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/oopStorage.hpp"
28 #include "gc/shared/oopStorageSet.hpp"
29 #include "jfr/jfrEvents.hpp"
30 #include "jfr/support/jfrThreadId.hpp"
31 #include "logging/log.hpp"
32 #include "logging/logStream.hpp"
33 #include "memory/allocation.inline.hpp"
34 #include "memory/resourceArea.hpp"
35 #include "oops/markWord.hpp"
36 #include "oops/oop.inline.hpp"
37 #include "oops/oopHandle.inline.hpp"
38 #include "oops/weakHandle.inline.hpp"
39 #include "prims/jvmtiDeferredUpdates.hpp"
40 #include "prims/jvmtiExport.hpp"
41 #include "runtime/atomic.hpp"
42 #include "runtime/globals.hpp"
43 #include "runtime/handles.inline.hpp"
44 #include "runtime/interfaceSupport.inline.hpp"
45 #include "runtime/javaThread.inline.hpp"
46 #include "runtime/mutexLocker.hpp"
47 #include "runtime/objectMonitor.hpp"
48 #include "runtime/objectMonitor.inline.hpp"
49 #include "runtime/orderAccess.hpp"
50 #include "runtime/osThread.hpp"
51 #include "runtime/perfData.hpp"
52 #include "runtime/safefetch.hpp"
53 #include "runtime/safepointMechanism.inline.hpp"
54 #include "runtime/sharedRuntime.hpp"
55 #include "services/threadService.hpp"
56 #include "utilities/dtrace.hpp"
57 #include "utilities/globalDefinitions.hpp"
58 #include "utilities/macros.hpp"
59 #include "utilities/preserveException.hpp"
60 #if INCLUDE_JFR
61 #include "jfr/support/jfrFlush.hpp"
62 #endif
63
64 #ifdef DTRACE_ENABLED
65
66 // Only bother with this argument setup if dtrace is available
67 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
68
69
70 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
71 char* bytes = nullptr; \
72 int len = 0; \
73 jlong jtid = SharedRuntime::get_java_tid(thread); \
74 Symbol* klassname = obj->klass()->name(); \
75 if (klassname != nullptr) { \
76 bytes = (char*)klassname->bytes(); \
77 len = klassname->utf8_length(); \
78 }
79
80 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
81 { \
82 if (DTraceMonitorProbes) { \
83 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
84 HOTSPOT_MONITOR_WAIT(jtid, \
85 (monitor), bytes, len, (millis)); \
86 } \
87 }
88
89 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
90 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
91 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
92 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
93 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
94
95 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
96 { \
97 if (DTraceMonitorProbes) { \
98 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
99 HOTSPOT_MONITOR_##probe(jtid, \
100 (uintptr_t)(monitor), bytes, len); \
101 } \
102 }
103
104 #else // ndef DTRACE_ENABLED
105
106 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
107 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
108
109 #endif // ndef DTRACE_ENABLED
110
111 // Tunables ...
112 // The knob* variables are effectively final. Once set they should
113 // never be modified hence. Consider using __read_mostly with GCC.
114
115 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool -
116
117 static int Knob_Bonus = 100; // spin success bonus
118 static int Knob_BonusB = 100; // spin success bonus
119 static int Knob_Penalty = 200; // spin failure penalty
120 static int Knob_Poverty = 1000;
121 static int Knob_FixedSpin = 0;
122 static int Knob_PreSpin = 10; // 20-100 likely better
123
124 DEBUG_ONLY(static volatile bool InitDone = false;)
125
126 OopStorage* ObjectMonitor::_oop_storage = nullptr;
127
128 // -----------------------------------------------------------------------------
129 // Theory of operations -- Monitors lists, thread residency, etc:
130 //
131 // * A thread acquires ownership of a monitor by successfully
132 // CAS()ing the _owner field from null to non-null.
133 //
134 // * Invariant: A thread appears on at most one monitor list --
135 // cxq, EntryList or WaitSet -- at any one time.
136 //
137 // * Contending threads "push" themselves onto the cxq with CAS
138 // and then spin/park.
139 //
140 // * After a contending thread eventually acquires the lock it must
141 // dequeue itself from either the EntryList or the cxq.
142 //
143 // * The exiting thread identifies and unparks an "heir presumptive"
144 // tentative successor thread on the EntryList. Critically, the
145 // exiting thread doesn't unlink the successor thread from the EntryList.
146 // After having been unparked, the wakee will recontend for ownership of
147 // the monitor. The successor (wakee) will either acquire the lock or
148 // re-park itself.
149 //
150 // Succession is provided for by a policy of competitive handoff.
151 // The exiting thread does _not_ grant or pass ownership to the
152 // successor thread. (This is also referred to as "handoff" succession").
153 // Instead the exiting thread releases ownership and possibly wakes
154 // a successor, so the successor can (re)compete for ownership of the lock.
155 // If the EntryList is empty but the cxq is populated the exiting
156 // thread will drain the cxq into the EntryList. It does so by
157 // by detaching the cxq (installing null with CAS) and folding
158 // the threads from the cxq into the EntryList. The EntryList is
159 // doubly linked, while the cxq is singly linked because of the
160 // CAS-based "push" used to enqueue recently arrived threads (RATs).
161 //
162 // * Concurrency invariants:
163 //
164 // -- only the monitor owner may access or mutate the EntryList.
165 // The mutex property of the monitor itself protects the EntryList
166 // from concurrent interference.
167 // -- Only the monitor owner may detach the cxq.
168 //
169 // * The monitor entry list operations avoid locks, but strictly speaking
170 // they're not lock-free. Enter is lock-free, exit is not.
171 // For a description of 'Methods and apparatus providing non-blocking access
172 // to a resource,' see U.S. Pat. No. 7844973.
173 //
174 // * The cxq can have multiple concurrent "pushers" but only one concurrent
175 // detaching thread. This mechanism is immune from the ABA corruption.
176 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
177 //
178 // * Taken together, the cxq and the EntryList constitute or form a
179 // single logical queue of threads stalled trying to acquire the lock.
180 // We use two distinct lists to improve the odds of a constant-time
181 // dequeue operation after acquisition (in the ::enter() epilogue) and
182 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm).
183 // A key desideratum is to minimize queue & monitor metadata manipulation
184 // that occurs while holding the monitor lock -- that is, we want to
185 // minimize monitor lock holds times. Note that even a small amount of
186 // fixed spinning will greatly reduce the # of enqueue-dequeue operations
187 // on EntryList|cxq. That is, spinning relieves contention on the "inner"
188 // locks and monitor metadata.
189 //
190 // Cxq points to the set of Recently Arrived Threads attempting entry.
191 // Because we push threads onto _cxq with CAS, the RATs must take the form of
192 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when
193 // the unlocking thread notices that EntryList is null but _cxq is != null.
194 //
195 // The EntryList is ordered by the prevailing queue discipline and
196 // can be organized in any convenient fashion, such as a doubly-linked list or
197 // a circular doubly-linked list. Critically, we want insert and delete operations
198 // to operate in constant-time. If we need a priority queue then something akin
199 // to Solaris' sleepq would work nicely. Viz.,
200 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
201 // Queue discipline is enforced at ::exit() time, when the unlocking thread
202 // drains the cxq into the EntryList, and orders or reorders the threads on the
203 // EntryList accordingly.
204 //
205 // Barring "lock barging", this mechanism provides fair cyclic ordering,
206 // somewhat similar to an elevator-scan.
207 //
208 // * The monitor synchronization subsystem avoids the use of native
209 // synchronization primitives except for the narrow platform-specific
210 // park-unpark abstraction. See the comments in os_solaris.cpp regarding
211 // the semantics of park-unpark. Put another way, this monitor implementation
212 // depends only on atomic operations and park-unpark. The monitor subsystem
213 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
214 // underlying OS manages the READY<->RUN transitions.
215 //
216 // * Waiting threads reside on the WaitSet list -- wait() puts
217 // the caller onto the WaitSet.
218 //
219 // * notify() or notifyAll() simply transfers threads from the WaitSet to
220 // either the EntryList or cxq. Subsequent exit() operations will
221 // unpark the notifyee. Unparking a notifee in notify() is inefficient -
222 // it's likely the notifyee would simply impale itself on the lock held
223 // by the notifier.
224 //
225 // * An interesting alternative is to encode cxq as (List,LockByte) where
226 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary
227 // variable, like _recursions, in the scheme. The threads or Events that form
228 // the list would have to be aligned in 256-byte addresses. A thread would
229 // try to acquire the lock or enqueue itself with CAS, but exiting threads
230 // could use a 1-0 protocol and simply STB to set the LockByte to 0.
231 // Note that is is *not* word-tearing, but it does presume that full-word
232 // CAS operations are coherent with intermix with STB operations. That's true
233 // on most common processors.
234 //
235 // * See also http://blogs.sun.com/dave
236
237
238 // Check that object() and set_object() are called from the right context:
239 static void check_object_context() {
240 #ifdef ASSERT
241 Thread* self = Thread::current();
242 if (self->is_Java_thread()) {
243 // Mostly called from JavaThreads so sanity check the thread state.
244 JavaThread* jt = JavaThread::cast(self);
245 switch (jt->thread_state()) {
246 case _thread_in_vm: // the usual case
247 case _thread_in_Java: // during deopt
248 break;
249 default:
250 fatal("called from an unsafe thread state");
251 }
252 assert(jt->is_active_Java_thread(), "must be active JavaThread");
253 } else {
254 // However, ThreadService::get_current_contended_monitor()
255 // can call here via the VMThread so sanity check it.
256 assert(self->is_VM_thread(), "must be");
257 }
258 #endif // ASSERT
259 }
260
261 ObjectMonitor::ObjectMonitor(oop object) :
262 _header(markWord::zero()),
263 _object(_oop_storage, object),
264 _owner(nullptr),
265 _previous_owner_tid(0),
266 _next_om(nullptr),
267 _recursions(0),
268 _EntryList(nullptr),
269 _cxq(nullptr),
270 _succ(nullptr),
271 _Responsible(nullptr),
272 _SpinDuration(ObjectMonitor::Knob_SpinLimit),
273 _contentions(0),
274 _WaitSet(nullptr),
275 _waiters(0),
276 _WaitSetLock(0)
277 { }
278
279 ObjectMonitor::~ObjectMonitor() {
280 _object.release(_oop_storage);
281 }
282
283 oop ObjectMonitor::object() const {
284 check_object_context();
285 return _object.resolve();
286 }
287
288 oop ObjectMonitor::object_peek() const {
289 return _object.peek();
290 }
291
292 void ObjectMonitor::ExitOnSuspend::operator()(JavaThread* current) {
293 if (current->is_suspended()) {
294 _om->_recursions = 0;
295 _om->_succ = nullptr;
296 // Don't need a full fence after clearing successor here because of the call to exit().
