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 "gc/shared/collectedHeap.hpp"
28 #include "jfr/jfrEvents.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/basicLock.inline.hpp"
39 #include "runtime/frame.inline.hpp"
40 #include "runtime/globals.hpp"
41 #include "runtime/handles.inline.hpp"
42 #include "runtime/handshake.hpp"
43 #include "runtime/interfaceSupport.inline.hpp"
44 #include "runtime/javaThread.hpp"
45 #include "runtime/lightweightSynchronizer.hpp"
46 #include "runtime/lockStack.inline.hpp"
47 #include "runtime/mutexLocker.hpp"
48 #include "runtime/objectMonitor.hpp"
49 #include "runtime/objectMonitor.inline.hpp"
50 #include "runtime/os.inline.hpp"
51 #include "runtime/osThread.hpp"
52 #include "runtime/perfData.hpp"
53 #include "runtime/safepointMechanism.inline.hpp"
54 #include "runtime/safepointVerifiers.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #include "runtime/stubRoutines.hpp"
57 #include "runtime/synchronizer.inline.hpp"
58 #include "runtime/threads.hpp"
59 #include "runtime/timer.hpp"
60 #include "runtime/trimNativeHeap.hpp"
61 #include "runtime/vframe.hpp"
62 #include "runtime/vmThread.hpp"
63 #include "utilities/align.hpp"
64 #include "utilities/dtrace.hpp"
65 #include "utilities/events.hpp"
66 #include "utilities/globalDefinitions.hpp"
67 #include "utilities/fastHash.hpp"
68 #include "utilities/linkedlist.hpp"
69 #include "utilities/preserveException.hpp"
70
71 class ObjectMonitorDeflationLogging;
72
73 void MonitorList::add(ObjectMonitor* m) {
74 ObjectMonitor* head;
75 do {
76 head = Atomic::load(&_head);
77 m->set_next_om(head);
78 } while (Atomic::cmpxchg(&_head, head, m) != head);
79
80 size_t count = Atomic::add(&_count, 1u);
81 if (count > max()) {
82 Atomic::inc(&_max);
83 }
84 }
85
86 size_t MonitorList::count() const {
87 return Atomic::load(&_count);
88 }
89
90 size_t MonitorList::max() const {
91 return Atomic::load(&_max);
92 }
93
94 class ObjectMonitorDeflationSafepointer : public StackObj {
95 JavaThread* const _current;
96 ObjectMonitorDeflationLogging* const _log;
97
98 public:
99 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log)
100 : _current(current), _log(log) {}
101
102 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter);
103 };
104
105 // Walk the in-use list and unlink deflated ObjectMonitors.
106 // Returns the number of unlinked ObjectMonitors.
107 size_t MonitorList::unlink_deflated(size_t deflated_count,
108 GrowableArray<ObjectMonitor*>* unlinked_list,
109 ObjectMonitorDeflationSafepointer* safepointer) {
110 size_t unlinked_count = 0;
111 ObjectMonitor* prev = nullptr;
112 ObjectMonitor* m = Atomic::load_acquire(&_head);
113
114 while (m != nullptr) {
115 if (m->is_being_async_deflated()) {
116 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can
117 // modify the list once per batch. The batch starts at "m".
118 size_t unlinked_batch = 0;
119 ObjectMonitor* next = m;
120 // Look for at most MonitorUnlinkBatch monitors, or the number of
121 // deflated and not unlinked monitors, whatever comes first.
122 assert(deflated_count >= unlinked_count, "Sanity: underflow");
123 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch);
124 do {
125 ObjectMonitor* next_next = next->next_om();
126 unlinked_batch++;
127 unlinked_list->append(next);
128 next = next_next;
129 if (unlinked_batch >= unlinked_batch_limit) {
130 // Reached the max batch, so bail out of the gathering loop.
131 break;
132 }
133 if (prev == nullptr && Atomic::load(&_head) != m) {
134 // Current batch used to be at head, but it is not at head anymore.
135 // Bail out and figure out where we currently are. This avoids long
136 // walks searching for new prev during unlink under heavy list inserts.
137 break;
138 }
139 } while (next != nullptr && next->is_being_async_deflated());
140
141 // Unlink the found batch.
142 if (prev == nullptr) {
143 // The current batch is the first batch, so there is a chance that it starts at head.
144 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head.
145 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, m, next);
146 if (prev_head != m) {
147 // Something must have updated the head. Figure out the actual prev for this batch.
148 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) {
149 prev = n;
150 }
151 assert(prev != nullptr, "Should have found the prev for the current batch");
152 prev->set_next_om(next);
153 }
154 } else {
155 // The current batch is preceded by another batch. This guarantees the current batch
156 // does not start at head. Unlink the entire current batch without updating the head.
157 assert(Atomic::load(&_head) != m, "Sanity");
158 prev->set_next_om(next);
159 }
160
161 unlinked_count += unlinked_batch;
162 if (unlinked_count >= deflated_count) {
163 // Reached the max so bail out of the searching loop.
164 // There should be no more deflated monitors left.
165 break;
166 }
167 m = next;
168 } else {
169 prev = m;
170 m = m->next_om();
171 }
172
173 // Must check for a safepoint/handshake and honor it.
174 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count);
175 }
176
177 #ifdef ASSERT
178 // Invariant: the code above should unlink all deflated monitors.
179 // The code that runs after this unlinking does not expect deflated monitors.
180 // Notably, attempting to deflate the already deflated monitor would break.
181 {
182 ObjectMonitor* m = Atomic::load_acquire(&_head);
183 while (m != nullptr) {
184 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked");
185 m = m->next_om();
186 }
187 }
188 #endif
189
190 Atomic::sub(&_count, unlinked_count);
191 return unlinked_count;
192 }
193
194 MonitorList::Iterator MonitorList::iterator() const {
195 return Iterator(Atomic::load_acquire(&_head));
196 }
197
198 ObjectMonitor* MonitorList::Iterator::next() {
199 ObjectMonitor* current = _current;
200 _current = current->next_om();
201 return current;
202 }
203
204 // The "core" versions of monitor enter and exit reside in this file.
205 // The interpreter and compilers contain specialized transliterated
206 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp
207 // fast_lock(...) for instance. If you make changes here, make sure to modify the
208 // interpreter, and both C1 and C2 fast-path inline locking code emission.
209 //
210 // -----------------------------------------------------------------------------
211
212 #ifdef DTRACE_ENABLED
213
214 // Only bother with this argument setup if dtrace is available
215 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
216
217 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
218 char* bytes = nullptr; \
219 int len = 0; \
220 jlong jtid = SharedRuntime::get_java_tid(thread); \
221 Symbol* klassname = obj->klass()->name(); \
222 if (klassname != nullptr) { \
223 bytes = (char*)klassname->bytes(); \
224 len = klassname->utf8_length(); \
225 }
226
227 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
228 { \
229 if (DTraceMonitorProbes) { \
230 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
231 HOTSPOT_MONITOR_WAIT(jtid, \
232 (uintptr_t)(monitor), bytes, len, (millis)); \
233 } \
234 }
235
236 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY
237 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL
238 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
239
240 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
241 { \
242 if (DTraceMonitorProbes) { \
243 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
244 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
245 (uintptr_t)(monitor), bytes, len); \
246 } \
247 }
248
249 #else // ndef DTRACE_ENABLED
250
251 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
252 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
253
254 #endif // ndef DTRACE_ENABLED
255
256 // This exists only as a workaround of dtrace bug 6254741
257 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) {
258 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
259 return 0;
260 }
261
262 static constexpr size_t inflation_lock_count() {
263 return 256;
264 }
265
266 // Static storage for an array of PlatformMutex.
267 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)];
268
269 static inline PlatformMutex* inflation_lock(size_t index) {
270 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]);
271 }
272
273 void ObjectSynchronizer::initialize() {
274 for (size_t i = 0; i < inflation_lock_count(); i++) {
275 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex();
276 }
277 // Start the ceiling with the estimate for one thread.
278 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate);
279
280 // Start the timer for deflations, so it does not trigger immediately.
281 _last_async_deflation_time_ns = os::javaTimeNanos();
282
283 if (LockingMode == LM_LIGHTWEIGHT) {
284 LightweightSynchronizer::initialize();
285 }
286 }
287
288 MonitorList ObjectSynchronizer::_in_use_list;
289 // monitors_used_above_threshold() policy is as follows:
290 //
291 // The ratio of the current _in_use_list count to the ceiling is used
292 // to determine if we are above MonitorUsedDeflationThreshold and need
293 // to do an async monitor deflation cycle. The ceiling is increased by
294 // AvgMonitorsPerThreadEstimate when a thread is added to the system
295 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is
296 // removed from the system.
297 //
298 // Note: If the _in_use_list max exceeds the ceiling, then
299 // monitors_used_above_threshold() will use the in_use_list max instead
300 // of the thread count derived ceiling because we have used more
301 // ObjectMonitors than the estimated average.
302 //
303 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax
304 // no-progress async monitor deflation cycles in a row, then the ceiling
305 // is adjusted upwards by monitors_used_above_threshold().
306 //
307 // Start the ceiling with the estimate for one thread in initialize()
308 // which is called after cmd line options are processed.
309 static size_t _in_use_list_ceiling = 0;
310 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false;
311 bool volatile ObjectSynchronizer::_is_final_audit = false;
312 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0;
313 static uintx _no_progress_cnt = 0;
314 static bool _no_progress_skip_increment = false;
315
316 // =====================> Quick functions
317
318 // The quick_* forms are special fast-path variants used to improve
319 // performance. In the simplest case, a "quick_*" implementation could
320 // simply return false, in which case the caller will perform the necessary
321 // state transitions and call the slow-path form.
