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/perfData.hpp"
52 #include "runtime/safepointMechanism.inline.hpp"
53 #include "runtime/safepointVerifiers.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #include "runtime/stubRoutines.hpp"
56 #include "runtime/synchronizer.inline.hpp"
57 #include "runtime/threads.hpp"
58 #include "runtime/timer.hpp"
59 #include "runtime/trimNativeHeap.hpp"
60 #include "runtime/vframe.hpp"
61 #include "runtime/vmThread.hpp"
62 #include "utilities/align.hpp"
63 #include "utilities/dtrace.hpp"
64 #include "utilities/events.hpp"
65 #include "utilities/globalCounter.inline.hpp"
66 #include "utilities/globalDefinitions.hpp"
67 #include "utilities/linkedlist.hpp"
68 #include "utilities/preserveException.hpp"
69
70 class ObjectMonitorDeflationLogging;
71
72 void MonitorList::add(ObjectMonitor* m) {
73 ObjectMonitor* head;
74 do {
75 head = Atomic::load(&_head);
76 m->set_next_om(head);
77 } while (Atomic::cmpxchg(&_head, head, m) != head);
78
79 size_t count = Atomic::add(&_count, 1u);
80 if (count > max()) {
81 Atomic::inc(&_max);
82 }
83 }
84
85 size_t MonitorList::count() const {
86 return Atomic::load(&_count);
87 }
88
89 size_t MonitorList::max() const {
90 return Atomic::load(&_max);
91 }
92
93 class ObjectMonitorDeflationSafepointer : public StackObj {
94 JavaThread* const _current;
95 ObjectMonitorDeflationLogging* const _log;
96
97 public:
98 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log)
99 : _current(current), _log(log) {}
100
101 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter);
102 };
103
104 // Walk the in-use list and unlink deflated ObjectMonitors.
105 // Returns the number of unlinked ObjectMonitors.
106 size_t MonitorList::unlink_deflated(size_t deflated_count,
107 GrowableArray<ObjectMonitor*>* unlinked_list,
108 ObjectMonitorDeflationSafepointer* safepointer) {
109 size_t unlinked_count = 0;
110 ObjectMonitor* prev = nullptr;
111 ObjectMonitor* m = Atomic::load_acquire(&_head);
112
113 while (m != nullptr) {
114 if (m->is_being_async_deflated()) {
115 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can
116 // modify the list once per batch. The batch starts at "m".
117 size_t unlinked_batch = 0;
118 ObjectMonitor* next = m;
119 // Look for at most MonitorUnlinkBatch monitors, or the number of
120 // deflated and not unlinked monitors, whatever comes first.
121 assert(deflated_count >= unlinked_count, "Sanity: underflow");
122 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch);
123 do {
124 ObjectMonitor* next_next = next->next_om();
125 unlinked_batch++;
126 unlinked_list->append(next);
127 next = next_next;
128 if (unlinked_batch >= unlinked_batch_limit) {
129 // Reached the max batch, so bail out of the gathering loop.
130 break;
131 }
132 if (prev == nullptr && Atomic::load(&_head) != m) {
133 // Current batch used to be at head, but it is not at head anymore.
134 // Bail out and figure out where we currently are. This avoids long
135 // walks searching for new prev during unlink under heavy list inserts.
136 break;
137 }
138 } while (next != nullptr && next->is_being_async_deflated());
139
140 // Unlink the found batch.
141 if (prev == nullptr) {
142 // The current batch is the first batch, so there is a chance that it starts at head.
143 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head.
144 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, m, next);
145 if (prev_head != m) {
146 // Something must have updated the head. Figure out the actual prev for this batch.
147 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) {
148 prev = n;
149 }
150 assert(prev != nullptr, "Should have found the prev for the current batch");
151 prev->set_next_om(next);
152 }
153 } else {
154 // The current batch is preceded by another batch. This guarantees the current batch
155 // does not start at head. Unlink the entire current batch without updating the head.
156 assert(Atomic::load(&_head) != m, "Sanity");
157 prev->set_next_om(next);
158 }
159
160 unlinked_count += unlinked_batch;
161 if (unlinked_count >= deflated_count) {
162 // Reached the max so bail out of the searching loop.
163 // There should be no more deflated monitors left.
164 break;
165 }
166 m = next;
167 } else {
168 prev = m;
169 m = m->next_om();
170 }
171
172 // Must check for a safepoint/handshake and honor it.
173 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count);
174 }
175
176 #ifdef ASSERT
177 // Invariant: the code above should unlink all deflated monitors.
178 // The code that runs after this unlinking does not expect deflated monitors.
179 // Notably, attempting to deflate the already deflated monitor would break.
180 {
181 ObjectMonitor* m = Atomic::load_acquire(&_head);
182 while (m != nullptr) {
183 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked");
184 m = m->next_om();
185 }
186 }
187 #endif
188
189 Atomic::sub(&_count, unlinked_count);
190 return unlinked_count;
191 }
192
193 MonitorList::Iterator MonitorList::iterator() const {
194 return Iterator(Atomic::load_acquire(&_head));
195 }
196
197 ObjectMonitor* MonitorList::Iterator::next() {
198 ObjectMonitor* current = _current;
199 _current = current->next_om();
200 return current;
201 }
202
203 // The "core" versions of monitor enter and exit reside in this file.
204 // The interpreter and compilers contain specialized transliterated
205 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp
206 // fast_lock(...) for instance. If you make changes here, make sure to modify the
207 // interpreter, and both C1 and C2 fast-path inline locking code emission.
208 //
209 // -----------------------------------------------------------------------------
210
211 #ifdef DTRACE_ENABLED
212
213 // Only bother with this argument setup if dtrace is available
214 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
215
216 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
217 char* bytes = nullptr; \
218 int len = 0; \
219 jlong jtid = SharedRuntime::get_java_tid(thread); \
220 Symbol* klassname = obj->klass()->name(); \
221 if (klassname != nullptr) { \
222 bytes = (char*)klassname->bytes(); \
223 len = klassname->utf8_length(); \
224 }
225
226 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
227 { \
228 if (DTraceMonitorProbes) { \
229 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
230 HOTSPOT_MONITOR_WAIT(jtid, \
231 (uintptr_t)(monitor), bytes, len, (millis)); \
232 } \
233 }
234
235 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY
236 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL
237 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
238
239 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
240 { \
241 if (DTraceMonitorProbes) { \
242 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
243 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
244 (uintptr_t)(monitor), bytes, len); \
245 } \
246 }
247
248 #else // ndef DTRACE_ENABLED
249
250 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
251 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
252
253 #endif // ndef DTRACE_ENABLED
254
255 // This exists only as a workaround of dtrace bug 6254741
256 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) {
257 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
258 return 0;
259 }
260
261 static constexpr size_t inflation_lock_count() {
262 return 256;
263 }
264
265 // Static storage for an array of PlatformMutex.
266 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)];
267
268 static inline PlatformMutex* inflation_lock(size_t index) {
269 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]);
270 }
271
272 void ObjectSynchronizer::initialize() {
273 for (size_t i = 0; i < inflation_lock_count(); i++) {
274 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex();
275 }
276 // Start the ceiling with the estimate for one thread.
277 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate);
278
279 // Start the timer for deflations, so it does not trigger immediately.
280 _last_async_deflation_time_ns = os::javaTimeNanos();
281
282 if (LockingMode == LM_LIGHTWEIGHT) {
283 LightweightSynchronizer::initialize();
284 }
285 }
286
287 MonitorList ObjectSynchronizer::_in_use_list;
288 // monitors_used_above_threshold() policy is as follows:
289 //
290 // The ratio of the current _in_use_list count to the ceiling is used
291 // to determine if we are above MonitorUsedDeflationThreshold and need
292 // to do an async monitor deflation cycle. The ceiling is increased by
293 // AvgMonitorsPerThreadEstimate when a thread is added to the system
294 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is
295 // removed from the system.
296 //
297 // Note: If the _in_use_list max exceeds the ceiling, then
298 // monitors_used_above_threshold() will use the in_use_list max instead
299 // of the thread count derived ceiling because we have used more
300 // ObjectMonitors than the estimated average.
301 //
302 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax
303 // no-progress async monitor deflation cycles in a row, then the ceiling
304 // is adjusted upwards by monitors_used_above_threshold().
305 //
306 // Start the ceiling with the estimate for one thread in initialize()
307 // which is called after cmd line options are processed.
308 static size_t _in_use_list_ceiling = 0;
309 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false;
310 bool volatile ObjectSynchronizer::_is_final_audit = false;
311 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0;
312 static uintx _no_progress_cnt = 0;
313 static bool _no_progress_skip_increment = false;
314
315 // =====================> Quick functions
316
317 // The quick_* forms are special fast-path variants used to improve
318 // performance. In the simplest case, a "quick_*" implementation could
319 // simply return false, in which case the caller will perform the necessary
320 // state transitions and call the slow-path form.
321 // The fast-path is designed to handle frequently arising cases in an efficient
322 // manner and is just a degenerate "optimistic" variant of the slow-path.
323 // returns true -- to indicate the call was satisfied.
324 // returns false -- to indicate the call needs the services of the slow-path.
325 // A no-loitering ordinance is in effect for code in the quick_* family
326 // operators: safepoints or indefinite blocking (blocking that might span a
327 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon
328 // entry.
