1 /*
2 * Copyright (c) 1998, 2024, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "classfile/vmSymbols.hpp"
27 #include "gc/shared/collectedHeap.hpp"
28 #include "jfr/jfrEvents.hpp"
29 #include "logging/log.hpp"
30 #include "logging/logStream.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/padded.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "memory/universe.hpp"
35 #include "oops/markWord.hpp"
36 #include "oops/oop.inline.hpp"
37 #include "runtime/atomic.hpp"
38 #include "runtime/basicLock.inline.hpp"
39 #include "runtime/frame.inline.hpp"
40 #include "runtime/globals.hpp"
41 #include "runtime/handles.inline.hpp"
42 #include "runtime/handshake.hpp"
43 #include "runtime/interfaceSupport.inline.hpp"
44 #include "runtime/javaThread.hpp"
45 #include "runtime/lightweightSynchronizer.hpp"
46 #include "runtime/lockStack.inline.hpp"
47 #include "runtime/mutexLocker.hpp"
48 #include "runtime/objectMonitor.hpp"
49 #include "runtime/objectMonitor.inline.hpp"
50 #include "runtime/os.inline.hpp"
51 #include "runtime/osThread.hpp"
52 #include "runtime/perfData.hpp"
53 #include "runtime/safepointMechanism.inline.hpp"
54 #include "runtime/safepointVerifiers.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #include "runtime/stubRoutines.hpp"
57 #include "runtime/synchronizer.inline.hpp"
58 #include "runtime/threads.hpp"
59 #include "runtime/timer.hpp"
60 #include "runtime/trimNativeHeap.hpp"
61 #include "runtime/vframe.hpp"
62 #include "runtime/vmThread.hpp"
63 #include "utilities/align.hpp"
64 #include "utilities/dtrace.hpp"
65 #include "utilities/events.hpp"
66 #include "utilities/globalDefinitions.hpp"
67 #include "utilities/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->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, JavaThread* current,
404 BasicLock * lock) {
405 assert(current->thread_state() == _thread_in_Java, "invariant");
406
407 if (useHeavyMonitors()) {
408 return false; // Slow path
409 }
410
411 if (LockingMode == LM_LIGHTWEIGHT) {
412 return LightweightSynchronizer::quick_enter(obj, current, lock);
413 }
414
415 assert(LockingMode == LM_LEGACY, "legacy mode below");
416
417 const markWord mark = obj->mark();
418
419 if (mark.has_monitor()) {
420
421 ObjectMonitor* const m = read_monitor(mark);
422 // An async deflation or GC can race us before we manage to make
423 // the ObjectMonitor busy by setting the owner below. If we detect
424 // that race we just bail out to the slow-path here.
425 if (m->object_peek() == nullptr) {
426 return false;
427 }
428 JavaThread* const owner = static_cast<JavaThread*>(m->owner_raw());
429
430 // Lock contention and Transactional Lock Elision (TLE) diagnostics
431 // and observability
432 // Case: light contention possibly amenable to TLE
433 // Case: TLE inimical operations such as nested/recursive synchronization
434
435 if (owner == current) {
436 m->_recursions++;
437 current->inc_held_monitor_count();
438 return true;
439 }
440
441 // This Java Monitor is inflated so obj's header will never be
442 // displaced to this thread's BasicLock. Make the displaced header
443 // non-null so this BasicLock is not seen as recursive nor as
444 // being locked. We do this unconditionally so that this thread's
445 // BasicLock cannot be mis-interpreted by any stack walkers. For
446 // performance reasons, stack walkers generally first check for
447 // stack-locking in the object's header, the second check is for
448 // recursive stack-locking in the displaced header in the BasicLock,
449 // and last are the inflated Java Monitor (ObjectMonitor) checks.
450 lock->set_displaced_header(markWord::unused_mark());
451
452 if (owner == nullptr && m->try_set_owner_from(nullptr, current) == nullptr) {
453 assert(m->_recursions == 0, "invariant");
454 current->inc_held_monitor_count();
455 return true;
456 }
457 }
458
459 // Note that we could inflate in quick_enter.
460 // This is likely a useful optimization
461 // Critically, in quick_enter() we must not:
462 // -- block indefinitely, or
463 // -- reach a safepoint
464
465 return false; // revert to slow-path
466 }
467
468 // Handle notifications when synchronizing on value based classes
469 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) {
470 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
471 frame last_frame = locking_thread->last_frame();
472 bool bcp_was_adjusted = false;
473 // Don't decrement bcp if it points to the frame's first instruction. This happens when
474 // handle_sync_on_value_based_class() is called because of a synchronized method. There
475 // is no actual monitorenter instruction in the byte code in this case.
476 if (last_frame.is_interpreted_frame() &&
477 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) {
478 // adjust bcp to point back to monitorenter so that we print the correct line numbers
479 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1);
480 bcp_was_adjusted = true;
481 }
482
483 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) {
484 ResourceMark rm;
485 stringStream ss;
486 locking_thread->print_active_stack_on(&ss);
487 char* base = (char*)strstr(ss.base(), "at");
488 char* newline = (char*)strchr(ss.base(), '\n');
489 if (newline != nullptr) {
490 *newline = '\0';
491 }
492 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base);
493 } else {
494 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses");
495 ResourceMark rm;
496 Log(valuebasedclasses) vblog;
497
498 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name());
499 if (locking_thread->has_last_Java_frame()) {
500 LogStream info_stream(vblog.info());
501 locking_thread->print_active_stack_on(&info_stream);
502 } else {
503 vblog.info("Cannot find the last Java frame");
504 }
505
506 EventSyncOnValueBasedClass event;
507 if (event.should_commit()) {
508 event.set_valueBasedClass(obj->klass());
509 event.commit();
510 }
511 }
512
513 if (bcp_was_adjusted) {
514 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1);
515 }
516 }
517
518 // -----------------------------------------------------------------------------
519 // Monitor Enter/Exit
520
521 void ObjectSynchronizer::enter_for(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
522 // When called with locking_thread != Thread::current() some mechanism must synchronize
523 // the locking_thread with respect to the current thread. Currently only used when
524 // deoptimizing and re-locking locks. See Deoptimization::relock_objects
525 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
526
527 if (LockingMode == LM_LIGHTWEIGHT) {
528 return LightweightSynchronizer::enter_for(obj, lock, locking_thread);
529 }
530
531 if (!enter_fast_impl(obj, lock, locking_thread)) {
532 // Inflated ObjectMonitor::enter_for is required
533
534 // An async deflation can race after the inflate_for() call and before
535 // enter_for() can make the ObjectMonitor busy. enter_for() returns false
536 // if we have lost the race to async deflation and we simply try again.
537 while (true) {
538 ObjectMonitor* monitor = inflate_for(locking_thread, obj(), inflate_cause_monitor_enter);
539 if (monitor->enter_for(locking_thread)) {
540 return;
541 }
542 assert(monitor->is_being_async_deflated(), "must be");
543 }
544 }
545 }
546
547 void ObjectSynchronizer::enter_legacy(Handle obj, BasicLock* lock, JavaThread* current) {
548 if (!enter_fast_impl(obj, lock, current)) {
549 // Inflated ObjectMonitor::enter is required
550
551 // An async deflation can race after the inflate() call and before
552 // enter() can make the ObjectMonitor busy. enter() returns false if
553 // we have lost the race to async deflation and we simply try again.
554 while (true) {
555 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter);
556 if (monitor->enter(current)) {
557 return;
558 }
559 }
560 }
561 }
562
563 // The interpreter and compiler assembly code tries to lock using the fast path
564 // of this algorithm. Make sure to update that code if the following function is
565 // changed. The implementation is extremely sensitive to race condition. Be careful.
