address) monitor->owner();
925 }
926
927 if (owner != NULL) {
928 // owning_thread_from_monitor_owner() may also return NULL here
929 return Threads::owning_thread_from_monitor_owner(owner, doLock);
930 }
931
932 // Unlocked case, header in place
933 // Cannot have assertion since this object may have been
934 // locked by another thread when reaching here.
935 // assert(mark->is_neutral(), "sanity check");
936
937 return NULL;
938 }
939
940 // Visitors ...
941
942 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
943 PaddedEnd<ObjectMonitor> * block =
944 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList);
945 while (block != NULL) {
946 assert(block->object() == CHAINMARKER, "must be a block header");
947 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
948 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
949 oop object = (oop)mid->object();
950 if (object != NULL) {
951 closure->do_monitor(mid);
952 }
953 }
954 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
955 }
956 }
957
958 // Get the next block in the block list.
959 static inline ObjectMonitor* next(ObjectMonitor* block) {
960 assert(block->object() == CHAINMARKER || block->object() == CLAIMEDMARKER, "must be a valid block header");
961 block = block->FreeNext;
962 assert(block == NULL || block->object() == CHAINMARKER || block->object() == CLAIMEDMARKER, "must be a valid block header");
963 return block;
964 }
965
966
967 void ObjectSynchronizer::oops_do(OopClosure* f) {
968 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
969 PaddedEnd<ObjectMonitor> * block =
970 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList);
971 for (; block != NULL; block = (PaddedEnd<ObjectMonitor> *)next(block)) {
972 assert(block->object() == CHAINMARKER, "must be a block header");
973 for (int i = 1; i < _BLOCKSIZE; i++) {
974 ObjectMonitor* mid = (ObjectMonitor *)&block[i];
975 if (mid->object() != NULL) {
976 f->do_oop((oop*)mid->object_addr());
977 }
978 }
979 }
980 }
981
982
983 // -----------------------------------------------------------------------------
984 // ObjectMonitor Lifecycle
985 // -----------------------
986 // Inflation unlinks monitors from the global gFreeList and
987 // associates them with objects. Deflation -- which occurs at
988 // STW-time -- disassociates idle monitors from objects. Such
989 // scavenged monitors are returned to the gFreeList.
990 //
991 // The global list is protected by gListLock. All the critical sections
992 // are short and operate in constant-time.
993 //
994 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
995 //
996 // Lifecycle:
997 // -- unassigned and on the global free list
998 // -- unassigned and on a thread's private omFreeList
999 // -- assigned to an object. The object is inflated and the mark refers
1000 // to the objectmonitor.
1001
1002
1003 // Constraining monitor pool growth via MonitorBound ...
1004 //
1005 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
1006 // the rate of scavenging is driven primarily by GC. As such, we can find
1007 // an inordinate number of monitors in circulation.
1008 // To avoid that scenario we can artificially induce a STW safepoint
1009 // if the pool appears to be growing past some reasonable bound.
1010 // Generally we favor time in space-time tradeoffs, but as there's no
1011 // natural back-pressure on the # of extant monitors we need to impose some
1012 // type of limit. Beware that if MonitorBound is set to too low a value
1013 // we could just loop. In addition, if MonitorBound is set to a low value
1014 // we'll incur more safepoints, which are harmful to performance.
1015 // See also: GuaranteedSafepointInterval
1016 //
1017 // The current implementation uses asynchronous VM operations.
1018
1019 static void InduceScavenge(Thread * Self, const char * Whence) {
1020 // Induce STW safepoint to trim monitors
1021 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
1022 // More precisely, trigger an asynchronous STW safepoint as the number
1023 // of active monitors passes the specified threshold.
1024 // TODO: assert thread state is reasonable
1025
1026 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
1027 if (ObjectMonitor::Knob_Verbose) {
1028 tty->print_cr("INFO: Monitor scavenge - Induced STW @%s (%d)",
1029 Whence, ForceMonitorScavenge) ;
1030 tty->flush();
1031 }
1032 // Induce a 'null' safepoint to scavenge monitors
1033 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
1034 // to the VMthread and have a lifespan longer than that of this activation record.
1035 // The VMThread will delete the op when completed.
1036 VMThread::execute(new VM_ForceAsyncSafepoint());
1037
1038 if (ObjectMonitor::Knob_Verbose) {
1039 tty->print_cr("INFO: Monitor scavenge - STW posted @%s (%d)",
1040 Whence, ForceMonitorScavenge) ;
1041 tty->flush();
1042 }
1043 }
1044 }
1045
1046 void ObjectSynchronizer::verifyInUse(Thread *Self) {
1047 ObjectMonitor* mid;
1048 int in_use_tally = 0;
1049 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) {
1050 in_use_tally++;
1051 }
1052 assert(in_use_tally == Self->omInUseCount, "in-use count off");
1053
1054 int free_tally = 0;
1055 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) {
1056 free_tally++;
1057 }
1058 assert(free_tally == Self->omFreeCount, "free count off");
1059 }
1060
1061 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) {
1062 // A large MAXPRIVATE value reduces both list lock contention
1063 // and list coherency traffic, but also tends to increase the
1064 // number of objectMonitors in circulation as well as the STW
1065 // scavenge costs. As usual, we lean toward time in space-time
1066 // tradeoffs.
