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/linkedlist.hpp"
67 #include "utilities/preserveException.hpp"
68
69 class ObjectMonitorDeflationLogging;
70
71 void MonitorList::add(ObjectMonitor* m) {
72 ObjectMonitor* head;
73 do {
74 head = Atomic::load(&_head);
75 m->set_next_om(head);
76 } while (Atomic::cmpxchg(&_head, head, m) != head);
77
78 size_t count = Atomic::add(&_count, 1u, memory_order_relaxed);
79 size_t old_max;
80 do {
81 old_max = Atomic::load(&_max);
82 if (count <= old_max) {
83 break;
84 }
85 } while (Atomic::cmpxchg(&_max, old_max, count, memory_order_relaxed) != old_max);
86 }
87
88 size_t MonitorList::count() const {
89 return Atomic::load(&_count);
90 }
91
92 size_t MonitorList::max() const {
93 return Atomic::load(&_max);
94 }
95
96 class ObjectMonitorDeflationSafepointer : public StackObj {
97 JavaThread* const _current;
98 ObjectMonitorDeflationLogging* const _log;
99
100 public:
101 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log)
102 : _current(current), _log(log) {}
103
104 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter);
105 };
106
107 // Walk the in-use list and unlink deflated ObjectMonitors.
108 // Returns the number of unlinked ObjectMonitors.
109 size_t MonitorList::unlink_deflated(size_t deflated_count,
110 GrowableArray<ObjectMonitor*>* unlinked_list,
111 ObjectMonitorDeflationSafepointer* safepointer) {
112 size_t unlinked_count = 0;
113 ObjectMonitor* prev = nullptr;
114 ObjectMonitor* m = Atomic::load_acquire(&_head);
115
116 while (m != nullptr) {
117 if (m->is_being_async_deflated()) {
118 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can
119 // modify the list once per batch. The batch starts at "m".
120 size_t unlinked_batch = 0;
121 ObjectMonitor* next = m;
122 // Look for at most MonitorUnlinkBatch monitors, or the number of
123 // deflated and not unlinked monitors, whatever comes first.
124 assert(deflated_count >= unlinked_count, "Sanity: underflow");
125 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch);
126 do {
127 ObjectMonitor* next_next = next->next_om();
128 unlinked_batch++;
129 unlinked_list->append(next);
130 next = next_next;
131 if (unlinked_batch >= unlinked_batch_limit) {
132 // Reached the max batch, so bail out of the gathering loop.
133 break;
134 }
135 if (prev == nullptr && Atomic::load(&_head) != m) {
136 // Current batch used to be at head, but it is not at head anymore.
137 // Bail out and figure out where we currently are. This avoids long
138 // walks searching for new prev during unlink under heavy list inserts.
139 break;
140 }
141 } while (next != nullptr && next->is_being_async_deflated());
142
143 // Unlink the found batch.
144 if (prev == nullptr) {
145 // The current batch is the first batch, so there is a chance that it starts at head.
146 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head.
147 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, m, next);
148 if (prev_head != m) {
149 // Something must have updated the head. Figure out the actual prev for this batch.
150 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) {
151 prev = n;
152 }
153 assert(prev != nullptr, "Should have found the prev for the current batch");
154 prev->set_next_om(next);
155 }
156 } else {
157 // The current batch is preceded by another batch. This guarantees the current batch
158 // does not start at head. Unlink the entire current batch without updating the head.
159 assert(Atomic::load(&_head) != m, "Sanity");
160 prev->set_next_om(next);
161 }
162
163 unlinked_count += unlinked_batch;
164 if (unlinked_count >= deflated_count) {
165 // Reached the max so bail out of the searching loop.
166 // There should be no more deflated monitors left.
167 break;
168 }
169 m = next;
170 } else {
171 prev = m;
172 m = m->next_om();
173 }
174
175 // Must check for a safepoint/handshake and honor it.
176 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count);
177 }
178
179 #ifdef ASSERT
180 // Invariant: the code above should unlink all deflated monitors.
181 // The code that runs after this unlinking does not expect deflated monitors.
182 // Notably, attempting to deflate the already deflated monitor would break.
183 {
184 ObjectMonitor* m = Atomic::load_acquire(&_head);
185 while (m != nullptr) {
186 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked");
187 m = m->next_om();
188 }
189 }
190 #endif
191
192 Atomic::sub(&_count, unlinked_count);
193 return unlinked_count;
194 }
195
196 MonitorList::Iterator MonitorList::iterator() const {
197 return Iterator(Atomic::load_acquire(&_head));
198 }
199
200 ObjectMonitor* MonitorList::Iterator::next() {
201 ObjectMonitor* current = _current;
202 _current = current->next_om();
203 return current;
204 }
205
206 // The "core" versions of monitor enter and exit reside in this file.
207 // The interpreter and compilers contain specialized transliterated
208 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp
209 // fast_lock(...) for instance. If you make changes here, make sure to modify the
210 // interpreter, and both C1 and C2 fast-path inline locking code emission.
211 //
212 // -----------------------------------------------------------------------------
213
214 #ifdef DTRACE_ENABLED
215
216 // Only bother with this argument setup if dtrace is available
217 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
218
219 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
220 char* bytes = nullptr; \
221 int len = 0; \
222 jlong jtid = SharedRuntime::get_java_tid(thread); \
223 Symbol* klassname = obj->klass()->name(); \
224 if (klassname != nullptr) { \
225 bytes = (char*)klassname->bytes(); \
226 len = klassname->utf8_length(); \
227 }
228
229 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
230 { \
231 if (DTraceMonitorProbes) { \
232 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
233 HOTSPOT_MONITOR_WAIT(jtid, \
234 (uintptr_t)(monitor), bytes, len, (millis)); \
235 } \
236 }
237
238 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY
239 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL
240 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
241
242 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
243 { \
244 if (DTraceMonitorProbes) { \
245 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
246 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
247 (uintptr_t)(monitor), bytes, len); \
248 } \
249 }
250
251 #else // ndef DTRACE_ENABLED
252
253 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
254 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
255
256 #endif // ndef DTRACE_ENABLED
257
258 // This exists only as a workaround of dtrace bug 6254741
259 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) {
260 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
261 return 0;
262 }
263
264 static constexpr size_t inflation_lock_count() {
265 return 256;
266 }
267
268 // Static storage for an array of PlatformMutex.
269 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)];
270
271 static inline PlatformMutex* inflation_lock(size_t index) {
272 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]);
273 }
274
275 void ObjectSynchronizer::initialize() {
276 for (size_t i = 0; i < inflation_lock_count(); i++) {
277 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex();
278 }
279 // Start the ceiling with the estimate for one thread.
280 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate);
281
282 // Start the timer for deflations, so it does not trigger immediately.
283 _last_async_deflation_time_ns = os::javaTimeNanos();
284
285 if (LockingMode == LM_LIGHTWEIGHT) {
286 LightweightSynchronizer::initialize();
287 }
288 }
289
290 MonitorList ObjectSynchronizer::_in_use_list;
291 // monitors_used_above_threshold() policy is as follows:
292 //
293 // The ratio of the current _in_use_list count to the ceiling is used
294 // to determine if we are above MonitorUsedDeflationThreshold and need
295 // to do an async monitor deflation cycle. The ceiling is increased by
296 // AvgMonitorsPerThreadEstimate when a thread is added to the system
297 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is
298 // removed from the system.
299 //
300 // Note: If the _in_use_list max exceeds the ceiling, then
301 // monitors_used_above_threshold() will use the in_use_list max instead
302 // of the thread count derived ceiling because we have used more
303 // ObjectMonitors than the estimated average.
304 //
305 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax
306 // no-progress async monitor deflation cycles in a row, then the ceiling
307 // is adjusted upwards by monitors_used_above_threshold().
308 //
309 // Start the ceiling with the estimate for one thread in initialize()
310 // which is called after cmd line options are processed.
311 static size_t _in_use_list_ceiling = 0;
312 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false;
313 bool volatile ObjectSynchronizer::_is_final_audit = false;
314 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0;
315 static uintx _no_progress_cnt = 0;
316 static bool _no_progress_skip_increment = false;
317
318 // =====================> Quick functions
319
320 // The quick_* forms are special fast-path variants used to improve
321 // performance. In the simplest case, a "quick_*" implementation could
322 // simply return false, in which case the caller will perform the necessary
323 // state transitions and call the slow-path form.
324 // The fast-path is designed to handle frequently arising cases in an efficient
325 // manner and is just a degenerate "optimistic" variant of the slow-path.
