1 /*
2 * Copyright (c) 1998, 2024, 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 "jfr/jfrEvents.hpp"
28 #include "gc/shared/suspendibleThreadSet.hpp"
29 #include "logging/log.hpp"
30 #include "logging/logStream.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/padded.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "memory/universe.hpp"
35 #include "oops/markWord.hpp"
36 #include "oops/oop.inline.hpp"
37 #include "runtime/atomic.hpp"
38 #include "runtime/biasedLocking.hpp"
39 #include "runtime/globals.hpp"
40 #include "runtime/handles.inline.hpp"
41 #include "runtime/handshake.hpp"
42 #include "runtime/interfaceSupport.inline.hpp"
43 #include "runtime/lockStack.inline.hpp"
44 #include "runtime/mutexLocker.hpp"
45 #include "runtime/objectMonitor.hpp"
46 #include "runtime/objectMonitor.inline.hpp"
47 #include "runtime/os.inline.hpp"
48 #include "runtime/osThread.hpp"
49 #include "runtime/perfData.hpp"
50 #include "runtime/safepointMechanism.inline.hpp"
51 #include "runtime/safepointVerifiers.hpp"
52 #include "runtime/sharedRuntime.hpp"
53 #include "runtime/stubRoutines.hpp"
54 #include "runtime/synchronizer.hpp"
55 #include "runtime/thread.inline.hpp"
56 #include "runtime/timer.hpp"
57 #include "runtime/trimNativeHeap.hpp"
58 #include "runtime/vframe.hpp"
59 #include "runtime/vmThread.hpp"
60 #include "utilities/align.hpp"
61 #include "utilities/dtrace.hpp"
62 #include "utilities/events.hpp"
63 #include "utilities/globalDefinitions.hpp"
64 #include "utilities/preserveException.hpp"
65
66 void MonitorList::add(ObjectMonitor* m) {
67 ObjectMonitor* head;
68 do {
69 head = Atomic::load(&_head);
70 m->set_next_om(head);
71 } while (Atomic::cmpxchg(&_head, head, m) != head);
72
73 size_t count = Atomic::add(&_count, 1u);
74 if (count > max()) {
75 Atomic::inc(&_max);
76 }
77 }
78
79 size_t MonitorList::count() const {
80 return Atomic::load(&_count);
81 }
82
83 size_t MonitorList::max() const {
84 return Atomic::load(&_max);
85 }
86
87 // Walk the in-use list and unlink (at most MonitorDeflationMax) deflated
88 // ObjectMonitors. Returns the number of unlinked ObjectMonitors.
89 size_t MonitorList::unlink_deflated(Thread* current, LogStream* ls,
90 elapsedTimer* timer_p,
91 GrowableArray<ObjectMonitor*>* unlinked_list) {
92 size_t unlinked_count = 0;
93 ObjectMonitor* prev = NULL;
94 ObjectMonitor* head = Atomic::load_acquire(&_head);
95 ObjectMonitor* m = head;
96 // The in-use list head can be NULL during the final audit.
97 while (m != NULL) {
98 if (m->is_being_async_deflated()) {
99 // Find next live ObjectMonitor.
100 ObjectMonitor* next = m;
101 do {
102 ObjectMonitor* next_next = next->next_om();
103 unlinked_count++;
104 unlinked_list->append(next);
105 next = next_next;
106 if (unlinked_count >= (size_t)MonitorDeflationMax) {
107 // Reached the max so bail out on the gathering loop.
108 break;
109 }
110 } while (next != NULL && next->is_being_async_deflated());
111 if (prev == NULL) {
112 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, head, next);
113 if (prev_head != head) {
114 // Find new prev ObjectMonitor that just got inserted.
115 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) {
116 prev = n;
117 }
118 prev->set_next_om(next);
119 }
120 } else {
121 prev->set_next_om(next);
122 }
123 if (unlinked_count >= (size_t)MonitorDeflationMax) {
124 // Reached the max so bail out on the searching loop.
125 break;
126 }
127 m = next;
128 } else {
129 prev = m;
130 m = m->next_om();
131 }
132
133 if (current->is_Java_thread()) {
134 // A JavaThread must check for a safepoint/handshake and honor it.
135 ObjectSynchronizer::chk_for_block_req(current->as_Java_thread(), "unlinking",
136 "unlinked_count", unlinked_count,
137 ls, timer_p);
138 }
139 }
140 Atomic::sub(&_count, unlinked_count);
141 return unlinked_count;
142 }
143
144 MonitorList::Iterator MonitorList::iterator() const {
145 return Iterator(Atomic::load_acquire(&_head));
146 }
147
148 ObjectMonitor* MonitorList::Iterator::next() {
149 ObjectMonitor* current = _current;
150 _current = current->next_om();
151 return current;
152 }
153
154 // The "core" versions of monitor enter and exit reside in this file.
155 // The interpreter and compilers contain specialized transliterated
156 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp
157 // fast_lock(...) for instance. If you make changes here, make sure to modify the
158 // interpreter, and both C1 and C2 fast-path inline locking code emission.
159 //
160 // -----------------------------------------------------------------------------
161
162 #ifdef DTRACE_ENABLED
163
164 // Only bother with this argument setup if dtrace is available
165 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
166
167 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
168 char* bytes = NULL; \
169 int len = 0; \
170 jlong jtid = SharedRuntime::get_java_tid(thread); \
171 Symbol* klassname = obj->klass()->name(); \
172 if (klassname != NULL) { \
173 bytes = (char*)klassname->bytes(); \
174 len = klassname->utf8_length(); \
175 }
176
177 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
178 { \
179 if (DTraceMonitorProbes) { \
180 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
181 HOTSPOT_MONITOR_WAIT(jtid, \
182 (uintptr_t)(monitor), bytes, len, (millis)); \
183 } \
184 }
185
186 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY
187 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL
188 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
189
190 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
191 { \
192 if (DTraceMonitorProbes) { \
193 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
194 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
195 (uintptr_t)(monitor), bytes, len); \
196 } \
197 }
198
199 #else // ndef DTRACE_ENABLED
200
201 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
202 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
203
204 #endif // ndef DTRACE_ENABLED
205
206 // This exists only as a workaround of dtrace bug 6254741
207 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
208 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
209 return 0;
210 }
211
212 static const int NINFLATIONLOCKS = 256;
213 static os::PlatformMutex* gInflationLocks[NINFLATIONLOCKS];
214
215 void ObjectSynchronizer::initialize() {
216 for (int i = 0; i < NINFLATIONLOCKS; i++) {
217 gInflationLocks[i] = new os::PlatformMutex();
218 }
219 // Start the ceiling with the estimate for one thread.
220 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate);
221
222 // Start the timer for deflations, so it does not trigger immediately.
223 _last_async_deflation_time_ns = os::javaTimeNanos();
224 }
225
226 MonitorList ObjectSynchronizer::_in_use_list;
227 // monitors_used_above_threshold() policy is as follows:
228 //
229 // The ratio of the current _in_use_list count to the ceiling is used
230 // to determine if we are above MonitorUsedDeflationThreshold and need
231 // to do an async monitor deflation cycle. The ceiling is increased by
232 // AvgMonitorsPerThreadEstimate when a thread is added to the system
233 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is
234 // removed from the system.
235 //
236 // Note: If the _in_use_list max exceeds the ceiling, then
237 // monitors_used_above_threshold() will use the in_use_list max instead
238 // of the thread count derived ceiling because we have used more
239 // ObjectMonitors than the estimated average.
240 //
241 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax
242 // no-progress async monitor deflation cycles in a row, then the ceiling
243 // is adjusted upwards by monitors_used_above_threshold().
244 //
245 // Start the ceiling with the estimate for one thread in initialize()
246 // which is called after cmd line options are processed.
247 static size_t _in_use_list_ceiling = 0;
248 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false;
249 bool volatile ObjectSynchronizer::_is_final_audit = false;
250 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0;
251 static uintx _no_progress_cnt = 0;
252 static bool _no_progress_skip_increment = false;
253
254 // =====================> Quick functions
255
256 // The quick_* forms are special fast-path variants used to improve
257 // performance. In the simplest case, a "quick_*" implementation could
258 // simply return false, in which case the caller will perform the necessary
259 // state transitions and call the slow-path form.
260 // The fast-path is designed to handle frequently arising cases in an efficient
261 // manner and is just a degenerate "optimistic" variant of the slow-path.
262 // returns true -- to indicate the call was satisfied.
263 // returns false -- to indicate the call needs the services of the slow-path.
264 // A no-loitering ordinance is in effect for code in the quick_* family
265 // operators: safepoints or indefinite blocking (blocking that might span a
266 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon
267 // entry.
268 //
269 // Consider: An interesting optimization is to have the JIT recognize the
270 // following common idiom:
271 // synchronized (someobj) { .... ; notify(); }
272 // That is, we find a notify() or notifyAll() call that immediately precedes
273 // the monitorexit operation. In that case the JIT could fuse the operations
274 // into a single notifyAndExit() runtime primitive.
275
276 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) {
277 assert(current->thread_state() == _thread_in_Java, "invariant");
278 NoSafepointVerifier nsv;
279 if (obj == NULL) return false; // slow-path for invalid obj
280 const markWord mark = obj->mark();
281
282 if (LockingMode == LM_LIGHTWEIGHT) {
283 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) {
284 // Degenerate notify
285 // fast-locked by caller so by definition the implied waitset is empty.
286 return true;
287 }
288 } else if (LockingMode == LM_LEGACY) {
289 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
290 // Degenerate notify
291 // stack-locked by caller so by definition the implied waitset is empty.
292 return true;
293 }
294 }
295
296 if (mark.has_monitor()) {
297 ObjectMonitor* const mon = mark.monitor();
298 assert(mon->object() == oop(obj), "invariant");
299 if (mon->owner() != current) return false; // slow-path for IMS exception
300
301 if (mon->first_waiter() != NULL) {
302 // We have one or more waiters. Since this is an inflated monitor
303 // that we own, we can transfer one or more threads from the waitset
304 // to the entrylist here and now, avoiding the slow-path.