566 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
567 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer");
568
569 if (obj->klass()->is_value_based()) {
570 handle_sync_on_value_based_class(obj, locking_thread);
571 }
572
573 locking_thread->inc_held_monitor_count();
574
575 if (!useHeavyMonitors()) {
576 if (LockingMode == LM_LEGACY) {
577 markWord mark = obj->mark();
578 if (mark.is_unlocked()) {
579 // Anticipate successful CAS -- the ST of the displaced mark must
580 // be visible <= the ST performed by the CAS.
581 lock->set_displaced_header(mark);
582 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) {
583 return true;
584 }
585 } else if (mark.has_locker() &&
586 locking_thread->is_lock_owned((address) mark.locker())) {
587 assert(lock != mark.locker(), "must not re-lock the same lock");
588 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock");
589 lock->set_displaced_header(markWord::from_pointer(nullptr));
590 return true;
591 }
592
593 // The object header will never be displaced to this lock,
594 // so it does not matter what the value is, except that it
595 // must be non-zero to avoid looking like a re-entrant lock,
596 // and must not look locked either.
597 lock->set_displaced_header(markWord::unused_mark());
598
599 // Failed to fast lock.
600 return false;
601 }
602 } else if (VerifyHeavyMonitors) {
603 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
604 }
605
606 return false;
607 }
608
609 void ObjectSynchronizer::exit_legacy(oop object, BasicLock* lock, JavaThread* current) {
610 assert(LockingMode != LM_LIGHTWEIGHT, "Use LightweightSynchronizer");
611
612 if (!useHeavyMonitors()) {
613 markWord mark = object->mark();
614 if (LockingMode == LM_LEGACY) {
615 markWord dhw = lock->displaced_header();
616 if (dhw.value() == 0) {
617 // If the displaced header is null, then this exit matches up with
618 // a recursive enter. No real work to do here except for diagnostics.
619 #ifndef PRODUCT
620 if (mark != markWord::INFLATING()) {
621 // Only do diagnostics if we are not racing an inflation. Simply
622 // exiting a recursive enter of a Java Monitor that is being
623 // inflated is safe; see the has_monitor() comment below.
624 assert(!mark.is_unlocked(), "invariant");
625 assert(!mark.has_locker() ||
626 current->is_lock_owned((address)mark.locker()), "invariant");
627 if (mark.has_monitor()) {
628 // The BasicLock's displaced_header is marked as a recursive
629 // enter and we have an inflated Java Monitor (ObjectMonitor).
630 // This is a special case where the Java Monitor was inflated
631 // after this thread entered the stack-lock recursively. When a
632 // Java Monitor is inflated, we cannot safely walk the Java
633 // Monitor owner's stack and update the BasicLocks because a
634 // Java Monitor can be asynchronously inflated by a thread that
635 // does not own the Java Monitor.
636 ObjectMonitor* m = read_monitor(mark);
637 assert(m->object()->mark() == mark, "invariant");
638 assert(m->is_entered(current), "invariant");
639 }
640 }
641 #endif
642 return;
643 }
644
645 if (mark == markWord::from_pointer(lock)) {
646 // If the object is stack-locked by the current thread, try to
647 // swing the displaced header from the BasicLock back to the mark.
648 assert(dhw.is_neutral(), "invariant");
649 if (object->cas_set_mark(dhw, mark) == mark) {
650 return;
651 }
652 }
653 }
654 } else if (VerifyHeavyMonitors) {
655 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
656 }
657
658 // We have to take the slow-path of possible inflation and then exit.
659 // The ObjectMonitor* can't be async deflated until ownership is
660 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
661 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal);
662 assert(!monitor->is_owner_anonymous(), "must not be");
663 monitor->exit(current);
664 }
665
666 // -----------------------------------------------------------------------------
667 // JNI locks on java objects
668 // NOTE: must use heavy weight monitor to handle jni monitor enter
669 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) {
670 if (obj->klass()->is_value_based()) {
671 handle_sync_on_value_based_class(obj, current);
672 }
673
674 // the current locking is from JNI instead of Java code
675 current->set_current_pending_monitor_is_from_java(false);
676 // An async deflation can race after the inflate() call and before
677 // enter() can make the ObjectMonitor busy. enter() returns false if
678 // we have lost the race to async deflation and we simply try again.
679 while (true) {
680 ObjectMonitor* monitor;
681 bool entered;
682 if (LockingMode == LM_LIGHTWEIGHT) {
683 entered = LightweightSynchronizer::inflate_and_enter(obj(), current, current, inflate_cause_jni_enter) != nullptr;
684 } else {
685 monitor = inflate(current, obj(), inflate_cause_jni_enter);
686 entered = monitor->enter(current);
687 }
688
689 if (entered) {
690 current->inc_held_monitor_count(1, true);
691 break;
692 }
693 }
694 current->set_current_pending_monitor_is_from_java(true);
695 }
696
697 // NOTE: must use heavy weight monitor to handle jni monitor exit
698 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) {
699 JavaThread* current = THREAD;
700
701 ObjectMonitor* monitor;
702 if (LockingMode == LM_LIGHTWEIGHT) {
703 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj, inflate_cause_jni_exit, CHECK);
704 } else {
705 // The ObjectMonitor* can't be async deflated until ownership is
706 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
707 monitor = inflate(current, obj, inflate_cause_jni_exit);
708 }
709 // If this thread has locked the object, exit the monitor. We
710 // intentionally do not use CHECK on check_owner because we must exit the
711 // monitor even if an exception was already pending.
712 if (monitor->check_owner(THREAD)) {
713 monitor->exit(current);
714 current->dec_held_monitor_count(1, true);
715 }
716 }
717
718 // -----------------------------------------------------------------------------
719 // Internal VM locks on java objects
720 // standard constructor, allows locking failures
721 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) {
722 _thread = thread;
723 _thread->check_for_valid_safepoint_state();
724 _obj = obj;
725
726 if (_obj() != nullptr) {
727 ObjectSynchronizer::enter(_obj, &_lock, _thread);
728 }
729 }
730
731 ObjectLocker::~ObjectLocker() {
732 if (_obj() != nullptr) {
733 ObjectSynchronizer::exit(_obj(), &_lock, _thread);
734 }
735 }
736
737
738 // -----------------------------------------------------------------------------
739 // Wait/Notify/NotifyAll
740 // NOTE: must use heavy weight monitor to handle wait()
741
742 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
743 JavaThread* current = THREAD;
744 if (millis < 0) {
745 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
746 }
747
748 ObjectMonitor* monitor;
749 if (LockingMode == LM_LIGHTWEIGHT) {
750 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_wait, CHECK_0);
751 } else {
752 // The ObjectMonitor* can't be async deflated because the _waiters
753 // field is incremented before ownership is dropped and decremented
754 // after ownership is regained.
755 monitor = inflate(current, obj(), inflate_cause_wait);
756 }
757
758 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis);
759 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code
760
761 // This dummy call is in place to get around dtrace bug 6254741. Once
762 // that's fixed we can uncomment the following line, remove the call
763 // and change this function back into a "void" func.
764 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
765 int ret_code = dtrace_waited_probe(monitor, obj, THREAD);
766 return ret_code;
767 }
768
769 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
770 JavaThread* current = THREAD;
771
772 markWord mark = obj->mark();
773 if (LockingMode == LM_LIGHTWEIGHT) {
774 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
775 // Not inflated so there can't be any waiters to notify.
776 return;
777 }
778 } else if (LockingMode == LM_LEGACY) {
779 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
780 // Not inflated so there can't be any waiters to notify.
781 return;
782 }
783 }
784
785 ObjectMonitor* monitor;
786 if (LockingMode == LM_LIGHTWEIGHT) {
787 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK);
788 } else {
789 // The ObjectMonitor* can't be async deflated until ownership is
790 // dropped by the calling thread.
791 monitor = inflate(current, obj(), inflate_cause_notify);
792 }
793 monitor->notify(CHECK);
794 }
795
796 // NOTE: see comment of notify()
797 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
798 JavaThread* current = THREAD;
799
800 markWord mark = obj->mark();
801 if (LockingMode == LM_LIGHTWEIGHT) {
802 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
803 // Not inflated so there can't be any waiters to notify.
