1 /*
2 * Copyright (c) 1998, 2025, 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 "classfile/vmSymbols.hpp"
26 #include "gc/shared/collectedHeap.hpp"
27 #include "jfr/jfrEvents.hpp"
28 #include "logging/log.hpp"
29 #include "logging/logStream.hpp"
30 #include "memory/allocation.inline.hpp"
31 #include "memory/padded.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "memory/universe.hpp"
34 #include "oops/markWord.hpp"
35 #include "oops/oop.inline.hpp"
36 #include "runtime/atomic.hpp"
37 #include "runtime/basicLock.inline.hpp"
38 #include "runtime/frame.inline.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/javaThread.hpp"
44 #include "runtime/lightweightSynchronizer.hpp"
45 #include "runtime/lockStack.inline.hpp"
46 #include "runtime/mutexLocker.hpp"
47 #include "runtime/objectMonitor.hpp"
48 #include "runtime/objectMonitor.inline.hpp"
49 #include "runtime/os.inline.hpp"
50 #include "runtime/osThread.hpp"
51 #include "runtime/safepointMechanism.inline.hpp"
52 #include "runtime/safepointVerifiers.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "runtime/stubRoutines.hpp"
55 #include "runtime/synchronizer.inline.hpp"
56 #include "runtime/threads.hpp"
57 #include "runtime/timer.hpp"
58 #include "runtime/trimNativeHeap.hpp"
59 #include "runtime/vframe.hpp"
60 #include "runtime/vmThread.hpp"
61 #include "utilities/align.hpp"
62 #include "utilities/dtrace.hpp"
63 #include "utilities/events.hpp"
64 #include "utilities/globalCounter.inline.hpp"
65 #include "utilities/globalDefinitions.hpp"
66 #include "utilities/fastHash.hpp"
67 #include "utilities/linkedlist.hpp"
68 #include "utilities/preserveException.hpp"
69
70 class ObjectMonitorDeflationLogging;
71
72 void MonitorList::add(ObjectMonitor* m) {
73 ObjectMonitor* head;
74 do {
75 head = Atomic::load(&_head);
76 m->set_next_om(head);
77 } while (Atomic::cmpxchg(&_head, head, m) != head);
78
79 size_t count = Atomic::add(&_count, 1u, memory_order_relaxed);
80 size_t old_max;
81 do {
82 old_max = Atomic::load(&_max);
83 if (count <= old_max) {
84 break;
85 }
86 } while (Atomic::cmpxchg(&_max, old_max, count, memory_order_relaxed) != old_max);
87 }
88
89 size_t MonitorList::count() const {
90 return Atomic::load(&_count);
91 }
92
93 size_t MonitorList::max() const {
94 return Atomic::load(&_max);
95 }
96
97 class ObjectMonitorDeflationSafepointer : public StackObj {
98 JavaThread* const _current;
99 ObjectMonitorDeflationLogging* const _log;
100
101 public:
102 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log)
103 : _current(current), _log(log) {}
104
105 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter);
106 };
107
108 // Walk the in-use list and unlink deflated ObjectMonitors.
109 // Returns the number of unlinked ObjectMonitors.
110 size_t MonitorList::unlink_deflated(size_t deflated_count,
111 GrowableArray<ObjectMonitor*>* unlinked_list,
112 ObjectMonitorDeflationSafepointer* safepointer) {
113 size_t unlinked_count = 0;
114 ObjectMonitor* prev = nullptr;
115 ObjectMonitor* m = Atomic::load_acquire(&_head);
116
117 while (m != nullptr) {
118 if (m->is_being_async_deflated()) {
119 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can
120 // modify the list once per batch. The batch starts at "m".
121 size_t unlinked_batch = 0;
122 ObjectMonitor* next = m;
123 // Look for at most MonitorUnlinkBatch monitors, or the number of
124 // deflated and not unlinked monitors, whatever comes first.
125 assert(deflated_count >= unlinked_count, "Sanity: underflow");
126 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch);
127 do {
128 ObjectMonitor* next_next = next->next_om();
129 unlinked_batch++;
130 unlinked_list->append(next);
131 next = next_next;
132 if (unlinked_batch >= unlinked_batch_limit) {
133 // Reached the max batch, so bail out of the gathering loop.
134 break;
135 }
136 if (prev == nullptr && Atomic::load(&_head) != m) {
137 // Current batch used to be at head, but it is not at head anymore.
138 // Bail out and figure out where we currently are. This avoids long
139 // walks searching for new prev during unlink under heavy list inserts.
140 break;
141 }
142 } while (next != nullptr && next->is_being_async_deflated());
143
144 // Unlink the found batch.
145 if (prev == nullptr) {
146 // The current batch is the first batch, so there is a chance that it starts at head.
147 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head.
148 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, m, next);
149 if (prev_head != m) {
150 // Something must have updated the head. Figure out the actual prev for this batch.
151 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) {
152 prev = n;
153 }
154 assert(prev != nullptr, "Should have found the prev for the current batch");
155 prev->set_next_om(next);
156 }
157 } else {
158 // The current batch is preceded by another batch. This guarantees the current batch
159 // does not start at head. Unlink the entire current batch without updating the head.
160 assert(Atomic::load(&_head) != m, "Sanity");
161 prev->set_next_om(next);
162 }
163
164 unlinked_count += unlinked_batch;
165 if (unlinked_count >= deflated_count) {
166 // Reached the max so bail out of the searching loop.
167 // There should be no more deflated monitors left.
168 break;
169 }
170 m = next;
171 } else {
172 prev = m;
173 m = m->next_om();
174 }
175
176 // Must check for a safepoint/handshake and honor it.
177 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count);
178 }
179
180 #ifdef ASSERT
181 // Invariant: the code above should unlink all deflated monitors.
182 // The code that runs after this unlinking does not expect deflated monitors.
183 // Notably, attempting to deflate the already deflated monitor would break.
184 {
185 ObjectMonitor* m = Atomic::load_acquire(&_head);
186 while (m != nullptr) {
187 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked");
188 m = m->next_om();
189 }
190 }
191 #endif
192
193 Atomic::sub(&_count, unlinked_count);
194 return unlinked_count;
195 }
196
197 MonitorList::Iterator MonitorList::iterator() const {
198 return Iterator(Atomic::load_acquire(&_head));
199 }
200
201 ObjectMonitor* MonitorList::Iterator::next() {
202 ObjectMonitor* current = _current;
203 _current = current->next_om();
204 return current;
205 }
206
207 // The "core" versions of monitor enter and exit reside in this file.
208 // The interpreter and compilers contain specialized transliterated
209 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp
210 // fast_lock(...) for instance. If you make changes here, make sure to modify the
211 // interpreter, and both C1 and C2 fast-path inline locking code emission.
212 //
213 // -----------------------------------------------------------------------------
214
215 #ifdef DTRACE_ENABLED
216
217 // Only bother with this argument setup if dtrace is available
218 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
219
220 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
221 char* bytes = nullptr; \
222 int len = 0; \
223 jlong jtid = SharedRuntime::get_java_tid(thread); \
224 Symbol* klassname = obj->klass()->name(); \
225 if (klassname != nullptr) { \
226 bytes = (char*)klassname->bytes(); \
227 len = klassname->utf8_length(); \
228 }
229
230 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
231 { \
232 if (DTraceMonitorProbes) { \
233 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
234 HOTSPOT_MONITOR_WAIT(jtid, \
235 (uintptr_t)(monitor), bytes, len, (millis)); \
236 } \
237 }
238
239 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY
240 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL
241 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
242
243 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
244 { \
245 if (DTraceMonitorProbes) { \
246 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
247 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
248 (uintptr_t)(monitor), bytes, len); \
249 } \
250 }
251
252 #else // ndef DTRACE_ENABLED
253
254 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
255 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
256
257 #endif // ndef DTRACE_ENABLED
258
259 // This exists only as a workaround of dtrace bug 6254741
260 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) {
261 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
262 return 0;
263 }
264
265 static constexpr size_t inflation_lock_count() {
266 return 256;
267 }
268
269 // Static storage for an array of PlatformMutex.
270 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)];
271
272 static inline PlatformMutex* inflation_lock(size_t index) {
273 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]);
274 }
275
276 void ObjectSynchronizer::initialize() {
277 for (size_t i = 0; i < inflation_lock_count(); i++) {
278 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex();
279 }
280 // Start the ceiling with the estimate for one thread.
281 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate);
282
283 // Start the timer for deflations, so it does not trigger immediately.
284 _last_async_deflation_time_ns = os::javaTimeNanos();
285
286 if (LockingMode == LM_LIGHTWEIGHT) {
287 LightweightSynchronizer::initialize();
288 }
289 }
290
291 MonitorList ObjectSynchronizer::_in_use_list;
292 // monitors_used_above_threshold() policy is as follows:
293 //
294 // The ratio of the current _in_use_list count to the ceiling is used
295 // to determine if we are above MonitorUsedDeflationThreshold and need
296 // to do an async monitor deflation cycle. The ceiling is increased by
297 // AvgMonitorsPerThreadEstimate when a thread is added to the system
298 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is
299 // removed from the system.
300 //
301 // Note: If the _in_use_list max exceeds the ceiling, then
302 // monitors_used_above_threshold() will use the in_use_list max instead
303 // of the thread count derived ceiling because we have used more
304 // ObjectMonitors than the estimated average.
305 //
306 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax
307 // no-progress async monitor deflation cycles in a row, then the ceiling
308 // is adjusted upwards by monitors_used_above_threshold().
309 //
310 // Start the ceiling with the estimate for one thread in initialize()
311 // which is called after cmd line options are processed.