322 // The fast-path is designed to handle frequently arising cases in an efficient
323 // manner and is just a degenerate "optimistic" variant of the slow-path.
324 // returns true -- to indicate the call was satisfied.
325 // returns false -- to indicate the call needs the services of the slow-path.
326 // A no-loitering ordinance is in effect for code in the quick_* family
327 // operators: safepoints or indefinite blocking (blocking that might span a
328 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon
329 // entry.
330 //
331 // Consider: An interesting optimization is to have the JIT recognize the
332 // following common idiom:
333 // synchronized (someobj) { .... ; notify(); }
334 // That is, we find a notify() or notifyAll() call that immediately precedes
335 // the monitorexit operation. In that case the JIT could fuse the operations
336 // into a single notifyAndExit() runtime primitive.
337
338 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) {
339 assert(current->thread_state() == _thread_in_Java, "invariant");
340 NoSafepointVerifier nsv;
341 if (obj == nullptr) return false; // slow-path for invalid obj
342 const markWord mark = obj->mark();
343
344 if (LockingMode == LM_LIGHTWEIGHT) {
345 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) {
346 // Degenerate notify
347 // fast-locked by caller so by definition the implied waitset is empty.
348 return true;
349 }
350 } else if (LockingMode == LM_LEGACY) {
351 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
352 // Degenerate notify
353 // stack-locked by caller so by definition the implied waitset is empty.
354 return true;
355 }
356 }
357
358 if (mark.has_monitor()) {
359 ObjectMonitor* const mon = read_monitor(current, obj, mark);
360 if (LockingMode == LM_LIGHTWEIGHT && mon == nullptr) {
361 // Racing with inflation/deflation go slow path
362 return false;
363 }
364 assert(mon->object() == oop(obj), "invariant");
365 if (!mon->has_owner(current)) return false; // slow-path for IMS exception
366
367 if (mon->first_waiter() != nullptr) {
368 // We have one or more waiters. Since this is an inflated monitor
369 // that we own, we can transfer one or more threads from the waitset
370 // to the entrylist here and now, avoiding the slow-path.
371 if (all) {
372 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, current);
373 } else {
374 DTRACE_MONITOR_PROBE(notify, mon, obj, current);
375 }
376 int free_count = 0;
377 do {
378 mon->INotify(current);
379 ++free_count;
380 } while (mon->first_waiter() != nullptr && all);
381 OM_PERFDATA_OP(Notifications, inc(free_count));
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 if (LockingMode == LM_LIGHTWEIGHT) {
412 return LightweightSynchronizer::quick_enter(obj, lock, current);
413 }
414
415 assert(LockingMode == LM_LEGACY, "legacy mode below");
416
417 const markWord mark = obj->mark();
418
419 if (mark.has_monitor()) {
420
421 ObjectMonitor* const m = read_monitor(mark);
422 // An async deflation or GC can race us before we manage to make
423 // the ObjectMonitor busy by setting the owner below. If we detect
424 // that race we just bail out to the slow-path here.
425 if (m->object_peek() == nullptr) {
426 return false;
427 }
428
429 // Lock contention and Transactional Lock Elision (TLE) diagnostics
430 // and observability
431 // Case: light contention possibly amenable to TLE
432 // Case: TLE inimical operations such as nested/recursive synchronization
433
434 if (m->has_owner(current)) {
435 m->_recursions++;
436 current->inc_held_monitor_count();
437 return true;
438 }
439
440 // This Java Monitor is inflated so obj's header will never be
441 // displaced to this thread's BasicLock. Make the displaced header
442 // non-null so this BasicLock is not seen as recursive nor as
443 // being locked. We do this unconditionally so that this thread's
444 // BasicLock cannot be mis-interpreted by any stack walkers. For
445 // performance reasons, stack walkers generally first check for
446 // stack-locking in the object's header, the second check is for
447 // recursive stack-locking in the displaced header in the BasicLock,
448 // and last are the inflated Java Monitor (ObjectMonitor) checks.
449 lock->set_displaced_header(markWord::unused_mark());
450
451 if (!m->has_owner() && m->try_set_owner(current)) {
452 assert(m->_recursions == 0, "invariant");
453 current->inc_held_monitor_count();
454 return true;
455 }
456 }
457
458 // Note that we could inflate in quick_enter.
459 // This is likely a useful optimization
460 // Critically, in quick_enter() we must not:
461 // -- block indefinitely, or
462 // -- reach a safepoint
463
464 return false; // revert to slow-path
465 }
466
467 // Handle notifications when synchronizing on value based classes
468 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) {
469 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
470 frame last_frame = locking_thread->last_frame();
471 bool bcp_was_adjusted = false;
472 // Don't decrement bcp if it points to the frame's first instruction. This happens when
473 // handle_sync_on_value_based_class() is called because of a synchronized method. There
474 // is no actual monitorenter instruction in the byte code in this case.
475 if (last_frame.is_interpreted_frame() &&
476 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) {
477 // adjust bcp to point back to monitorenter so that we print the correct line numbers
478 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1);
479 bcp_was_adjusted = true;
480 }
481
482 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) {
483 ResourceMark rm;
484 stringStream ss;
485 locking_thread->print_active_stack_on(&ss);
486 char* base = (char*)strstr(ss.base(), "at");
487 char* newline = (char*)strchr(ss.base(), '\n');
488 if (newline != nullptr) {
489 *newline = '\0';
490 }
491 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base);
492 } else {
493 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses");
494 ResourceMark rm;
495 Log(valuebasedclasses) vblog;
496
497 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name());
498 if (locking_thread->has_last_Java_frame()) {
499 LogStream info_stream(vblog.info());
500 locking_thread->print_active_stack_on(&info_stream);
501 } else {
502 vblog.info("Cannot find the last Java frame");
503 }
504
505 EventSyncOnValueBasedClass event;
506 if (event.should_commit()) {
507 event.set_valueBasedClass(obj->klass());
508 event.commit();
509 }
510 }
511
512 if (bcp_was_adjusted) {
513 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1);
514 }
515 }
516
517 // -----------------------------------------------------------------------------
518 // Monitor Enter/Exit
519
520 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
521 // When called with locking_thread != Thread::current() some mechanism must synchronize
522 // the locking_thread with respect to the current thread. Currently only used when
523 // deoptimizing and re-locking locks. See Deoptimization::relock_objects
524 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
525
526 if (LockingMode == LM_LIGHTWEIGHT) {
527 return LightweightSynchronizer::enter_for(obj, lock, locking_thread);
528 }
529
530 if (!enter_fast_impl(obj, lock, locking_thread)) {
531 // Inflated ObjectMonitor::enter_for is required
532
533 // An async deflation can race after the inflate_for() call and before
534 // enter_for() can make the ObjectMonitor busy. enter_for() returns false
535 // if we have lost the race to async deflation and we simply try again.
536 while (true) {
537 ObjectMonitor* monitor = inflate_for(locking_thread, obj(), inflate_cause_monitor_enter);
538 if (monitor->enter_for(locking_thread)) {
539 return;
540 }
541 assert(monitor->is_being_async_deflated(), "must be");
542 }
543 }
544 }
545
546 void ObjectSynchronizer::enter_legacy(Handle obj, BasicLock* lock, JavaThread* current) {
547 if (!enter_fast_impl(obj, lock, current)) {
548 // Inflated ObjectMonitor::enter is required
549
550 // An async deflation can race after the inflate() call and before
551 // enter() can make the ObjectMonitor busy. enter() returns false if
552 // we have lost the race to async deflation and we simply try again.
553 while (true) {
554 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter);
555 if (monitor->enter(current)) {
556 return;
557 }
558 }
559 }
560 }
561
562 // The interpreter and compiler assembly code tries to lock using the fast path
563 // of this algorithm. Make sure to update that code if the following function is
564 // changed. The implementation is extremely sensitive to race condition. Be careful.
565 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
566 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer");
567
568 if (obj->klass()->is_value_based()) {
569 handle_sync_on_value_based_class(obj, locking_thread);
570 }
571
572 locking_thread->inc_held_monitor_count();
573
574 if (!useHeavyMonitors()) {
575 if (LockingMode == LM_LEGACY) {
576 markWord mark = obj->mark();
577 if (mark.is_unlocked()) {
578 // Anticipate successful CAS -- the ST of the displaced mark must
579 // be visible <= the ST performed by the CAS.
580 lock->set_displaced_header(mark);
581 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) {
582 return true;
583 }
584 } else if (mark.has_locker() &&
585 locking_thread->is_lock_owned((address) mark.locker())) {
586 assert(lock != mark.locker(), "must not re-lock the same lock");
587 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock");
588 lock->set_displaced_header(markWord::from_pointer(nullptr));
589 return true;
590 }
591
592 // The object header will never be displaced to this lock,
593 // so it does not matter what the value is, except that it
594 // must be non-zero to avoid looking like a re-entrant lock,
595 // and must not look locked either.
596 lock->set_displaced_header(markWord::unused_mark());
597
598 // Failed to fast lock.