329 //
330 // Consider: An interesting optimization is to have the JIT recognize the
331 // following common idiom:
332 // synchronized (someobj) { .... ; notify(); }
333 // That is, we find a notify() or notifyAll() call that immediately precedes
334 // the monitorexit operation. In that case the JIT could fuse the operations
335 // into a single notifyAndExit() runtime primitive.
336
337 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) {
338 assert(current->thread_state() == _thread_in_Java, "invariant");
339 NoSafepointVerifier nsv;
340 if (obj == nullptr) return false; // slow-path for invalid obj
341 const markWord mark = obj->mark();
342
343 if (LockingMode == LM_LIGHTWEIGHT) {
344 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) {
345 // Degenerate notify
346 // fast-locked by caller so by definition the implied waitset is empty.
347 return true;
348 }
349 } else if (LockingMode == LM_LEGACY) {
350 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
351 // Degenerate notify
352 // stack-locked by caller so by definition the implied waitset is empty.
353 return true;
354 }
355 }
356
357 if (mark.has_monitor()) {
358 ObjectMonitor* const mon = read_monitor(current, obj, mark);
359 if (LockingMode == LM_LIGHTWEIGHT && mon == nullptr) {
360 // Racing with inflation/deflation go slow path
361 return false;
362 }
363 assert(mon->object() == oop(obj), "invariant");
364 if (!mon->has_owner(current)) return false; // slow-path for IMS exception
365
366 if (mon->first_waiter() != nullptr) {
367 // We have one or more waiters. Since this is an inflated monitor
368 // that we own, we can transfer one or more threads from the waitset
369 // to the entrylist here and now, avoiding the slow-path.
370 if (all) {
371 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, current);
372 } else {
373 DTRACE_MONITOR_PROBE(notify, mon, obj, current);
374 }
375 int free_count = 0;
376 do {
377 mon->INotify(current);
378 ++free_count;
379 } while (mon->first_waiter() != nullptr && all);
380 OM_PERFDATA_OP(Notifications, inc(free_count));
381 }
382 return true;
383 }
384
385 // other IMS exception states take the slow-path
386 return false;
387 }
388
389 static bool useHeavyMonitors() {
390 #if defined(X86) || defined(AARCH64) || defined(PPC64) || defined(RISCV64) || defined(S390)
391 return LockingMode == LM_MONITOR;
392 #else
393 return false;
394 #endif
395 }
396
397 // The LockNode emitted directly at the synchronization site would have
398 // been too big if it were to have included support for the cases of inflated
399 // recursive enter and exit, so they go here instead.
400 // Note that we can't safely call AsyncPrintJavaStack() from within
401 // quick_enter() as our thread state remains _in_Java.
402
403 bool ObjectSynchronizer::quick_enter_legacy(oop obj, BasicLock* lock, JavaThread* current) {
404 assert(current->thread_state() == _thread_in_Java, "invariant");
405
406 if (useHeavyMonitors()) {
407 return false; // Slow path
408 }
409
410 if (LockingMode == LM_LIGHTWEIGHT) {
411 return LightweightSynchronizer::quick_enter(obj, lock, current);
412 }
413
414 assert(LockingMode == LM_LEGACY, "legacy mode below");
415
416 const markWord mark = obj->mark();
417
418 if (mark.has_monitor()) {
419
420 ObjectMonitor* const m = read_monitor(mark);
421 // An async deflation or GC can race us before we manage to make
422 // the ObjectMonitor busy by setting the owner below. If we detect
423 // that race we just bail out to the slow-path here.
424 if (m->object_peek() == nullptr) {
425 return false;
426 }
427
428 // Lock contention and Transactional Lock Elision (TLE) diagnostics
429 // and observability
430 // Case: light contention possibly amenable to TLE
431 // Case: TLE inimical operations such as nested/recursive synchronization
432
433 if (m->has_owner(current)) {
434 m->_recursions++;
435 current->inc_held_monitor_count();
436 return true;
437 }
438
439 // This Java Monitor is inflated so obj's header will never be
440 // displaced to this thread's BasicLock. Make the displaced header
441 // non-null so this BasicLock is not seen as recursive nor as
442 // being locked. We do this unconditionally so that this thread's
443 // BasicLock cannot be mis-interpreted by any stack walkers. For
444 // performance reasons, stack walkers generally first check for
445 // stack-locking in the object's header, the second check is for
446 // recursive stack-locking in the displaced header in the BasicLock,
447 // and last are the inflated Java Monitor (ObjectMonitor) checks.
448 lock->set_displaced_header(markWord::unused_mark());
449
450 if (!m->has_owner() && m->try_set_owner(current)) {
451 assert(m->_recursions == 0, "invariant");
452 current->inc_held_monitor_count();
453 return true;
454 }
455 }
456
457 // Note that we could inflate in quick_enter.
458 // This is likely a useful optimization
459 // Critically, in quick_enter() we must not:
460 // -- block indefinitely, or
461 // -- reach a safepoint
462
463 return false; // revert to slow-path
464 }
465
466 // Handle notifications when synchronizing on value based classes
467 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) {
468 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
469 frame last_frame = locking_thread->last_frame();
470 bool bcp_was_adjusted = false;
471 // Don't decrement bcp if it points to the frame's first instruction. This happens when
472 // handle_sync_on_value_based_class() is called because of a synchronized method. There
473 // is no actual monitorenter instruction in the byte code in this case.
474 if (last_frame.is_interpreted_frame() &&
475 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) {
476 // adjust bcp to point back to monitorenter so that we print the correct line numbers
477 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1);
478 bcp_was_adjusted = true;
479 }
480
481 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) {
482 ResourceMark rm;
483 stringStream ss;
484 locking_thread->print_active_stack_on(&ss);
485 char* base = (char*)strstr(ss.base(), "at");
486 char* newline = (char*)strchr(ss.base(), '\n');
487 if (newline != nullptr) {
488 *newline = '\0';
489 }
490 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base);
491 } else {
492 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses");
493 ResourceMark rm;
494 Log(valuebasedclasses) vblog;
495
496 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name());
497 if (locking_thread->has_last_Java_frame()) {
498 LogStream info_stream(vblog.info());
499 locking_thread->print_active_stack_on(&info_stream);
500 } else {
501 vblog.info("Cannot find the last Java frame");
502 }
503
504 EventSyncOnValueBasedClass event;
505 if (event.should_commit()) {
506 event.set_valueBasedClass(obj->klass());
507 event.commit();
508 }
509 }
510
511 if (bcp_was_adjusted) {
512 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1);
513 }
514 }
515
516 // -----------------------------------------------------------------------------
517 // Monitor Enter/Exit
518
519 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
520 // When called with locking_thread != Thread::current() some mechanism must synchronize
521 // the locking_thread with respect to the current thread. Currently only used when
522 // deoptimizing and re-locking locks. See Deoptimization::relock_objects
523 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
524
525 if (LockingMode == LM_LIGHTWEIGHT) {
526 return LightweightSynchronizer::enter_for(obj, lock, locking_thread);
527 }
528
529 if (!enter_fast_impl(obj, lock, locking_thread)) {
530 // Inflated ObjectMonitor::enter_for is required
531
532 // An async deflation can race after the inflate_for() call and before
533 // enter_for() can make the ObjectMonitor busy. enter_for() returns false
534 // if we have lost the race to async deflation and we simply try again.
535 while (true) {
536 ObjectMonitor* monitor = inflate_for(locking_thread, obj(), inflate_cause_monitor_enter);
537 if (monitor->enter_for(locking_thread)) {
538 return;
539 }
540 assert(monitor->is_being_async_deflated(), "must be");
541 }
542 }
543 }
544
545 void ObjectSynchronizer::enter_legacy(Handle obj, BasicLock* lock, JavaThread* current) {
546 if (!enter_fast_impl(obj, lock, current)) {
547 // Inflated ObjectMonitor::enter is required
548
549 // An async deflation can race after the inflate() call and before
550 // enter() can make the ObjectMonitor busy. enter() returns false if
551 // we have lost the race to async deflation and we simply try again.
552 while (true) {
553 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter);
554 if (monitor->enter(current)) {
555 return;
556 }
557 }
558 }
559 }
560
561 // The interpreter and compiler assembly code tries to lock using the fast path
562 // of this algorithm. Make sure to update that code if the following function is
563 // changed. The implementation is extremely sensitive to race condition. Be careful.
564 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
565 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer");
566
567 if (obj->klass()->is_value_based()) {
568 handle_sync_on_value_based_class(obj, locking_thread);
569 }
570
571 locking_thread->inc_held_monitor_count();
572
573 if (!useHeavyMonitors()) {
574 if (LockingMode == LM_LEGACY) {
575 markWord mark = obj->mark();
576 if (mark.is_unlocked()) {
577 // Anticipate successful CAS -- the ST of the displaced mark must
578 // be visible <= the ST performed by the CAS.
579 lock->set_displaced_header(mark);
580 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) {
581 return true;
582 }
583 } else if (mark.has_locker() &&
584 locking_thread->is_lock_owned((address) mark.locker())) {
585 assert(lock != mark.locker(), "must not re-lock the same lock");
586 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock");
587 lock->set_displaced_header(markWord::from_pointer(nullptr));
588 return true;
589 }
590
591 // The object header will never be displaced to this lock,
592 // so it does not matter what the value is, except that it
593 // must be non-zero to avoid looking like a re-entrant lock,
594 // and must not look locked either.
