1 /*
2 * Copyright (c) 1998, 2018, 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/allocation.inline.hpp"
29 #include "memory/metaspaceShared.hpp"
30 #include "memory/padded.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/markOop.hpp"
33 #include "oops/oop.inline.hpp"
34 #include "runtime/atomic.hpp"
35 #include "runtime/biasedLocking.hpp"
36 #include "runtime/handles.inline.hpp"
37 #include "runtime/interfaceSupport.inline.hpp"
38 #include "runtime/mutexLocker.hpp"
39 #include "runtime/objectMonitor.hpp"
40 #include "runtime/objectMonitor.inline.hpp"
41 #include "runtime/osThread.hpp"
42 #include "runtime/safepointVerifiers.hpp"
43 #include "runtime/sharedRuntime.hpp"
44 #include "runtime/stubRoutines.hpp"
45 #include "runtime/synchronizer.hpp"
46 #include "runtime/thread.inline.hpp"
47 #include "runtime/vframe.hpp"
48 #include "runtime/vmThread.hpp"
49 #include "trace/traceMacros.hpp"
50 #include "trace/tracing.hpp"
51 #include "utilities/align.hpp"
52 #include "utilities/dtrace.hpp"
53 #include "utilities/events.hpp"
54 #include "utilities/preserveException.hpp"
55
56 // The "core" versions of monitor enter and exit reside in this file.
57 // The interpreter and compilers contain specialized transliterated
58 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(),
59 // for instance. If you make changes here, make sure to modify the
60 // interpreter, and both C1 and C2 fast-path inline locking code emission.
61 //
62 // -----------------------------------------------------------------------------
63
64 #ifdef DTRACE_ENABLED
65
66 // Only bother with this argument setup if dtrace is available
67 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
68
69 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
70 char* bytes = NULL; \
71 int len = 0; \
72 jlong jtid = SharedRuntime::get_java_tid(thread); \
73 Symbol* klassname = ((oop)(obj))->klass()->name(); \
74 if (klassname != NULL) { \
75 bytes = (char*)klassname->bytes(); \
76 len = klassname->utf8_length(); \
77 }
78
79 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
80 { \
81 if (DTraceMonitorProbes) { \
82 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
83 HOTSPOT_MONITOR_WAIT(jtid, \
84 (uintptr_t)(monitor), bytes, len, (millis)); \
85 } \
86 }
87
88 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY
89 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL
90 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
91
92 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
93 { \
94 if (DTraceMonitorProbes) { \
95 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
96 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
97 (uintptr_t)(monitor), bytes, len); \
98 } \
99 }
100
101 #else // ndef DTRACE_ENABLED
102
103 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
104 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
105
106 #endif // ndef DTRACE_ENABLED
107
108 // This exists only as a workaround of dtrace bug 6254741
109 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
110 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
111 return 0;
112 }
113
114 #define NINFLATIONLOCKS 256
115 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS];
116
117 // global list of blocks of monitors
118 PaddedEnd<ObjectMonitor> * volatile ObjectSynchronizer::gBlockList = NULL;
119 // global monitor free list
120 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL;
121 // global monitor in-use list, for moribund threads,
122 // monitors they inflated need to be scanned for deflation
123 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL;
124 // count of entries in gOmInUseList
125 int ObjectSynchronizer::gOmInUseCount = 0;
126
127 static volatile intptr_t gListLock = 0; // protects global monitor lists
128 static volatile int gMonitorFreeCount = 0; // # on gFreeList
129 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation
130
131 static void post_monitor_inflate_event(EventJavaMonitorInflate&,
132 const oop,
133 const ObjectSynchronizer::InflateCause);
134
135 #define CHAINMARKER (cast_to_oop<intptr_t>(-1))
136
137
138 // =====================> Quick functions
139
140 // The quick_* forms are special fast-path variants used to improve
141 // performance. In the simplest case, a "quick_*" implementation could
142 // simply return false, in which case the caller will perform the necessary
143 // state transitions and call the slow-path form.
144 // The fast-path is designed to handle frequently arising cases in an efficient
145 // manner and is just a degenerate "optimistic" variant of the slow-path.
146 // returns true -- to indicate the call was satisfied.
147 // returns false -- to indicate the call needs the services of the slow-path.
148 // A no-loitering ordinance is in effect for code in the quick_* family
149 // operators: safepoints or indefinite blocking (blocking that might span a
150 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon
151 // entry.
152 //
153 // Consider: An interesting optimization is to have the JIT recognize the
154 // following common idiom:
155 // synchronized (someobj) { .... ; notify(); }
156 // That is, we find a notify() or notifyAll() call that immediately precedes
157 // the monitorexit operation. In that case the JIT could fuse the operations
158 // into a single notifyAndExit() runtime primitive.
159
160 bool ObjectSynchronizer::quick_notify(oopDesc * obj, Thread * self, bool all) {
161 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
162 assert(self->is_Java_thread(), "invariant");
163 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant");
164 NoSafepointVerifier nsv;
165 if (obj == NULL) return false; // slow-path for invalid obj
166 const markOop mark = obj->mark();
167
168 if (mark->has_locker() && self->is_lock_owned((address)mark->locker())) {
169 // Degenerate notify
170 // stack-locked by caller so by definition the implied waitset is empty.
171 return true;
172 }
173
174 if (mark->has_monitor()) {
175 ObjectMonitor * const mon = mark->monitor();
176 assert(mon->object() == obj, "invariant");
177 if (mon->owner() != self) return false; // slow-path for IMS exception
178
179 if (mon->first_waiter() != NULL) {
180 // We have one or more waiters. Since this is an inflated monitor
181 // that we own, we can transfer one or more threads from the waitset
182 // to the entrylist here and now, avoiding the slow-path.
183 if (all) {
184 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self);
185 } else {
186 DTRACE_MONITOR_PROBE(notify, mon, obj, self);
187 }
188 int tally = 0;
189 do {
190 mon->INotify(self);
191 ++tally;
192 } while (mon->first_waiter() != NULL && all);
193 OM_PERFDATA_OP(Notifications, inc(tally));
194 }
195 return true;
196 }
197
198 // biased locking and any other IMS exception states take the slow-path
199 return false;
200 }
201
202
203 // The LockNode emitted directly at the synchronization site would have
204 // been too big if it were to have included support for the cases of inflated
205 // recursive enter and exit, so they go here instead.
206 // Note that we can't safely call AsyncPrintJavaStack() from within
207 // quick_enter() as our thread state remains _in_Java.
208
209 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self,
210 BasicLock * lock) {
211 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
212 assert(Self->is_Java_thread(), "invariant");
213 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant");
214 NoSafepointVerifier nsv;
215 if (obj == NULL) return false; // Need to throw NPE
216 const markOop mark = obj->mark();
217
218 if (mark->has_monitor()) {
219 ObjectMonitor * const m = mark->monitor();
220 assert(m->object() == obj, "invariant");
221 Thread * const owner = (Thread *) m->_owner;
222
223 // Lock contention and Transactional Lock Elision (TLE) diagnostics
224 // and observability
225 // Case: light contention possibly amenable to TLE
226 // Case: TLE inimical operations such as nested/recursive synchronization
227
228 if (owner == Self) {
229 m->_recursions++;
230 return true;
231 }
232
233 // This Java Monitor is inflated so obj's header will never be
234 // displaced to this thread's BasicLock. Make the displaced header
235 // non-NULL so this BasicLock is not seen as recursive nor as
236 // being locked. We do this unconditionally so that this thread's
237 // BasicLock cannot be mis-interpreted by any stack walkers. For
238 // performance reasons, stack walkers generally first check for
239 // Biased Locking in the object's header, the second check is for
240 // stack-locking in the object's header, the third check is for
241 // recursive stack-locking in the displaced header in the BasicLock,
242 // and last are the inflated Java Monitor (ObjectMonitor) checks.