599 return false;
600 }
601 } else if (VerifyHeavyMonitors) {
602 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
603 }
604
605 return false;
606 }
607
608 void ObjectSynchronizer::exit_legacy(oop object, BasicLock* lock, JavaThread* current) {
609 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer");
610
611 if (!useHeavyMonitors()) {
612 markWord mark = object->mark();
613 if (LockingMode == LM_LEGACY) {
614 markWord dhw = lock->displaced_header();
615 if (dhw.value() == 0) {
616 // If the displaced header is null, then this exit matches up with
617 // a recursive enter. No real work to do here except for diagnostics.
618 #ifndef PRODUCT
619 if (mark != markWord::INFLATING()) {
620 // Only do diagnostics if we are not racing an inflation. Simply
621 // exiting a recursive enter of a Java Monitor that is being
622 // inflated is safe; see the has_monitor() comment below.
623 assert(!mark.is_unlocked(), "invariant");
624 assert(!mark.has_locker() ||
625 current->is_lock_owned((address)mark.locker()), "invariant");
626 if (mark.has_monitor()) {
627 // The BasicLock's displaced_header is marked as a recursive
628 // enter and we have an inflated Java Monitor (ObjectMonitor).
629 // This is a special case where the Java Monitor was inflated
630 // after this thread entered the stack-lock recursively. When a
631 // Java Monitor is inflated, we cannot safely walk the Java
632 // Monitor owner's stack and update the BasicLocks because a
633 // Java Monitor can be asynchronously inflated by a thread that
634 // does not own the Java Monitor.
635 ObjectMonitor* m = read_monitor(mark);
636 assert(m->object()->mark() == mark, "invariant");
637 assert(m->is_entered(current), "invariant");
638 }
639 }
640 #endif
641 return;
642 }
643
644 if (mark == markWord::from_pointer(lock)) {
645 // If the object is stack-locked by the current thread, try to
646 // swing the displaced header from the BasicLock back to the mark.
647 assert(dhw.is_neutral(), "invariant");
648 if (object->cas_set_mark(dhw, mark) == mark) {
649 return;
650 }
651 }
652 }
653 } else if (VerifyHeavyMonitors) {
654 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
655 }
656
657 // We have to take the slow-path of possible inflation and then exit.
658 // The ObjectMonitor* can't be async deflated until ownership is
659 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
660 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal);
661 assert(!monitor->has_anonymous_owner(), "must not be");
662 monitor->exit(current);
663 }
664
665 // -----------------------------------------------------------------------------
666 // JNI locks on java objects
667 // NOTE: must use heavy weight monitor to handle jni monitor enter
668 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) {
669 // Top native frames in the stack will not be seen if we attempt
670 // preemption, since we start walking from the last Java anchor.
671 NoPreemptMark npm(current);
672
673 if (obj->klass()->is_value_based()) {
674 handle_sync_on_value_based_class(obj, current);
675 }
676
677 // the current locking is from JNI instead of Java code
678 current->set_current_pending_monitor_is_from_java(false);
679 // An async deflation can race after the inflate() call and before
680 // enter() can make the ObjectMonitor busy. enter() returns false if
681 // we have lost the race to async deflation and we simply try again.
682 while (true) {
683 ObjectMonitor* monitor;
684 bool entered;
685 if (LockingMode == LM_LIGHTWEIGHT) {
686 entered = LightweightSynchronizer::inflate_and_enter(obj(), inflate_cause_jni_enter, current, current) != nullptr;
687 } else {
688 monitor = inflate(current, obj(), inflate_cause_jni_enter);
689 entered = monitor->enter(current);
690 }
691
692 if (entered) {
693 current->inc_held_monitor_count(1, true);
694 break;
695 }
696 }
697 current->set_current_pending_monitor_is_from_java(true);
698 }
699
700 // NOTE: must use heavy weight monitor to handle jni monitor exit
701 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) {
702 JavaThread* current = THREAD;
703
704 ObjectMonitor* monitor;
705 if (LockingMode == LM_LIGHTWEIGHT) {
706 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj, inflate_cause_jni_exit, CHECK);
707 } else {
708 // The ObjectMonitor* can't be async deflated until ownership is
709 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
710 monitor = inflate(current, obj, inflate_cause_jni_exit);
711 }
712 // If this thread has locked the object, exit the monitor. We
713 // intentionally do not use CHECK on check_owner because we must exit the
714 // monitor even if an exception was already pending.
715 if (monitor->check_owner(THREAD)) {
716 monitor->exit(current);
717 current->dec_held_monitor_count(1, true);
718 }
719 }
720
721 // -----------------------------------------------------------------------------
722 // Internal VM locks on java objects
723 // standard constructor, allows locking failures
724 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) : _npm(thread) {
725 _thread = thread;
726 _thread->check_for_valid_safepoint_state();
727 _obj = obj;
728
729 if (_obj() != nullptr) {
730 ObjectSynchronizer::enter(_obj, &_lock, _thread);
731 }
732 }
733
734 ObjectLocker::~ObjectLocker() {
735 if (_obj() != nullptr) {
736 ObjectSynchronizer::exit(_obj(), &_lock, _thread);
737 }
738 }
739
740
741 // -----------------------------------------------------------------------------
742 // Wait/Notify/NotifyAll
743 // NOTE: must use heavy weight monitor to handle wait()
744
745 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
746 JavaThread* current = THREAD;
747 if (millis < 0) {
748 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
749 }
750
751 ObjectMonitor* monitor;
752 if (LockingMode == LM_LIGHTWEIGHT) {
753 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK_0);
754 } else {
755 // The ObjectMonitor* can't be async deflated because the _waiters
756 // field is incremented before ownership is dropped and decremented
757 // after ownership is regained.
758 monitor = inflate(current, obj(), inflate_cause_wait);
759 }
760
761 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis);
762 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code
763
764 // This dummy call is in place to get around dtrace bug 6254741. Once
765 // that's fixed we can uncomment the following line, remove the call
766 // and change this function back into a "void" func.
767 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
768 int ret_code = dtrace_waited_probe(monitor, obj, THREAD);
769 return ret_code;
770 }
771
772 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
773 if (millis < 0) {
774 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
775 }
776
777 ObjectMonitor* monitor;
778 if (LockingMode == LM_LIGHTWEIGHT) {
779 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK);
780 } else {
781 monitor = inflate(THREAD, obj(), inflate_cause_wait);
782 }
783 monitor->wait(millis, false, THREAD);
784 }
785
786
787 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
788 JavaThread* current = THREAD;
789
790 markWord mark = obj->mark();
791 if (LockingMode == LM_LIGHTWEIGHT) {
792 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
793 // Not inflated so there can't be any waiters to notify.
794 return;
795 }
796 } else if (LockingMode == LM_LEGACY) {
797 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
798 // Not inflated so there can't be any waiters to notify.
799 return;
800 }
801 }
802
803 ObjectMonitor* monitor;
804 if (LockingMode == LM_LIGHTWEIGHT) {
805 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK);
806 } else {
807 // The ObjectMonitor* can't be async deflated until ownership is
808 // dropped by the calling thread.
809 monitor = inflate(current, obj(), inflate_cause_notify);
810 }
811 monitor->notify(CHECK);
812 }
813
814 // NOTE: see comment of notify()
815 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
816 JavaThread* current = THREAD;
817
818 markWord mark = obj->mark();
819 if (LockingMode == LM_LIGHTWEIGHT) {
820 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
821 // Not inflated so there can't be any waiters to notify.
822 return;
823 }
824 } else if (LockingMode == LM_LEGACY) {
825 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
826 // Not inflated so there can't be any waiters to notify.
827 return;
828 }
829 }
830
831 ObjectMonitor* monitor;
832 if (LockingMode == LM_LIGHTWEIGHT) {
833 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK);
834 } else {
835 // The ObjectMonitor* can't be async deflated until ownership is
836 // dropped by the calling thread.
837 monitor = inflate(current, obj(), inflate_cause_notify);
838 }
839 monitor->notifyAll(CHECK);
840 }
841
842 // -----------------------------------------------------------------------------
843 // Hash Code handling
844
845 struct SharedGlobals {
846 char _pad_prefix[OM_CACHE_LINE_SIZE];
847 // This is a highly shared mostly-read variable.
848 // To avoid false-sharing it needs to be the sole occupant of a cache line.
849 volatile int stw_random;
850 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int));
851 // Hot RW variable -- Sequester to avoid false-sharing
852 volatile int hc_sequence;
853 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int));
854 };
855
856 static SharedGlobals GVars;
857
858 static markWord read_stable_mark(oop obj) {
859 markWord mark = obj->mark_acquire();
860 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) {
861 // New lightweight locking does not use the markWord::INFLATING() protocol.
862 return mark; // normal fast-path return
863 }
864
865 int its = 0;
866 for (;;) {
867 markWord mark = obj->mark_acquire();
868 if (!mark.is_being_inflated()) {
869 return mark; // normal fast-path return
870 }
871
872 // The object is being inflated by some other thread.
873 // The caller of read_stable_mark() must wait for inflation to complete.
874 // Avoid live-lock.
875
876 ++its;
877 if (its > 10000 || !os::is_MP()) {
878 if (its & 1) {
879 os::naked_yield();
880 } else {
881 // Note that the following code attenuates the livelock problem but is not
882 // a complete remedy. A more complete solution would require that the inflating
883 // thread hold the associated inflation lock. The following code simply restricts
884 // the number of spinners to at most one. We'll have N-2 threads blocked
885 // on the inflationlock, 1 thread holding the inflation lock and using
886 // a yield/park strategy, and 1 thread in the midst of inflation.