1067 const int MAXPRIVATE = 1024;
1068 for (;;) {
1069 ObjectMonitor * m;
1070
1071 // 1: try to allocate from the thread's local omFreeList.
1072 // Threads will attempt to allocate first from their local list, then
1073 // from the global list, and only after those attempts fail will the thread
1074 // attempt to instantiate new monitors. Thread-local free lists take
1075 // heat off the gListLock and improve allocation latency, as well as reducing
1076 // coherency traffic on the shared global list.
1077 m = Self->omFreeList;
1078 if (m != NULL) {
1079 Self->omFreeList = m->FreeNext;
1080 Self->omFreeCount--;
1081 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene
1082 guarantee(m->object() == NULL, "invariant");
1083 if (MonitorInUseLists) {
1084 m->FreeNext = Self->omInUseList;
1085 Self->omInUseList = m;
1086 Self->omInUseCount++;
1087 if (ObjectMonitor::Knob_VerifyInUse) {
1088 verifyInUse(Self);
1089 }
1090 } else {
1091 m->FreeNext = NULL;
1092 }
1093 return m;
1094 }
1095
1096 // 2: try to allocate from the global gFreeList
1097 // CONSIDER: use muxTry() instead of muxAcquire().
1098 // If the muxTry() fails then drop immediately into case 3.
1099 // If we're using thread-local free lists then try
1100 // to reprovision the caller's free list.
1101 if (gFreeList != NULL) {
1102 // Reprovision the thread's omFreeList.
1103 // Use bulk transfers to reduce the allocation rate and heat
1104 // on various locks.
1105 Thread::muxAcquire(&gListLock, "omAlloc");
1106 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) {
1107 gMonitorFreeCount--;
1108 ObjectMonitor * take = gFreeList;
1109 gFreeList = take->FreeNext;
1110 guarantee(take->object() == NULL, "invariant");
1111 guarantee(!take->is_busy(), "invariant");
1112 take->Recycle();
1113 omRelease(Self, take, false);
1114 }
1115 Thread::muxRelease(&gListLock);
1116 Self->omFreeProvision += 1 + (Self->omFreeProvision/2);
1117 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE;
1118 TEVENT(omFirst - reprovision);
1119
1120 const int mx = MonitorBound;
1121 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) {
1122 // We can't safely induce a STW safepoint from omAlloc() as our thread
1123 // state may not be appropriate for such activities and callers may hold
1124 // naked oops, so instead we defer the action.
1125 InduceScavenge(Self, "omAlloc");
1126 }
1127 continue;
1128 }
1129
1130 // 3: allocate a block of new ObjectMonitors
1131 // Both the local and global free lists are empty -- resort to malloc().
1132 // In the current implementation objectMonitors are TSM - immortal.
1133 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want
1134 // each ObjectMonitor to start at the beginning of a cache line,
1135 // so we use align_size_up().
1136 // A better solution would be to use C++ placement-new.
1137 // BEWARE: As it stands currently, we don't run the ctors!
1138 assert(_BLOCKSIZE > 1, "invariant");
1139 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE;
1140 PaddedEnd<ObjectMonitor> * temp;
1141 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1);
1142 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size,
1143 mtInternal);
1144 temp = (PaddedEnd<ObjectMonitor> *)
1145 align_size_up((intptr_t)real_malloc_addr,
1146 DEFAULT_CACHE_LINE_SIZE);
1147
1148 // NOTE: (almost) no way to recover if allocation failed.
1149 // We might be able to induce a STW safepoint and scavenge enough
1150 // objectMonitors to permit progress.
1151 if (temp == NULL) {
1152 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR,
1153 "Allocate ObjectMonitors");
1154 }
1155 (void)memset((void *) temp, 0, neededsize);
1156
1157 // Format the block.
1158 // initialize the linked list, each monitor points to its next
1159 // forming the single linked free list, the very first monitor
1160 // will points to next block, which forms the block list.
1161 // The trick of using the 1st element in the block as gBlockList
1162 // linkage should be reconsidered. A better implementation would
1163 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1164
1165 for (int i = 1; i < _BLOCKSIZE; i++) {
1166 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1];
1167 }
1168
1169 // terminate the last monitor as the end of list
1170 temp[_BLOCKSIZE - 1].FreeNext = NULL;
1171
1172 // Element [0] is reserved for global list linkage
1173 temp[0].set_object(CHAINMARKER);
1174
1175 // Consider carving out this thread's current request from the
1176 // block in hand. This avoids some lock traffic and redundant
1177 // list activity.
1178
1179 // Acquire the gListLock to manipulate gBlockList and gFreeList.
1180 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1181 Thread::muxAcquire(&gListLock, "omAlloc [2]");
1182 gMonitorPopulation += _BLOCKSIZE-1;
1183 gMonitorFreeCount += _BLOCKSIZE-1;
1184
1185 // Add the new block to the list of extant blocks (gBlockList).
1186 // The very first objectMonitor in a block is reserved and dedicated.
1187 // It serves as blocklist "next" linkage.
1188 temp[0].FreeNext = gBlockList;
1189 // There are lock-free uses of gBlockList so make sure that
1190 // the previous stores happen before we update gBlockList.