595 lock->set_displaced_header(markWord::unused_mark());
596
597 // Failed to fast lock.
598 return false;
599 }
600 } else if (VerifyHeavyMonitors) {
601 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
602 }
603
604 return false;
605 }
606
607 void ObjectSynchronizer::exit_legacy(oop object, BasicLock* lock, JavaThread* current) {
608 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer");
609
610 if (!useHeavyMonitors()) {
611 markWord mark = object->mark();
612 if (LockingMode == LM_LEGACY) {
613 markWord dhw = lock->displaced_header();
614 if (dhw.value() == 0) {
615 // If the displaced header is null, then this exit matches up with
616 // a recursive enter. No real work to do here except for diagnostics.
617 #ifndef PRODUCT
618 if (mark != markWord::INFLATING()) {
619 // Only do diagnostics if we are not racing an inflation. Simply
620 // exiting a recursive enter of a Java Monitor that is being
621 // inflated is safe; see the has_monitor() comment below.
622 assert(!mark.is_unlocked(), "invariant");
623 assert(!mark.has_locker() ||
624 current->is_lock_owned((address)mark.locker()), "invariant");
625 if (mark.has_monitor()) {
626 // The BasicLock's displaced_header is marked as a recursive
627 // enter and we have an inflated Java Monitor (ObjectMonitor).
628 // This is a special case where the Java Monitor was inflated
629 // after this thread entered the stack-lock recursively. When a
630 // Java Monitor is inflated, we cannot safely walk the Java
631 // Monitor owner's stack and update the BasicLocks because a
632 // Java Monitor can be asynchronously inflated by a thread that
633 // does not own the Java Monitor.
634 ObjectMonitor* m = read_monitor(mark);
635 assert(m->object()->mark() == mark, "invariant");
636 assert(m->is_entered(current), "invariant");
637 }
638 }
639 #endif
640 return;
641 }
642
643 if (mark == markWord::from_pointer(lock)) {
644 // If the object is stack-locked by the current thread, try to
645 // swing the displaced header from the BasicLock back to the mark.
646 assert(dhw.is_neutral(), "invariant");
647 if (object->cas_set_mark(dhw, mark) == mark) {
648 return;
649 }
650 }
651 }
652 } else if (VerifyHeavyMonitors) {
653 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
654 }
655
656 // We have to take the slow-path of possible inflation and then exit.
657 // The ObjectMonitor* can't be async deflated until ownership is
658 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
659 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal);
660 assert(!monitor->has_anonymous_owner(), "must not be");
661 monitor->exit(current);
662 }
663
664 // -----------------------------------------------------------------------------
665 // JNI locks on java objects
666 // NOTE: must use heavy weight monitor to handle jni monitor enter
667 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) {
668 // Top native frames in the stack will not be seen if we attempt
669 // preemption, since we start walking from the last Java anchor.
670 NoPreemptMark npm(current);
671
672 if (obj->klass()->is_value_based()) {
673 handle_sync_on_value_based_class(obj, current);
674 }
675
676 // the current locking is from JNI instead of Java code
677 current->set_current_pending_monitor_is_from_java(false);
678 // An async deflation can race after the inflate() call and before
679 // enter() can make the ObjectMonitor busy. enter() returns false if
680 // we have lost the race to async deflation and we simply try again.
681 while (true) {
682 ObjectMonitor* monitor;
683 bool entered;
684 if (LockingMode == LM_LIGHTWEIGHT) {
685 entered = LightweightSynchronizer::inflate_and_enter(obj(), inflate_cause_jni_enter, current, current) != nullptr;
686 } else {
687 monitor = inflate(current, obj(), inflate_cause_jni_enter);
688 entered = monitor->enter(current);
689 }
690
691 if (entered) {
692 current->inc_held_monitor_count(1, true);
693 break;
694 }
695 }
696 current->set_current_pending_monitor_is_from_java(true);
697 }
698
699 // NOTE: must use heavy weight monitor to handle jni monitor exit
700 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) {
701 JavaThread* current = THREAD;
702
703 ObjectMonitor* monitor;
704 if (LockingMode == LM_LIGHTWEIGHT) {
705 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj, inflate_cause_jni_exit, CHECK);
706 } else {
707 // The ObjectMonitor* can't be async deflated until ownership is
708 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
709 monitor = inflate(current, obj, inflate_cause_jni_exit);
710 }
711 // If this thread has locked the object, exit the monitor. We
712 // intentionally do not use CHECK on check_owner because we must exit the
713 // monitor even if an exception was already pending.
714 if (monitor->check_owner(THREAD)) {
715 monitor->exit(current);
716 current->dec_held_monitor_count(1, true);
717 }
718 }
719
720 // -----------------------------------------------------------------------------
721 // Internal VM locks on java objects
722 // standard constructor, allows locking failures
723 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) : _npm(thread) {
724 _thread = thread;
725 _thread->check_for_valid_safepoint_state();
726 _obj = obj;
727
728 if (_obj() != nullptr) {
729 ObjectSynchronizer::enter(_obj, &_lock, _thread);
730 }
731 }
732
733 ObjectLocker::~ObjectLocker() {
734 if (_obj() != nullptr) {
735 ObjectSynchronizer::exit(_obj(), &_lock, _thread);
736 }
737 }
738
739
740 // -----------------------------------------------------------------------------
741 // Wait/Notify/NotifyAll
742 // NOTE: must use heavy weight monitor to handle wait()
743
744 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
745 JavaThread* current = THREAD;
746 if (millis < 0) {
747 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
748 }
749
750 ObjectMonitor* monitor;
751 if (LockingMode == LM_LIGHTWEIGHT) {
752 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK_0);
753 } else {
754 // The ObjectMonitor* can't be async deflated because the _waiters
755 // field is incremented before ownership is dropped and decremented
756 // after ownership is regained.
757 monitor = inflate(current, obj(), inflate_cause_wait);
758 }
759
760 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis);
761 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code
762
763 // This dummy call is in place to get around dtrace bug 6254741. Once
764 // that's fixed we can uncomment the following line, remove the call
765 // and change this function back into a "void" func.
766 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
767 int ret_code = dtrace_waited_probe(monitor, obj, THREAD);
768 return ret_code;
769 }
770
771 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
772 if (millis < 0) {
773 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
774 }
775
776 ObjectMonitor* monitor;
777 if (LockingMode == LM_LIGHTWEIGHT) {
778 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK);
779 } else {
780 monitor = inflate(THREAD, obj(), inflate_cause_wait);
781 }
782 monitor->wait(millis, false, THREAD);
783 }
784
785
786 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
787 JavaThread* current = THREAD;
788
789 markWord mark = obj->mark();
790 if (LockingMode == LM_LIGHTWEIGHT) {
791 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
792 // Not inflated so there can't be any waiters to notify.
793 return;
794 }
795 } else if (LockingMode == LM_LEGACY) {
796 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
797 // Not inflated so there can't be any waiters to notify.
798 return;
799 }
800 }
801
802 ObjectMonitor* monitor;
803 if (LockingMode == LM_LIGHTWEIGHT) {
804 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK);
805 } else {
806 // The ObjectMonitor* can't be async deflated until ownership is
807 // dropped by the calling thread.
808 monitor = inflate(current, obj(), inflate_cause_notify);
809 }
810 monitor->notify(CHECK);
811 }
812
813 // NOTE: see comment of notify()
814 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
815 JavaThread* current = THREAD;
816
817 markWord mark = obj->mark();
818 if (LockingMode == LM_LIGHTWEIGHT) {
819 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
820 // Not inflated so there can't be any waiters to notify.
821 return;
822 }
823 } else if (LockingMode == LM_LEGACY) {
824 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
825 // Not inflated so there can't be any waiters to notify.
826 return;
827 }
828 }
829
830 ObjectMonitor* monitor;
831 if (LockingMode == LM_LIGHTWEIGHT) {
832 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK);
833 } else {
834 // The ObjectMonitor* can't be async deflated until ownership is
835 // dropped by the calling thread.
836 monitor = inflate(current, obj(), inflate_cause_notify);
837 }
838 monitor->notifyAll(CHECK);
839 }
840
841 // -----------------------------------------------------------------------------
842 // Hash Code handling
843
844 struct SharedGlobals {
845 char _pad_prefix[OM_CACHE_LINE_SIZE];
846 // This is a highly shared mostly-read variable.
847 // To avoid false-sharing it needs to be the sole occupant of a cache line.
848 volatile int stw_random;
849 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int));
850 // Hot RW variable -- Sequester to avoid false-sharing
851 volatile int hc_sequence;
852 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int));
853 };
854
855 static SharedGlobals GVars;
856
857 static markWord read_stable_mark(oop obj) {
858 markWord mark = obj->mark_acquire();
859 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) {
860 // New lightweight locking does not use the markWord::INFLATING() protocol.
861 return mark; // normal fast-path return
862 }
863
864 int its = 0;
865 for (;;) {
866 markWord mark = obj->mark_acquire();
867 if (!mark.is_being_inflated()) {
868 return mark; // normal fast-path return
869 }
870
871 // The object is being inflated by some other thread.
872 // The caller of read_stable_mark() must wait for inflation to complete.
873 // Avoid live-lock.