326 // returns true -- to indicate the call was satisfied.
327 // returns false -- to indicate the call needs the services of the slow-path.
328 // A no-loitering ordinance is in effect for code in the quick_* family
329 // operators: safepoints or indefinite blocking (blocking that might span a
330 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon
331 // entry.
332 //
333 // Consider: An interesting optimization is to have the JIT recognize the
334 // following common idiom:
335 // synchronized (someobj) { .... ; notify(); }
336 // That is, we find a notify() or notifyAll() call that immediately precedes
337 // the monitorexit operation. In that case the JIT could fuse the operations
338 // into a single notifyAndExit() runtime primitive.
339
340 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) {
341 assert(current->thread_state() == _thread_in_Java, "invariant");
342 NoSafepointVerifier nsv;
343 if (obj == nullptr) return false; // slow-path for invalid obj
344 const markWord mark = obj->mark();
345
346 if (LockingMode == LM_LIGHTWEIGHT) {
347 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) {
348 // Degenerate notify
349 // fast-locked by caller so by definition the implied waitset is empty.
350 return true;
351 }
352 } else if (LockingMode == LM_LEGACY) {
353 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
354 // Degenerate notify
355 // stack-locked by caller so by definition the implied waitset is empty.
356 return true;
357 }
358 }
359
360 if (mark.has_monitor()) {
361 ObjectMonitor* const mon = read_monitor(current, obj, mark);
362 if (LockingMode == LM_LIGHTWEIGHT && mon == nullptr) {
363 // Racing with inflation/deflation go slow path
364 return false;
365 }
366 assert(mon->object() == oop(obj), "invariant");
367 if (!mon->has_owner(current)) return false; // slow-path for IMS exception
368
369 if (mon->first_waiter() != nullptr) {
370 // We have one or more waiters. Since this is an inflated monitor
371 // that we own, we quickly notify them here and now, avoiding the slow-path.
372 if (all) {
373 mon->quick_notifyAll(current);
374 } else {
375 mon->quick_notify(current);
376 }
377 }
378 return true;
379 }
380
381 // other IMS exception states take the slow-path
382 return false;
383 }
384
385 static bool useHeavyMonitors() {
386 #if defined(X86) || defined(AARCH64) || defined(PPC64) || defined(RISCV64) || defined(S390)
387 return LockingMode == LM_MONITOR;
388 #else
389 return false;
390 #endif
391 }
392
393 // The LockNode emitted directly at the synchronization site would have
394 // been too big if it were to have included support for the cases of inflated
395 // recursive enter and exit, so they go here instead.
396 // Note that we can't safely call AsyncPrintJavaStack() from within
397 // quick_enter() as our thread state remains _in_Java.
398
399 bool ObjectSynchronizer::quick_enter_legacy(oop obj, BasicLock* lock, JavaThread* current) {
400 assert(current->thread_state() == _thread_in_Java, "invariant");
401
402 if (useHeavyMonitors()) {
403 return false; // Slow path
404 }
405
406 assert(LockingMode == LM_LEGACY, "legacy mode below");
407
408 const markWord mark = obj->mark();
409
410 if (mark.has_monitor()) {
411
412 ObjectMonitor* const m = read_monitor(mark);
413 // An async deflation or GC can race us before we manage to make
414 // the ObjectMonitor busy by setting the owner below. If we detect
415 // that race we just bail out to the slow-path here.
416 if (m->object_peek() == nullptr) {
417 return false;
418 }
419
420 // Lock contention and Transactional Lock Elision (TLE) diagnostics
421 // and observability
422 // Case: light contention possibly amenable to TLE
423 // Case: TLE inimical operations such as nested/recursive synchronization
424
425 if (m->has_owner(current)) {
426 m->increment_recursions(current);
427 current->inc_held_monitor_count();
428 return true;
429 }
430
431 // This Java Monitor is inflated so obj's header will never be
432 // displaced to this thread's BasicLock. Make the displaced header
433 // non-null so this BasicLock is not seen as recursive nor as
434 // being locked. We do this unconditionally so that this thread's
435 // BasicLock cannot be mis-interpreted by any stack walkers. For
436 // performance reasons, stack walkers generally first check for
437 // stack-locking in the object's header, the second check is for
438 // recursive stack-locking in the displaced header in the BasicLock,
439 // and last are the inflated Java Monitor (ObjectMonitor) checks.
440 lock->set_displaced_header(markWord::unused_mark());
441
442 if (!m->has_owner() && m->try_set_owner(current)) {
443 assert(m->recursions() == 0, "invariant");
444 current->inc_held_monitor_count();
445 return true;
446 }
447 }
448
449 // Note that we could inflate in quick_enter.
450 // This is likely a useful optimization
451 // Critically, in quick_enter() we must not:
452 // -- block indefinitely, or
453 // -- reach a safepoint
454
455 return false; // revert to slow-path
456 }
457
458 // Handle notifications when synchronizing on value based classes
459 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) {
460 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
461 frame last_frame = locking_thread->last_frame();
462 bool bcp_was_adjusted = false;
463 // Don't decrement bcp if it points to the frame's first instruction. This happens when
464 // handle_sync_on_value_based_class() is called because of a synchronized method. There
465 // is no actual monitorenter instruction in the byte code in this case.
466 if (last_frame.is_interpreted_frame() &&
467 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) {
468 // adjust bcp to point back to monitorenter so that we print the correct line numbers
469 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1);
470 bcp_was_adjusted = true;
471 }
472
473 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) {
474 ResourceMark rm;
475 stringStream ss;
476 locking_thread->print_active_stack_on(&ss);
477 char* base = (char*)strstr(ss.base(), "at");
478 char* newline = (char*)strchr(ss.base(), '\n');
479 if (newline != nullptr) {
480 *newline = '\0';
481 }
482 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base);
483 } else {
484 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses");
485 ResourceMark rm;
486 Log(valuebasedclasses) vblog;
487
488 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name());
489 if (locking_thread->has_last_Java_frame()) {
490 LogStream info_stream(vblog.info());
491 locking_thread->print_active_stack_on(&info_stream);
492 } else {
493 vblog.info("Cannot find the last Java frame");
494 }
495
496 EventSyncOnValueBasedClass event;
497 if (event.should_commit()) {
498 event.set_valueBasedClass(obj->klass());
499 event.commit();
500 }
501 }
502
503 if (bcp_was_adjusted) {
504 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1);
505 }
506 }
507
508 // -----------------------------------------------------------------------------
509 // Monitor Enter/Exit
510
511 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
512 // When called with locking_thread != Thread::current() some mechanism must synchronize
513 // the locking_thread with respect to the current thread. Currently only used when
514 // deoptimizing and re-locking locks. See Deoptimization::relock_objects
515 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
516
517 if (LockingMode == LM_LIGHTWEIGHT) {
518 return LightweightSynchronizer::enter_for(obj, lock, locking_thread);
519 }
520
521 if (!enter_fast_impl(obj, lock, locking_thread)) {
522 // Inflated ObjectMonitor::enter_for is required
523
524 // An async deflation can race after the inflate_for() call and before
525 // enter_for() can make the ObjectMonitor busy. enter_for() returns false
526 // if we have lost the race to async deflation and we simply try again.
527 while (true) {
528 ObjectMonitor* monitor = inflate_for(locking_thread, obj(), inflate_cause_monitor_enter);
529 if (monitor->enter_for(locking_thread)) {
530 return;
531 }
532 assert(monitor->is_being_async_deflated(), "must be");
533 }
534 }
535 }
536
537 void ObjectSynchronizer::enter_legacy(Handle obj, BasicLock* lock, JavaThread* current) {
538 if (!enter_fast_impl(obj, lock, current)) {
539 // Inflated ObjectMonitor::enter is required
540
541 // An async deflation can race after the inflate() call and before
542 // enter() can make the ObjectMonitor busy. enter() returns false if
543 // we have lost the race to async deflation and we simply try again.
544 while (true) {
545 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter);
546 if (monitor->enter(current)) {
547 return;
548 }
549 }
550 }
551 }
552
553 // The interpreter and compiler assembly code tries to lock using the fast path
554 // of this algorithm. Make sure to update that code if the following function is
555 // changed. The implementation is extremely sensitive to race condition. Be careful.