305 if (all) {
306 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, current);
307 } else {
308 DTRACE_MONITOR_PROBE(notify, mon, obj, current);
309 }
310 int free_count = 0;
311 do {
312 mon->INotify(current);
313 ++free_count;
314 } while (mon->first_waiter() != NULL && all);
315 OM_PERFDATA_OP(Notifications, inc(free_count));
316 }
317 return true;
318 }
319
320 // biased locking and any other IMS exception states take the slow-path
321 return false;
322 }
323
324
325 // The LockNode emitted directly at the synchronization site would have
326 // been too big if it were to have included support for the cases of inflated
327 // recursive enter and exit, so they go here instead.
328 // Note that we can't safely call AsyncPrintJavaStack() from within
329 // quick_enter() as our thread state remains _in_Java.
330
331 bool ObjectSynchronizer::quick_enter(oop obj, JavaThread* current,
332 BasicLock * lock) {
333 assert(current->thread_state() == _thread_in_Java, "invariant");
334 NoSafepointVerifier nsv;
335 if (obj == NULL) return false; // Need to throw NPE
336
337 if (obj->klass()->is_value_based()) {
338 return false;
339 }
340
341 if (LockingMode == LM_LIGHTWEIGHT) {
342 LockStack& lock_stack = current->lock_stack();
343 if (lock_stack.is_full()) {
344 // Always go into runtime if the lock stack is full.
345 return false;
346 }
347 if (lock_stack.try_recursive_enter(obj)) {
348 // Recursive lock successful.
349 return true;
350 }
351 }
352
353 const markWord mark = obj->mark();
354
355 if (mark.has_monitor()) {
356 ObjectMonitor* const m = mark.monitor();
357 // An async deflation or GC can race us before we manage to make
358 // the ObjectMonitor busy by setting the owner below. If we detect
359 // that race we just bail out to the slow-path here.
360 if (m->object_peek() == NULL) {
361 return false;
362 }
363 JavaThread* const owner = (JavaThread*) m->owner_raw();
364
365 // Lock contention and Transactional Lock Elision (TLE) diagnostics
366 // and observability
367 // Case: light contention possibly amenable to TLE
368 // Case: TLE inimical operations such as nested/recursive synchronization
369
370 if (owner == current) {
371 m->_recursions++;
372 return true;
373 }
374
375 if (LockingMode != LM_LIGHTWEIGHT) {
376 // This Java Monitor is inflated so obj's header will never be
377 // displaced to this thread's BasicLock. Make the displaced header
378 // non-NULL so this BasicLock is not seen as recursive nor as
379 // being locked. We do this unconditionally so that this thread's
380 // BasicLock cannot be mis-interpreted by any stack walkers. For
381 // performance reasons, stack walkers generally first check for
382 // Biased Locking in the object's header, the second check is for
383 // stack-locking in the object's header, the third check is for
384 // recursive stack-locking in the displaced header in the BasicLock,
385 // and last are the inflated Java Monitor (ObjectMonitor) checks.
386 lock->set_displaced_header(markWord::unused_mark());
387 }
388
389 if (owner == NULL && m->try_set_owner_from(NULL, current) == NULL) {
390 assert(m->_recursions == 0, "invariant");
391 return true;
392 }
393 }
394
395 // Note that we could inflate in quick_enter.
396 // This is likely a useful optimization
397 // Critically, in quick_enter() we must not:
398 // -- perform bias revocation, or
399 // -- block indefinitely, or
400 // -- reach a safepoint
401
402 return false; // revert to slow-path
403 }
404
405 // Handle notifications when synchronizing on value based classes
406 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) {
407 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
408 frame last_frame = locking_thread->last_frame();
409 bool bcp_was_adjusted = false;
410 // Don't decrement bcp if it points to the frame's first instruction. This happens when
411 // handle_sync_on_value_based_class() is called because of a synchronized method. There
412 // is no actual monitorenter instruction in the byte code in this case.
413 if (last_frame.is_interpreted_frame() &&
414 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) {
415 // adjust bcp to point back to monitorenter so that we print the correct line numbers
416 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1);
417 bcp_was_adjusted = true;
418 }
419
420 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) {
421 ResourceMark rm;
422 stringStream ss;
423 locking_thread->print_stack_on(&ss);
424 char* base = (char*)strstr(ss.base(), "at");
425 char* newline = (char*)strchr(ss.base(), '\n');
426 if (newline != NULL) {
427 *newline = '\0';
428 }
429 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base);
430 } else {
431 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses");
432 ResourceMark rm;
433 Log(valuebasedclasses) vblog;
434
435 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name());
436 if (locking_thread->has_last_Java_frame()) {
437 LogStream info_stream(vblog.info());
438 locking_thread->print_stack_on(&info_stream);
439 } else {
440 vblog.info("Cannot find the last Java frame");
441 }
442
443 EventSyncOnValueBasedClass event;
444 if (event.should_commit()) {
445 event.set_valueBasedClass(obj->klass());
446 event.commit();
447 }
448 }
449
450 if (bcp_was_adjusted) {
451 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1);
452 }
453 }
454
455 static bool useHeavyMonitors() {
456 #if defined(X86) || defined(AARCH64) || defined(PPC64) || defined(RISCV64)
457 return LockingMode == LM_MONITOR;
458 #else
459 return false;
460 #endif
461 }
462
463 // -----------------------------------------------------------------------------
464 // Monitor Enter/Exit
465
466 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
467 // When called with locking_thread != Thread::current() some mechanism must synchronize
468 // the locking_thread with respect to the current thread. Currently only used when
469 // deoptimizing and re-locking locks. See Deoptimization::relock_objects
470 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
471 if (!enter_fast_impl(obj, lock, locking_thread)) {
472 // Inflated ObjectMonitor::enter_for is required
473
474 // An async deflation can race after the inflate_for() call and before
475 // enter_for() can make the ObjectMonitor busy. enter_for() returns false
476 // if we have lost the race to async deflation and we simply try again.
477 while (true) {
478 ObjectMonitor* monitor = inflate_for(locking_thread, obj(), inflate_cause_monitor_enter);
479 if (monitor->enter_for(locking_thread)) {
480 return;
481 }
482 assert(monitor->is_being_async_deflated(), "must be");
483 }
484 }
485 }
486
487 void ObjectSynchronizer::enter(Handle obj, BasicLock* lock, JavaThread* current) {
488 assert(current == Thread::current(), "must be");
489 if (!enter_fast_impl(obj, lock, current)) {
490 // Inflated ObjectMonitor::enter is required
491
492 // An async deflation can race after the inflate() call and before
493 // enter() can make the ObjectMonitor busy. enter() returns false if
494 // we have lost the race to async deflation and we simply try again.
495 while (true) {
496 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter);
497 if (monitor->enter(current)) {
498 return;
499 }
500 }
501 }
502 }
503
504 // The interpreter and compiler assembly code tries to lock using the fast path
505 // of this algorithm. Make sure to update that code if the following function is
506 // changed. The implementation is extremely sensitive to race condition. Be careful.
507 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
508
509 if (obj->klass()->is_value_based()) {
510 handle_sync_on_value_based_class(obj, locking_thread);
511 }
512
513 if (!useHeavyMonitors()) {
514 if (LockingMode == LM_LIGHTWEIGHT) {
515 // Fast-locking does not use the 'lock' argument.
516 LockStack& lock_stack = locking_thread->lock_stack();
517 if (lock_stack.is_full()) {
518 // We unconditionally make room on the lock stack by inflating
519 // the least recently locked object on the lock stack.
520
521 // About the choice to inflate least recently locked object.
522 // First we must chose to inflate a lock, either some lock on
523 // the lock-stack or the lock that is currently being entered
524 // (which may or may not be on the lock-stack).
525 // Second the best lock to inflate is a lock which is entered
526 // in a control flow where there are only a very few locks being
527 // used, as the costly part of inflated locking is inflation,
528 // not locking. But this property is entirely program dependent.
529 // Third inflating the lock currently being entered on when it
530 // is not present on the lock-stack will result in a still full
531 // lock-stack. This creates a scenario where every deeper nested
532 // monitorenter must call into the runtime.
533 // The rational here is as follows:
534 // Because we cannot (currently) figure out the second, and want
535 // to avoid the third, we inflate a lock on the lock-stack.
536 // The least recently locked lock is chosen as it is the lock
537 // with the longest critical section.
538
539 log_info(monitorinflation)("LockStack capacity exceeded, inflating.");
540 ObjectMonitor* monitor = inflate_for(locking_thread, lock_stack.bottom(), inflate_cause_vm_internal);
541 assert(monitor->owner() == Thread::current(), "must be owner=" PTR_FORMAT " current=" PTR_FORMAT " mark=" PTR_FORMAT,
542 p2i(monitor->owner()), p2i(Thread::current()), monitor->object()->mark_acquire().value());
543 assert(!lock_stack.is_full(), "must have made room here");
544 }
545
546 markWord mark = obj()->mark_acquire();
547 while (mark.is_neutral()) {
548 // Retry until a lock state change has been observed. cas_set_mark() may collide with non lock bits modifications.
549 // Try to swing into 'fast-locked' state.
550 assert(!lock_stack.contains(obj()), "thread must not already hold the lock");
551 const markWord locked_mark = mark.set_fast_locked();
552 const markWord old_mark = obj()->cas_set_mark(locked_mark, mark);
553 if (old_mark == mark) {
554 // Successfully fast-locked, push object to lock-stack and return.
555 lock_stack.push(obj());
556 return true;
557 }
558 mark = old_mark;
559 }
560
561 if (mark.is_fast_locked() && lock_stack.try_recursive_enter(obj())) {
562 // Recursive lock successful.
563 return true;
564 }
565
566 // Failed to fast lock.
567 return false;
568 } else if (LockingMode == LM_LEGACY) {
569 if (UseBiasedLocking) {
570 BiasedLocking::revoke(locking_thread, obj);
571 }
572
573 markWord mark = obj->mark();
574 if (mark.is_neutral()) {
575 // Anticipate successful CAS -- the ST of the displaced mark must
576 // be visible <= the ST performed by the CAS.