243 lock->set_displaced_header(markOopDesc::unused_mark());
244
245 if (owner == NULL && Atomic::replace_if_null(Self, &(m->_owner))) {
246 assert(m->_recursions == 0, "invariant");
247 assert(m->_owner == Self, "invariant");
248 return true;
249 }
250 }
251
252 // Note that we could inflate in quick_enter.
253 // This is likely a useful optimization
254 // Critically, in quick_enter() we must not:
255 // -- perform bias revocation, or
256 // -- block indefinitely, or
257 // -- reach a safepoint
258
259 return false; // revert to slow-path
260 }
261
262 // -----------------------------------------------------------------------------
263 // Fast Monitor Enter/Exit
264 // This the fast monitor enter. The interpreter and compiler use
265 // some assembly copies of this code. Make sure update those code
266 // if the following function is changed. The implementation is
267 // extremely sensitive to race condition. Be careful.
268
269 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock,
270 bool attempt_rebias, TRAPS) {
271 if (UseBiasedLocking) {
272 if (!SafepointSynchronize::is_at_safepoint()) {
273 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD);
274 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) {
275 return;
276 }
277 } else {
278 assert(!attempt_rebias, "can not rebias toward VM thread");
279 BiasedLocking::revoke_at_safepoint(obj);
280 }
281 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
282 }
283
284 slow_enter(obj, lock, THREAD);
285 }
286
287 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) {
288 markOop mark = object->mark();
289 // We cannot check for Biased Locking if we are racing an inflation.
290 assert(mark == markOopDesc::INFLATING() ||
291 !mark->has_bias_pattern(), "should not see bias pattern here");
292
293 markOop dhw = lock->displaced_header();
294 if (dhw == NULL) {
295 // If the displaced header is NULL, then this exit matches up with
296 // a recursive enter. No real work to do here except for diagnostics.
297 #ifndef PRODUCT
298 if (mark != markOopDesc::INFLATING()) {
299 // Only do diagnostics if we are not racing an inflation. Simply
300 // exiting a recursive enter of a Java Monitor that is being
301 // inflated is safe; see the has_monitor() comment below.
302 assert(!mark->is_neutral(), "invariant");
303 assert(!mark->has_locker() ||
304 THREAD->is_lock_owned((address)mark->locker()), "invariant");
305 if (mark->has_monitor()) {
306 // The BasicLock's displaced_header is marked as a recursive
307 // enter and we have an inflated Java Monitor (ObjectMonitor).
308 // This is a special case where the Java Monitor was inflated
309 // after this thread entered the stack-lock recursively. When a
310 // Java Monitor is inflated, we cannot safely walk the Java
311 // Monitor owner's stack and update the BasicLocks because a
312 // Java Monitor can be asynchronously inflated by a thread that
313 // does not own the Java Monitor.
314 ObjectMonitor * m = mark->monitor();
315 assert(((oop)(m->object()))->mark() == mark, "invariant");
316 assert(m->is_entered(THREAD), "invariant");
317 }
318 }
319 #endif
320 return;
321 }
322
323 if (mark == (markOop) lock) {
324 // If the object is stack-locked by the current thread, try to
325 // swing the displaced header from the BasicLock back to the mark.
326 assert(dhw->is_neutral(), "invariant");
327 if (object->cas_set_mark(dhw, mark) == mark) {
328 TEVENT(fast_exit: release stack-lock);
329 return;
330 }
331 }
332
333 // We have to take the slow-path of possible inflation and then exit.
334 ObjectSynchronizer::inflate(THREAD,
335 object,
336 inflate_cause_vm_internal)->exit(true, THREAD);
337 }
338
339 // -----------------------------------------------------------------------------
340 // Interpreter/Compiler Slow Case
341 // This routine is used to handle interpreter/compiler slow case
342 // We don't need to use fast path here, because it must have been
343 // failed in the interpreter/compiler code.
344 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
345 markOop mark = obj->mark();
346 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
347
348 if (mark->is_neutral()) {
349 // Anticipate successful CAS -- the ST of the displaced mark must
350 // be visible <= the ST performed by the CAS.
351 lock->set_displaced_header(mark);
352 if (mark == obj()->cas_set_mark((markOop) lock, mark)) {
353 TEVENT(slow_enter: release stacklock);
354 return;
355 }
356 // Fall through to inflate() ...
357 } else if (mark->has_locker() &&
358 THREAD->is_lock_owned((address)mark->locker())) {
359 assert(lock != mark->locker(), "must not re-lock the same lock");
360 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
361 lock->set_displaced_header(NULL);
362 return;
363 }
364
365 // The object header will never be displaced to this lock,
366 // so it does not matter what the value is, except that it
367 // must be non-zero to avoid looking like a re-entrant lock,
368 // and must not look locked either.
369 lock->set_displaced_header(markOopDesc::unused_mark());
370 ObjectSynchronizer::inflate(THREAD,
371 obj(),
372 inflate_cause_monitor_enter)->enter(THREAD);
373 }
374
375 // This routine is used to handle interpreter/compiler slow case
376 // We don't need to use fast path here, because it must have
377 // failed in the interpreter/compiler code. Simply use the heavy
378 // weight monitor should be ok, unless someone find otherwise.
379 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
380 fast_exit(object, lock, THREAD);
381 }
382
383 // -----------------------------------------------------------------------------
384 // Class Loader support to workaround deadlocks on the class loader lock objects
385 // Also used by GC
386 // complete_exit()/reenter() are used to wait on a nested lock
387 // i.e. to give up an outer lock completely and then re-enter
388 // Used when holding nested locks - lock acquisition order: lock1 then lock2
389 // 1) complete_exit lock1 - saving recursion count
390 // 2) wait on lock2
391 // 3) when notified on lock2, unlock lock2
392 // 4) reenter lock1 with original recursion count
393 // 5) lock lock2
394 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
395 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
396 TEVENT(complete_exit);
397 if (UseBiasedLocking) {
398 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
399 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
400 }
401
402 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
403 obj(),
404 inflate_cause_vm_internal);
405
406 return monitor->complete_exit(THREAD);
407 }
408
409 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
410 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
411 TEVENT(reenter);
412 if (UseBiasedLocking) {
413 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
414 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
415 }
416
417 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
418 obj(),
419 inflate_cause_vm_internal);
420
421 monitor->reenter(recursion, THREAD);
422 }
423 // -----------------------------------------------------------------------------
424 // JNI locks on java objects
425 // NOTE: must use heavy weight monitor to handle jni monitor enter
426 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) {
427 // the current locking is from JNI instead of Java code
428 TEVENT(jni_enter);
429 if (UseBiasedLocking) {
430 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
431 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
432 }
433 THREAD->set_current_pending_monitor_is_from_java(false);
434 ObjectSynchronizer::inflate(THREAD, obj(), inflate_cause_jni_enter)->enter(THREAD);
435 THREAD->set_current_pending_monitor_is_from_java(true);
436 }
437
438 // NOTE: must use heavy weight monitor to handle jni monitor exit
439 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
440 TEVENT(jni_exit);
441 if (UseBiasedLocking) {
442 Handle h_obj(THREAD, obj);
443 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD);
444 obj = h_obj();
445 }
446 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
447
448 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
449 obj,
450 inflate_cause_jni_exit);
451 // If this thread has locked the object, exit the monitor. Note: can't use
452 // monitor->check(CHECK); must exit even if an exception is pending.