804 return;
805 }
806 } else if (LockingMode == LM_LEGACY) {
807 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
808 // Not inflated so there can't be any waiters to notify.
809 return;
810 }
811 }
812
813 ObjectMonitor* monitor;
814 if (LockingMode == LM_LIGHTWEIGHT) {
815 monitor = LightweightSynchronizer::inflate_locked_or_imse(obj(), inflate_cause_notify, CHECK);
816 } else {
817 // The ObjectMonitor* can't be async deflated until ownership is
818 // dropped by the calling thread.
819 monitor = inflate(current, obj(), inflate_cause_notify);
820 }
821 monitor->notifyAll(CHECK);
822 }
823
824 // -----------------------------------------------------------------------------
825 // Hash Code handling
826
827 struct SharedGlobals {
828 char _pad_prefix[OM_CACHE_LINE_SIZE];
829 // This is a highly shared mostly-read variable.
830 // To avoid false-sharing it needs to be the sole occupant of a cache line.
831 volatile int stw_random;
832 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int));
833 // Hot RW variable -- Sequester to avoid false-sharing
834 volatile int hc_sequence;
835 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int));
836 };
837
838 static SharedGlobals GVars;
839
840 static markWord read_stable_mark(oop obj) {
841 markWord mark = obj->mark_acquire();
842 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) {
843 // New lightweight locking does not use the markWord::INFLATING() protocol.
844 return mark; // normal fast-path return
845 }
846
847 int its = 0;
848 for (;;) {
849 markWord mark = obj->mark_acquire();
850 if (!mark.is_being_inflated()) {
851 return mark; // normal fast-path return
852 }
853
854 // The object is being inflated by some other thread.
855 // The caller of read_stable_mark() must wait for inflation to complete.
856 // Avoid live-lock.
857
858 ++its;
859 if (its > 10000 || !os::is_MP()) {
860 if (its & 1) {
861 os::naked_yield();
862 } else {
863 // Note that the following code attenuates the livelock problem but is not
864 // a complete remedy. A more complete solution would require that the inflating
865 // thread hold the associated inflation lock. The following code simply restricts
866 // the number of spinners to at most one. We'll have N-2 threads blocked
867 // on the inflationlock, 1 thread holding the inflation lock and using
868 // a yield/park strategy, and 1 thread in the midst of inflation.
869 // A more refined approach would be to change the encoding of INFLATING
870 // to allow encapsulation of a native thread pointer. Threads waiting for
871 // inflation to complete would use CAS to push themselves onto a singly linked
872 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
873 // and calling park(). When inflation was complete the thread that accomplished inflation
874 // would detach the list and set the markword to inflated with a single CAS and
875 // then for each thread on the list, set the flag and unpark() the thread.
876
877 // Index into the lock array based on the current object address.
878 static_assert(is_power_of_2(inflation_lock_count()), "must be");
879 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1);
880 int YieldThenBlock = 0;
881 assert(ix < inflation_lock_count(), "invariant");
882 inflation_lock(ix)->lock();
883 while (obj->mark_acquire() == markWord::INFLATING()) {
884 // Beware: naked_yield() is advisory and has almost no effect on some platforms
885 // so we periodically call current->_ParkEvent->park(1).
886 // We use a mixed spin/yield/block mechanism.
887 if ((YieldThenBlock++) >= 16) {
888 Thread::current()->_ParkEvent->park(1);
889 } else {
890 os::naked_yield();
891 }
892 }
893 inflation_lock(ix)->unlock();
894 }
895 } else {
896 SpinPause(); // SMP-polite spinning
897 }
898 }
899 }
900
901 // hashCode() generation :
902 //
903 // Possibilities:
904 // * MD5Digest of {obj,stw_random}
905 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function.
906 // * A DES- or AES-style SBox[] mechanism
907 // * One of the Phi-based schemes, such as:
908 // 2654435761 = 2^32 * Phi (golden ratio)
909 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ;
910 // * A variation of Marsaglia's shift-xor RNG scheme.
911 // * (obj ^ stw_random) is appealing, but can result
912 // in undesirable regularity in the hashCode values of adjacent objects
913 // (objects allocated back-to-back, in particular). This could potentially
914 // result in hashtable collisions and reduced hashtable efficiency.
915 // There are simple ways to "diffuse" the middle address bits over the
916 // generated hashCode values:
917
918 static intptr_t get_next_hash(Thread* current, oop obj) {
919 intptr_t value = 0;
920 if (hashCode == 0) {
921 // This form uses global Park-Miller RNG.
922 // On MP system we'll have lots of RW access to a global, so the
923 // mechanism induces lots of coherency traffic.
924 value = os::random();
925 } else if (hashCode == 1) {
926 // This variation has the property of being stable (idempotent)
927 // between STW operations. This can be useful in some of the 1-0
928 // synchronization schemes.
929 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3;
930 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random;
931 } else if (hashCode == 2) {
932 value = 1; // for sensitivity testing
933 } else if (hashCode == 3) {
934 value = ++GVars.hc_sequence;
935 } else if (hashCode == 4) {
936 value = cast_from_oop<intptr_t>(obj);
937 } else {
938 // Marsaglia's xor-shift scheme with thread-specific state
939 // This is probably the best overall implementation -- we'll
940 // likely make this the default in future releases.
941 unsigned t = current->_hashStateX;
942 t ^= (t << 11);
943 current->_hashStateX = current->_hashStateY;
944 current->_hashStateY = current->_hashStateZ;
945 current->_hashStateZ = current->_hashStateW;
946 unsigned v = current->_hashStateW;
947 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
948 current->_hashStateW = v;
949 value = v;
950 }
951
952 value &= UseCompactObjectHeaders ? markWord::hash_mask_compact : markWord::hash_mask;
953 if (value == 0) value = 0xBAD;
954 assert(value != markWord::no_hash, "invariant");
955 return value;
956 }
957
958 static intptr_t install_hash_code(Thread* current, oop obj) {
959 assert(UseObjectMonitorTable && LockingMode == LM_LIGHTWEIGHT, "must be");
960
961 markWord mark = obj->mark_acquire();
962 for(;;) {
963 intptr_t hash = mark.hash();
964 if (hash != 0) {
965 return hash;
966 }
967
968 hash = get_next_hash(current, obj);
969 const markWord old_mark = mark;
970 const markWord new_mark = old_mark.copy_set_hash(hash);
971
972 mark = obj->cas_set_mark(new_mark, old_mark);
973 if (old_mark == mark) {
974 return hash;
975 }
976 }
977 }
978
979 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) {
980 // Since the monitor isn't in the object header, it can simply be installed.
981 if (UseObjectMonitorTable) {
982 return install_hash_code(current, obj);
983 }
984
985 while (true) {
986 ObjectMonitor* monitor = nullptr;
987 markWord temp, test;
988 intptr_t hash;
989 markWord mark = read_stable_mark(obj);
990 if (VerifyHeavyMonitors) {
991 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)");
992 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
993 }
994 if (mark.is_unlocked() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) {
995 hash = mark.hash();
996 if (hash != 0) { // if it has a hash, just return it
997 return hash;
998 }
999 hash = get_next_hash(current, obj); // get a new hash
1000 temp = mark.copy_set_hash(hash); // merge the hash into header
1001 // try to install the hash
1002 test = obj->cas_set_mark(temp, mark);
1003 if (test == mark) { // if the hash was installed, return it
1004 return hash;
1005 }
1006 if (LockingMode == LM_LIGHTWEIGHT) {
1007 // CAS failed, retry
1008 continue;
1009 }
1010 // Failed to install the hash. It could be that another thread
1011 // installed the hash just before our attempt or inflation has
1012 // occurred or... so we fall thru to inflate the monitor for
1013 // stability and then install the hash.