297 _om->exit(current, false /* not_suspended */);
298 _om_exited = true;
299
300 current->set_current_pending_monitor(_om);
301 }
302 }
303
304 void ObjectMonitor::ClearSuccOnSuspend::operator()(JavaThread* current) {
305 if (current->is_suspended()) {
306 if (_om->_succ == current) {
307 _om->_succ = nullptr;
308 OrderAccess::fence(); // always do a full fence when successor is cleared
309 }
310 }
311 }
312
313 // -----------------------------------------------------------------------------
314 // Enter support
315
316 bool ObjectMonitor::enter_for(JavaThread* locking_thread) {
317 // Used by ObjectSynchronizer::enter_for to enter for another thread.
318 // The monitor is private to or already owned by locking_thread which must be suspended.
319 // So this code may only contend with deflation.
320 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
321
322 // Block out deflation as soon as possible.
323 add_to_contentions(1);
324
325 bool success = false;
326 if (!is_being_async_deflated()) {
327 void* prev_owner = try_set_owner_from(nullptr, locking_thread);
328
329 if (prev_owner == nullptr) {
330 assert(_recursions == 0, "invariant");
331 success = true;
332 } else if (prev_owner == locking_thread) {
333 _recursions++;
334 success = true;
335 } else if (prev_owner == DEFLATER_MARKER) {
336 // Racing with deflation.
337 prev_owner = try_set_owner_from(DEFLATER_MARKER, locking_thread);
338 if (prev_owner == DEFLATER_MARKER) {
339 // Cancelled deflation. Increment contentions as part of the deflation protocol.
340 add_to_contentions(1);
341 success = true;
342 } else if (prev_owner == nullptr) {
343 // At this point we cannot race with deflation as we have both incremented
344 // contentions, seen contention > 0 and seen a DEFLATER_MARKER.
345 // success will only be false if this races with something other than
346 // deflation.
347 prev_owner = try_set_owner_from(nullptr, locking_thread);
348 success = prev_owner == nullptr;
349 }
350 } else if (LockingMode == LM_LEGACY && locking_thread->is_lock_owned((address)prev_owner)) {
351 assert(_recursions == 0, "must be");
352 _recursions = 1;
353 set_owner_from_BasicLock(prev_owner, locking_thread);
354 success = true;
355 }
356 assert(success, "Failed to enter_for: locking_thread=" INTPTR_FORMAT
357 ", this=" INTPTR_FORMAT "{owner=" INTPTR_FORMAT "}, observed owner: " INTPTR_FORMAT,
358 p2i(locking_thread), p2i(this), p2i(owner_raw()), p2i(prev_owner));
359 } else {
360 // Async deflation is in progress and our contentions increment
361 // above lost the race to async deflation. Undo the work and
362 // force the caller to retry.
363 const oop l_object = object();
364 if (l_object != nullptr) {
365 // Attempt to restore the header/dmw to the object's header so that
366 // we only retry once if the deflater thread happens to be slow.
367 install_displaced_markword_in_object(l_object);
368 }
369 }
370
371 add_to_contentions(-1);
372
373 assert(!success || owner_raw() == locking_thread, "must be");
374
375 return success;
376 }
377
378 bool ObjectMonitor::enter(JavaThread* current) {
379 assert(current == JavaThread::current(), "must be");
380 // The following code is ordered to check the most common cases first
381 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
382
383 void* cur = try_set_owner_from(nullptr, current);
384 if (cur == nullptr) {
385 assert(_recursions == 0, "invariant");
386 return true;
387 }
388
389 if (cur == current) {
390 // TODO-FIXME: check for integer overflow! BUGID 6557169.
391 _recursions++;
392 return true;
393 }
394
395 if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
396 assert(_recursions == 0, "internal state error");
397 _recursions = 1;
398 set_owner_from_BasicLock(cur, current); // Convert from BasicLock* to Thread*.
399 return true;
400 }
401
402 // We've encountered genuine contention.
403
404 // Try one round of spinning *before* enqueueing current
405 // and before going through the awkward and expensive state
406 // transitions. The following spin is strictly optional ...
407 // Note that if we acquire the monitor from an initial spin
408 // we forgo posting JVMTI events and firing DTRACE probes.
409 if (TrySpin(current) > 0) {
410 assert(owner_raw() == current, "must be current: owner=" INTPTR_FORMAT, p2i(owner_raw()));
411 assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions);
412 assert(object()->mark() == markWord::encode(this),
413 "object mark must match encoded this: mark=" INTPTR_FORMAT
414 ", encoded this=" INTPTR_FORMAT, object()->mark().value(),
415 markWord::encode(this).value());
416 return true;
417 }
418
419 assert(owner_raw() != current, "invariant");
420 assert(_succ != current, "invariant");
421 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
422 assert(current->thread_state() != _thread_blocked, "invariant");
423
424 // Keep track of contention for JVM/TI and M&M queries.
425 add_to_contentions(1);
426 if (is_being_async_deflated()) {
427 // Async deflation is in progress and our contentions increment
428 // above lost the race to async deflation. Undo the work and
429 // force the caller to retry.
430 const oop l_object = object();
431 if (l_object != nullptr) {
432 // Attempt to restore the header/dmw to the object's header so that
433 // we only retry once if the deflater thread happens to be slow.
434 install_displaced_markword_in_object(l_object);
435 }
436 add_to_contentions(-1);
437 return false;
438 }
439
440 JFR_ONLY(JfrConditionalFlush<EventJavaMonitorEnter> flush(current);)
441 EventJavaMonitorEnter event;
442 if (event.is_started()) {
443 event.set_monitorClass(object()->klass());
444 // Set an address that is 'unique enough', such that events close in
445 // time and with the same address are likely (but not guaranteed) to
446 // belong to the same object.
447 event.set_address((uintptr_t)this);
448 }
449
450 { // Change java thread status to indicate blocked on monitor enter.
451 JavaThreadBlockedOnMonitorEnterState jtbmes(current, this);
452
453 assert(current->current_pending_monitor() == nullptr, "invariant");
454 current->set_current_pending_monitor(this);
455
456 DTRACE_MONITOR_PROBE(contended__enter, this, object(), current);
457 if (JvmtiExport::should_post_monitor_contended_enter()) {
458 JvmtiExport::post_monitor_contended_enter(current, this);
459
460 // The current thread does not yet own the monitor and does not
461 // yet appear on any queues that would get it made the successor.
462 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
463 // handler cannot accidentally consume an unpark() meant for the
464 // ParkEvent associated with this ObjectMonitor.
465 }
466
467 OSThreadContendState osts(current->osthread());
468
469 assert(current->thread_state() == _thread_in_vm, "invariant");
470
471 for (;;) {
472 ExitOnSuspend eos(this);
473 {
474 ThreadBlockInVMPreprocess<ExitOnSuspend> tbivs(current, eos, true /* allow_suspend */);
475 EnterI(current);
476 current->set_current_pending_monitor(nullptr);
477 // We can go to a safepoint at the end of this block. If we
478 // do a thread dump during that safepoint, then this thread will show
479 // as having "-locked" the monitor, but the OS and java.lang.Thread
480 // states will still report that the thread is blocked trying to
481 // acquire it.
482 // If there is a suspend request, ExitOnSuspend will exit the OM
483 // and set the OM as pending.
484 }
485 if (!eos.exited()) {
486 // ExitOnSuspend did not exit the OM
487 assert(owner_raw() == current, "invariant");
488 break;
489 }
490 }
491
492 // We've just gotten past the enter-check-for-suspend dance and we now own
493 // the monitor free and clear.
494 }
495
496 add_to_contentions(-1);
497 assert(contentions() >= 0, "must not be negative: contentions=%d", contentions());
498
499 // Must either set _recursions = 0 or ASSERT _recursions == 0.
500 assert(_recursions == 0, "invariant");
501 assert(owner_raw() == current, "invariant");
502 assert(_succ != current, "invariant");
503 assert(object()->mark() == markWord::encode(this), "invariant");
504
505 // The thread -- now the owner -- is back in vm mode.
506 // Report the glorious news via TI,DTrace and jvmstat.
507 // The probe effect is non-trivial. All the reportage occurs
508 // while we hold the monitor, increasing the length of the critical
509 // section. Amdahl's parallel speedup law comes vividly into play.
510 //
511 // Another option might be to aggregate the events (thread local or
512 // per-monitor aggregation) and defer reporting until a more opportune
513 // time -- such as next time some thread encounters contention but has
514 // yet to acquire the lock. While spinning that thread could
515 // spinning we could increment JVMStat counters, etc.
516
517 DTRACE_MONITOR_PROBE(contended__entered, this, object(), current);
518 if (JvmtiExport::should_post_monitor_contended_entered()) {
519 JvmtiExport::post_monitor_contended_entered(current, this);
520
521 // The current thread already owns the monitor and is not going to
522 // call park() for the remainder of the monitor enter protocol. So
523 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
524 // event handler consumed an unpark() issued by the thread that
525 // just exited the monitor.
526 }
527 if (event.should_commit()) {
528 event.set_previousOwner(_previous_owner_tid);
529 event.commit();
530 }
531 OM_PERFDATA_OP(ContendedLockAttempts, inc());
532 return true;
533 }
534
535 // Caveat: TryLock() is not necessarily serializing if it returns failure.
536 // Callers must compensate as needed.
537
538 int ObjectMonitor::TryLock(JavaThread* current) {
539 void* own = owner_raw();
540 if (own != nullptr) return 0;
541 if (try_set_owner_from(nullptr, current) == nullptr) {
542 assert(_recursions == 0, "invariant");
543 return 1;
544 }
545 // The lock had been free momentarily, but we lost the race to the lock.
546 // Interference -- the CAS failed.
547 // We can either return -1 or retry.
548 // Retry doesn't make as much sense because the lock was just acquired.
549 return -1;
550 }
551
552 // Deflate the specified ObjectMonitor if not in-use. Returns true if it
553 // was deflated and false otherwise.
554 //
555 // The async deflation protocol sets owner to DEFLATER_MARKER and
556 // makes contentions negative as signals to contending threads that
557 // an async deflation is in progress. There are a number of checks
558 // as part of the protocol to make sure that the calling thread has
559 // not lost the race to a contending thread.
560 //
561 // The ObjectMonitor has been successfully async deflated when:
562 // (contentions < 0)
563 // Contending threads that see that condition know to retry their operation.
564 //
565 bool ObjectMonitor::deflate_monitor() {
566 if (is_busy()) {
567 // Easy checks are first - the ObjectMonitor is busy so no deflation.
568 return false;
569 }
570
571 const oop obj = object_peek();
572
573 if (obj == nullptr) {
574 // If the object died, we can recycle the monitor without racing with
575 // Java threads. The GC already broke the association with the object.
576 set_owner_from(nullptr, DEFLATER_MARKER);
577 assert(contentions() >= 0, "must be non-negative: contentions=%d", contentions());
578 _contentions = INT_MIN; // minimum negative int
579 } else {
580 // Attempt async deflation protocol.
