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/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/lockStack.inline.hpp"
45 #include "runtime/mutexLocker.hpp"
46 #include "runtime/objectMonitor.hpp"
47 #include "runtime/objectMonitor.inline.hpp"
48 #include "runtime/os.inline.hpp"
49 #include "runtime/osThread.hpp"
50 #include "runtime/perfData.hpp"
51 #include "runtime/safepointMechanism.inline.hpp"
52 #include "runtime/safepointVerifiers.hpp"
53 #include "runtime/sharedRuntime.hpp"
54 #include "runtime/stubRoutines.hpp"
55 #include "runtime/synchronizer.hpp"
56 #include "runtime/threads.hpp"
57 #include "runtime/timer.hpp"
58 #include "runtime/trimNativeHeap.hpp"
59 #include "runtime/vframe.hpp"
60 #include "runtime/vmThread.hpp"
61 #include "utilities/align.hpp"
62 #include "utilities/dtrace.hpp"
63 #include "utilities/events.hpp"
64 #include "utilities/globalDefinitions.hpp"
65 #include "utilities/linkedlist.hpp"
66 #include "utilities/preserveException.hpp"
67
68 class ObjectMonitorDeflationLogging;
69
70 void MonitorList::add(ObjectMonitor* m) {
71 ObjectMonitor* head;
72 do {
73 head = Atomic::load(&_head);
74 m->set_next_om(head);
75 } while (Atomic::cmpxchg(&_head, head, m) != head);
76
77 size_t count = Atomic::add(&_count, 1u);
78 if (count > max()) {
79 Atomic::inc(&_max);
80 }
81 }
82
83 size_t MonitorList::count() const {
84 return Atomic::load(&_count);
85 }
86
87 size_t MonitorList::max() const {
88 return Atomic::load(&_max);
89 }
90
91 class ObjectMonitorDeflationSafepointer : public StackObj {
92 JavaThread* const _current;
93 ObjectMonitorDeflationLogging* const _log;
94
95 public:
96 ObjectMonitorDeflationSafepointer(JavaThread* current, ObjectMonitorDeflationLogging* log)
97 : _current(current), _log(log) {}
98
99 void block_for_safepoint(const char* op_name, const char* count_name, size_t counter);
100 };
101
102 // Walk the in-use list and unlink deflated ObjectMonitors.
103 // Returns the number of unlinked ObjectMonitors.
104 size_t MonitorList::unlink_deflated(size_t deflated_count,
105 GrowableArray<ObjectMonitor*>* unlinked_list,
106 ObjectMonitorDeflationSafepointer* safepointer) {
107 size_t unlinked_count = 0;
108 ObjectMonitor* prev = nullptr;
109 ObjectMonitor* m = Atomic::load_acquire(&_head);
110
111 while (m != nullptr) {
112 if (m->is_being_async_deflated()) {
113 // Find next live ObjectMonitor. Batch up the unlinkable monitors, so we can
114 // modify the list once per batch. The batch starts at "m".
115 size_t unlinked_batch = 0;
116 ObjectMonitor* next = m;
117 // Look for at most MonitorUnlinkBatch monitors, or the number of
118 // deflated and not unlinked monitors, whatever comes first.
119 assert(deflated_count >= unlinked_count, "Sanity: underflow");
120 size_t unlinked_batch_limit = MIN2<size_t>(deflated_count - unlinked_count, MonitorUnlinkBatch);
121 do {
122 ObjectMonitor* next_next = next->next_om();
123 unlinked_batch++;
124 unlinked_list->append(next);
125 next = next_next;
126 if (unlinked_batch >= unlinked_batch_limit) {
127 // Reached the max batch, so bail out of the gathering loop.
128 break;
129 }
130 if (prev == nullptr && Atomic::load(&_head) != m) {
131 // Current batch used to be at head, but it is not at head anymore.
132 // Bail out and figure out where we currently are. This avoids long
133 // walks searching for new prev during unlink under heavy list inserts.
134 break;
135 }
136 } while (next != nullptr && next->is_being_async_deflated());
137
138 // Unlink the found batch.
139 if (prev == nullptr) {
140 // The current batch is the first batch, so there is a chance that it starts at head.
141 // Optimistically assume no inserts happened, and try to unlink the entire batch from the head.
142 ObjectMonitor* prev_head = Atomic::cmpxchg(&_head, m, next);
143 if (prev_head != m) {
144 // Something must have updated the head. Figure out the actual prev for this batch.
145 for (ObjectMonitor* n = prev_head; n != m; n = n->next_om()) {
146 prev = n;
147 }
148 assert(prev != nullptr, "Should have found the prev for the current batch");
149 prev->set_next_om(next);
150 }
151 } else {
152 // The current batch is preceded by another batch. This guarantees the current batch
153 // does not start at head. Unlink the entire current batch without updating the head.
154 assert(Atomic::load(&_head) != m, "Sanity");
155 prev->set_next_om(next);
156 }
157
158 unlinked_count += unlinked_batch;
159 if (unlinked_count >= deflated_count) {
160 // Reached the max so bail out of the searching loop.
161 // There should be no more deflated monitors left.
162 break;
163 }
164 m = next;
165 } else {
166 prev = m;
167 m = m->next_om();
168 }
169
170 // Must check for a safepoint/handshake and honor it.
171 safepointer->block_for_safepoint("unlinking", "unlinked_count", unlinked_count);
172 }
173
174 #ifdef ASSERT
175 // Invariant: the code above should unlink all deflated monitors.
176 // The code that runs after this unlinking does not expect deflated monitors.
177 // Notably, attempting to deflate the already deflated monitor would break.
178 {
179 ObjectMonitor* m = Atomic::load_acquire(&_head);
180 while (m != nullptr) {
181 assert(!m->is_being_async_deflated(), "All deflated monitors should be unlinked");
182 m = m->next_om();
183 }
184 }
185 #endif
186
187 Atomic::sub(&_count, unlinked_count);
188 return unlinked_count;
189 }
190
191 MonitorList::Iterator MonitorList::iterator() const {
192 return Iterator(Atomic::load_acquire(&_head));
193 }
194
195 ObjectMonitor* MonitorList::Iterator::next() {
196 ObjectMonitor* current = _current;
197 _current = current->next_om();
198 return current;
199 }
200
201 // The "core" versions of monitor enter and exit reside in this file.
202 // The interpreter and compilers contain specialized transliterated
203 // variants of the enter-exit fast-path operations. See c2_MacroAssembler_x86.cpp
204 // fast_lock(...) for instance. If you make changes here, make sure to modify the
205 // interpreter, and both C1 and C2 fast-path inline locking code emission.
206 //
207 // -----------------------------------------------------------------------------
208
209 #ifdef DTRACE_ENABLED
210
211 // Only bother with this argument setup if dtrace is available
212 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
213
214 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
215 char* bytes = nullptr; \
216 int len = 0; \
217 jlong jtid = SharedRuntime::get_java_tid(thread); \
218 Symbol* klassname = obj->klass()->name(); \
219 if (klassname != nullptr) { \
220 bytes = (char*)klassname->bytes(); \
221 len = klassname->utf8_length(); \
222 }
223
224 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
225 { \
226 if (DTraceMonitorProbes) { \
227 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
228 HOTSPOT_MONITOR_WAIT(jtid, \
229 (uintptr_t)(monitor), bytes, len, (millis)); \
230 } \
231 }
232
233 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY
234 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL
235 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
236
237 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
238 { \
239 if (DTraceMonitorProbes) { \
240 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
241 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
242 (uintptr_t)(monitor), bytes, len); \
243 } \
244 }
245
246 #else // ndef DTRACE_ENABLED
247
248 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
249 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
250
251 #endif // ndef DTRACE_ENABLED
252
253 // This exists only as a workaround of dtrace bug 6254741
254 static int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, JavaThread* thr) {
255 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
256 return 0;
257 }
258
259 static constexpr size_t inflation_lock_count() {
260 return 256;
261 }
262
263 // Static storage for an array of PlatformMutex.
264 alignas(PlatformMutex) static uint8_t _inflation_locks[inflation_lock_count()][sizeof(PlatformMutex)];
265
266 static inline PlatformMutex* inflation_lock(size_t index) {
267 return reinterpret_cast<PlatformMutex*>(_inflation_locks[index]);
268 }
269
270 void ObjectSynchronizer::initialize() {
271 for (size_t i = 0; i < inflation_lock_count(); i++) {
272 ::new(static_cast<void*>(inflation_lock(i))) PlatformMutex();
273 }
274 // Start the ceiling with the estimate for one thread.
275 set_in_use_list_ceiling(AvgMonitorsPerThreadEstimate);
276
277 // Start the timer for deflations, so it does not trigger immediately.
278 _last_async_deflation_time_ns = os::javaTimeNanos();
279 }
280
281 MonitorList ObjectSynchronizer::_in_use_list;
282 // monitors_used_above_threshold() policy is as follows:
283 //
284 // The ratio of the current _in_use_list count to the ceiling is used
285 // to determine if we are above MonitorUsedDeflationThreshold and need
286 // to do an async monitor deflation cycle. The ceiling is increased by
287 // AvgMonitorsPerThreadEstimate when a thread is added to the system
288 // and is decreased by AvgMonitorsPerThreadEstimate when a thread is
289 // removed from the system.
290 //
291 // Note: If the _in_use_list max exceeds the ceiling, then
292 // monitors_used_above_threshold() will use the in_use_list max instead
293 // of the thread count derived ceiling because we have used more
294 // ObjectMonitors than the estimated average.
295 //
296 // Note: If deflate_idle_monitors() has NoAsyncDeflationProgressMax
297 // no-progress async monitor deflation cycles in a row, then the ceiling
298 // is adjusted upwards by monitors_used_above_threshold().
299 //
300 // Start the ceiling with the estimate for one thread in initialize()
301 // which is called after cmd line options are processed.
302 static size_t _in_use_list_ceiling = 0;
303 bool volatile ObjectSynchronizer::_is_async_deflation_requested = false;
304 bool volatile ObjectSynchronizer::_is_final_audit = false;
305 jlong ObjectSynchronizer::_last_async_deflation_time_ns = 0;
306 static uintx _no_progress_cnt = 0;
307 static bool _no_progress_skip_increment = false;
308
309 // =====================> Quick functions
310
311 // The quick_* forms are special fast-path variants used to improve
312 // performance. In the simplest case, a "quick_*" implementation could
313 // simply return false, in which case the caller will perform the necessary
314 // state transitions and call the slow-path form.
315 // The fast-path is designed to handle frequently arising cases in an efficient
316 // manner and is just a degenerate "optimistic" variant of the slow-path.
317 // returns true -- to indicate the call was satisfied.
318 // returns false -- to indicate the call needs the services of the slow-path.
319 // A no-loitering ordinance is in effect for code in the quick_* family
320 // operators: safepoints or indefinite blocking (blocking that might span a
321 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon
322 // entry.