1191 OrderAccess::release_store_ptr(&gBlockList, temp);
1192
1193 // Add the new string of objectMonitors to the global free list
1194 temp[_BLOCKSIZE - 1].FreeNext = gFreeList;
1195 gFreeList = temp + 1;
1196 Thread::muxRelease(&gListLock);
1197 TEVENT(Allocate block of monitors);
1198 }
1199 }
1200
1201 // Place "m" on the caller's private per-thread omFreeList.
1202 // In practice there's no need to clamp or limit the number of
1203 // monitors on a thread's omFreeList as the only time we'll call
1204 // omRelease is to return a monitor to the free list after a CAS
1205 // attempt failed. This doesn't allow unbounded #s of monitors to
1206 // accumulate on a thread's free list.
1207 //
1208 // Key constraint: all ObjectMonitors on a thread's free list and the global
1209 // free list must have their object field set to null. This prevents the
1210 // scavenger -- deflate_idle_monitors -- from reclaiming them.
1211
1212 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m,
1213 bool fromPerThreadAlloc) {
1214 guarantee(m->object() == NULL, "invariant");
1215 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor");
1216 // Remove from omInUseList
1217 if (MonitorInUseLists && fromPerThreadAlloc) {
1218 ObjectMonitor* cur_mid_in_use = NULL;
1219 bool extracted = false;
1220 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) {
1221 if (m == mid) {
1222 // extract from per-thread in-use list
1223 if (mid == Self->omInUseList) {
1224 Self->omInUseList = mid->FreeNext;
1225 } else if (cur_mid_in_use != NULL) {
1226 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1227 }
1228 extracted = true;
1229 Self->omInUseCount--;
1230 if (ObjectMonitor::Knob_VerifyInUse) {
1231 verifyInUse(Self);
1232 }
1233 break;
1234 }
1235 }
1236 assert(extracted, "Should have extracted from in-use list");
1237 }
1238
1239 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new
1240 m->FreeNext = Self->omFreeList;
1241 Self->omFreeList = m;
1242 Self->omFreeCount++;
1243 }
1244
1245 // Return the monitors of a moribund thread's local free list to
1246 // the global free list. Typically a thread calls omFlush() when
1247 // it's dying. We could also consider having the VM thread steal
1248 // monitors from threads that have not run java code over a few
1249 // consecutive STW safepoints. Relatedly, we might decay
1250 // omFreeProvision at STW safepoints.
1251 //
1252 // Also return the monitors of a moribund thread's omInUseList to
1253 // a global gOmInUseList under the global list lock so these
1254 // will continue to be scanned.
1255 //
1256 // We currently call omFlush() from the Thread:: dtor _after the thread
1257 // has been excised from the thread list and is no longer a mutator.
1258 // That means that omFlush() can run concurrently with a safepoint and
1259 // the scavenge operator. Calling omFlush() from JavaThread::exit() might
1260 // be a better choice as we could safely reason that that the JVM is
1261 // not at a safepoint at the time of the call, and thus there could
1262 // be not inopportune interleavings between omFlush() and the scavenge
1263 // operator.
1264
1265 void ObjectSynchronizer::omFlush(Thread * Self) {
1266 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL
1267 Self->omFreeList = NULL;
1268 ObjectMonitor * tail = NULL;
1269 int tally = 0;
1270 if (list != NULL) {
1271 ObjectMonitor * s;
1272 // The thread is going away, the per-thread free monitors
1273 // are freed via set_owner(NULL)
1274 // Link them to tail, which will be linked into the global free list
1275 // gFreeList below, under the gListLock
1276 for (s = list; s != NULL; s = s->FreeNext) {
1277 tally++;
1278 tail = s;
1279 guarantee(s->object() == NULL, "invariant");
1280 guarantee(!s->is_busy(), "invariant");
1281 s->set_owner(NULL); // redundant but good hygiene
1282 TEVENT(omFlush - Move one);
1283 }
1284 guarantee(tail != NULL && list != NULL, "invariant");
1285 }
1286
1287 ObjectMonitor * inUseList = Self->omInUseList;
1288 ObjectMonitor * inUseTail = NULL;
1289 int inUseTally = 0;
1290 if (inUseList != NULL) {
1291 Self->omInUseList = NULL;
1292 ObjectMonitor *cur_om;
1293 // The thread is going away, however the omInUseList inflated
1294 // monitors may still be in-use by other threads.
1295 // Link them to inUseTail, which will be linked into the global in-use list
1296 // gOmInUseList below, under the gListLock
1297 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) {
1298 inUseTail = cur_om;
1299 inUseTally++;
1300 }
1301 assert(Self->omInUseCount == inUseTally, "in-use count off");
1302 Self->omInUseCount = 0;
1303 guarantee(inUseTail != NULL && inUseList != NULL, "invariant");
1304 }
1305
1306 Thread::muxAcquire(&gListLock, "omFlush");
1307 if (tail != NULL) {
1308 tail->FreeNext = gFreeList;
1309 gFreeList = list;
1310 gMonitorFreeCount += tally;
1311 }
1312
1313 if (inUseTail != NULL) {
1314 inUseTail->FreeNext = gOmInUseList;
1315 gOmInUseList = inUseList;
1316 gOmInUseCount += inUseTally;
1317 }
1318
1319 Thread::muxRelease(&gListLock);
1320 TEVENT(omFlush);
1321 }
1322
1323 // Fast path code shared by multiple functions
1324 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) {
1325 markOop mark = obj->mark();
1326 if (mark->has_monitor()) {
1327 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid");
1328 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header");
1329 return mark->monitor();
1330 }
1331 return ObjectSynchronizer::inflate(Thread::current(),
1332 obj,
1333 inflate_cause_vm_internal);
1334 }
1335
1336 ObjectMonitor* ObjectSynchronizer::inflate(Thread * Self,
1337 oop object,
1338 const InflateCause cause) {
1339
1340 // Inflate mutates the heap ...