874
875 ++its;
876 if (its > 10000 || !os::is_MP()) {
877 if (its & 1) {
878 os::naked_yield();
879 } else {
880 // Note that the following code attenuates the livelock problem but is not
881 // a complete remedy. A more complete solution would require that the inflating
882 // thread hold the associated inflation lock. The following code simply restricts
883 // the number of spinners to at most one. We'll have N-2 threads blocked
884 // on the inflationlock, 1 thread holding the inflation lock and using
885 // a yield/park strategy, and 1 thread in the midst of inflation.
886 // A more refined approach would be to change the encoding of INFLATING
887 // to allow encapsulation of a native thread pointer. Threads waiting for
888 // inflation to complete would use CAS to push themselves onto a singly linked
889 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
890 // and calling park(). When inflation was complete the thread that accomplished inflation
891 // would detach the list and set the markword to inflated with a single CAS and
892 // then for each thread on the list, set the flag and unpark() the thread.
893
894 // Index into the lock array based on the current object address.
895 static_assert(is_power_of_2(inflation_lock_count()), "must be");
896 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1);
897 int YieldThenBlock = 0;
898 assert(ix < inflation_lock_count(), "invariant");
899 inflation_lock(ix)->lock();
900 while (obj->mark_acquire() == markWord::INFLATING()) {
901 // Beware: naked_yield() is advisory and has almost no effect on some platforms
902 // so we periodically call current->_ParkEvent->park(1).
903 // We use a mixed spin/yield/block mechanism.
904 if ((YieldThenBlock++) >= 16) {
905 Thread::current()->_ParkEvent->park(1);
906 } else {
907 os::naked_yield();
908 }
909 }
910 inflation_lock(ix)->unlock();
911 }
912 } else {
913 SpinPause(); // SMP-polite spinning
914 }
915 }
916 }
917
918 // hashCode() generation :
919 //
920 // Possibilities:
921 // * MD5Digest of {obj,stw_random}
922 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function.
923 // * A DES- or AES-style SBox[] mechanism
924 // * One of the Phi-based schemes, such as:
925 // 2654435761 = 2^32 * Phi (golden ratio)
926 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ;
927 // * A variation of Marsaglia's shift-xor RNG scheme.
928 // * (obj ^ stw_random) is appealing, but can result
929 // in undesirable regularity in the hashCode values of adjacent objects
930 // (objects allocated back-to-back, in particular). This could potentially
931 // result in hashtable collisions and reduced hashtable efficiency.
932 // There are simple ways to "diffuse" the middle address bits over the
933 // generated hashCode values:
934
935 static intptr_t get_next_hash(Thread* current, oop obj) {
936 intptr_t value = 0;
937 if (hashCode == 0) {
938 // This form uses global Park-Miller RNG.
939 // On MP system we'll have lots of RW access to a global, so the
940 // mechanism induces lots of coherency traffic.
941 value = os::random();
942 } else if (hashCode == 1) {
943 // This variation has the property of being stable (idempotent)
944 // between STW operations. This can be useful in some of the 1-0
945 // synchronization schemes.
946 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3;
947 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random;
948 } else if (hashCode == 2) {
949 value = 1; // for sensitivity testing
950 } else if (hashCode == 3) {
951 value = ++GVars.hc_sequence;
952 } else if (hashCode == 4) {
953 value = cast_from_oop<intptr_t>(obj);
954 } else {
955 // Marsaglia's xor-shift scheme with thread-specific state
956 // This is probably the best overall implementation -- we'll
957 // likely make this the default in future releases.
958 unsigned t = current->_hashStateX;
959 t ^= (t << 11);
960 current->_hashStateX = current->_hashStateY;
961 current->_hashStateY = current->_hashStateZ;
962 current->_hashStateZ = current->_hashStateW;
963 unsigned v = current->_hashStateW;
964 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
965 current->_hashStateW = v;
966 value = v;
967 }
968
969 value &= markWord::hash_mask;
970 if (value == 0) value = 0xBAD;
971 assert(value != markWord::no_hash, "invariant");
972 return value;
973 }
974
975 static intptr_t install_hash_code(Thread* current, oop obj) {
976 assert(UseObjectMonitorTable && LockingMode == LM_LIGHTWEIGHT, "must be");
977
978 markWord mark = obj->mark_acquire();
979 for (;;) {
980 intptr_t hash = mark.hash();
981 if (hash != 0) {
982 return hash;
983 }
984
985 hash = get_next_hash(current, obj);
986 const markWord old_mark = mark;
987 const markWord new_mark = old_mark.copy_set_hash(hash);
988
989 mark = obj->cas_set_mark(new_mark, old_mark);
990 if (old_mark == mark) {
991 return hash;
992 }
993 }
994 }
995
996 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) {
997 if (UseObjectMonitorTable) {
998 // Since the monitor isn't in the object header, the hash can simply be
999 // installed in the object header.
1000 return install_hash_code(current, obj);
1001 }
1002
1003 while (true) {
1004 ObjectMonitor* monitor = nullptr;
1005 markWord temp, test;
1006 intptr_t hash;
1007 markWord mark = read_stable_mark(obj);
1008 if (VerifyHeavyMonitors) {
1009 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)");
1010 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
1011 }
1012 if (mark.is_unlocked() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) {
1013 hash = mark.hash();
1014 if (hash != 0) { // if it has a hash, just return it
1015 return hash;
1016 }
1017 hash = get_next_hash(current, obj); // get a new hash
1018 temp = mark.copy_set_hash(hash); // merge the hash into header
1019 // try to install the hash
1020 test = obj->cas_set_mark(temp, mark);
1021 if (test == mark) { // if the hash was installed, return it
1022 return hash;
1023 }
1024 if (LockingMode == LM_LIGHTWEIGHT) {
1025 // CAS failed, retry
1026 continue;
1027 }
1028 // Failed to install the hash. It could be that another thread
1029 // installed the hash just before our attempt or inflation has
1030 // occurred or... so we fall thru to inflate the monitor for
1031 // stability and then install the hash.
1032 } else if (mark.has_monitor()) {
1033 monitor = mark.monitor();
1034 temp = monitor->header();
1035 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1036 hash = temp.hash();
1037 if (hash != 0) {
1038 // It has a hash.
1039
1040 // Separate load of dmw/header above from the loads in
1041 // is_being_async_deflated().
1042
1043 // dmw/header and _contentions may get written by different threads.
1044 // Make sure to observe them in the same order when having several observers.
1045 OrderAccess::loadload_for_IRIW();
1046
1047 if (monitor->is_being_async_deflated()) {
1048 // But we can't safely use the hash if we detect that async
1049 // deflation has occurred. So we attempt to restore the
1050 // header/dmw to the object's header so that we only retry
1051 // once if the deflater thread happens to be slow.
1052 monitor->install_displaced_markword_in_object(obj);
1053 continue;
1054 }
1055 return hash;
1056 }
1057 // Fall thru so we only have one place that installs the hash in
1058 // the ObjectMonitor.
1059 } else if (LockingMode == LM_LEGACY && mark.has_locker()
1060 && current->is_Java_thread()
1061 && JavaThread::cast(current)->is_lock_owned((address)mark.locker())) {
1062 // This is a stack-lock owned by the calling thread so fetch the
1063 // displaced markWord from the BasicLock on the stack.
1064 temp = mark.displaced_mark_helper();
1065 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1066 hash = temp.hash();
1067 if (hash != 0) { // if it has a hash, just return it
1068 return hash;
1069 }
1070 // WARNING:
1071 // The displaced header in the BasicLock on a thread's stack
1072 // is strictly immutable. It CANNOT be changed in ANY cases.
1073 // So we have to inflate the stack-lock into an ObjectMonitor
1074 // even if the current thread owns the lock. The BasicLock on
1075 // a thread's stack can be asynchronously read by other threads
1076 // during an inflate() call so any change to that stack memory
1077 // may not propagate to other threads correctly.
1078 }
1079
1080 // Inflate the monitor to set the hash.
1081
1082 // There's no need to inflate if the mark has already got a monitor.
1083 // NOTE: an async deflation can race after we get the monitor and
1084 // before we can update the ObjectMonitor's header with the hash
1085 // value below.
1086 monitor = mark.has_monitor() ? mark.monitor() : inflate(current, obj, inflate_cause_hash_code);
1087 // Load ObjectMonitor's header/dmw field and see if it has a hash.
1088 mark = monitor->header();
1089 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
1090 hash = mark.hash();
1091 if (hash == 0) { // if it does not have a hash
1092 hash = get_next_hash(current, obj); // get a new hash
1093 temp = mark.copy_set_hash(hash) ; // merge the hash into header
1094 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1095 uintptr_t v = Atomic::cmpxchg(monitor->metadata_addr(), mark.value(), temp.value());
1096 test = markWord(v);
1097 if (test != mark) {
1098 // The attempt to update the ObjectMonitor's header/dmw field
1099 // did not work. This can happen if another thread managed to
1100 // merge in the hash just before our cmpxchg().
1101 // If we add any new usages of the header/dmw field, this code
1102 // will need to be updated.
1103 hash = test.hash();
1104 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value());
1105 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash");
1106 }
1107 if (monitor->is_being_async_deflated() && !UseObjectMonitorTable) {
1108 // If we detect that async deflation has occurred, then we
1109 // attempt to restore the header/dmw to the object's header
1110 // so that we only retry once if the deflater thread happens
1111 // to be slow.
1112 monitor->install_displaced_markword_in_object(obj);
1113 continue;
1114 }
1115 }
1116 // We finally get the hash.