312 static size_t _in_use_list_ceiling = 0;
313 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false;
314 bool volatile ObjectSynchronizer::_is_final_audit = false;
315 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0;
316 static uintx _no_progress_cnt = 0;
317 static bool _no_progress_skip_increment = false;
318
319 // =====================> Quick functions
320
321 // The quick_* forms are special fast-path variants used to improve
322 // performance. In the simplest case, a "quick_*" implementation could
323 // simply return false, in which case the caller will perform the necessary
324 // state transitions and call the slow-path form.
325 // The fast-path is designed to handle frequently arising cases in an efficient
326 // manner and is just a degenerate "optimistic" variant of the slow-path.
327 // returns true -- to indicate the call was satisfied.
328 // returns false -- to indicate the call needs the services of the slow-path.
329 // A no-loitering ordinance is in effect for code in the quick_* family
330 // operators: safepoints or indefinite blocking (blocking that might span a
331 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon
332 // entry.
333 //
334 // Consider: An interesting optimization is to have the JIT recognize the
335 // following common idiom:
336 // synchronized (someobj) { .... ; notify(); }
337 // That is, we find a notify() or notifyAll() call that immediately precedes
338 // the monitorexit operation. In that case the JIT could fuse the operations
339 // into a single notifyAndExit() runtime primitive.
340
341 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) {
342 assert(current->thread_state() == _thread_in_Java, "invariant");
343 NoSafepointVerifier nsv;
344 if (obj == nullptr) return false; // slow-path for invalid obj
345 const markWord mark = obj->mark();
346
347 if (LockingMode == LM_LIGHTWEIGHT) {
348 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) {
349 // Degenerate notify
350 // fast-locked by caller so by definition the implied waitset is empty.
351 return true;
352 }
353 } else if (LockingMode == LM_LEGACY) {
354 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
355 // Degenerate notify
356 // stack-locked by caller so by definition the implied waitset is empty.
357 return true;
358 }
359 }
360
361 if (mark.has_monitor()) {
362 ObjectMonitor* const mon = read_monitor(current, obj, mark);
363 if (LockingMode == LM_LIGHTWEIGHT && mon == nullptr) {
364 // Racing with inflation/deflation go slow path
365 return false;
366 }
367 assert(mon->object() == oop(obj), "invariant");
368 if (!mon->has_owner(current)) return false; // slow-path for IMS exception
369
370 if (mon->first_waiter() != nullptr) {
371 // We have one or more waiters. Since this is an inflated monitor
372 // that we own, we can transfer one or more threads from the waitset
373 // to the entry_list here and now, avoiding the slow-path.
374 if (all) {
375 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, current);
376 } else {
377 DTRACE_MONITOR_PROBE(notify, mon, obj, current);
378 }
379 do {
380 mon->notify_internal(current);
381 } while (mon->first_waiter() != nullptr && all);
382 }
383 return true;
384 }
385
386 // other IMS exception states take the slow-path
387 return false;
388 }
389
390 static bool useHeavyMonitors() {
391 #if defined(X86) || defined(AARCH64) || defined(PPC64) || defined(RISCV64) || defined(S390)
392 return LockingMode == LM_MONITOR;
393 #else
394 return false;
395 #endif
396 }
397
398 // The LockNode emitted directly at the synchronization site would have
399 // been too big if it were to have included support for the cases of inflated
400 // recursive enter and exit, so they go here instead.
401 // Note that we can't safely call AsyncPrintJavaStack() from within
402 // quick_enter() as our thread state remains _in_Java.
403
404 bool ObjectSynchronizer::quick_enter_legacy(oop obj, BasicLock* lock, JavaThread* current) {
405 assert(current->thread_state() == _thread_in_Java, "invariant");
406
407 if (useHeavyMonitors()) {
408 return false; // Slow path
409 }
410
411 assert(LockingMode == LM_LEGACY, "legacy mode below");
412
413 const markWord mark = obj->mark();
414
415 if (mark.has_monitor()) {
416
417 ObjectMonitor* const m = read_monitor(mark);
418 // An async deflation or GC can race us before we manage to make
419 // the ObjectMonitor busy by setting the owner below. If we detect
420 // that race we just bail out to the slow-path here.
421 if (m->object_peek() == nullptr) {
422 return false;
423 }
424
425 // Lock contention and Transactional Lock Elision (TLE) diagnostics
426 // and observability
427 // Case: light contention possibly amenable to TLE
428 // Case: TLE inimical operations such as nested/recursive synchronization
429
430 if (m->has_owner(current)) {
431 m->_recursions++;
432 current->inc_held_monitor_count();
433 return true;
434 }
435
436 // This Java Monitor is inflated so obj's header will never be
437 // displaced to this thread's BasicLock. Make the displaced header
438 // non-null so this BasicLock is not seen as recursive nor as
439 // being locked. We do this unconditionally so that this thread's
440 // BasicLock cannot be mis-interpreted by any stack walkers. For
441 // performance reasons, stack walkers generally first check for
442 // stack-locking in the object's header, the second check is for
443 // recursive stack-locking in the displaced header in the BasicLock,
444 // and last are the inflated Java Monitor (ObjectMonitor) checks.
445 lock->set_displaced_header(markWord::unused_mark());
446
447 if (!m->has_owner() && m->try_set_owner(current)) {
448 assert(m->_recursions == 0, "invariant");
449 current->inc_held_monitor_count();
450 return true;
451 }
452 }
453
454 // Note that we could inflate in quick_enter.
455 // This is likely a useful optimization
456 // Critically, in quick_enter() we must not:
457 // -- block indefinitely, or
458 // -- reach a safepoint
459
460 return false; // revert to slow-path
461 }
462
463 // Handle notifications when synchronizing on value based classes
464 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) {
465 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
466 frame last_frame = locking_thread->last_frame();
467 bool bcp_was_adjusted = false;
468 // Don't decrement bcp if it points to the frame's first instruction. This happens when
469 // handle_sync_on_value_based_class() is called because of a synchronized method. There
470 // is no actual monitorenter instruction in the byte code in this case.
471 if (last_frame.is_interpreted_frame() &&
472 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) {
473 // adjust bcp to point back to monitorenter so that we print the correct line numbers
474 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1);
475 bcp_was_adjusted = true;
476 }
477
478 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) {
479 ResourceMark rm;
480 stringStream ss;
481 locking_thread->print_active_stack_on(&ss);
482 char* base = (char*)strstr(ss.base(), "at");
483 char* newline = (char*)strchr(ss.base(), '\n');
484 if (newline != nullptr) {
485 *newline = '\0';
486 }
487 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base);
488 } else {
489 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses");
490 ResourceMark rm;
491 Log(valuebasedclasses) vblog;
492
493 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name());
494 if (locking_thread->has_last_Java_frame()) {
495 LogStream info_stream(vblog.info());
496 locking_thread->print_active_stack_on(&info_stream);
497 } else {
498 vblog.info("Cannot find the last Java frame");
499 }
500
501 EventSyncOnValueBasedClass event;
502 if (event.should_commit()) {
503 event.set_valueBasedClass(obj->klass());
504 event.commit();
505 }
506 }
507
508 if (bcp_was_adjusted) {
509 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1);
510 }
511 }
512
513 // -----------------------------------------------------------------------------
514 // Monitor Enter/Exit
515
516 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
517 // When called with locking_thread != Thread::current() some mechanism must synchronize
518 // the locking_thread with respect to the current thread. Currently only used when
519 // deoptimizing and re-locking locks. See Deoptimization::relock_objects
520 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
521
522 if (LockingMode == LM_LIGHTWEIGHT) {
523 return LightweightSynchronizer::enter_for(obj, lock, locking_thread);
524 }
525
526 if (!enter_fast_impl(obj, lock, locking_thread)) {
527 // Inflated ObjectMonitor::enter_for is required
528
529 // An async deflation can race after the inflate_for() call and before
530 // enter_for() can make the ObjectMonitor busy. enter_for() returns false
531 // if we have lost the race to async deflation and we simply try again.
532 while (true) {
533 ObjectMonitor* monitor = inflate_for(locking_thread, obj(), inflate_cause_monitor_enter);
534 if (monitor->enter_for(locking_thread)) {
535 return;
536 }
537 assert(monitor->is_being_async_deflated(), "must be");
538 }
539 }
540 }
541
542 void ObjectSynchronizer::enter_legacy(Handle obj, BasicLock* lock, JavaThread* current) {
543 if (!enter_fast_impl(obj, lock, current)) {
544 // Inflated ObjectMonitor::enter is required
545
546 // An async deflation can race after the inflate() call and before
547 // enter() can make the ObjectMonitor busy. enter() returns false if
548 // we have lost the race to async deflation and we simply try again.
549 while (true) {
550 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter);
551 if (monitor->enter(current)) {
552 return;
553 }
554 }
555 }
556 }
557
558 // The interpreter and compiler assembly code tries to lock using the fast path
559 // of this algorithm. Make sure to update that code if the following function is
560 // changed. The implementation is extremely sensitive to race condition. Be careful.
561 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
562 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer");
563
564 if (obj->klass()->is_value_based()) {
565 handle_sync_on_value_based_class(obj, locking_thread);
566 }
567
568 locking_thread->inc_held_monitor_count();
569
570 if (!useHeavyMonitors()) {
571 if (LockingMode == LM_LEGACY) {
572 markWord mark = obj->mark();
573 if (mark.is_unlocked()) {
574 // Anticipate successful CAS -- the ST of the displaced mark must
575 // be visible <= the ST performed by the CAS.