1014 } else if (mark.has_monitor()) {
1015 monitor = mark.monitor();
1016 temp = monitor->header();
1017 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1018 hash = temp.hash();
1019 if (hash != 0) {
1020 // It has a hash.
1021
1022 // Separate load of dmw/header above from the loads in
1023 // is_being_async_deflated().
1024
1025 // dmw/header and _contentions may get written by different threads.
1026 // Make sure to observe them in the same order when having several observers.
1027 OrderAccess::loadload_for_IRIW();
1028
1029 if (monitor->is_being_async_deflated()) {
1030 // But we can't safely use the hash if we detect that async
1031 // deflation has occurred. So we attempt to restore the
1032 // header/dmw to the object's header so that we only retry
1033 // once if the deflater thread happens to be slow.
1034 monitor->install_displaced_markword_in_object(obj);
1035 continue;
1036 }
1037 return hash;
1038 }
1039 // Fall thru so we only have one place that installs the hash in
1040 // the ObjectMonitor.
1041 } else if (LockingMode == LM_LEGACY && mark.has_locker()
1042 && current->is_Java_thread()
1043 && JavaThread::cast(current)->is_lock_owned((address)mark.locker())) {
1044 // This is a stack-lock owned by the calling thread so fetch the
1045 // displaced markWord from the BasicLock on the stack.
1046 temp = mark.displaced_mark_helper();
1047 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1048 hash = temp.hash();
1049 if (hash != 0) { // if it has a hash, just return it
1050 return hash;
1051 }
1052 // WARNING:
1053 // The displaced header in the BasicLock on a thread's stack
1054 // is strictly immutable. It CANNOT be changed in ANY cases.
1055 // So we have to inflate the stack-lock into an ObjectMonitor
1056 // even if the current thread owns the lock. The BasicLock on
1057 // a thread's stack can be asynchronously read by other threads
1058 // during an inflate() call so any change to that stack memory
1059 // may not propagate to other threads correctly.
1060 }
1061
1062 // Inflate the monitor to set the hash.
1063
1064 // There's no need to inflate if the mark has already got a monitor.
1065 // NOTE: an async deflation can race after we get the monitor and
1066 // before we can update the ObjectMonitor's header with the hash
1067 // value below.
1068 monitor = mark.has_monitor() ? mark.monitor() : inflate(current, obj, inflate_cause_hash_code);
1069 // Load ObjectMonitor's header/dmw field and see if it has a hash.
1070 mark = monitor->header();
1071 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
1072 hash = mark.hash();
1073 if (hash == 0) { // if it does not have a hash
1074 hash = get_next_hash(current, obj); // get a new hash
1075 temp = mark.copy_set_hash(hash) ; // merge the hash into header
1076 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1077 uintptr_t v = Atomic::cmpxchg(monitor->metadata_addr(), mark.value(), temp.value());
1078 test = markWord(v);
1079 if (test != mark) {
1080 // The attempt to update the ObjectMonitor's header/dmw field
1081 // did not work. This can happen if another thread managed to
1082 // merge in the hash just before our cmpxchg().
1083 // If we add any new usages of the header/dmw field, this code
1084 // will need to be updated.
1085 hash = test.hash();
1086 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value());
1087 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash");
1088 }
1089 if (monitor->is_being_async_deflated() && !UseObjectMonitorTable) {
1090 // If we detect that async deflation has occurred, then we
1091 // attempt to restore the header/dmw to the object's header
1092 // so that we only retry once if the deflater thread happens
1093 // to be slow.
1094 monitor->install_displaced_markword_in_object(obj);
1095 continue;
1096 }
1097 }
1098 // We finally get the hash.
1099 return hash;
1100 }
1101 }
1102
1103 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current,
1104 Handle h_obj) {
1105 assert(current == JavaThread::current(), "Can only be called on current thread");
1106 oop obj = h_obj();
1107
1108 markWord mark = read_stable_mark(obj);
1109
1110 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1111 // stack-locked case, header points into owner's stack
1112 return current->is_lock_owned((address)mark.locker());
1113 }
1114
1115 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1116 // fast-locking case, see if lock is in current's lock stack
1117 return current->lock_stack().contains(h_obj());
1118 }
1119
1120 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) {
1121 ObjectMonitor* monitor = read_monitor(current, obj, mark);
1122 if (monitor != nullptr) {
1123 return monitor->is_entered(current) != 0;
1124 }
1125 // Racing with inflation/deflation, retry
1126 mark = obj->mark_acquire();
1127
1128 if (mark.is_fast_locked()) {
1129 // Some other thread fast_locked, current could not have held the lock
1130 return false;
1131 }
1132 }
1133
1134 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) {
1135 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
1136 // The first stage of async deflation does not affect any field
1137 // used by this comparison so the ObjectMonitor* is usable here.
1138 ObjectMonitor* monitor = read_monitor(mark);
1139 return monitor->is_entered(current) != 0;
1140 }
1141 // Unlocked case, header in place
1142 assert(mark.is_unlocked(), "sanity check");
1143 return false;
1144 }
1145
1146 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
1147 oop obj = h_obj();
1148 markWord mark = read_stable_mark(obj);
1149
1150 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1151 // stack-locked so header points into owner's stack.
1152 // owning_thread_from_monitor_owner() may also return null here:
1153 return Threads::owning_thread_from_monitor_owner(t_list, (address) mark.locker());
1154 }
1155
1156 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1157 // fast-locked so get owner from the object.
1158 // owning_thread_from_object() may also return null here:
1159 return Threads::owning_thread_from_object(t_list, h_obj());
1160 }
1161
1162 while (LockingMode == LM_LIGHTWEIGHT && mark.has_monitor()) {
1163 ObjectMonitor* monitor = read_monitor(Thread::current(), obj, mark);
1164 if (monitor != nullptr) {
1165 return Threads::owning_thread_from_monitor(t_list, monitor);
1166 }
1167 // Racing with inflation/deflation, retry
1168 mark = obj->mark_acquire();
1169
1170 if (mark.is_fast_locked()) {
1171 // Some other thread fast_locked
1172 return Threads::owning_thread_from_object(t_list, h_obj());
1173 }
1174 }
1175
1176 if (LockingMode != LM_LIGHTWEIGHT && mark.has_monitor()) {
1177 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
1178 // The first stage of async deflation does not affect any field
1179 // used by this comparison so the ObjectMonitor* is usable here.
1180 ObjectMonitor* monitor = read_monitor(mark);
1181 assert(monitor != nullptr, "monitor should be non-null");
1182 // owning_thread_from_monitor() may also return null here:
1183 return Threads::owning_thread_from_monitor(t_list, monitor);
1184 }
1185
1186 // Unlocked case, header in place
1187 // Cannot have assertion since this object may have been
1188 // locked by another thread when reaching here.
1189 // assert(mark.is_unlocked(), "sanity check");
1190
1191 return nullptr;
1192 }
1193
1194 // Visitors ...
1195
1196 // Iterate over all ObjectMonitors.
1197 template <typename Function>
1198 void ObjectSynchronizer::monitors_iterate(Function function) {
1199 MonitorList::Iterator iter = _in_use_list.iterator();
1200 while (iter.has_next()) {
1201 ObjectMonitor* monitor = iter.next();
1202 function(monitor);
1203 }
1204 }
1205
1206 // Iterate ObjectMonitors owned by any thread and where the owner `filter`
1207 // returns true.
1208 template <typename OwnerFilter>
1209 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) {
1210 monitors_iterate([&](ObjectMonitor* monitor) {
1211 // This function is only called at a safepoint or when the
1212 // target thread is suspended or when the target thread is
1213 // operating on itself. The current closures in use today are
1214 // only interested in an owned ObjectMonitor and ownership
1215 // cannot be dropped under the calling contexts so the
1216 // ObjectMonitor cannot be async deflated.