556 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
557 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer");
558
559 if (obj->klass()->is_value_based()) {
560 handle_sync_on_value_based_class(obj, locking_thread);
561 }
562
563 locking_thread->inc_held_monitor_count();
564
565 if (!useHeavyMonitors()) {
566 if (LockingMode == LM_LEGACY) {
567 markWord mark = obj->mark();
568 if (mark.is_unlocked()) {
569 // Anticipate successful CAS -- the ST of the displaced mark must
570 // be visible <= the ST performed by the CAS.
571 lock->set_displaced_header(mark);
572 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) {
573 return true;
574 }
575 } else if (mark.has_locker() &&
576 locking_thread->is_lock_owned((address) mark.locker())) {
577 assert(lock != mark.locker(), "must not re-lock the same lock");
578 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock");
579 lock->set_displaced_header(markWord::from_pointer(nullptr));
580 return true;
581 }
582
583 // The object header will never be displaced to this lock,
584 // so it does not matter what the value is, except that it
585 // must be non-zero to avoid looking like a re-entrant lock,
586 // and must not look locked either.
587 lock->set_displaced_header(markWord::unused_mark());
588
589 // Failed to fast lock.
590 return false;
591 }
592 } else if (VerifyHeavyMonitors) {
593 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
594 }
595
596 return false;
597 }
598
599 void ObjectSynchronizer::exit_legacy(oop object, BasicLock* lock, JavaThread* current) {
600 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer");
601
602 if (!useHeavyMonitors()) {
603 markWord mark = object->mark();
604 if (LockingMode == LM_LEGACY) {
605 markWord dhw = lock->displaced_header();
606 if (dhw.value() == 0) {
607 // If the displaced header is null, then this exit matches up with
608 // a recursive enter. No real work to do here except for diagnostics.
609 #ifndef PRODUCT
610 if (mark != markWord::INFLATING()) {
611 // Only do diagnostics if we are not racing an inflation. Simply
612 // exiting a recursive enter of a Java Monitor that is being
613 // inflated is safe; see the has_monitor() comment below.
614 assert(!mark.is_unlocked(), "invariant");
615 assert(!mark.has_locker() ||
616 current->is_lock_owned((address)mark.locker()), "invariant");
617 if (mark.has_monitor()) {
618 // The BasicLock's displaced_header is marked as a recursive
619 // enter and we have an inflated Java Monitor (ObjectMonitor).
620 // This is a special case where the Java Monitor was inflated
621 // after this thread entered the stack-lock recursively. When a
622 // Java Monitor is inflated, we cannot safely walk the Java
623 // Monitor owner's stack and update the BasicLocks because a
624 // Java Monitor can be asynchronously inflated by a thread that
625 // does not own the Java Monitor.
626 ObjectMonitor* m = read_monitor(mark);
627 assert(m->object()->mark() == mark, "invariant");
628 assert(m->is_entered(current), "invariant");
629 }
630 }
631 #endif
632 return;
633 }
634
635 if (mark == markWord::from_pointer(lock)) {
636 // If the object is stack-locked by the current thread, try to
637 // swing the displaced header from the BasicLock back to the mark.
638 assert(dhw.is_neutral(), "invariant");
639 if (object->cas_set_mark(dhw, mark) == mark) {
640 return;
641 }
642 }
643 }
644 } else if (VerifyHeavyMonitors) {
645 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
646 }
647
648 // We have to take the slow-path of possible inflation and then exit.
649 // The ObjectMonitor* can't be async deflated until ownership is
650 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
651 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal);
652 assert(!monitor->has_anonymous_owner(), "must not be");
653 monitor->exit(current);
654 }
655
656 // -----------------------------------------------------------------------------
657 // JNI locks on java objects
658 // NOTE: must use heavy weight monitor to handle jni monitor enter
659 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) {
660 // Top native frames in the stack will not be seen if we attempt
661 // preemption, since we start walking from the last Java anchor.
662 NoPreemptMark npm(current);
663
664 if (obj->klass()->is_value_based()) {
665 handle_sync_on_value_based_class(obj, current);
666 }
667
668 // the current locking is from JNI instead of Java code
669 current->set_current_pending_monitor_is_from_java(false);
670 // An async deflation can race after the inflate() call and before
671 // enter() can make the ObjectMonitor busy. enter() returns false if
672 // we have lost the race to async deflation and we simply try again.
673 while (true) {
674 ObjectMonitor* monitor;
675 bool entered;
676 if (LockingMode == LM_LIGHTWEIGHT) {
677 BasicLock lock;
678 entered = LightweightSynchronizer::inflate_and_enter(obj(), &lock, inflate_cause_jni_enter, current, current) != nullptr;
679 } else {
680 monitor = inflate(current, obj(), inflate_cause_jni_enter);
681 entered = monitor->enter(current);
682 }
683
684 if (entered) {
685 current->inc_held_monitor_count(1, true);
686 break;
687 }
688 }
689 current->set_current_pending_monitor_is_from_java(true);
690 }
691
692 // NOTE: must use heavy weight monitor to handle jni monitor exit
693 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) {
694 JavaThread* current = THREAD;
695
696 ObjectMonitor* monitor;
697 if (LockingMode == LM_LIGHTWEIGHT) {
698 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj, inflate_cause_jni_exit, CHECK);
699 } else {
700 // The ObjectMonitor* can't be async deflated until ownership is
701 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
702 monitor = inflate(current, obj, inflate_cause_jni_exit);
703 }
704 // If this thread has locked the object, exit the monitor. We
705 // intentionally do not use CHECK on check_owner because we must exit the
706 // monitor even if an exception was already pending.
707 if (monitor->check_owner(THREAD)) {
708 monitor->exit(current);
709 current->dec_held_monitor_count(1, true);
710 }
711 }
712
713 // -----------------------------------------------------------------------------
714 // Internal VM locks on java objects
715 // standard constructor, allows locking failures
716 ObjectLocker::ObjectLocker(Handle obj, TRAPS) : _thread(THREAD), _obj(obj),
717 _npm(_thread, _thread->at_preemptable_init() /* ignore_mark */), _skip_exit(false) {
718 assert(!_thread->preempting(), "");
719
720 _thread->check_for_valid_safepoint_state();
721
722 if (_obj() != nullptr) {
723 ObjectSynchronizer::enter(_obj, &_lock, _thread);
724
725 if (_thread->preempting()) {
726 // If preemption was cancelled we acquired the monitor after freezing
727 // the frames. Redoing the vm call later in thaw will require us to
728 // release it since the call should look like the original one. We
729 // do it in ~ObjectLocker to reduce the window of time we hold the
730 // monitor since we can't do anything useful with it now, and would
731 // otherwise just force other vthreads to preempt in case they try
732 // to acquire this monitor.
733 _skip_exit = !_thread->preemption_cancelled();
734 _thread->set_pending_preempted_exception();
735 }
736 }
737 }
738
739 ObjectLocker::~ObjectLocker() {
740 if (_obj() != nullptr && !_skip_exit) {
741 ObjectSynchronizer::exit(_obj(), &_lock, _thread);
742 }
743 }
744
745 void ObjectLocker::wait_uninterruptibly(TRAPS) {
746 ObjectSynchronizer::waitUninterruptibly(_obj, 0, _thread);
747 if (_thread->preempting()) {
748 _skip_exit = true;
749 _thread->set_pending_preempted_exception();
750 _thread->set_class_to_be_initialized(nullptr);
751 }
752 }
753
754 // -----------------------------------------------------------------------------
755 // Wait/Notify/NotifyAll
756 // NOTE: must use heavy weight monitor to handle wait()
757
758 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
759 JavaThread* current = THREAD;
760 if (millis < 0) {
761 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
762 }
763
764 ObjectMonitor* monitor;
765 if (LockingMode == LM_LIGHTWEIGHT) {
766 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK_0);
767 } else {
768 // The ObjectMonitor* can't be async deflated because the _waiters
769 // field is incremented before ownership is dropped and decremented
770 // after ownership is regained.
771 monitor = inflate(current, obj(), inflate_cause_wait);
772 }
773
774 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis);
775 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code
776
777 // This dummy call is in place to get around dtrace bug 6254741. Once
778 // that's fixed we can uncomment the following line, remove the call
779 // and change this function back into a "void" func.
780 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
781 int ret_code = dtrace_waited_probe(monitor, obj, THREAD);
782 return ret_code;
783 }
784
785 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
786 assert(millis >= 0, "timeout value is negative");
787
788 ObjectMonitor* monitor;
789 if (LockingMode == LM_LIGHTWEIGHT) {
790 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK);
791 } else {
792 monitor = inflate(THREAD, obj(), inflate_cause_wait);
793 }
794 monitor->wait(millis, false, THREAD);
795 }
796
797
798 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
799 JavaThread* current = THREAD;
800
801 markWord mark = obj->mark();
802 if (LockingMode == LM_LIGHTWEIGHT) {
803 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
804 // Not inflated so there can't be any waiters to notify.