1341 // Relaxing assertion for bug 6320749.
1342 assert(Universe::verify_in_progress() ||
1343 !SafepointSynchronize::is_at_safepoint(), "invariant");
1344
1345 EventJavaMonitorInflate event;
1346
1347 for (;;) {
1348 const markOop mark = object->mark();
1349 assert(!mark->has_bias_pattern(), "invariant");
1350
1351 // The mark can be in one of the following states:
1352 // * Inflated - just return
1353 // * Stack-locked - coerce it to inflated
1354 // * INFLATING - busy wait for conversion to complete
1355 // * Neutral - aggressively inflate the object.
1356 // * BIASED - Illegal. We should never see this
1357
1358 // CASE: inflated
1359 if (mark->has_monitor()) {
1360 ObjectMonitor * inf = mark->monitor();
1361 assert(inf->header()->is_neutral(), "invariant");
1362 assert(oopDesc::equals((oop) inf->object(), object), "invariant");
1363 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1364 event.cancel(); // let's not post an inflation event, unless we did the deed ourselves
1365 return inf;
1366 }
1367
1368 // CASE: inflation in progress - inflating over a stack-lock.
1369 // Some other thread is converting from stack-locked to inflated.
1370 // Only that thread can complete inflation -- other threads must wait.
1371 // The INFLATING value is transient.
1372 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1373 // We could always eliminate polling by parking the thread on some auxiliary list.
1374 if (mark == markOopDesc::INFLATING()) {
1375 TEVENT(Inflate: spin while INFLATING);
1376 ReadStableMark(object);
1377 continue;
1378 }
1379
1380 // CASE: stack-locked
1381 // Could be stack-locked either by this thread or by some other thread.
1382 //
1383 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1384 // to install INFLATING into the mark word. We originally installed INFLATING,
1385 // allocated the objectmonitor, and then finally STed the address of the
1386 // objectmonitor into the mark. This was correct, but artificially lengthened
1387 // the interval in which INFLATED appeared in the mark, thus increasing
1388 // the odds of inflation contention.
1389 //
1390 // We now use per-thread private objectmonitor free lists.
1391 // These list are reprovisioned from the global free list outside the
1392 // critical INFLATING...ST interval. A thread can transfer
1393 // multiple objectmonitors en-mass from the global free list to its local free list.
1394 // This reduces coherency traffic and lock contention on the global free list.
1395 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1396 // before or after the CAS(INFLATING) operation.
1397 // See the comments in omAlloc().
1398
1399 if (mark->has_locker()) {
1400 ObjectMonitor * m = omAlloc(Self);
1401 // Optimistically prepare the objectmonitor - anticipate successful CAS
1402 // We do this before the CAS in order to minimize the length of time
1403 // in which INFLATING appears in the mark.
1404 m->Recycle();
1405 m->_Responsible = NULL;
1406 m->_recursions = 0;
1407 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1408
1409 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark);
1410 if (cmp != mark) {
1411 omRelease(Self, m, true);
1412 continue; // Interference -- just retry
1413 }
1414
1415 // We've successfully installed INFLATING (0) into the mark-word.
1416 // This is the only case where 0 will appear in a mark-word.
1417 // Only the singular thread that successfully swings the mark-word
1418 // to 0 can perform (or more precisely, complete) inflation.
1419 //
1420 // Why do we CAS a 0 into the mark-word instead of just CASing the
1421 // mark-word from the stack-locked value directly to the new inflated state?
1422 // Consider what happens when a thread unlocks a stack-locked object.
1423 // It attempts to use CAS to swing the displaced header value from the
1424 // on-stack basiclock back into the object header. Recall also that the
1425 // header value (hashcode, etc) can reside in (a) the object header, or
1426 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1427 // header in an objectMonitor. The inflate() routine must copy the header
1428 // value from the basiclock on the owner's stack to the objectMonitor, all
1429 // the while preserving the hashCode stability invariants. If the owner
1430 // decides to release the lock while the value is 0, the unlock will fail
1431 // and control will eventually pass from slow_exit() to inflate. The owner
1432 // will then spin, waiting for the 0 value to disappear. Put another way,
1433 // the 0 causes the owner to stall if the owner happens to try to
1434 // drop the lock (restoring the header from the basiclock to the object)
1435 // while inflation is in-progress. This protocol avoids races that might
1436 // would otherwise permit hashCode values to change or "flicker" for an object.
1437 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1438 // 0 serves as a "BUSY" inflate-in-progress indicator.
1439
1440
1441 // fetch the displaced mark from the owner's stack.