1217 if (monitor->has_owner() && filter(monitor->owner_raw())) {
1218 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating");
1219
1220 closure->do_monitor(monitor);
1221 }
1222 });
1223 }
1224
1225 // Iterate ObjectMonitors where the owner == thread; this does NOT include
1226 // ObjectMonitors where owner is set to a stack-lock address in thread.
1227 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) {
1228 auto thread_filter = [&](void* owner) { return owner == thread; };
1229 return owned_monitors_iterate_filtered(closure, thread_filter);
1230 }
1231
1232 // Iterate ObjectMonitors owned by any thread.
1233 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) {
1234 auto all_filter = [&](void* owner) { return true; };
1235 return owned_monitors_iterate_filtered(closure, all_filter);
1236 }
1237
1238 static bool monitors_used_above_threshold(MonitorList* list) {
1239 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy
1240 return false;
1241 }
1242 // Start with ceiling based on a per-thread estimate:
1243 size_t ceiling = ObjectSynchronizer::in_use_list_ceiling();
1244 size_t old_ceiling = ceiling;
1245 if (ceiling < list->max()) {
1246 // The max used by the system has exceeded the ceiling so use that:
1247 ceiling = list->max();
1248 }
1249 size_t monitors_used = list->count();
1250 if (monitors_used == 0) { // empty list is easy
1251 return false;
1252 }
1253 if (NoAsyncDeflationProgressMax != 0 &&
1254 _no_progress_cnt >= NoAsyncDeflationProgressMax) {
1255 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0;
1256 size_t new_ceiling = ceiling / remainder + 1;
1257 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling);
1258 log_info(monitorinflation)("Too many deflations without progress; "
1259 "bumping in_use_list_ceiling from " SIZE_FORMAT
1260 " to " SIZE_FORMAT, old_ceiling, new_ceiling);
1261 _no_progress_cnt = 0;
1262 ceiling = new_ceiling;
1263 }
1264
1265 // Check if our monitor usage is above the threshold:
1266 size_t monitor_usage = (monitors_used * 100LL) / ceiling;
1267 if (int(monitor_usage) > MonitorUsedDeflationThreshold) {
1268 log_info(monitorinflation)("monitors_used=" SIZE_FORMAT ", ceiling=" SIZE_FORMAT
1269 ", monitor_usage=" SIZE_FORMAT ", threshold=%d",
1270 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold);
1271 return true;
1272 }
1273
1274 return false;
1275 }
1276
1277 size_t ObjectSynchronizer::in_use_list_ceiling() {
1278 return _in_use_list_ceiling;
1279 }
1280
1281 void ObjectSynchronizer::dec_in_use_list_ceiling() {
1282 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1283 }
1284
1285 void ObjectSynchronizer::inc_in_use_list_ceiling() {
1286 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1287 }
1288
1289 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) {
1290 _in_use_list_ceiling = new_value;
1291 }
1292
1293 bool ObjectSynchronizer::is_async_deflation_needed() {
1294 if (is_async_deflation_requested()) {
1295 // Async deflation request.
1296 log_info(monitorinflation)("Async deflation needed: explicit request");
1297 return true;
1298 }
1299
1300 jlong time_since_last = time_since_last_async_deflation_ms();
1301
1302 if (AsyncDeflationInterval > 0 &&
1303 time_since_last > AsyncDeflationInterval &&
1304 monitors_used_above_threshold(&_in_use_list)) {
1305 // It's been longer than our specified deflate interval and there
1306 // are too many monitors in use. We don't deflate more frequently
1307 // than AsyncDeflationInterval (unless is_async_deflation_requested)
1308 // in order to not swamp the MonitorDeflationThread.
1309 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold");
1310 return true;
1311 }
1312
1313 if (GuaranteedAsyncDeflationInterval > 0 &&
1314 time_since_last > GuaranteedAsyncDeflationInterval) {
1315 // It's been longer than our specified guaranteed deflate interval.
1316 // We need to clean up the used monitors even if the threshold is
1317 // not reached, to keep the memory utilization at bay when many threads
1318 // touched many monitors.
1319 log_info(monitorinflation)("Async deflation needed: guaranteed interval (" INTX_FORMAT " ms) "
1320 "is greater than time since last deflation (" JLONG_FORMAT " ms)",
1321 GuaranteedAsyncDeflationInterval, time_since_last);
1322
1323 // If this deflation has no progress, then it should not affect the no-progress
1324 // tracking, otherwise threshold heuristics would think it was triggered, experienced
1325 // no progress, and needs to backoff more aggressively. In this "no progress" case,
1326 // the generic code would bump the no-progress counter, and we compensate for that
1327 // by telling it to skip the update.
1328 //
1329 // If this deflation has progress, then it should let non-progress tracking
1330 // know about this, otherwise the threshold heuristics would kick in, potentially
1331 // experience no-progress due to aggressive cleanup by this deflation, and think
1332 // it is still in no-progress stride. In this "progress" case, the generic code would
1333 // zero the counter, and we allow it to happen.
1334 _no_progress_skip_increment = true;
1335
1336 return true;
1337 }
1338
1339 return false;
1340 }
1341
1342 void ObjectSynchronizer::request_deflate_idle_monitors() {
1343 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag);
1344 set_is_async_deflation_requested(true);
1345 ml.notify_all();
1346 }
1347
1348 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() {
1349 JavaThread* current = JavaThread::current();
1350 bool ret_code = false;
1351
1352 jlong last_time = last_async_deflation_time_ns();
1353
1354 request_deflate_idle_monitors();
1355
1356 const int N_CHECKS = 5;
1357 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds
1358 if (last_async_deflation_time_ns() > last_time) {
1359 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i);
1360 ret_code = true;
1361 break;
1362 }
1363 {
1364 // JavaThread has to honor the blocking protocol.
1365 ThreadBlockInVM tbivm(current);
1366 os::naked_short_sleep(999); // sleep for almost 1 second
1367 }
1368 }
1369 if (!ret_code) {
1370 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS);
1371 }
1372
1373 return ret_code;
1374 }
1375
1376 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() {
1377 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS);
1378 }
1379
1380 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1381 const oop obj,
1382 ObjectSynchronizer::InflateCause cause) {
1383 assert(event != nullptr, "invariant");
1384 event->set_monitorClass(obj->klass());
1385 event->set_address((uintptr_t)(void*)obj);
1386 event->set_cause((u1)cause);
1387 event->commit();
1388 }
1389
1390 // Fast path code shared by multiple functions
1391 void ObjectSynchronizer::inflate_helper(oop obj) {
1392 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter");
1393 markWord mark = obj->mark_acquire();
1394 if (mark.has_monitor()) {
1395 ObjectMonitor* monitor = read_monitor(mark);
1396 markWord dmw = monitor->header();
1397 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value());
1398 return;
1399 }
1400 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal);
1401 }
1402
1403 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) {
1404 assert(current == Thread::current(), "must be");
1405 assert(LockingMode != LM_LIGHTWEIGHT, "only inflate through enter");
1406 return inflate_impl(obj, cause);
1407 }
1408
1409 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) {
1410 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be");
1411 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_for");
1412 return inflate_impl(obj, cause);
1413 }
1414
1415 ObjectMonitor* ObjectSynchronizer::inflate_impl(oop object, const InflateCause cause) {
1416 assert(LockingMode != LM_LIGHTWEIGHT, "LM_LIGHTWEIGHT cannot use inflate_impl");
1417 EventJavaMonitorInflate event;
1418
1419 for (;;) {
1420 const markWord mark = object->mark_acquire();
1421
1422 // The mark can be in one of the following states:
1423 // * inflated - Just return it.
1424 // * stack-locked - Coerce it to inflated from stack-locked.
1425 // * INFLATING - Busy wait for conversion from stack-locked to
1426 // inflated.
1427 // * unlocked - Aggressively inflate the object.
1428
1429 // CASE: inflated
1430 if (mark.has_monitor()) {
1431 ObjectMonitor* inf = mark.monitor();
1432 markWord dmw = inf->header();
1433 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1434 return inf;
1435 }
1436
1437 // CASE: inflation in progress - inflating over a stack-lock.