805 return;
806 }
807 } else if (LockingMode == LM_LEGACY) {
808 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
809 // Not inflated so there can't be any waiters to notify.
810 return;
811 }
812 }
813
814 ObjectMonitor* monitor;
815 if (LockingMode == LM_LIGHTWEIGHT) {
816 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK);
817 } else {
818 // The ObjectMonitor* can't be async deflated until ownership is
819 // dropped by the calling thread.
820 monitor = inflate(current, obj(), inflate_cause_notify);
821 }
822 monitor->notify(CHECK);
823 }
824
825 // NOTE: see comment of notify()
826 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
827 JavaThread* current = THREAD;
828
829 markWord mark = obj->mark();
830 if (LockingMode == LM_LIGHTWEIGHT) {
831 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
832 // Not inflated so there can't be any waiters to notify.
833 return;
834 }
835 } else if (LockingMode == LM_LEGACY) {
836 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
837 // Not inflated so there can't be any waiters to notify.
838 return;
839 }
840 }
841
842 ObjectMonitor* monitor;
843 if (LockingMode == LM_LIGHTWEIGHT) {
844 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK);
845 } else {
846 // The ObjectMonitor* can't be async deflated until ownership is
847 // dropped by the calling thread.
848 monitor = inflate(current, obj(), inflate_cause_notify);
849 }
850 monitor->notifyAll(CHECK);
851 }
852
853 // -----------------------------------------------------------------------------
854 // Hash Code handling
855
856 struct SharedGlobals {
857 char _pad_prefix[OM_CACHE_LINE_SIZE];
858 // This is a highly shared mostly-read variable.
859 // To avoid false-sharing it needs to be the sole occupant of a cache line.
860 volatile int stw_random;
861 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int));
862 // Hot RW variable -- Sequester to avoid false-sharing
863 volatile int hc_sequence;
864 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int));
865 };
866
867 static SharedGlobals GVars;
868
869 static markWord read_stable_mark(oop obj) {
870 markWord mark = obj->mark_acquire();
871 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) {
872 // New lightweight locking does not use the markWord::INFLATING() protocol.
873 return mark; // normal fast-path return
874 }
875
876 int its = 0;
877 for (;;) {
878 markWord mark = obj->mark_acquire();
879 if (!mark.is_being_inflated()) {
880 return mark; // normal fast-path return
881 }
882
883 // The object is being inflated by some other thread.
884 // The caller of read_stable_mark() must wait for inflation to complete.
885 // Avoid live-lock.
886
887 ++its;
888 if (its > 10000 || !os::is_MP()) {
889 if (its & 1) {
890 os::naked_yield();
891 } else {
892 // Note that the following code attenuates the livelock problem but is not
893 // a complete remedy. A more complete solution would require that the inflating
894 // thread hold the associated inflation lock. The following code simply restricts
895 // the number of spinners to at most one. We'll have N-2 threads blocked
896 // on the inflationlock, 1 thread holding the inflation lock and using
897 // a yield/park strategy, and 1 thread in the midst of inflation.
898 // A more refined approach would be to change the encoding of INFLATING
899 // to allow encapsulation of a native thread pointer. Threads waiting for
900 // inflation to complete would use CAS to push themselves onto a singly linked
901 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
902 // and calling park(). When inflation was complete the thread that accomplished inflation
903 // would detach the list and set the markword to inflated with a single CAS and
904 // then for each thread on the list, set the flag and unpark() the thread.
905
906 // Index into the lock array based on the current object address.
907 static_assert(is_power_of_2(inflation_lock_count()), "must be");
908 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1);
909 int YieldThenBlock = 0;
910 assert(ix < inflation_lock_count(), "invariant");
911 inflation_lock(ix)->lock();
912 while (obj->mark_acquire() == markWord::INFLATING()) {
913 // Beware: naked_yield() is advisory and has almost no effect on some platforms
914 // so we periodically call current->_ParkEvent->park(1).
915 // We use a mixed spin/yield/block mechanism.
916 if ((YieldThenBlock++) >= 16) {
917 Thread::current()->_ParkEvent->park(1);
918 } else {
919 os::naked_yield();
920 }
921 }
922 inflation_lock(ix)->unlock();
923 }
924 } else {
925 SpinPause(); // SMP-polite spinning
926 }
927 }
928 }
929
930 // hashCode() generation :
931 //
932 // Possibilities:
933 // * MD5Digest of {obj,stw_random}
934 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function.
935 // * A DES- or AES-style SBox[] mechanism
936 // * One of the Phi-based schemes, such as:
937 // 2654435761 = 2^32 * Phi (golden ratio)
938 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ;
939 // * A variation of Marsaglia's shift-xor RNG scheme.
940 // * (obj ^ stw_random) is appealing, but can result
941 // in undesirable regularity in the hashCode values of adjacent objects
942 // (objects allocated back-to-back, in particular). This could potentially
943 // result in hashtable collisions and reduced hashtable efficiency.
944 // There are simple ways to "diffuse" the middle address bits over the
945 // generated hashCode values:
946
947 static intptr_t get_next_hash(Thread* current, oop obj) {
948 intptr_t value = 0;
949 if (hashCode == 0) {
950 // This form uses global Park-Miller RNG.
951 // On MP system we'll have lots of RW access to a global, so the
952 // mechanism induces lots of coherency traffic.
953 value = os::random();
954 } else if (hashCode == 1) {
955 // This variation has the property of being stable (idempotent)
956 // between STW operations. This can be useful in some of the 1-0
957 // synchronization schemes.
958 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3;
959 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random;
960 } else if (hashCode == 2) {
961 value = 1; // for sensitivity testing
962 } else if (hashCode == 3) {
963 value = ++GVars.hc_sequence;
964 } else if (hashCode == 4) {
965 value = cast_from_oop<intptr_t>(obj);
966 } else {
967 // Marsaglia's xor-shift scheme with thread-specific state
968 // This is probably the best overall implementation -- we'll
969 // likely make this the default in future releases.
970 unsigned t = current->_hashStateX;
971 t ^= (t << 11);
972 current->_hashStateX = current->_hashStateY;
973 current->_hashStateY = current->_hashStateZ;
974 current->_hashStateZ = current->_hashStateW;
975 unsigned v = current->_hashStateW;
976 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
977 current->_hashStateW = v;
978 value = v;
979 }
980
981 value &= markWord::hash_mask;
982 if (value == 0) value = 0xBAD;
983 assert(value != markWord::no_hash, "invariant");
984 return value;
985 }
986
987 static intptr_t install_hash_code(Thread* current, oop obj) {
988 assert(UseObjectMonitorTable && LockingMode == LM_LIGHTWEIGHT, "must be");
989
990 markWord mark = obj->mark_acquire();
991 for (;;) {
992 intptr_t hash = mark.hash();
993 if (hash != 0) {
994 return hash;
995 }
996
997 hash = get_next_hash(current, obj);
998 const markWord old_mark = mark;
999 const markWord new_mark = old_mark.copy_set_hash(hash);
1000
1001 mark = obj->cas_set_mark(new_mark, old_mark);
1002 if (old_mark == mark) {
1003 return hash;
1004 }
1005 }
1006 }
1007
1008 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) {
1009 if (UseObjectMonitorTable) {
1010 // Since the monitor isn't in the object header, the hash can simply be
1011 // installed in the object header.
1012 return install_hash_code(current, obj);
1013 }
1014
1015 while (true) {
1016 ObjectMonitor* monitor = nullptr;
1017 markWord temp, test;
1018 intptr_t hash;
1019 markWord mark = read_stable_mark(obj);
1020 if (VerifyHeavyMonitors) {
1021 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)");
1022 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
1023 }
1024 if (mark.is_unlocked() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) {
1025 hash = mark.hash();
1026 if (hash != 0) { // if it has a hash, just return it
1027 return hash;
1028 }
1029 hash = get_next_hash(current, obj); // get a new hash
1030 temp = mark.copy_set_hash(hash); // merge the hash into header
1031 // try to install the hash
1032 test = obj->cas_set_mark(temp, mark);
1033 if (test == mark) { // if the hash was installed, return it
1034 return hash;
1035 }
1036 if (LockingMode == LM_LIGHTWEIGHT) {
1037 // CAS failed, retry
1038 continue;
1039 }
1040 // Failed to install the hash. It could be that another thread
1041 // installed the hash just before our attempt or inflation has
1042 // occurred or... so we fall thru to inflate the monitor for
1043 // stability and then install the hash.