577 lock->set_displaced_header(mark);
578 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) {
579 return true;
580 }
581 } else if (mark.has_locker() &&
582 locking_thread->is_lock_owned((address) mark.locker())) {
583 assert(lock != mark.locker(), "must not re-lock the same lock");
584 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock");
585 lock->set_displaced_header(markWord::from_pointer(NULL));
586 return true;
587 }
588
589 // The object header will never be displaced to this lock,
590 // so it does not matter what the value is, except that it
591 // must be non-zero to avoid looking like a re-entrant lock,
592 // and must not look locked either.
593 lock->set_displaced_header(markWord::unused_mark());
594
595 // Failed to fast lock.
596 return false;
597 }
598 }
599
600 return false;
601 }
602
603 void ObjectSynchronizer::exit(oop object, BasicLock* lock, JavaThread* current) {
604 if (!useHeavyMonitors()) {
605 markWord mark = object->mark();
606 if (LockingMode == LM_LIGHTWEIGHT) {
607 // Fast-locking does not use the 'lock' argument.
608 LockStack& lock_stack = current->lock_stack();
609 if (mark.is_fast_locked() && lock_stack.try_recursive_exit(object)) {
610 // Recursively unlocked.
611 return;
612 }
613
614 if (mark.is_fast_locked() && lock_stack.is_recursive(object)) {
615 // This lock is recursive but is not at the top of the lock stack so we're
616 // doing an unbalanced exit. We have to fall thru to inflation below and
617 // let ObjectMonitor::exit() do the unlock.
618 } else {
619 while (mark.is_fast_locked()) {
620 // Retry until a lock state change has been observed. cas_set_mark() may collide with non lock bits modifications.
621 const markWord unlocked_mark = mark.set_unlocked();
622 const markWord old_mark = object->cas_set_mark(unlocked_mark, mark);
623 if (old_mark == mark) {
624 size_t recursions = lock_stack.remove(object) - 1;
625 assert(recursions == 0, "must not be recursive here");
626 return;
627 }
628 mark = old_mark;
629 }
630 }
631 } else if (LockingMode == LM_LEGACY) {
632 markWord dhw = lock->displaced_header();
633 if (dhw.value() == 0) {
634 // If the displaced header is NULL, then this exit matches up with
635 // a recursive enter. No real work to do here except for diagnostics.
636 #ifndef PRODUCT
637 if (mark != markWord::INFLATING()) {
638 // Only do diagnostics if we are not racing an inflation. Simply
639 // exiting a recursive enter of a Java Monitor that is being
640 // inflated is safe; see the has_monitor() comment below.
641 assert(!mark.is_neutral(), "invariant");
642 assert(!mark.has_locker() ||
643 current->is_lock_owned((address)mark.locker()), "invariant");
644 if (mark.has_monitor()) {
645 // The BasicLock's displaced_header is marked as a recursive
646 // enter and we have an inflated Java Monitor (ObjectMonitor).
647 // This is a special case where the Java Monitor was inflated
648 // after this thread entered the stack-lock recursively. When a
649 // Java Monitor is inflated, we cannot safely walk the Java
650 // Monitor owner's stack and update the BasicLocks because a
651 // Java Monitor can be asynchronously inflated by a thread that
652 // does not own the Java Monitor.
653 ObjectMonitor* m = mark.monitor();
654 assert(m->object()->mark() == mark, "invariant");
655 assert(m->is_entered(current), "invariant");
656 }
657 }
658 #endif
659 return;
660 }
661
662 if (mark == markWord::from_pointer(lock)) {
663 // If the object is stack-locked by the current thread, try to
664 // swing the displaced header from the BasicLock back to the mark.
665 assert(dhw.is_neutral(), "invariant");
666 if (object->cas_set_mark(dhw, mark) == mark) {
667 return;
668 }
669 }
670 }
671 }
672
673 // We have to take the slow-path of possible inflation and then exit.
674 // The ObjectMonitor* can't be async deflated until ownership is
675 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
676 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal);
677 assert(!monitor->is_owner_anonymous(), "must not be");
678 monitor->exit(current);
679 }
680
681 // -----------------------------------------------------------------------------
682 // Class Loader support to workaround deadlocks on the class loader lock objects
683 // Also used by GC
684 // complete_exit()/reenter() are used to wait on a nested lock
685 // i.e. to give up an outer lock completely and then re-enter
686 // Used when holding nested locks - lock acquisition order: lock1 then lock2
687 // 1) complete_exit lock1 - saving recursion count
688 // 2) wait on lock2
689 // 3) when notified on lock2, unlock lock2
690 // 4) reenter lock1 with original recursion count
691 // 5) lock lock2
692 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
693 intx ObjectSynchronizer::complete_exit(Handle obj, JavaThread* current) {
694 if (UseBiasedLocking) {
695 BiasedLocking::revoke(current, obj);
696 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
697 }
698
699 // The ObjectMonitor* can't be async deflated until ownership is
700 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
701 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_vm_internal);
702 intptr_t ret_code = monitor->complete_exit(current);
703 return ret_code;
704 }
705
706 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
707 void ObjectSynchronizer::reenter(Handle obj, intx recursions, JavaThread* current) {
708 if (UseBiasedLocking) {
709 BiasedLocking::revoke(current, obj);
710 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
711 }
712
713 // An async deflation can race after the inflate() call and before
714 // reenter() -> enter() can make the ObjectMonitor busy. reenter() ->
715 // enter() returns false if we have lost the race to async deflation
716 // and we simply try again.
717 while (true) {
718 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_vm_internal);
719 if (monitor->reenter(recursions, current)) {
720 return;
721 }
722 }
723 }
724
725 // -----------------------------------------------------------------------------
726 // JNI locks on java objects
727 // NOTE: must use heavy weight monitor to handle jni monitor enter
728 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) {
729 if (obj->klass()->is_value_based()) {
730 handle_sync_on_value_based_class(obj, current);
731 }
732
733 // the current locking is from JNI instead of Java code
734 if (UseBiasedLocking) {
735 BiasedLocking::revoke(current, obj);
736 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
737 }
738 current->set_current_pending_monitor_is_from_java(false);
739 // An async deflation can race after the inflate() call and before
740 // enter() can make the ObjectMonitor busy. enter() returns false if
741 // we have lost the race to async deflation and we simply try again.
742 while (true) {
743 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_jni_enter);
744 if (monitor->enter(current)) {
745 break;
746 }
747 }
748 current->set_current_pending_monitor_is_from_java(true);
749 }
750
751 // NOTE: must use heavy weight monitor to handle jni monitor exit
752 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) {
753 JavaThread* current = THREAD;
754 if (UseBiasedLocking) {
755 Handle h_obj(current, obj);
756 BiasedLocking::revoke(current, h_obj);
757 obj = h_obj();
758 }
759 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
760
761 // The ObjectMonitor* can't be async deflated until ownership is
762 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
763 ObjectMonitor* monitor = inflate(current, obj, inflate_cause_jni_exit);
764 // If this thread has locked the object, exit the monitor. We
765 // intentionally do not use CHECK on check_owner because we must exit the
766 // monitor even if an exception was already pending.
767 if (monitor->check_owner(THREAD)) {
768 monitor->exit(current);
769 }
770 }
771
772 // -----------------------------------------------------------------------------
773 // Internal VM locks on java objects
774 // standard constructor, allows locking failures
775 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) {
776 _thread = thread;
777 _thread->check_for_valid_safepoint_state();
778 _obj = obj;
779
780 if (_obj() != NULL) {
781 ObjectSynchronizer::enter(_obj, &_lock, _thread);
782 }
783 }
784
785 ObjectLocker::~ObjectLocker() {
786 if (_obj() != NULL) {
787 ObjectSynchronizer::exit(_obj(), &_lock, _thread);
788 }
789 }
790
791
792 // -----------------------------------------------------------------------------
793 // Wait/Notify/NotifyAll
794 // NOTE: must use heavy weight monitor to handle wait()
795 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
796 JavaThread* current = THREAD;
797 if (UseBiasedLocking) {
798 BiasedLocking::revoke(current, obj);
799 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
800 }
801 if (millis < 0) {
802 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
803 }
804 // The ObjectMonitor* can't be async deflated because the _waiters
805 // field is incremented before ownership is dropped and decremented
806 // after ownership is regained.
807 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_wait);
808
809 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis);
810 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code
811
812 // This dummy call is in place to get around dtrace bug 6254741. Once
813 // that's fixed we can uncomment the following line, remove the call
814 // and change this function back into a "void" func.
815 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
816 int ret_code = dtrace_waited_probe(monitor, obj, THREAD);
817 return ret_code;
818 }
819
820 // No exception are possible in this case as we only use this internally when locking is
821 // correct and we have to wait until notified - so no interrupts or timeouts.
822 void ObjectSynchronizer::wait_uninterruptibly(Handle obj, JavaThread* current) {
823 if (UseBiasedLocking) {
824 BiasedLocking::revoke(current, obj);
825 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
826 }
827 // The ObjectMonitor* can't be async deflated because the _waiters
828 // field is incremented before ownership is dropped and decremented
829 // after ownership is regained.
830 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_wait);
831 monitor->wait(0 /* wait-forever */, false /* not interruptible */, current);
832 }
833
834 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
835 JavaThread* current = THREAD;
836 if (UseBiasedLocking) {
837 BiasedLocking::revoke(current, obj);
838 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
839 }
840
841 markWord mark = obj->mark();
842 if (LockingMode == LM_LIGHTWEIGHT) {
843 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
844 // Not inflated so there can't be any waiters to notify.
845 return;
846 }
847 } else if (LockingMode == LM_LEGACY) {
848 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
849 // Not inflated so there can't be any waiters to notify.
850 return;
851 }
852 }
853 // The ObjectMonitor* can't be async deflated until ownership is
854 // dropped by the calling thread.