1117 return hash;
1118 }
1119 }
1120
1121 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current,
1122 Handle h_obj) {
1123 assert(current == JavaThread::current(), "Can only be called on current thread");
1124 oop obj = h_obj();
1125
1126 markWord mark = read_stable_mark(obj);
1127
1128 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1129 // stack-locked case, header points into owner's stack
1130 return current->is_lock_owned((address)mark.locker());
1131 }
1132
1133 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1134 // fast-locking case, see if lock is in current's lock stack
1135 return current->lock_stack().contains(h_obj());
1136 }
1137
1138 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) {
1139 ObjectMonitor* monitor = read_monitor(current, obj, mark);
1140 if (monitor != nullptr) {
1141 return monitor->is_entered(current) != 0;
1142 }
1143 // Racing with inflation/deflation, retry
1144 mark = obj->mark_acquire();
1145
1146 if (mark.is_fast_locked()) {
1147 // Some other thread fast_locked, current could not have held the lock
1148 return false;
1149 }
1150 }
1151
1152 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) {
1153 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
1154 // The first stage of async deflation does not affect any field
1155 // used by this comparison so the ObjectMonitor* is usable here.
1156 ObjectMonitor* monitor = read_monitor(mark);
1157 return monitor->is_entered(current) != 0;
1158 }
1159 // Unlocked case, header in place
1160 assert(mark.is_unlocked(), "sanity check");
1161 return false;
1162 }
1163
1164 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
1165 oop obj = h_obj();
1166 markWord mark = read_stable_mark(obj);
1167
1168 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1169 // stack-locked so header points into owner's stack.
1170 // owning_thread_from_monitor_owner() may also return null here:
1171 return Threads::owning_thread_from_stacklock(t_list, (address) mark.locker());
1172 }
1173
1174 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1175 // fast-locked so get owner from the object.
1176 // owning_thread_from_object() may also return null here:
1177 return Threads::owning_thread_from_object(t_list, h_obj());
1178 }
1179
1180 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) {
1181 ObjectMonitor* monitor = read_monitor(Thread::current(), obj, mark);
1182 if (monitor != nullptr) {
1183 return Threads::owning_thread_from_monitor(t_list, monitor);
1184 }
1185 // Racing with inflation/deflation, retry
1186 mark = obj->mark_acquire();
1187
1188 if (mark.is_fast_locked()) {
1189 // Some other thread fast_locked
1190 return Threads::owning_thread_from_object(t_list, h_obj());
1191 }
1192 }
1193
1194 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) {
1195 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
1196 // The first stage of async deflation does not affect any field
1197 // used by this comparison so the ObjectMonitor* is usable here.
1198 ObjectMonitor* monitor = read_monitor(mark);
1199 assert(monitor != nullptr, "monitor should be non-null");
1200 // owning_thread_from_monitor() may also return null here:
1201 return Threads::owning_thread_from_monitor(t_list, monitor);
1202 }
1203
1204 // Unlocked case, header in place
1205 // Cannot have assertion since this object may have been
1206 // locked by another thread when reaching here.
1207 // assert(mark.is_unlocked(), "sanity check");
1208
1209 return nullptr;
1210 }
1211
1212 // Visitors ...
1213
1214 // Iterate over all ObjectMonitors.
1215 template <typename Function>
1216 void ObjectSynchronizer::monitors_iterate(Function function) {
1217 MonitorList::Iterator iter = _in_use_list.iterator();
1218 while (iter.has_next()) {
1219 ObjectMonitor* monitor = iter.next();
1220 function(monitor);
1221 }
1222 }
1223
1224 // Iterate ObjectMonitors owned by any thread and where the owner `filter`
1225 // returns true.
1226 template <typename OwnerFilter>
1227 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) {
1228 monitors_iterate([&](ObjectMonitor* monitor) {
1229 // This function is only called at a safepoint or when the
1230 // target thread is suspended or when the target thread is
1231 // operating on itself. The current closures in use today are
1232 // only interested in an owned ObjectMonitor and ownership
1233 // cannot be dropped under the calling contexts so the
1234 // ObjectMonitor cannot be async deflated.
1235 if (monitor->has_owner() && filter(monitor)) {
1236 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating");
1237
1238 closure->do_monitor(monitor);
1239 }
1240 });
1241 }
1242
1243 // Iterate ObjectMonitors where the owner == thread; this does NOT include
1244 // ObjectMonitors where owner is set to a stack-lock address in thread.
1245 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) {
1246 int64_t key = ObjectMonitor::owner_id_from(thread);
1247 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; };
1248 return owned_monitors_iterate_filtered(closure, thread_filter);
1249 }
1250
1251 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, oop vthread) {
1252 int64_t key = ObjectMonitor::owner_id_from(vthread);
1253 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->owner() == key; };
1254 return owned_monitors_iterate_filtered(closure, thread_filter);
1255 }
1256
1257 // Iterate ObjectMonitors owned by any thread.
1258 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) {
1259 auto all_filter = [&](ObjectMonitor* monitor) { return true; };
1260 return owned_monitors_iterate_filtered(closure, all_filter);
1261 }
1262
1263 static bool monitors_used_above_threshold(MonitorList* list) {
1264 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy
1265 return false;
1266 }
1267 size_t monitors_used = list->count();
1268 if (monitors_used == 0) { // empty list is easy
1269 return false;
1270 }
1271 size_t old_ceiling = ObjectSynchronizer::in_use_list_ceiling();
1272 // Make sure that we use a ceiling value that is not lower than
1273 // previous, not lower than the recorded max used by the system, and
1274 // not lower than the current number of monitors in use (which can
1275 // race ahead of max). The result is guaranteed > 0.
1276 size_t ceiling = MAX3(old_ceiling, list->max(), monitors_used);
1277
1278 // Check if our monitor usage is above the threshold:
1279 size_t monitor_usage = (monitors_used * 100LL) / ceiling;
1280 if (int(monitor_usage) > MonitorUsedDeflationThreshold) {
1281 // Deflate monitors if over the threshold percentage, unless no
1282 // progress on previous deflations.
1283 bool is_above_threshold = true;
1284
1285 // Check if it's time to adjust the in_use_list_ceiling up, due
1286 // to too many async deflation attempts without any progress.
1287 if (NoAsyncDeflationProgressMax != 0 &&
1288 _no_progress_cnt >= NoAsyncDeflationProgressMax) {
1289 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0;
1290 size_t delta = (size_t)(ceiling * remainder) + 1;
1291 size_t new_ceiling = (ceiling > SIZE_MAX - delta)
1292 ? SIZE_MAX // Overflow, let's clamp new_ceiling.
1293 : ceiling + delta;
1294
1295 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling);
1296 log_info(monitorinflation)("Too many deflations without progress; "
1297 "bumping in_use_list_ceiling from %zu"
1298 " to %zu", old_ceiling, new_ceiling);
1299 _no_progress_cnt = 0;
1300 ceiling = new_ceiling;
1301
1302 // Check if our monitor usage is still above the threshold:
1303 monitor_usage = (monitors_used * 100LL) / ceiling;
1304 is_above_threshold = int(monitor_usage) > MonitorUsedDeflationThreshold;
1305 }
1306 log_info(monitorinflation)("monitors_used=%zu, ceiling=%zu"
1307 ", monitor_usage=%zu, threshold=%d",
1308 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold);
1309 return is_above_threshold;
1310 }
1311
1312 return false;
1313 }
1314
1315 size_t ObjectSynchronizer::in_use_list_ceiling() {
1316 return _in_use_list_ceiling;
1317 }
1318
1319 void ObjectSynchronizer::dec_in_use_list_ceiling() {
1320 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1321 }
1322
1323 void ObjectSynchronizer::inc_in_use_list_ceiling() {
1324 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1325 }
1326
1327 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) {
1328 _in_use_list_ceiling = new_value;
1329 }
1330
1331 bool ObjectSynchronizer::is_async_deflation_needed() {
1332 if (is_async_deflation_requested()) {
1333 // Async deflation request.
1334 log_info(monitorinflation)("Async deflation needed: explicit request");
1335 return true;
1336 }
1337
1338 jlong time_since_last = time_since_last_async_deflation_ms();
1339
1340 if (AsyncDeflationInterval > 0 &&
1341 time_since_last > AsyncDeflationInterval &&
1342 monitors_used_above_threshold(&_in_use_list)) {
1343 // It's been longer than our specified deflate interval and there
1344 // are too many monitors in use. We don't deflate more frequently
1345 // than AsyncDeflationInterval (unless is_async_deflation_requested)
1346 // in order to not swamp the MonitorDeflationThread.
1347 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold");
1348 return true;
1349 }
1350
1351 if (GuaranteedAsyncDeflationInterval > 0 &&
1352 time_since_last > GuaranteedAsyncDeflationInterval) {
1353 // It's been longer than our specified guaranteed deflate interval.
1354 // We need to clean up the used monitors even if the threshold is
1355 // not reached, to keep the memory utilization at bay when many threads
1356 // touched many monitors.
1357 log_info(monitorinflation)("Async deflation needed: guaranteed interval (%zd ms) "
1358 "is greater than time since last deflation (" JLONG_FORMAT " ms)",
1359 GuaranteedAsyncDeflationInterval, time_since_last);
1360
1361 // If this deflation has no progress, then it should not affect the no-progress
1362 // tracking, otherwise threshold heuristics would think it was triggered, experienced
1363 // no progress, and needs to backoff more aggressively. In this "no progress" case,
1364 // the generic code would bump the no-progress counter, and we compensate for that
1365 // by telling it to skip the update.