1044 } else if (mark.has_monitor()) {
1045 monitor = mark.monitor();
1046 temp = monitor->header();
1047 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1048 hash = temp.hash();
1049 if (hash != 0) {
1050 // It has a hash.
1051
1052 // Separate load of dmw/header above from the loads in
1053 // is_being_async_deflated().
1054
1055 // dmw/header and _contentions may get written by different threads.
1056 // Make sure to observe them in the same order when having several observers.
1057 OrderAccess::loadload_for_IRIW();
1058
1059 if (monitor->is_being_async_deflated()) {
1060 // But we can't safely use the hash if we detect that async
1061 // deflation has occurred. So we attempt to restore the
1062 // header/dmw to the object's header so that we only retry
1063 // once if the deflater thread happens to be slow.
1064 monitor->install_displaced_markword_in_object(obj);
1065 continue;
1066 }
1067 return hash;
1068 }
1069 // Fall thru so we only have one place that installs the hash in
1070 // the ObjectMonitor.
1071 } else if (LockingMode == LM_LEGACY && mark.has_locker()
1072 && current->is_Java_thread()
1073 && JavaThread::cast(current)->is_lock_owned((address)mark.locker())) {
1074 // This is a stack-lock owned by the calling thread so fetch the
1075 // displaced markWord from the BasicLock on the stack.
1076 temp = mark.displaced_mark_helper();
1077 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1078 hash = temp.hash();
1079 if (hash != 0) { // if it has a hash, just return it
1080 return hash;
1081 }
1082 // WARNING:
1083 // The displaced header in the BasicLock on a thread's stack
1084 // is strictly immutable. It CANNOT be changed in ANY cases.
1085 // So we have to inflate the stack-lock into an ObjectMonitor
1086 // even if the current thread owns the lock. The BasicLock on
1087 // a thread's stack can be asynchronously read by other threads
1088 // during an inflate() call so any change to that stack memory
1089 // may not propagate to other threads correctly.
1090 }
1091
1092 // Inflate the monitor to set the hash.
1093
1094 // There's no need to inflate if the mark has already got a monitor.
1095 // NOTE: an async deflation can race after we get the monitor and
1096 // before we can update the ObjectMonitor's header with the hash
1097 // value below.
1098 monitor = mark.has_monitor() ? mark.monitor() : inflate(current, obj, inflate_cause_hash_code);
1099 // Load ObjectMonitor's header/dmw field and see if it has a hash.
1100 mark = monitor->header();
1101 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
1102 hash = mark.hash();
1103 if (hash == 0) { // if it does not have a hash
1104 hash = get_next_hash(current, obj); // get a new hash
1105 temp = mark.copy_set_hash(hash) ; // merge the hash into header
1106 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1107 uintptr_t v = Atomic::cmpxchg(monitor->metadata_addr(), mark.value(), temp.value());
1108 test = markWord(v);
1109 if (test != mark) {
1110 // The attempt to update the ObjectMonitor's header/dmw field
1111 // did not work. This can happen if another thread managed to
1112 // merge in the hash just before our cmpxchg().
1113 // If we add any new usages of the header/dmw field, this code
1114 // will need to be updated.
1115 hash = test.hash();
1116 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value());
1117 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash");
1118 }
1119 if (monitor->is_being_async_deflated() && !UseObjectMonitorTable) {
1120 // If we detect that async deflation has occurred, then we
1121 // attempt to restore the header/dmw to the object's header
1122 // so that we only retry once if the deflater thread happens
1123 // to be slow.
1124 monitor->install_displaced_markword_in_object(obj);
1125 continue;
1126 }
1127 }
1128 // We finally get the hash.
1129 return hash;
1130 }
1131 }
1132
1133 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current,
1134 Handle h_obj) {
1135 assert(current == JavaThread::current(), "Can only be called on current thread");
1136 oop obj = h_obj();
1137
1138 markWord mark = read_stable_mark(obj);
1139
1140 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1141 // stack-locked case, header points into owner's stack
1142 return current->is_lock_owned((address)mark.locker());
1143 }
1144
1145 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1146 // fast-locking case, see if lock is in current's lock stack
1147 return current->lock_stack().contains(h_obj());
1148 }
1149
1150 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) {
1151 ObjectMonitor* monitor = read_monitor(current, obj, mark);
1152 if (monitor != nullptr) {
1153 return monitor->is_entered(current) != 0;
1154 }
1155 // Racing with inflation/deflation, retry
1156 mark = obj->mark_acquire();
1157
1158 if (mark.is_fast_locked()) {
1159 // Some other thread fast_locked, current could not have held the lock
1160 return false;
1161 }
1162 }
1163
1164 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) {
1165 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
1166 // The first stage of async deflation does not affect any field
1167 // used by this comparison so the ObjectMonitor* is usable here.
1168 ObjectMonitor* monitor = read_monitor(mark);
1169 return monitor->is_entered(current) != 0;
1170 }
1171 // Unlocked case, header in place
1172 assert(mark.is_unlocked(), "sanity check");
1173 return false;
1174 }
1175
1176 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
1177 oop obj = h_obj();
1178 markWord mark = read_stable_mark(obj);
1179
1180 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1181 // stack-locked so header points into owner's stack.
1182 // owning_thread_from_monitor_owner() may also return null here:
1183 return Threads::owning_thread_from_stacklock(t_list, (address) mark.locker());
1184 }
1185
1186 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1187 // fast-locked so get owner from the object.
1188 // owning_thread_from_object() may also return null here:
1189 return Threads::owning_thread_from_object(t_list, h_obj());
1190 }
1191
1192 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) {
1193 ObjectMonitor* monitor = read_monitor(Thread::current(), obj, mark);
1194 if (monitor != nullptr) {
1195 return Threads::owning_thread_from_monitor(t_list, monitor);
1196 }
1197 // Racing with inflation/deflation, retry
1198 mark = obj->mark_acquire();
1199
1200 if (mark.is_fast_locked()) {
1201 // Some other thread fast_locked
1202 return Threads::owning_thread_from_object(t_list, h_obj());
1203 }
1204 }
1205
1206 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) {
1207 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
1208 // The first stage of async deflation does not affect any field
1209 // used by this comparison so the ObjectMonitor* is usable here.
1210 ObjectMonitor* monitor = read_monitor(mark);
1211 assert(monitor != nullptr, "monitor should be non-null");
1212 // owning_thread_from_monitor() may also return null here:
1213 return Threads::owning_thread_from_monitor(t_list, monitor);
1214 }
1215
1216 // Unlocked case, header in place
1217 // Cannot have assertion since this object may have been
1218 // locked by another thread when reaching here.
1219 // assert(mark.is_unlocked(), "sanity check");
1220
1221 return nullptr;
1222 }
1223
1224 // Visitors ...
1225
1226 // Iterate over all ObjectMonitors.
1227 template <typename Function>
1228 void ObjectSynchronizer::monitors_iterate(Function function) {
1229 MonitorList::Iterator iter = _in_use_list.iterator();
1230 while (iter.has_next()) {
1231 ObjectMonitor* monitor = iter.next();
1232 function(monitor);
1233 }
1234 }
1235
1236 // Iterate ObjectMonitors owned by any thread and where the owner `filter`
1237 // returns true.
1238 template <typename OwnerFilter>
1239 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) {
1240 monitors_iterate([&](ObjectMonitor* monitor) {
1241 // This function is only called at a safepoint or when the
1242 // target thread is suspended or when the target thread is
1243 // operating on itself. The current closures in use today are
1244 // only interested in an owned ObjectMonitor and ownership
1245 // cannot be dropped under the calling contexts so the
1246 // ObjectMonitor cannot be async deflated.
1247 if (monitor->has_owner() && filter(monitor)) {
1248 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating");
1249
1250 closure->do_monitor(monitor);
1251 }
1252 });
1253 }
1254
1255 // Iterate ObjectMonitors where the owner == thread; this does NOT include
1256 // ObjectMonitors where owner is set to a stack-lock address in thread.