581
582 // Set a null owner to DEFLATER_MARKER to force any contending thread
583 // through the slow path. This is just the first part of the async
584 // deflation dance.
585 if (try_set_owner_from(nullptr, DEFLATER_MARKER) != nullptr) {
586 // The owner field is no longer null so we lost the race since the
587 // ObjectMonitor is now busy.
588 return false;
589 }
590
591 if (contentions() > 0 || _waiters != 0) {
592 // Another thread has raced to enter the ObjectMonitor after
593 // is_busy() above or has already entered and waited on
594 // it which makes it busy so no deflation. Restore owner to
595 // null if it is still DEFLATER_MARKER.
596 if (try_set_owner_from(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) {
597 // Deferred decrement for the JT EnterI() that cancelled the async deflation.
598 add_to_contentions(-1);
599 }
600 return false;
601 }
602
603 // Make a zero contentions field negative to force any contending threads
604 // to retry. This is the second part of the async deflation dance.
605 if (Atomic::cmpxchg(&_contentions, 0, INT_MIN) != 0) {
606 // Contentions was no longer 0 so we lost the race since the
607 // ObjectMonitor is now busy. Restore owner to null if it is
608 // still DEFLATER_MARKER:
609 if (try_set_owner_from(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) {
610 // Deferred decrement for the JT EnterI() that cancelled the async deflation.
611 add_to_contentions(-1);
612 }
613 return false;
614 }
615 }
616
617 // Sanity checks for the races:
618 guarantee(owner_is_DEFLATER_MARKER(), "must be deflater marker");
619 guarantee(contentions() < 0, "must be negative: contentions=%d",
620 contentions());
621 guarantee(_waiters == 0, "must be 0: waiters=%d", _waiters);
622 guarantee(_cxq == nullptr, "must be no contending threads: cxq="
623 INTPTR_FORMAT, p2i(_cxq));
624 guarantee(_EntryList == nullptr,
625 "must be no entering threads: EntryList=" INTPTR_FORMAT,
626 p2i(_EntryList));
627
628 if (obj != nullptr) {
629 if (log_is_enabled(Trace, monitorinflation)) {
630 ResourceMark rm;
631 log_trace(monitorinflation)("deflate_monitor: object=" INTPTR_FORMAT
632 ", mark=" INTPTR_FORMAT ", type='%s'",
633 p2i(obj), obj->mark().value(),
634 obj->klass()->external_name());
635 }
636
637 // Install the old mark word if nobody else has already done it.
638 install_displaced_markword_in_object(obj);
639 }
640
641 // We leave owner == DEFLATER_MARKER and contentions < 0
642 // to force any racing threads to retry.
643 return true; // Success, ObjectMonitor has been deflated.
644 }
645
646 // Install the displaced mark word (dmw) of a deflating ObjectMonitor
647 // into the header of the object associated with the monitor. This
648 // idempotent method is called by a thread that is deflating a
649 // monitor and by other threads that have detected a race with the
650 // deflation process.
651 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) {
652 // This function must only be called when (owner == DEFLATER_MARKER
653 // && contentions <= 0), but we can't guarantee that here because
654 // those values could change when the ObjectMonitor gets moved from
655 // the global free list to a per-thread free list.
656
657 guarantee(obj != nullptr, "must be non-null");
658
659 // Separate loads in is_being_async_deflated(), which is almost always
660 // called before this function, from the load of dmw/header below.
661
662 // _contentions and dmw/header may get written by different threads.
663 // Make sure to observe them in the same order when having several observers.
664 OrderAccess::loadload_for_IRIW();
665
666 const oop l_object = object_peek();
667 if (l_object == nullptr) {
668 // ObjectMonitor's object ref has already been cleared by async
669 // deflation or GC so we're done here.
670 return;
671 }
672 assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj="
673 INTPTR_FORMAT, p2i(l_object), p2i(obj));
674
675 markWord dmw = header();
676 // The dmw has to be neutral (not null, not locked and not marked).
677 assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value());
678
679 // Install displaced mark word if the object's header still points
680 // to this ObjectMonitor. More than one racing caller to this function
681 // can rarely reach this point, but only one can win.
682 markWord res = obj->cas_set_mark(dmw, markWord::encode(this));
683 if (res != markWord::encode(this)) {
684 // This should be rare so log at the Info level when it happens.
685 log_info(monitorinflation)("install_displaced_markword_in_object: "
686 "failed cas_set_mark: new_mark=" INTPTR_FORMAT
687 ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT,
688 dmw.value(), markWord::encode(this).value(),
689 res.value());
690 }
691
692 // Note: It does not matter which thread restored the header/dmw
693 // into the object's header. The thread deflating the monitor just
694 // wanted the object's header restored and it is. The threads that
695 // detected a race with the deflation process also wanted the
696 // object's header restored before they retry their operation and
697 // because it is restored they will only retry once.
698 }
699
700 // Convert the fields used by is_busy() to a string that can be
701 // used for diagnostic output.
702 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) {
703 ss->print("is_busy: waiters=%d"
704 ", contentions=%d"
705 ", owner=" PTR_FORMAT
706 ", cxq=" PTR_FORMAT
707 ", EntryList=" PTR_FORMAT,
708 _waiters,
709 (contentions() > 0 ? contentions() : 0),
710 owner_is_DEFLATER_MARKER()
711 // We report null instead of DEFLATER_MARKER here because is_busy()
712 // ignores DEFLATER_MARKER values.
713 ? p2i(nullptr)
714 : p2i(owner_raw()),
715 p2i(_cxq),
716 p2i(_EntryList));
717 return ss->base();
718 }
719
720 #define MAX_RECHECK_INTERVAL 1000
721
722 void ObjectMonitor::EnterI(JavaThread* current) {
723 assert(current->thread_state() == _thread_blocked, "invariant");
724
725 // Try the lock - TATAS
726 if (TryLock (current) > 0) {
727 assert(_succ != current, "invariant");
728 assert(owner_raw() == current, "invariant");
729 assert(_Responsible != current, "invariant");
730 return;
731 }
732
733 if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
734 // Cancelled the in-progress async deflation by changing owner from
735 // DEFLATER_MARKER to current. As part of the contended enter protocol,
736 // contentions was incremented to a positive value before EnterI()
737 // was called and that prevents the deflater thread from winning the
738 // last part of the 2-part async deflation protocol. After EnterI()
739 // returns to enter(), contentions is decremented because the caller
740 // now owns the monitor. We bump contentions an extra time here to
741 // prevent the deflater thread from winning the last part of the
742 // 2-part async deflation protocol after the regular decrement
743 // occurs in enter(). The deflater thread will decrement contentions
744 // after it recognizes that the async deflation was cancelled.
745 add_to_contentions(1);
746 assert(_succ != current, "invariant");
747 assert(_Responsible != current, "invariant");
748 return;
749 }
750
751 assert(InitDone, "Unexpectedly not initialized");
752
753 // We try one round of spinning *before* enqueueing current.
754 //
755 // If the _owner is ready but OFFPROC we could use a YieldTo()
756 // operation to donate the remainder of this thread's quantum
757 // to the owner. This has subtle but beneficial affinity
758 // effects.
759
760 if (TrySpin(current) > 0) {
761 assert(owner_raw() == current, "invariant");
762 assert(_succ != current, "invariant");
763 assert(_Responsible != current, "invariant");
764 return;
765 }
766
767 // The Spin failed -- Enqueue and park the thread ...
768 assert(_succ != current, "invariant");
769 assert(owner_raw() != current, "invariant");
770 assert(_Responsible != current, "invariant");
771
772 // Enqueue "current" on ObjectMonitor's _cxq.
773 //
774 // Node acts as a proxy for current.
775 // As an aside, if were to ever rewrite the synchronization code mostly
776 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
777 // Java objects. This would avoid awkward lifecycle and liveness issues,
778 // as well as eliminate a subset of ABA issues.
779 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
780
781 ObjectWaiter node(current);
782 current->_ParkEvent->reset();
783 node._prev = (ObjectWaiter*) 0xBAD;
784 node.TState = ObjectWaiter::TS_CXQ;
785
786 // Push "current" onto the front of the _cxq.
787 // Once on cxq/EntryList, current stays on-queue until it acquires the lock.
788 // Note that spinning tends to reduce the rate at which threads
789 // enqueue and dequeue on EntryList|cxq.
790 ObjectWaiter* nxt;
791 for (;;) {
792 node._next = nxt = _cxq;
793 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break;
794
795 // Interference - the CAS failed because _cxq changed. Just retry.
796 // As an optional optimization we retry the lock.
797 if (TryLock (current) > 0) {
798 assert(_succ != current, "invariant");
799 assert(owner_raw() == current, "invariant");
800 assert(_Responsible != current, "invariant");
801 return;
802 }
803 }
804
805 // Check for cxq|EntryList edge transition to non-null. This indicates
806 // the onset of contention. While contention persists exiting threads
807 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit
808 // operations revert to the faster 1-0 mode. This enter operation may interleave
809 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
810 // arrange for one of the contending thread to use a timed park() operations
811 // to detect and recover from the race. (Stranding is form of progress failure
812 // where the monitor is unlocked but all the contending threads remain parked).
813 // That is, at least one of the contended threads will periodically poll _owner.
814 // One of the contending threads will become the designated "Responsible" thread.
815 // The Responsible thread uses a timed park instead of a normal indefinite park
816 // operation -- it periodically wakes and checks for and recovers from potential
817 // strandings admitted by 1-0 exit operations. We need at most one Responsible
818 // thread per-monitor at any given moment. Only threads on cxq|EntryList may
819 // be responsible for a monitor.
820 //
821 // Currently, one of the contended threads takes on the added role of "Responsible".
822 // A viable alternative would be to use a dedicated "stranding checker" thread
823 // that periodically iterated over all the threads (or active monitors) and unparked
824 // successors where there was risk of stranding. This would help eliminate the
825 // timer scalability issues we see on some platforms as we'd only have one thread
826 // -- the checker -- parked on a timer.
827
828 if (nxt == nullptr && _EntryList == nullptr) {
829 // Try to assume the role of responsible thread for the monitor.
830 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=current }
831 Atomic::replace_if_null(&_Responsible, current);
832 }
833
834 // The lock might have been released while this thread was occupied queueing
835 // itself onto _cxq. To close the race and avoid "stranding" and
836 // progress-liveness failure we must resample-retry _owner before parking.
837 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
838 // In this case the ST-MEMBAR is accomplished with CAS().
839 //
840 // TODO: Defer all thread state transitions until park-time.
841 // Since state transitions are heavy and inefficient we'd like
842 // to defer the state transitions until absolutely necessary,
843 // and in doing so avoid some transitions ...