887 // A more refined approach would be to change the encoding of INFLATING
888 // to allow encapsulation of a native thread pointer. Threads waiting for
889 // inflation to complete would use CAS to push themselves onto a singly linked
890 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
891 // and calling park(). When inflation was complete the thread that accomplished inflation
892 // would detach the list and set the markword to inflated with a single CAS and
893 // then for each thread on the list, set the flag and unpark() the thread.
894
895 // Index into the lock array based on the current object address.
896 static_assert(is_power_of_2(inflation_lock_count()), "must be");
897 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1);
898 int YieldThenBlock = 0;
899 assert(ix < inflation_lock_count(), "invariant");
900 inflation_lock(ix)->lock();
901 while (obj->mark_acquire() == markWord::INFLATING()) {
902 // Beware: naked_yield() is advisory and has almost no effect on some platforms
903 // so we periodically call current->_ParkEvent->park(1).
904 // We use a mixed spin/yield/block mechanism.
905 if ((YieldThenBlock++) >= 16) {
906 Thread::current()->_ParkEvent->park(1);
907 } else {
908 os::naked_yield();
909 }
910 }
911 inflation_lock(ix)->unlock();
912 }
913 } else {
914 SpinPause(); // SMP-polite spinning
915 }
916 }
917 }
918
919 // hashCode() generation :
920 //
921 // Possibilities:
922 // * MD5Digest of {obj,stw_random}
923 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function.
924 // * A DES- or AES-style SBox[] mechanism
925 // * One of the Phi-based schemes, such as:
926 // 2654435761 = 2^32 * Phi (golden ratio)
927 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ;
928 // * A variation of Marsaglia's shift-xor RNG scheme.
929 // * (obj ^ stw_random) is appealing, but can result
930 // in undesirable regularity in the hashCode values of adjacent objects
931 // (objects allocated back-to-back, in particular). This could potentially
932 // result in hashtable collisions and reduced hashtable efficiency.
933 // There are simple ways to "diffuse" the middle address bits over the
934 // generated hashCode values:
935
936 intptr_t ObjectSynchronizer::get_next_hash(Thread* current, oop obj) {
937 intptr_t value = 0;
938 if (hashCode == 0) {
939 // This form uses global Park-Miller RNG.
940 // On MP system we'll have lots of RW access to a global, so the
941 // mechanism induces lots of coherency traffic.
942 value = os::random();
943 } else if (hashCode == 1) {
944 // This variation has the property of being stable (idempotent)
945 // between STW operations. This can be useful in some of the 1-0
946 // synchronization schemes.
947 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3;
948 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random;
949 } else if (hashCode == 2) {
950 value = 1; // for sensitivity testing
951 } else if (hashCode == 3) {
952 value = ++GVars.hc_sequence;
953 } else if (hashCode == 4) {
954 value = cast_from_oop<intptr_t>(obj);
955 } else if (hashCode == 5) {
956 // Marsaglia's xor-shift scheme with thread-specific state
957 // This is probably the best overall implementation -- we'll
958 // likely make this the default in future releases.
959 unsigned t = current->_hashStateX;
960 t ^= (t << 11);
961 current->_hashStateX = current->_hashStateY;
962 current->_hashStateY = current->_hashStateZ;
963 current->_hashStateZ = current->_hashStateW;
964 unsigned v = current->_hashStateW;
965 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
966 current->_hashStateW = v;
967 value = v;
968 } else {
969 assert(UseCompactObjectHeaders, "Only with compact i-hash");
970 #ifdef _LP64
971 uint64_t val = cast_from_oop<uint64_t>(obj);
972 uint32_t hash = FastHash::get_hash32((uint32_t)val, (uint32_t)(val >> 32));
973 #else
974 uint32_t val = cast_from_oop<uint32_t>(obj);
975 uint32_t hash = FastHash::get_hash32(val, UCONST64(0xAAAAAAAA));
976 #endif
977 value= static_cast<intptr_t>(hash);
978 }
979
980 value &= markWord::hash_mask;
981 if (hashCode != 6 && value == 0) value = 0xBAD;
982 assert(value != markWord::no_hash || hashCode == 6, "invariant");
983 return value;
984 }
985
986 static intptr_t install_hash_code(Thread* current, oop obj) {
987 assert(UseObjectMonitorTable && LockingMode == LM_LIGHTWEIGHT, "must be");
988
989 markWord mark = obj->mark_acquire();
990 for (;;) {
991 if (UseCompactObjectHeaders) {
992 if (mark.is_hashed()) {
993 return LightweightSynchronizer::get_hash(mark, obj);
994 }
995 intptr_t hash = ObjectSynchronizer::get_next_hash(current, obj); // get a new hash
996 markWord new_mark;
997 if (mark.is_not_hashed_expanded()) {
998 new_mark = mark.set_hashed_expanded();
999 int offset = mark.klass()->hash_offset_in_bytes(obj);
1000 obj->int_field_put(offset, (jint) hash);
1001 } else {
1002 new_mark = mark.set_hashed_not_expanded();
1003 }
1004 markWord old_mark = obj->cas_set_mark(new_mark, mark);
1005 if (old_mark == mark) {
1006 return hash;
1007 }
1008 mark = old_mark;
1009 } else {
1010 intptr_t hash = mark.hash();
1011 if (hash != 0) {
1012 return hash;
1013 }
1014
1015 hash = ObjectSynchronizer::get_next_hash(current, obj);
1016 const markWord old_mark = mark;
1017 const markWord new_mark = old_mark.copy_set_hash(hash);
1018
1019 mark = obj->cas_set_mark(new_mark, old_mark);
1020 if (old_mark == mark) {
1021 return hash;
1022 }
1023 }
1024 }
1025 }
1026
1027 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) {
1028 if (UseObjectMonitorTable) {
1029 // Since the monitor isn't in the object header, the hash can simply be
1030 // installed in the object header.
1031 return install_hash_code(current, obj);
1032 }
1033
1034 while (true) {
1035 ObjectMonitor* monitor = nullptr;
1036 markWord temp, test;
1037 intptr_t hash;
1038 markWord mark = read_stable_mark(obj);
1039 if (VerifyHeavyMonitors) {
1040 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)");
1041 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
1042 }
1043 if (mark.is_unlocked() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) {
1044 hash = mark.hash();
1045 if (hash != 0) { // if it has a hash, just return it
1046 return hash;
1047 }
1048 hash = get_next_hash(current, obj); // get a new hash
1049 temp = mark.copy_set_hash(hash); // merge the hash into header
1050 // try to install the hash
1051 test = obj->cas_set_mark(temp, mark);
1052 if (test == mark) { // if the hash was installed, return it
1053 return hash;
1054 }
1055 if (LockingMode == LM_LIGHTWEIGHT) {
1056 // CAS failed, retry
1057 continue;
1058 }
1059 // Failed to install the hash. It could be that another thread
1060 // installed the hash just before our attempt or inflation has
1061 // occurred or... so we fall thru to inflate the monitor for
1062 // stability and then install the hash.
1063 } else if (mark.has_monitor()) {
1064 monitor = mark.monitor();
1065 temp = monitor->header();
1066 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1067 hash = temp.hash();
1068 if (hash != 0) {
1069 // It has a hash.
1070
1071 // Separate load of dmw/header above from the loads in
1072 // is_being_async_deflated().
1073
1074 // dmw/header and _contentions may get written by different threads.
1075 // Make sure to observe them in the same order when having several observers.
1076 OrderAccess::loadload_for_IRIW();
1077
1078 if (monitor->is_being_async_deflated()) {
1079 // But we can't safely use the hash if we detect that async
1080 // deflation has occurred. So we attempt to restore the
1081 // header/dmw to the object's header so that we only retry
1082 // once if the deflater thread happens to be slow.
1083 monitor->install_displaced_markword_in_object(obj);
1084 continue;
1085 }
1086 return hash;
1087 }
1088 // Fall thru so we only have one place that installs the hash in
1089 // the ObjectMonitor.
1090 } else if (LockingMode == LM_LEGACY && mark.has_locker()
1091 && current->is_Java_thread()
1092 && JavaThread::cast(current)->is_lock_owned((address)mark.locker())) {
1093 // This is a stack-lock owned by the calling thread so fetch the
1094 // displaced markWord from the BasicLock on the stack.
1095 temp = mark.displaced_mark_helper();
1096 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1097 hash = temp.hash();
1098 if (hash != 0) { // if it has a hash, just return it
1099 return hash;
1100 }
1101 // WARNING:
1102 // The displaced header in the BasicLock on a thread's stack
1103 // is strictly immutable. It CANNOT be changed in ANY cases.
1104 // So we have to inflate the stack-lock into an ObjectMonitor
1105 // even if the current thread owns the lock. The BasicLock on
1106 // a thread's stack can be asynchronously read by other threads
1107 // during an inflate() call so any change to that stack memory
1108 // may not propagate to other threads correctly.
1109 }
1110
1111 // Inflate the monitor to set the hash.
1112
1113 // There's no need to inflate if the mark has already got a monitor.
1114 // NOTE: an async deflation can race after we get the monitor and
1115 // before we can update the ObjectMonitor's header with the hash
1116 // value below.
1117 monitor = mark.has_monitor() ? mark.monitor() : inflate(current, obj, inflate_cause_hash_code);
1118 // Load ObjectMonitor's header/dmw field and see if it has a hash.