1257 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) {
1258 int64_t key = ObjectMonitor::owner_id_from(thread);
1259 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; };
1260 return owned_monitors_iterate_filtered(closure, thread_filter);
1261 }
1262
1263 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, oop vthread) {
1264 int64_t key = ObjectMonitor::owner_id_from(vthread);
1265 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; };
1266 return owned_monitors_iterate_filtered(closure, thread_filter);
1267 }
1268
1269 // Iterate ObjectMonitors owned by any thread.
1270 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) {
1271 auto all_filter = [&](ObjectMonitor* monitor) { return true; };
1272 return owned_monitors_iterate_filtered(closure, all_filter);
1273 }
1274
1275 static bool monitors_used_above_threshold(MonitorList* list) {
1276 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy
1277 return false;
1278 }
1279 size_t monitors_used = list->count();
1280 if (monitors_used == 0) { // empty list is easy
1281 return false;
1282 }
1283 size_t old_ceiling = ObjectSynchronizer::in_use_list_ceiling();
1284 // Make sure that we use a ceiling value that is not lower than
1285 // previous, not lower than the recorded max used by the system, and
1286 // not lower than the current number of monitors in use (which can
1287 // race ahead of max). The result is guaranteed > 0.
1288 size_t ceiling = MAX3(old_ceiling, list->max(), monitors_used);
1289
1290 // Check if our monitor usage is above the threshold:
1291 size_t monitor_usage = (monitors_used * 100LL) / ceiling;
1292 if (int(monitor_usage) > MonitorUsedDeflationThreshold) {
1293 // Deflate monitors if over the threshold percentage, unless no
1294 // progress on previous deflations.
1295 bool is_above_threshold = true;
1296
1297 // Check if it's time to adjust the in_use_list_ceiling up, due
1298 // to too many async deflation attempts without any progress.
1299 if (NoAsyncDeflationProgressMax != 0 &&
1300 _no_progress_cnt >= NoAsyncDeflationProgressMax) {
1301 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0;
1302 size_t delta = (size_t)(ceiling * remainder) + 1;
1303 size_t new_ceiling = (ceiling > SIZE_MAX - delta)
1304 ? SIZE_MAX // Overflow, let's clamp new_ceiling.
1305 : ceiling + delta;
1306
1307 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling);
1308 log_info(monitorinflation)("Too many deflations without progress; "
1309 "bumping in_use_list_ceiling from %zu"
1310 " to %zu", old_ceiling, new_ceiling);
1311 _no_progress_cnt = 0;
1312 ceiling = new_ceiling;
1313
1314 // Check if our monitor usage is still above the threshold:
1315 monitor_usage = (monitors_used * 100LL) / ceiling;
1316 is_above_threshold = int(monitor_usage) > MonitorUsedDeflationThreshold;
1317 }
1318 log_info(monitorinflation)("monitors_used=%zu, ceiling=%zu"
1319 ", monitor_usage=%zu, threshold=%d",
1320 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold);
1321 return is_above_threshold;
1322 }
1323
1324 return false;
1325 }
1326
1327 size_t ObjectSynchronizer::in_use_list_count() {
1328 return _in_use_list.count();
1329 }
1330
1331 size_t ObjectSynchronizer::in_use_list_max() {
1332 return _in_use_list.max();
1333 }
1334
1335 size_t ObjectSynchronizer::in_use_list_ceiling() {
1336 return _in_use_list_ceiling;
1337 }
1338
1339 void ObjectSynchronizer::dec_in_use_list_ceiling() {
1340 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1341 }
1342
1343 void ObjectSynchronizer::inc_in_use_list_ceiling() {
1344 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1345 }
1346
1347 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) {
1348 _in_use_list_ceiling = new_value;
1349 }
1350
1351 bool ObjectSynchronizer::is_async_deflation_needed() {
1352 if (is_async_deflation_requested()) {
1353 // Async deflation request.
1354 log_info(monitorinflation)("Async deflation needed: explicit request");
1355 return true;
1356 }
1357
1358 jlong time_since_last = time_since_last_async_deflation_ms();
1359
1360 if (AsyncDeflationInterval > 0 &&
1361 time_since_last > AsyncDeflationInterval &&
1362 monitors_used_above_threshold(&_in_use_list)) {
1363 // It's been longer than our specified deflate interval and there
1364 // are too many monitors in use. We don't deflate more frequently
1365 // than AsyncDeflationInterval (unless is_async_deflation_requested)
1366 // in order to not swamp the MonitorDeflationThread.
1367 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold");
1368 return true;
1369 }
1370
1371 if (GuaranteedAsyncDeflationInterval > 0 &&
1372 time_since_last > GuaranteedAsyncDeflationInterval) {
1373 // It's been longer than our specified guaranteed deflate interval.
1374 // We need to clean up the used monitors even if the threshold is
1375 // not reached, to keep the memory utilization at bay when many threads
1376 // touched many monitors.
1377 log_info(monitorinflation)("Async deflation needed: guaranteed interval (%zd ms) "
1378 "is greater than time since last deflation (" JLONG_FORMAT " ms)",
1379 GuaranteedAsyncDeflationInterval, time_since_last);
1380
1381 // If this deflation has no progress, then it should not affect the no-progress
1382 // tracking, otherwise threshold heuristics would think it was triggered, experienced
1383 // no progress, and needs to backoff more aggressively. In this "no progress" case,
1384 // the generic code would bump the no-progress counter, and we compensate for that
1385 // by telling it to skip the update.
1386 //
1387 // If this deflation has progress, then it should let non-progress tracking
1388 // know about this, otherwise the threshold heuristics would kick in, potentially
1389 // experience no-progress due to aggressive cleanup by this deflation, and think
1390 // it is still in no-progress stride. In this "progress" case, the generic code would
1391 // zero the counter, and we allow it to happen.
1392 _no_progress_skip_increment = true;
1393
1394 return true;
1395 }
1396
1397 return false;
1398 }
1399
1400 void ObjectSynchronizer::request_deflate_idle_monitors() {
1401 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag);
1402 set_is_async_deflation_requested(true);
1403 ml.notify_all();
1404 }
1405
1406 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() {
1407 JavaThread* current = JavaThread::current();
1408 bool ret_code = false;
1409
1410 jlong last_time = last_async_deflation_time_ns();
1411
1412 request_deflate_idle_monitors();
1413
1414 const int N_CHECKS = 5;
1415 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds
1416 if (last_async_deflation_time_ns() > last_time) {
1417 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i);
1418 ret_code = true;
1419 break;
1420 }
1421 {
1422 // JavaThread has to honor the blocking protocol.
1423 ThreadBlockInVM tbivm(current);
1424 os::naked_short_sleep(999); // sleep for almost 1 second
1425 }
1426 }
1427 if (!ret_code) {
1428 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS);
1429 }
1430
1431 return ret_code;
1432 }
1433
1434 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() {
1435 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS);
1436 }
1437
1438 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1439 const oop obj,
1440 ObjectSynchronizer::InflateCause cause) {
1441 assert(event != nullptr, "invariant");
1442 const Klass* monitor_klass = obj->klass();
1443 if (ObjectMonitor::is_jfr_excluded(monitor_klass)) {
1444 return;
1445 }
1446 event->set_monitorClass(monitor_klass);
1447 event->set_address((uintptr_t)(void*)obj);
1448 event->set_cause((u1)cause);
1449 event->commit();
1450 }
1451
1452 // Fast path code shared by multiple functions
1453 void ObjectSynchronizer::inflate_helper(oop obj) {
1454 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter");
1455 markWord mark = obj->mark_acquire();
1456 if (mark.has_monitor()) {
1457 ObjectMonitor* monitor = read_monitor(mark);
1458 markWord dmw = monitor->header();
1459 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value());
1460 return;
1461 }
1462 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal);
1463 }
1464
1465 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) {
1466 assert(current == Thread::current(), "must be");
1467 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter");
1468 return inflate_impl(current->is_Java_thread() ? JavaThread::cast(current) : nullptr, obj, cause);
1469 }
1470
1471 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) {
1472 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be");
1473 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_for");
1474 return inflate_impl(thread, obj, cause);
1475 }
1476
1477 ObjectMonitor* ObjectSynchronizer::inflate_impl(JavaThread* locking_thread, oop object, const InflateCause cause) {
1478 // The JavaThread* locking_thread requires that the locking_thread == Thread::current() or
1479 // is suspended throughout the call by some other mechanism.
1480 // The thread might be nullptr when called from a non JavaThread. (As may still be
1481 // the case from FastHashCode). However it is only important for correctness that the
1482 // thread is set when called from ObjectSynchronizer::enter from the owning thread,
1483 // ObjectSynchronizer::enter_for from any thread, or ObjectSynchronizer::exit.