855 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_notify);
856 monitor->notify(CHECK);
857 }
858
859 // NOTE: see comment of notify()
860 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
861 JavaThread* current = THREAD;
862 if (UseBiasedLocking) {
863 BiasedLocking::revoke(current, obj);
864 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
865 }
866
867 markWord mark = obj->mark();
868 if (LockingMode == LM_LIGHTWEIGHT) {
869 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
870 // Not inflated so there can't be any waiters to notify.
871 return;
872 }
873 } else if (LockingMode == LM_LEGACY) {
874 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
875 // Not inflated so there can't be any waiters to notify.
876 return;
877 }
878 }
879 // The ObjectMonitor* can't be async deflated until ownership is
880 // dropped by the calling thread.
881 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_notify);
882 monitor->notifyAll(CHECK);
883 }
884
885 // -----------------------------------------------------------------------------
886 // Hash Code handling
887
888 struct SharedGlobals {
889 char _pad_prefix[OM_CACHE_LINE_SIZE];
890 // This is a highly shared mostly-read variable.
891 // To avoid false-sharing it needs to be the sole occupant of a cache line.
892 volatile int stw_random;
893 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int));
894 // Hot RW variable -- Sequester to avoid false-sharing
895 volatile int hc_sequence;
896 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int));
897 };
898
899 static SharedGlobals GVars;
900
901 static markWord read_stable_mark(oop obj) {
902 markWord mark = obj->mark_acquire();
903 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) {
904 // New lightweight locking does not use the markWord::INFLATING() protocol.
905 return mark; // normal fast-path return
906 }
907
908 int its = 0;
909 for (;;) {
910 markWord mark = obj->mark_acquire();
911 if (!mark.is_being_inflated()) {
912 return mark; // normal fast-path return
913 }
914
915 // The object is being inflated by some other thread.
916 // The caller of read_stable_mark() must wait for inflation to complete.
917 // Avoid live-lock.
918
919 ++its;
920 if (its > 10000 || !os::is_MP()) {
921 if (its & 1) {
922 os::naked_yield();
923 } else {
924 // Note that the following code attenuates the livelock problem but is not
925 // a complete remedy. A more complete solution would require that the inflating
926 // thread hold the associated inflation lock. The following code simply restricts
927 // the number of spinners to at most one. We'll have N-2 threads blocked
928 // on the inflationlock, 1 thread holding the inflation lock and using
929 // a yield/park strategy, and 1 thread in the midst of inflation.
930 // A more refined approach would be to change the encoding of INFLATING
931 // to allow encapsulation of a native thread pointer. Threads waiting for
932 // inflation to complete would use CAS to push themselves onto a singly linked
933 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
934 // and calling park(). When inflation was complete the thread that accomplished inflation
935 // would detach the list and set the markword to inflated with a single CAS and
936 // then for each thread on the list, set the flag and unpark() the thread.
937
938 // Index into the lock array based on the current object address.
939 static_assert(is_power_of_2(NINFLATIONLOCKS), "must be");
940 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1);
941 int YieldThenBlock = 0;
942 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant");
943 gInflationLocks[ix]->lock();
944 while (obj->mark_acquire() == markWord::INFLATING()) {
945 // Beware: naked_yield() is advisory and has almost no effect on some platforms
946 // so we periodically call current->_ParkEvent->park(1).
947 // We use a mixed spin/yield/block mechanism.
948 if ((YieldThenBlock++) >= 16) {
949 Thread::current()->_ParkEvent->park(1);
950 } else {
951 os::naked_yield();
952 }
953 }
954 gInflationLocks[ix]->unlock();
955 }
956 } else {
957 SpinPause(); // SMP-polite spinning
958 }
959 }
960 }
961
962 // hashCode() generation :
963 //
964 // Possibilities:
965 // * MD5Digest of {obj,stw_random}
966 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function.
967 // * A DES- or AES-style SBox[] mechanism
968 // * One of the Phi-based schemes, such as:
969 // 2654435761 = 2^32 * Phi (golden ratio)
970 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ;
971 // * A variation of Marsaglia's shift-xor RNG scheme.
972 // * (obj ^ stw_random) is appealing, but can result
973 // in undesirable regularity in the hashCode values of adjacent objects
974 // (objects allocated back-to-back, in particular). This could potentially
975 // result in hashtable collisions and reduced hashtable efficiency.
976 // There are simple ways to "diffuse" the middle address bits over the
977 // generated hashCode values:
978
979 static inline intptr_t get_next_hash(Thread* current, oop obj) {
980 intptr_t value = 0;
981 if (hashCode == 0) {
982 // This form uses global Park-Miller RNG.
983 // On MP system we'll have lots of RW access to a global, so the
984 // mechanism induces lots of coherency traffic.
985 value = os::random();
986 } else if (hashCode == 1) {
987 // This variation has the property of being stable (idempotent)
988 // between STW operations. This can be useful in some of the 1-0
989 // synchronization schemes.
990 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3;
991 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random;
992 } else if (hashCode == 2) {
993 value = 1; // for sensitivity testing
994 } else if (hashCode == 3) {
995 value = ++GVars.hc_sequence;
996 } else if (hashCode == 4) {
997 value = cast_from_oop<intptr_t>(obj);
998 } else {
999 // Marsaglia's xor-shift scheme with thread-specific state
1000 // This is probably the best overall implementation -- we'll
1001 // likely make this the default in future releases.
1002 unsigned t = current->_hashStateX;
1003 t ^= (t << 11);
1004 current->_hashStateX = current->_hashStateY;
1005 current->_hashStateY = current->_hashStateZ;
1006 current->_hashStateZ = current->_hashStateW;
1007 unsigned v = current->_hashStateW;
1008 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
1009 current->_hashStateW = v;
1010 value = v;
1011 }
1012
1013 value &= UseCompactObjectHeaders ? markWord::hash_mask_compact : markWord::hash_mask;
1014 if (value == 0) value = 0xBAD;
1015 assert(value != markWord::no_hash, "invariant");
1016 return value;
1017 }
1018
1019 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) {
1020 if (UseBiasedLocking) {
1021 // NOTE: many places throughout the JVM do not expect a safepoint
1022 // to be taken here. However, we only ever bias Java instances and all
1023 // of the call sites of identity_hash that might revoke biases have
1024 // been checked to make sure they can handle a safepoint. The
1025 // added check of the bias pattern is to avoid useless calls to
1026 // thread-local storage.
1027 if (obj->mark().has_bias_pattern()) {
1028 // Handle for oop obj in case of STW safepoint
1029 Handle hobj(current, obj);
1030 if (SafepointSynchronize::is_at_safepoint()) {
1031 BiasedLocking::revoke_at_safepoint(hobj);
1032 } else {
1033 BiasedLocking::revoke(current->as_Java_thread(), hobj);
1034 }
1035 obj = hobj();
1036 assert(!obj->mark().has_bias_pattern(), "biases should be revoked by now");
1037 }
1038 }
1039
1040 while (true) {
1041 ObjectMonitor* monitor = NULL;
1042 markWord temp, test;
1043 intptr_t hash;
1044 markWord mark = read_stable_mark(obj);
1045
1046 // object should remain ineligible for biased locking
1047 assert(!mark.has_bias_pattern(), "invariant");
1048
1049 if (mark.is_neutral() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) {
1050 hash = mark.hash();
1051 if (hash != 0) { // if it has a hash, just return it
1052 return hash;
1053 }
1054 hash = get_next_hash(current, obj); // get a new hash
1055 temp = mark.copy_set_hash(hash); // merge the hash into header
1056 // try to install the hash
1057 test = obj->cas_set_mark(temp, mark);
1058 if (test == mark) { // if the hash was installed, return it
1059 return hash;
1060 }
1061 if (LockingMode == LM_LIGHTWEIGHT) {
1062 // CAS failed, retry
1063 continue;
1064 }
1065 // Failed to install the hash. It could be that another thread
1066 // installed the hash just before our attempt or inflation has
1067 // occurred or... so we fall thru to inflate the monitor for
1068 // stability and then install the hash.
1069 } else if (mark.has_monitor()) {
1070 monitor = mark.monitor();
1071 temp = monitor->header();
1072 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1073 hash = temp.hash();
1074 if (hash != 0) {
1075 // It has a hash.
1076
1077 // Separate load of dmw/header above from the loads in
1078 // is_being_async_deflated().
1079
1080 // dmw/header and _contentions may get written by different threads.
1081 // Make sure to observe them in the same order when having several observers.
1082 OrderAccess::loadload_for_IRIW();
1083
1084 if (monitor->is_being_async_deflated()) {
1085 // But we can't safely use the hash if we detect that async
1086 // deflation has occurred. So we attempt to restore the
1087 // header/dmw to the object's header so that we only retry
1088 // once if the deflater thread happens to be slow.
1089 monitor->install_displaced_markword_in_object(obj);
1090 continue;
1091 }
1092 return hash;
1093 }
1094 // Fall thru so we only have one place that installs the hash in
1095 // the ObjectMonitor.
1096 } else if (LockingMode == LM_LEGACY && mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
1097 // This is a stack lock owned by the calling thread so fetch the
1098 // displaced markWord from the BasicLock on the stack.
1099 temp = mark.displaced_mark_helper();
1100 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1101 hash = temp.hash();
1102 if (hash != 0) { // if it has a hash, just return it
1103 return hash;
1104 }
1105 // WARNING:
1106 // The displaced header in the BasicLock on a thread's stack
1107 // is strictly immutable. It CANNOT be changed in ANY cases.
1108 // So we have to inflate the stack lock into an ObjectMonitor
1109 // even if the current thread owns the lock. The BasicLock on
1110 // a thread's stack can be asynchronously read by other threads
1111 // during an inflate() call so any change to that stack memory
1112 // may not propagate to other threads correctly.
1113 }
1114
1115 // Inflate the monitor to set the hash.
1116
1117 // An async deflation can race after the inflate() call and before we
1118 // can update the ObjectMonitor's header with the hash value below.
1119 monitor = inflate(current, obj, inflate_cause_hash_code);
1120 // Load ObjectMonitor's header/dmw field and see if it has a hash.