1119 mark = monitor->header();
1120 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
1121 hash = mark.hash();
1122 if (hash == 0) { // if it does not have a hash
1123 hash = get_next_hash(current, obj); // get a new hash
1124 temp = mark.copy_set_hash(hash) ; // merge the hash into header
1125 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1126 uintptr_t v = Atomic::cmpxchg(monitor->metadata_addr(), mark.value(), temp.value());
1127 test = markWord(v);
1128 if (test != mark) {
1129 // The attempt to update the ObjectMonitor's header/dmw field
1130 // did not work. This can happen if another thread managed to
1131 // merge in the hash just before our cmpxchg().
1132 // If we add any new usages of the header/dmw field, this code
1133 // will need to be updated.
1134 hash = test.hash();
1135 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value());
1136 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash");
1137 }
1138 if (monitor->is_being_async_deflated() && !UseObjectMonitorTable) {
1139 // If we detect that async deflation has occurred, then we
1140 // attempt to restore the header/dmw to the object's header
1141 // so that we only retry once if the deflater thread happens
1142 // to be slow.
1143 monitor->install_displaced_markword_in_object(obj);
1144 continue;
1145 }
1146 }
1147 // We finally get the hash.
1148 return hash;
1149 }
1150 }
1151
1152 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current,
1153 Handle h_obj) {
1154 assert(current == JavaThread::current(), "Can only be called on current thread");
1155 oop obj = h_obj();
1156
1157 markWord mark = read_stable_mark(obj);
1158
1159 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1160 // stack-locked case, header points into owner's stack
1161 return current->is_lock_owned((address)mark.locker());
1162 }
1163
1164 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1165 // fast-locking case, see if lock is in current's lock stack
1166 return current->lock_stack().contains(h_obj());
1167 }
1168
1169 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) {
1170 ObjectMonitor* monitor = read_monitor(current, obj, mark);
1171 if (monitor != nullptr) {
1172 return monitor->is_entered(current) != 0;
1173 }
1174 // Racing with inflation/deflation, retry
1175 mark = obj->mark_acquire();
1176
1177 if (mark.is_fast_locked()) {
1178 // Some other thread fast_locked, current could not have held the lock
1179 return false;
1180 }
1181 }
1182
1183 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) {
1184 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
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 = read_monitor(mark);
1188 return monitor->is_entered(current) != 0;
1189 }
1190 // Unlocked case, header in place
1191 assert(mark.is_unlocked(), "sanity check");
1192 return false;
1193 }
1194
1195 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
1196 oop obj = h_obj();
1197 markWord mark = read_stable_mark(obj);
1198
1199 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1200 // stack-locked so header points into owner's stack.
1201 // owning_thread_from_monitor_owner() may also return null here:
1202 return Threads::owning_thread_from_stacklock(t_list, (address) mark.locker());
1203 }
1204
1205 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1206 // fast-locked so get owner from the object.
1207 // owning_thread_from_object() may also return null here:
1208 return Threads::owning_thread_from_object(t_list, h_obj());
1209 }
1210
1211 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) {
1212 ObjectMonitor* monitor = read_monitor(Thread::current(), obj, mark);
1213 if (monitor != nullptr) {
1214 return Threads::owning_thread_from_monitor(t_list, monitor);
1215 }
1216 // Racing with inflation/deflation, retry
1217 mark = obj->mark_acquire();
1218
1219 if (mark.is_fast_locked()) {
1220 // Some other thread fast_locked
1221 return Threads::owning_thread_from_object(t_list, h_obj());
1222 }
1223 }
1224
1225 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) {
1226 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
1227 // The first stage of async deflation does not affect any field
1228 // used by this comparison so the ObjectMonitor* is usable here.
1229 ObjectMonitor* monitor = read_monitor(mark);
1230 assert(monitor != nullptr, "monitor should be non-null");
1231 // owning_thread_from_monitor() may also return null here:
1232 return Threads::owning_thread_from_monitor(t_list, monitor);
1233 }
1234
1235 // Unlocked case, header in place
1236 // Cannot have assertion since this object may have been
1237 // locked by another thread when reaching here.
1238 // assert(mark.is_unlocked(), "sanity check");
1239
1240 return nullptr;
1241 }
1242
1243 // Visitors ...
1244
1245 // Iterate over all ObjectMonitors.
1246 template <typename Function>
1247 void ObjectSynchronizer::monitors_iterate(Function function) {
1248 MonitorList::Iterator iter = _in_use_list.iterator();
1249 while (iter.has_next()) {
1250 ObjectMonitor* monitor = iter.next();
1251 function(monitor);
1252 }
1253 }
1254
1255 // Iterate ObjectMonitors owned by any thread and where the owner `filter`
1256 // returns true.
1257 template <typename OwnerFilter>
1258 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) {
1259 monitors_iterate([&](ObjectMonitor* monitor) {
1260 // This function is only called at a safepoint or when the
1261 // target thread is suspended or when the target thread is
1262 // operating on itself. The current closures in use today are
1263 // only interested in an owned ObjectMonitor and ownership
1264 // cannot be dropped under the calling contexts so the
1265 // ObjectMonitor cannot be async deflated.
1266 if (monitor->has_owner() && filter(monitor)) {
1267 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating");
1268
1269 closure->do_monitor(monitor);
1270 }
1271 });
1272 }
1273
1274 // Iterate ObjectMonitors where the owner == thread; this does NOT include
1275 // ObjectMonitors where owner is set to a stack-lock address in thread.
1276 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) {
1277 int64_t key = ObjectMonitor::owner_id_from(thread);
1278 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; };
1279 return owned_monitors_iterate_filtered(closure, thread_filter);
1280 }
1281
1282 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, oop vthread) {
1283 int64_t key = ObjectMonitor::owner_id_from(vthread);
1284 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; };
1285 return owned_monitors_iterate_filtered(closure, thread_filter);
1286 }
1287
1288 // Iterate ObjectMonitors owned by any thread.
1289 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) {
1290 auto all_filter = [&](ObjectMonitor* monitor) { return true; };
1291 return owned_monitors_iterate_filtered(closure, all_filter);
1292 }
1293
1294 static bool monitors_used_above_threshold(MonitorList* list) {
1295 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy
1296 return false;
1297 }
1298 // Start with ceiling based on a per-thread estimate:
1299 size_t ceiling = ObjectSynchronizer::in_use_list_ceiling();
1300 size_t old_ceiling = ceiling;
1301 if (ceiling < list->max()) {
1302 // The max used by the system has exceeded the ceiling so use that:
1303 ceiling = list->max();
1304 }
1305 size_t monitors_used = list->count();
1306 if (monitors_used == 0) { // empty list is easy
1307 return false;
1308 }
1309 if (NoAsyncDeflationProgressMax != 0 &&
1310 _no_progress_cnt >= NoAsyncDeflationProgressMax) {
1311 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0;
1312 size_t new_ceiling = ceiling + (size_t)((double)ceiling * remainder) + 1;
1313 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling);
1314 log_info(monitorinflation)("Too many deflations without progress; "
1315 "bumping in_use_list_ceiling from " SIZE_FORMAT
1316 " to " SIZE_FORMAT, old_ceiling, new_ceiling);
1317 _no_progress_cnt = 0;
1318 ceiling = new_ceiling;
1319 }
1320
1321 // Check if our monitor usage is above the threshold:
1322 size_t monitor_usage = (monitors_used * 100LL) / ceiling;
1323 if (int(monitor_usage) > MonitorUsedDeflationThreshold) {
1324 log_info(monitorinflation)("monitors_used=" SIZE_FORMAT ", ceiling=" SIZE_FORMAT
1325 ", monitor_usage=" SIZE_FORMAT ", threshold=%d",
1326 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold);
1327 return true;
1328 }
1329
1330 return false;
1331 }
1332
1333 size_t ObjectSynchronizer::in_use_list_ceiling() {
1334 return _in_use_list_ceiling;
1335 }
1336
1337 void ObjectSynchronizer::dec_in_use_list_ceiling() {
1338 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1339 }
1340
1341 void ObjectSynchronizer::inc_in_use_list_ceiling() {
1342 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1343 }
1344
1345 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) {
1346 _in_use_list_ceiling = new_value;
1347 }
1348
1349 bool ObjectSynchronizer::is_async_deflation_needed() {
1350 if (is_async_deflation_requested()) {
1351 // Async deflation request.
1352 log_info(monitorinflation)("Async deflation needed: explicit request");
1353 return true;
1354 }
1355
1356 jlong time_since_last = time_since_last_async_deflation_ms();
1357
1358 if (AsyncDeflationInterval > 0 &&
1359 time_since_last > AsyncDeflationInterval &&
1360 monitors_used_above_threshold(&_in_use_list)) {
1361 // It's been longer than our specified deflate interval and there
1362 // are too many monitors in use. We don't deflate more frequently
1363 // than AsyncDeflationInterval (unless is_async_deflation_requested)
1364 // in order to not swamp the MonitorDeflationThread.
1365 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold");
1366 return true;
1367 }
1368
1369 if (GuaranteedAsyncDeflationInterval > 0 &&
1370 time_since_last > GuaranteedAsyncDeflationInterval) {
1371 // It's been longer than our specified guaranteed deflate interval.
1372 // We need to clean up the used monitors even if the threshold is
1373 // not reached, to keep the memory utilization at bay when many threads
1374 // touched many monitors.
1375 log_info(monitorinflation)("Async deflation needed: guaranteed interval (" INTX_FORMAT " ms) "
1376 "is greater than time since last deflation (" JLONG_FORMAT " ms)",
1377 GuaranteedAsyncDeflationInterval, time_since_last);
1378
1379 // If this deflation has no progress, then it should not affect the no-progress
1380 // tracking, otherwise threshold heuristics would think it was triggered, experienced
1381 // no progress, and needs to backoff more aggressively. In this "no progress" case,
1382 // the generic code would bump the no-progress counter, and we compensate for that
1383 // by telling it to skip the update.