844
845 int nWakeups = 0;
846 int recheckInterval = 1;
847
848 for (;;) {
849
850 if (TryLock(current) > 0) break;
851 assert(owner_raw() != current, "invariant");
852
853 // park self
854 if (_Responsible == current) {
855 current->_ParkEvent->park((jlong) recheckInterval);
856 // Increase the recheckInterval, but clamp the value.
857 recheckInterval *= 8;
858 if (recheckInterval > MAX_RECHECK_INTERVAL) {
859 recheckInterval = MAX_RECHECK_INTERVAL;
860 }
861 } else {
862 current->_ParkEvent->park();
863 }
864
865 if (TryLock(current) > 0) break;
866
867 if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
868 // Cancelled the in-progress async deflation by changing owner from
869 // DEFLATER_MARKER to current. As part of the contended enter protocol,
870 // contentions was incremented to a positive value before EnterI()
871 // was called and that prevents the deflater thread from winning the
872 // last part of the 2-part async deflation protocol. After EnterI()
873 // returns to enter(), contentions is decremented because the caller
874 // now owns the monitor. We bump contentions an extra time here to
875 // prevent the deflater thread from winning the last part of the
876 // 2-part async deflation protocol after the regular decrement
877 // occurs in enter(). The deflater thread will decrement contentions
878 // after it recognizes that the async deflation was cancelled.
879 add_to_contentions(1);
880 break;
881 }
882
883 // The lock is still contested.
884 // Keep a tally of the # of futile wakeups.
885 // Note that the counter is not protected by a lock or updated by atomics.
886 // That is by design - we trade "lossy" counters which are exposed to
887 // races during updates for a lower probe effect.
888
889 // This PerfData object can be used in parallel with a safepoint.
890 // See the work around in PerfDataManager::destroy().
891 OM_PERFDATA_OP(FutileWakeups, inc());
892 ++nWakeups;
893
894 // Assuming this is not a spurious wakeup we'll normally find _succ == current.
895 // We can defer clearing _succ until after the spin completes
896 // TrySpin() must tolerate being called with _succ == current.
897 // Try yet another round of adaptive spinning.
898 if (TrySpin(current) > 0) break;
899
900 // We can find that we were unpark()ed and redesignated _succ while
901 // we were spinning. That's harmless. If we iterate and call park(),
902 // park() will consume the event and return immediately and we'll
903 // just spin again. This pattern can repeat, leaving _succ to simply
904 // spin on a CPU.
905
906 if (_succ == current) _succ = nullptr;
907
908 // Invariant: after clearing _succ a thread *must* retry _owner before parking.
909 OrderAccess::fence();
910 }
911
912 // Egress :
913 // current has acquired the lock -- Unlink current from the cxq or EntryList.
914 // Normally we'll find current on the EntryList .
915 // From the perspective of the lock owner (this thread), the
916 // EntryList is stable and cxq is prepend-only.
917 // The head of cxq is volatile but the interior is stable.
918 // In addition, current.TState is stable.
919
920 assert(owner_raw() == current, "invariant");
921
922 UnlinkAfterAcquire(current, &node);
923 if (_succ == current) _succ = nullptr;
924
925 assert(_succ != current, "invariant");
926 if (_Responsible == current) {
927 _Responsible = nullptr;
928 OrderAccess::fence(); // Dekker pivot-point
929
930 // We may leave threads on cxq|EntryList without a designated
931 // "Responsible" thread. This is benign. When this thread subsequently
932 // exits the monitor it can "see" such preexisting "old" threads --
933 // threads that arrived on the cxq|EntryList before the fence, above --
934 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads
935 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
936 // non-null and elect a new "Responsible" timer thread.
937 //
938 // This thread executes:
939 // ST Responsible=null; MEMBAR (in enter epilogue - here)
940 // LD cxq|EntryList (in subsequent exit)
941 //
942 // Entering threads in the slow/contended path execute:
943 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
944 // The (ST cxq; MEMBAR) is accomplished with CAS().
945 //
946 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
947 // exit operation from floating above the ST Responsible=null.
948 }
949
950 // We've acquired ownership with CAS().
951 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
952 // But since the CAS() this thread may have also stored into _succ,
953 // EntryList, cxq or Responsible. These meta-data updates must be
954 // visible __before this thread subsequently drops the lock.
955 // Consider what could occur if we didn't enforce this constraint --
956 // STs to monitor meta-data and user-data could reorder with (become
957 // visible after) the ST in exit that drops ownership of the lock.
958 // Some other thread could then acquire the lock, but observe inconsistent
959 // or old monitor meta-data and heap data. That violates the JMM.
960 // To that end, the 1-0 exit() operation must have at least STST|LDST
961 // "release" barrier semantics. Specifically, there must be at least a
962 // STST|LDST barrier in exit() before the ST of null into _owner that drops
963 // the lock. The barrier ensures that changes to monitor meta-data and data
964 // protected by the lock will be visible before we release the lock, and
965 // therefore before some other thread (CPU) has a chance to acquire the lock.
966 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
967 //
968 // Critically, any prior STs to _succ or EntryList must be visible before
969 // the ST of null into _owner in the *subsequent* (following) corresponding
970 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily
971 // execute a serializing instruction.
972
973 return;
974 }
975
976 // ReenterI() is a specialized inline form of the latter half of the
977 // contended slow-path from EnterI(). We use ReenterI() only for
978 // monitor reentry in wait().
979 //
980 // In the future we should reconcile EnterI() and ReenterI().
981
982 void ObjectMonitor::ReenterI(JavaThread* current, ObjectWaiter* currentNode) {
983 assert(current != nullptr, "invariant");
984 assert(currentNode != nullptr, "invariant");
985 assert(currentNode->_thread == current, "invariant");
986 assert(_waiters > 0, "invariant");
987 assert(object()->mark() == markWord::encode(this), "invariant");
988
989 assert(current->thread_state() != _thread_blocked, "invariant");
990
991 int nWakeups = 0;
992 for (;;) {
993 ObjectWaiter::TStates v = currentNode->TState;
994 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
995 assert(owner_raw() != current, "invariant");
996
997 if (TryLock(current) > 0) break;
998 if (TrySpin(current) > 0) break;
999
1000 {
1001 OSThreadContendState osts(current->osthread());
1002
1003 assert(current->thread_state() == _thread_in_vm, "invariant");
1004
1005 {
1006 ClearSuccOnSuspend csos(this);
1007 ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1008 current->_ParkEvent->park();
1009 }
1010 }
1011
1012 // Try again, but just so we distinguish between futile wakeups and
1013 // successful wakeups. The following test isn't algorithmically
1014 // necessary, but it helps us maintain sensible statistics.
1015 if (TryLock(current) > 0) break;
1016
1017 // The lock is still contested.
1018 // Keep a tally of the # of futile wakeups.
1019 // Note that the counter is not protected by a lock or updated by atomics.
1020 // That is by design - we trade "lossy" counters which are exposed to
1021 // races during updates for a lower probe effect.
1022 ++nWakeups;
1023
1024 // Assuming this is not a spurious wakeup we'll normally
1025 // find that _succ == current.
1026 if (_succ == current) _succ = nullptr;
1027
1028 // Invariant: after clearing _succ a contending thread
1029 // *must* retry _owner before parking.
1030 OrderAccess::fence();
1031
1032 // This PerfData object can be used in parallel with a safepoint.
1033 // See the work around in PerfDataManager::destroy().
1034 OM_PERFDATA_OP(FutileWakeups, inc());
1035 }
1036
1037 // current has acquired the lock -- Unlink current from the cxq or EntryList .
1038 // Normally we'll find current on the EntryList.
1039 // Unlinking from the EntryList is constant-time and atomic-free.
1040 // From the perspective of the lock owner (this thread), the
1041 // EntryList is stable and cxq is prepend-only.
1042 // The head of cxq is volatile but the interior is stable.
1043 // In addition, current.TState is stable.
1044
1045 assert(owner_raw() == current, "invariant");
1046 assert(object()->mark() == markWord::encode(this), "invariant");
1047 UnlinkAfterAcquire(current, currentNode);
1048 if (_succ == current) _succ = nullptr;
1049 assert(_succ != current, "invariant");
1050 currentNode->TState = ObjectWaiter::TS_RUN;
1051 OrderAccess::fence(); // see comments at the end of EnterI()
1052 }
1053
1054 // By convention we unlink a contending thread from EntryList|cxq immediately
1055 // after the thread acquires the lock in ::enter(). Equally, we could defer
1056 // unlinking the thread until ::exit()-time.
1057
1058 void ObjectMonitor::UnlinkAfterAcquire(JavaThread* current, ObjectWaiter* currentNode) {
1059 assert(owner_raw() == current, "invariant");
1060 assert(currentNode->_thread == current, "invariant");
1061
1062 if (currentNode->TState == ObjectWaiter::TS_ENTER) {
1063 // Normal case: remove current from the DLL EntryList .
1064 // This is a constant-time operation.
1065 ObjectWaiter* nxt = currentNode->_next;
1066 ObjectWaiter* prv = currentNode->_prev;
1067 if (nxt != nullptr) nxt->_prev = prv;
1068 if (prv != nullptr) prv->_next = nxt;
1069 if (currentNode == _EntryList) _EntryList = nxt;
1070 assert(nxt == nullptr || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
1071 assert(prv == nullptr || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
1072 } else {
1073 assert(currentNode->TState == ObjectWaiter::TS_CXQ, "invariant");
1074 // Inopportune interleaving -- current is still on the cxq.
1075 // This usually means the enqueue of self raced an exiting thread.
1076 // Normally we'll find current near the front of the cxq, so
1077 // dequeueing is typically fast. If needbe we can accelerate
1078 // this with some MCS/CHL-like bidirectional list hints and advisory
1079 // back-links so dequeueing from the interior will normally operate
1080 // in constant-time.
1081 // Dequeue current from either the head (with CAS) or from the interior
1082 // with a linear-time scan and normal non-atomic memory operations.
1083 // CONSIDER: if current is on the cxq then simply drain cxq into EntryList
1084 // and then unlink current from EntryList. We have to drain eventually,
1085 // so it might as well be now.
1086
1087 ObjectWaiter* v = _cxq;
1088 assert(v != nullptr, "invariant");
1089 if (v != currentNode || Atomic::cmpxchg(&_cxq, v, currentNode->_next) != v) {
1090 // The CAS above can fail from interference IFF a "RAT" arrived.
1091 // In that case current must be in the interior and can no longer be
1092 // at the head of cxq.
1093 if (v == currentNode) {
1094 assert(_cxq != v, "invariant");
1095 v = _cxq; // CAS above failed - start scan at head of list
1096 }
1097 ObjectWaiter* p;
1098 ObjectWaiter* q = nullptr;
1099 for (p = v; p != nullptr && p != currentNode; p = p->_next) {
1100 q = p;
1101 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
1102 }
1103 assert(v != currentNode, "invariant");
1104 assert(p == currentNode, "Node not found on cxq");
1105 assert(p != _cxq, "invariant");
1106 assert(q != nullptr, "invariant");
1107 assert(q->_next == p, "invariant");
1108 q->_next = p->_next;
1109 }
1110 }
1111
1112 #ifdef ASSERT
1113 // Diagnostic hygiene ...