1121 mark = monitor->header();
1122 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
1123 hash = mark.hash();
1124 if (hash == 0) { // if it does not have a hash
1125 hash = get_next_hash(current, obj); // get a new hash
1126 temp = mark.copy_set_hash(hash) ; // merge the hash into header
1127 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1128 uintptr_t v = Atomic::cmpxchg((volatile uintptr_t*)monitor->header_addr(), mark.value(), temp.value());
1129 test = markWord(v);
1130 if (test != mark) {
1131 // The attempt to update the ObjectMonitor's header/dmw field
1132 // did not work. This can happen if another thread managed to
1133 // merge in the hash just before our cmpxchg().
1134 // If we add any new usages of the header/dmw field, this code
1135 // will need to be updated.
1136 hash = test.hash();
1137 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value());
1138 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash");
1139 }
1140 if (monitor->is_being_async_deflated()) {
1141 // If we detect that async deflation has occurred, then we
1142 // attempt to restore the header/dmw to the object's header
1143 // so that we only retry once if the deflater thread happens
1144 // to be slow.
1145 monitor->install_displaced_markword_in_object(obj);
1146 continue;
1147 }
1148 }
1149 // We finally get the hash.
1150 return hash;
1151 }
1152 }
1153
1154 // Deprecated -- use FastHashCode() instead.
1155
1156 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
1157 return FastHashCode(Thread::current(), obj());
1158 }
1159
1160
1161 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current,
1162 Handle h_obj) {
1163 if (UseBiasedLocking) {
1164 BiasedLocking::revoke(current, h_obj);
1165 assert(!h_obj->mark().has_bias_pattern(), "biases should be revoked by now");
1166 }
1167
1168 assert(current == JavaThread::current(), "Can only be called on current thread");
1169 oop obj = h_obj();
1170
1171 markWord mark = read_stable_mark(obj);
1172
1173 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1174 // stack-locked case, header points into owner's stack
1175 return current->is_lock_owned((address)mark.locker());
1176 }
1177
1178 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1179 // fast-locking case, see if lock is in current's lock stack
1180 return current->lock_stack().contains(h_obj());
1181 }
1182
1183 // Contended case, header points to ObjectMonitor (tagged pointer)
1184 if (mark.has_monitor()) {
1185 // The first stage of async deflation does not affect any field
1186 // used by this comparison so the ObjectMonitor* is usable here.
1187 ObjectMonitor* monitor = mark.monitor();
1188 return monitor->is_entered(current) != 0;
1189 }
1190 // Unlocked case, header in place
1191 assert(mark.is_neutral(), "sanity check");
1192 return false;
1193 }
1194
1195 // FIXME: jvmti should call this
1196 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
1197 if (UseBiasedLocking) {
1198 if (SafepointSynchronize::is_at_safepoint()) {
1199 BiasedLocking::revoke_at_safepoint(h_obj);
1200 } else {
1201 BiasedLocking::revoke(JavaThread::current(), h_obj);
1202 }
1203 assert(!h_obj->mark().has_bias_pattern(), "biases should be revoked by now");
1204 }
1205
1206 oop obj = h_obj();
1207 markWord mark = read_stable_mark(obj);
1208
1209 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1210 // stack-locked so header points into owner's stack.
1211 // owning_thread_from_monitor_owner() may also return null here:
1212 return Threads::owning_thread_from_monitor_owner(t_list, (address) mark.locker());
1213 }
1214
1215 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1216 // fast-locked so get owner from the object.
1217 // owning_thread_from_object() may also return null here:
1218 return Threads::owning_thread_from_object(t_list, h_obj());
1219 }
1220
1221 // Contended case, header points to ObjectMonitor (tagged pointer)
1222 if (mark.has_monitor()) {
1223 // The first stage of async deflation does not affect any field
1224 // used by this comparison so the ObjectMonitor* is usable here.
1225 ObjectMonitor* monitor = mark.monitor();
1226 assert(monitor != NULL, "monitor should be non-null");
1227 // owning_thread_from_monitor() may also return null here:
1228 return Threads::owning_thread_from_monitor(t_list, monitor);
1229 }
1230
1231 // Unlocked case, header in place
1232 // Cannot have assertion since this object may have been
1233 // locked by another thread when reaching here.
1234 // assert(mark.is_neutral(), "sanity check");
1235
1236 return NULL;
1237 }
1238
1239 // Visitors ...
1240
1241 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure, JavaThread* thread) {
1242 MonitorList::Iterator iter = _in_use_list.iterator();
1243 while (iter.has_next()) {
1244 ObjectMonitor* mid = iter.next();
1245 if (mid->owner() != thread) {
1246 continue;
1247 }
1248 if (!mid->is_being_async_deflated() && mid->object_peek() != NULL) {
1249 // Only process with closure if the object is set.
1250
1251 // monitors_iterate() is only called at a safepoint or when the
1252 // target thread is suspended or when the target thread is
1253 // operating on itself. The current closures in use today are
1254 // only interested in an owned ObjectMonitor and ownership
1255 // cannot be dropped under the calling contexts so the
1256 // ObjectMonitor cannot be async deflated.
1257 closure->do_monitor(mid);
1258 }
1259 }
1260 }
1261
1262 static bool monitors_used_above_threshold(MonitorList* list) {
1263 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy
1264 return false;
1265 }
1266 // Start with ceiling based on a per-thread estimate:
1267 size_t ceiling = ObjectSynchronizer::in_use_list_ceiling();
1268 size_t old_ceiling = ceiling;
1269 if (ceiling < list->max()) {
1270 // The max used by the system has exceeded the ceiling so use that:
1271 ceiling = list->max();
1272 }
1273 size_t monitors_used = list->count();
1274 if (monitors_used == 0) { // empty list is easy
1275 return false;
1276 }
1277 if (NoAsyncDeflationProgressMax != 0 &&
1278 _no_progress_cnt >= NoAsyncDeflationProgressMax) {
1279 float remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0;
1280 size_t new_ceiling = ceiling + (ceiling * remainder) + 1;
1281 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling);
1282 log_info(monitorinflation)("Too many deflations without progress; "
1283 "bumping in_use_list_ceiling from " SIZE_FORMAT
1284 " to " SIZE_FORMAT, old_ceiling, new_ceiling);
1285 _no_progress_cnt = 0;
1286 ceiling = new_ceiling;
1287 }
1288
1289 // Check if our monitor usage is above the threshold:
1290 size_t monitor_usage = (monitors_used * 100LL) / ceiling;
1291 if (int(monitor_usage) > MonitorUsedDeflationThreshold) {
1292 log_info(monitorinflation)("monitors_used=" SIZE_FORMAT ", ceiling=" SIZE_FORMAT
1293 ", monitor_usage=" SIZE_FORMAT ", threshold=" INTX_FORMAT,
1294 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold);
1295 return true;
1296 }
1297
1298 return false;
1299 }
1300
1301 size_t ObjectSynchronizer::in_use_list_ceiling() {
1302 return _in_use_list_ceiling;
1303 }
1304
1305 void ObjectSynchronizer::dec_in_use_list_ceiling() {
1306 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1307 }
1308
1309 void ObjectSynchronizer::inc_in_use_list_ceiling() {
1310 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1311 }
1312
1313 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) {
1314 _in_use_list_ceiling = new_value;
1315 }
1316
1317 bool ObjectSynchronizer::is_async_deflation_needed() {
1318 if (is_async_deflation_requested()) {
1319 // Async deflation request.
1320 log_info(monitorinflation)("Async deflation needed: explicit request");
1321 return true;
1322 }
1323
1324 jlong time_since_last = time_since_last_async_deflation_ms();
1325
1326 if (AsyncDeflationInterval > 0 &&
1327 time_since_last > AsyncDeflationInterval &&
1328 monitors_used_above_threshold(&_in_use_list)) {
1329 // It's been longer than our specified deflate interval and there
1330 // are too many monitors in use. We don't deflate more frequently
1331 // than AsyncDeflationInterval (unless is_async_deflation_requested)
1332 // in order to not swamp the MonitorDeflationThread.
1333 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold");
1334 return true;
1335 }
1336
1337 if (GuaranteedAsyncDeflationInterval > 0 &&
1338 time_since_last > GuaranteedAsyncDeflationInterval) {
1339 // It's been longer than our specified guaranteed deflate interval.
1340 // We need to clean up the used monitors even if the threshold is
1341 // not reached, to keep the memory utilization at bay when many threads
1342 // touched many monitors.
1343 log_info(monitorinflation)("Async deflation needed: guaranteed interval (" INTX_FORMAT " ms) "
1344 "is greater than time since last deflation (" JLONG_FORMAT " ms)",
1345 GuaranteedAsyncDeflationInterval, time_since_last);
1346
1347 // If this deflation has no progress, then it should not affect the no-progress
1348 // tracking, otherwise threshold heuristics would think it was triggered, experienced
1349 // no progress, and needs to backoff more aggressively. In this "no progress" case,
1350 // the generic code would bump the no-progress counter, and we compensate for that
1351 // by telling it to skip the update.
1352 //
1353 // If this deflation has progress, then it should let non-progress tracking
1354 // know about this, otherwise the threshold heuristics would kick in, potentially
1355 // experience no-progress due to aggressive cleanup by this deflation, and think
1356 // it is still in no-progress stride. In this "progress" case, the generic code would
1357 // zero the counter, and we allow it to happen.
1358 _no_progress_skip_increment = true;
1359
1360 return true;
1361 }
1362
1363 return false;
1364 }
1365
1366 bool ObjectSynchronizer::request_deflate_idle_monitors() {
1367 JavaThread* current = JavaThread::current();
1368 bool ret_code = false;
1369
1370 jlong last_time = last_async_deflation_time_ns();
1371 set_is_async_deflation_requested(true);
1372 {
1373 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag);
1374 ml.notify_all();
1375 }
1376 const int N_CHECKS = 5;
1377 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds
1378 if (last_async_deflation_time_ns() > last_time) {
1379 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i);
1380 ret_code = true;
1381 break;
1382 }
1383 {
1384 // JavaThread has to honor the blocking protocol.