1384 //
1385 // If this deflation has progress, then it should let non-progress tracking
1386 // know about this, otherwise the threshold heuristics would kick in, potentially
1387 // experience no-progress due to aggressive cleanup by this deflation, and think
1388 // it is still in no-progress stride. In this "progress" case, the generic code would
1389 // zero the counter, and we allow it to happen.
1390 _no_progress_skip_increment = true;
1391
1392 return true;
1393 }
1394
1395 return false;
1396 }
1397
1398 void ObjectSynchronizer::request_deflate_idle_monitors() {
1399 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag);
1400 set_is_async_deflation_requested(true);
1401 ml.notify_all();
1402 }
1403
1404 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() {
1405 JavaThread* current = JavaThread::current();
1406 bool ret_code = false;
1407
1408 jlong last_time = last_async_deflation_time_ns();
1409
1410 request_deflate_idle_monitors();
1411
1412 const int N_CHECKS = 5;
1413 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds
1414 if (last_async_deflation_time_ns() > last_time) {
1415 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i);
1416 ret_code = true;
1417 break;
1418 }
1419 {
1420 // JavaThread has to honor the blocking protocol.
1421 ThreadBlockInVM tbivm(current);
1422 os::naked_short_sleep(999); // sleep for almost 1 second
1423 }
1424 }
1425 if (!ret_code) {
1426 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS);
1427 }
1428
1429 return ret_code;
1430 }
1431
1432 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() {
1433 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS);
1434 }
1435
1436 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1437 const oop obj,
1438 ObjectSynchronizer::InflateCause cause) {
1439 assert(event != nullptr, "invariant");
1440 event->set_monitorClass(obj->klass());
1441 event->set_address((uintptr_t)(void*)obj);
1442 event->set_cause((u1)cause);
1443 event->commit();
1444 }
1445
1446 // Fast path code shared by multiple functions
1447 void ObjectSynchronizer::inflate_helper(oop obj) {
1448 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter");
1449 markWord mark = obj->mark_acquire();
1450 if (mark.has_monitor()) {
1451 ObjectMonitor* monitor = read_monitor(mark);
1452 markWord dmw = monitor->header();
1453 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value());
1454 return;
1455 }
1456 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal);
1457 }
1458
1459 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) {
1460 assert(current == Thread::current(), "must be");
1461 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter");
1462 return inflate_impl(current->is_Java_thread() ? JavaThread::cast(current) : nullptr, obj, cause);
1463 }
1464
1465 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) {
1466 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be");
1467 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_for");
1468 return inflate_impl(thread, obj, cause);
1469 }
1470
1471 ObjectMonitor* ObjectSynchronizer::inflate_impl(JavaThread* locking_thread, oop object, const InflateCause cause) {
1472 // The JavaThread* locking_thread requires that the locking_thread == Thread::current() or
1473 // is suspended throughout the call by some other mechanism.
1474 // The thread might be nullptr when called from a non JavaThread. (As may still be
1475 // the case from FastHashCode). However it is only important for correctness that the
1476 // thread is set when called from ObjectSynchronizer::enter from the owning thread,
1477 // ObjectSynchronizer::enter_for from any thread, or ObjectSynchronizer::exit.
1478 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_impl");
1479 EventJavaMonitorInflate event;
1480
1481 for (;;) {
1482 const markWord mark = object->mark_acquire();
1483
1484 // The mark can be in one of the following states:
1485 // * inflated - If the ObjectMonitor owner is anonymous and the
1486 // locking_thread owns the object lock, then we
1487 // make the locking_thread the ObjectMonitor owner.
1488 // * stack-locked - Coerce it to inflated from stack-locked.
1489 // * INFLATING - Busy wait for conversion from stack-locked to
1490 // inflated.
1491 // * unlocked - Aggressively inflate the object.
1492
1493 // CASE: inflated
1494 if (mark.has_monitor()) {
1495 ObjectMonitor* inf = mark.monitor();
1496 markWord dmw = inf->header();
1497 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1498 if (inf->has_anonymous_owner() && locking_thread != nullptr) {
1499 assert(LockingMode == LM_LEGACY, "invariant");
1500 if (locking_thread->is_lock_owned((address)inf->stack_locker())) {
1501 inf->set_stack_locker(nullptr);
1502 inf->set_owner_from_anonymous(locking_thread);
1503 }
1504 }
1505 return inf;
1506 }
1507
1508 // CASE: inflation in progress - inflating over a stack-lock.
1509 // Some other thread is converting from stack-locked to inflated.
1510 // Only that thread can complete inflation -- other threads must wait.
1511 // The INFLATING value is transient.
1512 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1513 // We could always eliminate polling by parking the thread on some auxiliary list.
1514 if (mark == markWord::INFLATING()) {
1515 read_stable_mark(object);
1516 continue;
1517 }
1518
1519 // CASE: stack-locked
1520 // Could be stack-locked either by current or by some other thread.
1521 //
1522 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting
1523 // to install INFLATING into the mark word. We originally installed INFLATING,
1524 // allocated the ObjectMonitor, and then finally STed the address of the
1525 // ObjectMonitor into the mark. This was correct, but artificially lengthened
1526 // the interval in which INFLATING appeared in the mark, thus increasing
1527 // the odds of inflation contention. If we lose the race to set INFLATING,
1528 // then we just delete the ObjectMonitor and loop around again.
1529 //
1530 LogStreamHandle(Trace, monitorinflation) lsh;
1531 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1532 ObjectMonitor* m = new ObjectMonitor(object);
1533 // Optimistically prepare the ObjectMonitor - anticipate successful CAS
1534 // We do this before the CAS in order to minimize the length of time
1535 // in which INFLATING appears in the mark.
1536
1537 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark);
1538 if (cmp != mark) {
1539 delete m;
1540 continue; // Interference -- just retry
1541 }
1542
1543 // We've successfully installed INFLATING (0) into the mark-word.
1544 // This is the only case where 0 will appear in a mark-word.
1545 // Only the singular thread that successfully swings the mark-word
1546 // to 0 can perform (or more precisely, complete) inflation.
1547 //
1548 // Why do we CAS a 0 into the mark-word instead of just CASing the
1549 // mark-word from the stack-locked value directly to the new inflated state?
1550 // Consider what happens when a thread unlocks a stack-locked object.
1551 // It attempts to use CAS to swing the displaced header value from the
1552 // on-stack BasicLock back into the object header. Recall also that the
1553 // header value (hash code, etc) can reside in (a) the object header, or
1554 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1555 // header in an ObjectMonitor. The inflate() routine must copy the header
1556 // value from the BasicLock on the owner's stack to the ObjectMonitor, all
1557 // the while preserving the hashCode stability invariants. If the owner
1558 // decides to release the lock while the value is 0, the unlock will fail
1559 // and control will eventually pass from slow_exit() to inflate. The owner
1560 // will then spin, waiting for the 0 value to disappear. Put another way,
1561 // the 0 causes the owner to stall if the owner happens to try to
1562 // drop the lock (restoring the header from the BasicLock to the object)
1563 // while inflation is in-progress. This protocol avoids races that might
1564 // would otherwise permit hashCode values to change or "flicker" for an object.
1565 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable.
1566 // 0 serves as a "BUSY" inflate-in-progress indicator.
1567
1568
1569 // fetch the displaced mark from the owner's stack.
1570 // The owner can't die or unwind past the lock while our INFLATING
1571 // object is in the mark. Furthermore the owner can't complete
1572 // an unlock on the object, either.
1573 markWord dmw = mark.displaced_mark_helper();
1574 // Catch if the object's header is not neutral (not locked and
1575 // not marked is what we care about here).
1576 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1577
1578 // Setup monitor fields to proper values -- prepare the monitor
1579 m->set_header(dmw);
1580
1581 // Note that a thread can inflate an object
1582 // that it has stack-locked -- as might happen in wait() -- directly
1583 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom.
1584 if (locking_thread != nullptr && locking_thread->is_lock_owned((address)mark.locker())) {
1585 m->set_owner(locking_thread);
1586 } else {
1587 // Use ANONYMOUS_OWNER to indicate that the owner is the BasicLock on the stack,
1588 // and set the stack locker field in the monitor.
1589 m->set_stack_locker(mark.locker());
1590 m->set_anonymous_owner();
1591 }
1592 // TODO-FIXME: assert BasicLock->dhw != 0.
1593
1594 // Must preserve store ordering. The monitor state must
1595 // be stable at the time of publishing the monitor address.
1596 guarantee(object->mark() == markWord::INFLATING(), "invariant");
1597 // Release semantics so that above set_object() is seen first.
1598 object->release_set_mark(markWord::encode(m));
1599
1600 // Once ObjectMonitor is configured and the object is associated
1601 // with the ObjectMonitor, it is safe to allow async deflation:
1602 _in_use_list.add(m);
1603
1604 // Hopefully the performance counters are allocated on distinct cache lines
1605 // to avoid false sharing on MP systems ...