1442 // The owner can't die or unwind past the lock while our INFLATING
1443 // object is in the mark. Furthermore the owner can't complete
1444 // an unlock on the object, either.
1445 markOop dmw = mark->displaced_mark_helper();
1446 assert(dmw->is_neutral(), "invariant");
1447
1448 // Setup monitor fields to proper values -- prepare the monitor
1449 m->set_header(dmw);
1450
1451 // Optimization: if the mark->locker stack address is associated
1452 // with this thread we could simply set m->_owner = Self.
1453 // Note that a thread can inflate an object
1454 // that it has stack-locked -- as might happen in wait() -- directly
1455 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1456 m->set_owner(mark->locker());
1457 m->set_object(object);
1458 // TODO-FIXME: assert BasicLock->dhw != 0.
1459
1460 // Must preserve store ordering. The monitor state must
1461 // be stable at the time of publishing the monitor address.
1462 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant");
1463 object->release_set_mark(markOopDesc::encode(m));
1464
1465 // Hopefully the performance counters are allocated on distinct cache lines
1466 // to avoid false sharing on MP systems ...
1467 OM_PERFDATA_OP(Inflations, inc());
1468 TEVENT(Inflate: overwrite stacklock);
1469 if (log_is_enabled(Debug, monitorinflation)) {
1470 if (object->is_instance()) {
1471 ResourceMark rm;
1472 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1473 p2i(object), p2i(object->mark()),
1474 object->klass()->external_name());
1475 }
1476 }
1477 if (event.should_commit()) {
1478 post_monitor_inflate_event(event, object, cause);
1479 }
1480 return m;
1481 }
1482
1483 // CASE: neutral
1484 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1485 // If we know we're inflating for entry it's better to inflate by swinging a
1486 // pre-locked objectMonitor pointer into the object header. A successful
1487 // CAS inflates the object *and* confers ownership to the inflating thread.
1488 // In the current implementation we use a 2-step mechanism where we CAS()
1489 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1490 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1491 // would be useful.
1492
1493 assert(mark->is_neutral(), "invariant");
1494 ObjectMonitor * m = omAlloc(Self);
1495 // prepare m for installation - set monitor to initial state
1496 m->Recycle();
1497 m->set_header(mark);
1498 m->set_owner(NULL);
1499 m->set_object(object);
1500 m->_recursions = 0;
1501 m->_Responsible = NULL;
1502 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1503
1504 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) {
1505 m->set_object(NULL);
1506 m->set_owner(NULL);
1507 m->Recycle();
1508 omRelease(Self, m, true);
1509 m = NULL;
1510 continue;
1511 // interference - the markword changed - just retry.
1512 // The state-transitions are one-way, so there's no chance of
1513 // live-lock -- "Inflated" is an absorbing state.
1514 }
1515
1516 // Hopefully the performance counters are allocated on distinct
1517 // cache lines to avoid false sharing on MP systems ...
1518 OM_PERFDATA_OP(Inflations, inc());
1519 TEVENT(Inflate: overwrite neutral);
1520 if (log_is_enabled(Debug, monitorinflation)) {
1521 if (object->is_instance()) {
1522 ResourceMark rm;
1523 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1524 p2i(object), p2i(object->mark()),
1525 object->klass()->external_name());
1526 }
1527 }
1528 if (event.should_commit()) {
1529 post_monitor_inflate_event(event, object, cause);
1530 }
1531 return m;
1532 }
1533 }
1534
1535
1536 // Deflate_idle_monitors() is called at all safepoints, immediately
1537 // after all mutators are stopped, but before any objects have moved.
1538 // It traverses the list of known monitors, deflating where possible.
1539 // The scavenged monitor are returned to the monitor free list.
1540 //
1541 // Beware that we scavenge at *every* stop-the-world point.
1542 // Having a large number of monitors in-circulation negatively
1543 // impacts the performance of some applications (e.g., PointBase).
1544 // Broadly, we want to minimize the # of monitors in circulation.
1545 //
1546 // We have added a flag, MonitorInUseLists, which creates a list
1547 // of active monitors for each thread. deflate_idle_monitors()
1548 // only scans the per-thread in-use lists. omAlloc() puts all
1549 // assigned monitors on the per-thread list. deflate_idle_monitors()
1550 // returns the non-busy monitors to the global free list.
1551 // When a thread dies, omFlush() adds the list of active monitors for
1552 // that thread to a global gOmInUseList acquiring the
1553 // global list lock. deflate_idle_monitors() acquires the global
1554 // list lock to scan for non-busy monitors to the global free list.
1555 // An alternative could have used a single global in-use list. The
1556 // downside would have been the additional cost of acquiring the global list lock
1557 // for every omAlloc().
1558 //
1559 // Perversely, the heap size -- and thus the STW safepoint rate --
1560 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1561 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1562 // This is an unfortunate aspect of this design.
1563
1564 enum ManifestConstants {
1565 ClearResponsibleAtSTW = 0
1566 };
1567
1568 // Deflate a single monitor if not in-use
1569 // Return true if deflated, false if in-use
1570 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1571 ObjectMonitor** freeHeadp,
1572 ObjectMonitor** freeTailp) {
1573 bool deflated;
1574 // Normal case ... The monitor is associated with obj.