1438 // Some other thread is converting from stack-locked to inflated.
1439 // Only that thread can complete inflation -- other threads must wait.
1440 // The INFLATING value is transient.
1441 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1442 // We could always eliminate polling by parking the thread on some auxiliary list.
1443 if (mark == markWord::INFLATING()) {
1444 read_stable_mark(object);
1445 continue;
1446 }
1447
1448 // CASE: stack-locked
1449 // Could be stack-locked either by current or by some other thread.
1450 //
1451 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting
1452 // to install INFLATING into the mark word. We originally installed INFLATING,
1453 // allocated the ObjectMonitor, and then finally STed the address of the
1454 // ObjectMonitor into the mark. This was correct, but artificially lengthened
1455 // the interval in which INFLATING appeared in the mark, thus increasing
1456 // the odds of inflation contention. If we lose the race to set INFLATING,
1457 // then we just delete the ObjectMonitor and loop around again.
1458 //
1459 LogStreamHandle(Trace, monitorinflation) lsh;
1460 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1461 ObjectMonitor* m = new ObjectMonitor(object);
1462 // Optimistically prepare the ObjectMonitor - anticipate successful CAS
1463 // We do this before the CAS in order to minimize the length of time
1464 // in which INFLATING appears in the mark.
1465
1466 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark);
1467 if (cmp != mark) {
1468 delete m;
1469 continue; // Interference -- just retry
1470 }
1471
1472 // We've successfully installed INFLATING (0) into the mark-word.
1473 // This is the only case where 0 will appear in a mark-word.
1474 // Only the singular thread that successfully swings the mark-word
1475 // to 0 can perform (or more precisely, complete) inflation.
1476 //
1477 // Why do we CAS a 0 into the mark-word instead of just CASing the
1478 // mark-word from the stack-locked value directly to the new inflated state?
1479 // Consider what happens when a thread unlocks a stack-locked object.
1480 // It attempts to use CAS to swing the displaced header value from the
1481 // on-stack BasicLock back into the object header. Recall also that the
1482 // header value (hash code, etc) can reside in (a) the object header, or
1483 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1484 // header in an ObjectMonitor. The inflate() routine must copy the header
1485 // value from the BasicLock on the owner's stack to the ObjectMonitor, all
1486 // the while preserving the hashCode stability invariants. If the owner
1487 // decides to release the lock while the value is 0, the unlock will fail
1488 // and control will eventually pass from slow_exit() to inflate. The owner
1489 // will then spin, waiting for the 0 value to disappear. Put another way,
1490 // the 0 causes the owner to stall if the owner happens to try to
1491 // drop the lock (restoring the header from the BasicLock to the object)
1492 // while inflation is in-progress. This protocol avoids races that might
1493 // would otherwise permit hashCode values to change or "flicker" for an object.
1494 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable.
1495 // 0 serves as a "BUSY" inflate-in-progress indicator.
1496
1497
1498 // fetch the displaced mark from the owner's stack.
1499 // The owner can't die or unwind past the lock while our INFLATING
1500 // object is in the mark. Furthermore the owner can't complete
1501 // an unlock on the object, either.
1502 markWord dmw = mark.displaced_mark_helper();
1503 // Catch if the object's header is not neutral (not locked and
1504 // not marked is what we care about here).
1505 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1506
1507 // Setup monitor fields to proper values -- prepare the monitor
1508 m->set_header(dmw);
1509
1510 // Optimization: if the mark.locker stack address is associated
1511 // with this thread we could simply set m->_owner = current.
1512 // Note that a thread can inflate an object
1513 // that it has stack-locked -- as might happen in wait() -- directly
1514 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom.
1515 m->set_owner_from(nullptr, mark.locker());
1516 // TODO-FIXME: assert BasicLock->dhw != 0.
1517
1518 // Must preserve store ordering. The monitor state must
1519 // be stable at the time of publishing the monitor address.
1520 guarantee(object->mark() == markWord::INFLATING(), "invariant");
1521 // Release semantics so that above set_object() is seen first.
1522 object->release_set_mark(markWord::encode(m));
1523
1524 // Once ObjectMonitor is configured and the object is associated
1525 // with the ObjectMonitor, it is safe to allow async deflation:
1526 _in_use_list.add(m);
1527
1528 // Hopefully the performance counters are allocated on distinct cache lines
1529 // to avoid false sharing on MP systems ...
1530 OM_PERFDATA_OP(Inflations, inc());
1531 if (log_is_enabled(Trace, monitorinflation)) {
1532 ResourceMark rm;
1533 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1534 INTPTR_FORMAT ", type='%s'", p2i(object),
1535 object->mark().value(), object->klass()->external_name());
1536 }
1537 if (event.should_commit()) {
1538 post_monitor_inflate_event(&event, object, cause);
1539 }
1540 return m;
1541 }
1542
1543 // CASE: unlocked
1544 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1545 // If we know we're inflating for entry it's better to inflate by swinging a
1546 // pre-locked ObjectMonitor pointer into the object header. A successful
1547 // CAS inflates the object *and* confers ownership to the inflating thread.
1548 // In the current implementation we use a 2-step mechanism where we CAS()
1549 // to inflate and then CAS() again to try to swing _owner from null to current.
1550 // An inflateTry() method that we could call from enter() would be useful.
1551
1552 assert(mark.is_unlocked(), "invariant: header=" INTPTR_FORMAT, mark.value());
1553 ObjectMonitor* m = new ObjectMonitor(object);
1554 // prepare m for installation - set monitor to initial state
1555 m->set_header(mark);
1556
1557 if (object->cas_set_mark(markWord::encode(m), mark) != mark) {
1558 delete m;
1559 m = nullptr;
1560 continue;
1561 // interference - the markword changed - just retry.
1562 // The state-transitions are one-way, so there's no chance of
1563 // live-lock -- "Inflated" is an absorbing state.
1564 }
1565
1566 // Once the ObjectMonitor is configured and object is associated
1567 // with the ObjectMonitor, it is safe to allow async deflation:
1568 _in_use_list.add(m);
1569
1570 // Hopefully the performance counters are allocated on distinct
1571 // cache lines to avoid false sharing on MP systems ...
1572 OM_PERFDATA_OP(Inflations, inc());
1573 if (log_is_enabled(Trace, monitorinflation)) {
1574 ResourceMark rm;
1575 lsh.print_cr("inflate(unlocked): object=" INTPTR_FORMAT ", mark="
1576 INTPTR_FORMAT ", type='%s'", p2i(object),
1577 object->mark().value(), object->klass()->external_name());
1578 }
1579 if (event.should_commit()) {
1580 post_monitor_inflate_event(&event, object, cause);
1581 }
1582 return m;
1583 }
1584 }
1585
1586 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle
1587 // ObjectMonitors. Returns the number of deflated ObjectMonitors.
1588 //
1589 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) {
1590 MonitorList::Iterator iter = _in_use_list.iterator();
1591 size_t deflated_count = 0;
1592 Thread* current = Thread::current();
1593
1594 while (iter.has_next()) {
1595 if (deflated_count >= (size_t)MonitorDeflationMax) {
1596 break;
1597 }
1598 ObjectMonitor* mid = iter.next();
1599 if (mid->deflate_monitor(current)) {
1600 deflated_count++;
1601 }
1602
1603 // Must check for a safepoint/handshake and honor it.