1606 OM_PERFDATA_OP(Inflations, inc());
1607 if (log_is_enabled(Trace, monitorinflation)) {
1608 ResourceMark rm;
1609 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1610 INTPTR_FORMAT ", type='%s'", p2i(object),
1611 object->mark().value(), object->klass()->external_name());
1612 }
1613 if (event.should_commit()) {
1614 post_monitor_inflate_event(&event, object, cause);
1615 }
1616 return m;
1617 }
1618
1619 // CASE: unlocked
1620 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1621 // If we know we're inflating for entry it's better to inflate by swinging a
1622 // pre-locked ObjectMonitor pointer into the object header. A successful
1623 // CAS inflates the object *and* confers ownership to the inflating thread.
1624 // In the current implementation we use a 2-step mechanism where we CAS()
1625 // to inflate and then CAS() again to try to swing _owner from null to current.
1626 // An inflateTry() method that we could call from enter() would be useful.
1627
1628 assert(mark.is_unlocked(), "invariant: header=" INTPTR_FORMAT, mark.value());
1629 ObjectMonitor* m = new ObjectMonitor(object);
1630 // prepare m for installation - set monitor to initial state
1631 m->set_header(mark);
1632
1633 if (object->cas_set_mark(markWord::encode(m), mark) != mark) {
1634 delete m;
1635 m = nullptr;
1636 continue;
1637 // interference - the markword changed - just retry.
1638 // The state-transitions are one-way, so there's no chance of
1639 // live-lock -- "Inflated" is an absorbing state.
1640 }
1641
1642 // Once the ObjectMonitor is configured and object is associated
1643 // with the ObjectMonitor, it is safe to allow async deflation:
1644 _in_use_list.add(m);
1645
1646 // Hopefully the performance counters are allocated on distinct
1647 // cache lines to avoid false sharing on MP systems ...
1648 OM_PERFDATA_OP(Inflations, inc());
1649 if (log_is_enabled(Trace, monitorinflation)) {
1650 ResourceMark rm;
1651 lsh.print_cr("inflate(unlocked): object=" INTPTR_FORMAT ", mark="
1652 INTPTR_FORMAT ", type='%s'", p2i(object),
1653 object->mark().value(), object->klass()->external_name());
1654 }
1655 if (event.should_commit()) {
1656 post_monitor_inflate_event(&event, object, cause);
1657 }
1658 return m;
1659 }
1660 }
1661
1662 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle
1663 // ObjectMonitors. Returns the number of deflated ObjectMonitors.
1664 //
1665 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) {
1666 MonitorList::Iterator iter = _in_use_list.iterator();
1667 size_t deflated_count = 0;
1668 Thread* current = Thread::current();
1669
1670 while (iter.has_next()) {
1671 if (deflated_count >= (size_t)MonitorDeflationMax) {
1672 break;
1673 }
1674 ObjectMonitor* mid = iter.next();
1675 if (mid->deflate_monitor(current)) {
1676 deflated_count++;
1677 }
1678
1679 // Must check for a safepoint/handshake and honor it.
1680 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count);
1681 }
1682
1683 return deflated_count;
1684 }
1685
1686 class HandshakeForDeflation : public HandshakeClosure {
1687 public:
1688 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {}
1689
1690 void do_thread(Thread* thread) {
1691 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread="
1692 INTPTR_FORMAT, p2i(thread));
1693 if (thread->is_Java_thread()) {
1694 // Clear OM cache
1695 JavaThread* jt = JavaThread::cast(thread);
1696 jt->om_clear_monitor_cache();
1697 }
1698 }
1699 };
1700
1701 class VM_RendezvousGCThreads : public VM_Operation {
1702 public:
1703 bool evaluate_at_safepoint() const override { return false; }
1704 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; }
1705 void doit() override {
1706 Universe::heap()->safepoint_synchronize_begin();
1707 Universe::heap()->safepoint_synchronize_end();
1708 };
1709 };
1710
1711 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list,
1712 ObjectMonitorDeflationSafepointer* safepointer) {
1713 NativeHeapTrimmer::SuspendMark sm("monitor deletion");
1714 size_t deleted_count = 0;
1715 for (ObjectMonitor* monitor: *delete_list) {
1716 delete monitor;
1717 deleted_count++;
1718 // A JavaThread must check for a safepoint/handshake and honor it.
1719 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count);
1720 }
1721 return deleted_count;
1722 }
1723
1724 class ObjectMonitorDeflationLogging: public StackObj {
1725 LogStreamHandle(Debug, monitorinflation) _debug;
1726 LogStreamHandle(Info, monitorinflation) _info;
1727 LogStream* _stream;
1728 elapsedTimer _timer;
1729
1730 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); }
1731 size_t count() const { return ObjectSynchronizer::_in_use_list.count(); }
1732 size_t max() const { return ObjectSynchronizer::_in_use_list.max(); }
1733
1734 public:
1735 ObjectMonitorDeflationLogging()
1736 : _debug(), _info(), _stream(nullptr) {
1737 if (_debug.is_enabled()) {
1738 _stream = &_debug;
1739 } else if (_info.is_enabled()) {
1740 _stream = &_info;
1741 }
1742 }
1743
1744 void begin() {
1745 if (_stream != nullptr) {
1746 _stream->print_cr("begin deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1747 ceiling(), count(), max());
1748 _timer.start();
1749 }
1750 }
1751
1752 void before_handshake(size_t unlinked_count) {
1753 if (_stream != nullptr) {
1754 _timer.stop();
1755 _stream->print_cr("before handshaking: unlinked_count=" SIZE_FORMAT
1756 ", in_use_list stats: ceiling=" SIZE_FORMAT ", count="
1757 SIZE_FORMAT ", max=" SIZE_FORMAT,
1758 unlinked_count, ceiling(), count(), max());
1759 }
1760 }
1761
1762 void after_handshake() {
1763 if (_stream != nullptr) {
1764 _stream->print_cr("after handshaking: in_use_list stats: ceiling="
1765 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1766 ceiling(), count(), max());
1767 _timer.start();
1768 }
1769 }
1770
1771 void end(size_t deflated_count, size_t unlinked_count) {
1772 if (_stream != nullptr) {
1773 _timer.stop();
1774 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) {
1775 _stream->print_cr("deflated_count=" SIZE_FORMAT ", {unlinked,deleted}_count=" SIZE_FORMAT " monitors in %3.7f secs",
1776 deflated_count, unlinked_count, _timer.seconds());
1777 }
1778 _stream->print_cr("end deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1779 ceiling(), count(), max());
1780 }
1781 }
1782
1783 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) {
1784 if (_stream != nullptr) {
1785 _timer.stop();
1786 _stream->print_cr("pausing %s: %s=" SIZE_FORMAT ", in_use_list stats: ceiling="
1787 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1788 op_name, cnt_name, cnt, ceiling(), count(), max());
1789 }
1790 }
1791
1792 void after_block_for_safepoint(const char* op_name) {
1793 if (_stream != nullptr) {
1794 _stream->print_cr("resuming %s: in_use_list stats: ceiling=" SIZE_FORMAT
1795 ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, op_name,
1796 ceiling(), count(), max());
1797 _timer.start();
1798 }
1799 }
1800 };
1801
1802 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) {
1803 if (!SafepointMechanism::should_process(_current)) {
1804 return;
1805 }
1806
1807 // A safepoint/handshake has started.
1808 _log->before_block_for_safepoint(op_name, count_name, counter);
1809
1810 {
1811 // Honor block request.
1812 ThreadBlockInVM tbivm(_current);
1813 }
1814
1815 _log->after_block_for_safepoint(op_name);
1816 }
1817
1818 // This function is called by the MonitorDeflationThread to deflate
1819 // ObjectMonitors.
1820 size_t ObjectSynchronizer::deflate_idle_monitors() {
1821 JavaThread* current = JavaThread::current();
1822 assert(current->is_monitor_deflation_thread(), "The only monitor deflater");
1823
1824 // The async deflation request has been processed.
1825 _last_async_deflation_time_ns = os::javaTimeNanos();
1826 set_is_async_deflation_requested(false);
1827
1828 ObjectMonitorDeflationLogging log;
1829 ObjectMonitorDeflationSafepointer safepointer(current, &log);
1830
1831 log.begin();
1832
1833 // Deflate some idle ObjectMonitors.
1834 size_t deflated_count = deflate_monitor_list(&safepointer);
1835
1836 // Unlink the deflated ObjectMonitors from the in-use list.
1837 size_t unlinked_count = 0;
1838 size_t deleted_count = 0;
1839 if (deflated_count > 0) {
1840 ResourceMark rm(current);
1841 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count);
1842 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer);
1843
1844 #ifdef ASSERT
1845 if (UseObjectMonitorTable) {
1846 for (ObjectMonitor* monitor : delete_list) {
1847 assert(!LightweightSynchronizer::contains_monitor(current, monitor), "Should have been removed");
1848 }
1849 }
1850 #endif
1851
1852 log.before_handshake(unlinked_count);
1853
1854 // A JavaThread needs to handshake in order to safely free the
1855 // ObjectMonitors that were deflated in this cycle.
1856 HandshakeForDeflation hfd_hc;
1857 Handshake::execute(&hfd_hc);
1858 // Also, we sync and desync GC threads around the handshake, so that they can
1859 // safely read the mark-word and look-through to the object-monitor, without
1860 // being afraid that the object-monitor is going away.
1861 VM_RendezvousGCThreads sync_gc;
1862 VMThread::execute(&sync_gc);
1863
1864 log.after_handshake();
1865
1866 // After the handshake, safely free the ObjectMonitors that were
1867 // deflated and unlinked in this cycle.
1868
1869 // Delete the unlinked ObjectMonitors.