1366 //
1367 // If this deflation has progress, then it should let non-progress tracking
1368 // know about this, otherwise the threshold heuristics would kick in, potentially
1369 // experience no-progress due to aggressive cleanup by this deflation, and think
1370 // it is still in no-progress stride. In this "progress" case, the generic code would
1371 // zero the counter, and we allow it to happen.
1372 _no_progress_skip_increment = true;
1373
1374 return true;
1375 }
1376
1377 return false;
1378 }
1379
1380 void ObjectSynchronizer::request_deflate_idle_monitors() {
1381 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag);
1382 set_is_async_deflation_requested(true);
1383 ml.notify_all();
1384 }
1385
1386 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() {
1387 JavaThread* current = JavaThread::current();
1388 bool ret_code = false;
1389
1390 jlong last_time = last_async_deflation_time_ns();
1391
1392 request_deflate_idle_monitors();
1393
1394 const int N_CHECKS = 5;
1395 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds
1396 if (last_async_deflation_time_ns() > last_time) {
1397 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i);
1398 ret_code = true;
1399 break;
1400 }
1401 {
1402 // JavaThread has to honor the blocking protocol.
1403 ThreadBlockInVM tbivm(current);
1404 os::naked_short_sleep(999); // sleep for almost 1 second
1405 }
1406 }
1407 if (!ret_code) {
1408 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS);
1409 }
1410
1411 return ret_code;
1412 }
1413
1414 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() {
1415 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS);
1416 }
1417
1418 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1419 const oop obj,
1420 ObjectSynchronizer::InflateCause cause) {
1421 assert(event != nullptr, "invariant");
1422 event->set_monitorClass(obj->klass());
1423 event->set_address((uintptr_t)(void*)obj);
1424 event->set_cause((u1)cause);
1425 event->commit();
1426 }
1427
1428 // Fast path code shared by multiple functions
1429 void ObjectSynchronizer::inflate_helper(oop obj) {
1430 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter");
1431 markWord mark = obj->mark_acquire();
1432 if (mark.has_monitor()) {
1433 ObjectMonitor* monitor = read_monitor(mark);
1434 markWord dmw = monitor->header();
1435 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value());
1436 return;
1437 }
1438 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal);
1439 }
1440
1441 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) {
1442 assert(current == Thread::current(), "must be");
1443 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter");
1444 return inflate_impl(current->is_Java_thread() ? JavaThread::cast(current) : nullptr, obj, cause);
1445 }
1446
1447 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) {
1448 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be");
1449 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_for");
1450 return inflate_impl(thread, obj, cause);
1451 }
1452
1453 ObjectMonitor* ObjectSynchronizer::inflate_impl(JavaThread* locking_thread, oop object, const InflateCause cause) {
1454 // The JavaThread* locking_thread requires that the locking_thread == Thread::current() or
1455 // is suspended throughout the call by some other mechanism.
1456 // The thread might be nullptr when called from a non JavaThread. (As may still be
1457 // the case from FastHashCode). However it is only important for correctness that the
1458 // thread is set when called from ObjectSynchronizer::enter from the owning thread,
1459 // ObjectSynchronizer::enter_for from any thread, or ObjectSynchronizer::exit.
1460 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_impl");
1461 EventJavaMonitorInflate event;
1462
1463 for (;;) {
1464 const markWord mark = object->mark_acquire();
1465
1466 // The mark can be in one of the following states:
1467 // * inflated - If the ObjectMonitor owner is anonymous and the
1468 // locking_thread owns the object lock, then we
1469 // make the locking_thread the ObjectMonitor owner.
1470 // * stack-locked - Coerce it to inflated from stack-locked.
1471 // * INFLATING - Busy wait for conversion from stack-locked to
1472 // inflated.
1473 // * unlocked - Aggressively inflate the object.
1474
1475 // CASE: inflated
1476 if (mark.has_monitor()) {
1477 ObjectMonitor* inf = mark.monitor();
1478 markWord dmw = inf->header();
1479 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1480 if (inf->has_anonymous_owner() && locking_thread != nullptr) {
1481 assert(LockingMode == LM_LEGACY, "invariant");
1482 if (locking_thread->is_lock_owned((address)inf->stack_locker())) {
1483 inf->set_stack_locker(nullptr);
1484 inf->set_owner_from_anonymous(locking_thread);
1485 }
1486 }
1487 return inf;
1488 }
1489
1490 // CASE: inflation in progress - inflating over a stack-lock.
1491 // Some other thread is converting from stack-locked to inflated.
1492 // Only that thread can complete inflation -- other threads must wait.
1493 // The INFLATING value is transient.
1494 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1495 // We could always eliminate polling by parking the thread on some auxiliary list.
1496 if (mark == markWord::INFLATING()) {
1497 read_stable_mark(object);
1498 continue;
1499 }
1500
1501 // CASE: stack-locked
1502 // Could be stack-locked either by current or by some other thread.
1503 //
1504 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting
1505 // to install INFLATING into the mark word. We originally installed INFLATING,
1506 // allocated the ObjectMonitor, and then finally STed the address of the
1507 // ObjectMonitor into the mark. This was correct, but artificially lengthened
1508 // the interval in which INFLATING appeared in the mark, thus increasing
1509 // the odds of inflation contention. If we lose the race to set INFLATING,
1510 // then we just delete the ObjectMonitor and loop around again.
1511 //
1512 LogStreamHandle(Trace, monitorinflation) lsh;
1513 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1514 ObjectMonitor* m = new ObjectMonitor(object);
1515 // Optimistically prepare the ObjectMonitor - anticipate successful CAS
1516 // We do this before the CAS in order to minimize the length of time
1517 // in which INFLATING appears in the mark.
1518
1519 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark);
1520 if (cmp != mark) {
1521 delete m;
1522 continue; // Interference -- just retry
1523 }
1524
1525 // We've successfully installed INFLATING (0) into the mark-word.
1526 // This is the only case where 0 will appear in a mark-word.
1527 // Only the singular thread that successfully swings the mark-word
1528 // to 0 can perform (or more precisely, complete) inflation.
1529 //
1530 // Why do we CAS a 0 into the mark-word instead of just CASing the
1531 // mark-word from the stack-locked value directly to the new inflated state?
1532 // Consider what happens when a thread unlocks a stack-locked object.
1533 // It attempts to use CAS to swing the displaced header value from the
1534 // on-stack BasicLock back into the object header. Recall also that the
1535 // header value (hash code, etc) can reside in (a) the object header, or
1536 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1537 // header in an ObjectMonitor. The inflate() routine must copy the header
1538 // value from the BasicLock on the owner's stack to the ObjectMonitor, all
1539 // the while preserving the hashCode stability invariants. If the owner
1540 // decides to release the lock while the value is 0, the unlock will fail
1541 // and control will eventually pass from slow_exit() to inflate. The owner
1542 // will then spin, waiting for the 0 value to disappear. Put another way,
1543 // the 0 causes the owner to stall if the owner happens to try to
1544 // drop the lock (restoring the header from the BasicLock to the object)
1545 // while inflation is in-progress. This protocol avoids races that might
1546 // would otherwise permit hashCode values to change or "flicker" for an object.
1547 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable.
1548 // 0 serves as a "BUSY" inflate-in-progress indicator.
1549
1550
1551 // fetch the displaced mark from the owner's stack.
1552 // The owner can't die or unwind past the lock while our INFLATING
1553 // object is in the mark. Furthermore the owner can't complete
1554 // an unlock on the object, either.
1555 markWord dmw = mark.displaced_mark_helper();
1556 // Catch if the object's header is not neutral (not locked and
1557 // not marked is what we care about here).
1558 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1559
1560 // Setup monitor fields to proper values -- prepare the monitor
1561 m->set_header(dmw);
1562
1563 // Note that a thread can inflate an object
1564 // that it has stack-locked -- as might happen in wait() -- directly
1565 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom.
1566 if (locking_thread != nullptr && locking_thread->is_lock_owned((address)mark.locker())) {
1567 m->set_owner(locking_thread);
1568 } else {
1569 // Use ANONYMOUS_OWNER to indicate that the owner is the BasicLock on the stack,
1570 // and set the stack locker field in the monitor.
1571 m->set_stack_locker(mark.locker());
1572 m->set_anonymous_owner();
1573 }
1574 // TODO-FIXME: assert BasicLock->dhw != 0.
1575
1576 // Must preserve store ordering. The monitor state must
1577 // be stable at the time of publishing the monitor address.
1578 guarantee(object->mark() == markWord::INFLATING(), "invariant");
1579 // Release semantics so that above set_object() is seen first.
1580 object->release_set_mark(markWord::encode(m));
1581
1582 // Once ObjectMonitor is configured and the object is associated
1583 // with the ObjectMonitor, it is safe to allow async deflation:
1584 _in_use_list.add(m);
1585
1586 // Hopefully the performance counters are allocated on distinct cache lines
1587 // to avoid false sharing on MP systems ...