1385 ThreadBlockInVM tbivm(current);
1386 os::naked_short_sleep(999); // sleep for almost 1 second
1387 }
1388 }
1389 if (!ret_code) {
1390 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS);
1391 }
1392
1393 return ret_code;
1394 }
1395
1396 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() {
1397 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS);
1398 }
1399
1400 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1401 const oop obj,
1402 ObjectSynchronizer::InflateCause cause) {
1403 assert(event != NULL, "invariant");
1404 event->set_monitorClass(obj->klass());
1405 event->set_address((uintptr_t)(void*)obj);
1406 event->set_cause((u1)cause);
1407 event->commit();
1408 }
1409
1410 // Fast path code shared by multiple functions
1411 void ObjectSynchronizer::inflate_helper(oop obj) {
1412 markWord mark = obj->mark_acquire();
1413 if (mark.has_monitor()) {
1414 ObjectMonitor* monitor = mark.monitor();
1415 markWord dmw = monitor->header();
1416 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value());
1417 return;
1418 }
1419 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal);
1420 }
1421
1422 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) {
1423 assert(current == Thread::current(), "must be");
1424 if (LockingMode == LM_LIGHTWEIGHT && current->is_Java_thread()) {
1425 return inflate_impl(current->as_Java_thread(), obj, cause);
1426 }
1427 return inflate_impl(nullptr, obj, cause);
1428 }
1429
1430 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) {
1431 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be");
1432 return inflate_impl(thread, obj, cause);
1433 }
1434
1435 ObjectMonitor* ObjectSynchronizer::inflate_impl(JavaThread* inflating_thread, oop object, const InflateCause cause) {
1436 // The JavaThread* inflating_thread parameter is only used by LM_LIGHTWEIGHT and requires
1437 // that the inflating_thread == Thread::current() or is suspended throughout the call by
1438 // some other mechanism.
1439 // Even with LM_LIGHTWEIGHT the thread might be nullptr when called from a non
1440 // JavaThread. (As may still be the case from FastHashCode). However it is only
1441 // important for the correctness of the LM_LIGHTWEIGHT algorithm that the thread
1442 // is set when called from ObjectSynchronizer::enter from the owning thread,
1443 // ObjectSynchronizer::enter_for from any thread, or ObjectSynchronizer::exit.
1444 EventJavaMonitorInflate event;
1445
1446 for (;;) {
1447 const markWord mark = object->mark_acquire();
1448 assert(!mark.has_bias_pattern(), "invariant");
1449
1450 // The mark can be in one of the following states:
1451 // * inflated - Just return if using stack-locking.
1452 // If using fast-locking and the ObjectMonitor owner
1453 // is anonymous and the inflating_thread owns the
1454 // object lock, then we make the inflating_thread
1455 // the ObjectMonitor owner and remove the lock from
1456 // the inflating_thread's lock stack.
1457 // * fast-locked - Coerce it to inflated from fast-locked.
1458 // * stack-locked - Coerce it to inflated from stack-locked.
1459 // * INFLATING - busy wait for conversion to complete
1460 // * Neutral - aggressively inflate the object.
1461 // * BIASED - Illegal. We should never see this
1462
1463 // CASE: inflated
1464 if (mark.has_monitor()) {
1465 ObjectMonitor* inf = mark.monitor();
1466 markWord dmw = inf->header();
1467 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1468 if (LockingMode == LM_LIGHTWEIGHT && inf->is_owner_anonymous() &&
1469 inflating_thread != nullptr && inflating_thread->lock_stack().contains(object)) {
1470 inf->set_owner_from_anonymous(inflating_thread);
1471 size_t removed = inflating_thread->lock_stack().remove(object);
1472 inf->set_recursions(removed - 1);
1473 }
1474 return inf;
1475 }
1476
1477 // CASE: inflation in progress - inflating over a stack-lock.
1478 // Some other thread is converting from stack-locked to inflated.
1479 // Only that thread can complete inflation -- other threads must wait.
1480 // The INFLATING value is transient.
1481 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1482 // We could always eliminate polling by parking the thread on some auxiliary list.
1483 if (LockingMode != LM_LIGHTWEIGHT) {
1484 // New lightweight locking does not use INFLATING.
1485 // CASE: inflation in progress - inflating over a stack-lock.
1486 // Some other thread is converting from stack-locked to inflated.
1487 // Only that thread can complete inflation -- other threads must wait.
1488 // The INFLATING value is transient.
1489 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1490 // We could always eliminate polling by parking the thread on some auxiliary list.
1491 if (mark == markWord::INFLATING()) {
1492 read_stable_mark(object);
1493 continue;
1494 }
1495 }
1496
1497 // CASE: fast-locked
1498 // Could be fast-locked either by the inflating_thread or by some other thread.
1499 //
1500 // Note that we allocate the ObjectMonitor speculatively, _before_
1501 // attempting to set the object's mark to the new ObjectMonitor. If
1502 // the inflating_thread owns the monitor, then we set the ObjectMonitor's
1503 // owner to the inflating_thread. Otherwise, we set the ObjectMonitor's owner
1504 // to anonymous. If we lose the race to set the object's mark to the
1505 // new ObjectMonitor, then we just delete it and loop around again.
1506 //
1507 LogStreamHandle(Trace, monitorinflation) lsh;
1508 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1509 ObjectMonitor* monitor = new ObjectMonitor(object);
1510 monitor->set_header(mark.set_unlocked());
1511 bool own = inflating_thread != nullptr && inflating_thread->lock_stack().contains(object);
1512 if (own) {
1513 // Owned by us.
1514 monitor->set_owner_from(NULL, inflating_thread);
1515 } else {
1516 // Owned by somebody else.
1517 monitor->set_owner_anonymous();
1518 }
1519 markWord monitor_mark = markWord::encode(monitor);
1520 markWord old_mark = object->cas_set_mark(monitor_mark, mark);
1521 if (old_mark == mark) {
1522 // Success! Return inflated monitor.
1523 if (own) {
1524 size_t removed = inflating_thread->lock_stack().remove(object);
1525 monitor->set_recursions(removed - 1);
1526 }
1527 // Once the ObjectMonitor is configured and object is associated
1528 // with the ObjectMonitor, it is safe to allow async deflation:
1529 _in_use_list.add(monitor);
1530
1531 // Hopefully the performance counters are allocated on distinct
1532 // cache lines to avoid false sharing on MP systems ...
1533 OM_PERFDATA_OP(Inflations, inc());
1534 if (log_is_enabled(Trace, monitorinflation)) {
1535 ResourceMark rm;
1536 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1537 INTPTR_FORMAT ", type='%s'", p2i(object),
1538 object->mark().value(), object->klass()->external_name());
1539 }
1540 if (event.should_commit()) {
1541 post_monitor_inflate_event(&event, object, cause);
1542 }
1543 return monitor;
1544 } else {
1545 delete monitor;
1546 continue; // Interference -- just retry
1547 }
1548 }
1549
1550 // CASE: stack-locked
1551 // Could be stack-locked either by this thread or by some other thread.
1552 //
1553 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting
1554 // to install INFLATING into the mark word. We originally installed INFLATING,
1555 // allocated the ObjectMonitor, and then finally STed the address of the
1556 // ObjectMonitor into the mark. This was correct, but artificially lengthened
1557 // the interval in which INFLATING appeared in the mark, thus increasing
1558 // the odds of inflation contention.
1559
1560 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1561 assert(LockingMode != LM_LIGHTWEIGHT, "cannot happen with new lightweight locking");
1562 ObjectMonitor* m = new ObjectMonitor(object);
1563 // Optimistically prepare the ObjectMonitor - anticipate successful CAS
1564 // We do this before the CAS in order to minimize the length of time
1565 // in which INFLATING appears in the mark.
1566
1567 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark);
1568 if (cmp != mark) {
1569 delete m;
1570 continue; // Interference -- just retry
1571 }
1572
1573 // We've successfully installed INFLATING (0) into the mark-word.
1574 // This is the only case where 0 will appear in a mark-word.
1575 // Only the singular thread that successfully swings the mark-word
1576 // to 0 can perform (or more precisely, complete) inflation.
1577 //
1578 // Why do we CAS a 0 into the mark-word instead of just CASing the
1579 // mark-word from the stack-locked value directly to the new inflated state?
1580 // Consider what happens when a thread unlocks a stack-locked object.
1581 // It attempts to use CAS to swing the displaced header value from the
1582 // on-stack BasicLock back into the object header. Recall also that the
1583 // header value (hash code, etc) can reside in (a) the object header, or
1584 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1585 // header in an ObjectMonitor. The inflate() routine must copy the header
1586 // value from the BasicLock on the owner's stack to the ObjectMonitor, all
1587 // the while preserving the hashCode stability invariants. If the owner
1588 // decides to release the lock while the value is 0, the unlock will fail
1589 // and control will eventually pass from slow_exit() to inflate. The owner
1590 // will then spin, waiting for the 0 value to disappear. Put another way,
1591 // the 0 causes the owner to stall if the owner happens to try to
1592 // drop the lock (restoring the header from the BasicLock to the object)
1593 // while inflation is in-progress. This protocol avoids races that might
1594 // would otherwise permit hashCode values to change or "flicker" for an object.
1595 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable.
1596 // 0 serves as a "BUSY" inflate-in-progress indicator.
1597
1598
1599 // fetch the displaced mark from the owner's stack.
1600 // The owner can't die or unwind past the lock while our INFLATING
1601 // object is in the mark. Furthermore the owner can't complete
1602 // an unlock on the object, either.
1603 markWord dmw = mark.displaced_mark_helper();
1604 // Catch if the object's header is not neutral (not locked and
1605 // not marked is what we care about here).
1606 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1607
1608 // Setup monitor fields to proper values -- prepare the monitor
1609 m->set_header(dmw);
1610
1611 // Optimization: if the mark.locker stack address is associated
1612 // with this thread we could simply set m->_owner = current.
1613 // Note that a thread can inflate an object
1614 // that it has stack-locked -- as might happen in wait() -- directly
1615 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1616 m->set_owner_from(NULL, mark.locker());
1617 // TODO-FIXME: assert BasicLock->dhw != 0.