323 //
324 // Consider: An interesting optimization is to have the JIT recognize the
325 // following common idiom:
326 // synchronized (someobj) { .... ; notify(); }
327 // That is, we find a notify() or notifyAll() call that immediately precedes
328 // the monitorexit operation. In that case the JIT could fuse the operations
329 // into a single notifyAndExit() runtime primitive.
330
331 bool ObjectSynchronizer::quick_notify(oopDesc* obj, JavaThread* current, bool all) {
332 assert(current->thread_state() == _thread_in_Java, "invariant");
333 NoSafepointVerifier nsv;
334 if (obj == nullptr) return false; // slow-path for invalid obj
335 const markWord mark = obj->mark();
336
337 if (LockingMode == LM_LIGHTWEIGHT) {
338 if (mark.is_fast_locked() && current->lock_stack().contains(cast_to_oop(obj))) {
339 // Degenerate notify
340 // fast-locked by caller so by definition the implied waitset is empty.
341 return true;
342 }
343 } else if (LockingMode == LM_LEGACY) {
344 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
345 // Degenerate notify
346 // stack-locked by caller so by definition the implied waitset is empty.
347 return true;
348 }
349 }
350
351 if (mark.has_monitor()) {
352 ObjectMonitor* const mon = mark.monitor();
353 assert(mon->object() == oop(obj), "invariant");
354 if (!mon->is_owner(current)) return false; // slow-path for IMS exception
355
356 if (mon->first_waiter() != nullptr) {
357 // We have one or more waiters. Since this is an inflated monitor
358 // that we own, we can transfer one or more threads from the waitset
359 // to the entrylist here and now, avoiding the slow-path.
360 if (all) {
361 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, current);
362 } else {
363 DTRACE_MONITOR_PROBE(notify, mon, obj, current);
364 }
365 int free_count = 0;
366 do {
367 mon->INotify(current);
368 ++free_count;
369 } while (mon->first_waiter() != nullptr && all);
370 OM_PERFDATA_OP(Notifications, inc(free_count));
371 }
372 return true;
373 }
374
375 // other IMS exception states take the slow-path
376 return false;
377 }
378
379
380 // The LockNode emitted directly at the synchronization site would have
381 // been too big if it were to have included support for the cases of inflated
382 // recursive enter and exit, so they go here instead.
383 // Note that we can't safely call AsyncPrintJavaStack() from within
384 // quick_enter() as our thread state remains _in_Java.
385
386 bool ObjectSynchronizer::quick_enter(oop obj, JavaThread* current,
387 BasicLock * lock) {
388 assert(current->thread_state() == _thread_in_Java, "invariant");
389 NoSafepointVerifier nsv;
390 if (obj == nullptr) return false; // Need to throw NPE
391
392 if (obj->klass()->is_value_based()) {
393 return false;
394 }
395
396 if (LockingMode == LM_LIGHTWEIGHT) {
397 LockStack& lock_stack = current->lock_stack();
398 if (lock_stack.is_full()) {
399 // Always go into runtime if the lock stack is full.
400 return false;
401 }
402 if (lock_stack.try_recursive_enter(obj)) {
403 // Recursive lock successful.
404 NOT_LOOM_MONITOR_SUPPORT(current->inc_held_monitor_count();)
405 return true;
406 }
407 }
408
409 const markWord mark = obj->mark();
410
411 if (mark.has_monitor()) {
412 ObjectMonitor* const m = mark.monitor();
413 // An async deflation or GC can race us before we manage to make
414 // the ObjectMonitor busy by setting the owner below. If we detect
415 // that race we just bail out to the slow-path here.
416 if (m->object_peek() == nullptr) {
417 return false;
418 }
419
420 // Lock contention and Transactional Lock Elision (TLE) diagnostics
421 // and observability
422 // Case: light contention possibly amenable to TLE
423 // Case: TLE inimical operations such as nested/recursive synchronization
424
425 if (m->is_owner(current)) {
426 m->_recursions++;
427 NOT_LOOM_MONITOR_SUPPORT(current->inc_held_monitor_count();)
428 return true;
429 }
430
431 if (LockingMode != LM_LIGHTWEIGHT) {
432 // This Java Monitor is inflated so obj's header will never be
433 // displaced to this thread's BasicLock. Make the displaced header
434 // non-null so this BasicLock is not seen as recursive nor as
435 // being locked. We do this unconditionally so that this thread's
436 // BasicLock cannot be mis-interpreted by any stack walkers. For
437 // performance reasons, stack walkers generally first check for
438 // stack-locking in the object's header, the second check is for
439 // recursive stack-locking in the displaced header in the BasicLock,
440 // and last are the inflated Java Monitor (ObjectMonitor) checks.
441 lock->set_displaced_header(markWord::unused_mark());
442 }
443
444 if (!m->has_owner() && m->try_set_owner_from(nullptr, current) == nullptr) {
445 assert(m->_recursions == 0, "invariant");
446 NOT_LOOM_MONITOR_SUPPORT(current->inc_held_monitor_count();)
447 return true;
448 }
449 }
450
451 // Note that we could inflate in quick_enter.
452 // This is likely a useful optimization
453 // Critically, in quick_enter() we must not:
454 // -- block indefinitely, or
455 // -- reach a safepoint
456
457 return false; // revert to slow-path
458 }
459
460 // Handle notifications when synchronizing on value based classes
461 void ObjectSynchronizer::handle_sync_on_value_based_class(Handle obj, JavaThread* locking_thread) {
462 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
463 frame last_frame = locking_thread->last_frame();
464 bool bcp_was_adjusted = false;
465 // Don't decrement bcp if it points to the frame's first instruction. This happens when
466 // handle_sync_on_value_based_class() is called because of a synchronized method. There
467 // is no actual monitorenter instruction in the byte code in this case.
468 if (last_frame.is_interpreted_frame() &&
469 (last_frame.interpreter_frame_method()->code_base() < last_frame.interpreter_frame_bcp())) {
470 // adjust bcp to point back to monitorenter so that we print the correct line numbers
471 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() - 1);
472 bcp_was_adjusted = true;
473 }
474
475 if (DiagnoseSyncOnValueBasedClasses == FATAL_EXIT) {
476 ResourceMark rm;
477 stringStream ss;
478 locking_thread->print_active_stack_on(&ss);
479 char* base = (char*)strstr(ss.base(), "at");
480 char* newline = (char*)strchr(ss.base(), '\n');
481 if (newline != nullptr) {
482 *newline = '\0';
483 }
484 fatal("Synchronizing on object " INTPTR_FORMAT " of klass %s %s", p2i(obj()), obj->klass()->external_name(), base);
485 } else {
486 assert(DiagnoseSyncOnValueBasedClasses == LOG_WARNING, "invalid value for DiagnoseSyncOnValueBasedClasses");
487 ResourceMark rm;
488 Log(valuebasedclasses) vblog;
489
490 vblog.info("Synchronizing on object " INTPTR_FORMAT " of klass %s", p2i(obj()), obj->klass()->external_name());
491 if (locking_thread->has_last_Java_frame()) {
492 LogStream info_stream(vblog.info());
493 locking_thread->print_active_stack_on(&info_stream);
494 } else {
495 vblog.info("Cannot find the last Java frame");
496 }
497
498 EventSyncOnValueBasedClass event;
499 if (event.should_commit()) {
500 event.set_valueBasedClass(obj->klass());
501 event.commit();
502 }
503 }
504
505 if (bcp_was_adjusted) {
506 last_frame.interpreter_frame_set_bcp(last_frame.interpreter_frame_bcp() + 1);
507 }
508 }
509
510 static bool useHeavyMonitors() {
511 #if defined(X86) || defined(AARCH64) || defined(PPC64) || defined(RISCV64) || defined(S390)
512 return LockingMode == LM_MONITOR;
513 #else
514 return false;
515 #endif
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 NOT_LOOM_MONITOR_SUPPORT(locking_thread->inc_held_monitor_count();)
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(Handle obj, BasicLock* lock, JavaThread* current) {
546 assert(current == Thread::current(), "must be");
547
548 NOT_LOOM_MONITOR_SUPPORT(current->inc_held_monitor_count();)
549
550 if (!enter_fast_impl(obj, lock, current)) {
551 // Inflated ObjectMonitor::enter is required
552
553 // An async deflation can race after the inflate() call and before
554 // enter() can make the ObjectMonitor busy. enter() returns false if
555 // we have lost the race to async deflation and we simply try again.
556 while (true) {
557 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_monitor_enter);
558 if (monitor->enter(current)) {
559 return;
560 }
561 }
562 }
563 }
564
565 // The interpreter and compiler assembly code tries to lock using the fast path
566 // of this algorithm. Make sure to update that code if the following function is
567 // changed. The implementation is extremely sensitive to race condition. Be careful.
568 bool ObjectSynchronizer::enter_fast_impl(Handle obj, BasicLock* lock, JavaThread* locking_thread) {
569 if (obj->klass()->is_value_based()) {
570 handle_sync_on_value_based_class(obj, locking_thread);
571 }
572
573 if (!useHeavyMonitors()) {
574 if (LockingMode == LM_LIGHTWEIGHT) {
575 // Fast-locking does not use the 'lock' argument.
576 LockStack& lock_stack = locking_thread->lock_stack();
577 if (lock_stack.is_full()) {
578 // We unconditionally make room on the lock stack by inflating
579 // the least recently locked object on the lock stack.
580
581 // About the choice to inflate least recently locked object.
582 // First we must chose to inflate a lock, either some lock on
583 // the lock-stack or the lock that is currently being entered
584 // (which may or may not be on the lock-stack).
585 // Second the best lock to inflate is a lock which is entered
586 // in a control flow where there are only a very few locks being
587 // used, as the costly part of inflated locking is inflation,
588 // not locking. But this property is entirely program dependent.
589 // Third inflating the lock currently being entered on when it
590 // is not present on the lock-stack will result in a still full
591 // lock-stack. This creates a scenario where every deeper nested
592 // monitorenter must call into the runtime.
593 // The rational here is as follows:
594 // Because we cannot (currently) figure out the second, and want
595 // to avoid the third, we inflate a lock on the lock-stack.
596 // The least recently locked lock is chosen as it is the lock
597 // with the longest critical section.
598
599 log_info(monitorinflation)("LockStack capacity exceeded, inflating.");
600 ObjectMonitor* monitor = inflate_for(locking_thread, lock_stack.bottom(), inflate_cause_vm_internal);
601 assert(monitor->is_owner(JavaThread::current()), "must be owner=" PTR_FORMAT " current=" PTR_FORMAT " mark=" PTR_FORMAT,
602 p2i(monitor->owner()), p2i(Thread::current()), monitor->object()->mark_acquire().value());
603 assert(!lock_stack.is_full(), "must have made room here");
604 }
605
606 markWord mark = obj()->mark_acquire();
607 while (mark.is_unlocked()) {
608 // Retry until a lock state change has been observed. cas_set_mark() may collide with non lock bits modifications.