1588 OM_PERFDATA_OP(Inflations, inc());
1589 if (log_is_enabled(Trace, monitorinflation)) {
1590 ResourceMark rm;
1591 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1592 INTPTR_FORMAT ", type='%s'", p2i(object),
1593 object->mark().value(), object->klass()->external_name());
1594 }
1595 if (event.should_commit()) {
1596 post_monitor_inflate_event(&event, object, cause);
1597 }
1598 return m;
1599 }
1600
1601 // CASE: unlocked
1602 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1603 // If we know we're inflating for entry it's better to inflate by swinging a
1604 // pre-locked ObjectMonitor pointer into the object header. A successful
1605 // CAS inflates the object *and* confers ownership to the inflating thread.
1606 // In the current implementation we use a 2-step mechanism where we CAS()
1607 // to inflate and then CAS() again to try to swing _owner from null to current.
1608 // An inflateTry() method that we could call from enter() would be useful.
1609
1610 assert(mark.is_unlocked(), "invariant: header=" INTPTR_FORMAT, mark.value());
1611 ObjectMonitor* m = new ObjectMonitor(object);
1612 // prepare m for installation - set monitor to initial state
1613 m->set_header(mark);
1614
1615 if (object->cas_set_mark(markWord::encode(m), mark) != mark) {
1616 delete m;
1617 m = nullptr;
1618 continue;
1619 // interference - the markword changed - just retry.
1620 // The state-transitions are one-way, so there's no chance of
1621 // live-lock -- "Inflated" is an absorbing state.
1622 }
1623
1624 // Once the ObjectMonitor is configured and object is associated
1625 // with the ObjectMonitor, it is safe to allow async deflation:
1626 _in_use_list.add(m);
1627
1628 // Hopefully the performance counters are allocated on distinct
1629 // cache lines to avoid false sharing on MP systems ...
1630 OM_PERFDATA_OP(Inflations, inc());
1631 if (log_is_enabled(Trace, monitorinflation)) {
1632 ResourceMark rm;
1633 lsh.print_cr("inflate(unlocked): object=" INTPTR_FORMAT ", mark="
1634 INTPTR_FORMAT ", type='%s'", p2i(object),
1635 object->mark().value(), object->klass()->external_name());
1636 }
1637 if (event.should_commit()) {
1638 post_monitor_inflate_event(&event, object, cause);
1639 }
1640 return m;
1641 }
1642 }
1643
1644 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle
1645 // ObjectMonitors. Returns the number of deflated ObjectMonitors.
1646 //
1647 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) {
1648 MonitorList::Iterator iter = _in_use_list.iterator();
1649 size_t deflated_count = 0;
1650 Thread* current = Thread::current();
1651
1652 while (iter.has_next()) {
1653 if (deflated_count >= (size_t)MonitorDeflationMax) {
1654 break;
1655 }
1656 ObjectMonitor* mid = iter.next();
1657 if (mid->deflate_monitor(current)) {
1658 deflated_count++;
1659 }
1660
1661 // Must check for a safepoint/handshake and honor it.
1662 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count);
1663 }
1664
1665 return deflated_count;
1666 }
1667
1668 class HandshakeForDeflation : public HandshakeClosure {
1669 public:
1670 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {}
1671
1672 void do_thread(Thread* thread) {
1673 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread="
1674 INTPTR_FORMAT, p2i(thread));
1675 if (thread->is_Java_thread()) {
1676 // Clear OM cache
1677 JavaThread* jt = JavaThread::cast(thread);
1678 jt->om_clear_monitor_cache();
1679 }
1680 }
1681 };
1682
1683 class VM_RendezvousGCThreads : public VM_Operation {
1684 public:
1685 bool evaluate_at_safepoint() const override { return false; }
1686 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; }
1687 void doit() override {
1688 Universe::heap()->safepoint_synchronize_begin();
1689 Universe::heap()->safepoint_synchronize_end();
1690 };
1691 };
1692
1693 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list,
1694 ObjectMonitorDeflationSafepointer* safepointer) {
1695 NativeHeapTrimmer::SuspendMark sm("monitor deletion");
1696 size_t deleted_count = 0;
1697 for (ObjectMonitor* monitor: *delete_list) {
1698 delete monitor;
1699 deleted_count++;
1700 // A JavaThread must check for a safepoint/handshake and honor it.
1701 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count);
1702 }
1703 return deleted_count;
1704 }
1705
1706 class ObjectMonitorDeflationLogging: public StackObj {
1707 LogStreamHandle(Debug, monitorinflation) _debug;
1708 LogStreamHandle(Info, monitorinflation) _info;
1709 LogStream* _stream;
1710 elapsedTimer _timer;
1711
1712 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); }
1713 size_t count() const { return ObjectSynchronizer::_in_use_list.count(); }
1714 size_t max() const { return ObjectSynchronizer::_in_use_list.max(); }
1715
1716 public:
1717 ObjectMonitorDeflationLogging()
1718 : _debug(), _info(), _stream(nullptr) {
1719 if (_debug.is_enabled()) {
1720 _stream = &_debug;
1721 } else if (_info.is_enabled()) {
1722 _stream = &_info;
1723 }
1724 }
1725
1726 void begin() {
1727 if (_stream != nullptr) {
1728 _stream->print_cr("begin deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu",
1729 ceiling(), count(), max());
1730 _timer.start();
1731 }
1732 }
1733
1734 void before_handshake(size_t unlinked_count) {
1735 if (_stream != nullptr) {
1736 _timer.stop();
1737 _stream->print_cr("before handshaking: unlinked_count=%zu"
1738 ", in_use_list stats: ceiling=%zu, count="
1739 "%zu, max=%zu",
1740 unlinked_count, ceiling(), count(), max());
1741 }
1742 }
1743
1744 void after_handshake() {
1745 if (_stream != nullptr) {
1746 _stream->print_cr("after handshaking: in_use_list stats: ceiling="
1747 "%zu, count=%zu, max=%zu",
1748 ceiling(), count(), max());
1749 _timer.start();
1750 }
1751 }
1752
1753 void end(size_t deflated_count, size_t unlinked_count) {
1754 if (_stream != nullptr) {
1755 _timer.stop();
1756 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) {
1757 _stream->print_cr("deflated_count=%zu, {unlinked,deleted}_count=%zu monitors in %3.7f secs",
1758 deflated_count, unlinked_count, _timer.seconds());
1759 }
1760 _stream->print_cr("end deflating: in_use_list stats: ceiling=%zu, count=%zu, max=%zu",
1761 ceiling(), count(), max());
1762 }
1763 }
1764
1765 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) {
1766 if (_stream != nullptr) {
1767 _timer.stop();
1768 _stream->print_cr("pausing %s: %s=%zu, in_use_list stats: ceiling="
1769 "%zu, count=%zu, max=%zu",
1770 op_name, cnt_name, cnt, ceiling(), count(), max());
1771 }
1772 }
1773
1774 void after_block_for_safepoint(const char* op_name) {
1775 if (_stream != nullptr) {
1776 _stream->print_cr("resuming %s: in_use_list stats: ceiling=%zu"
1777 ", count=%zu, max=%zu", op_name,
1778 ceiling(), count(), max());
1779 _timer.start();
1780 }
1781 }
1782 };
1783
1784 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) {
1785 if (!SafepointMechanism::should_process(_current)) {
1786 return;
1787 }
1788
1789 // A safepoint/handshake has started.
1790 _log->before_block_for_safepoint(op_name, count_name, counter);
1791
1792 {
1793 // Honor block request.
1794 ThreadBlockInVM tbivm(_current);
1795 }
1796
1797 _log->after_block_for_safepoint(op_name);
1798 }
1799
1800 // This function is called by the MonitorDeflationThread to deflate
1801 // ObjectMonitors.
1802 size_t ObjectSynchronizer::deflate_idle_monitors() {
1803 JavaThread* current = JavaThread::current();
1804 assert(current->is_monitor_deflation_thread(), "The only monitor deflater");
1805
1806 // The async deflation request has been processed.
1807 _last_async_deflation_time_ns = os::javaTimeNanos();
1808 set_is_async_deflation_requested(false);
1809
1810 ObjectMonitorDeflationLogging log;
1811 ObjectMonitorDeflationSafepointer safepointer(current, &log);
1812
1813 log.begin();
1814
1815 // Deflate some idle ObjectMonitors.
1816 size_t deflated_count = deflate_monitor_list(&safepointer);
1817
1818 // Unlink the deflated ObjectMonitors from the in-use list.
1819 size_t unlinked_count = 0;
1820 size_t deleted_count = 0;
1821 if (deflated_count > 0) {
1822 ResourceMark rm(current);
1823 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count);
1824 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer);
1825
1826 #ifdef ASSERT
1827 if (UseObjectMonitorTable) {
1828 for (ObjectMonitor* monitor : delete_list) {
1829 assert(!LightweightSynchronizer::contains_monitor(current, monitor), "Should have been removed");
1830 }
1831 }
1832 #endif
1833
1834 log.before_handshake(unlinked_count);
1835
1836 // A JavaThread needs to handshake in order to safely free the
1837 // ObjectMonitors that were deflated in this cycle.
1838 HandshakeForDeflation hfd_hc;
1839 Handshake::execute(&hfd_hc);
1840 // Also, we sync and desync GC threads around the handshake, so that they can
1841 // safely read the mark-word and look-through to the object-monitor, without
1842 // being afraid that the object-monitor is going away.
1843 VM_RendezvousGCThreads sync_gc;
1844 VMThread::execute(&sync_gc);
1845
1846 log.after_handshake();
1847
1848 // After the handshake, safely free the ObjectMonitors that were
1849 // deflated and unlinked in this cycle.