576 lock->set_displaced_header(mark);
577 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) {
578 return true;
579 }
580 } else if (mark.has_locker() &&
581 locking_thread->is_lock_owned((address) mark.locker())) {
582 assert(lock != mark.locker(), "must not re-lock the same lock");
583 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock");
584 lock->set_displaced_header(markWord::from_pointer(nullptr));
585 return true;
586 }
587
588 // The object header will never be displaced to this lock,
589 // so it does not matter what the value is, except that it
590 // must be non-zero to avoid looking like a re-entrant lock,
591 // and must not look locked either.
592 lock->set_displaced_header(markWord::unused_mark());
593
594 // Failed to fast lock.
595 return false;
596 }
597 } else if (VerifyHeavyMonitors) {
598 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
599 }
600
601 return false;
602 }
603
604 void ObjectSynchronizer::exit_legacy(oop object, BasicLock* lock, JavaThread* current) {
605 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer");
606
607 if (!useHeavyMonitors()) {
608 markWord mark = object->mark();
609 if (LockingMode == LM_LEGACY) {
610 markWord dhw = lock->displaced_header();
611 if (dhw.value() == 0) {
612 // If the displaced header is null, then this exit matches up with
613 // a recursive enter. No real work to do here except for diagnostics.
614 #ifndef PRODUCT
615 if (mark != markWord::INFLATING()) {
616 // Only do diagnostics if we are not racing an inflation. Simply
617 // exiting a recursive enter of a Java Monitor that is being
618 // inflated is safe; see the has_monitor() comment below.
619 assert(!mark.is_unlocked(), "invariant");
620 assert(!mark.has_locker() ||
621 current->is_lock_owned((address)mark.locker()), "invariant");
622 if (mark.has_monitor()) {
623 // The BasicLock's displaced_header is marked as a recursive
624 // enter and we have an inflated Java Monitor (ObjectMonitor).
625 // This is a special case where the Java Monitor was inflated
626 // after this thread entered the stack-lock recursively. When a
627 // Java Monitor is inflated, we cannot safely walk the Java
628 // Monitor owner's stack and update the BasicLocks because a
629 // Java Monitor can be asynchronously inflated by a thread that
630 // does not own the Java Monitor.
631 ObjectMonitor* m = read_monitor(mark);
632 assert(m->object()->mark() == mark, "invariant");
633 assert(m->is_entered(current), "invariant");
634 }
635 }
636 #endif
637 return;
638 }
639
640 if (mark == markWord::from_pointer(lock)) {
641 // If the object is stack-locked by the current thread, try to
642 // swing the displaced header from the BasicLock back to the mark.
643 assert(dhw.is_neutral(), "invariant");
644 if (object->cas_set_mark(dhw, mark) == mark) {
645 return;
646 }
647 }
648 }
649 } else if (VerifyHeavyMonitors) {
650 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
651 }
652
653 // We have to take the slow-path of possible inflation and then exit.
654 // The ObjectMonitor* can't be async deflated until ownership is
655 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
656 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal);
657 assert(!monitor->has_anonymous_owner(), "must not be");
658 monitor->exit(current);
659 }
660
661 // -----------------------------------------------------------------------------
662 // JNI locks on java objects
663 // NOTE: must use heavy weight monitor to handle jni monitor enter
664 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) {
665 // Top native frames in the stack will not be seen if we attempt
666 // preemption, since we start walking from the last Java anchor.
667 NoPreemptMark npm(current);
668
669 if (obj->klass()->is_value_based()) {
670 handle_sync_on_value_based_class(obj, current);
671 }
672
673 // the current locking is from JNI instead of Java code
674 current->set_current_pending_monitor_is_from_java(false);
675 // An async deflation can race after the inflate() call and before
676 // enter() can make the ObjectMonitor busy. enter() returns false if
677 // we have lost the race to async deflation and we simply try again.
678 while (true) {
679 ObjectMonitor* monitor;
680 bool entered;
681 if (LockingMode == LM_LIGHTWEIGHT) {
682 entered = LightweightSynchronizer::inflate_and_enter(obj(), inflate_cause_jni_enter, current, current) != nullptr;
683 } else {
684 monitor = inflate(current, obj(), inflate_cause_jni_enter);
685 entered = monitor->enter(current);
686 }
687
688 if (entered) {
689 current->inc_held_monitor_count(1, true);
690 break;
691 }
692 }
693 current->set_current_pending_monitor_is_from_java(true);
694 }
695
696 // NOTE: must use heavy weight monitor to handle jni monitor exit
697 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) {
698 JavaThread* current = THREAD;
699
700 ObjectMonitor* monitor;
701 if (LockingMode == LM_LIGHTWEIGHT) {
702 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj, inflate_cause_jni_exit, CHECK);
703 } else {
704 // The ObjectMonitor* can't be async deflated until ownership is
705 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
706 monitor = inflate(current, obj, inflate_cause_jni_exit);
707 }
708 // If this thread has locked the object, exit the monitor. We
709 // intentionally do not use CHECK on check_owner because we must exit the
710 // monitor even if an exception was already pending.
711 if (monitor->check_owner(THREAD)) {
712 monitor->exit(current);
713 current->dec_held_monitor_count(1, true);
714 }
715 }
716
717 // -----------------------------------------------------------------------------
718 // Internal VM locks on java objects
719 // standard constructor, allows locking failures
720 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) : _npm(thread) {
721 _thread = thread;
722 _thread->check_for_valid_safepoint_state();
723 _obj = obj;
724
725 if (_obj() != nullptr) {
726 ObjectSynchronizer::enter(_obj, &_lock, _thread);
727 }
728 }
729
730 ObjectLocker::~ObjectLocker() {
731 if (_obj() != nullptr) {
732 ObjectSynchronizer::exit(_obj(), &_lock, _thread);
733 }
734 }
735
736
737 // -----------------------------------------------------------------------------
738 // Wait/Notify/NotifyAll
739 // NOTE: must use heavy weight monitor to handle wait()
740
741 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
742 JavaThread* current = THREAD;
743 if (millis < 0) {
744 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
745 }
746
747 ObjectMonitor* monitor;
748 if (LockingMode == LM_LIGHTWEIGHT) {
749 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK_0);
750 } else {
751 // The ObjectMonitor* can't be async deflated because the _waiters
752 // field is incremented before ownership is dropped and decremented
753 // after ownership is regained.
754 monitor = inflate(current, obj(), inflate_cause_wait);
755 }
756
757 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis);
758 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code
759
760 // This dummy call is in place to get around dtrace bug 6254741. Once
761 // that's fixed we can uncomment the following line, remove the call
762 // and change this function back into a "void" func.
763 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
764 int ret_code = dtrace_waited_probe(monitor, obj, THREAD);
765 return ret_code;
766 }
767
768 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
769 if (millis < 0) {
770 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
771 }
772
773 ObjectMonitor* monitor;
774 if (LockingMode == LM_LIGHTWEIGHT) {
775 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK);
776 } else {
777 monitor = inflate(THREAD, obj(), inflate_cause_wait);
778 }
779 monitor->wait(millis, false, THREAD);
780 }
781
782
783 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
784 JavaThread* current = THREAD;
785
786 markWord mark = obj->mark();
787 if (LockingMode == LM_LIGHTWEIGHT) {
788 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
789 // Not inflated so there can't be any waiters to notify.
790 return;
791 }
792 } else if (LockingMode == LM_LEGACY) {
793 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
794 // Not inflated so there can't be any waiters to notify.
795 return;
796 }
797 }
798
799 ObjectMonitor* monitor;
800 if (LockingMode == LM_LIGHTWEIGHT) {
801 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK);
802 } else {
803 // The ObjectMonitor* can't be async deflated until ownership is
804 // dropped by the calling thread.
805 monitor = inflate(current, obj(), inflate_cause_notify);
806 }
807 monitor->notify(CHECK);
808 }
809
810 // NOTE: see comment of notify()
811 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
812 JavaThread* current = THREAD;
813
814 markWord mark = obj->mark();
815 if (LockingMode == LM_LIGHTWEIGHT) {
816 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
817 // Not inflated so there can't be any waiters to notify.
818 return;
819 }
820 } else if (LockingMode == LM_LEGACY) {
821 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
822 // Not inflated so there can't be any waiters to notify.
823 return;
824 }
825 }
826
827 ObjectMonitor* monitor;
828 if (LockingMode == LM_LIGHTWEIGHT) {
829 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK);
830 } else {
831 // The ObjectMonitor* can't be async deflated until ownership is
832 // dropped by the calling thread.
833 monitor = inflate(current, obj(), inflate_cause_notify);
834 }
835 monitor->notifyAll(CHECK);
836 }
837
838 // -----------------------------------------------------------------------------
839 // Hash Code handling
840
841 struct SharedGlobals {
842 char _pad_prefix[OM_CACHE_LINE_SIZE];
843 // This is a highly shared mostly-read variable.
844 // To avoid false-sharing it needs to be the sole occupant of a cache line.
845 volatile int stw_random;
846 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int));
847 // Hot RW variable -- Sequester to avoid false-sharing
848 volatile int hc_sequence;
849 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int));
850 };
851
852 static SharedGlobals GVars;
853
854 static markWord read_stable_mark(oop obj) {
855 markWord mark = obj->mark_acquire();
856 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) {
857 // New lightweight locking does not use the markWord::INFLATING() protocol.
858 return mark; // normal fast-path return
859 }
860
861 int its = 0;
862 for (;;) {
863 markWord mark = obj->mark_acquire();
864 if (!mark.is_being_inflated()) {
865 return mark; // normal fast-path return
866 }
867
868 // The object is being inflated by some other thread.