609 // Try to swing into 'fast-locked' state.
610 assert(!lock_stack.contains(obj()), "thread must not already hold the lock");
611 const markWord locked_mark = mark.set_fast_locked();
612 const markWord old_mark = obj()->cas_set_mark(locked_mark, mark);
613 if (old_mark == mark) {
614 // Successfully fast-locked, push object to lock-stack and return.
615 lock_stack.push(obj());
616 return true;
617 }
618 mark = old_mark;
619 }
620
621 if (mark.is_fast_locked() && lock_stack.try_recursive_enter(obj())) {
622 // Recursive lock successful.
623 return true;
624 }
625
626 // Failed to fast lock.
627 return false;
628 } else if (LockingMode == LM_LEGACY) {
629 markWord mark = obj->mark();
630 if (mark.is_unlocked()) {
631 // Anticipate successful CAS -- the ST of the displaced mark must
632 // be visible <= the ST performed by the CAS.
633 lock->set_displaced_header(mark);
634 if (mark == obj()->cas_set_mark(markWord::from_pointer(lock), mark)) {
635 LOOM_MONITOR_SUPPORT_ONLY(locking_thread->inc_held_monitor_count();)
636 return true;
637 }
638 } else if (mark.has_locker() &&
639 locking_thread->is_lock_owned((address) mark.locker())) {
640 assert(lock != mark.locker(), "must not re-lock the same lock");
641 assert(lock != (BasicLock*) obj->mark().value(), "don't relock with same BasicLock");
642 lock->set_displaced_header(markWord::from_pointer(nullptr));
643 return true;
644 }
645
646 // The object header will never be displaced to this lock,
647 // so it does not matter what the value is, except that it
648 // must be non-zero to avoid looking like a re-entrant lock,
649 // and must not look locked either.
650 lock->set_displaced_header(markWord::unused_mark());
651
652 // Failed to fast lock.
653 return false;
654 }
655 } else if (VerifyHeavyMonitors) {
656 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
657 }
658
659 return false;
660 }
661
662 void ObjectSynchronizer::exit(oop object, BasicLock* lock, JavaThread* current) {
663 NOT_LOOM_MONITOR_SUPPORT(current->dec_held_monitor_count();)
664
665 if (!useHeavyMonitors()) {
666 markWord mark = object->mark();
667 if (LockingMode == LM_LIGHTWEIGHT) {
668 // Fast-locking does not use the 'lock' argument.
669 LockStack& lock_stack = current->lock_stack();
670 if (mark.is_fast_locked() && lock_stack.try_recursive_exit(object)) {
671 // Recursively unlocked.
672 return;
673 }
674
675 if (mark.is_fast_locked() && lock_stack.is_recursive(object)) {
676 // This lock is recursive but is not at the top of the lock stack so we're
677 // doing an unbalanced exit. We have to fall thru to inflation below and
678 // let ObjectMonitor::exit() do the unlock.
679 } else {
680 while (mark.is_fast_locked()) {
681 // Retry until a lock state change has been observed. cas_set_mark() may collide with non lock bits modifications.
682 const markWord unlocked_mark = mark.set_unlocked();
683 const markWord old_mark = object->cas_set_mark(unlocked_mark, mark);
684 if (old_mark == mark) {
685 size_t recursions = lock_stack.remove(object) - 1;
686 assert(recursions == 0, "must not be recursive here");
687 return;
688 }
689 mark = old_mark;
690 }
691 }
692 } else if (LockingMode == LM_LEGACY) {
693 markWord dhw = lock->displaced_header();
694 if (dhw.value() == 0) {
695 // If the displaced header is null, then this exit matches up with
696 // a recursive enter. No real work to do here except for diagnostics.
697 #ifndef PRODUCT
698 if (mark != markWord::INFLATING()) {
699 // Only do diagnostics if we are not racing an inflation. Simply
700 // exiting a recursive enter of a Java Monitor that is being
701 // inflated is safe; see the has_monitor() comment below.
702 assert(!mark.is_unlocked(), "invariant");
703 assert(!mark.has_locker() ||
704 current->is_lock_owned((address)mark.locker()), "invariant");
705 if (mark.has_monitor()) {
706 // The BasicLock's displaced_header is marked as a recursive
707 // enter and we have an inflated Java Monitor (ObjectMonitor).
708 // This is a special case where the Java Monitor was inflated
709 // after this thread entered the stack-lock recursively. When a
710 // Java Monitor is inflated, we cannot safely walk the Java
711 // Monitor owner's stack and update the BasicLocks because a
712 // Java Monitor can be asynchronously inflated by a thread that
713 // does not own the Java Monitor.
714 ObjectMonitor* m = mark.monitor();
715 assert(m->object()->mark() == mark, "invariant");
716 assert(m->is_entered(current), "invariant");
717 }
718 }
719 #endif
720 return;
721 }
722
723 if (mark == markWord::from_pointer(lock)) {
724 // If the object is stack-locked by the current thread, try to
725 // swing the displaced header from the BasicLock back to the mark.
726 assert(dhw.is_neutral(), "invariant");
727 if (object->cas_set_mark(dhw, mark) == mark) {
728 LOOM_MONITOR_SUPPORT_ONLY(current->dec_held_monitor_count();)
729 return;
730 }
731 }
732 }
733 } else if (VerifyHeavyMonitors) {
734 guarantee((object->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
735 }
736
737 // We have to take the slow-path of possible inflation and then exit.
738 // The ObjectMonitor* can't be async deflated until ownership is
739 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
740 ObjectMonitor* monitor = inflate(current, object, inflate_cause_vm_internal);
741 assert(!monitor->is_owner_anonymous(), "must not be");
742 monitor->exit(current);
743 }
744
745 // -----------------------------------------------------------------------------
746 // JNI locks on java objects
747 // NOTE: must use heavy weight monitor to handle jni monitor enter
748 void ObjectSynchronizer::jni_enter(Handle obj, JavaThread* current) {
749 if (obj->klass()->is_value_based()) {
750 handle_sync_on_value_based_class(obj, current);
751 }
752
753 // the current locking is from JNI instead of Java code
754 current->set_current_pending_monitor_is_from_java(false);
755 // An async deflation can race after the inflate() call and before
756 // enter() can make the ObjectMonitor busy. enter() returns false if
757 // we have lost the race to async deflation and we simply try again.
758 while (true) {
759 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_jni_enter);
760 if (monitor->enter(current)) {
761 current->inc_held_monitor_count(1, true);
762 break;
763 }
764 }
765 current->set_current_pending_monitor_is_from_java(true);
766 }
767
768 // NOTE: must use heavy weight monitor to handle jni monitor exit
769 void ObjectSynchronizer::jni_exit(oop obj, TRAPS) {
770 JavaThread* current = THREAD;
771
772 // The ObjectMonitor* can't be async deflated until ownership is
773 // dropped inside exit() and the ObjectMonitor* must be !is_busy().
774 ObjectMonitor* monitor = inflate(current, obj, inflate_cause_jni_exit);
775 // If this thread has locked the object, exit the monitor. We
776 // intentionally do not use CHECK on check_owner because we must exit the
777 // monitor even if an exception was already pending.
778 if (monitor->check_owner(THREAD)) {
779 monitor->exit(current);
780 current->dec_held_monitor_count(1, true);
781 }
782 }
783
784 // -----------------------------------------------------------------------------
785 // Internal VM locks on java objects
786 // standard constructor, allows locking failures
787 ObjectLocker::ObjectLocker(Handle obj, JavaThread* thread) {
788 _thread = thread;
789 _thread->check_for_valid_safepoint_state();
790 DEBUG_ONLY(_thread->inc_obj_locker_count();)
791 _obj = obj;
792
793 if (_obj() != nullptr) {
794 ObjectSynchronizer::enter(_obj, &_lock, _thread);
795 }
796 }
797
798 ObjectLocker::~ObjectLocker() {
799 DEBUG_ONLY(_thread->dec_obj_locker_count();)
800 if (_obj() != nullptr) {
801 ObjectSynchronizer::exit(_obj(), &_lock, _thread);
802 }
803 }
804
805
806 // -----------------------------------------------------------------------------
807 // Wait/Notify/NotifyAll
808 // NOTE: must use heavy weight monitor to handle wait()
809 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
810 JavaThread* current = THREAD;
811 if (millis < 0) {
812 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
813 }
814 // The ObjectMonitor* can't be async deflated because the _waiters
815 // field is incremented before ownership is dropped and decremented
816 // after ownership is regained.
817 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_wait);
818
819 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), current, millis);
820 monitor->wait(millis, true, THREAD); // Not CHECK as we need following code
821
822 // This dummy call is in place to get around dtrace bug 6254741. Once
823 // that's fixed we can uncomment the following line, remove the call
824 // and change this function back into a "void" func.
825 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
826 int ret_code = dtrace_waited_probe(monitor, obj, THREAD);
827 return ret_code;
828 }
829
830 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
831 if (millis < 0) {
832 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
833 }
834 ObjectSynchronizer::inflate(THREAD,
835 obj(),
836 inflate_cause_wait)->wait(millis, false, THREAD);
837 }
838
839
840 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
841 JavaThread* current = THREAD;
842
843 markWord mark = obj->mark();
844 if (LockingMode == LM_LIGHTWEIGHT) {
845 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
846 // Not inflated so there can't be any waiters to notify.
847 return;
848 }
849 } else if (LockingMode == LM_LEGACY) {
850 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
851 // Not inflated so there can't be any waiters to notify.
852 return;
853 }
854 }
855 // The ObjectMonitor* can't be async deflated until ownership is
856 // dropped by the calling thread.
857 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_notify);
858 monitor->notify(CHECK);
859 }
860
861 // NOTE: see comment of notify()
862 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
863 JavaThread* current = THREAD;
864
865 markWord mark = obj->mark();
866 if (LockingMode == LM_LIGHTWEIGHT) {
867 if ((mark.is_fast_locked() && current->lock_stack().contains(obj()))) {
868 // Not inflated so there can't be any waiters to notify.
869 return;
870 }
871 } else if (LockingMode == LM_LEGACY) {
872 if (mark.has_locker() && current->is_lock_owned((address)mark.locker())) {
873 // Not inflated so there can't be any waiters to notify.
874 return;
875 }
876 }
877 // The ObjectMonitor* can't be async deflated until ownership is
878 // dropped by the calling thread.
879 ObjectMonitor* monitor = inflate(current, obj(), inflate_cause_notify);
880 monitor->notifyAll(CHECK);
881 }
882
883 // -----------------------------------------------------------------------------
884 // Hash Code handling
885
886 struct SharedGlobals {
887 char _pad_prefix[OM_CACHE_LINE_SIZE];
888 // This is a highly shared mostly-read variable.
889 // To avoid false-sharing it needs to be the sole occupant of a cache line.