1575 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant");
1576 guarantee(mid == obj->mark()->monitor(), "invariant");
1577 guarantee(mid->header()->is_neutral(), "invariant");
1578
1579 if (mid->is_busy()) {
1580 if (ClearResponsibleAtSTW) mid->_Responsible = NULL;
1581 deflated = false;
1582 } else {
1583 // Deflate the monitor if it is no longer being used
1584 // It's idle - scavenge and return to the global free list
1585 // plain old deflation ...
1586 TEVENT(deflate_idle_monitors - scavenge1);
1587 if (log_is_enabled(Debug, monitorinflation)) {
1588 if (obj->is_instance()) {
1589 ResourceMark rm;
1590 log_debug(monitorinflation)("Deflating object " INTPTR_FORMAT " , "
1591 "mark " INTPTR_FORMAT " , type %s",
1592 p2i(obj), p2i(obj->mark()),
1593 obj->klass()->external_name());
1594 }
1595 }
1596
1597 // Restore the header back to obj
1598 obj->release_set_mark(mid->header());
1599 mid->clear();
1600
1601 assert(mid->object() == NULL, "invariant");
1602
1603 // Move the object to the working free list defined by freeHeadp, freeTailp
1604 if (*freeHeadp == NULL) *freeHeadp = mid;
1605 if (*freeTailp != NULL) {
1606 ObjectMonitor * prevtail = *freeTailp;
1607 assert(prevtail->FreeNext == NULL, "cleaned up deflated?");
1608 prevtail->FreeNext = mid;
1609 }
1610 *freeTailp = mid;
1611 deflated = true;
1612 }
1613 return deflated;
1614 }
1615
1616 // Walk a given monitor list, and deflate idle monitors
1617 // The given list could be a per-thread list or a global list
1618 // Caller acquires gListLock
1619 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp,
1620 ObjectMonitor** freeHeadp,
1621 ObjectMonitor** freeTailp) {
1622 ObjectMonitor* mid;
1623 ObjectMonitor* next;
1624 ObjectMonitor* cur_mid_in_use = NULL;
1625 int deflated_count = 0;
1626
1627 for (mid = *listHeadp; mid != NULL;) {
1628 oop obj = (oop) mid->object();
1629 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) {
1630 // if deflate_monitor succeeded,
1631 // extract from per-thread in-use list
1632 if (mid == *listHeadp) {
1633 *listHeadp = mid->FreeNext;
1634 } else if (cur_mid_in_use != NULL) {
1635 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1636 }
1637 next = mid->FreeNext;
1638 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list
1639 mid = next;
1640 deflated_count++;
1641 } else {
1642 cur_mid_in_use = mid;
1643 mid = mid->FreeNext;
1644 }
1645 }
1646 return deflated_count;
1647 }
1648
1649 void ObjectSynchronizer::deflate_idle_monitors() {
1650 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1651 int nInuse = 0; // currently associated with objects
1652 int nInCirculation = 0; // extant
1653 int nScavenged = 0; // reclaimed
1654 bool deflated = false;
1655
1656 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1657 ObjectMonitor * freeTailp = NULL;
1658
1659 TEVENT(deflate_idle_monitors);
1660 // Prevent omFlush from changing mids in Thread dtor's during deflation
1661 // And in case the vm thread is acquiring a lock during a safepoint
1662 // See e.g. 6320749
1663 Thread::muxAcquire(&gListLock, "scavenge - return");
1664
1665 if (MonitorInUseLists) {
1666 int inUse = 0;
1667 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) {
1668 nInCirculation+= cur->omInUseCount;
1669 int deflated_count = deflate_monitor_list(cur->omInUseList_addr(), &freeHeadp, &freeTailp);
1670 cur->omInUseCount-= deflated_count;
1671 if (ObjectMonitor::Knob_VerifyInUse) {
1672 verifyInUse(cur);
1673 }
1674 nScavenged += deflated_count;
1675 nInuse += cur->omInUseCount;
1676 }
1677
1678 // For moribund threads, scan gOmInUseList
1679 if (gOmInUseList) {
1680 nInCirculation += gOmInUseCount;
1681 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp);
1682 gOmInUseCount-= deflated_count;
1683 nScavenged += deflated_count;
1684 nInuse += gOmInUseCount;
1685 }
1686
1687 } else {
1688 PaddedEnd<ObjectMonitor> * block =
1689 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList);
1690 for (; block != NULL; block = (PaddedEnd<ObjectMonitor> *)next(block)) {
1691 // Iterate over all extant monitors - Scavenge all idle monitors.
1692 assert(block->object() == CHAINMARKER, "must be a block header");
1693 nInCirculation += _BLOCKSIZE;
1694 for (int i = 1; i < _BLOCKSIZE; i++) {
1695 ObjectMonitor* mid = (ObjectMonitor*)&block[i];
1696 oop obj = (oop)mid->object();
1697
1698 if (obj == NULL) {
1699 // The monitor is not associated with an object.
1700 // The monitor should either be a thread-specific private
1701 // free list or the global free list.
1702 // obj == NULL IMPLIES mid->is_busy() == 0
1703 guarantee(!mid->is_busy(), "invariant");
1704 continue;
1705 }
1706 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp);
1707
1708 if (deflated) {
1709 mid->FreeNext = NULL;
1710 nScavenged++;
1711 } else {
1712 nInuse++;
1713 }
1714 }
1715 }
1716 }
1717
1718 gMonitorFreeCount += nScavenged;
1719
1720 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree.