1114 currentNode->_prev = (ObjectWaiter*) 0xBAD;
1115 currentNode->_next = (ObjectWaiter*) 0xBAD;
1116 currentNode->TState = ObjectWaiter::TS_RUN;
1117 #endif
1118 }
1119
1120 // -----------------------------------------------------------------------------
1121 // Exit support
1122 //
1123 // exit()
1124 // ~~~~~~
1125 // Note that the collector can't reclaim the objectMonitor or deflate
1126 // the object out from underneath the thread calling ::exit() as the
1127 // thread calling ::exit() never transitions to a stable state.
1128 // This inhibits GC, which in turn inhibits asynchronous (and
1129 // inopportune) reclamation of "this".
1130 //
1131 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
1132 // There's one exception to the claim above, however. EnterI() can call
1133 // exit() to drop a lock if the acquirer has been externally suspended.
1134 // In that case exit() is called with _thread_state == _thread_blocked,
1135 // but the monitor's _contentions field is > 0, which inhibits reclamation.
1136 //
1137 // 1-0 exit
1138 // ~~~~~~~~
1139 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
1140 // the fast-path operators have been optimized so the common ::exit()
1141 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
1142 // The code emitted by fast_unlock() elides the usual MEMBAR. This
1143 // greatly improves latency -- MEMBAR and CAS having considerable local
1144 // latency on modern processors -- but at the cost of "stranding". Absent the
1145 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
1146 // ::enter() path, resulting in the entering thread being stranding
1147 // and a progress-liveness failure. Stranding is extremely rare.
1148 // We use timers (timed park operations) & periodic polling to detect
1149 // and recover from stranding. Potentially stranded threads periodically
1150 // wake up and poll the lock. See the usage of the _Responsible variable.
1151 //
1152 // The CAS() in enter provides for safety and exclusion, while the CAS or
1153 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking
1154 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
1155 // We detect and recover from stranding with timers.
1156 //
1157 // If a thread transiently strands it'll park until (a) another
1158 // thread acquires the lock and then drops the lock, at which time the
1159 // exiting thread will notice and unpark the stranded thread, or, (b)
1160 // the timer expires. If the lock is high traffic then the stranding latency
1161 // will be low due to (a). If the lock is low traffic then the odds of
1162 // stranding are lower, although the worst-case stranding latency
1163 // is longer. Critically, we don't want to put excessive load in the
1164 // platform's timer subsystem. We want to minimize both the timer injection
1165 // rate (timers created/sec) as well as the number of timers active at
1166 // any one time. (more precisely, we want to minimize timer-seconds, which is
1167 // the integral of the # of active timers at any instant over time).
1168 // Both impinge on OS scalability. Given that, at most one thread parked on
1169 // a monitor will use a timer.
1170 //
1171 // There is also the risk of a futile wake-up. If we drop the lock
1172 // another thread can reacquire the lock immediately, and we can
1173 // then wake a thread unnecessarily. This is benign, and we've
1174 // structured the code so the windows are short and the frequency
1175 // of such futile wakups is low.
1176
1177 void ObjectMonitor::exit(JavaThread* current, bool not_suspended) {
1178 void* cur = owner_raw();
1179 if (current != cur) {
1180 if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1181 assert(_recursions == 0, "invariant");
1182 set_owner_from_BasicLock(cur, current); // Convert from BasicLock* to Thread*.
1183 _recursions = 0;
1184 } else {
1185 // Apparent unbalanced locking ...
1186 // Naively we'd like to throw IllegalMonitorStateException.
1187 // As a practical matter we can neither allocate nor throw an
1188 // exception as ::exit() can be called from leaf routines.
1189 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
1190 // Upon deeper reflection, however, in a properly run JVM the only
1191 // way we should encounter this situation is in the presence of
1192 // unbalanced JNI locking. TODO: CheckJNICalls.
1193 // See also: CR4414101
1194 #ifdef ASSERT
1195 LogStreamHandle(Error, monitorinflation) lsh;
1196 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT
1197 " is exiting an ObjectMonitor it does not own.", p2i(current));
1198 lsh.print_cr("The imbalance is possibly caused by JNI locking.");
1199 print_debug_style_on(&lsh);
1200 assert(false, "Non-balanced monitor enter/exit!");
1201 #endif
1202 return;
1203 }
1204 }
1205
1206 if (_recursions != 0) {
1207 _recursions--; // this is simple recursive enter
1208 return;
1209 }
1210
1211 // Invariant: after setting Responsible=null an thread must execute
1212 // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
1213 _Responsible = nullptr;
1214
1215 #if INCLUDE_JFR
1216 // get the owner's thread id for the MonitorEnter event
1217 // if it is enabled and the thread isn't suspended
1218 if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
1219 _previous_owner_tid = JFR_THREAD_ID(current);
1220 }
1221 #endif
1222
1223 for (;;) {
1224 assert(current == owner_raw(), "invariant");
1225
1226 // Drop the lock.
1227 // release semantics: prior loads and stores from within the critical section
1228 // must not float (reorder) past the following store that drops the lock.
1229 // Uses a storeload to separate release_store(owner) from the
1230 // successor check. The try_set_owner() below uses cmpxchg() so
1231 // we get the fence down there.
1232 release_clear_owner(current);
1233 OrderAccess::storeload();
1234
1235 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != nullptr) {
1236 return;
1237 }
1238 // Other threads are blocked trying to acquire the lock.
1239
1240 // Normally the exiting thread is responsible for ensuring succession,
1241 // but if other successors are ready or other entering threads are spinning
1242 // then this thread can simply store null into _owner and exit without
1243 // waking a successor. The existence of spinners or ready successors
1244 // guarantees proper succession (liveness). Responsibility passes to the
1245 // ready or running successors. The exiting thread delegates the duty.
1246 // More precisely, if a successor already exists this thread is absolved
1247 // of the responsibility of waking (unparking) one.
1248 //
1249 // The _succ variable is critical to reducing futile wakeup frequency.
1250 // _succ identifies the "heir presumptive" thread that has been made
1251 // ready (unparked) but that has not yet run. We need only one such
1252 // successor thread to guarantee progress.
1253 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1254 // section 3.3 "Futile Wakeup Throttling" for details.
1255 //
1256 // Note that spinners in Enter() also set _succ non-null.
1257 // In the current implementation spinners opportunistically set
1258 // _succ so that exiting threads might avoid waking a successor.
1259 // Another less appealing alternative would be for the exiting thread
1260 // to drop the lock and then spin briefly to see if a spinner managed
1261 // to acquire the lock. If so, the exiting thread could exit
1262 // immediately without waking a successor, otherwise the exiting
1263 // thread would need to dequeue and wake a successor.
1264 // (Note that we'd need to make the post-drop spin short, but no
1265 // shorter than the worst-case round-trip cache-line migration time.
1266 // The dropped lock needs to become visible to the spinner, and then
1267 // the acquisition of the lock by the spinner must become visible to
1268 // the exiting thread).
1269
1270 // It appears that an heir-presumptive (successor) must be made ready.
1271 // Only the current lock owner can manipulate the EntryList or
1272 // drain _cxq, so we need to reacquire the lock. If we fail
1273 // to reacquire the lock the responsibility for ensuring succession
1274 // falls to the new owner.
1275 //
1276 if (try_set_owner_from(nullptr, current) != nullptr) {
1277 return;
1278 }
1279
1280 guarantee(owner_raw() == current, "invariant");
1281
1282 ObjectWaiter* w = nullptr;
1283
1284 w = _EntryList;
1285 if (w != nullptr) {
1286 // I'd like to write: guarantee (w->_thread != current).
1287 // But in practice an exiting thread may find itself on the EntryList.
1288 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and
1289 // then calls exit(). Exit release the lock by setting O._owner to null.
1290 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The
1291 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1292 // release the lock "O". T2 resumes immediately after the ST of null into
1293 // _owner, above. T2 notices that the EntryList is populated, so it
1294 // reacquires the lock and then finds itself on the EntryList.
1295 // Given all that, we have to tolerate the circumstance where "w" is
1296 // associated with current.
1297 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1298 ExitEpilog(current, w);
1299 return;
1300 }
1301
1302 // If we find that both _cxq and EntryList are null then just
1303 // re-run the exit protocol from the top.
1304 w = _cxq;
1305 if (w == nullptr) continue;
1306
1307 // Drain _cxq into EntryList - bulk transfer.
1308 // First, detach _cxq.
1309 // The following loop is tantamount to: w = swap(&cxq, nullptr)
1310 for (;;) {
1311 assert(w != nullptr, "Invariant");
1312 ObjectWaiter* u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)nullptr);
1313 if (u == w) break;
1314 w = u;
1315 }
1316
1317 assert(w != nullptr, "invariant");
1318 assert(_EntryList == nullptr, "invariant");
1319
1320 // Convert the LIFO SLL anchored by _cxq into a DLL.
1321 // The list reorganization step operates in O(LENGTH(w)) time.
1322 // It's critical that this step operate quickly as
1323 // "current" still holds the outer-lock, restricting parallelism
1324 // and effectively lengthening the critical section.
1325 // Invariant: s chases t chases u.
1326 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1327 // we have faster access to the tail.
1328
1329 _EntryList = w;
1330 ObjectWaiter* q = nullptr;
1331 ObjectWaiter* p;
1332 for (p = w; p != nullptr; p = p->_next) {
1333 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1334 p->TState = ObjectWaiter::TS_ENTER;
1335 p->_prev = q;
1336 q = p;
1337 }
1338
1339 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = nullptr
1340 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1341
1342 // See if we can abdicate to a spinner instead of waking a thread.
1343 // A primary goal of the implementation is to reduce the
1344 // context-switch rate.
1345 if (_succ != nullptr) continue;
1346
1347 w = _EntryList;
1348 if (w != nullptr) {
1349 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1350 ExitEpilog(current, w);
1351 return;
1352 }
1353 }
1354 }
1355
1356 void ObjectMonitor::ExitEpilog(JavaThread* current, ObjectWaiter* Wakee) {
1357 assert(owner_raw() == current, "invariant");
1358
1359 // Exit protocol:
1360 // 1. ST _succ = wakee
1361 // 2. membar #loadstore|#storestore;
1362 // 2. ST _owner = nullptr
1363 // 3. unpark(wakee)
1364
1365 _succ = Wakee->_thread;
1366 ParkEvent * Trigger = Wakee->_event;
1367
1368 // Hygiene -- once we've set _owner = nullptr we can't safely dereference Wakee again.
1369 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1370 // out-of-scope (non-extant).
1371 Wakee = nullptr;
1372
1373 // Drop the lock.
1374 // Uses a fence to separate release_store(owner) from the LD in unpark().