453 if (monitor->check(THREAD)) {
454 monitor->exit(true, THREAD);
455 }
456 }
457
458 // -----------------------------------------------------------------------------
459 // Internal VM locks on java objects
460 // standard constructor, allows locking failures
461 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
462 _dolock = doLock;
463 _thread = thread;
464 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
465 _obj = obj;
466
467 if (_dolock) {
468 TEVENT(ObjectLocker);
469
470 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
471 }
472 }
473
474 ObjectLocker::~ObjectLocker() {
475 if (_dolock) {
476 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
477 }
478 }
479
480
481 // -----------------------------------------------------------------------------
482 // Wait/Notify/NotifyAll
483 // NOTE: must use heavy weight monitor to handle wait()
484 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
485 if (UseBiasedLocking) {
486 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
487 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
488 }
489 if (millis < 0) {
490 TEVENT(wait - throw IAX);
491 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
492 }
493 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
494 obj(),
495 inflate_cause_wait);
496
497 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
498 monitor->wait(millis, true, THREAD);
499
500 // This dummy call is in place to get around dtrace bug 6254741. Once
501 // that's fixed we can uncomment the following line, remove the call
502 // and change this function back into a "void" func.
503 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
504 return dtrace_waited_probe(monitor, obj, THREAD);
505 }
506
507 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
508 if (UseBiasedLocking) {
509 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
510 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
511 }
512 if (millis < 0) {
513 TEVENT(wait - throw IAX);
514 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
515 }
516 ObjectSynchronizer::inflate(THREAD,
517 obj(),
518 inflate_cause_wait)->wait(millis, false, THREAD);
519 }
520
521 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
522 if (UseBiasedLocking) {
523 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
524 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
525 }
526
527 markOop mark = obj->mark();
528 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
529 return;
530 }
531 ObjectSynchronizer::inflate(THREAD,
532 obj(),
533 inflate_cause_notify)->notify(THREAD);
534 }
535
536 // NOTE: see comment of notify()
537 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
538 if (UseBiasedLocking) {
539 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
540 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
541 }
542
543 markOop mark = obj->mark();
544 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
545 return;
546 }
547 ObjectSynchronizer::inflate(THREAD,
548 obj(),
549 inflate_cause_notify)->notifyAll(THREAD);
550 }
551
552 // -----------------------------------------------------------------------------
553 // Hash Code handling
554 //
555 // Performance concern:
556 // OrderAccess::storestore() calls release() which at one time stored 0
557 // into the global volatile OrderAccess::dummy variable. This store was
558 // unnecessary for correctness. Many threads storing into a common location
559 // causes considerable cache migration or "sloshing" on large SMP systems.
560 // As such, I avoided using OrderAccess::storestore(). In some cases
561 // OrderAccess::fence() -- which incurs local latency on the executing
562 // processor -- is a better choice as it scales on SMP systems.
563 //
564 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for
565 // a discussion of coherency costs. Note that all our current reference
566 // platforms provide strong ST-ST order, so the issue is moot on IA32,
567 // x64, and SPARC.
568 //
569 // As a general policy we use "volatile" to control compiler-based reordering
570 // and explicit fences (barriers) to control for architectural reordering
571 // performed by the CPU(s) or platform.
572
573 struct SharedGlobals {
574 char _pad_prefix[DEFAULT_CACHE_LINE_SIZE];
575 // These are highly shared mostly-read variables.
576 // To avoid false-sharing they need to be the sole occupants of a cache line.
577 volatile int stwRandom;
578 volatile int stwCycle;
579 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int) * 2);
580 // Hot RW variable -- Sequester to avoid false-sharing
581 volatile int hcSequence;
582 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int));
583 };
584
585 static SharedGlobals GVars;
586 static int MonitorScavengeThreshold = 1000000;
587 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending
588
589 static markOop ReadStableMark(oop obj) {
590 markOop mark = obj->mark();
591 if (!mark->is_being_inflated()) {
592 return mark; // normal fast-path return
593 }
594
595 int its = 0;
596 for (;;) {
597 markOop mark = obj->mark();
598 if (!mark->is_being_inflated()) {
599 return mark; // normal fast-path return
600 }
601
602 // The object is being inflated by some other thread.
603 // The caller of ReadStableMark() must wait for inflation to complete.
604 // Avoid live-lock
605 // TODO: consider calling SafepointSynchronize::do_call_back() while
606 // spinning to see if there's a safepoint pending. If so, immediately
607 // yielding or blocking would be appropriate. Avoid spinning while
608 // there is a safepoint pending.
609 // TODO: add inflation contention performance counters.
610 // TODO: restrict the aggregate number of spinners.
611
612 ++its;
613 if (its > 10000 || !os::is_MP()) {
614 if (its & 1) {
615 os::naked_yield();
616 TEVENT(Inflate: INFLATING - yield);
617 } else {
618 // Note that the following code attenuates the livelock problem but is not
619 // a complete remedy. A more complete solution would require that the inflating
620 // thread hold the associated inflation lock. The following code simply restricts
621 // the number of spinners to at most one. We'll have N-2 threads blocked
622 // on the inflationlock, 1 thread holding the inflation lock and using
623 // a yield/park strategy, and 1 thread in the midst of inflation.
624 // A more refined approach would be to change the encoding of INFLATING
625 // to allow encapsulation of a native thread pointer. Threads waiting for
626 // inflation to complete would use CAS to push themselves onto a singly linked
627 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
628 // and calling park(). When inflation was complete the thread that accomplished inflation
629 // would detach the list and set the markword to inflated with a single CAS and
630 // then for each thread on the list, set the flag and unpark() the thread.
631 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease
632 // wakes at most one thread whereas we need to wake the entire list.
633 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1);
634 int YieldThenBlock = 0;
635 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant");
636 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant");
637 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock");
638 while (obj->mark() == markOopDesc::INFLATING()) {
639 // Beware: NakedYield() is advisory and has almost no effect on some platforms
640 // so we periodically call Self->_ParkEvent->park(1).
641 // We use a mixed spin/yield/block mechanism.
642 if ((YieldThenBlock++) >= 16) {
643 Thread::current()->_ParkEvent->park(1);
644 } else {
645 os::naked_yield();
646 }
647 }
648 Thread::muxRelease(gInflationLocks + ix);
649 TEVENT(Inflate: INFLATING - yield/park);
650 }
651 } else {
652 SpinPause(); // SMP-polite spinning
653 }
654 }
655 }
656
657 // hashCode() generation :
658 //
659 // Possibilities:
660 // * MD5Digest of {obj,stwRandom}
661 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function.
662 // * A DES- or AES-style SBox[] mechanism
663 // * One of the Phi-based schemes, such as:
664 // 2654435761 = 2^32 * Phi (golden ratio)
665 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ;
666 // * A variation of Marsaglia's shift-xor RNG scheme.
667 // * (obj ^ stwRandom) is appealing, but can result
668 // in undesirable regularity in the hashCode values of adjacent objects
669 // (objects allocated back-to-back, in particular). This could potentially
670 // result in hashtable collisions and reduced hashtable efficiency.
671 // There are simple ways to "diffuse" the middle address bits over the
672 // generated hashCode values:
673
674 static inline intptr_t get_next_hash(Thread * Self, oop obj) {
675 intptr_t value = 0;
676 if (hashCode == 0) {
677 // This form uses global Park-Miller 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 = Atomic::cmpxchg(temp, monitor->header_addr(), 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(ThreadsList * t_list, Handle h_obj) {
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(t_list, owner);
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 = OrderAccess::load_acquire(&gBlockList);
944 while (block != NULL) {
945 assert(block->object() == CHAINMARKER, "must be a block header");
946 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
947 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
948 oop object = (oop)mid->object();
949 if (object != NULL) {
950 closure->do_monitor(mid);
951 }
952 }
953 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
954 }
955 }
956
957 // Get the next block in the block list.