1850
1851 // Delete the unlinked ObjectMonitors.
1852 deleted_count = delete_monitors(&delete_list, &safepointer);
1853 assert(unlinked_count == deleted_count, "must be");
1854 }
1855
1856 log.end(deflated_count, unlinked_count);
1857
1858 OM_PERFDATA_OP(MonExtant, set_value(_in_use_list.count()));
1859 OM_PERFDATA_OP(Deflations, inc(deflated_count));
1860
1861 GVars.stw_random = os::random();
1862
1863 if (deflated_count != 0) {
1864 _no_progress_cnt = 0;
1865 } else if (_no_progress_skip_increment) {
1866 _no_progress_skip_increment = false;
1867 } else {
1868 _no_progress_cnt++;
1869 }
1870
1871 return deflated_count;
1872 }
1873
1874 // Monitor cleanup on JavaThread::exit
1875
1876 // Iterate through monitor cache and attempt to release thread's monitors
1877 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1878 private:
1879 JavaThread* _thread;
1880
1881 public:
1882 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {}
1883 void do_monitor(ObjectMonitor* mid) {
1884 intx rec = mid->complete_exit(_thread);
1885 _thread->dec_held_monitor_count(rec + 1);
1886 }
1887 };
1888
1889 // Release all inflated monitors owned by current thread. Lightweight monitors are
1890 // ignored. This is meant to be called during JNI thread detach which assumes
1891 // all remaining monitors are heavyweight. All exceptions are swallowed.
1892 // Scanning the extant monitor list can be time consuming.
1893 // A simple optimization is to add a per-thread flag that indicates a thread
1894 // called jni_monitorenter() during its lifetime.
1895 //
1896 // Instead of NoSafepointVerifier it might be cheaper to
1897 // use an idiom of the form:
1898 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1899 // <code that must not run at safepoint>
1900 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1901 // Since the tests are extremely cheap we could leave them enabled
1902 // for normal product builds.
1903
1904 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) {
1905 assert(current == JavaThread::current(), "must be current Java thread");
1906 NoSafepointVerifier nsv;
1907 ReleaseJavaMonitorsClosure rjmc(current);
1908 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current);
1909 assert(!current->has_pending_exception(), "Should not be possible");
1910 current->clear_pending_exception();
1911 assert(current->held_monitor_count() == 0, "Should not be possible");
1912 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count.
1913 current->clear_jni_monitor_count();
1914 }
1915
1916 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1917 switch (cause) {
1918 case inflate_cause_vm_internal: return "VM Internal";
1919 case inflate_cause_monitor_enter: return "Monitor Enter";
1920 case inflate_cause_wait: return "Monitor Wait";
1921 case inflate_cause_notify: return "Monitor Notify";
1922 case inflate_cause_hash_code: return "Monitor Hash Code";
1923 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1924 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1925 default:
1926 ShouldNotReachHere();
1927 }
1928 return "Unknown";
1929 }
1930
1931 //------------------------------------------------------------------------------
1932 // Debugging code
1933
1934 u_char* ObjectSynchronizer::get_gvars_addr() {
1935 return (u_char*)&GVars;
1936 }
1937
1938 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() {
1939 return (u_char*)&GVars.hc_sequence;
1940 }
1941
1942 size_t ObjectSynchronizer::get_gvars_size() {
1943 return sizeof(SharedGlobals);
1944 }
1945
1946 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() {
1947 return (u_char*)&GVars.stw_random;
1948 }
1949
1950 // Do the final audit and print of ObjectMonitor stats; must be done
1951 // by the VMThread at VM exit time.
1952 void ObjectSynchronizer::do_final_audit_and_print_stats() {
1953 assert(Thread::current()->is_VM_thread(), "sanity check");
1954
1955 if (is_final_audit()) { // Only do the audit once.
1956 return;
1957 }
1958 set_is_final_audit();
1959 log_info(monitorinflation)("Starting the final audit.");
1960
1961 if (log_is_enabled(Info, monitorinflation)) {
1962 LogStreamHandle(Info, monitorinflation) ls;
1963 audit_and_print_stats(&ls, true /* on_exit */);
1964 }
1965 }
1966
1967 // This function can be called by the MonitorDeflationThread or it can be called when
1968 // we are trying to exit the VM. The list walker functions can run in parallel with
1969 // the other list operations.
1970 // Calls to this function can be added in various places as a debugging
1971 // aid.
1972 //
1973 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) {
1974 int error_cnt = 0;
1975
1976 ls->print_cr("Checking in_use_list:");
1977 chk_in_use_list(ls, &error_cnt);
1978
1979 if (error_cnt == 0) {
1980 ls->print_cr("No errors found in in_use_list checks.");
1981 } else {
1982 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt);
1983 }
1984
1985 // When exiting, only log the interesting entries at the Info level.
1986 // When called at intervals by the MonitorDeflationThread, log output
1987 // at the Trace level since there can be a lot of it.
1988 if (!on_exit && log_is_enabled(Trace, monitorinflation)) {
1989 LogStreamHandle(Trace, monitorinflation) ls_tr;
1990 log_in_use_monitor_details(&ls_tr, true /* log_all */);
1991 } else if (on_exit) {
1992 log_in_use_monitor_details(ls, false /* log_all */);
1993 }
1994
1995 ls->flush();
1996
1997 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt);
1998 }
1999
2000 // Check the in_use_list; log the results of the checks.
2001 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) {
2002 size_t l_in_use_count = _in_use_list.count();
2003 size_t l_in_use_max = _in_use_list.max();
2004 out->print_cr("count=%zu, max=%zu", l_in_use_count,
2005 l_in_use_max);
2006
2007 size_t ck_in_use_count = 0;
2008 MonitorList::Iterator iter = _in_use_list.iterator();
2009 while (iter.has_next()) {
2010 ObjectMonitor* mid = iter.next();
2011 chk_in_use_entry(mid, out, error_cnt_p);
2012 ck_in_use_count++;
2013 }
2014
2015 if (l_in_use_count == ck_in_use_count) {
2016 out->print_cr("in_use_count=%zu equals ck_in_use_count=%zu",
2017 l_in_use_count, ck_in_use_count);
2018 } else {
2019 out->print_cr("WARNING: in_use_count=%zu is not equal to "
2020 "ck_in_use_count=%zu", l_in_use_count,
2021 ck_in_use_count);
2022 }
2023
2024 size_t ck_in_use_max = _in_use_list.max();
2025 if (l_in_use_max == ck_in_use_max) {
2026 out->print_cr("in_use_max=%zu equals ck_in_use_max=%zu",
2027 l_in_use_max, ck_in_use_max);
2028 } else {
2029 out->print_cr("WARNING: in_use_max=%zu is not equal to "
2030 "ck_in_use_max=%zu", l_in_use_max, ck_in_use_max);
2031 }
2032 }
2033
2034 // Check an in-use monitor entry; log any errors.
2035 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out,
2036 int* error_cnt_p) {
2037 if (n->owner_is_DEFLATER_MARKER()) {
2038 // This could happen when monitor deflation blocks for a safepoint.
2039 return;
2040 }
2041
2042
2043 if (n->metadata() == 0) {
2044 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must "
2045 "have non-null _metadata (header/hash) field.", p2i(n));
2046 *error_cnt_p = *error_cnt_p + 1;
2047 }
2048
2049 const oop obj = n->object_peek();
2050 if (obj == nullptr) {
2051 return;
2052 }
2053
2054 const markWord mark = obj->mark();
2055 if (!mark.has_monitor()) {
2056 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2057 "object does not think it has a monitor: obj="
2058 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n),
2059 p2i(obj), mark.value());
2060 *error_cnt_p = *error_cnt_p + 1;
2061 return;
2062 }
2063
2064 ObjectMonitor* const obj_mon = read_monitor(Thread::current(), obj, mark);
2065 if (n != obj_mon) {
2066 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2067 "object does not refer to the same monitor: obj="
2068 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon="
2069 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon));
2070 *error_cnt_p = *error_cnt_p + 1;
2071 }
2072 }
2073
2074 // Log details about ObjectMonitors on the in_use_list. The 'BHL'
2075 // flags indicate why the entry is in-use, 'object' and 'object type'
2076 // indicate the associated object and its type.
2077 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) {
2078 if (_in_use_list.count() > 0) {
2079 stringStream ss;
2080 out->print_cr("In-use monitor info%s:", log_all ? "" : " (eliding idle monitors)");
2081 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2082 out->print_cr("%18s %s %18s %18s",
2083 "monitor", "BHL", "object", "object type");
2084 out->print_cr("================== === ================== ==================");
2085
2086 auto is_interesting = [&](ObjectMonitor* monitor) {
2087 return log_all || monitor->has_owner() || monitor->is_busy();
2088 };
2089
2090 monitors_iterate([&](ObjectMonitor* monitor) {
2091 if (is_interesting(monitor)) {
2092 const oop obj = monitor->object_peek();
2093 const intptr_t hash = UseObjectMonitorTable ? monitor->hash() : monitor->header().hash();
2094 ResourceMark rm;
2095 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor),
2096 monitor->is_busy(), hash != 0, monitor->has_owner(),
2097 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name());
2098 if (monitor->is_busy()) {
2099 out->print(" (%s)", monitor->is_busy_to_string(&ss));
2100 ss.reset();
2101 }
2102 out->cr();
2103 }
2104 });
2105 }
2106
2107 out->flush();
2108 }