1375 release_clear_owner(current);
1376 OrderAccess::fence();
1377
1378 DTRACE_MONITOR_PROBE(contended__exit, this, object(), current);
1379 Trigger->unpark();
1380
1381 // Maintain stats and report events to JVMTI
1382 OM_PERFDATA_OP(Parks, inc());
1383 }
1384
1385 // complete_exit exits a lock returning recursion count
1386 // complete_exit requires an inflated monitor
1387 // The _owner field is not always the Thread addr even with an
1388 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1389 // thread due to contention.
1390 intx ObjectMonitor::complete_exit(JavaThread* current) {
1391 assert(InitDone, "Unexpectedly not initialized");
1392
1393 void* cur = owner_raw();
1394 if (current != cur) {
1395 if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1396 assert(_recursions == 0, "internal state error");
1397 set_owner_from_BasicLock(cur, current); // Convert from BasicLock* to Thread*.
1398 _recursions = 0;
1399 }
1400 }
1401
1402 guarantee(current == owner_raw(), "complete_exit not owner");
1403 intx save = _recursions; // record the old recursion count
1404 _recursions = 0; // set the recursion level to be 0
1405 exit(current); // exit the monitor
1406 guarantee(owner_raw() != current, "invariant");
1407 return save;
1408 }
1409
1410 // Checks that the current THREAD owns this monitor and causes an
1411 // immediate return if it doesn't. We don't use the CHECK macro
1412 // because we want the IMSE to be the only exception that is thrown
1413 // from the call site when false is returned. Any other pending
1414 // exception is ignored.
1415 #define CHECK_OWNER() \
1416 do { \
1417 if (!check_owner(THREAD)) { \
1418 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \
1419 return; \
1420 } \
1421 } while (false)
1422
1423 // Returns true if the specified thread owns the ObjectMonitor.
1424 // Otherwise returns false and throws IllegalMonitorStateException
1425 // (IMSE). If there is a pending exception and the specified thread
1426 // is not the owner, that exception will be replaced by the IMSE.
1427 bool ObjectMonitor::check_owner(TRAPS) {
1428 JavaThread* current = THREAD;
1429 void* cur = owner_raw();
1430 assert(cur != anon_owner_ptr(), "no anon owner here");
1431 if (cur == current) {
1432 return true;
1433 }
1434 if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1435 set_owner_from_BasicLock(cur, current); // Convert from BasicLock* to Thread*.
1436 _recursions = 0;
1437 return true;
1438 }
1439 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(),
1440 "current thread is not owner", false);
1441 }
1442
1443 static inline bool is_excluded(const Klass* monitor_klass) {
1444 assert(monitor_klass != nullptr, "invariant");
1445 NOT_JFR_RETURN_(false);
1446 JFR_ONLY(return vmSymbols::jfr_chunk_rotation_monitor() == monitor_klass->name();)
1447 }
1448
1449 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1450 ObjectMonitor* monitor,
1451 uint64_t notifier_tid,
1452 jlong timeout,
1453 bool timedout) {
1454 assert(event != nullptr, "invariant");
1455 assert(monitor != nullptr, "invariant");
1456 const Klass* monitor_klass = monitor->object()->klass();
1457 if (is_excluded(monitor_klass)) {
1458 return;
1459 }
1460 event->set_monitorClass(monitor_klass);
1461 event->set_timeout(timeout);
1462 // Set an address that is 'unique enough', such that events close in
1463 // time and with the same address are likely (but not guaranteed) to
1464 // belong to the same object.
1465 event->set_address((uintptr_t)monitor);
1466 event->set_notifier(notifier_tid);
1467 event->set_timedOut(timedout);
1468 event->commit();
1469 }
1470
1471 // -----------------------------------------------------------------------------
1472 // Wait/Notify/NotifyAll
1473 //
1474 // Note: a subset of changes to ObjectMonitor::wait()
1475 // will need to be replicated in complete_exit
1476 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1477 JavaThread* current = THREAD;
1478
1479 assert(InitDone, "Unexpectedly not initialized");
1480
1481 CHECK_OWNER(); // Throws IMSE if not owner.
1482
1483 EventJavaMonitorWait event;
1484
1485 // check for a pending interrupt
1486 if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1487 // post monitor waited event. Note that this is past-tense, we are done waiting.
1488 if (JvmtiExport::should_post_monitor_waited()) {
1489 // Note: 'false' parameter is passed here because the
1490 // wait was not timed out due to thread interrupt.
1491 JvmtiExport::post_monitor_waited(current, this, false);
1492
1493 // In this short circuit of the monitor wait protocol, the
1494 // current thread never drops ownership of the monitor and
1495 // never gets added to the wait queue so the current thread
1496 // cannot be made the successor. This means that the
1497 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1498 // consume an unpark() meant for the ParkEvent associated with
1499 // this ObjectMonitor.
1500 }
1501 if (event.should_commit()) {
1502 post_monitor_wait_event(&event, this, 0, millis, false);
1503 }
1504 THROW(vmSymbols::java_lang_InterruptedException());
1505 return;
1506 }
1507
1508 current->set_current_waiting_monitor(this);
1509
1510 // create a node to be put into the queue
1511 // Critically, after we reset() the event but prior to park(), we must check
1512 // for a pending interrupt.
1513 ObjectWaiter node(current);
1514 node.TState = ObjectWaiter::TS_WAIT;
1515 current->_ParkEvent->reset();
1516 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag
1517
1518 // Enter the waiting queue, which is a circular doubly linked list in this case
1519 // but it could be a priority queue or any data structure.
1520 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only
1521 // by the owner of the monitor *except* in the case where park()
1522 // returns because of a timeout of interrupt. Contention is exceptionally rare
1523 // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1524
1525 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1526 AddWaiter(&node);
1527 Thread::SpinRelease(&_WaitSetLock);
1528
1529 _Responsible = nullptr;
1530
1531 intx save = _recursions; // record the old recursion count
1532 _waiters++; // increment the number of waiters
1533 _recursions = 0; // set the recursion level to be 1
1534 exit(current); // exit the monitor
1535 guarantee(owner_raw() != current, "invariant");
1536
1537 // The thread is on the WaitSet list - now park() it.
1538 // On MP systems it's conceivable that a brief spin before we park
1539 // could be profitable.
1540 //
1541 // TODO-FIXME: change the following logic to a loop of the form
1542 // while (!timeout && !interrupted && _notified == 0) park()
1543
1544 int ret = OS_OK;
1545 int WasNotified = 0;
1546
1547 // Need to check interrupt state whilst still _thread_in_vm
1548 bool interrupted = interruptible && current->is_interrupted(false);
1549
1550 { // State transition wrappers
1551 OSThread* osthread = current->osthread();
1552 OSThreadWaitState osts(osthread, true);
1553
1554 assert(current->thread_state() == _thread_in_vm, "invariant");
1555
1556 {
1557 ClearSuccOnSuspend csos(this);
1558 ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1559 if (interrupted || HAS_PENDING_EXCEPTION) {
1560 // Intentionally empty
1561 } else if (node._notified == 0) {
1562 if (millis <= 0) {
1563 current->_ParkEvent->park();
1564 } else {
1565 ret = current->_ParkEvent->park(millis);
1566 }
1567 }
1568 }
1569
1570 // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1571 // from the WaitSet to the EntryList.
1572 // See if we need to remove Node from the WaitSet.
1573 // We use double-checked locking to avoid grabbing _WaitSetLock
1574 // if the thread is not on the wait queue.
1575 //
1576 // Note that we don't need a fence before the fetch of TState.
1577 // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1578 // written by the is thread. (perhaps the fetch might even be satisfied
1579 // by a look-aside into the processor's own store buffer, although given
1580 // the length of the code path between the prior ST and this load that's
1581 // highly unlikely). If the following LD fetches a stale TS_WAIT value
1582 // then we'll acquire the lock and then re-fetch a fresh TState value.
1583 // That is, we fail toward safety.
1584
1585 if (node.TState == ObjectWaiter::TS_WAIT) {
1586 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1587 if (node.TState == ObjectWaiter::TS_WAIT) {
1588 DequeueSpecificWaiter(&node); // unlink from WaitSet
1589 assert(node._notified == 0, "invariant");
1590 node.TState = ObjectWaiter::TS_RUN;
1591 }
1592 Thread::SpinRelease(&_WaitSetLock);
1593 }
1594
1595 // The thread is now either on off-list (TS_RUN),
1596 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1597 // The Node's TState variable is stable from the perspective of this thread.
1598 // No other threads will asynchronously modify TState.
1599 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1600 OrderAccess::loadload();
1601 if (_succ == current) _succ = nullptr;
1602 WasNotified = node._notified;
1603
1604 // Reentry phase -- reacquire the monitor.
1605 // re-enter contended monitor after object.wait().
1606 // retain OBJECT_WAIT state until re-enter successfully completes
1607 // Thread state is thread_in_vm and oop access is again safe,
1608 // although the raw address of the object may have changed.
1609 // (Don't cache naked oops over safepoints, of course).
1610
1611 // post monitor waited event. Note that this is past-tense, we are done waiting.
1612 if (JvmtiExport::should_post_monitor_waited()) {
1613 JvmtiExport::post_monitor_waited(current, this, ret == OS_TIMEOUT);
1614
1615 if (node._notified != 0 && _succ == current) {
1616 // In this part of the monitor wait-notify-reenter protocol it
1617 // is possible (and normal) for another thread to do a fastpath
1618 // monitor enter-exit while this thread is still trying to get
1619 // to the reenter portion of the protocol.
1620 //
1621 // The ObjectMonitor was notified and the current thread is
1622 // the successor which also means that an unpark() has already
1623 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1624 // consume the unpark() that was done when the successor was
1625 // set because the same ParkEvent is shared between Java
1626 // monitors and JVM/TI RawMonitors (for now).
1627 //
1628 // We redo the unpark() to ensure forward progress, i.e., we
1629 // don't want all pending threads hanging (parked) with none
1630 // entering the unlocked monitor.
1631 node._event->unpark();
1632 }
1633 }
1634
1635 if (event.should_commit()) {
1636 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1637 }
1638
1639 OrderAccess::fence();
1640
1641 assert(owner_raw() != current, "invariant");
1642 ObjectWaiter::TStates v = node.TState;
1643 if (v == ObjectWaiter::TS_RUN) {
1644 enter(current);
1645 } else {
1646 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1647 ReenterI(current, &node);
1648 node.wait_reenter_end(this);
1649 }
1650
1651 // current has reacquired the lock.
1652 // Lifecycle - the node representing current must not appear on any queues.
1653 // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1654 // want residual elements associated with this thread left on any lists.
1655 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1656 assert(owner_raw() == current, "invariant");
1657 assert(_succ != current, "invariant");
1658 } // OSThreadWaitState()
1659
1660 current->set_current_waiting_monitor(nullptr);
1661
1662 guarantee(_recursions == 0, "invariant");
1663 int relock_count = JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(current);
1664 _recursions = save // restore the old recursion count
1665 + relock_count; // increased by the deferred relock count
1666 current->inc_held_monitor_count(relock_count); // Deopt never entered these counts.