1870 deleted_count = delete_monitors(&delete_list, &safepointer);
1871 assert(unlinked_count == deleted_count, "must be");
1872 }
1873
1874 log.end(deflated_count, unlinked_count);
1875
1876 OM_PERFDATA_OP(MonExtant, set_value(_in_use_list.count()));
1877 OM_PERFDATA_OP(Deflations, inc(deflated_count));
1878
1879 GVars.stw_random = os::random();
1880
1881 if (deflated_count != 0) {
1882 _no_progress_cnt = 0;
1883 } else if (_no_progress_skip_increment) {
1884 _no_progress_skip_increment = false;
1885 } else {
1886 _no_progress_cnt++;
1887 }
1888
1889 return deflated_count;
1890 }
1891
1892 // Monitor cleanup on JavaThread::exit
1893
1894 // Iterate through monitor cache and attempt to release thread's monitors
1895 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1896 private:
1897 JavaThread* _thread;
1898
1899 public:
1900 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {}
1901 void do_monitor(ObjectMonitor* mid) {
1902 intx rec = mid->complete_exit(_thread);
1903 _thread->dec_held_monitor_count(rec + 1);
1904 }
1905 };
1906
1907 // Release all inflated monitors owned by current thread. Lightweight monitors are
1908 // ignored. This is meant to be called during JNI thread detach which assumes
1909 // all remaining monitors are heavyweight. All exceptions are swallowed.
1910 // Scanning the extant monitor list can be time consuming.
1911 // A simple optimization is to add a per-thread flag that indicates a thread
1912 // called jni_monitorenter() during its lifetime.
1913 //
1914 // Instead of NoSafepointVerifier it might be cheaper to
1915 // use an idiom of the form:
1916 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1917 // <code that must not run at safepoint>
1918 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1919 // Since the tests are extremely cheap we could leave them enabled
1920 // for normal product builds.
1921
1922 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) {
1923 assert(current == JavaThread::current(), "must be current Java thread");
1924 NoSafepointVerifier nsv;
1925 ReleaseJavaMonitorsClosure rjmc(current);
1926 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current);
1927 assert(!current->has_pending_exception(), "Should not be possible");
1928 current->clear_pending_exception();
1929 assert(current->held_monitor_count() == 0, "Should not be possible");
1930 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count.
1931 current->clear_jni_monitor_count();
1932 }
1933
1934 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1935 switch (cause) {
1936 case inflate_cause_vm_internal: return "VM Internal";
1937 case inflate_cause_monitor_enter: return "Monitor Enter";
1938 case inflate_cause_wait: return "Monitor Wait";
1939 case inflate_cause_notify: return "Monitor Notify";
1940 case inflate_cause_hash_code: return "Monitor Hash Code";
1941 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1942 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1943 default:
1944 ShouldNotReachHere();
1945 }
1946 return "Unknown";
1947 }
1948
1949 //------------------------------------------------------------------------------
1950 // Debugging code
1951
1952 u_char* ObjectSynchronizer::get_gvars_addr() {
1953 return (u_char*)&GVars;
1954 }
1955
1956 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() {
1957 return (u_char*)&GVars.hc_sequence;
1958 }
1959
1960 size_t ObjectSynchronizer::get_gvars_size() {
1961 return sizeof(SharedGlobals);
1962 }
1963
1964 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() {
1965 return (u_char*)&GVars.stw_random;
1966 }
1967
1968 // Do the final audit and print of ObjectMonitor stats; must be done
1969 // by the VMThread at VM exit time.
1970 void ObjectSynchronizer::do_final_audit_and_print_stats() {
1971 assert(Thread::current()->is_VM_thread(), "sanity check");
1972
1973 if (is_final_audit()) { // Only do the audit once.
1974 return;
1975 }
1976 set_is_final_audit();
1977 log_info(monitorinflation)("Starting the final audit.");
1978
1979 if (log_is_enabled(Info, monitorinflation)) {
1980 LogStreamHandle(Info, monitorinflation) ls;
1981 audit_and_print_stats(&ls, true /* on_exit */);
1982 }
1983 }
1984
1985 // This function can be called by the MonitorDeflationThread or it can be called when
1986 // we are trying to exit the VM. The list walker functions can run in parallel with
1987 // the other list operations.
1988 // Calls to this function can be added in various places as a debugging
1989 // aid.
1990 //
1991 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) {
1992 int error_cnt = 0;
1993
1994 ls->print_cr("Checking in_use_list:");
1995 chk_in_use_list(ls, &error_cnt);
1996
1997 if (error_cnt == 0) {
1998 ls->print_cr("No errors found in in_use_list checks.");
1999 } else {
2000 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt);
2001 }
2002
2003 // When exiting, only log the interesting entries at the Info level.
2004 // When called at intervals by the MonitorDeflationThread, log output
2005 // at the Trace level since there can be a lot of it.
2006 if (!on_exit && log_is_enabled(Trace, monitorinflation)) {
2007 LogStreamHandle(Trace, monitorinflation) ls_tr;
2008 log_in_use_monitor_details(&ls_tr, true /* log_all */);
2009 } else if (on_exit) {
2010 log_in_use_monitor_details(ls, false /* log_all */);
2011 }
2012
2013 ls->flush();
2014
2015 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt);
2016 }
2017
2018 // Check the in_use_list; log the results of the checks.
2019 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) {
2020 size_t l_in_use_count = _in_use_list.count();
2021 size_t l_in_use_max = _in_use_list.max();
2022 out->print_cr("count=" SIZE_FORMAT ", max=" SIZE_FORMAT, l_in_use_count,
2023 l_in_use_max);
2024
2025 size_t ck_in_use_count = 0;
2026 MonitorList::Iterator iter = _in_use_list.iterator();
2027 while (iter.has_next()) {
2028 ObjectMonitor* mid = iter.next();
2029 chk_in_use_entry(mid, out, error_cnt_p);
2030 ck_in_use_count++;
2031 }
2032
2033 if (l_in_use_count == ck_in_use_count) {
2034 out->print_cr("in_use_count=" SIZE_FORMAT " equals ck_in_use_count="
2035 SIZE_FORMAT, l_in_use_count, ck_in_use_count);
2036 } else {
2037 out->print_cr("WARNING: in_use_count=" SIZE_FORMAT " is not equal to "
2038 "ck_in_use_count=" SIZE_FORMAT, l_in_use_count,
2039 ck_in_use_count);
2040 }
2041
2042 size_t ck_in_use_max = _in_use_list.max();
2043 if (l_in_use_max == ck_in_use_max) {
2044 out->print_cr("in_use_max=" SIZE_FORMAT " equals ck_in_use_max="
2045 SIZE_FORMAT, l_in_use_max, ck_in_use_max);
2046 } else {
2047 out->print_cr("WARNING: in_use_max=" SIZE_FORMAT " is not equal to "
2048 "ck_in_use_max=" SIZE_FORMAT, l_in_use_max, ck_in_use_max);
2049 }
2050 }
2051
2052 // Check an in-use monitor entry; log any errors.
2053 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out,
2054 int* error_cnt_p) {
2055 if (n->owner_is_DEFLATER_MARKER()) {
2056 // This could happen when monitor deflation blocks for a safepoint.
2057 return;
2058 }
2059
2060
2061 if (n->metadata() == 0) {
2062 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must "
2063 "have non-null _metadata (header/hash) field.", p2i(n));
2064 *error_cnt_p = *error_cnt_p + 1;
2065 }
2066
2067 const oop obj = n->object_peek();
2068 if (obj == nullptr) {
2069 return;
2070 }
2071
2072 const markWord mark = obj->mark();
2073 if (!mark.has_monitor()) {
2074 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2075 "object does not think it has a monitor: obj="
2076 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n),
2077 p2i(obj), mark.value());
2078 *error_cnt_p = *error_cnt_p + 1;
2079 return;
2080 }
2081
2082 ObjectMonitor* const obj_mon = read_monitor(Thread::current(), obj, mark);
2083 if (n != obj_mon) {
2084 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2085 "object does not refer to the same monitor: obj="
2086 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon="
2087 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon));
2088 *error_cnt_p = *error_cnt_p + 1;
2089 }
2090 }
2091
2092 // Log details about ObjectMonitors on the in_use_list. The 'BHL'
2093 // flags indicate why the entry is in-use, 'object' and 'object type'
2094 // indicate the associated object and its type.
2095 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) {
2096 if (_in_use_list.count() > 0) {
2097 stringStream ss;
2098 out->print_cr("In-use monitor info%s:", log_all ? "" : " (eliding idle monitors)");
2099 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2100 out->print_cr("%18s %s %18s %18s",
2101 "monitor", "BHL", "object", "object type");
2102 out->print_cr("================== === ================== ==================");
2103
2104 auto is_interesting = [&](ObjectMonitor* monitor) {
2105 return log_all || monitor->has_owner() || monitor->is_busy();
2106 };
2107
2108 monitors_iterate([&](ObjectMonitor* monitor) {
2109 if (is_interesting(monitor)) {
2110 const oop obj = monitor->object_peek();
2111 const intptr_t hash = UseObjectMonitorTable ? monitor->hash() : monitor->header().hash();
2112 ResourceMark rm;
2113 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor),
2114 monitor->is_busy(), hash != 0, monitor->has_owner(),
2115 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name());
2116 if (monitor->is_busy()) {
2117 out->print(" (%s)", monitor->is_busy_to_string(&ss));
2118 ss.reset();
2119 }
2120 out->cr();
2121 }
2122 });
2123 }
2124
2125 out->flush();
2126 }