1618
1619 // Must preserve store ordering. The monitor state must
1620 // be stable at the time of publishing the monitor address.
1621 guarantee(object->mark() == markWord::INFLATING(), "invariant");
1622 // Release semantics so that above set_object() is seen first.
1623 object->release_set_mark(markWord::encode(m));
1624
1625 // Once ObjectMonitor is configured and the object is associated
1626 // with the ObjectMonitor, it is safe to allow async deflation:
1627 _in_use_list.add(m);
1628
1629 // Hopefully the performance counters are allocated on distinct cache lines
1630 // to avoid false sharing on MP systems ...
1631 OM_PERFDATA_OP(Inflations, inc());
1632 if (log_is_enabled(Trace, monitorinflation)) {
1633 ResourceMark rm;
1634 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1635 INTPTR_FORMAT ", type='%s'", p2i(object),
1636 object->mark().value(), object->klass()->external_name());
1637 }
1638 if (event.should_commit()) {
1639 post_monitor_inflate_event(&event, object, cause);
1640 }
1641 return m;
1642 }
1643
1644 // CASE: neutral
1645 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1646 // If we know we're inflating for entry it's better to inflate by swinging a
1647 // pre-locked ObjectMonitor pointer into the object header. A successful
1648 // CAS inflates the object *and* confers ownership to the inflating thread.
1649 // In the current implementation we use a 2-step mechanism where we CAS()
1650 // to inflate and then CAS() again to try to swing _owner from NULL to current.
1651 // An inflateTry() method that we could call from enter() would be useful.
1652
1653 // Catch if the object's header is not neutral (not locked and
1654 // not marked is what we care about here).
1655 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
1656 ObjectMonitor* m = new ObjectMonitor(object);
1657 // prepare m for installation - set monitor to initial state
1658 m->set_header(mark);
1659
1660 if (object->cas_set_mark(markWord::encode(m), mark) != mark) {
1661 delete m;
1662 m = NULL;
1663 continue;
1664 // interference - the markword changed - just retry.
1665 // The state-transitions are one-way, so there's no chance of
1666 // live-lock -- "Inflated" is an absorbing state.
1667 }
1668
1669 // Once the ObjectMonitor is configured and object is associated
1670 // with the ObjectMonitor, it is safe to allow async deflation:
1671 _in_use_list.add(m);
1672
1673 // Hopefully the performance counters are allocated on distinct
1674 // cache lines to avoid false sharing on MP systems ...
1675 OM_PERFDATA_OP(Inflations, inc());
1676 if (log_is_enabled(Trace, monitorinflation)) {
1677 ResourceMark rm;
1678 lsh.print_cr("inflate(neutral): object=" INTPTR_FORMAT ", mark="
1679 INTPTR_FORMAT ", type='%s'", p2i(object),
1680 object->mark().value(), object->klass()->external_name());
1681 }
1682 if (event.should_commit()) {
1683 post_monitor_inflate_event(&event, object, cause);
1684 }
1685 return m;
1686 }
1687 }
1688
1689 void ObjectSynchronizer::chk_for_block_req(JavaThread* current, const char* op_name,
1690 const char* cnt_name, size_t cnt,
1691 LogStream* ls, elapsedTimer* timer_p) {
1692 if (!SafepointMechanism::should_process(current)) {
1693 return;
1694 }
1695
1696 // A safepoint/handshake has started.
1697 if (ls != NULL) {
1698 timer_p->stop();
1699 ls->print_cr("pausing %s: %s=" SIZE_FORMAT ", in_use_list stats: ceiling="
1700 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1701 op_name, cnt_name, cnt, in_use_list_ceiling(),
1702 _in_use_list.count(), _in_use_list.max());
1703 }
1704
1705 {
1706 // Honor block request.
1707 ThreadBlockInVM tbivm(current);
1708 }
1709
1710 if (ls != NULL) {
1711 ls->print_cr("resuming %s: in_use_list stats: ceiling=" SIZE_FORMAT
1712 ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, op_name,
1713 in_use_list_ceiling(), _in_use_list.count(), _in_use_list.max());
1714 timer_p->start();
1715 }
1716 }
1717
1718 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle
1719 // ObjectMonitors. Returns the number of deflated ObjectMonitors.
1720 size_t ObjectSynchronizer::deflate_monitor_list(Thread* current, LogStream* ls,
1721 elapsedTimer* timer_p) {
1722 MonitorList::Iterator iter = _in_use_list.iterator();
1723 size_t deflated_count = 0;
1724
1725 while (iter.has_next()) {
1726 if (deflated_count >= (size_t)MonitorDeflationMax) {
1727 break;
1728 }
1729 ObjectMonitor* mid = iter.next();
1730 if (mid->deflate_monitor()) {
1731 deflated_count++;
1732 }
1733
1734 if (current->is_Java_thread()) {
1735 // A JavaThread must check for a safepoint/handshake and honor it.
1736 chk_for_block_req(current->as_Java_thread(), "deflation", "deflated_count",
1737 deflated_count, ls, timer_p);
1738 }
1739 }
1740
1741 return deflated_count;
1742 }
1743
1744 class HandshakeForDeflation : public HandshakeClosure {
1745 public:
1746 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {}
1747
1748 void do_thread(Thread* thread) {
1749 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread="
1750 INTPTR_FORMAT, p2i(thread));
1751 }
1752 };
1753
1754 class VM_RendezvousGCThreads : public VM_Operation {
1755 public:
1756 bool evaluate_at_safepoint() const override { return false; }
1757 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; }
1758 void doit() override {
1759 SuspendibleThreadSet::synchronize();
1760 SuspendibleThreadSet::desynchronize();
1761 };
1762 };
1763
1764 // This function is called by the MonitorDeflationThread to deflate
1765 // ObjectMonitors. It is also called via do_final_audit_and_print_stats()
1766 // by the VMThread.
1767 size_t ObjectSynchronizer::deflate_idle_monitors() {
1768 Thread* current = Thread::current();
1769 if (current->is_Java_thread()) {
1770 // The async deflation request has been processed.
1771 _last_async_deflation_time_ns = os::javaTimeNanos();
1772 set_is_async_deflation_requested(false);
1773 }
1774
1775 LogStreamHandle(Debug, monitorinflation) lsh_debug;
1776 LogStreamHandle(Info, monitorinflation) lsh_info;
1777 LogStream* ls = NULL;
1778 if (log_is_enabled(Debug, monitorinflation)) {
1779 ls = &lsh_debug;
1780 } else if (log_is_enabled(Info, monitorinflation)) {
1781 ls = &lsh_info;
1782 }
1783
1784 elapsedTimer timer;
1785 if (ls != NULL) {
1786 ls->print_cr("begin deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1787 in_use_list_ceiling(), _in_use_list.count(), _in_use_list.max());
1788 timer.start();
1789 }
1790
1791 // Deflate some idle ObjectMonitors.
1792 size_t deflated_count = deflate_monitor_list(current, ls, &timer);
1793 if (deflated_count > 0 || is_final_audit()) {
1794 // There are ObjectMonitors that have been deflated or this is the
1795 // final audit and all the remaining ObjectMonitors have been
1796 // deflated, BUT the MonitorDeflationThread blocked for the final
1797 // safepoint during unlinking.
1798
1799 // Unlink deflated ObjectMonitors from the in-use list.
1800 ResourceMark rm;
1801 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count);
1802 size_t unlinked_count = _in_use_list.unlink_deflated(current, ls, &timer,
1803 &delete_list);
1804 if (current->is_Java_thread()) {
1805 if (ls != NULL) {
1806 timer.stop();
1807 ls->print_cr("before handshaking: unlinked_count=" SIZE_FORMAT
1808 ", in_use_list stats: ceiling=" SIZE_FORMAT ", count="
1809 SIZE_FORMAT ", max=" SIZE_FORMAT,
1810 unlinked_count, in_use_list_ceiling(),
1811 _in_use_list.count(), _in_use_list.max());
1812 }
1813
1814 // A JavaThread needs to handshake in order to safely free the
1815 // ObjectMonitors that were deflated in this cycle.
1816 // Also, we sync and desync GC threads around the handshake, so that they can
1817 // safely read the mark-word and look-through to the object-monitor, without
1818 // being afraid that the object-monitor is going away.
1819 HandshakeForDeflation hfd_hc;
1820 Handshake::execute(&hfd_hc);
1821 VM_RendezvousGCThreads sync_gc;
1822 VMThread::execute(&sync_gc);
1823
1824 if (ls != NULL) {
1825 ls->print_cr("after handshaking: in_use_list stats: ceiling="
1826 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1827 in_use_list_ceiling(), _in_use_list.count(), _in_use_list.max());
1828 timer.start();
1829 }
1830 }
1831
1832 NativeHeapTrimmer::SuspendMark sm("monitor deletion");
1833
1834 // After the handshake, safely free the ObjectMonitors that were
1835 // deflated in this cycle.
1836 size_t deleted_count = 0;
1837 for (ObjectMonitor* monitor: delete_list) {
1838 delete monitor;
1839 deleted_count++;
1840
1841 if (current->is_Java_thread()) {
1842 // A JavaThread must check for a safepoint/handshake and honor it.
1843 chk_for_block_req(current->as_Java_thread(), "deletion", "deleted_count",
1844 deleted_count, ls, &timer);
1845 }
1846 }
1847 }
1848
1849 if (ls != NULL) {
1850 timer.stop();
1851 if (deflated_count != 0 || log_is_enabled(Debug, monitorinflation)) {
1852 ls->print_cr("deflated " SIZE_FORMAT " monitors in %3.7f secs",
1853 deflated_count, timer.seconds());
1854 }
1855 ls->print_cr("end deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1856 in_use_list_ceiling(), _in_use_list.count(), _in_use_list.max());
1857 }
1858
1859 OM_PERFDATA_OP(MonExtant, set_value(_in_use_list.count()));
1860 OM_PERFDATA_OP(Deflations, inc(deflated_count));
1861
1862 GVars.stw_random = os::random();
1863
1864 if (deflated_count != 0) {
1865 _no_progress_cnt = 0;
1866 } else if (_no_progress_skip_increment) {
1867 _no_progress_skip_increment = false;
1868 } else {
1869 _no_progress_cnt++;
1870 }
1871
1872 return deflated_count;
1873 }
1874
1875 // Monitor cleanup on JavaThread::exit
1876
1877 // Iterate through monitor cache and attempt to release thread's monitors
1878 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1879 private:
1880 JavaThread* _thread;
1881
1882 public:
1883 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {}
1884 void do_monitor(ObjectMonitor* mid) {
1885 (void)mid->complete_exit(_thread);
1886 }
1887 };
1888
1889 // Release all inflated monitors owned by current thread. Lightweight monitors are
1890 // ignored. This is meant to be called during JNI thread detach which assumes
1891 // all remaining monitors are heavyweight. All exceptions are swallowed.