1667 _waiters--; // decrement the number of waiters
1668
1669 // Verify a few postconditions
1670 assert(owner_raw() == current, "invariant");
1671 assert(_succ != current, "invariant");
1672 assert(object()->mark() == markWord::encode(this), "invariant");
1673
1674 // check if the notification happened
1675 if (!WasNotified) {
1676 // no, it could be timeout or Thread.interrupt() or both
1677 // check for interrupt event, otherwise it is timeout
1678 if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1679 THROW(vmSymbols::java_lang_InterruptedException());
1680 }
1681 }
1682
1683 // NOTE: Spurious wake up will be consider as timeout.
1684 // Monitor notify has precedence over thread interrupt.
1685 }
1686
1687
1688 // Consider:
1689 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1690 // then instead of transferring a thread from the WaitSet to the EntryList
1691 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1692
1693 void ObjectMonitor::INotify(JavaThread* current) {
1694 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1695 ObjectWaiter* iterator = DequeueWaiter();
1696 if (iterator != nullptr) {
1697 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1698 guarantee(iterator->_notified == 0, "invariant");
1699 // Disposition - what might we do with iterator ?
1700 // a. add it directly to the EntryList - either tail (policy == 1)
1701 // or head (policy == 0).
1702 // b. push it onto the front of the _cxq (policy == 2).
1703 // For now we use (b).
1704
1705 iterator->TState = ObjectWaiter::TS_ENTER;
1706
1707 iterator->_notified = 1;
1708 iterator->_notifier_tid = JFR_THREAD_ID(current);
1709
1710 ObjectWaiter* list = _EntryList;
1711 if (list != nullptr) {
1712 assert(list->_prev == nullptr, "invariant");
1713 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1714 assert(list != iterator, "invariant");
1715 }
1716
1717 // prepend to cxq
1718 if (list == nullptr) {
1719 iterator->_next = iterator->_prev = nullptr;
1720 _EntryList = iterator;
1721 } else {
1722 iterator->TState = ObjectWaiter::TS_CXQ;
1723 for (;;) {
1724 ObjectWaiter* front = _cxq;
1725 iterator->_next = front;
1726 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) {
1727 break;
1728 }
1729 }
1730 }
1731
1732 // _WaitSetLock protects the wait queue, not the EntryList. We could
1733 // move the add-to-EntryList operation, above, outside the critical section
1734 // protected by _WaitSetLock. In practice that's not useful. With the
1735 // exception of wait() timeouts and interrupts the monitor owner
1736 // is the only thread that grabs _WaitSetLock. There's almost no contention
1737 // on _WaitSetLock so it's not profitable to reduce the length of the
1738 // critical section.
1739
1740 iterator->wait_reenter_begin(this);
1741 }
1742 Thread::SpinRelease(&_WaitSetLock);
1743 }
1744
1745 // Consider: a not-uncommon synchronization bug is to use notify() when
1746 // notifyAll() is more appropriate, potentially resulting in stranded
1747 // threads; this is one example of a lost wakeup. A useful diagnostic
1748 // option is to force all notify() operations to behave as notifyAll().
1749 //
1750 // Note: We can also detect many such problems with a "minimum wait".
1751 // When the "minimum wait" is set to a small non-zero timeout value
1752 // and the program does not hang whereas it did absent "minimum wait",
1753 // that suggests a lost wakeup bug.
1754
1755 void ObjectMonitor::notify(TRAPS) {
1756 JavaThread* current = THREAD;
1757 CHECK_OWNER(); // Throws IMSE if not owner.
1758 if (_WaitSet == nullptr) {
1759 return;
1760 }
1761 DTRACE_MONITOR_PROBE(notify, this, object(), current);
1762 INotify(current);
1763 OM_PERFDATA_OP(Notifications, inc(1));
1764 }
1765
1766
1767 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1768 // from the waitset to the EntryList. This could be done more efficiently with a
1769 // single bulk transfer but in practice it's not time-critical. Beware too,
1770 // that in prepend-mode we invert the order of the waiters. Let's say that the
1771 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1772 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1773
1774 void ObjectMonitor::notifyAll(TRAPS) {
1775 JavaThread* current = THREAD;
1776 CHECK_OWNER(); // Throws IMSE if not owner.
1777 if (_WaitSet == nullptr) {
1778 return;
1779 }
1780
1781 DTRACE_MONITOR_PROBE(notifyAll, this, object(), current);
1782 int tally = 0;
1783 while (_WaitSet != nullptr) {
1784 tally++;
1785 INotify(current);
1786 }
1787
1788 OM_PERFDATA_OP(Notifications, inc(tally));
1789 }
1790
1791 // -----------------------------------------------------------------------------
1792 // Adaptive Spinning Support
1793 //
1794 // Adaptive spin-then-block - rational spinning
1795 //
1796 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1797 // algorithm. On high order SMP systems it would be better to start with
1798 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH,
1799 // a contending thread could enqueue itself on the cxq and then spin locally
1800 // on a thread-specific variable such as its ParkEvent._Event flag.
1801 // That's left as an exercise for the reader. Note that global spinning is
1802 // not problematic on Niagara, as the L2 cache serves the interconnect and
1803 // has both low latency and massive bandwidth.
1804 //
1805 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1806 // acquisition attempts where we opt to spin -- at 100% and vary the spin count
1807 // (duration) or we can fix the count at approximately the duration of
1808 // a context switch and vary the frequency. Of course we could also
1809 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1810 // For a description of 'Adaptive spin-then-block mutual exclusion in
1811 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1812 //
1813 // This implementation varies the duration "D", where D varies with
1814 // the success rate of recent spin attempts. (D is capped at approximately
1815 // length of a round-trip context switch). The success rate for recent
1816 // spin attempts is a good predictor of the success rate of future spin
1817 // attempts. The mechanism adapts automatically to varying critical
1818 // section length (lock modality), system load and degree of parallelism.
1819 // D is maintained per-monitor in _SpinDuration and is initialized
1820 // optimistically. Spin frequency is fixed at 100%.
1821 //
1822 // Note that _SpinDuration is volatile, but we update it without locks
1823 // or atomics. The code is designed so that _SpinDuration stays within
1824 // a reasonable range even in the presence of races. The arithmetic
1825 // operations on _SpinDuration are closed over the domain of legal values,
1826 // so at worst a race will install and older but still legal value.
1827 // At the very worst this introduces some apparent non-determinism.
1828 // We might spin when we shouldn't or vice-versa, but since the spin
1829 // count are relatively short, even in the worst case, the effect is harmless.
1830 //
1831 // Care must be taken that a low "D" value does not become an
1832 // an absorbing state. Transient spinning failures -- when spinning
1833 // is overall profitable -- should not cause the system to converge
1834 // on low "D" values. We want spinning to be stable and predictable
1835 // and fairly responsive to change and at the same time we don't want
1836 // it to oscillate, become metastable, be "too" non-deterministic,
1837 // or converge on or enter undesirable stable absorbing states.
1838 //
1839 // We implement a feedback-based control system -- using past behavior
1840 // to predict future behavior. We face two issues: (a) if the
1841 // input signal is random then the spin predictor won't provide optimal
1842 // results, and (b) if the signal frequency is too high then the control
1843 // system, which has some natural response lag, will "chase" the signal.
1844 // (b) can arise from multimodal lock hold times. Transient preemption
1845 // can also result in apparent bimodal lock hold times.
1846 // Although sub-optimal, neither condition is particularly harmful, as
1847 // in the worst-case we'll spin when we shouldn't or vice-versa.
1848 // The maximum spin duration is rather short so the failure modes aren't bad.
1849 // To be conservative, I've tuned the gain in system to bias toward
1850 // _not spinning. Relatedly, the system can sometimes enter a mode where it
1851 // "rings" or oscillates between spinning and not spinning. This happens
1852 // when spinning is just on the cusp of profitability, however, so the
1853 // situation is not dire. The state is benign -- there's no need to add
1854 // hysteresis control to damp the transition rate between spinning and
1855 // not spinning.
1856
1857 // Spinning: Fixed frequency (100%), vary duration
1858 int ObjectMonitor::TrySpin(JavaThread* current) {
1859 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning.
1860 int ctr = Knob_FixedSpin;
1861 if (ctr != 0) {
1862 while (--ctr >= 0) {
1863 if (TryLock(current) > 0) return 1;
1864 SpinPause();
1865 }
1866 return 0;
1867 }
1868
1869 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1870 if (TryLock(current) > 0) {
1871 // Increase _SpinDuration ...
1872 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1873 // Raising _SpurDuration to the poverty line is key.
1874 int x = _SpinDuration;
1875 if (x < Knob_SpinLimit) {
1876 if (x < Knob_Poverty) x = Knob_Poverty;
1877 _SpinDuration = x + Knob_BonusB;
1878 }
1879 return 1;
1880 }
1881 SpinPause();
1882 }
1883
1884 // Admission control - verify preconditions for spinning
1885 //
1886 // We always spin a little bit, just to prevent _SpinDuration == 0 from
1887 // becoming an absorbing state. Put another way, we spin briefly to
1888 // sample, just in case the system load, parallelism, contention, or lock
1889 // modality changed.
1890 //
1891 // Consider the following alternative:
1892 // Periodically set _SpinDuration = _SpinLimit and try a long/full
1893 // spin attempt. "Periodically" might mean after a tally of
1894 // the # of failed spin attempts (or iterations) reaches some threshold.
1895 // This takes us into the realm of 1-out-of-N spinning, where we
1896 // hold the duration constant but vary the frequency.
1897
1898 ctr = _SpinDuration;
1899 if (ctr <= 0) return 0;
1900
1901 // We're good to spin ... spin ingress.
1902 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1903 // when preparing to LD...CAS _owner, etc and the CAS is likely
1904 // to succeed.
1905 if (_succ == nullptr) {
1906 _succ = current;
1907 }
1908 Thread* prv = nullptr;
1909
1910 // There are three ways to exit the following loop:
1911 // 1. A successful spin where this thread has acquired the lock.
1912 // 2. Spin failure with prejudice
1913 // 3. Spin failure without prejudice
1914
1915 while (--ctr >= 0) {
1916
1917 // Periodic polling -- Check for pending GC
1918 // Threads may spin while they're unsafe.
1919 // We don't want spinning threads to delay the JVM from reaching
1920 // a stop-the-world safepoint or to steal cycles from GC.
1921 // If we detect a pending safepoint we abort in order that
1922 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1923 // this thread, if safe, doesn't steal cycles from GC.
1924 // This is in keeping with the "no loitering in runtime" rule.
1925 // We periodically check to see if there's a safepoint pending.