890 volatile int stw_random;
891 DEFINE_PAD_MINUS_SIZE(1, OM_CACHE_LINE_SIZE, sizeof(volatile int));
892 // Hot RW variable -- Sequester to avoid false-sharing
893 volatile int hc_sequence;
894 DEFINE_PAD_MINUS_SIZE(2, OM_CACHE_LINE_SIZE, sizeof(volatile int));
895 };
896
897 static SharedGlobals GVars;
898
899 static markWord read_stable_mark(oop obj) {
900 markWord mark = obj->mark_acquire();
901 if (!mark.is_being_inflated() || LockingMode == LM_LIGHTWEIGHT) {
902 // New lightweight locking does not use the markWord::INFLATING() protocol.
903 return mark; // normal fast-path return
904 }
905
906 int its = 0;
907 for (;;) {
908 markWord mark = obj->mark_acquire();
909 if (!mark.is_being_inflated()) {
910 return mark; // normal fast-path return
911 }
912
913 // The object is being inflated by some other thread.
914 // The caller of read_stable_mark() must wait for inflation to complete.
915 // Avoid live-lock.
916
917 ++its;
918 if (its > 10000 || !os::is_MP()) {
919 if (its & 1) {
920 os::naked_yield();
921 } else {
922 // Note that the following code attenuates the livelock problem but is not
923 // a complete remedy. A more complete solution would require that the inflating
924 // thread hold the associated inflation lock. The following code simply restricts
925 // the number of spinners to at most one. We'll have N-2 threads blocked
926 // on the inflationlock, 1 thread holding the inflation lock and using
927 // a yield/park strategy, and 1 thread in the midst of inflation.
928 // A more refined approach would be to change the encoding of INFLATING
929 // to allow encapsulation of a native thread pointer. Threads waiting for
930 // inflation to complete would use CAS to push themselves onto a singly linked
931 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
932 // and calling park(). When inflation was complete the thread that accomplished inflation
933 // would detach the list and set the markword to inflated with a single CAS and
934 // then for each thread on the list, set the flag and unpark() the thread.
935
936 // Index into the lock array based on the current object address.
937 static_assert(is_power_of_2(inflation_lock_count()), "must be");
938 size_t ix = (cast_from_oop<intptr_t>(obj) >> 5) & (inflation_lock_count() - 1);
939 int YieldThenBlock = 0;
940 assert(ix < inflation_lock_count(), "invariant");
941 inflation_lock(ix)->lock();
942 while (obj->mark_acquire() == markWord::INFLATING()) {
943 // Beware: naked_yield() is advisory and has almost no effect on some platforms
944 // so we periodically call current->_ParkEvent->park(1).
945 // We use a mixed spin/yield/block mechanism.
946 if ((YieldThenBlock++) >= 16) {
947 Thread::current()->_ParkEvent->park(1);
948 } else {
949 os::naked_yield();
950 }
951 }
952 inflation_lock(ix)->unlock();
953 }
954 } else {
955 SpinPause(); // SMP-polite spinning
956 }
957 }
958 }
959
960 // hashCode() generation :
961 //
962 // Possibilities:
963 // * MD5Digest of {obj,stw_random}
964 // * CRC32 of {obj,stw_random} or any linear-feedback shift register function.
965 // * A DES- or AES-style SBox[] mechanism
966 // * One of the Phi-based schemes, such as:
967 // 2654435761 = 2^32 * Phi (golden ratio)
968 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stw_random ;
969 // * A variation of Marsaglia's shift-xor RNG scheme.
970 // * (obj ^ stw_random) is appealing, but can result
971 // in undesirable regularity in the hashCode values of adjacent objects
972 // (objects allocated back-to-back, in particular). This could potentially
973 // result in hashtable collisions and reduced hashtable efficiency.
974 // There are simple ways to "diffuse" the middle address bits over the
975 // generated hashCode values:
976
977 static inline intptr_t get_next_hash(Thread* current, oop obj) {
978 intptr_t value = 0;
979 if (hashCode == 0) {
980 // This form uses global Park-Miller RNG.
981 // On MP system we'll have lots of RW access to a global, so the
982 // mechanism induces lots of coherency traffic.
983 value = os::random();
984 } else if (hashCode == 1) {
985 // This variation has the property of being stable (idempotent)
986 // between STW operations. This can be useful in some of the 1-0
987 // synchronization schemes.
988 intptr_t addr_bits = cast_from_oop<intptr_t>(obj) >> 3;
989 value = addr_bits ^ (addr_bits >> 5) ^ GVars.stw_random;
990 } else if (hashCode == 2) {
991 value = 1; // for sensitivity testing
992 } else if (hashCode == 3) {
993 value = ++GVars.hc_sequence;
994 } else if (hashCode == 4) {
995 value = cast_from_oop<intptr_t>(obj);
996 } else {
997 // Marsaglia's xor-shift scheme with thread-specific state
998 // This is probably the best overall implementation -- we'll
999 // likely make this the default in future releases.
1000 unsigned t = current->_hashStateX;
1001 t ^= (t << 11);
1002 current->_hashStateX = current->_hashStateY;
1003 current->_hashStateY = current->_hashStateZ;
1004 current->_hashStateZ = current->_hashStateW;
1005 unsigned v = current->_hashStateW;
1006 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
1007 current->_hashStateW = v;
1008 value = v;
1009 }
1010
1011 value &= markWord::hash_mask;
1012 if (value == 0) value = 0xBAD;
1013 assert(value != markWord::no_hash, "invariant");
1014 return value;
1015 }
1016
1017 intptr_t ObjectSynchronizer::FastHashCode(Thread* current, oop obj) {
1018
1019 while (true) {
1020 ObjectMonitor* monitor = nullptr;
1021 markWord temp, test;
1022 intptr_t hash;
1023 markWord mark = read_stable_mark(obj);
1024 if (VerifyHeavyMonitors) {
1025 assert(LockingMode == LM_MONITOR, "+VerifyHeavyMonitors requires LockingMode == 0 (LM_MONITOR)");
1026 guarantee((obj->mark().value() & markWord::lock_mask_in_place) != markWord::locked_value, "must not be lightweight/stack-locked");
1027 }
1028 if (mark.is_unlocked() || (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked())) {
1029 hash = mark.hash();
1030 if (hash != 0) { // if it has a hash, just return it
1031 return hash;
1032 }
1033 hash = get_next_hash(current, obj); // get a new hash
1034 temp = mark.copy_set_hash(hash); // merge the hash into header
1035 // try to install the hash
1036 test = obj->cas_set_mark(temp, mark);
1037 if (test == mark) { // if the hash was installed, return it
1038 return hash;
1039 }
1040 if (LockingMode == LM_LIGHTWEIGHT) {
1041 // CAS failed, retry
1042 continue;
1043 }
1044 // Failed to install the hash. It could be that another thread
1045 // installed the hash just before our attempt or inflation has
1046 // occurred or... so we fall thru to inflate the monitor for
1047 // stability and then install the hash.
1048 } else if (mark.has_monitor()) {
1049 monitor = mark.monitor();
1050 temp = monitor->header();
1051 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1052 hash = temp.hash();
1053 if (hash != 0) {
1054 // It has a hash.
1055
1056 // Separate load of dmw/header above from the loads in
1057 // is_being_async_deflated().
1058
1059 // dmw/header and _contentions may get written by different threads.
1060 // Make sure to observe them in the same order when having several observers.
1061 OrderAccess::loadload_for_IRIW();
1062
1063 if (monitor->is_being_async_deflated()) {
1064 // But we can't safely use the hash if we detect that async
1065 // deflation has occurred. So we attempt to restore the
1066 // header/dmw to the object's header so that we only retry
1067 // once if the deflater thread happens to be slow.
1068 monitor->install_displaced_markword_in_object(obj);
1069 continue;
1070 }
1071 return hash;
1072 }
1073 // Fall thru so we only have one place that installs the hash in
1074 // the ObjectMonitor.
1075 } else if (LockingMode == LM_LEGACY && mark.has_locker()
1076 && current->is_Java_thread()
1077 && JavaThread::cast(current)->is_lock_owned((address)mark.locker())) {
1078 // This is a stack-lock owned by the calling thread so fetch the
1079 // displaced markWord from the BasicLock on the stack.
1080 temp = mark.displaced_mark_helper();
1081 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1082 hash = temp.hash();
1083 if (hash != 0) { // if it has a hash, just return it
1084 return hash;
1085 }
1086 // WARNING:
1087 // The displaced header in the BasicLock on a thread's stack
1088 // is strictly immutable. It CANNOT be changed in ANY cases.
1089 // So we have to inflate the stack-lock into an ObjectMonitor
1090 // even if the current thread owns the lock. The BasicLock on
1091 // a thread's stack can be asynchronously read by other threads
1092 // during an inflate() call so any change to that stack memory
1093 // may not propagate to other threads correctly.
1094 }
1095
1096 // Inflate the monitor to set the hash.
1097
1098 // There's no need to inflate if the mark has already got a monitor.
1099 // NOTE: an async deflation can race after we get the monitor and
1100 // before we can update the ObjectMonitor's header with the hash
1101 // value below.
1102 monitor = mark.has_monitor() ? mark.monitor() : inflate(current, obj, inflate_cause_hash_code);
1103 // Load ObjectMonitor's header/dmw field and see if it has a hash.
1104 mark = monitor->header();
1105 assert(mark.is_neutral(), "invariant: header=" INTPTR_FORMAT, mark.value());
1106 hash = mark.hash();
1107 if (hash == 0) { // if it does not have a hash
1108 hash = get_next_hash(current, obj); // get a new hash
1109 temp = mark.copy_set_hash(hash) ; // merge the hash into header
1110 assert(temp.is_neutral(), "invariant: header=" INTPTR_FORMAT, temp.value());
1111 uintptr_t v = Atomic::cmpxchg((volatile uintptr_t*)monitor->header_addr(), mark.value(), temp.value());
1112 test = markWord(v);
1113 if (test != mark) {
1114 // The attempt to update the ObjectMonitor's header/dmw field
1115 // did not work. This can happen if another thread managed to
1116 // merge in the hash just before our cmpxchg().
1117 // If we add any new usages of the header/dmw field, this code
1118 // will need to be updated.
1119 hash = test.hash();
1120 assert(test.is_neutral(), "invariant: header=" INTPTR_FORMAT, test.value());
1121 assert(hash != 0, "should only have lost the race to a thread that set a non-zero hash");
1122 }
1123 if (monitor->is_being_async_deflated()) {
1124 // If we detect that async deflation has occurred, then we
1125 // attempt to restore the header/dmw to the object's header
1126 // so that we only retry once if the deflater thread happens
1127 // to be slow.
1128 monitor->install_displaced_markword_in_object(obj);
1129 continue;
1130 }
1131 }
1132 // We finally get the hash.