869 // The caller of read_stable_mark() must wait for inflation to complete.
870 // Avoid live-lock.
871
872 ++its;
873 if (its > 10000 || !os::is_MP()) {
874 if (its & 1) {
875 os::naked_yield();
876 } else {
877 // Note that the following code attenuates the livelock problem but is not
878 // a complete remedy. A more complete solution would require that the inflating
879 // thread hold the associated inflation lock. The following code simply restricts
880 // the number of spinners to at most one. We'll have N-2 threads blocked
881 // on the inflationlock, 1 thread holding the inflation lock and using
882 // a yield/park strategy, and 1 thread in the midst of inflation.
883 // A more refined approach would be to change the encoding of INFLATING
884 // to allow encapsulation of a native thread pointer. Threads waiting for
885 // inflation to complete would use CAS to push themselves onto a singly linked
886 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
887 // and calling park(). When inflation was complete the thread that accomplished inflation
888 // would detach the list and set the markword to inflated with a single CAS and
889 // then for each thread on the list, set the flag and unpark() the thread.
890
891 // Index into the lock array based on the current object address.
892 static_assert(is_power_of_2(inflation_lock_count()), "must be");
893 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1);
894 int YieldThenBlock = 0;
895 assert(ix < inflation_lock_count(), "invariant");
896 inflation_lock(ix)->lock();
897 while (obj->mark_acquire() == markWord::INFLATING()) {
898 // Beware: naked_yield() is advisory and has almost no effect on some platforms
899 // so we periodically call current->_ParkEvent->park(1).
900 // We use a mixed spin/yield/block mechanism.
901 if ((YieldThenBlock++) >= 16) {
902 Thread::current()->_ParkEvent->park(1);
903 } else {
904 os::naked_yield();
905 }
906 }
907 inflation_lock(ix)->unlock();
908 }
909 } else {
910 SpinPause(); // SMP-polite spinning
911 }
912 }
913 }
914
915 // hashCode() generation :
916 //
917 // Possibilities:
918 // * MD5Digest of {obj,stw_random}
919 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function.
920 // * A DES- or AES-style SBox[] mechanism
921 // * One of the Phi-based schemes, such as:
922 // 2654435761 = 2^32 * Phi (golden ratio)
923 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ;
924 // * A variation of Marsaglia's shift-xor RNG scheme.
925 // * (obj ^ stw_random) is appealing, but can result
926 // in undesirable regularity in the hashCode values of adjacent objects
927 // (objects allocated back-to-back, in particular). This could potentially
928 // result in hashtable collisions and reduced hashtable efficiency.
929 // There are simple ways to "diffuse" the middle address bits over the
930 // generated hashCode values:
931
932 intptr_t ObjectSynchronizer::get_next_hash(Thread* current, oop obj) {
933 intptr_t value = 0;
934 if (hashCode == 0) {
935 // This form uses global Park-Miller RNG.
936 // On MP system we'll have lots of RW access to a global, so the
937 // mechanism induces lots of coherency traffic.
938 value = os::random();
939 } else if (hashCode == 1) {
940 // This variation has the property of being stable (idempotent)
941 // between STW operations. This can be useful in some of the 1-0
942 // synchronization schemes.
943 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3;
944 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random;
945 } else if (hashCode == 2) {
946 value = 1; // for sensitivity testing
947 } else if (hashCode == 3) {
948 value = ++GVars.hc_sequence;
949 } else if (hashCode == 4) {
950 value = cast_from_oop<intptr_t>(obj);
951 } else if (hashCode == 5) {
952 // Marsaglia's xor-shift scheme with thread-specific state
953 // This is probably the best overall implementation -- we'll
954 // likely make this the default in future releases.
955 unsigned t = current->_hashStateX;
956 t ^= (t << 11);
957 current->_hashStateX = current->_hashStateY;
958 current->_hashStateY = current->_hashStateZ;
959 current->_hashStateZ = current->_hashStateW;
960 unsigned v = current->_hashStateW;
961 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
962 current->_hashStateW = v;
963 value = v;
964 } else {
965 assert(UseCompactObjectHeaders, "Only with compact i-hash");
966 #ifdef _LP64
967 uint64_t val = cast_from_oop<uint64_t>(obj);
968 uint32_t hash = FastHash::get_hash32((uint32_t)val, (uint32_t)(val >> 32));
969 #else
970 uint32_t val = cast_from_oop<uint32_t>(obj);
971 uint32_t hash = FastHash::get_hash32(val, UCONST64(0xAAAAAAAA));
972 #endif
973 value= static_cast<intptr_t>(hash);
974 }
975
976 value &= markWord::hash_mask;
977 if (hashCode != 6 && value == 0) value = 0xBAD;
978 assert(value != markWord::no_hash || hashCode == 6, "invariant");
979 return value;
980 }
981
982 static intptr_t install_hash_code(Thread* current, oop obj) {
983 assert(UseObjectMonitorTable && LockingMode == LM_LIGHTWEIGHT, "must be");
984
985 markWord mark = obj->mark_acquire();
986 for (;;) {
987 if (UseCompactObjectHeaders) {
988 if (mark.is_hashed()) {
989 return LightweightSynchronizer::get_hash(mark, obj);
990 }
991 intptr_t hash = ObjectSynchronizer::get_next_hash(current, obj); // get a new hash
992 markWord new_mark;
993 if (mark.is_not_hashed_expanded()) {
994 new_mark = mark.set_hashed_expanded();
995 int offset = mark.klass()->hash_offset_in_bytes(obj, mark);
996 obj->int_field_put(offset, (jint) hash);
997 } else {
998 new_mark = mark.set_hashed_not_expanded();
999 }
1000 markWord old_mark = obj->cas_set_mark(new_mark, mark);
1001 if (old_mark == mark) {
1002 return hash;
1003 }
1004 mark = old_mark;
1005 } else {
1006 intptr_t hash = mark.hash();
1007 if (hash != 0) {
1008 return hash;
1009 }
1010
1011 hash = ObjectSynchronizer::get_next_hash(current, obj);
1012 const markWord old_mark = mark;
1013 const markWord new_mark = old_mark.copy_set_hash(hash);
1014
1015 mark = obj->cas_set_mark(new_mark, old_mark);
1016 if (old_mark == mark) {
1017 return hash;
1018 }
1019 }
1020 }
1021 }
1022
1023 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) {
1024 if (UseObjectMonitorTable) {
1025 // Since the monitor isn't in the object header, the hash can simply be
1026 // installed in the object header.
1027 return install_hash_code(current, obj);
1028 }
1029
1030 while (true) {
1031 ObjectMonitor* monitor = nullptr;
1032 markWord temp, test;
1033 intptr_t hash;
1034 markWord mark = read_stable_mark(obj);
1035 if (VerifyHeavyMonitors) {
1036 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)");
1037 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
1038 }
1039 if (mark.is_unlocked() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) {
1040 hash = mark.hash();
1041 if (hash != 0) { // if it has a hash, just return it
1042 return hash;
1043 }
1044 hash = get_next_hash(current, obj); // get a new hash
1045 temp = mark.copy_set_hash(hash); // merge the hash into header
1046 // try to install the hash
1047 test = obj->cas_set_mark(temp, mark);
1048 if (test == mark) { // if the hash was installed, return it
1049 return hash;
1050 }
1051 if (LockingMode == LM_LIGHTWEIGHT) {
1052 // CAS failed, retry
1053 continue;
1054 }
1055 // Failed to install the hash. It could be that another thread
1056 // installed the hash just before our attempt or inflation has
1057 // occurred or... so we fall thru to inflate the monitor for
1058 // stability and then install the hash.
1059 } else if (mark.has_monitor()) {
1060 monitor = mark.monitor();
1061 temp = monitor->header();
1062 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1063 hash = temp.hash();
1064 if (hash != 0) {
1065 // It has a hash.
1066
1067 // Separate load of dmw/header above from the loads in
1068 // is_being_async_deflated().
1069
1070 // dmw/header and _contentions may get written by different threads.
1071 // Make sure to observe them in the same order when having several observers.
1072 OrderAccess::loadload_for_IRIW();
1073
1074 if (monitor->is_being_async_deflated()) {
1075 // But we can't safely use the hash if we detect that async
1076 // deflation has occurred. So we attempt to restore the
1077 // header/dmw to the object's header so that we only retry
1078 // once if the deflater thread happens to be slow.
1079 monitor->install_displaced_markword_in_object(obj);
1080 continue;
1081 }
1082 return hash;
1083 }
1084 // Fall thru so we only have one place that installs the hash in
1085 // the ObjectMonitor.
1086 } else if (LockingMode == LM_LEGACY && mark.has_locker()
1087 && current->is_Java_thread()
1088 && JavaThread::cast(current)->is_lock_owned((address)mark.locker())) {
1089 // This is a stack-lock owned by the calling thread so fetch the
1090 // displaced markWord from the BasicLock on the stack.
1091 temp = mark.displaced_mark_helper();
1092 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1093 hash = temp.hash();
1094 if (hash != 0) { // if it has a hash, just return it
1095 return hash;
1096 }
1097 // WARNING:
1098 // The displaced header in the BasicLock on a thread's stack
1099 // is strictly immutable. It CANNOT be changed in ANY cases.
1100 // So we have to inflate the stack-lock into an ObjectMonitor
1101 // even if the current thread owns the lock. The BasicLock on
1102 // a thread's stack can be asynchronously read by other threads
1103 // during an inflate() call so any change to that stack memory
1104 // may not propagate to other threads correctly.
1105 }
1106
1107 // Inflate the monitor to set the hash.
1108
1109 // There's no need to inflate if the mark has already got a monitor.