1604 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count);
1605 }
1606
1607 return deflated_count;
1608 }
1609
1610 class HandshakeForDeflation : public HandshakeClosure {
1611 public:
1612 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {}
1613
1614 void do_thread(Thread* thread) {
1615 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread="
1616 INTPTR_FORMAT, p2i(thread));
1617 if (thread->is_Java_thread()) {
1618 // Clear OM cache
1619 JavaThread* jt = JavaThread::cast(thread);
1620 jt->om_clear_monitor_cache();
1621 }
1622 }
1623 };
1624
1625 class VM_RendezvousGCThreads : public VM_Operation {
1626 public:
1627 bool evaluate_at_safepoint() const override { return false; }
1628 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; }
1629 void doit() override {
1630 Universe::heap()->safepoint_synchronize_begin();
1631 Universe::heap()->safepoint_synchronize_end();
1632 };
1633 };
1634
1635 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list,
1636 ObjectMonitorDeflationSafepointer* safepointer) {
1637 NativeHeapTrimmer::SuspendMark sm("monitor deletion");
1638 size_t deleted_count = 0;
1639 for (ObjectMonitor* monitor: *delete_list) {
1640 delete monitor;
1641 deleted_count++;
1642 // A JavaThread must check for a safepoint/handshake and honor it.
1643 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count);
1644 }
1645 return deleted_count;
1646 }
1647
1648 class ObjectMonitorDeflationLogging: public StackObj {
1649 LogStreamHandle(Debug, monitorinflation) _debug;
1650 LogStreamHandle(Info, monitorinflation) _info;
1651 LogStream* _stream;
1652 elapsedTimer _timer;
1653
1654 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); }
1655 size_t count() const { return ObjectSynchronizer::_in_use_list.count(); }
1656 size_t max() const { return ObjectSynchronizer::_in_use_list.max(); }
1657
1658 public:
1659 ObjectMonitorDeflationLogging()
1660 : _debug(), _info(), _stream(nullptr) {
1661 if (_debug.is_enabled()) {
1662 _stream = &_debug;
1663 } else if (_info.is_enabled()) {
1664 _stream = &_info;
1665 }
1666 }
1667
1668 void begin() {
1669 if (_stream != nullptr) {
1670 _stream->print_cr("begin deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1671 ceiling(), count(), max());
1672 _timer.start();
1673 }
1674 }
1675
1676 void before_handshake(size_t unlinked_count) {
1677 if (_stream != nullptr) {
1678 _timer.stop();
1679 _stream->print_cr("before handshaking: unlinked_count=" SIZE_FORMAT
1680 ", in_use_list stats: ceiling=" SIZE_FORMAT ", count="
1681 SIZE_FORMAT ", max=" SIZE_FORMAT,
1682 unlinked_count, ceiling(), count(), max());
1683 }
1684 }
1685
1686 void after_handshake() {
1687 if (_stream != nullptr) {
1688 _stream->print_cr("after handshaking: in_use_list stats: ceiling="
1689 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1690 ceiling(), count(), max());
1691 _timer.start();
1692 }
1693 }
1694
1695 void end(size_t deflated_count, size_t unlinked_count) {
1696 if (_stream != nullptr) {
1697 _timer.stop();
1698 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) {
1699 _stream->print_cr("deflated_count=" SIZE_FORMAT ", {unlinked,deleted}_count=" SIZE_FORMAT " monitors in %3.7f secs",
1700 deflated_count, unlinked_count, _timer.seconds());
1701 }
1702 _stream->print_cr("end deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1703 ceiling(), count(), max());
1704 }
1705 }
1706
1707 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) {
1708 if (_stream != nullptr) {
1709 _timer.stop();
1710 _stream->print_cr("pausing %s: %s=" SIZE_FORMAT ", in_use_list stats: ceiling="
1711 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1712 op_name, cnt_name, cnt, ceiling(), count(), max());
1713 }
1714 }
1715
1716 void after_block_for_safepoint(const char* op_name) {
1717 if (_stream != nullptr) {
1718 _stream->print_cr("resuming %s: in_use_list stats: ceiling=" SIZE_FORMAT
1719 ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, op_name,
1720 ceiling(), count(), max());
1721 _timer.start();
1722 }
1723 }
1724 };
1725
1726 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) {
1727 if (!SafepointMechanism::should_process(_current)) {
1728 return;
1729 }
1730
1731 // A safepoint/handshake has started.
1732 _log->before_block_for_safepoint(op_name, count_name, counter);
1733
1734 {
1735 // Honor block request.
1736 ThreadBlockInVM tbivm(_current);
1737 }
1738
1739 _log->after_block_for_safepoint(op_name);
1740 }
1741
1742 // This function is called by the MonitorDeflationThread to deflate
1743 // ObjectMonitors.
1744 size_t ObjectSynchronizer::deflate_idle_monitors() {
1745 JavaThread* current = JavaThread::current();
1746 assert(current->is_monitor_deflation_thread(), "The only monitor deflater");
1747
1748 // The async deflation request has been processed.
1749 _last_async_deflation_time_ns = os::javaTimeNanos();
1750 set_is_async_deflation_requested(false);
1751
1752 ObjectMonitorDeflationLogging log;
1753 ObjectMonitorDeflationSafepointer safepointer(current, &log);
1754
1755 log.begin();
1756
1757 // Deflate some idle ObjectMonitors.
1758 size_t deflated_count = deflate_monitor_list(&safepointer);
1759
1760 // Unlink the deflated ObjectMonitors from the in-use list.
1761 size_t unlinked_count = 0;
1762 size_t deleted_count = 0;
1763 if (deflated_count > 0) {
1764 ResourceMark rm(current);
1765 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count);
1766 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer);
1767
1768 #ifdef ASSERT
1769 if (UseObjectMonitorTable) {
1770 for (ObjectMonitor* monitor : delete_list) {
1771 assert(!LightweightSynchronizer::contains_monitor(current, monitor), "Should have been removed");
1772 }
1773 }
1774 #endif
1775
1776 log.before_handshake(unlinked_count);
1777
1778 // A JavaThread needs to handshake in order to safely free the
1779 // ObjectMonitors that were deflated in this cycle.
1780 HandshakeForDeflation hfd_hc;
1781 Handshake::execute(&hfd_hc);
1782 // Also, we sync and desync GC threads around the handshake, so that they can
1783 // safely read the mark-word and look-through to the object-monitor, without
1784 // being afraid that the object-monitor is going away.
1785 VM_RendezvousGCThreads sync_gc;
1786 VMThread::execute(&sync_gc);
1787
1788 log.after_handshake();
1789
1790 // After the handshake, safely free the ObjectMonitors that were
1791 // deflated and unlinked in this cycle.
1792
1793 // Delete the unlinked ObjectMonitors.
1794 deleted_count = delete_monitors(&delete_list, &safepointer);
1795 assert(unlinked_count == deleted_count, "must be");
1796 }
1797
1798 log.end(deflated_count, unlinked_count);
1799
1800 OM_PERFDATA_OP(MonExtant, set_value(_in_use_list.count()));
1801 OM_PERFDATA_OP(Deflations, inc(deflated_count));
1802
1803 GVars.stw_random = os::random();
1804
1805 if (deflated_count != 0) {
1806 _no_progress_cnt = 0;
1807 } else if (_no_progress_skip_increment) {
1808 _no_progress_skip_increment = false;
1809 } else {
1810 _no_progress_cnt++;
1811 }
1812
1813 return deflated_count;
1814 }
1815
1816 // Monitor cleanup on JavaThread::exit
1817
1818 // Iterate through monitor cache and attempt to release thread's monitors
1819 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1820 private:
1821 JavaThread* _thread;
1822
1823 public:
1824 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {}
1825 void do_monitor(ObjectMonitor* mid) {
1826 intx rec = mid->complete_exit(_thread);
1827 _thread->dec_held_monitor_count(rec + 1);
1828 }
1829 };
1830
1831 // Release all inflated monitors owned by current thread. Lightweight monitors are
1832 // ignored. This is meant to be called during JNI thread detach which assumes
1833 // all remaining monitors are heavyweight. All exceptions are swallowed.
1834 // Scanning the extant monitor list can be time consuming.
1835 // A simple optimization is to add a per-thread flag that indicates a thread
1836 // called jni_monitorenter() during its lifetime.