1721
1722 if (ObjectMonitor::Knob_Verbose) {
1723 tty->print_cr("INFO: Deflate: InCirc=%d InUse=%d Scavenged=%d "
1724 "ForceMonitorScavenge=%d : pop=%d free=%d",
1725 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge,
1726 gMonitorPopulation, gMonitorFreeCount);
1727 tty->flush();
1728 }
1729
1730 ForceMonitorScavenge = 0; // Reset
1731
1732 // Move the scavenged monitors back to the global free list.
1733 if (freeHeadp != NULL) {
1734 guarantee(freeTailp != NULL && nScavenged > 0, "invariant");
1735 assert(freeTailp->FreeNext == NULL, "invariant");
1736 // constant-time list splice - prepend scavenged segment to gFreeList
1737 freeTailp->FreeNext = gFreeList;
1738 gFreeList = freeHeadp;
1739 }
1740 Thread::muxRelease(&gListLock);
1741
1742 OM_PERFDATA_OP(Deflations, inc(nScavenged));
1743 OM_PERFDATA_OP(MonExtant, set_value(nInCirculation));
1744
1745 // TODO: Add objectMonitor leak detection.
1746 // Audit/inventory the objectMonitors -- make sure they're all accounted for.
1747 GVars.stwRandom = os::random();
1748 GVars.stwCycle++;
1749 }
1750
1751 // Monitor cleanup on JavaThread::exit
1752
1753 // Iterate through monitor cache and attempt to release thread's monitors
1754 // Gives up on a particular monitor if an exception occurs, but continues
1755 // the overall iteration, swallowing the exception.
1756 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1757 private:
1758 TRAPS;
1759
1760 public:
1761 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {}
1762 void do_monitor(ObjectMonitor* mid) {
1763 if (mid->owner() == THREAD) {
1764 if (ObjectMonitor::Knob_VerifyMatch != 0) {
1765 ResourceMark rm;
1766 Handle obj((oop) mid->object());
1767 tty->print("INFO: unexpected locked object:");
1768 javaVFrame::print_locked_object_class_name(tty, obj, "locked");
1769 fatal("exiting JavaThread=" INTPTR_FORMAT
1770 " unexpectedly owns ObjectMonitor=" INTPTR_FORMAT,
1771 p2i(THREAD), p2i(mid));
1772 }
1773 (void)mid->complete_exit(CHECK);
1774 }
1775 }
1776 };
1777
1778 // Release all inflated monitors owned by THREAD. Lightweight monitors are
1779 // ignored. This is meant to be called during JNI thread detach which assumes
1780 // all remaining monitors are heavyweight. All exceptions are swallowed.
1781 // Scanning the extant monitor list can be time consuming.
1782 // A simple optimization is to add a per-thread flag that indicates a thread
1783 // called jni_monitorenter() during its lifetime.
1784 //
1785 // Instead of No_Savepoint_Verifier it might be cheaper to
1786 // use an idiom of the form:
1787 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1788 // <code that must not run at safepoint>
1789 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1790 // Since the tests are extremely cheap we could leave them enabled
1791 // for normal product builds.
1792
1793 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) {
1794 assert(THREAD == JavaThread::current(), "must be current Java thread");
1795 NoSafepointVerifier nsv;
1796 ReleaseJavaMonitorsClosure rjmc(THREAD);
1797 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread");
1798 ObjectSynchronizer::monitors_iterate(&rjmc);
1799 Thread::muxRelease(&gListLock);
1800 THREAD->clear_pending_exception();
1801 }
1802
1803 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1804 switch (cause) {
1805 case inflate_cause_vm_internal: return "VM Internal";
1806 case inflate_cause_monitor_enter: return "Monitor Enter";
1807 case inflate_cause_wait: return "Monitor Wait";
1808 case inflate_cause_notify: return "Monitor Notify";
1809 case inflate_cause_hash_code: return "Monitor Hash Code";
1810 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1811 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1812 default:
1813 ShouldNotReachHere();
1814 }
1815 return "Unknown";
1816 }
1817
1818 static void post_monitor_inflate_event(EventJavaMonitorInflate& event,
1819 const oop obj,
1820 const ObjectSynchronizer::InflateCause cause) {
1821 #if INCLUDE_TRACE
1822 assert(event.should_commit(), "check outside");
1823 event.set_monitorClass(obj->klass());
1824 event.set_address((TYPE_ADDRESS)(uintptr_t)(void*)obj);
1825 event.set_cause((u1)cause);
1826 event.commit();
1827 #endif
1828 }
1829
1830 //------------------------------------------------------------------------------
1831 // Debugging code
1832
1833 void ObjectSynchronizer::sanity_checks(const bool verbose,
1834 const uint cache_line_size,
1835 int *error_cnt_ptr,
1836 int *warning_cnt_ptr) {
1837 u_char *addr_begin = (u_char*)&GVars;
1838 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom;
1839 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence;
1840
1841 if (verbose) {
1842 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT,
1843 sizeof(SharedGlobals));
1844 }
1845
1846 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin);
1847 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom);
1848
1849 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin);
1850 if (verbose) {
1851 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence);
1852 }
1853
1854 if (cache_line_size != 0) {
1855 // We were able to determine the L1 data cache line size so
1856 // do some cache line specific sanity checks
1857
1858 if (offset_stwRandom < cache_line_size) {
1859 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer "
1860 "to the struct beginning than a cache line which permits "
1861 "false sharing.");
1862 (*warning_cnt_ptr)++;
1863 }
1864
1865 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) {
1866 tty->print_cr("WARNING: the SharedGlobals.stwRandom and "
1867 "SharedGlobals.hcSequence fields are closer than a cache "
1868 "line which permits false sharing.");
1869 (*warning_cnt_ptr)++;
1870 }
1871
1872 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) {
1873 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer "
1874 "to the struct end than a cache line which permits false "
1875 "sharing.");
1876 (*warning_cnt_ptr)++;
1877 }
1878 }
1879 }
1880
1881 #ifndef PRODUCT
1882
1883 // Verify all monitors in the monitor cache, the verification is weak.