1110 // NOTE: an async deflation can race after we get the monitor and
1111 // before we can update the ObjectMonitor's header with the hash
1112 // value below.
1113 monitor = mark.has_monitor() ? mark.monitor() : inflate(current, obj, inflate_cause_hash_code);
1114 // Load ObjectMonitor's header/dmw field and see if it has a hash.
1115 mark = monitor->header();
1116 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
1117 hash = mark.hash();
1118 if (hash == 0) { // if it does not have a hash
1119 hash = get_next_hash(current, obj); // get a new hash
1120 temp = mark.copy_set_hash(hash) ; // merge the hash into header
1121 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1122 uintptr_t v = Atomic::cmpxchg(monitor->metadata_addr(), mark.value(), temp.value());
1123 test = markWord(v);
1124 if (test != mark) {
1125 // The attempt to update the ObjectMonitor's header/dmw field
1126 // did not work. This can happen if another thread managed to
1127 // merge in the hash just before our cmpxchg().
1128 // If we add any new usages of the header/dmw field, this code
1129 // will need to be updated.
1130 hash = test.hash();
1131 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value());
1132 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash");
1133 }
1134 if (monitor->is_being_async_deflated() && !UseObjectMonitorTable) {
1135 // If we detect that async deflation has occurred, then we
1136 // attempt to restore the header/dmw to the object's header
1137 // so that we only retry once if the deflater thread happens
1138 // to be slow.
1139 monitor->install_displaced_markword_in_object(obj);
1140 continue;
1141 }
1142 }
1143 // We finally get the hash.
1144 return hash;
1145 }
1146 }
1147
1148 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current,
1149 Handle h_obj) {
1150 assert(current == JavaThread::current(), "Can only be called on current thread");
1151 oop obj = h_obj();
1152
1153 markWord mark = read_stable_mark(obj);
1154
1155 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1156 // stack-locked case, header points into owner's stack
1157 return current->is_lock_owned((address)mark.locker());
1158 }
1159
1160 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1161 // fast-locking case, see if lock is in current's lock stack
1162 return current->lock_stack().contains(h_obj());
1163 }
1164
1165 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) {
1166 ObjectMonitor* monitor = read_monitor(current, obj, mark);
1167 if (monitor != nullptr) {
1168 return monitor->is_entered(current) != 0;
1169 }
1170 // Racing with inflation/deflation, retry
1171 mark = obj->mark_acquire();
1172
1173 if (mark.is_fast_locked()) {
1174 // Some other thread fast_locked, current could not have held the lock
1175 return false;
1176 }
1177 }
1178
1179 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) {
1180 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
1181 // The first stage of async deflation does not affect any field
1182 // used by this comparison so the ObjectMonitor* is usable here.
1183 ObjectMonitor* monitor = read_monitor(mark);
1184 return monitor->is_entered(current) != 0;
1185 }
1186 // Unlocked case, header in place
1187 assert(mark.is_unlocked(), "sanity check");
1188 return false;
1189 }
1190
1191 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
1192 oop obj = h_obj();
1193 markWord mark = read_stable_mark(obj);
1194
1195 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1196 // stack-locked so header points into owner's stack.
1197 // owning_thread_from_monitor_owner() may also return null here:
1198 return Threads::owning_thread_from_stacklock(t_list, (address) mark.locker());
1199 }
1200
1201 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1202 // fast-locked so get owner from the object.
1203 // owning_thread_from_object() may also return null here:
1204 return Threads::owning_thread_from_object(t_list, h_obj());
1205 }
1206
1207 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) {
1208 ObjectMonitor* monitor = read_monitor(Thread::current(), obj, mark);
1209 if (monitor != nullptr) {
1210 return Threads::owning_thread_from_monitor(t_list, monitor);
1211 }
1212 // Racing with inflation/deflation, retry
1213 mark = obj->mark_acquire();
1214
1215 if (mark.is_fast_locked()) {
1216 // Some other thread fast_locked
1217 return Threads::owning_thread_from_object(t_list, h_obj());
1218 }
1219 }
1220
1221 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) {
1222 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
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 = read_monitor(mark);
1226 assert(monitor != nullptr, "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_unlocked(), "sanity check");
1235
1236 return nullptr;
1237 }
1238
1239 // Visitors ...
1240
1241 // Iterate over all ObjectMonitors.
1242 template <typename Function>
1243 void ObjectSynchronizer::monitors_iterate(Function function) {
1244 MonitorList::Iterator iter = _in_use_list.iterator();
1245 while (iter.has_next()) {
1246 ObjectMonitor* monitor = iter.next();
1247 function(monitor);
1248 }
1249 }
1250
1251 // Iterate ObjectMonitors owned by any thread and where the owner `filter`
1252 // returns true.
1253 template <typename OwnerFilter>
1254 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) {
1255 monitors_iterate([&](ObjectMonitor* monitor) {
1256 // This function is only called at a safepoint or when the
1257 // target thread is suspended or when the target thread is
1258 // operating on itself. The current closures in use today are
1259 // only interested in an owned ObjectMonitor and ownership
1260 // cannot be dropped under the calling contexts so the
1261 // ObjectMonitor cannot be async deflated.
1262 if (monitor->has_owner() && filter(monitor)) {
1263 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating");
1264
1265 closure->do_monitor(monitor);
1266 }
1267 });
1268 }
1269
1270 // Iterate ObjectMonitors where the owner == thread; this does NOT include
1271 // ObjectMonitors where owner is set to a stack-lock address in thread.
1272 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) {
1273 int64_t key = ObjectMonitor::owner_id_from(thread);
1274 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; };
1275 return owned_monitors_iterate_filtered(closure, thread_filter);
1276 }
1277
1278 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, oop vthread) {
1279 int64_t key = ObjectMonitor::owner_id_from(vthread);
1280 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; };
1281 return owned_monitors_iterate_filtered(closure, thread_filter);
1282 }
1283
1284 // Iterate ObjectMonitors owned by any thread.
1285 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) {
1286 auto all_filter = [&](ObjectMonitor* monitor) { return true; };
1287 return owned_monitors_iterate_filtered(closure, all_filter);
1288 }
1289
1290 static bool monitors_used_above_threshold(MonitorList* list) {
1291 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy
1292 return false;
1293 }
1294 size_t monitors_used = list->count();
1295 if (monitors_used == 0) { // empty list is easy
1296 return false;
1297 }
1298 size_t old_ceiling = ObjectSynchronizer::in_use_list_ceiling();
1299 // Make sure that we use a ceiling value that is not lower than
1300 // previous, not lower than the recorded max used by the system, and
1301 // not lower than the current number of monitors in use (which can
1302 // race ahead of max). The result is guaranteed > 0.
1303 size_t ceiling = MAX3(old_ceiling, list->max(), monitors_used);
1304
1305 // Check if our monitor usage is above the threshold:
1306 size_t monitor_usage = (monitors_used * 100LL) / ceiling;
1307 if (int(monitor_usage) > MonitorUsedDeflationThreshold) {
1308 // Deflate monitors if over the threshold percentage, unless no
1309 // progress on previous deflations.
1310 bool is_above_threshold = true;
1311
1312 // Check if it's time to adjust the in_use_list_ceiling up, due
1313 // to too many async deflation attempts without any progress.
1314 if (NoAsyncDeflationProgressMax != 0 &&
1315 _no_progress_cnt >= NoAsyncDeflationProgressMax) {
1316 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0;
1317 size_t delta = (size_t)(ceiling * remainder) + 1;
1318 size_t new_ceiling = (ceiling > SIZE_MAX - delta)
1319 ? SIZE_MAX // Overflow, let's clamp new_ceiling.
1320 : ceiling + delta;
1321
1322 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling);
1323 log_info(monitorinflation)("Too many deflations without progress; "
1324 "bumping in_use_list_ceiling from %zu"
1325 " to %zu", old_ceiling, new_ceiling);
1326 _no_progress_cnt = 0;
1327 ceiling = new_ceiling;
1328
1329 // Check if our monitor usage is still above the threshold:
1330 monitor_usage = (monitors_used * 100LL) / ceiling;
1331 is_above_threshold = int(monitor_usage) > MonitorUsedDeflationThreshold;
1332 }
1333 log_info(monitorinflation)("monitors_used=%zu, ceiling=%zu"
1334 ", monitor_usage=%zu, threshold=%d",
1335 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold);
1336 return is_above_threshold;
1337 }
1338
1339 return false;
1340 }
1341
1342 size_t ObjectSynchronizer::in_use_list_count() {
1343 return _in_use_list.count();
1344 }
1345
1346 size_t ObjectSynchronizer::in_use_list_max() {
1347 return _in_use_list.max();
1348 }
1349
1350 size_t ObjectSynchronizer::in_use_list_ceiling() {
1351 return _in_use_list_ceiling;
1352 }
1353
1354 void ObjectSynchronizer::dec_in_use_list_ceiling() {
1355 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1356 }
1357
1358 void ObjectSynchronizer::inc_in_use_list_ceiling() {
1359 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1360 }
1361
1362 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) {
1363 _in_use_list_ceiling = new_value;
1364 }
1365
1366 bool ObjectSynchronizer::is_async_deflation_needed() {
1367 if (is_async_deflation_requested()) {
1368 // Async deflation request.
1369 log_info(monitorinflation)("Async deflation needed: explicit request");
1370 return true;
1371 }
1372
1373 jlong time_since_last = time_since_last_async_deflation_ms();
1374
1375 if (AsyncDeflationInterval > 0 &&
1376 time_since_last > AsyncDeflationInterval &&
1377 monitors_used_above_threshold(&_in_use_list)) {
1378 // It's been longer than our specified deflate interval and there
1379 // are too many monitors in use. We don't deflate more frequently
1380 // than AsyncDeflationInterval (unless is_async_deflation_requested)
1381 // in order to not swamp the MonitorDeflationThread.
