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