1884 void ObjectSynchronizer::verify() {
1885 PaddedEnd<ObjectMonitor> * block =
1886 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList);
1887 while (block != NULL) {
1888 assert(block->object() == CHAINMARKER, "must be a block header");
1889 for (int i = 1; i < _BLOCKSIZE; i++) {
1890 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
1891 oop object = (oop)mid->object();
1892 if (object != NULL) {
1893 mid->verify();
1894 }
1895 }
1896 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
1897 }
1898 }
1899
1900 // Check if monitor belongs to the monitor cache
1901 // The list is grow-only so it's *relatively* safe to traverse
1902 // the list of extant blocks without taking a lock.
1903
1904 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) {
1905 PaddedEnd<ObjectMonitor> * block =
1906 (PaddedEnd<ObjectMonitor> *)OrderAccess::load_ptr_acquire(&gBlockList);
1907 while (block != NULL) {
1908 assert(block->object() == CHAINMARKER, "must be a block header");
1909 if (monitor > (ObjectMonitor *)&block[0] &&
1910 monitor < (ObjectMonitor *)&block[_BLOCKSIZE]) {
1911 address mon = (address)monitor;
1912 address blk = (address)block;
1913 size_t diff = mon - blk;
1914 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "must be aligned");
1915 return 1;
1916 }
1917 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
1918 }
1919 return 0;
1920 }
1921
1922 #endif
1923
1924
1925 ParallelObjectSynchronizerIterator ObjectSynchronizer::parallel_iterator() {
1926 return ParallelObjectSynchronizerIterator(gBlockList);
1927 }
1928
1929 // ParallelObjectSynchronizerIterator implementation
1930 ParallelObjectSynchronizerIterator::ParallelObjectSynchronizerIterator(ObjectMonitor * head)
1931 : _head(head), _cur(head) {
1932 assert(SafepointSynchronize::is_at_safepoint(), "Must at safepoint");
1933 }
1934
1935 ParallelObjectSynchronizerIterator::~ParallelObjectSynchronizerIterator() {
1936 assert(SafepointSynchronize::is_at_safepoint(), "Must at safepoint");
1937 PaddedEnd<ObjectMonitor>* block = (PaddedEnd<ObjectMonitor>*)_head;
1938 for (; block != NULL; block = (PaddedEnd<ObjectMonitor> *)next(block)) {
1939 assert(block->object() == CLAIMEDMARKER, "Must be a claimed block");
1940 // Restore chainmarker
1941 block->set_object(CHAINMARKER);
1942 }
1943 }
1944
1945 void* ParallelObjectSynchronizerIterator::claim() {
1946 PaddedEnd<ObjectMonitor>* my_cur = (PaddedEnd<ObjectMonitor>*)_cur;
1947 PaddedEnd<ObjectMonitor>* next_block;
1948
1949 while (true) {
1950 if (my_cur == NULL) return NULL;
1951
1952 if (my_cur->object() == CHAINMARKER) {
1953 if (my_cur->cas_set_object(CLAIMEDMARKER, CHAINMARKER) == CHAINMARKER) {
1954 return (void*)my_cur;
1955 }
1956 } else {
1957 assert(my_cur->object() == CLAIMEDMARKER, "Must be");
1958 }
1959
1960 next_block = (PaddedEnd<ObjectMonitor> *)next(my_cur);
1961 my_cur = (PaddedEnd<ObjectMonitor> *)Atomic::cmpxchg_ptr(next_block, &_cur, my_cur);
1962 }
1963 }
1964
1965 bool ParallelObjectSynchronizerIterator::parallel_oops_do(OopClosure* f) {
1966 PaddedEnd<ObjectMonitor>* block = (PaddedEnd<ObjectMonitor>*)claim();
1967 if (block != NULL) {
1968 assert(block->object() == CLAIMEDMARKER, "Must be a claimed block");
1969 for (int i = 1; i < ObjectSynchronizer::_BLOCKSIZE; i++) {
1970 ObjectMonitor* mid = (ObjectMonitor *)&block[i];
1971 if (mid->object() != NULL) {
1972 f->do_oop((oop*)mid->object_addr());
1973 }
1974 }
1975 return true;
1976 }
1977 return false;
1978 }