1382 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold");
1383 return true;
1384 }
1385
1386 if (GuaranteedAsyncDeflationInterval > 0 &&
1387 time_since_last > GuaranteedAsyncDeflationInterval) {
1388 // It's been longer than our specified guaranteed deflate interval.
1389 // We need to clean up the used monitors even if the threshold is
1390 // not reached, to keep the memory utilization at bay when many threads
1391 // touched many monitors.
1392 log_info(monitorinflation)("Async deflation needed: guaranteed interval (%zd ms) "
1393 "is greater than time since last deflation (" JLONG_FORMAT " ms)",
1394 GuaranteedAsyncDeflationInterval, time_since_last);
1395
1396 // If this deflation has no progress, then it should not affect the no-progress
1397 // tracking, otherwise threshold heuristics would think it was triggered, experienced
1398 // no progress, and needs to backoff more aggressively. In this "no progress" case,
1399 // the generic code would bump the no-progress counter, and we compensate for that
1400 // by telling it to skip the update.
1401 //
1402 // If this deflation has progress, then it should let non-progress tracking
1403 // know about this, otherwise the threshold heuristics would kick in, potentially
1404 // experience no-progress due to aggressive cleanup by this deflation, and think
1405 // it is still in no-progress stride. In this "progress" case, the generic code would
1406 // zero the counter, and we allow it to happen.
1407 _no_progress_skip_increment = true;
1408
1409 return true;
1410 }
1411
1412 return false;
1413 }
1414
1415 void ObjectSynchronizer::request_deflate_idle_monitors() {
1416 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag);
1417 set_is_async_deflation_requested(true);
1418 ml.notify_all();
1419 }
1420
1421 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() {
1422 JavaThread* current = JavaThread::current();
1423 bool ret_code = false;
1424
1425 jlong last_time = last_async_deflation_time_ns();
1426
1427 request_deflate_idle_monitors();
1428
1429 const int N_CHECKS = 5;
1430 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds
1431 if (last_async_deflation_time_ns() > last_time) {
1432 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i);
1433 ret_code = true;
1434 break;
1435 }
1436 {
1437 // JavaThread has to honor the blocking protocol.
1438 ThreadBlockInVM tbivm(current);
1439 os::naked_short_sleep(999); // sleep for almost 1 second
1440 }
1441 }
1442 if (!ret_code) {
1443 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS);
1444 }
1445
1446 return ret_code;
1447 }
1448
1449 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() {
1450 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS);
1451 }
1452
1453 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1454 const oop obj,
1455 ObjectSynchronizer::InflateCause cause) {
1456 assert(event != nullptr, "invariant");
1457 const Klass* monitor_klass = obj->klass();
1458 if (ObjectMonitor::is_jfr_excluded(monitor_klass)) {
1459 return;
1460 }
1461 event->set_monitorClass(monitor_klass);
1462 event->set_address((uintptr_t)(void*)obj);
1463 event->set_cause((u1)cause);
1464 event->commit();
1465 }
1466
1467 // Fast path code shared by multiple functions
1468 void ObjectSynchronizer::inflate_helper(oop obj) {
1469 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter");
1470 markWord mark = obj->mark_acquire();
1471 if (mark.has_monitor()) {
1472 ObjectMonitor* monitor = read_monitor(mark);
1473 markWord dmw = monitor->header();
1474 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value());
1475 return;
1476 }
1477 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal);
1478 }
1479
1480 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) {
1481 assert(current == Thread::current(), "must be");
1482 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter");
1483 return inflate_impl(current->is_Java_thread() ? JavaThread::cast(current) : nullptr, obj, cause);
1484 }
1485
1486 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) {
1487 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be");
1488 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_for");
1489 return inflate_impl(thread, obj, cause);
1490 }
1491
1492 ObjectMonitor* ObjectSynchronizer::inflate_impl(JavaThread* locking_thread, oop object, const InflateCause cause) {
1493 // The JavaThread* locking_thread requires that the locking_thread == Thread::current() or
1494 // is suspended throughout the call by some other mechanism.
1495 // The thread might be nullptr when called from a non JavaThread. (As may still be
1496 // the case from FastHashCode). However it is only important for correctness that the
1497 // thread is set when called from ObjectSynchronizer::enter from the owning thread,
1498 // ObjectSynchronizer::enter_for from any thread, or ObjectSynchronizer::exit.
1499 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_impl");
1500 EventJavaMonitorInflate event;
1501
1502 for (;;) {
1503 const markWord mark = object->mark_acquire();
1504
1505 // The mark can be in one of the following states:
1506 // * inflated - If the ObjectMonitor owner is anonymous and the
1507 // locking_thread owns the object lock, then we
1508 // make the locking_thread the ObjectMonitor owner.
1509 // * stack-locked - Coerce it to inflated from stack-locked.
1510 // * INFLATING - Busy wait for conversion from stack-locked to
1511 // inflated.
1512 // * unlocked - Aggressively inflate the object.
1513
1514 // CASE: inflated
1515 if (mark.has_monitor()) {
1516 ObjectMonitor* inf = mark.monitor();
1517 markWord dmw = inf->header();
1518 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1519 if (inf->has_anonymous_owner() && locking_thread != nullptr) {
1520 assert(LockingMode == LM_LEGACY, "invariant");
1521 if (locking_thread->is_lock_owned((address)inf->stack_locker())) {
1522 inf->set_stack_locker(nullptr);
1523 inf->set_owner_from_anonymous(locking_thread);
1524 }
1525 }
1526 return inf;
1527 }
1528
1529 // CASE: inflation in progress - inflating over a stack-lock.
1530 // Some other thread is converting from stack-locked to inflated.
1531 // Only that thread can complete inflation -- other threads must wait.
1532 // The INFLATING value is transient.
1533 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1534 // We could always eliminate polling by parking the thread on some auxiliary list.
1535 if (mark == markWord::INFLATING()) {
1536 read_stable_mark(object);
1537 continue;
1538 }
1539
1540 // CASE: stack-locked
1541 // Could be stack-locked either by current or by some other thread.
1542 //
1543 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting
1544 // to install INFLATING into the mark word. We originally installed INFLATING,
1545 // allocated the ObjectMonitor, and then finally STed the address of the
1546 // ObjectMonitor into the mark. This was correct, but artificially lengthened
1547 // the interval in which INFLATING appeared in the mark, thus increasing
1548 // the odds of inflation contention. If we lose the race to set INFLATING,
1549 // then we just delete the ObjectMonitor and loop around again.
1550 //
1551 LogStreamHandle(Trace, monitorinflation) lsh;
1552 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1553 ObjectMonitor* m = new ObjectMonitor(object);
1554 // Optimistically prepare the ObjectMonitor - anticipate successful CAS
1555 // We do this before the CAS in order to minimize the length of time
1556 // in which INFLATING appears in the mark.
1557
1558 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark);
1559 if (cmp != mark) {
1560 delete m;
1561 continue; // Interference -- just retry
1562 }
1563
1564 // We've successfully installed INFLATING (0) into the mark-word.
1565 // This is the only case where 0 will appear in a mark-word.
1566 // Only the singular thread that successfully swings the mark-word
1567 // to 0 can perform (or more precisely, complete) inflation.
1568 //
1569 // Why do we CAS a 0 into the mark-word instead of just CASing the
1570 // mark-word from the stack-locked value directly to the new inflated state?
1571 // Consider what happens when a thread unlocks a stack-locked object.
1572 // It attempts to use CAS to swing the displaced header value from the
1573 // on-stack BasicLock back into the object header. Recall also that the
1574 // header value (hash code, etc) can reside in (a) the object header, or
1575 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1576 // header in an ObjectMonitor. The inflate() routine must copy the header
1577 // value from the BasicLock on the owner's stack to the ObjectMonitor, all
1578 // the while preserving the hashCode stability invariants. If the owner
1579 // decides to release the lock while the value is 0, the unlock will fail
1580 // and control will eventually pass from slow_exit() to inflate. The owner
1581 // will then spin, waiting for the 0 value to disappear. Put another way,
1582 // the 0 causes the owner to stall if the owner happens to try to
1583 // drop the lock (restoring the header from the BasicLock to the object)
1584 // while inflation is in-progress. This protocol avoids races that might
1585 // would otherwise permit hashCode values to change or "flicker" for an object.
1586 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable.
1587 // 0 serves as a "BUSY" inflate-in-progress indicator.
1588
1589
1590 // fetch the displaced mark from the owner's stack.
1591 // The owner can't die or unwind past the lock while our INFLATING
1592 // object is in the mark. Furthermore the owner can't complete
1593 // an unlock on the object, either.
1594 markWord dmw = mark.displaced_mark_helper();
1595 // Catch if the object's header is not neutral (not locked and
1596 // not marked is what we care about here).
1597 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1598
1599 // Setup monitor fields to proper values -- prepare the monitor
1600 m->set_header(dmw);
1601
1602 // Note that a thread can inflate an object
1603 // that it has stack-locked -- as might happen in wait() -- directly
1604 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom.
1605 if (locking_thread != nullptr && locking_thread->is_lock_owned((address)mark.locker())) {
1606 m->set_owner(locking_thread);
1607 } else {
1608 // Use ANONYMOUS_OWNER to indicate that the owner is the BasicLock on the stack,
1609 // and set the stack locker field in the monitor.