958 static inline PaddedEnd<ObjectMonitor>* next(PaddedEnd<ObjectMonitor>* block) {
959 assert(block->object() == CHAINMARKER, "must be a block header");
960 block = (PaddedEnd<ObjectMonitor>*) block->FreeNext;
961 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
962 return block;
963 }
964
965 static bool monitors_used_above_threshold() {
966 if (gMonitorPopulation == 0) {
967 return false;
968 }
969 int monitors_used = gMonitorPopulation - gMonitorFreeCount;
970 int monitor_usage = (monitors_used * 100LL) / gMonitorPopulation;
971 return monitor_usage > MonitorUsedDeflationThreshold;
972 }
973
974 bool ObjectSynchronizer::is_cleanup_needed() {
975 if (MonitorUsedDeflationThreshold > 0) {
976 return monitors_used_above_threshold();
977 }
978 return false;
979 }
980
981 void ObjectSynchronizer::oops_do(OopClosure* f) {
982 if (MonitorInUseLists) {
983 // When using thread local monitor lists, we only scan the
984 // global used list here (for moribund threads), and
985 // the thread-local monitors in Thread::oops_do().
986 global_used_oops_do(f);
987 } else {
988 global_oops_do(f);
989 }
990 }
991
992 void ObjectSynchronizer::global_oops_do(OopClosure* f) {
993 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
994 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
995 for (; block != NULL; block = next(block)) {
996 assert(block->object() == CHAINMARKER, "must be a block header");
997 for (int i = 1; i < _BLOCKSIZE; i++) {
998 ObjectMonitor* mid = (ObjectMonitor *)&block[i];
999 if (mid->object() != NULL) {
1000 f->do_oop((oop*)mid->object_addr());
1001 }
1002 }
1003 }
1004 }
1005
1006 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) {
1007 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1008 list_oops_do(gOmInUseList, f);
1009 }
1010
1011 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) {
1012 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1013 list_oops_do(thread->omInUseList, f);
1014 }
1015
1016 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) {
1017 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1018 ObjectMonitor* mid;
1019 for (mid = list; mid != NULL; mid = mid->FreeNext) {
1020 if (mid->object() != NULL) {
1021 f->do_oop((oop*)mid->object_addr());
1022 }
1023 }
1024 }
1025
1026
1027 // -----------------------------------------------------------------------------
1028 // ObjectMonitor Lifecycle
1029 // -----------------------
1030 // Inflation unlinks monitors from the global gFreeList and
1031 // associates them with objects. Deflation -- which occurs at
1032 // STW-time -- disassociates idle monitors from objects. Such
1033 // scavenged monitors are returned to the gFreeList.
1034 //
1035 // The global list is protected by gListLock. All the critical sections
1036 // are short and operate in constant-time.
1037 //
1038 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
1039 //
1040 // Lifecycle:
1041 // -- unassigned and on the global free list
1042 // -- unassigned and on a thread's private omFreeList
1043 // -- assigned to an object. The object is inflated and the mark refers
1044 // to the objectmonitor.
1045
1046
1047 // Constraining monitor pool growth via MonitorBound ...
1048 //
1049 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
1050 // the rate of scavenging is driven primarily by GC. As such, we can find
1051 // an inordinate number of monitors in circulation.
1052 // To avoid that scenario we can artificially induce a STW safepoint
1053 // if the pool appears to be growing past some reasonable bound.
1054 // Generally we favor time in space-time tradeoffs, but as there's no
1055 // natural back-pressure on the # of extant monitors we need to impose some
1056 // type of limit. Beware that if MonitorBound is set to too low a value
1057 // we could just loop. In addition, if MonitorBound is set to a low value
1058 // we'll incur more safepoints, which are harmful to performance.
1059 // See also: GuaranteedSafepointInterval
1060 //
1061 // The current implementation uses asynchronous VM operations.
1062
1063 static void InduceScavenge(Thread * Self, const char * Whence) {
1064 // Induce STW safepoint to trim monitors
1065 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
1066 // More precisely, trigger an asynchronous STW safepoint as the number
1067 // of active monitors passes the specified threshold.
1068 // TODO: assert thread state is reasonable
1069
1070 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
1071 if (ObjectMonitor::Knob_Verbose) {
1072 tty->print_cr("INFO: Monitor scavenge - Induced STW @%s (%d)",
1073 Whence, ForceMonitorScavenge) ;
1074 tty->flush();
1075 }
1076 // Induce a 'null' safepoint to scavenge monitors
1077 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
1078 // to the VMthread and have a lifespan longer than that of this activation record.
1079 // The VMThread will delete the op when completed.
1080 VMThread::execute(new VM_ScavengeMonitors());
1081
1082 if (ObjectMonitor::Knob_Verbose) {
1083 tty->print_cr("INFO: Monitor scavenge - STW posted @%s (%d)",
1084 Whence, ForceMonitorScavenge) ;
1085 tty->flush();
1086 }
1087 }
1088 }
1089
1090 void ObjectSynchronizer::verifyInUse(Thread *Self) {
1091 ObjectMonitor* mid;
1092 int in_use_tally = 0;
1093 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) {
1094 in_use_tally++;
1095 }
1096 assert(in_use_tally == Self->omInUseCount, "in-use count off");
1097
1098 int free_tally = 0;
1099 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) {
1100 free_tally++;
1101 }
1102 assert(free_tally == Self->omFreeCount, "free count off");
1103 }
1104
1105 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) {
1106 // A large MAXPRIVATE value reduces both list lock contention
1107 // and list coherency traffic, but also tends to increase the
1108 // number of objectMonitors in circulation as well as the STW
1109 // scavenge costs. As usual, we lean toward time in space-time
1110 // tradeoffs.
1111 const int MAXPRIVATE = 1024;
1112 for (;;) {
1113 ObjectMonitor * m;
1114
1115 // 1: try to allocate from the thread's local omFreeList.
1116 // Threads will attempt to allocate first from their local list, then
1117 // from the global list, and only after those attempts fail will the thread
1118 // attempt to instantiate new monitors. Thread-local free lists take
1119 // heat off the gListLock and improve allocation latency, as well as reducing
1120 // coherency traffic on the shared global list.
1121 m = Self->omFreeList;
1122 if (m != NULL) {
1123 Self->omFreeList = m->FreeNext;
1124 Self->omFreeCount--;
1125 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene
1126 guarantee(m->object() == NULL, "invariant");
1127 if (MonitorInUseLists) {
1128 m->FreeNext = Self->omInUseList;
1129 Self->omInUseList = m;
1130 Self->omInUseCount++;
1131 if (ObjectMonitor::Knob_VerifyInUse) {
1132 verifyInUse(Self);
1133 }
1134 } else {
1135 m->FreeNext = NULL;
1136 }
1137 return m;
1138 }
1139
1140 // 2: try to allocate from the global gFreeList
1141 // CONSIDER: use muxTry() instead of muxAcquire().
1142 // If the muxTry() fails then drop immediately into case 3.
1143 // If we're using thread-local free lists then try
1144 // to reprovision the caller's free list.
1145 if (gFreeList != NULL) {
1146 // Reprovision the thread's omFreeList.
1147 // Use bulk transfers to reduce the allocation rate and heat
1148 // on various locks.
1149 Thread::muxAcquire(&gListLock, "omAlloc");
1150 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) {
1151 gMonitorFreeCount--;
1152 ObjectMonitor * take = gFreeList;
1153 gFreeList = take->FreeNext;
1154 guarantee(take->object() == NULL, "invariant");
1155 guarantee(!take->is_busy(), "invariant");
1156 take->Recycle();
1157 omRelease(Self, take, false);
1158 }
1159 Thread::muxRelease(&gListLock);
1160 Self->omFreeProvision += 1 + (Self->omFreeProvision/2);
1161 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE;
1162 TEVENT(omFirst - reprovision);
1163
1164 const int mx = MonitorBound;
1165 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) {
1166 // We can't safely induce a STW safepoint from omAlloc() as our thread
1167 // state may not be appropriate for such activities and callers may hold
1168 // naked oops, so instead we defer the action.