1484 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_impl");
1485 EventJavaMonitorInflate event;
1486
1487 for (;;) {
1488 const markWord mark = object->mark_acquire();
1489
1490 // The mark can be in one of the following states:
1491 // * inflated - If the ObjectMonitor owner is anonymous and the
1492 // locking_thread owns the object lock, then we
1493 // make the locking_thread the ObjectMonitor owner.
1494 // * stack-locked - Coerce it to inflated from stack-locked.
1495 // * INFLATING - Busy wait for conversion from stack-locked to
1496 // inflated.
1497 // * unlocked - Aggressively inflate the object.
1498
1499 // CASE: inflated
1500 if (mark.has_monitor()) {
1501 ObjectMonitor* inf = mark.monitor();
1502 markWord dmw = inf->header();
1503 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1504 if (inf->has_anonymous_owner() && locking_thread != nullptr) {
1505 assert(LockingMode == LM_LEGACY, "invariant");
1506 if (locking_thread->is_lock_owned((address)inf->stack_locker())) {
1507 inf->set_stack_locker(nullptr);
1508 inf->set_owner_from_anonymous(locking_thread);
1509 }
1510 }
1511 return inf;
1512 }
1513
1514 // CASE: inflation in progress - inflating over a stack-lock.
1515 // Some other thread is converting from stack-locked to inflated.
1516 // Only that thread can complete inflation -- other threads must wait.
1517 // The INFLATING value is transient.
1518 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1519 // We could always eliminate polling by parking the thread on some auxiliary list.
1520 if (mark == markWord::INFLATING()) {
1521 read_stable_mark(object);
1522 continue;
1523 }
1524
1525 // CASE: stack-locked
1526 // Could be stack-locked either by current or by some other thread.
1527 //
1528 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting
1529 // to install INFLATING into the mark word. We originally installed INFLATING,
1530 // allocated the ObjectMonitor, and then finally STed the address of the
1531 // ObjectMonitor into the mark. This was correct, but artificially lengthened
1532 // the interval in which INFLATING appeared in the mark, thus increasing
1533 // the odds of inflation contention. If we lose the race to set INFLATING,
1534 // then we just delete the ObjectMonitor and loop around again.
1535 //
1536 LogStreamHandle(Trace, monitorinflation) lsh;
1537 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1538 ObjectMonitor* m = new ObjectMonitor(object);
1539 // Optimistically prepare the ObjectMonitor - anticipate successful CAS
1540 // We do this before the CAS in order to minimize the length of time
1541 // in which INFLATING appears in the mark.
1542
1543 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark);
1544 if (cmp != mark) {
1545 delete m;
1546 continue; // Interference -- just retry
1547 }
1548
1549 // We've successfully installed INFLATING (0) into the mark-word.
1550 // This is the only case where 0 will appear in a mark-word.
1551 // Only the singular thread that successfully swings the mark-word
1552 // to 0 can perform (or more precisely, complete) inflation.
1553 //
1554 // Why do we CAS a 0 into the mark-word instead of just CASing the
1555 // mark-word from the stack-locked value directly to the new inflated state?
1556 // Consider what happens when a thread unlocks a stack-locked object.
1557 // It attempts to use CAS to swing the displaced header value from the
1558 // on-stack BasicLock back into the object header. Recall also that the
1559 // header value (hash code, etc) can reside in (a) the object header, or
1560 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1561 // header in an ObjectMonitor. The inflate() routine must copy the header
1562 // value from the BasicLock on the owner's stack to the ObjectMonitor, all
1563 // the while preserving the hashCode stability invariants. If the owner
1564 // decides to release the lock while the value is 0, the unlock will fail
1565 // and control will eventually pass from slow_exit() to inflate. The owner
1566 // will then spin, waiting for the 0 value to disappear. Put another way,
1567 // the 0 causes the owner to stall if the owner happens to try to
1568 // drop the lock (restoring the header from the BasicLock to the object)
1569 // while inflation is in-progress. This protocol avoids races that might
1570 // would otherwise permit hashCode values to change or "flicker" for an object.
1571 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable.
1572 // 0 serves as a "BUSY" inflate-in-progress indicator.
1573
1574
1575 // fetch the displaced mark from the owner's stack.
1576 // The owner can't die or unwind past the lock while our INFLATING
1577 // object is in the mark. Furthermore the owner can't complete
1578 // an unlock on the object, either.
1579 markWord dmw = mark.displaced_mark_helper();
1580 // Catch if the object's header is not neutral (not locked and
1581 // not marked is what we care about here).
1582 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1583
1584 // Setup monitor fields to proper values -- prepare the monitor
1585 m->set_header(dmw);
1586
1587 // Note that a thread can inflate an object
1588 // that it has stack-locked -- as might happen in wait() -- directly
1589 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom.
1590 if (locking_thread != nullptr && locking_thread->is_lock_owned((address)mark.locker())) {
1591 m->set_owner(locking_thread);
1592 } else {
1593 // Use ANONYMOUS_OWNER to indicate that the owner is the BasicLock on the stack,
1594 // and set the stack locker field in the monitor.
1595 m->set_stack_locker(mark.locker());
1596 m->set_anonymous_owner();
1597 }
1598 // TODO-FIXME: assert BasicLock->dhw != 0.
1599
1600 // Must preserve store ordering. The monitor state must
1601 // be stable at the time of publishing the monitor address.
1602 guarantee(object->mark() == markWord::INFLATING(), "invariant");
1603 // Release semantics so that above set_object() is seen first.
1604 object->release_set_mark(markWord::encode(m));
1605
1606 // Once ObjectMonitor is configured and the object is associated
1607 // with the ObjectMonitor, it is safe to allow async deflation:
1608 _in_use_list.add(m);
1609
1610 if (log_is_enabled(Trace, monitorinflation)) {
1611 ResourceMark rm;
1612 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1613 INTPTR_FORMAT ", type='%s'", p2i(object),
1614 object->mark().value(), object->klass()->external_name());
1615 }
1616 if (event.should_commit()) {
1617 post_monitor_inflate_event(&event, object, cause);
1618 }
1619 return m;
1620 }
1621
1622 // CASE: unlocked
1623 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1624 // If we know we're inflating for entry it's better to inflate by swinging a
1625 // pre-locked ObjectMonitor pointer into the object header. A successful
1626 // CAS inflates the object *and* confers ownership to the inflating thread.
1627 // In the current implementation we use a 2-step mechanism where we CAS()
1628 // to inflate and then CAS() again to try to swing _owner from null to current.
1629 // An inflateTry() method that we could call from enter() would be useful.
1630
1631 assert(mark.is_unlocked(), "invariant: header=" INTPTR_FORMAT, mark.value());
1632 ObjectMonitor* m = new ObjectMonitor(object);
1633 // prepare m for installation - set monitor to initial state
1634 m->set_header(mark);
1635
1636 if (object->cas_set_mark(markWord::encode(m), mark) != mark) {
1637 delete m;
1638 m = nullptr;
1639 continue;
1640 // interference - the markword changed - just retry.
1641 // The state-transitions are one-way, so there's no chance of
1642 // live-lock -- "Inflated" is an absorbing state.
1643 }
1644
1645 // Once the ObjectMonitor is configured and object is associated
1646 // with the ObjectMonitor, it is safe to allow async deflation:
1647 _in_use_list.add(m);
1648
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=%zu, count=%zu, max=%zu",
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=%zu"
1756 ", in_use_list stats: ceiling=%zu, count="
1757 "%zu, max=%zu",
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 "%zu, count=%zu, max=%zu",
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=%zu, {unlinked,deleted}_count=%zu monitors in %3.7f secs",
1776 deflated_count, unlinked_count, _timer.seconds());
1777 }
1778 _stream->print_cr("end deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu",
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=%zu, in_use_list stats: ceiling="
1787 "%zu, count=%zu, max=%zu",
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=%zu"
1795 ", count=%zu, max=%zu", 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 GVars.stw_random = os::random();
1877
1878 if (deflated_count != 0) {
1879 _no_progress_cnt = 0;
1880 } else if (_no_progress_skip_increment) {
1881 _no_progress_skip_increment = false;
1882 } else {
1883 _no_progress_cnt++;
1884 }
1885
1886 return deflated_count;
1887 }
1888
1889 // Monitor cleanup on JavaThread::exit
1890
1891 // Iterate through monitor cache and attempt to release thread's monitors
1892 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1893 private:
1894 JavaThread* _thread;
1895
1896 public:
1897 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {}
1898 void do_monitor(ObjectMonitor* mid) {
1899 intx rec = mid->complete_exit(_thread);
1900 _thread->dec_held_monitor_count(rec + 1);
1901 }
1902 };
1903
1904 // Release all inflated monitors owned by current thread. Lightweight monitors are
1905 // ignored. This is meant to be called during JNI thread detach which assumes
1906 // all remaining monitors are heavyweight. All exceptions are swallowed.