1610 m->set_stack_locker(mark.locker());
1611 m->set_anonymous_owner();
1612 }
1613 // TODO-FIXME: assert BasicLock->dhw != 0.
1614
1615 // Must preserve store ordering. The monitor state must
1616 // be stable at the time of publishing the monitor address.
1617 guarantee(object->mark() == markWord::INFLATING(), "invariant");
1618 // Release semantics so that above set_object() is seen first.
1619 object->release_set_mark(markWord::encode(m));
1620
1621 // Once ObjectMonitor is configured and the object is associated
1622 // with the ObjectMonitor, it is safe to allow async deflation:
1623 _in_use_list.add(m);
1624
1625 if (log_is_enabled(Trace, monitorinflation)) {
1626 ResourceMark rm;
1627 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1628 INTPTR_FORMAT ", type='%s'", p2i(object),
1629 object->mark().value(), object->klass()->external_name());
1630 }
1631 if (event.should_commit()) {
1632 post_monitor_inflate_event(&event, object, cause);
1633 }
1634 return m;
1635 }
1636
1637 // CASE: unlocked
1638 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1639 // If we know we're inflating for entry it's better to inflate by swinging a
1640 // pre-locked ObjectMonitor pointer into the object header. A successful
1641 // CAS inflates the object *and* confers ownership to the inflating thread.
1642 // In the current implementation we use a 2-step mechanism where we CAS()
1643 // to inflate and then CAS() again to try to swing _owner from null to current.
1644 // An inflateTry() method that we could call from enter() would be useful.
1645
1646 assert(mark.is_unlocked(), "invariant: header=" INTPTR_FORMAT, mark.value());
1647 ObjectMonitor* m = new ObjectMonitor(object);
1648 // prepare m for installation - set monitor to initial state
1649 m->set_header(mark);
1650
1651 if (object->cas_set_mark(markWord::encode(m), mark) != mark) {
1652 delete m;
1653 m = nullptr;
1654 continue;
1655 // interference - the markword changed - just retry.
1656 // The state-transitions are one-way, so there's no chance of
1657 // live-lock -- "Inflated" is an absorbing state.
1658 }
1659
1660 // Once the ObjectMonitor is configured and object is associated
1661 // with the ObjectMonitor, it is safe to allow async deflation:
1662 _in_use_list.add(m);
1663
1664 if (log_is_enabled(Trace, monitorinflation)) {
1665 ResourceMark rm;
1666 lsh.print_cr("inflate(unlocked): object=" INTPTR_FORMAT ", mark="
1667 INTPTR_FORMAT ", type='%s'", p2i(object),
1668 object->mark().value(), object->klass()->external_name());
1669 }
1670 if (event.should_commit()) {
1671 post_monitor_inflate_event(&event, object, cause);
1672 }
1673 return m;
1674 }
1675 }
1676
1677 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle
1678 // ObjectMonitors. Returns the number of deflated ObjectMonitors.
1679 //
1680 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) {
1681 MonitorList::Iterator iter = _in_use_list.iterator();
1682 size_t deflated_count = 0;
1683 Thread* current = Thread::current();
1684
1685 while (iter.has_next()) {
1686 if (deflated_count >= (size_t)MonitorDeflationMax) {
1687 break;
1688 }
1689 ObjectMonitor* mid = iter.next();
1690 if (mid->deflate_monitor(current)) {
1691 deflated_count++;
1692 }
1693
1694 // Must check for a safepoint/handshake and honor it.
1695 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count);
1696 }
1697
1698 return deflated_count;
1699 }
1700
1701 class HandshakeForDeflation : public HandshakeClosure {
1702 public:
1703 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {}
1704
1705 void do_thread(Thread* thread) {
1706 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread="
1707 INTPTR_FORMAT, p2i(thread));
1708 if (thread->is_Java_thread()) {
1709 // Clear OM cache
1710 JavaThread* jt = JavaThread::cast(thread);
1711 jt->om_clear_monitor_cache();
1712 }
1713 }
1714 };
1715
1716 class VM_RendezvousGCThreads : public VM_Operation {
1717 public:
1718 bool evaluate_at_safepoint() const override { return false; }
1719 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; }
1720 void doit() override {
1721 Universe::heap()->safepoint_synchronize_begin();
1722 Universe::heap()->safepoint_synchronize_end();
1723 };
1724 };
1725
1726 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list,
1727 ObjectMonitorDeflationSafepointer* safepointer) {
1728 NativeHeapTrimmer::SuspendMark sm("monitor deletion");
1729 size_t deleted_count = 0;
1730 for (ObjectMonitor* monitor: *delete_list) {
1731 delete monitor;
1732 deleted_count++;
1733 // A JavaThread must check for a safepoint/handshake and honor it.
1734 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count);
1735 }
1736 return deleted_count;
1737 }
1738
1739 class ObjectMonitorDeflationLogging: public StackObj {
1740 LogStreamHandle(Debug, monitorinflation) _debug;
1741 LogStreamHandle(Info, monitorinflation) _info;
1742 LogStream* _stream;
1743 elapsedTimer _timer;
1744
1745 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); }
1746 size_t count() const { return ObjectSynchronizer::in_use_list_count(); }
1747 size_t max() const { return ObjectSynchronizer::in_use_list_max(); }
1748
1749 public:
1750 ObjectMonitorDeflationLogging()
1751 : _debug(), _info(), _stream(nullptr) {
1752 if (_debug.is_enabled()) {
1753 _stream = &_debug;
1754 } else if (_info.is_enabled()) {
1755 _stream = &_info;
1756 }
1757 }
1758
1759 void begin() {
1760 if (_stream != nullptr) {
1761 _stream->print_cr("begin deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu",
1762 ceiling(), count(), max());
1763 _timer.start();
1764 }
1765 }
1766
1767 void before_handshake(size_t unlinked_count) {
1768 if (_stream != nullptr) {
1769 _timer.stop();
1770 _stream->print_cr("before handshaking: unlinked_count=%zu"
1771 ", in_use_list stats: ceiling=%zu, count="
1772 "%zu, max=%zu",
1773 unlinked_count, ceiling(), count(), max());
1774 }
1775 }
1776
1777 void after_handshake() {
1778 if (_stream != nullptr) {
1779 _stream->print_cr("after handshaking: in_use_list stats: ceiling="
1780 "%zu, count=%zu, max=%zu",
1781 ceiling(), count(), max());
1782 _timer.start();
1783 }
1784 }
1785
1786 void end(size_t deflated_count, size_t unlinked_count) {
1787 if (_stream != nullptr) {
1788 _timer.stop();
1789 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) {
1790 _stream->print_cr("deflated_count=%zu, {unlinked,deleted}_count=%zu monitors in %3.7f secs",
1791 deflated_count, unlinked_count, _timer.seconds());
1792 }
1793 _stream->print_cr("end deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu",
1794 ceiling(), count(), max());
1795 }
1796 }
1797
1798 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) {
1799 if (_stream != nullptr) {
1800 _timer.stop();
1801 _stream->print_cr("pausing %s: %s=%zu, in_use_list stats: ceiling="
1802 "%zu, count=%zu, max=%zu",
1803 op_name, cnt_name, cnt, ceiling(), count(), max());
1804 }
1805 }
1806
1807 void after_block_for_safepoint(const char* op_name) {
1808 if (_stream != nullptr) {
1809 _stream->print_cr("resuming %s: in_use_list stats: ceiling=%zu"
1810 ", count=%zu, max=%zu", op_name,
1811 ceiling(), count(), max());
1812 _timer.start();
1813 }
1814 }
1815 };
1816
1817 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) {
1818 if (!SafepointMechanism::should_process(_current)) {
1819 return;
1820 }
1821
1822 // A safepoint/handshake has started.
1823 _log->before_block_for_safepoint(op_name, count_name, counter);
1824
1825 {
1826 // Honor block request.
1827 ThreadBlockInVM tbivm(_current);
1828 }
1829
1830 _log->after_block_for_safepoint(op_name);
1831 }
1832
1833 // This function is called by the MonitorDeflationThread to deflate
1834 // ObjectMonitors.
1835 size_t ObjectSynchronizer::deflate_idle_monitors() {
1836 JavaThread* current = JavaThread::current();
1837 assert(current->is_monitor_deflation_thread(), "The only monitor deflater");
1838
1839 // The async deflation request has been processed.
1840 _last_async_deflation_time_ns = os::javaTimeNanos();
1841 set_is_async_deflation_requested(false);
1842
1843 ObjectMonitorDeflationLogging log;
1844 ObjectMonitorDeflationSafepointer safepointer(current, &log);
1845
1846 log.begin();
1847
1848 // Deflate some idle ObjectMonitors.
1849 size_t deflated_count = deflate_monitor_list(&safepointer);
1850
1851 // Unlink the deflated ObjectMonitors from the in-use list.
1852 size_t unlinked_count = 0;
1853 size_t deleted_count = 0;
1854 if (deflated_count > 0) {
1855 ResourceMark rm(current);
1856 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count);
1857 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer);
1858
1859 #ifdef ASSERT
1860 if (UseObjectMonitorTable) {
1861 for (ObjectMonitor* monitor : delete_list) {
1862 assert(!LightweightSynchronizer::contains_monitor(current, monitor), "Should have been removed");
1863 }
1864 }
1865 #endif
1866
1867 log.before_handshake(unlinked_count);
1868
1869 // A JavaThread needs to handshake in order to safely free the
1870 // ObjectMonitors that were deflated in this cycle.
1871 HandshakeForDeflation hfd_hc;
1872 Handshake::execute(&hfd_hc);
1873 // Also, we sync and desync GC threads around the handshake, so that they can
1874 // safely read the mark-word and look-through to the object-monitor, without
1875 // being afraid that the object-monitor is going away.