1133 return hash;
1134 }
1135 }
1136
1137 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* current,
1138 Handle h_obj) {
1139 assert(current == JavaThread::current(), "Can only be called on current thread");
1140 oop obj = h_obj();
1141
1142 markWord mark = read_stable_mark(obj);
1143
1144 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1145 // stack-locked case, header points into owner's stack
1146 return current->is_lock_owned((address)mark.locker());
1147 }
1148
1149 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1150 // fast-locking case, see if lock is in current's lock stack
1151 return current->lock_stack().contains(h_obj());
1152 }
1153
1154 if (mark.has_monitor()) {
1155 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
1156 // The first stage of async deflation does not affect any field
1157 // used by this comparison so the ObjectMonitor* is usable here.
1158 ObjectMonitor* monitor = mark.monitor();
1159 return monitor->is_entered(current) != 0;
1160 }
1161 // Unlocked case, header in place
1162 assert(mark.is_unlocked(), "sanity check");
1163 return false;
1164 }
1165
1166 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
1167 oop obj = h_obj();
1168 markWord mark = read_stable_mark(obj);
1169
1170 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1171 // stack-locked so header points into owner's stack.
1172 // owning_thread_from_monitor_owner() may also return null here:
1173 return Threads::owning_thread_from_stacklock(t_list, (address) mark.locker());
1174 }
1175
1176 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1177 // fast-locked so get owner from the object.
1178 // owning_thread_from_object() may also return null here:
1179 return Threads::owning_thread_from_object(t_list, h_obj());
1180 }
1181
1182 if (mark.has_monitor()) {
1183 // Inflated monitor so header points to ObjectMonitor (tagged pointer).
1184 // The first stage of async deflation does not affect any field
1185 // used by this comparison so the ObjectMonitor* is usable here.
1186 ObjectMonitor* monitor = mark.monitor();
1187 assert(monitor != nullptr, "monitor should be non-null");
1188 // owning_thread_from_monitor() may also return null here:
1189 return Threads::owning_thread_from_monitor(t_list, monitor);
1190 }
1191
1192 // Unlocked case, header in place
1193 // Cannot have assertion since this object may have been
1194 // locked by another thread when reaching here.
1195 // assert(mark.is_unlocked(), "sanity check");
1196
1197 return nullptr;
1198 }
1199
1200 // Visitors ...
1201
1202 // Iterate over all ObjectMonitors.
1203 template <typename Function>
1204 void ObjectSynchronizer::monitors_iterate(Function function) {
1205 MonitorList::Iterator iter = _in_use_list.iterator();
1206 while (iter.has_next()) {
1207 ObjectMonitor* monitor = iter.next();
1208 function(monitor);
1209 }
1210 }
1211
1212 // Iterate ObjectMonitors owned by any thread and where the owner `filter`
1213 // returns true.
1214 template <typename OwnerFilter>
1215 void ObjectSynchronizer::owned_monitors_iterate_filtered(MonitorClosure* closure, OwnerFilter filter) {
1216 monitors_iterate([&](ObjectMonitor* monitor) {
1217 // This function is only called at a safepoint or when the
1218 // target thread is suspended or when the target thread is
1219 // operating on itself. The current closures in use today are
1220 // only interested in an owned ObjectMonitor and ownership
1221 // cannot be dropped under the calling contexts so the
1222 // ObjectMonitor cannot be async deflated.
1223 if (monitor->has_owner() && filter(monitor)) {
1224 assert(!monitor->is_being_async_deflated(), "Owned monitors should not be deflating");
1225
1226 closure->do_monitor(monitor);
1227 }
1228 });
1229 }
1230
1231 // Iterate ObjectMonitors where the owner == thread; this does NOT include
1232 // ObjectMonitors where owner is set to a stack-lock address in thread.
1233 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure, JavaThread* thread) {
1234 auto thread_filter = [&](ObjectMonitor* monitor) { return monitor->is_owner(thread); };
1235 return owned_monitors_iterate_filtered(closure, thread_filter);
1236 }
1237
1238 // Iterate ObjectMonitors owned by any thread.
1239 void ObjectSynchronizer::owned_monitors_iterate(MonitorClosure* closure) {
1240 auto all_filter = [&](ObjectMonitor* monitor) { return true; };
1241 return owned_monitors_iterate_filtered(closure, all_filter);
1242 }
1243
1244 static bool monitors_used_above_threshold(MonitorList* list) {
1245 if (MonitorUsedDeflationThreshold == 0) { // disabled case is easy
1246 return false;
1247 }
1248 // Start with ceiling based on a per-thread estimate:
1249 size_t ceiling = ObjectSynchronizer::in_use_list_ceiling();
1250 size_t old_ceiling = ceiling;
1251 if (ceiling < list->max()) {
1252 // The max used by the system has exceeded the ceiling so use that:
1253 ceiling = list->max();
1254 }
1255 size_t monitors_used = list->count();
1256 if (monitors_used == 0) { // empty list is easy
1257 return false;
1258 }
1259 if (NoAsyncDeflationProgressMax != 0 &&
1260 _no_progress_cnt >= NoAsyncDeflationProgressMax) {
1261 double remainder = (100.0 - MonitorUsedDeflationThreshold) / 100.0;
1262 size_t new_ceiling = ceiling + (size_t)((double)ceiling * remainder) + 1;
1263 ObjectSynchronizer::set_in_use_list_ceiling(new_ceiling);
1264 log_info(monitorinflation)("Too many deflations without progress; "
1265 "bumping in_use_list_ceiling from " SIZE_FORMAT
1266 " to " SIZE_FORMAT, old_ceiling, new_ceiling);
1267 _no_progress_cnt = 0;
1268 ceiling = new_ceiling;
1269 }
1270
1271 // Check if our monitor usage is above the threshold:
1272 size_t monitor_usage = (monitors_used * 100LL) / ceiling;
1273 if (int(monitor_usage) > MonitorUsedDeflationThreshold) {
1274 log_info(monitorinflation)("monitors_used=" SIZE_FORMAT ", ceiling=" SIZE_FORMAT
1275 ", monitor_usage=" SIZE_FORMAT ", threshold=%d",
1276 monitors_used, ceiling, monitor_usage, MonitorUsedDeflationThreshold);
1277 return true;
1278 }
1279
1280 return false;
1281 }
1282
1283 size_t ObjectSynchronizer::in_use_list_ceiling() {
1284 return _in_use_list_ceiling;
1285 }
1286
1287 void ObjectSynchronizer::dec_in_use_list_ceiling() {
1288 Atomic::sub(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1289 }
1290
1291 void ObjectSynchronizer::inc_in_use_list_ceiling() {
1292 Atomic::add(&_in_use_list_ceiling, AvgMonitorsPerThreadEstimate);
1293 }
1294
1295 void ObjectSynchronizer::set_in_use_list_ceiling(size_t new_value) {
1296 _in_use_list_ceiling = new_value;
1297 }
1298
1299 bool ObjectSynchronizer::is_async_deflation_needed() {
1300 if (is_async_deflation_requested()) {
1301 // Async deflation request.
1302 log_info(monitorinflation)("Async deflation needed: explicit request");
1303 return true;
1304 }
1305
1306 jlong time_since_last = time_since_last_async_deflation_ms();
1307
1308 if (AsyncDeflationInterval > 0 &&
1309 time_since_last > AsyncDeflationInterval &&
1310 monitors_used_above_threshold(&_in_use_list)) {
1311 // It's been longer than our specified deflate interval and there
1312 // are too many monitors in use. We don't deflate more frequently
1313 // than AsyncDeflationInterval (unless is_async_deflation_requested)
1314 // in order to not swamp the MonitorDeflationThread.
1315 log_info(monitorinflation)("Async deflation needed: monitors used are above the threshold");
1316 return true;
1317 }
1318
1319 if (GuaranteedAsyncDeflationInterval > 0 &&
1320 time_since_last > GuaranteedAsyncDeflationInterval) {
1321 // It's been longer than our specified guaranteed deflate interval.
1322 // We need to clean up the used monitors even if the threshold is
1323 // not reached, to keep the memory utilization at bay when many threads
1324 // touched many monitors.
1325 log_info(monitorinflation)("Async deflation needed: guaranteed interval (" INTX_FORMAT " ms) "
1326 "is greater than time since last deflation (" JLONG_FORMAT " ms)",
1327 GuaranteedAsyncDeflationInterval, time_since_last);
1328
1329 // If this deflation has no progress, then it should not affect the no-progress
1330 // tracking, otherwise threshold heuristics would think it was triggered, experienced
1331 // no progress, and needs to backoff more aggressively. In this "no progress" case,
1332 // the generic code would bump the no-progress counter, and we compensate for that
1333 // by telling it to skip the update.
1334 //
1335 // If this deflation has progress, then it should let non-progress tracking
1336 // know about this, otherwise the threshold heuristics would kick in, potentially
1337 // experience no-progress due to aggressive cleanup by this deflation, and think
1338 // it is still in no-progress stride. In this "progress" case, the generic code would
1339 // zero the counter, and we allow it to happen.
1340 _no_progress_skip_increment = true;
1341
1342 return true;
1343 }
1344
1345 return false;
1346 }
1347
1348 void ObjectSynchronizer::request_deflate_idle_monitors() {
1349 MonitorLocker ml(MonitorDeflation_lock, Mutex::_no_safepoint_check_flag);
1350 set_is_async_deflation_requested(true);
1351 ml.notify_all();
1352 }
1353
1354 bool ObjectSynchronizer::request_deflate_idle_monitors_from_wb() {
1355 JavaThread* current = JavaThread::current();
1356 bool ret_code = false;
1357
1358 jlong last_time = last_async_deflation_time_ns();
1359
1360 request_deflate_idle_monitors();
1361
1362 const int N_CHECKS = 5;
1363 for (int i = 0; i < N_CHECKS; i++) { // sleep for at most 5 seconds
1364 if (last_async_deflation_time_ns() > last_time) {
1365 log_info(monitorinflation)("Async Deflation happened after %d check(s).", i);
1366 ret_code = true;
1367 break;
1368 }
1369 {
1370 // JavaThread has to honor the blocking protocol.