1926 if ((ctr & 0xFF) == 0) {
1927 // Can't call SafepointMechanism::should_process() since that
1928 // might update the poll values and we could be in a thread_blocked
1929 // state here which is not allowed so just check the poll.
1930 if (SafepointMechanism::local_poll_armed(current)) {
1931 goto Abort; // abrupt spin egress
1932 }
1933 SpinPause();
1934 }
1935
1936 // Probe _owner with TATAS
1937 // If this thread observes the monitor transition or flicker
1938 // from locked to unlocked to locked, then the odds that this
1939 // thread will acquire the lock in this spin attempt go down
1940 // considerably. The same argument applies if the CAS fails
1941 // or if we observe _owner change from one non-null value to
1942 // another non-null value. In such cases we might abort
1943 // the spin without prejudice or apply a "penalty" to the
1944 // spin count-down variable "ctr", reducing it by 100, say.
1945
1946 JavaThread* ox = static_cast<JavaThread*>(owner_raw());
1947 if (ox == nullptr) {
1948 ox = static_cast<JavaThread*>(try_set_owner_from(nullptr, current));
1949 if (ox == nullptr) {
1950 // The CAS succeeded -- this thread acquired ownership
1951 // Take care of some bookkeeping to exit spin state.
1952 if (_succ == current) {
1953 _succ = nullptr;
1954 }
1955
1956 // Increase _SpinDuration :
1957 // The spin was successful (profitable) so we tend toward
1958 // longer spin attempts in the future.
1959 // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1960 // If we acquired the lock early in the spin cycle it
1961 // makes sense to increase _SpinDuration proportionally.
1962 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1963 int x = _SpinDuration;
1964 if (x < Knob_SpinLimit) {
1965 if (x < Knob_Poverty) x = Knob_Poverty;
1966 _SpinDuration = x + Knob_Bonus;
1967 }
1968 return 1;
1969 }
1970
1971 // The CAS failed ... we can take any of the following actions:
1972 // * penalize: ctr -= CASPenalty
1973 // * exit spin with prejudice -- goto Abort;
1974 // * exit spin without prejudice.
1975 // * Since CAS is high-latency, retry again immediately.
1976 prv = ox;
1977 goto Abort;
1978 }
1979
1980 // Did lock ownership change hands ?
1981 if (ox != prv && prv != nullptr) {
1982 goto Abort;
1983 }
1984 prv = ox;
1985
1986 if (_succ == nullptr) {
1987 _succ = current;
1988 }
1989 }
1990
1991 // Spin failed with prejudice -- reduce _SpinDuration.
1992 // TODO: Use an AIMD-like policy to adjust _SpinDuration.
1993 // AIMD is globally stable.
1994 {
1995 int x = _SpinDuration;
1996 if (x > 0) {
1997 // Consider an AIMD scheme like: x -= (x >> 3) + 100
1998 // This is globally sample and tends to damp the response.
1999 x -= Knob_Penalty;
2000 if (x < 0) x = 0;
2001 _SpinDuration = x;
2002 }
2003 }
2004
2005 Abort:
2006 if (_succ == current) {
2007 _succ = nullptr;
2008 // Invariant: after setting succ=null a contending thread
2009 // must recheck-retry _owner before parking. This usually happens
2010 // in the normal usage of TrySpin(), but it's safest
2011 // to make TrySpin() as foolproof as possible.
2012 OrderAccess::fence();
2013 if (TryLock(current) > 0) return 1;
2014 }
2015 return 0;
2016 }
2017
2018
2019 // -----------------------------------------------------------------------------
2020 // WaitSet management ...
2021
2022 ObjectWaiter::ObjectWaiter(JavaThread* current) {
2023 _next = nullptr;
2024 _prev = nullptr;
2025 _notified = 0;
2026 _notifier_tid = 0;
2027 TState = TS_RUN;
2028 _thread = current;
2029 _event = _thread->_ParkEvent;
2030 _active = false;
2031 assert(_event != nullptr, "invariant");
2032 }
2033
2034 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2035 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(_thread, mon);
2036 }
2037
2038 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2039 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(_thread, _active);
2040 }
2041
2042 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2043 assert(node != nullptr, "should not add null node");
2044 assert(node->_prev == nullptr, "node already in list");
2045 assert(node->_next == nullptr, "node already in list");
2046 // put node at end of queue (circular doubly linked list)
2047 if (_WaitSet == nullptr) {
2048 _WaitSet = node;
2049 node->_prev = node;
2050 node->_next = node;
2051 } else {
2052 ObjectWaiter* head = _WaitSet;
2053 ObjectWaiter* tail = head->_prev;
2054 assert(tail->_next == head, "invariant check");
2055 tail->_next = node;
2056 head->_prev = node;
2057 node->_next = head;
2058 node->_prev = tail;
2059 }
2060 }
2061
2062 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2063 // dequeue the very first waiter
2064 ObjectWaiter* waiter = _WaitSet;
2065 if (waiter) {
2066 DequeueSpecificWaiter(waiter);
2067 }
2068 return waiter;
2069 }
2070
2071 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2072 assert(node != nullptr, "should not dequeue nullptr node");
2073 assert(node->_prev != nullptr, "node already removed from list");
2074 assert(node->_next != nullptr, "node already removed from list");
2075 // when the waiter has woken up because of interrupt,
2076 // timeout or other spurious wake-up, dequeue the
2077 // waiter from waiting list
2078 ObjectWaiter* next = node->_next;
2079 if (next == node) {
2080 assert(node->_prev == node, "invariant check");
2081 _WaitSet = nullptr;
2082 } else {
2083 ObjectWaiter* prev = node->_prev;
2084 assert(prev->_next == node, "invariant check");
2085 assert(next->_prev == node, "invariant check");
2086 next->_prev = prev;
2087 prev->_next = next;
2088 if (_WaitSet == node) {
2089 _WaitSet = next;
2090 }
2091 }
2092 node->_next = nullptr;
2093 node->_prev = nullptr;
2094 }
2095
2096 // -----------------------------------------------------------------------------
2097 // PerfData support
2098 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = nullptr;
2099 PerfCounter * ObjectMonitor::_sync_FutileWakeups = nullptr;
2100 PerfCounter * ObjectMonitor::_sync_Parks = nullptr;
2101 PerfCounter * ObjectMonitor::_sync_Notifications = nullptr;
2102 PerfCounter * ObjectMonitor::_sync_Inflations = nullptr;
2103 PerfCounter * ObjectMonitor::_sync_Deflations = nullptr;
2104 PerfLongVariable * ObjectMonitor::_sync_MonExtant = nullptr;
2105
2106 // One-shot global initialization for the sync subsystem.
2107 // We could also defer initialization and initialize on-demand
2108 // the first time we call ObjectSynchronizer::inflate().
2109 // Initialization would be protected - like so many things - by
2110 // the MonitorCache_lock.
2111
2112 void ObjectMonitor::Initialize() {
2113 assert(!InitDone, "invariant");
2114
2115 if (!os::is_MP()) {
2116 Knob_SpinLimit = 0;
2117 Knob_PreSpin = 0;
2118 Knob_FixedSpin = -1;
2119 }
2120
2121 if (UsePerfData) {
2122 EXCEPTION_MARK;
2123 #define NEWPERFCOUNTER(n) \
2124 { \
2125 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \
2126 CHECK); \
2127 }
2128 #define NEWPERFVARIABLE(n) \
2129 { \
2130 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \
2131 CHECK); \
2132 }
2133 NEWPERFCOUNTER(_sync_Inflations);
2134 NEWPERFCOUNTER(_sync_Deflations);
2135 NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2136 NEWPERFCOUNTER(_sync_FutileWakeups);
2137 NEWPERFCOUNTER(_sync_Parks);
2138 NEWPERFCOUNTER(_sync_Notifications);
2139 NEWPERFVARIABLE(_sync_MonExtant);
2140 #undef NEWPERFCOUNTER
2141 #undef NEWPERFVARIABLE
2142 }
2143
2144 _oop_storage = OopStorageSet::create_weak("ObjectSynchronizer Weak", mtSynchronizer);
2145
2146 DEBUG_ONLY(InitDone = true;)
2147 }
2148
2149 void ObjectMonitor::print_on(outputStream* st) const {
2150 // The minimal things to print for markWord printing, more can be added for debugging and logging.
2151 st->print("{contentions=0x%08x,waiters=0x%08x"
2152 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}",
2153 contentions(), waiters(), recursions(),
2154 p2i(owner()));
2155 }
2156 void ObjectMonitor::print() const { print_on(tty); }
2157
2158 #ifdef ASSERT
2159 // Print the ObjectMonitor like a debugger would:
2160 //
2161 // (ObjectMonitor) 0x00007fdfb6012e40 = {
2162 // _header = 0x0000000000000001
2163 // _object = 0x000000070ff45fd0
2164 // _pad_buf0 = {
2165 // [0] = '\0'
2166 // ...
2167 // [43] = '\0'
2168 // }
2169 // _owner = 0x0000000000000000
2170 // _previous_owner_tid = 0
2171 // _pad_buf1 = {
2172 // [0] = '\0'
2173 // ...
2174 // [47] = '\0'
2175 // }
2176 // _next_om = 0x0000000000000000
2177 // _recursions = 0
2178 // _EntryList = 0x0000000000000000
2179 // _cxq = 0x0000000000000000
2180 // _succ = 0x0000000000000000
2181 // _Responsible = 0x0000000000000000
2182 // _SpinDuration = 5000
2183 // _contentions = 0
2184 // _WaitSet = 0x0000700009756248
2185 // _waiters = 1
2186 // _WaitSetLock = 0
2187 // }
2188 //
2189 void ObjectMonitor::print_debug_style_on(outputStream* st) const {
2190 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this));
2191 st->print_cr(" _header = " INTPTR_FORMAT, header().value());
2192 st->print_cr(" _object = " INTPTR_FORMAT, p2i(object_peek()));
2193 st->print_cr(" _pad_buf0 = {");
2194 st->print_cr(" [0] = '\\0'");
2195 st->print_cr(" ...");
2196 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1);
2197 st->print_cr(" }");
2198 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(owner_raw()));
2199 st->print_cr(" _previous_owner_tid = " UINT64_FORMAT, _previous_owner_tid);
2200 st->print_cr(" _pad_buf1 = {");
2201 st->print_cr(" [0] = '\\0'");
2202 st->print_cr(" ...");
2203 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1);
2204 st->print_cr(" }");
2205 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(next_om()));
2206 st->print_cr(" _recursions = " INTX_FORMAT, _recursions);
2207 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList));
2208 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq));
2209 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ));
2210 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible));
2211 st->print_cr(" _SpinDuration = %d", _SpinDuration);
2212 st->print_cr(" _contentions = %d", contentions());
2213 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet));
2214 st->print_cr(" _waiters = %d", _waiters);
2215 st->print_cr(" _WaitSetLock = %d", _WaitSetLock);
2216 st->print_cr("}");
2217 }
2218 #endif