1169 InduceScavenge(Self, "omAlloc");
1170 }
1171 continue;
1172 }
1173
1174 // 3: allocate a block of new ObjectMonitors
1175 // Both the local and global free lists are empty -- resort to malloc().
1176 // In the current implementation objectMonitors are TSM - immortal.
1177 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want
1178 // each ObjectMonitor to start at the beginning of a cache line,
1179 // so we use align_up().
1180 // A better solution would be to use C++ placement-new.
1181 // BEWARE: As it stands currently, we don't run the ctors!
1182 assert(_BLOCKSIZE > 1, "invariant");
1183 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE;
1184 PaddedEnd<ObjectMonitor> * temp;
1185 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1);
1186 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size,
1187 mtInternal);
1188 temp = (PaddedEnd<ObjectMonitor> *)
1189 align_up(real_malloc_addr, DEFAULT_CACHE_LINE_SIZE);
1190
1191 // NOTE: (almost) no way to recover if allocation failed.
1192 // We might be able to induce a STW safepoint and scavenge enough
1193 // objectMonitors to permit progress.
1194 if (temp == NULL) {
1195 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR,
1196 "Allocate ObjectMonitors");
1197 }
1198 (void)memset((void *) temp, 0, neededsize);
1199
1200 // Format the block.
1201 // initialize the linked list, each monitor points to its next
1202 // forming the single linked free list, the very first monitor
1203 // will points to next block, which forms the block list.
1204 // The trick of using the 1st element in the block as gBlockList
1205 // linkage should be reconsidered. A better implementation would
1206 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1207
1208 for (int i = 1; i < _BLOCKSIZE; i++) {
1209 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1];
1210 }
1211
1212 // terminate the last monitor as the end of list
1213 temp[_BLOCKSIZE - 1].FreeNext = NULL;
1214
1215 // Element [0] is reserved for global list linkage
1216 temp[0].set_object(CHAINMARKER);
1217
1218 // Consider carving out this thread's current request from the
1219 // block in hand. This avoids some lock traffic and redundant
1220 // list activity.
1221
1222 // Acquire the gListLock to manipulate gBlockList and gFreeList.
1223 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1224 Thread::muxAcquire(&gListLock, "omAlloc [2]");
1225 gMonitorPopulation += _BLOCKSIZE-1;
1226 gMonitorFreeCount += _BLOCKSIZE-1;
1227
1228 // Add the new block to the list of extant blocks (gBlockList).
1229 // The very first objectMonitor in a block is reserved and dedicated.
1230 // It serves as blocklist "next" linkage.
1231 temp[0].FreeNext = gBlockList;
1232 // There are lock-free uses of gBlockList so make sure that
1233 // the previous stores happen before we update gBlockList.
1234 OrderAccess::release_store(&gBlockList, temp);
1235
1236 // Add the new string of objectMonitors to the global free list
1237 temp[_BLOCKSIZE - 1].FreeNext = gFreeList;
1238 gFreeList = temp + 1;
1239 Thread::muxRelease(&gListLock);
1240 TEVENT(Allocate block of monitors);
1241 }
1242 }
1243
1244 // Place "m" on the caller's private per-thread omFreeList.
1245 // In practice there's no need to clamp or limit the number of
1246 // monitors on a thread's omFreeList as the only time we'll call
1247 // omRelease is to return a monitor to the free list after a CAS
1248 // attempt failed. This doesn't allow unbounded #s of monitors to
1249 // accumulate on a thread's free list.
1250 //
1251 // Key constraint: all ObjectMonitors on a thread's free list and the global
1252 // free list must have their object field set to null. This prevents the
1253 // scavenger -- deflate_idle_monitors -- from reclaiming them.
1254
1255 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m,
1256 bool fromPerThreadAlloc) {
1257 guarantee(m->object() == NULL, "invariant");
1258 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor");
1259 // Remove from omInUseList
1260 if (MonitorInUseLists && fromPerThreadAlloc) {
1261 ObjectMonitor* cur_mid_in_use = NULL;
1262 bool extracted = false;
1263 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) {
1264 if (m == mid) {
1265 // extract from per-thread in-use list
1266 if (mid == Self->omInUseList) {
1267 Self->omInUseList = mid->FreeNext;
1268 } else if (cur_mid_in_use != NULL) {
1269 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1270 }
1271 extracted = true;
1272 Self->omInUseCount--;
1273 if (ObjectMonitor::Knob_VerifyInUse) {
1274 verifyInUse(Self);
1275 }
1276 break;
1277 }
1278 }
1279 assert(extracted, "Should have extracted from in-use list");
1280 }
1281
1282 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new
1283 m->FreeNext = Self->omFreeList;
1284 Self->omFreeList = m;
1285 Self->omFreeCount++;
1286 }
1287
1288 // Return the monitors of a moribund thread's local free list to
1289 // the global free list. Typically a thread calls omFlush() when
1290 // it's dying. We could also consider having the VM thread steal
1291 // monitors from threads that have not run java code over a few
1292 // consecutive STW safepoints. Relatedly, we might decay
1293 // omFreeProvision at STW safepoints.
1294 //
1295 // Also return the monitors of a moribund thread's omInUseList to
1296 // a global gOmInUseList under the global list lock so these
1297 // will continue to be scanned.
1298 //
1299 // We currently call omFlush() from Threads::remove() _before the thread
1300 // has been excised from the thread list and is no longer a mutator.
1301 // This means that omFlush() can not run concurrently with a safepoint and
1302 // interleave with the scavenge operator. In particular, this ensures that
1303 // the thread's monitors are scanned by a GC safepoint, either via
1304 // Thread::oops_do() (if safepoint happens before omFlush()) or via
1305 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's
1306 // monitors have been transferred to the global in-use list).
1307
1308 void ObjectSynchronizer::omFlush(Thread * Self) {
1309 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL
1310 Self->omFreeList = NULL;
1311 ObjectMonitor * tail = NULL;
1312 int tally = 0;
1313 if (list != NULL) {
1314 ObjectMonitor * s;
1315 // The thread is going away, the per-thread free monitors
1316 // are freed via set_owner(NULL)
1317 // Link them to tail, which will be linked into the global free list
1318 // gFreeList below, under the gListLock
1319 for (s = list; s != NULL; s = s->FreeNext) {
1320 tally++;
1321 tail = s;
1322 guarantee(s->object() == NULL, "invariant");
1323 guarantee(!s->is_busy(), "invariant");
1324 s->set_owner(NULL); // redundant but good hygiene
1325 TEVENT(omFlush - Move one);
1326 }
1327 guarantee(tail != NULL && list != NULL, "invariant");
1328 }
1329
1330 ObjectMonitor * inUseList = Self->omInUseList;
1331 ObjectMonitor * inUseTail = NULL;
1332 int inUseTally = 0;
1333 if (inUseList != NULL) {
1334 Self->omInUseList = NULL;
1335 ObjectMonitor *cur_om;
1336 // The thread is going away, however the omInUseList inflated
1337 // monitors may still be in-use by other threads.
1338 // Link them to inUseTail, which will be linked into the global in-use list
1339 // gOmInUseList below, under the gListLock
1340 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) {
1341 inUseTail = cur_om;
1342 inUseTally++;
1343 }
1344 assert(Self->omInUseCount == inUseTally, "in-use count off");
1345 Self->omInUseCount = 0;
1346 guarantee(inUseTail != NULL && inUseList != NULL, "invariant");
1347 }
1348
1349 Thread::muxAcquire(&gListLock, "omFlush");
1350 if (tail != NULL) {
1351 tail->FreeNext = gFreeList;
1352 gFreeList = list;
1353 gMonitorFreeCount += tally;
1354 assert(Self->omFreeCount == tally, "free-count off");
1355 Self->omFreeCount = 0;
1356 }
1357
1358 if (inUseTail != NULL) {
1359 inUseTail->FreeNext = gOmInUseList;
1360 gOmInUseList = inUseList;
1361 gOmInUseCount += inUseTally;
1362 }
1363
1364 Thread::muxRelease(&gListLock);
1365 TEVENT(omFlush);
1366 }
1367
1368 // Fast path code shared by multiple functions
1369 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) {
1370 markOop mark = obj->mark();
1371 if (mark->has_monitor()) {
1372 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid");
1373 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header");
1374 return mark->monitor();
1375 }
1376 return ObjectSynchronizer::inflate(Thread::current(),
1377 obj,
1378 inflate_cause_vm_internal);
1379 }
1380
1381 ObjectMonitor* ObjectSynchronizer::inflate(Thread * Self,
1382 oop object,
1383 const InflateCause cause) {
1384
1385 // Inflate mutates the heap ...