1907 // Scanning the extant monitor list can be time consuming.
1908 // A simple optimization is to add a per-thread flag that indicates a thread
1909 // called jni_monitorenter() during its lifetime.
1910 //
1911 // Instead of NoSafepointVerifier it might be cheaper to
1912 // use an idiom of the form:
1913 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1914 // <code that must not run at safepoint>
1915 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1916 // Since the tests are extremely cheap we could leave them enabled
1917 // for normal product builds.
1918
1919 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) {
1920 assert(current == JavaThread::current(), "must be current Java thread");
1921 NoSafepointVerifier nsv;
1922 ReleaseJavaMonitorsClosure rjmc(current);
1923 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current);
1924 assert(!current->has_pending_exception(), "Should not be possible");
1925 current->clear_pending_exception();
1926 assert(current->held_monitor_count() == 0, "Should not be possible");
1927 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count.
1928 current->clear_jni_monitor_count();
1929 }
1930
1931 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1932 switch (cause) {
1933 case inflate_cause_vm_internal: return "VM Internal";
1934 case inflate_cause_monitor_enter: return "Monitor Enter";
1935 case inflate_cause_wait: return "Monitor Wait";
1936 case inflate_cause_notify: return "Monitor Notify";
1937 case inflate_cause_hash_code: return "Monitor Hash Code";
1938 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1939 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1940 default:
1941 ShouldNotReachHere();
1942 }
1943 return "Unknown";
1944 }
1945
1946 //------------------------------------------------------------------------------
1947 // Debugging code
1948
1949 u_char* ObjectSynchronizer::get_gvars_addr() {
1950 return (u_char*)&GVars;
1951 }
1952
1953 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() {
1954 return (u_char*)&GVars.hc_sequence;
1955 }
1956
1957 size_t ObjectSynchronizer::get_gvars_size() {
1958 return sizeof(SharedGlobals);
1959 }
1960
1961 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() {
1962 return (u_char*)&GVars.stw_random;
1963 }
1964
1965 // Do the final audit and print of ObjectMonitor stats; must be done
1966 // by the VMThread at VM exit time.
1967 void ObjectSynchronizer::do_final_audit_and_print_stats() {
1968 assert(Thread::current()->is_VM_thread(), "sanity check");
1969
1970 if (is_final_audit()) { // Only do the audit once.
1971 return;
1972 }
1973 set_is_final_audit();
1974 log_info(monitorinflation)("Starting the final audit.");
1975
1976 if (log_is_enabled(Info, monitorinflation)) {
1977 LogStreamHandle(Info, monitorinflation) ls;
1978 audit_and_print_stats(&ls, true /* on_exit */);
1979 }
1980 }
1981
1982 // This function can be called by the MonitorDeflationThread or it can be called when
1983 // we are trying to exit the VM. The list walker functions can run in parallel with
1984 // the other list operations.
1985 // Calls to this function can be added in various places as a debugging
1986 // aid.
1987 //
1988 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) {
1989 int error_cnt = 0;
1990
1991 ls->print_cr("Checking in_use_list:");
1992 chk_in_use_list(ls, &error_cnt);
1993
1994 if (error_cnt == 0) {
1995 ls->print_cr("No errors found in in_use_list checks.");
1996 } else {
1997 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt);
1998 }
1999
2000 // When exiting, only log the interesting entries at the Info level.
2001 // When called at intervals by the MonitorDeflationThread, log output
2002 // at the Trace level since there can be a lot of it.
2003 if (!on_exit && log_is_enabled(Trace, monitorinflation)) {
2004 LogStreamHandle(Trace, monitorinflation) ls_tr;
2005 log_in_use_monitor_details(&ls_tr, true /* log_all */);
2006 } else if (on_exit) {
2007 log_in_use_monitor_details(ls, false /* log_all */);
2008 }
2009
2010 ls->flush();
2011
2012 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt);
2013 }
2014
2015 // Check the in_use_list; log the results of the checks.
2016 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) {
2017 size_t l_in_use_count = _in_use_list.count();
2018 size_t l_in_use_max = _in_use_list.max();
2019 out->print_cr("count=%zu, max=%zu", l_in_use_count,
2020 l_in_use_max);
2021
2022 size_t ck_in_use_count = 0;
2023 MonitorList::Iterator iter = _in_use_list.iterator();
2024 while (iter.has_next()) {
2025 ObjectMonitor* mid = iter.next();
2026 chk_in_use_entry(mid, out, error_cnt_p);
2027 ck_in_use_count++;
2028 }
2029
2030 if (l_in_use_count == ck_in_use_count) {
2031 out->print_cr("in_use_count=%zu equals ck_in_use_count=%zu",
2032 l_in_use_count, ck_in_use_count);
2033 } else {
2034 out->print_cr("WARNING: in_use_count=%zu is not equal to "
2035 "ck_in_use_count=%zu", l_in_use_count,
2036 ck_in_use_count);
2037 }
2038
2039 size_t ck_in_use_max = _in_use_list.max();
2040 if (l_in_use_max == ck_in_use_max) {
2041 out->print_cr("in_use_max=%zu equals ck_in_use_max=%zu",
2042 l_in_use_max, ck_in_use_max);
2043 } else {
2044 out->print_cr("WARNING: in_use_max=%zu is not equal to "
2045 "ck_in_use_max=%zu", l_in_use_max, ck_in_use_max);
2046 }
2047 }
2048
2049 // Check an in-use monitor entry; log any errors.
2050 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out,
2051 int* error_cnt_p) {
2052 if (n->owner_is_DEFLATER_MARKER()) {
2053 // This could happen when monitor deflation blocks for a safepoint.
2054 return;
2055 }
2056
2057
2058 if (n->metadata() == 0) {
2059 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must "
2060 "have non-null _metadata (header/hash) field.", p2i(n));
2061 *error_cnt_p = *error_cnt_p + 1;
2062 }
2063
2064 const oop obj = n->object_peek();
2065 if (obj == nullptr) {
2066 return;
2067 }
2068
2069 const markWord mark = obj->mark();
2070 if (!mark.has_monitor()) {
2071 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2072 "object does not think it has a monitor: obj="
2073 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n),
2074 p2i(obj), mark.value());
2075 *error_cnt_p = *error_cnt_p + 1;
2076 return;
2077 }
2078
2079 ObjectMonitor* const obj_mon = read_monitor(Thread::current(), obj, mark);
2080 if (n != obj_mon) {
2081 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2082 "object does not refer to the same monitor: obj="
2083 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon="
2084 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon));
2085 *error_cnt_p = *error_cnt_p + 1;
2086 }
2087 }
2088
2089 // Log details about ObjectMonitors on the in_use_list. The 'BHL'
2090 // flags indicate why the entry is in-use, 'object' and 'object type'
2091 // indicate the associated object and its type.
2092 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) {
2093 if (_in_use_list.count() > 0) {
2094 stringStream ss;
2095 out->print_cr("In-use monitor info%s:", log_all ? "" : " (eliding idle monitors)");
2096 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2097 out->print_cr("%18s %s %18s %18s",
2098 "monitor", "BHL", "object", "object type");
2099 out->print_cr("================== === ================== ==================");
2100
2101 auto is_interesting = [&](ObjectMonitor* monitor) {
2102 return log_all || monitor->has_owner() || monitor->is_busy();
2103 };
2104
2105 monitors_iterate([&](ObjectMonitor* monitor) {
2106 if (is_interesting(monitor)) {
2107 const oop obj = monitor->object_peek();
2108 const intptr_t hash = UseObjectMonitorTable ? monitor->hash() : monitor->header().hash();
2109 ResourceMark rm;
2110 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor),
2111 monitor->is_busy(), hash != 0, monitor->has_owner(),
2112 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name());
2113 if (monitor->is_busy()) {
2114 out->print(" (%s)", monitor->is_busy_to_string(&ss));
2115 ss.reset();
2116 }
2117 out->cr();
2118 }
2119 });
2120 }
2121
2122 out->flush();
2123 }