1371 ThreadBlockInVM tbivm(current);
1372 os::naked_short_sleep(999); // sleep for almost 1 second
1373 }
1374 }
1375 if (!ret_code) {
1376 log_info(monitorinflation)("Async Deflation DID NOT happen after %d checks.", N_CHECKS);
1377 }
1378
1379 return ret_code;
1380 }
1381
1382 jlong ObjectSynchronizer::time_since_last_async_deflation_ms() {
1383 return (os::javaTimeNanos() - last_async_deflation_time_ns()) / (NANOUNITS / MILLIUNITS);
1384 }
1385
1386 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1387 const oop obj,
1388 ObjectSynchronizer::InflateCause cause) {
1389 assert(event != nullptr, "invariant");
1390 event->set_monitorClass(obj->klass());
1391 event->set_address((uintptr_t)(void*)obj);
1392 event->set_cause((u1)cause);
1393 event->commit();
1394 }
1395
1396 // Fast path code shared by multiple functions
1397 void ObjectSynchronizer::inflate_helper(oop obj) {
1398 markWord mark = obj->mark_acquire();
1399 if (mark.has_monitor()) {
1400 ObjectMonitor* monitor = mark.monitor();
1401 markWord dmw = monitor->header();
1402 assert(dmw.is_neutral(), "sanity check: header=" INTPTR_FORMAT, dmw.value());
1403 return;
1404 }
1405 (void)inflate(Thread::current(), obj, inflate_cause_vm_internal);
1406 }
1407
1408 ObjectMonitor* ObjectSynchronizer::inflate(Thread* current, oop obj, const InflateCause cause) {
1409 assert(current == Thread::current(), "must be");
1410 return inflate_impl(current->is_Java_thread() ? JavaThread::cast(current) : nullptr, obj, cause);
1411 }
1412
1413 ObjectMonitor* ObjectSynchronizer::inflate_for(JavaThread* thread, oop obj, const InflateCause cause) {
1414 assert(thread == Thread::current() || thread->is_obj_deopt_suspend(), "must be");
1415 return inflate_impl(thread, obj, cause);
1416 }
1417
1418 ObjectMonitor* ObjectSynchronizer::inflate_impl(JavaThread* inflating_thread, oop object, const InflateCause cause) {
1419 // The JavaThread* inflating_thread requires that the inflating_thread == Thread::current() or
1420 // is suspended throughout the call by some other mechanism.
1421 // The thread might be nullptr when called from a non JavaThread. (As may still be
1422 // the case from FastHashCode). However it is only important for correctness that the
1423 // thread is set when called from ObjectSynchronizer::enter from the owning thread,
1424 // ObjectSynchronizer::enter_for from any thread, or ObjectSynchronizer::exit.
1425 EventJavaMonitorInflate event;
1426
1427 for (;;) {
1428 const markWord mark = object->mark_acquire();
1429
1430 // The mark can be in one of the following states:
1431 // * inflated - If the ObjectMonitor owner is anonymous and the
1432 // inflating_thread owns the object lock, then we
1433 // make the inflating_thread the ObjectMonitor owner.
1434 // For LM_LIGHTWEIGHT we also remove the lock from
1435 // the inflating_thread's lock stack.
1436 // * fast-locked - Coerce it to inflated from fast-locked.
1437 // * stack-locked - Coerce it to inflated from stack-locked.
1438 // * INFLATING - Busy wait for conversion from stack-locked to
1439 // inflated.
1440 // * unlocked - Aggressively inflate the object.
1441
1442 // CASE: inflated
1443 if (mark.has_monitor()) {
1444 ObjectMonitor* inf = mark.monitor();
1445 markWord dmw = inf->header();
1446 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1447 if (inf->is_owner_anonymous() && inflating_thread != nullptr) {
1448 if (LockingMode == LM_LIGHTWEIGHT) {
1449 if (inflating_thread->lock_stack().contains(object)) {
1450 inf->set_owner_from_anonymous(inflating_thread);
1451 size_t removed = inflating_thread->lock_stack().remove(object);
1452 inf->set_recursions(removed - 1);
1453 }
1454 } else {
1455 assert(LockingMode == LM_LEGACY, "invariant");
1456 if (inflating_thread->is_lock_owned((address)inf->stack_locker())) {
1457 inf->set_owner_from_BasicLock(inflating_thread);
1458 // Decrement monitor count now since this monitor is okay for freezing
1459 LOOM_MONITOR_SUPPORT_ONLY(inflating_thread->dec_held_monitor_count();)
1460 }
1461 }
1462 }
1463 return inf;
1464 }
1465
1466 if (LockingMode != LM_LIGHTWEIGHT) {
1467 // New lightweight locking does not use INFLATING.
1468 // CASE: inflation in progress - inflating over a stack-lock.
1469 // Some other thread is converting from stack-locked to inflated.
1470 // Only that thread can complete inflation -- other threads must wait.
1471 // The INFLATING value is transient.
1472 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1473 // We could always eliminate polling by parking the thread on some auxiliary list.
1474 if (mark == markWord::INFLATING()) {
1475 read_stable_mark(object);
1476 continue;
1477 }
1478 }
1479
1480 // CASE: fast-locked
1481 // Could be fast-locked either by the inflating_thread or by some other thread.
1482 //
1483 // Note that we allocate the ObjectMonitor speculatively, _before_
1484 // attempting to set the object's mark to the new ObjectMonitor. If
1485 // the inflating_thread owns the monitor, then we set the ObjectMonitor's
1486 // owner to the inflating_thread. Otherwise, we set the ObjectMonitor's owner
1487 // to anonymous. If we lose the race to set the object's mark to the
1488 // new ObjectMonitor, then we just delete it and loop around again.
1489 //
1490 LogStreamHandle(Trace, monitorinflation) lsh;
1491 if (LockingMode == LM_LIGHTWEIGHT && mark.is_fast_locked()) {
1492 ObjectMonitor* monitor = new ObjectMonitor(object);
1493 monitor->set_header(mark.set_unlocked());
1494 bool own = inflating_thread != nullptr && inflating_thread->lock_stack().contains(object);
1495 if (own) {
1496 // Owned by inflating_thread.
1497 monitor->set_owner_from(nullptr, inflating_thread);
1498 } else {
1499 // Owned by somebody else.
1500 monitor->set_owner_anonymous();
1501 }
1502 markWord monitor_mark = markWord::encode(monitor);
1503 markWord old_mark = object->cas_set_mark(monitor_mark, mark);
1504 if (old_mark == mark) {
1505 // Success! Return inflated monitor.
1506 if (own) {
1507 size_t removed = inflating_thread->lock_stack().remove(object);
1508 monitor->set_recursions(removed - 1);
1509 }
1510 // Once the ObjectMonitor is configured and object is associated
1511 // with the ObjectMonitor, it is safe to allow async deflation:
1512 _in_use_list.add(monitor);
1513
1514 // Hopefully the performance counters are allocated on distinct
1515 // cache lines to avoid false sharing on MP systems ...
1516 OM_PERFDATA_OP(Inflations, inc());
1517 if (log_is_enabled(Trace, monitorinflation)) {
1518 ResourceMark rm;
1519 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1520 INTPTR_FORMAT ", type='%s'", p2i(object),
1521 object->mark().value(), object->klass()->external_name());
1522 }
1523 if (event.should_commit()) {
1524 post_monitor_inflate_event(&event, object, cause);
1525 }
1526 return monitor;
1527 } else {
1528 delete monitor;
1529 continue; // Interference -- just retry
1530 }
1531 }
1532
1533 // CASE: stack-locked
1534 // Could be stack-locked either by current or by some other thread.
1535 //
1536 // Note that we allocate the ObjectMonitor speculatively, _before_ attempting
1537 // to install INFLATING into the mark word. We originally installed INFLATING,
1538 // allocated the ObjectMonitor, and then finally STed the address of the
1539 // ObjectMonitor into the mark. This was correct, but artificially lengthened
1540 // the interval in which INFLATING appeared in the mark, thus increasing
1541 // the odds of inflation contention. If we lose the race to set INFLATING,
1542 // then we just delete the ObjectMonitor and loop around again.
1543 //
1544 if (LockingMode == LM_LEGACY && mark.has_locker()) {
1545 assert(LockingMode != LM_LIGHTWEIGHT, "cannot happen with new lightweight locking");
1546 ObjectMonitor* m = new ObjectMonitor(object);
1547 // Optimistically prepare the ObjectMonitor - anticipate successful CAS
1548 // We do this before the CAS in order to minimize the length of time
1549 // in which INFLATING appears in the mark.
1550
1551 markWord cmp = object->cas_set_mark(markWord::INFLATING(), mark);
1552 if (cmp != mark) {
1553 delete m;
1554 continue; // Interference -- just retry
1555 }
1556
1557 // We've successfully installed INFLATING (0) into the mark-word.
1558 // This is the only case where 0 will appear in a mark-word.
1559 // Only the singular thread that successfully swings the mark-word
1560 // to 0 can perform (or more precisely, complete) inflation.
1561 //
1562 // Why do we CAS a 0 into the mark-word instead of just CASing the
1563 // mark-word from the stack-locked value directly to the new inflated state?
1564 // Consider what happens when a thread unlocks a stack-locked object.
1565 // It attempts to use CAS to swing the displaced header value from the
1566 // on-stack BasicLock back into the object header. Recall also that the
1567 // header value (hash code, etc) can reside in (a) the object header, or
1568 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1569 // header in an ObjectMonitor. The inflate() routine must copy the header
1570 // value from the BasicLock on the owner's stack to the ObjectMonitor, all
1571 // the while preserving the hashCode stability invariants. If the owner
1572 // decides to release the lock while the value is 0, the unlock will fail
1573 // and control will eventually pass from slow_exit() to inflate. The owner
1574 // will then spin, waiting for the 0 value to disappear. Put another way,
1575 // the 0 causes the owner to stall if the owner happens to try to
1576 // drop the lock (restoring the header from the BasicLock to the object)
1577 // while inflation is in-progress. This protocol avoids races that might
1578 // would otherwise permit hashCode values to change or "flicker" for an object.
1579 // Critically, while object->mark is 0 mark.displaced_mark_helper() is stable.
1580 // 0 serves as a "BUSY" inflate-in-progress indicator.
1581
1582
1583 // fetch the displaced mark from the owner's stack.
1584 // The owner can't die or unwind past the lock while our INFLATING
1585 // object is in the mark. Furthermore the owner can't complete
1586 // an unlock on the object, either.
1587 markWord dmw = mark.displaced_mark_helper();
1588 // Catch if the object's header is not neutral (not locked and
1589 // not marked is what we care about here).
1590 assert(dmw.is_neutral(), "invariant: header=" INTPTR_FORMAT, dmw.value());
1591
1592 // Setup monitor fields to proper values -- prepare the monitor
1593 m->set_header(dmw);
1594
1595 // Note that a thread can inflate an object
1596 // that it has stack-locked -- as might happen in wait() -- directly
1597 // with CAS. That is, we can avoid the xchg-nullptr .... ST idiom.
1598 if (inflating_thread != nullptr && inflating_thread->is_lock_owned((address)mark.locker())) {
1599 m->set_owner_from(nullptr, inflating_thread);
1600 // Decrement monitor count now since this monitor is okay for freezing
1601 LOOM_MONITOR_SUPPORT_ONLY(inflating_thread->dec_held_monitor_count();)
1602 } else {
1603 // Use ANONYMOUS_OWNER to indicate that the owner is the BasicLock on the stack,
1604 // and set the stack locker field in the monitor.
1605 m->set_stack_locker(mark.locker());
1606 m->set_owner_anonymous(); // second
1607 }
1608 // TODO-FIXME: assert BasicLock->dhw != 0.
1609
1610 // Must preserve store ordering. The monitor state must
1611 // be stable at the time of publishing the monitor address.
1612 guarantee(object->mark() == markWord::INFLATING(), "invariant");
1613 // Release semantics so that above set_object() is seen first.