1386 // Relaxing assertion for bug 6320749.
1387 assert(Universe::verify_in_progress() ||
1388 !SafepointSynchronize::is_at_safepoint(), "invariant");
1389
1390 EventJavaMonitorInflate event;
1391
1392 for (;;) {
1393 const markOop mark = object->mark();
1394 assert(!mark->has_bias_pattern(), "invariant");
1395
1396 // The mark can be in one of the following states:
1397 // * Inflated - just return
1398 // * Stack-locked - coerce it to inflated
1399 // * INFLATING - busy wait for conversion to complete
1400 // * Neutral - aggressively inflate the object.
1401 // * BIASED - Illegal. We should never see this
1402
1403 // CASE: inflated
1404 if (mark->has_monitor()) {
1405 ObjectMonitor * inf = mark->monitor();
1406 assert(inf->header()->is_neutral(), "invariant");
1407 assert(inf->object() == object, "invariant");
1408 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1409 return inf;
1410 }
1411
1412 // CASE: inflation in progress - inflating over a stack-lock.
1413 // Some other thread is converting from stack-locked to inflated.
1414 // Only that thread can complete inflation -- other threads must wait.
1415 // The INFLATING value is transient.
1416 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1417 // We could always eliminate polling by parking the thread on some auxiliary list.
1418 if (mark == markOopDesc::INFLATING()) {
1419 TEVENT(Inflate: spin while INFLATING);
1420 ReadStableMark(object);
1421 continue;
1422 }
1423
1424 // CASE: stack-locked
1425 // Could be stack-locked either by this thread or by some other thread.
1426 //
1427 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1428 // to install INFLATING into the mark word. We originally installed INFLATING,
1429 // allocated the objectmonitor, and then finally STed the address of the
1430 // objectmonitor into the mark. This was correct, but artificially lengthened
1431 // the interval in which INFLATED appeared in the mark, thus increasing
1432 // the odds of inflation contention.
1433 //
1434 // We now use per-thread private objectmonitor free lists.
1435 // These list are reprovisioned from the global free list outside the
1436 // critical INFLATING...ST interval. A thread can transfer
1437 // multiple objectmonitors en-mass from the global free list to its local free list.
1438 // This reduces coherency traffic and lock contention on the global free list.
1439 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1440 // before or after the CAS(INFLATING) operation.
1441 // See the comments in omAlloc().
1442
1443 if (mark->has_locker()) {
1444 ObjectMonitor * m = omAlloc(Self);
1445 // Optimistically prepare the objectmonitor - anticipate successful CAS
1446 // We do this before the CAS in order to minimize the length of time
1447 // in which INFLATING appears in the mark.
1448 m->Recycle();
1449 m->_Responsible = NULL;
1450 m->_recursions = 0;
1451 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1452
1453 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark);
1454 if (cmp != mark) {
1455 omRelease(Self, m, true);
1456 continue; // Interference -- just retry
1457 }
1458
1459 // We've successfully installed INFLATING (0) into the mark-word.
1460 // This is the only case where 0 will appear in a mark-word.
1461 // Only the singular thread that successfully swings the mark-word
1462 // to 0 can perform (or more precisely, complete) inflation.
1463 //
1464 // Why do we CAS a 0 into the mark-word instead of just CASing the
1465 // mark-word from the stack-locked value directly to the new inflated state?
1466 // Consider what happens when a thread unlocks a stack-locked object.
1467 // It attempts to use CAS to swing the displaced header value from the
1468 // on-stack basiclock back into the object header. Recall also that the
1469 // header value (hashcode, etc) can reside in (a) the object header, or
1470 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1471 // header in an objectMonitor. The inflate() routine must copy the header
1472 // value from the basiclock on the owner's stack to the objectMonitor, all
1473 // the while preserving the hashCode stability invariants. If the owner
1474 // decides to release the lock while the value is 0, the unlock will fail
1475 // and control will eventually pass from slow_exit() to inflate. The owner
1476 // will then spin, waiting for the 0 value to disappear. Put another way,
1477 // the 0 causes the owner to stall if the owner happens to try to
1478 // drop the lock (restoring the header from the basiclock to the object)
1479 // while inflation is in-progress. This protocol avoids races that might
1480 // would otherwise permit hashCode values to change or "flicker" for an object.
1481 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1482 // 0 serves as a "BUSY" inflate-in-progress indicator.
1483
1484
1485 // fetch the displaced mark from the owner's stack.
1486 // The owner can't die or unwind past the lock while our INFLATING
1487 // object is in the mark. Furthermore the owner can't complete
1488 // an unlock on the object, either.
1489 markOop dmw = mark->displaced_mark_helper();
1490 assert(dmw->is_neutral(), "invariant");
1491
1492 // Setup monitor fields to proper values -- prepare the monitor
1493 m->set_header(dmw);
1494
1495 // Optimization: if the mark->locker stack address is associated
1496 // with this thread we could simply set m->_owner = Self.
1497 // Note that a thread can inflate an object
1498 // that it has stack-locked -- as might happen in wait() -- directly
1499 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1500 m->set_owner(mark->locker());
1501 m->set_object(object);
1502 // TODO-FIXME: assert BasicLock->dhw != 0.
1503
1504 // Must preserve store ordering. The monitor state must
1505 // be stable at the time of publishing the monitor address.
1506 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant");
1507 object->release_set_mark(markOopDesc::encode(m));
1508
1509 // Hopefully the performance counters are allocated on distinct cache lines
1510 // to avoid false sharing on MP systems ...
1511 OM_PERFDATA_OP(Inflations, inc());
1512 TEVENT(Inflate: overwrite stacklock);
1513 if (log_is_enabled(Debug, monitorinflation)) {
1514 if (object->is_instance()) {
1515 ResourceMark rm;
1516 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1517 p2i(object), p2i(object->mark()),
1518 object->klass()->external_name());
1519 }
1520 }
1521 if (event.should_commit()) {
1522 post_monitor_inflate_event(event, object, cause);
1523 }
1524 return m;
1525 }
1526
1527 // CASE: neutral
1528 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1529 // If we know we're inflating for entry it's better to inflate by swinging a
1530 // pre-locked objectMonitor pointer into the object header. A successful
1531 // CAS inflates the object *and* confers ownership to the inflating thread.
1532 // In the current implementation we use a 2-step mechanism where we CAS()
1533 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1534 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1535 // would be useful.
1536
1537 assert(mark->is_neutral(), "invariant");
1538 ObjectMonitor * m = omAlloc(Self);
1539 // prepare m for installation - set monitor to initial state
1540 m->Recycle();
1541 m->set_header(mark);
1542 m->set_owner(NULL);
1543 m->set_object(object);
1544 m->_recursions = 0;
1545 m->_Responsible = NULL;
1546 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1547
1548 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) {
1549 m->set_object(NULL);
1550 m->set_owner(NULL);
1551 m->Recycle();
1552 omRelease(Self, m, true);
1553 m = NULL;
1554 continue;
1555 // interference - the markword changed - just retry.
1556 // The state-transitions are one-way, so there's no chance of
1557 // live-lock -- "Inflated" is an absorbing state.