1892 // Scanning the extant monitor list can be time consuming.
1893 // A simple optimization is to add a per-thread flag that indicates a thread
1894 // called jni_monitorenter() during its lifetime.
1895 //
1896 // Instead of NoSafepointVerifier it might be cheaper to
1897 // use an idiom of the form:
1898 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1899 // <code that must not run at safepoint>
1900 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1901 // Since the tests are extremely cheap we could leave them enabled
1902 // for normal product builds.
1903
1904 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) {
1905 assert(current == JavaThread::current(), "must be current Java thread");
1906 NoSafepointVerifier nsv;
1907 ReleaseJavaMonitorsClosure rjmc(current);
1908 ObjectSynchronizer::monitors_iterate(&rjmc, current);
1909 assert(!current->has_pending_exception(), "Should not be possible");
1910 current->clear_pending_exception();
1911 }
1912
1913 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1914 switch (cause) {
1915 case inflate_cause_vm_internal: return "VM Internal";
1916 case inflate_cause_monitor_enter: return "Monitor Enter";
1917 case inflate_cause_wait: return "Monitor Wait";
1918 case inflate_cause_notify: return "Monitor Notify";
1919 case inflate_cause_hash_code: return "Monitor Hash Code";
1920 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1921 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1922 default:
1923 ShouldNotReachHere();
1924 }
1925 return "Unknown";
1926 }
1927
1928 //------------------------------------------------------------------------------
1929 // Debugging code
1930
1931 u_char* ObjectSynchronizer::get_gvars_addr() {
1932 return (u_char*)&GVars;
1933 }
1934
1935 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() {
1936 return (u_char*)&GVars.hc_sequence;
1937 }
1938
1939 size_t ObjectSynchronizer::get_gvars_size() {
1940 return sizeof(SharedGlobals);
1941 }
1942
1943 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() {
1944 return (u_char*)&GVars.stw_random;
1945 }
1946
1947 // Do the final audit and print of ObjectMonitor stats; must be done
1948 // by the VMThread at VM exit time.
1949 void ObjectSynchronizer::do_final_audit_and_print_stats() {
1950 assert(Thread::current()->is_VM_thread(), "sanity check");
1951
1952 if (is_final_audit()) { // Only do the audit once.
1953 return;
1954 }
1955 set_is_final_audit();
1956 log_info(monitorinflation)("Starting the final audit.");
1957
1958 if (log_is_enabled(Info, monitorinflation)) {
1959 // Do a deflation in order to reduce the in-use monitor population
1960 // that is reported by ObjectSynchronizer::log_in_use_monitor_details()
1961 // which is called by ObjectSynchronizer::audit_and_print_stats().
1962 while (ObjectSynchronizer::deflate_idle_monitors() != 0) {
1963 ; // empty
1964 }
1965 // The other audit_and_print_stats() call is done at the Debug
1966 // level at a safepoint in SafepointSynchronize::do_cleanup_tasks.
1967 ObjectSynchronizer::audit_and_print_stats(true /* on_exit */);
1968 }
1969 }
1970
1971 // This function can be called at a safepoint or it can be called when
1972 // we are trying to exit the VM. When we are trying to exit the VM, the
1973 // list walker functions can run in parallel with the other list
1974 // operations so spin-locking is used for safety.
1975 //
1976 // Calls to this function can be added in various places as a debugging
1977 // aid; pass 'true' for the 'on_exit' parameter to have in-use monitor
1978 // details logged at the Info level and 'false' for the 'on_exit'
1979 // parameter to have in-use monitor details logged at the Trace level.
1980 //
1981 void ObjectSynchronizer::audit_and_print_stats(bool on_exit) {
1982 assert(on_exit || SafepointSynchronize::is_at_safepoint(), "invariant");
1983
1984 LogStreamHandle(Debug, monitorinflation) lsh_debug;
1985 LogStreamHandle(Info, monitorinflation) lsh_info;
1986 LogStreamHandle(Trace, monitorinflation) lsh_trace;
1987 LogStream* ls = NULL;
1988 if (log_is_enabled(Trace, monitorinflation)) {
1989 ls = &lsh_trace;
1990 } else if (log_is_enabled(Debug, monitorinflation)) {
1991 ls = &lsh_debug;
1992 } else if (log_is_enabled(Info, monitorinflation)) {
1993 ls = &lsh_info;
1994 }
1995 assert(ls != NULL, "sanity check");
1996
1997 int error_cnt = 0;
1998
1999 ls->print_cr("Checking in_use_list:");
2000 chk_in_use_list(ls, &error_cnt);
2001
2002 if (error_cnt == 0) {
2003 ls->print_cr("No errors found in in_use_list checks.");
2004 } else {
2005 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt);
2006 }
2007
2008 if ((on_exit && log_is_enabled(Info, monitorinflation)) ||
2009 (!on_exit && log_is_enabled(Trace, monitorinflation))) {
2010 // When exiting this log output is at the Info level. When called
2011 // at a safepoint, this log output is at the Trace level since
2012 // there can be a lot of it.
2013 log_in_use_monitor_details(ls);
2014 }
2015
2016 ls->flush();
2017
2018 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt);
2019 }
2020
2021 // Check the in_use_list; log the results of the checks.
2022 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) {
2023 size_t l_in_use_count = _in_use_list.count();
2024 size_t l_in_use_max = _in_use_list.max();
2025 out->print_cr("count=" SIZE_FORMAT ", max=" SIZE_FORMAT, l_in_use_count,
2026 l_in_use_max);
2027
2028 size_t ck_in_use_count = 0;
2029 MonitorList::Iterator iter = _in_use_list.iterator();
2030 while (iter.has_next()) {
2031 ObjectMonitor* mid = iter.next();
2032 chk_in_use_entry(mid, out, error_cnt_p);
2033 ck_in_use_count++;
2034 }
2035
2036 if (l_in_use_count == ck_in_use_count) {
2037 out->print_cr("in_use_count=" SIZE_FORMAT " equals ck_in_use_count="
2038 SIZE_FORMAT, l_in_use_count, ck_in_use_count);
2039 } else {
2040 out->print_cr("WARNING: in_use_count=" SIZE_FORMAT " is not equal to "
2041 "ck_in_use_count=" SIZE_FORMAT, l_in_use_count,
2042 ck_in_use_count);
2043 }
2044
2045 size_t ck_in_use_max = _in_use_list.max();
2046 if (l_in_use_max == ck_in_use_max) {
2047 out->print_cr("in_use_max=" SIZE_FORMAT " equals ck_in_use_max="
2048 SIZE_FORMAT, l_in_use_max, ck_in_use_max);
2049 } else {
2050 out->print_cr("WARNING: in_use_max=" SIZE_FORMAT " is not equal to "
2051 "ck_in_use_max=" SIZE_FORMAT, l_in_use_max, ck_in_use_max);
2052 }
2053 }
2054
2055 // Check an in-use monitor entry; log any errors.
2056 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out,
2057 int* error_cnt_p) {
2058 if (n->owner_is_DEFLATER_MARKER()) {
2059 // This should not happen, but if it does, it is not fatal.
2060 out->print_cr("WARNING: monitor=" INTPTR_FORMAT ": in-use monitor is "
2061 "deflated.", p2i(n));
2062 return;
2063 }
2064 if (n->header().value() == 0) {
2065 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must "
2066 "have non-NULL _header field.", p2i(n));
2067 *error_cnt_p = *error_cnt_p + 1;
2068 }
2069 const oop obj = n->object_peek();
2070 if (obj != NULL) {
2071 const markWord mark = obj->mark();
2072 if (!mark.has_monitor()) {
2073 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2074 "object does not think it has a monitor: obj="
2075 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n),
2076 p2i(obj), mark.value());
2077 *error_cnt_p = *error_cnt_p + 1;
2078 }
2079 ObjectMonitor* const obj_mon = mark.monitor();
2080 if (n != obj_mon) {
2081 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2082 "object does not refer to the same monitor: obj="
2083 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon="
2084 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon));
2085 *error_cnt_p = *error_cnt_p + 1;
2086 }
2087 }
2088 }
2089
2090 // Log details about ObjectMonitors on the in_use_list. The 'BHL'
2091 // flags indicate why the entry is in-use, 'object' and 'object type'
2092 // indicate the associated object and its type.
2093 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out) {
2094 stringStream ss;
2095 if (_in_use_list.count() > 0) {
2096 out->print_cr("In-use monitor info:");
2097 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2098 out->print_cr("%18s %s %18s %18s",
2099 "monitor", "BHL", "object", "object type");
2100 out->print_cr("================== === ================== ==================");
2101 MonitorList::Iterator iter = _in_use_list.iterator();
2102 while (iter.has_next()) {
2103 ObjectMonitor* mid = iter.next();
2104 const oop obj = mid->object_peek();
2105 const markWord mark = mid->header();
2106 ResourceMark rm;
2107 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(mid),
2108 mid->is_busy(), mark.hash() != 0, mid->owner() != NULL,
2109 p2i(obj), obj == NULL ? "" : obj->klass()->external_name());
2110 if (mid->is_busy()) {
2111 out->print(" (%s)", mid->is_busy_to_string(&ss));
2112 ss.reset();
2113 }
2114 out->cr();
2115 }
2116 }
2117
2118 out->flush();
2119 }