1614 object->release_set_mark(markWord::encode(m));
1615
1616 // Once ObjectMonitor is configured and the object is associated
1617 // with the ObjectMonitor, it is safe to allow async deflation:
1618 _in_use_list.add(m);
1619
1620 // Hopefully the performance counters are allocated on distinct cache lines
1621 // to avoid false sharing on MP systems ...
1622 OM_PERFDATA_OP(Inflations, inc());
1623 if (log_is_enabled(Trace, monitorinflation)) {
1624 ResourceMark rm;
1625 lsh.print_cr("inflate(has_locker): object=" INTPTR_FORMAT ", mark="
1626 INTPTR_FORMAT ", type='%s'", p2i(object),
1627 object->mark().value(), object->klass()->external_name());
1628 }
1629 if (event.should_commit()) {
1630 post_monitor_inflate_event(&event, object, cause);
1631 }
1632 return m;
1633 }
1634
1635 // CASE: unlocked
1636 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1637 // If we know we're inflating for entry it's better to inflate by swinging a
1638 // pre-locked ObjectMonitor pointer into the object header. A successful
1639 // CAS inflates the object *and* confers ownership to the inflating thread.
1640 // In the current implementation we use a 2-step mechanism where we CAS()
1641 // to inflate and then CAS() again to try to swing _owner from null to current.
1642 // An inflateTry() method that we could call from enter() would be useful.
1643
1644 assert(mark.is_unlocked(), "invariant: header=" INTPTR_FORMAT, mark.value());
1645 ObjectMonitor* m = new ObjectMonitor(object);
1646 // prepare m for installation - set monitor to initial state
1647 m->set_header(mark);
1648
1649 if (object->cas_set_mark(markWord::encode(m), mark) != mark) {
1650 delete m;
1651 m = nullptr;
1652 continue;
1653 // interference - the markword changed - just retry.
1654 // The state-transitions are one-way, so there's no chance of
1655 // live-lock -- "Inflated" is an absorbing state.
1656 }
1657
1658 // Once the ObjectMonitor is configured and object is associated
1659 // with the ObjectMonitor, it is safe to allow async deflation:
1660 _in_use_list.add(m);
1661
1662 // Hopefully the performance counters are allocated on distinct
1663 // cache lines to avoid false sharing on MP systems ...
1664 OM_PERFDATA_OP(Inflations, inc());
1665 if (log_is_enabled(Trace, monitorinflation)) {
1666 ResourceMark rm;
1667 lsh.print_cr("inflate(unlocked): object=" INTPTR_FORMAT ", mark="
1668 INTPTR_FORMAT ", type='%s'", p2i(object),
1669 object->mark().value(), object->klass()->external_name());
1670 }
1671 if (event.should_commit()) {
1672 post_monitor_inflate_event(&event, object, cause);
1673 }
1674 return m;
1675 }
1676 }
1677
1678 // Walk the in-use list and deflate (at most MonitorDeflationMax) idle
1679 // ObjectMonitors. Returns the number of deflated ObjectMonitors.
1680 //
1681 size_t ObjectSynchronizer::deflate_monitor_list(ObjectMonitorDeflationSafepointer* safepointer) {
1682 MonitorList::Iterator iter = _in_use_list.iterator();
1683 size_t deflated_count = 0;
1684
1685 while (iter.has_next()) {
1686 if (deflated_count >= (size_t)MonitorDeflationMax) {
1687 break;
1688 }
1689 ObjectMonitor* mid = iter.next();
1690 if (mid->deflate_monitor()) {
1691 deflated_count++;
1692 }
1693
1694 // Must check for a safepoint/handshake and honor it.
1695 safepointer->block_for_safepoint("deflation", "deflated_count", deflated_count);
1696 }
1697
1698 return deflated_count;
1699 }
1700
1701 class HandshakeForDeflation : public HandshakeClosure {
1702 public:
1703 HandshakeForDeflation() : HandshakeClosure("HandshakeForDeflation") {}
1704
1705 void do_thread(Thread* thread) {
1706 log_trace(monitorinflation)("HandshakeForDeflation::do_thread: thread="
1707 INTPTR_FORMAT, p2i(thread));
1708 }
1709 };
1710
1711 class VM_RendezvousGCThreads : public VM_Operation {
1712 public:
1713 bool evaluate_at_safepoint() const override { return false; }
1714 VMOp_Type type() const override { return VMOp_RendezvousGCThreads; }
1715 void doit() override {
1716 Universe::heap()->safepoint_synchronize_begin();
1717 Universe::heap()->safepoint_synchronize_end();
1718 };
1719 };
1720
1721 static size_t delete_monitors(GrowableArray<ObjectMonitor*>* delete_list,
1722 ObjectMonitorDeflationSafepointer* safepointer) {
1723 NativeHeapTrimmer::SuspendMark sm("monitor deletion");
1724 size_t deleted_count = 0;
1725 for (ObjectMonitor* monitor: *delete_list) {
1726 delete monitor;
1727 deleted_count++;
1728 // A JavaThread must check for a safepoint/handshake and honor it.
1729 safepointer->block_for_safepoint("deletion", "deleted_count", deleted_count);
1730 }
1731 return deleted_count;
1732 }
1733
1734 class ObjectMonitorDeflationLogging: public StackObj {
1735 LogStreamHandle(Debug, monitorinflation) _debug;
1736 LogStreamHandle(Info, monitorinflation) _info;
1737 LogStream* _stream;
1738 elapsedTimer _timer;
1739
1740 size_t ceiling() const { return ObjectSynchronizer::in_use_list_ceiling(); }
1741 size_t count() const { return ObjectSynchronizer::_in_use_list.count(); }
1742 size_t max() const { return ObjectSynchronizer::_in_use_list.max(); }
1743
1744 public:
1745 ObjectMonitorDeflationLogging()
1746 : _debug(), _info(), _stream(nullptr) {
1747 if (_debug.is_enabled()) {
1748 _stream = &_debug;
1749 } else if (_info.is_enabled()) {
1750 _stream = &_info;
1751 }
1752 }
1753
1754 void begin() {
1755 if (_stream != nullptr) {
1756 _stream->print_cr("begin deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1757 ceiling(), count(), max());
1758 _timer.start();
1759 }
1760 }
1761
1762 void before_handshake(size_t unlinked_count) {
1763 if (_stream != nullptr) {
1764 _timer.stop();
1765 _stream->print_cr("before handshaking: unlinked_count=" SIZE_FORMAT
1766 ", in_use_list stats: ceiling=" SIZE_FORMAT ", count="
1767 SIZE_FORMAT ", max=" SIZE_FORMAT,
1768 unlinked_count, ceiling(), count(), max());
1769 }
1770 }
1771
1772 void after_handshake() {
1773 if (_stream != nullptr) {
1774 _stream->print_cr("after handshaking: in_use_list stats: ceiling="
1775 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1776 ceiling(), count(), max());
1777 _timer.start();
1778 }
1779 }
1780
1781 void end(size_t deflated_count, size_t unlinked_count) {
1782 if (_stream != nullptr) {
1783 _timer.stop();
1784 if (deflated_count != 0 || unlinked_count != 0 || _debug.is_enabled()) {
1785 _stream->print_cr("deflated_count=" SIZE_FORMAT ", {unlinked,deleted}_count=" SIZE_FORMAT " monitors in %3.7f secs",
1786 deflated_count, unlinked_count, _timer.seconds());
1787 }
1788 _stream->print_cr("end deflating: in_use_list stats: ceiling=" SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1789 ceiling(), count(), max());
1790 }
1791 }
1792
1793 void before_block_for_safepoint(const char* op_name, const char* cnt_name, size_t cnt) {
1794 if (_stream != nullptr) {
1795 _timer.stop();
1796 _stream->print_cr("pausing %s: %s=" SIZE_FORMAT ", in_use_list stats: ceiling="
1797 SIZE_FORMAT ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT,
1798 op_name, cnt_name, cnt, ceiling(), count(), max());
1799 }
1800 }
1801
1802 void after_block_for_safepoint(const char* op_name) {
1803 if (_stream != nullptr) {
1804 _stream->print_cr("resuming %s: in_use_list stats: ceiling=" SIZE_FORMAT
1805 ", count=" SIZE_FORMAT ", max=" SIZE_FORMAT, op_name,
1806 ceiling(), count(), max());
1807 _timer.start();
1808 }
1809 }
1810 };
1811
1812 void ObjectMonitorDeflationSafepointer::block_for_safepoint(const char* op_name, const char* count_name, size_t counter) {
1813 if (!SafepointMechanism::should_process(_current)) {
1814 return;
1815 }
1816
1817 // A safepoint/handshake has started.
1818 _log->before_block_for_safepoint(op_name, count_name, counter);
1819
1820 {
1821 // Honor block request.
1822 ThreadBlockInVM tbivm(_current);
1823 }
1824
1825 _log->after_block_for_safepoint(op_name);
1826 }
1827
1828 // This function is called by the MonitorDeflationThread to deflate
1829 // ObjectMonitors.
1830 size_t ObjectSynchronizer::deflate_idle_monitors() {
1831 JavaThread* current = JavaThread::current();
1832 assert(current->is_monitor_deflation_thread(), "The only monitor deflater");
1833
1834 // The async deflation request has been processed.
1835 _last_async_deflation_time_ns = os::javaTimeNanos();
1836 set_is_async_deflation_requested(false);
1837
1838 ObjectMonitorDeflationLogging log;
1839 ObjectMonitorDeflationSafepointer safepointer(current, &log);
1840
1841 log.begin();
1842
1843 // Deflate some idle ObjectMonitors.
1844 size_t deflated_count = deflate_monitor_list(&safepointer);
1845
1846 // Unlink the deflated ObjectMonitors from the in-use list.
1847 size_t unlinked_count = 0;
1848 size_t deleted_count = 0;
1849 if (deflated_count > 0) {
1850 ResourceMark rm(current);
1851 GrowableArray<ObjectMonitor*> delete_list((int)deflated_count);
1852 unlinked_count = _in_use_list.unlink_deflated(deflated_count, &delete_list, &safepointer);
1853
1854 log.before_handshake(unlinked_count);
1855
1856 // A JavaThread needs to handshake in order to safely free the
1857 // ObjectMonitors that were deflated in this cycle.
1858 HandshakeForDeflation hfd_hc;
1859 Handshake::execute(&hfd_hc);
1860 // Also, we sync and desync GC threads around the handshake, so that they can
1861 // safely read the mark-word and look-through to the object-monitor, without
1862 // being afraid that the object-monitor is going away.
1863 VM_RendezvousGCThreads sync_gc;
1864 VMThread::execute(&sync_gc);
1865
1866 log.after_handshake();
1867
1868 // After the handshake, safely free the ObjectMonitors that were
1869 // deflated and unlinked in this cycle.
1870
1871 // Delete the unlinked ObjectMonitors.