1558 }
1559
1560 // Hopefully the performance counters are allocated on distinct
1561 // cache lines to avoid false sharing on MP systems ...
1562 OM_PERFDATA_OP(Inflations, inc());
1563 TEVENT(Inflate: overwrite neutral);
1564 if (log_is_enabled(Debug, monitorinflation)) {
1565 if (object->is_instance()) {
1566 ResourceMark rm;
1567 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1568 p2i(object), p2i(object->mark()),
1569 object->klass()->external_name());
1570 }
1571 }
1572 if (event.should_commit()) {
1573 post_monitor_inflate_event(event, object, cause);
1574 }
1575 return m;
1576 }
1577 }
1578
1579
1580 // Deflate_idle_monitors() is called at all safepoints, immediately
1581 // after all mutators are stopped, but before any objects have moved.
1582 // It traverses the list of known monitors, deflating where possible.
1583 // The scavenged monitor are returned to the monitor free list.
1584 //
1585 // Beware that we scavenge at *every* stop-the-world point.
1586 // Having a large number of monitors in-circulation negatively
1587 // impacts the performance of some applications (e.g., PointBase).
1588 // Broadly, we want to minimize the # of monitors in circulation.
1589 //
1590 // We have added a flag, MonitorInUseLists, which creates a list
1591 // of active monitors for each thread. deflate_idle_monitors()
1592 // only scans the per-thread in-use lists. omAlloc() puts all
1593 // assigned monitors on the per-thread list. deflate_idle_monitors()
1594 // returns the non-busy monitors to the global free list.
1595 // When a thread dies, omFlush() adds the list of active monitors for
1596 // that thread to a global gOmInUseList acquiring the
1597 // global list lock. deflate_idle_monitors() acquires the global
1598 // list lock to scan for non-busy monitors to the global free list.
1599 // An alternative could have used a single global in-use list. The
1600 // downside would have been the additional cost of acquiring the global list lock
1601 // for every omAlloc().
1602 //
1603 // Perversely, the heap size -- and thus the STW safepoint rate --
1604 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1605 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1606 // This is an unfortunate aspect of this design.
1607
1608 enum ManifestConstants {
1609 ClearResponsibleAtSTW = 0
1610 };
1611
1612 // Deflate a single monitor if not in-use
1613 // Return true if deflated, false if in-use
1614 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1615 ObjectMonitor** freeHeadp,
1616 ObjectMonitor** freeTailp) {
1617 bool deflated;
1618 // Normal case ... The monitor is associated with obj.
1619 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant");
1620 guarantee(mid == obj->mark()->monitor(), "invariant");
1621 guarantee(mid->header()->is_neutral(), "invariant");
1622
1623 if (mid->is_busy()) {
1624 if (ClearResponsibleAtSTW) mid->_Responsible = NULL;
1625 deflated = false;
1626 } else {
1627 // Deflate the monitor if it is no longer being used
1628 // It's idle - scavenge and return to the global free list
1629 // plain old deflation ...
1630 TEVENT(deflate_idle_monitors - scavenge1);
1631 if (log_is_enabled(Debug, monitorinflation)) {
1632 if (obj->is_instance()) {
1633 ResourceMark rm;
1634 log_debug(monitorinflation)("Deflating object " INTPTR_FORMAT " , "
1635 "mark " INTPTR_FORMAT " , type %s",
1636 p2i(obj), p2i(obj->mark()),
1637 obj->klass()->external_name());
1638 }
1639 }
1640
1641 // Restore the header back to obj
1642 obj->release_set_mark(mid->header());
1643 mid->clear();
1644
1645 assert(mid->object() == NULL, "invariant");
1646
1647 // Move the object to the working free list defined by freeHeadp, freeTailp
1648 if (*freeHeadp == NULL) *freeHeadp = mid;
1649 if (*freeTailp != NULL) {
1650 ObjectMonitor * prevtail = *freeTailp;
1651 assert(prevtail->FreeNext == NULL, "cleaned up deflated?");
1652 prevtail->FreeNext = mid;
1653 }
1654 *freeTailp = mid;
1655 deflated = true;
1656 }
1657 return deflated;
1658 }
1659
1660 // Walk a given monitor list, and deflate idle monitors
1661 // The given list could be a per-thread list or a global list
1662 // Caller acquires gListLock.
1663 //
1664 // In the case of parallel processing of thread local monitor lists,
1665 // work is done by Threads::parallel_threads_do() which ensures that
1666 // each Java thread is processed by exactly one worker thread, and
1667 // thus avoid conflicts that would arise when worker threads would
1668 // process the same monitor lists concurrently.
1669 //
1670 // See also ParallelSPCleanupTask and
1671 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and
1672 // Threads::parallel_java_threads_do() in thread.cpp.
1673 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp,
1674 ObjectMonitor** freeHeadp,
1675 ObjectMonitor** freeTailp) {
1676 ObjectMonitor* mid;
1677 ObjectMonitor* next;
1678 ObjectMonitor* cur_mid_in_use = NULL;
1679 int deflated_count = 0;
1680
1681 for (mid = *listHeadp; mid != NULL;) {
1682 oop obj = (oop) mid->object();
1683 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) {
1684 // if deflate_monitor succeeded,
1685 // extract from per-thread in-use list
1686 if (mid == *listHeadp) {
1687 *listHeadp = mid->FreeNext;
1688 } else if (cur_mid_in_use != NULL) {
1689 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1690 }
1691 next = mid->FreeNext;
1692 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list
1693 mid = next;
1694 deflated_count++;
1695 } else {
1696 cur_mid_in_use = mid;
1697 mid = mid->FreeNext;
1698 }
1699 }
1700 return deflated_count;
1701 }
1702
1703 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1704 counters->nInuse = 0; // currently associated with objects
1705 counters->nInCirculation = 0; // extant
1706 counters->nScavenged = 0; // reclaimed
1707 }
1708
1709 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) {
1710 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1711 bool deflated = false;
1712
1713 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1714 ObjectMonitor * freeTailp = NULL;
1715
1716 TEVENT(deflate_idle_monitors);
1717 // Prevent omFlush from changing mids in Thread dtor's during deflation
1718 // And in case the vm thread is acquiring a lock during a safepoint
1719 // See e.g. 6320749
1720 Thread::muxAcquire(&gListLock, "scavenge - return");
1721
1722 if (MonitorInUseLists) {
1723 // Note: the thread-local monitors lists get deflated in
1724 // a separate pass. See deflate_thread_local_monitors().
1725
1726 // For moribund threads, scan gOmInUseList
1727 if (gOmInUseList) {
1728 counters->nInCirculation += gOmInUseCount;
1729 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp);
1730 gOmInUseCount -= deflated_count;
1731 counters->nScavenged += deflated_count;
1732 counters->nInuse += gOmInUseCount;
1733 }
1734
1735 } else {
1736 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
1737 for (; block != NULL; block = next(block)) {
1738 // Iterate over all extant monitors - Scavenge all idle monitors.
1739 assert(block->object() == CHAINMARKER, "must be a block header");
1740 counters->nInCirculation += _BLOCKSIZE;
1741 for (int i = 1; i < _BLOCKSIZE; i++) {
1742 ObjectMonitor* mid = (ObjectMonitor*)&block[i];
1743 oop obj = (oop)mid->object();
1744
1745 if (obj == NULL) {
1746 // The monitor is not associated with an object.
1747 // The monitor should either be a thread-specific private
1748 // free list or the global free list.
1749 // obj == NULL IMPLIES mid->is_busy() == 0
1750 guarantee(!mid->is_busy(), "invariant");
1751 continue;
1752 }
1753 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp);
1754
1755 if (deflated) {
1756 mid->FreeNext = NULL;
1757 counters->nScavenged++;
1758 } else {
1759 counters->nInuse++;
1760 }
1761 }
1762 }
1763 }
1764
1765 // Move the scavenged monitors back to the global free list.