1876 VM_RendezvousGCThreads sync_gc;
1877 VMThread::execute(&sync_gc);
1878
1879 log.after_handshake();
1880
1881 // After the handshake, safely free the ObjectMonitors that were
1882 // deflated and unlinked in this cycle.
1883
1884 // Delete the unlinked ObjectMonitors.
1885 deleted_count = delete_monitors(&delete_list, &safepointer);
1886 assert(unlinked_count == deleted_count, "must be");
1887 }
1888
1889 log.end(deflated_count, unlinked_count);
1890
1891 GVars.stw_random = os::random();
1892
1893 if (deflated_count != 0) {
1894 _no_progress_cnt = 0;
1895 } else if (_no_progress_skip_increment) {
1896 _no_progress_skip_increment = false;
1897 } else {
1898 _no_progress_cnt++;
1899 }
1900
1901 return deflated_count;
1902 }
1903
1904 // Monitor cleanup on JavaThread::exit
1905
1906 // Iterate through monitor cache and attempt to release thread's monitors
1907 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1908 private:
1909 JavaThread* _thread;
1910
1911 public:
1912 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {}
1913 void do_monitor(ObjectMonitor* mid) {
1914 intx rec = mid->complete_exit(_thread);
1915 _thread->dec_held_monitor_count(rec + 1);
1916 }
1917 };
1918
1919 // Release all inflated monitors owned by current thread. Lightweight monitors are
1920 // ignored. This is meant to be called during JNI thread detach which assumes
1921 // all remaining monitors are heavyweight. All exceptions are swallowed.
1922 // Scanning the extant monitor list can be time consuming.
1923 // A simple optimization is to add a per-thread flag that indicates a thread
1924 // called jni_monitorenter() during its lifetime.
1925 //
1926 // Instead of NoSafepointVerifier it might be cheaper to
1927 // use an idiom of the form:
1928 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1929 // <code that must not run at safepoint>
1930 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1931 // Since the tests are extremely cheap we could leave them enabled
1932 // for normal product builds.
1933
1934 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) {
1935 assert(current == JavaThread::current(), "must be current Java thread");
1936 NoSafepointVerifier nsv;
1937 ReleaseJavaMonitorsClosure rjmc(current);
1938 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current);
1939 assert(!current->has_pending_exception(), "Should not be possible");
1940 current->clear_pending_exception();
1941 assert(current->held_monitor_count() == 0, "Should not be possible");
1942 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count.
1943 current->clear_jni_monitor_count();
1944 }
1945
1946 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1947 switch (cause) {
1948 case inflate_cause_vm_internal: return "VM Internal";
1949 case inflate_cause_monitor_enter: return "Monitor Enter";
1950 case inflate_cause_wait: return "Monitor Wait";
1951 case inflate_cause_notify: return "Monitor Notify";
1952 case inflate_cause_hash_code: return "Monitor Hash Code";
1953 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1954 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1955 default:
1956 ShouldNotReachHere();
1957 }
1958 return "Unknown";
1959 }
1960
1961 //------------------------------------------------------------------------------
1962 // Debugging code
1963
1964 u_char* ObjectSynchronizer::get_gvars_addr() {
1965 return (u_char*)&GVars;
1966 }
1967
1968 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() {
1969 return (u_char*)&GVars.hc_sequence;
1970 }
1971
1972 size_t ObjectSynchronizer::get_gvars_size() {
1973 return sizeof(SharedGlobals);
1974 }
1975
1976 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() {
1977 return (u_char*)&GVars.stw_random;
1978 }
1979
1980 // Do the final audit and print of ObjectMonitor stats; must be done
1981 // by the VMThread at VM exit time.
1982 void ObjectSynchronizer::do_final_audit_and_print_stats() {
1983 assert(Thread::current()->is_VM_thread(), "sanity check");
1984
1985 if (is_final_audit()) { // Only do the audit once.
1986 return;
1987 }
1988 set_is_final_audit();
1989 log_info(monitorinflation)("Starting the final audit.");
1990
1991 if (log_is_enabled(Info, monitorinflation)) {
1992 LogStreamHandle(Info, monitorinflation) ls;
1993 audit_and_print_stats(&ls, true /* on_exit */);
1994 }
1995 }
1996
1997 // This function can be called by the MonitorDeflationThread or it can be called when
1998 // we are trying to exit the VM. The list walker functions can run in parallel with
1999 // the other list operations.
2000 // Calls to this function can be added in various places as a debugging
2001 // aid.
2002 //
2003 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) {
2004 int error_cnt = 0;
2005
2006 ls->print_cr("Checking in_use_list:");
2007 chk_in_use_list(ls, &error_cnt);
2008
2009 if (error_cnt == 0) {
2010 ls->print_cr("No errors found in in_use_list checks.");
2011 } else {
2012 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt);
2013 }
2014
2015 // When exiting, only log the interesting entries at the Info level.
2016 // When called at intervals by the MonitorDeflationThread, log output
2017 // at the Trace level since there can be a lot of it.
2018 if (!on_exit && log_is_enabled(Trace, monitorinflation)) {
2019 LogStreamHandle(Trace, monitorinflation) ls_tr;
2020 log_in_use_monitor_details(&ls_tr, true /* log_all */);
2021 } else if (on_exit) {
2022 log_in_use_monitor_details(ls, false /* log_all */);
2023 }
2024
2025 ls->flush();
2026
2027 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt);
2028 }
2029
2030 // Check the in_use_list; log the results of the checks.
2031 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) {
2032 size_t l_in_use_count = _in_use_list.count();
2033 size_t l_in_use_max = _in_use_list.max();
2034 out->print_cr("count=%zu, max=%zu", l_in_use_count,
2035 l_in_use_max);
2036
2037 size_t ck_in_use_count = 0;
2038 MonitorList::Iterator iter = _in_use_list.iterator();
2039 while (iter.has_next()) {
2040 ObjectMonitor* mid = iter.next();
2041 chk_in_use_entry(mid, out, error_cnt_p);
2042 ck_in_use_count++;
2043 }
2044
2045 if (l_in_use_count == ck_in_use_count) {
2046 out->print_cr("in_use_count=%zu equals ck_in_use_count=%zu",
2047 l_in_use_count, ck_in_use_count);
2048 } else {
2049 out->print_cr("WARNING: in_use_count=%zu is not equal to "
2050 "ck_in_use_count=%zu", l_in_use_count,
2051 ck_in_use_count);
2052 }
2053
2054 size_t ck_in_use_max = _in_use_list.max();
2055 if (l_in_use_max == ck_in_use_max) {
2056 out->print_cr("in_use_max=%zu equals ck_in_use_max=%zu",
2057 l_in_use_max, ck_in_use_max);
2058 } else {
2059 out->print_cr("WARNING: in_use_max=%zu is not equal to "
2060 "ck_in_use_max=%zu", l_in_use_max, ck_in_use_max);
2061 }
2062 }
2063
2064 // Check an in-use monitor entry; log any errors.
2065 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out,
2066 int* error_cnt_p) {
2067 if (n->owner_is_DEFLATER_MARKER()) {
2068 // This could happen when monitor deflation blocks for a safepoint.
2069 return;
2070 }
2071
2072
2073 if (n->metadata() == 0) {
2074 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must "
2075 "have non-null _metadata (header/hash) field.", p2i(n));
2076 *error_cnt_p = *error_cnt_p + 1;
2077 }
2078
2079 const oop obj = n->object_peek();
2080 if (obj == nullptr) {
2081 return;
2082 }
2083
2084 const markWord mark = obj->mark();
2085 if (!mark.has_monitor()) {
2086 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2087 "object does not think it has a monitor: obj="
2088 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n),
2089 p2i(obj), mark.value());
2090 *error_cnt_p = *error_cnt_p + 1;
2091 return;
2092 }
2093
2094 ObjectMonitor* const obj_mon = read_monitor(Thread::current(), obj, mark);
2095 if (n != obj_mon) {
2096 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2097 "object does not refer to the same monitor: obj="
2098 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon="
2099 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon));
2100 *error_cnt_p = *error_cnt_p + 1;
2101 }
2102 }
2103
2104 // Log details about ObjectMonitors on the in_use_list. The 'BHL'
2105 // flags indicate why the entry is in-use, 'object' and 'object type'
2106 // indicate the associated object and its type.
2107 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) {
2108 if (_in_use_list.count() > 0) {
2109 stringStream ss;
2110 out->print_cr("In-use monitor info%s:", log_all ? "" : " (eliding idle monitors)");
2111 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2112 out->print_cr("%18s %s %18s %18s",
2113 "monitor", "BHL", "object", "object type");
2114 out->print_cr("================== === ================== ==================");
2115
2116 auto is_interesting = [&](ObjectMonitor* monitor) {
2117 return log_all || monitor->has_owner() || monitor->is_busy();
2118 };
2119
2120 monitors_iterate([&](ObjectMonitor* monitor) {
2121 if (is_interesting(monitor)) {
2122 const oop obj = monitor->object_peek();
2123 const intptr_t hash = UseObjectMonitorTable ? monitor->hash() : monitor->header().hash();
2124 ResourceMark rm;
2125 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor),
2126 monitor->is_busy(), hash != 0, monitor->has_owner(),
2127 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name());
2128 if (monitor->is_busy()) {
2129 out->print(" (%s)", monitor->is_busy_to_string(&ss));
2130 ss.reset();
2131 }
2132 out->cr();
2133 }
2134 });
2135 }
2136
2137 out->flush();
2138 }