1872 deleted_count = delete_monitors(&delete_list, &safepointer);
1873 assert(unlinked_count == deleted_count, "must be");
1874 }
1875
1876 log.end(deflated_count, unlinked_count);
1877
1878 OM_PERFDATA_OP(MonExtant, set_value(_in_use_list.count()));
1879 OM_PERFDATA_OP(Deflations, inc(deflated_count));
1880
1881 GVars.stw_random = os::random();
1882
1883 if (deflated_count != 0) {
1884 _no_progress_cnt = 0;
1885 } else if (_no_progress_skip_increment) {
1886 _no_progress_skip_increment = false;
1887 } else {
1888 _no_progress_cnt++;
1889 }
1890
1891 return deflated_count;
1892 }
1893
1894 // Monitor cleanup on JavaThread::exit
1895
1896 // Iterate through monitor cache and attempt to release thread's monitors
1897 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1898 private:
1899 JavaThread* _thread;
1900
1901 public:
1902 ReleaseJavaMonitorsClosure(JavaThread* thread) : _thread(thread) {}
1903 void do_monitor(ObjectMonitor* mid) {
1904 intx rec = mid->complete_exit(_thread);
1905 _thread->dec_held_monitor_count(NOT_LOOM_MONITOR_SUPPORT((rec + 1)));
1906 }
1907 };
1908
1909 // Release all inflated monitors owned by current thread. Lightweight monitors are
1910 // ignored. This is meant to be called during JNI thread detach which assumes
1911 // all remaining monitors are heavyweight. All exceptions are swallowed.
1912 // Scanning the extant monitor list can be time consuming.
1913 // A simple optimization is to add a per-thread flag that indicates a thread
1914 // called jni_monitorenter() during its lifetime.
1915 //
1916 // Instead of NoSafepointVerifier it might be cheaper to
1917 // use an idiom of the form:
1918 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1919 // <code that must not run at safepoint>
1920 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1921 // Since the tests are extremely cheap we could leave them enabled
1922 // for normal product builds.
1923
1924 void ObjectSynchronizer::release_monitors_owned_by_thread(JavaThread* current) {
1925 assert(current == JavaThread::current(), "must be current Java thread");
1926 NoSafepointVerifier nsv;
1927 ReleaseJavaMonitorsClosure rjmc(current);
1928 ObjectSynchronizer::owned_monitors_iterate(&rjmc, current);
1929 assert(!current->has_pending_exception(), "Should not be possible");
1930 current->clear_pending_exception();
1931 assert(current->held_monitor_count() == 0, "Should not be possible");
1932 // All monitors (including entered via JNI) have been unlocked above, so we need to clear jni count.
1933 current->clear_jni_monitor_count();
1934 }
1935
1936 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1937 switch (cause) {
1938 case inflate_cause_vm_internal: return "VM Internal";
1939 case inflate_cause_monitor_enter: return "Monitor Enter";
1940 case inflate_cause_wait: return "Monitor Wait";
1941 case inflate_cause_notify: return "Monitor Notify";
1942 case inflate_cause_hash_code: return "Monitor Hash Code";
1943 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1944 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1945 case inflate_cause_cont_freeze: return "Continuation Freeze";
1946 default:
1947 ShouldNotReachHere();
1948 }
1949 return "Unknown";
1950 }
1951
1952 //------------------------------------------------------------------------------
1953 // Debugging code
1954
1955 u_char* ObjectSynchronizer::get_gvars_addr() {
1956 return (u_char*)&GVars;
1957 }
1958
1959 u_char* ObjectSynchronizer::get_gvars_hc_sequence_addr() {
1960 return (u_char*)&GVars.hc_sequence;
1961 }
1962
1963 size_t ObjectSynchronizer::get_gvars_size() {
1964 return sizeof(SharedGlobals);
1965 }
1966
1967 u_char* ObjectSynchronizer::get_gvars_stw_random_addr() {
1968 return (u_char*)&GVars.stw_random;
1969 }
1970
1971 // Do the final audit and print of ObjectMonitor stats; must be done
1972 // by the VMThread at VM exit time.
1973 void ObjectSynchronizer::do_final_audit_and_print_stats() {
1974 assert(Thread::current()->is_VM_thread(), "sanity check");
1975
1976 if (is_final_audit()) { // Only do the audit once.
1977 return;
1978 }
1979 set_is_final_audit();
1980 log_info(monitorinflation)("Starting the final audit.");
1981
1982 if (log_is_enabled(Info, monitorinflation)) {
1983 LogStreamHandle(Info, monitorinflation) ls;
1984 audit_and_print_stats(&ls, true /* on_exit */);
1985 }
1986 }
1987
1988 // This function can be called by the MonitorDeflationThread or it can be called when
1989 // we are trying to exit the VM. The list walker functions can run in parallel with
1990 // the other list operations.
1991 // Calls to this function can be added in various places as a debugging
1992 // aid.
1993 //
1994 void ObjectSynchronizer::audit_and_print_stats(outputStream* ls, bool on_exit) {
1995 int error_cnt = 0;
1996
1997 ls->print_cr("Checking in_use_list:");
1998 chk_in_use_list(ls, &error_cnt);
1999
2000 if (error_cnt == 0) {
2001 ls->print_cr("No errors found in in_use_list checks.");
2002 } else {
2003 log_error(monitorinflation)("found in_use_list errors: error_cnt=%d", error_cnt);
2004 }
2005
2006 // When exiting, only log the interesting entries at the Info level.
2007 // When called at intervals by the MonitorDeflationThread, log output
2008 // at the Trace level since there can be a lot of it.
2009 if (!on_exit && log_is_enabled(Trace, monitorinflation)) {
2010 LogStreamHandle(Trace, monitorinflation) ls_tr;
2011 log_in_use_monitor_details(&ls_tr, true /* log_all */);
2012 } else if (on_exit) {
2013 log_in_use_monitor_details(ls, false /* log_all */);
2014 }
2015
2016 ls->flush();
2017
2018 guarantee(error_cnt == 0, "ERROR: found monitor list errors: error_cnt=%d", error_cnt);
2019 }
2020
2021 // Check the in_use_list; log the results of the checks.
2022 void ObjectSynchronizer::chk_in_use_list(outputStream* out, int *error_cnt_p) {
2023 size_t l_in_use_count = _in_use_list.count();
2024 size_t l_in_use_max = _in_use_list.max();
2025 out->print_cr("count=" SIZE_FORMAT ", max=" SIZE_FORMAT, l_in_use_count,
2026 l_in_use_max);
2027
2028 size_t ck_in_use_count = 0;
2029 MonitorList::Iterator iter = _in_use_list.iterator();
2030 while (iter.has_next()) {
2031 ObjectMonitor* mid = iter.next();
2032 chk_in_use_entry(mid, out, error_cnt_p);
2033 ck_in_use_count++;
2034 }
2035
2036 if (l_in_use_count == ck_in_use_count) {
2037 out->print_cr("in_use_count=" SIZE_FORMAT " equals ck_in_use_count="
2038 SIZE_FORMAT, l_in_use_count, ck_in_use_count);
2039 } else {
2040 out->print_cr("WARNING: in_use_count=" SIZE_FORMAT " is not equal to "
2041 "ck_in_use_count=" SIZE_FORMAT, l_in_use_count,
2042 ck_in_use_count);
2043 }
2044
2045 size_t ck_in_use_max = _in_use_list.max();
2046 if (l_in_use_max == ck_in_use_max) {
2047 out->print_cr("in_use_max=" SIZE_FORMAT " equals ck_in_use_max="
2048 SIZE_FORMAT, l_in_use_max, ck_in_use_max);
2049 } else {
2050 out->print_cr("WARNING: in_use_max=" SIZE_FORMAT " is not equal to "
2051 "ck_in_use_max=" SIZE_FORMAT, l_in_use_max, ck_in_use_max);
2052 }
2053 }
2054
2055 // Check an in-use monitor entry; log any errors.
2056 void ObjectSynchronizer::chk_in_use_entry(ObjectMonitor* n, outputStream* out,
2057 int* error_cnt_p) {
2058 if (n->owner_is_DEFLATER_MARKER()) {
2059 // This could happen when monitor deflation blocks for a safepoint.
2060 return;
2061 }
2062
2063 if (n->header().value() == 0) {
2064 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor must "
2065 "have non-null _header field.", p2i(n));
2066 *error_cnt_p = *error_cnt_p + 1;
2067 }
2068 const oop obj = n->object_peek();
2069 if (obj != nullptr) {
2070 const markWord mark = obj->mark();
2071 if (!mark.has_monitor()) {
2072 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2073 "object does not think it has a monitor: obj="
2074 INTPTR_FORMAT ", mark=" INTPTR_FORMAT, p2i(n),
2075 p2i(obj), mark.value());
2076 *error_cnt_p = *error_cnt_p + 1;
2077 }
2078 ObjectMonitor* const obj_mon = mark.monitor();
2079 if (n != obj_mon) {
2080 out->print_cr("ERROR: monitor=" INTPTR_FORMAT ": in-use monitor's "
2081 "object does not refer to the same monitor: obj="
2082 INTPTR_FORMAT ", mark=" INTPTR_FORMAT ", obj_mon="
2083 INTPTR_FORMAT, p2i(n), p2i(obj), mark.value(), p2i(obj_mon));
2084 *error_cnt_p = *error_cnt_p + 1;
2085 }
2086 }
2087 }
2088
2089 // Log details about ObjectMonitors on the in_use_list. The 'BHL'
2090 // flags indicate why the entry is in-use, 'object' and 'object type'
2091 // indicate the associated object and its type.
2092 void ObjectSynchronizer::log_in_use_monitor_details(outputStream* out, bool log_all) {
2093 if (_in_use_list.count() > 0) {
2094 stringStream ss;
2095 out->print_cr("In-use monitor info:");
2096 out->print_cr("(B -> is_busy, H -> has hash code, L -> lock status)");
2097 out->print_cr("%18s %s %18s %18s",
2098 "monitor", "BHL", "object", "object type");
2099 out->print_cr("================== === ================== ==================");
2100
2101 auto is_interesting = [&](ObjectMonitor* monitor) {
2102 return log_all || monitor->has_owner() || monitor->is_busy();
2103 };
2104
2105 monitors_iterate([&](ObjectMonitor* monitor) {
2106 if (is_interesting(monitor)) {
2107 const oop obj = monitor->object_peek();
2108 const markWord mark = monitor->header();
2109 ResourceMark rm;
2110 out->print(INTPTR_FORMAT " %d%d%d " INTPTR_FORMAT " %s", p2i(monitor),
2111 monitor->is_busy(), mark.hash() != 0, monitor->owner() != nullptr,
2112 p2i(obj), obj == nullptr ? "" : obj->klass()->external_name());
2113 if (monitor->is_busy()) {
2114 out->print(" (%s)", monitor->is_busy_to_string(&ss));
2115 ss.reset();
2116 }
2117 out->cr();
2118 }
2119 });
2120 }
2121
2122 out->flush();
2123 }