1766 if (freeHeadp != NULL) {
1767 guarantee(freeTailp != NULL && counters->nScavenged > 0, "invariant");
1768 assert(freeTailp->FreeNext == NULL, "invariant");
1769 // constant-time list splice - prepend scavenged segment to gFreeList
1770 freeTailp->FreeNext = gFreeList;
1771 gFreeList = freeHeadp;
1772 }
1773 Thread::muxRelease(&gListLock);
1774
1775 }
1776
1777 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1778 gMonitorFreeCount += counters->nScavenged;
1779
1780 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree.
1781
1782 if (ObjectMonitor::Knob_Verbose) {
1783 tty->print_cr("INFO: Deflate: InCirc=%d InUse=%d Scavenged=%d "
1784 "ForceMonitorScavenge=%d : pop=%d free=%d",
1785 counters->nInCirculation, counters->nInuse, counters->nScavenged, ForceMonitorScavenge,
1786 gMonitorPopulation, gMonitorFreeCount);
1787 tty->flush();
1788 }
1789
1790 ForceMonitorScavenge = 0; // Reset
1791
1792 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged));
1793 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation));
1794
1795 // TODO: Add objectMonitor leak detection.
1796 // Audit/inventory the objectMonitors -- make sure they're all accounted for.
1797 GVars.stwRandom = os::random();
1798 GVars.stwCycle++;
1799 }
1800
1801 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) {
1802 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1803 if (!MonitorInUseLists) return;
1804
1805 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1806 ObjectMonitor * freeTailp = NULL;
1807
1808 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp);
1809
1810 Thread::muxAcquire(&gListLock, "scavenge - return");
1811
1812 // Adjust counters
1813 counters->nInCirculation += thread->omInUseCount;
1814 thread->omInUseCount -= deflated_count;
1815 if (ObjectMonitor::Knob_VerifyInUse) {
1816 verifyInUse(thread);
1817 }
1818 counters->nScavenged += deflated_count;
1819 counters->nInuse += thread->omInUseCount;
1820
1821 // Move the scavenged monitors back to the global free list.
1822 if (freeHeadp != NULL) {
1823 guarantee(freeTailp != NULL && deflated_count > 0, "invariant");
1824 assert(freeTailp->FreeNext == NULL, "invariant");
1825
1826 // constant-time list splice - prepend scavenged segment to gFreeList
1827 freeTailp->FreeNext = gFreeList;
1828 gFreeList = freeHeadp;
1829 }
1830 Thread::muxRelease(&gListLock);
1831 }
1832
1833 // Monitor cleanup on JavaThread::exit
1834
1835 // Iterate through monitor cache and attempt to release thread's monitors
1836 // Gives up on a particular monitor if an exception occurs, but continues
1837 // the overall iteration, swallowing the exception.
1838 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1839 private:
1840 TRAPS;
1841
1842 public:
1843 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {}
1844 void do_monitor(ObjectMonitor* mid) {
1845 if (mid->owner() == THREAD) {
1846 if (ObjectMonitor::Knob_VerifyMatch != 0) {
1847 ResourceMark rm;
1848 Handle obj(THREAD, (oop) mid->object());
1849 tty->print("INFO: unexpected locked object:");
1850 javaVFrame::print_locked_object_class_name(tty, obj, "locked");
1851 fatal("exiting JavaThread=" INTPTR_FORMAT
1852 " unexpectedly owns ObjectMonitor=" INTPTR_FORMAT,
1853 p2i(THREAD), p2i(mid));
1854 }
1855 (void)mid->complete_exit(CHECK);
1856 }
1857 }
1858 };
1859
1860 // Release all inflated monitors owned by THREAD. Lightweight monitors are
1861 // ignored. This is meant to be called during JNI thread detach which assumes
1862 // all remaining monitors are heavyweight. All exceptions are swallowed.
1863 // Scanning the extant monitor list can be time consuming.
1864 // A simple optimization is to add a per-thread flag that indicates a thread
1865 // called jni_monitorenter() during its lifetime.
1866 //
1867 // Instead of No_Savepoint_Verifier it might be cheaper to
1868 // use an idiom of the form:
1869 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1870 // <code that must not run at safepoint>
1871 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1872 // Since the tests are extremely cheap we could leave them enabled
1873 // for normal product builds.
1874
1875 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) {
1876 assert(THREAD == JavaThread::current(), "must be current Java thread");
1877 NoSafepointVerifier nsv;
1878 ReleaseJavaMonitorsClosure rjmc(THREAD);
1879 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread");
1880 ObjectSynchronizer::monitors_iterate(&rjmc);
1881 Thread::muxRelease(&gListLock);
1882 THREAD->clear_pending_exception();
1883 }
1884
1885 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1886 switch (cause) {
1887 case inflate_cause_vm_internal: return "VM Internal";
1888 case inflate_cause_monitor_enter: return "Monitor Enter";
1889 case inflate_cause_wait: return "Monitor Wait";
1890 case inflate_cause_notify: return "Monitor Notify";
1891 case inflate_cause_hash_code: return "Monitor Hash Code";
1892 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1893 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1894 default:
1895 ShouldNotReachHere();
1896 }
1897 return "Unknown";
1898 }
1899
1900 static void post_monitor_inflate_event(EventJavaMonitorInflate& event,
1901 const oop obj,
1902 const ObjectSynchronizer::InflateCause cause) {
1903 #if INCLUDE_TRACE
1904 assert(event.should_commit(), "check outside");
1905 event.set_monitorClass(obj->klass());
1906 event.set_address((TYPE_ADDRESS)(uintptr_t)(void*)obj);
1907 event.set_cause((u1)cause);
1908 event.commit();
1909 #endif
1910 }
1911
1912 //------------------------------------------------------------------------------
1913 // Debugging code
1914
1915 void ObjectSynchronizer::sanity_checks(const bool verbose,
1916 const uint cache_line_size,
1917 int *error_cnt_ptr,
1918 int *warning_cnt_ptr) {
1919 u_char *addr_begin = (u_char*)&GVars;
1920 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom;
1921 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence;
1922
1923 if (verbose) {
1924 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT,
1925 sizeof(SharedGlobals));
1926 }
1927
1928 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin);
1929 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom);
1930
1931 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin);
1932 if (verbose) {
1933 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence);
1934 }
1935
1936 if (cache_line_size != 0) {
1937 // We were able to determine the L1 data cache line size so
1938 // do some cache line specific sanity checks
1939
1940 if (offset_stwRandom < cache_line_size) {
1941 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer "
1942 "to the struct beginning than a cache line which permits "
1943 "false sharing.");
1944 (*warning_cnt_ptr)++;
1945 }
1946
1947 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) {
1948 tty->print_cr("WARNING: the SharedGlobals.stwRandom and "
1949 "SharedGlobals.hcSequence fields are closer than a cache "
1950 "line which permits false sharing.");
1951 (*warning_cnt_ptr)++;
1952 }
1953
1954 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) {
1955 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer "
1956 "to the struct end than a cache line which permits false "
1957 "sharing.");
1958 (*warning_cnt_ptr)++;
1959 }
1960 }
1961 }
1962
1963 #ifndef PRODUCT
1964
1965 // Check if monitor belongs to the monitor cache
1966 // The list is grow-only so it's *relatively* safe to traverse
1967 // the list of extant blocks without taking a lock.
1968
1969 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) {
1970 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
1971 while (block != NULL) {
1972 assert(block->object() == CHAINMARKER, "must be a block header");
1973 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) {
1974 address mon = (address)monitor;
1975 address blk = (address)block;
1976 size_t diff = mon - blk;
1977 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "must be aligned");
1978 return 1;
1979 }
1980 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
1981 }
1982 return 0;
1983 }
1984
1985 #endif