1837 //
1838 // Instead of NoSafepointVerifier it might be cheaper to
1839 // use an idiom of the form:
1840 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1841 // <code that must not run at safepoint>
1842 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1843 // Since the tests are extremely cheap we could leave them enabled
1844 // for normal product builds.
1845
1846 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) {
1847 assert(current == JavaThread::current(), "must be current Java thread");
1848 NoSafepointVerifier nsv;
1849 ReleaseJavaMonitorsClosure rjmc(current);
1850 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current);
1851 assert(!current->has_pending_exception(), "Should not be possible");
1852 current->clear_pending_exception();
1853 assert(current->held_monitor_count() == 0, "Should not be possible");
1854 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count.
1855 current->clear_jni_monitor_count();
1856 }
1857
1858 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1859 switch (cause) {
1860 case inflate_cause_vm_internal: return "VM Internal";
1861 case inflate_cause_monitor_enter: return "Monitor Enter";
1862 case inflate_cause_wait: return "Monitor Wait";
1863 case inflate_cause_notify: return "Monitor Notify";
1864 case inflate_cause_hash_code: return "Monitor Hash Code";
1865 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1866 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1867 default:
1868 ShouldNotReachHere();
1869 }
1870 return "Unknown";
1871 }
1872
1873 //------------------------------------------------------------------------------
1874 // Debugging code
1875
1876 u_char* ObjectSynchronizer::get_gvars_addr() {
1877 return (u_char*)&GVars;
1878 }
1879
1880 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() {
1881 return (u_char*)&GVars.hc_sequence;
1882 }
1883
1884 size_t ObjectSynchronizer::get_gvars_size() {
1885 return sizeof(SharedGlobals);
1886 }
1887
1888 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() {
1889 return (u_char*)&GVars.stw_random;
1890 }
1891
1892 // Do the final audit and print of ObjectMonitor stats; must be done
1893 // by the VMThread at VM exit time.
1894 void ObjectSynchronizer::do_final_audit_and_print_stats() {
1895 assert(Thread::current()->is_VM_thread(), "sanity check");
1896
1897 if (is_final_audit()) { // Only do the audit once.
1898 return;
1899 }
1900 set_is_final_audit();
1901 log_info(monitorinflation)("Starting the final audit.");
1902
1903 if (log_is_enabled(Info, monitorinflation)) {
1904 LogStreamHandle(Info, monitorinflation) ls;
1905 audit_and_print_stats(&ls, true /* on_exit */);
1906 }
1907 }
1908
1909 // This function can be called by the MonitorDeflationThread or it can be called when
1910 // we are trying to exit the VM. The list walker functions can run in parallel with
1911 // the other list operations.
1912 // Calls to this function can be added in various places as a debugging
1913 // aid.
1914 //
1915 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) {
1916 int error_cnt = 0;
1917
1918 ls->print_cr("Checking in_use_list:");
1919 chk_in_use_list(ls, &error_cnt);
1920
1921 if (error_cnt == 0) {
1922 ls->print_cr("No errors found in in_use_list checks.");
1923 } else {
1924 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt);
1925 }
1926
1927 // When exiting, only log the interesting entries at the Info level.
1928 // When called at intervals by the MonitorDeflationThread, log output
1929 // at the Trace level since there can be a lot of it.
1930 if (!on_exit && log_is_enabled(Trace, monitorinflation)) {
1931 LogStreamHandle(Trace, monitorinflation) ls_tr;
1932 log_in_use_monitor_details(&ls_tr, true /* log_all */);
1933 } else if (on_exit) {
1934 log_in_use_monitor_details(ls, false /* log_all */);
1935 }
1936
1937 ls->flush();
1938
1939 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt);
1940 }
1941
1942 // Check the in_use_list; log the results of the checks.
1943 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) {
1944 size_t l_in_use_count = _in_use_list.count();
1945 size_t l_in_use_max = _in_use_list.max();
1946 out->print_cr("count=" SIZE_FORMAT ", max=" SIZE_FORMAT, l_in_use_count,
1947 l_in_use_max);
1948
1949 size_t ck_in_use_count = 0;
1950 MonitorList::Iterator iter = _in_use_list.iterator();
1951 while (iter.has_next()) {
1952 ObjectMonitor* mid = iter.next();
1953 chk_in_use_entry(mid, out, error_cnt_p);
1954 ck_in_use_count++;
1955 }
1956
1957 if (l_in_use_count == ck_in_use_count) {
1958 out->print_cr("in_use_count=" SIZE_FORMAT " equals ck_in_use_count="
1959 SIZE_FORMAT, l_in_use_count, ck_in_use_count);
1960 } else {
1961 out->print_cr("WARNING: in_use_count=" SIZE_FORMAT " is not equal to "
1962 "ck_in_use_count=" SIZE_FORMAT, l_in_use_count,
1963 ck_in_use_count);
1964 }
1965
1966 size_t ck_in_use_max = _in_use_list.max();
1967 if (l_in_use_max == ck_in_use_max) {
1968 out->print_cr("in_use_max=" SIZE_FORMAT " equals ck_in_use_max="
1969 SIZE_FORMAT, l_in_use_max, ck_in_use_max);
1970 } else {
1971 out->print_cr("WARNING: in_use_max=" SIZE_FORMAT " is not equal to "
1972 "ck_in_use_max=" SIZE_FORMAT, l_in_use_max, ck_in_use_max);
1973 }
1974 }
1975
1976 // Check an in-use monitor entry; log any errors.
1977 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out,
1978 int* error_cnt_p) {
1979 if (n->owner_is_DEFLATER_MARKER()) {
1980 // This could happen when monitor deflation blocks for a safepoint.
1981 return;
1982 }
1983
1984
1985 if (n->metadata() == 0) {
1986 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must "
1987 "have non-null _metadata (header/hash) field.", p2i(n));
1988 *error_cnt_p = *error_cnt_p + 1;
1989 }
1990
1991 const oop obj = n->object_peek();
1992 if (obj == nullptr) {
1993 return;
1994 }
1995
1996 const markWord mark = obj->mark();
1997 if (!mark.has_monitor()) {
1998 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
1999 "object does not think it has a monitor: obj="
2000 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n),
2001 p2i(obj), mark.value());
2002 *error_cnt_p = *error_cnt_p + 1;
2003 return;
2004 }
2005
2006 ObjectMonitor* const obj_mon = read_monitor(Thread::current(), obj, mark);
2007 if (n != obj_mon) {
2008 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2009 "object does not refer to the same monitor: obj="
2010 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon="
2011 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon));
2012 *error_cnt_p = *error_cnt_p + 1;
2013 }
2014 }
2015
2016 // Log details about ObjectMonitors on the in_use_list. The 'BHL'
2017 // flags indicate why the entry is in-use, 'object' and 'object type'
2018 // indicate the associated object and its type.
2019 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) {
2020 if (_in_use_list.count() > 0) {
2021 stringStream ss;
2022 out->print_cr("In-use monitor info:");
2023 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2024 out->print_cr("%18s %s %18s %18s",
2025 "monitor", "BHL", "object", "object type");
2026 out->print_cr("================== === ================== ==================");
2027
2028 auto is_interesting = [&](ObjectMonitor* monitor) {
2029 return log_all || monitor->has_owner() || monitor->is_busy();
2030 };
2031
2032 monitors_iterate([&](ObjectMonitor* monitor) {
2033 if (is_interesting(monitor)) {
2034 const oop obj = monitor->object_peek();
2035 const intptr_t hash = UseObjectMonitorTable ? monitor->hash() : monitor->header().hash();
2036 ResourceMark rm;
2037 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor),
2038 monitor->is_busy(), hash != 0, monitor->owner() != nullptr,
2039 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name());
2040 if (monitor->is_busy()) {
2041 out->print(" (%s)", monitor->is_busy_to_string(&ss));
2042 ss.reset();
2043 }
2044 out->cr();
2045 }
2046 });
2047 }
2048
2049 out->flush();
2050 }