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 DEBUG_ONLY(static volatile bool InitDone = false;)
112
113 OopStorage* ObjectMonitor::_oop_storage = nullptr;
114
115 // -----------------------------------------------------------------------------
116 // Theory of operations -- Monitors lists, thread residency, etc:
117 //
118 // * A thread acquires ownership of a monitor by successfully
119 // CAS()ing the _owner field from null to non-null.
120 //
121 // * Invariant: A thread appears on at most one monitor list --
122 // cxq, EntryList or WaitSet -- at any one time.
123 //
124 // * Contending threads "push" themselves onto the cxq with CAS
125 // and then spin/park.
126 //
127 // * After a contending thread eventually acquires the lock it must
128 // dequeue itself from either the EntryList or the cxq.
129 //
130 // * The exiting thread identifies and unparks an "heir presumptive"
131 // tentative successor thread on the EntryList. Critically, the
132 // exiting thread doesn't unlink the successor thread from the EntryList.
133 // After having been unparked, the wakee will recontend for ownership of
134 // the monitor. The successor (wakee) will either acquire the lock or
135 // re-park itself.
136 //
137 // Succession is provided for by a policy of competitive handoff.
138 // The exiting thread does _not_ grant or pass ownership to the
139 // successor thread. (This is also referred to as "handoff" succession").
140 // Instead the exiting thread releases ownership and possibly wakes
141 // a successor, so the successor can (re)compete for ownership of the lock.
142 // If the EntryList is empty but the cxq is populated the exiting
143 // thread will drain the cxq into the EntryList. It does so by
144 // by detaching the cxq (installing null with CAS) and folding
145 // the threads from the cxq into the EntryList. The EntryList is
146 // doubly linked, while the cxq is singly linked because of the
147 // CAS-based "push" used to enqueue recently arrived threads (RATs).
148 //
149 // * Concurrency invariants:
150 //
151 // -- only the monitor owner may access or mutate the EntryList.
152 // The mutex property of the monitor itself protects the EntryList
153 // from concurrent interference.
154 // -- Only the monitor owner may detach the cxq.
155 //
156 // * The monitor entry list operations avoid locks, but strictly speaking
157 // they're not lock-free. Enter is lock-free, exit is not.
158 // For a description of 'Methods and apparatus providing non-blocking access
159 // to a resource,' see U.S. Pat. No. 7844973.
160 //
161 // * The cxq can have multiple concurrent "pushers" but only one concurrent
162 // detaching thread. This mechanism is immune from the ABA corruption.
163 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
164 //
165 // * Taken together, the cxq and the EntryList constitute or form a
166 // single logical queue of threads stalled trying to acquire the lock.
167 // We use two distinct lists to improve the odds of a constant-time
168 // dequeue operation after acquisition (in the ::enter() epilogue) and
169 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm).
170 // A key desideratum is to minimize queue & monitor metadata manipulation
171 // that occurs while holding the monitor lock -- that is, we want to
172 // minimize monitor lock holds times. Note that even a small amount of
173 // fixed spinning will greatly reduce the # of enqueue-dequeue operations
174 // on EntryList|cxq. That is, spinning relieves contention on the "inner"
175 // locks and monitor metadata.
176 //
177 // Cxq points to the set of Recently Arrived Threads attempting entry.
178 // Because we push threads onto _cxq with CAS, the RATs must take the form of
179 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when
180 // the unlocking thread notices that EntryList is null but _cxq is != null.
181 //
182 // The EntryList is ordered by the prevailing queue discipline and
183 // can be organized in any convenient fashion, such as a doubly-linked list or
184 // a circular doubly-linked list. Critically, we want insert and delete operations
185 // to operate in constant-time. If we need a priority queue then something akin
186 // to Solaris' sleepq would work nicely. Viz.,
187 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
188 // Queue discipline is enforced at ::exit() time, when the unlocking thread
189 // drains the cxq into the EntryList, and orders or reorders the threads on the
190 // EntryList accordingly.
191 //
192 // Barring "lock barging", this mechanism provides fair cyclic ordering,
193 // somewhat similar to an elevator-scan.
194 //
195 // * The monitor synchronization subsystem avoids the use of native
196 // synchronization primitives except for the narrow platform-specific
197 // park-unpark abstraction. See the comments in os_solaris.cpp regarding
198 // the semantics of park-unpark. Put another way, this monitor implementation
199 // depends only on atomic operations and park-unpark. The monitor subsystem
200 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
201 // underlying OS manages the READY<->RUN transitions.
202 //
203 // * Waiting threads reside on the WaitSet list -- wait() puts
204 // the caller onto the WaitSet.
205 //
206 // * notify() or notifyAll() simply transfers threads from the WaitSet to
207 // either the EntryList or cxq. Subsequent exit() operations will
208 // unpark the notifyee. Unparking a notifee in notify() is inefficient -
209 // it's likely the notifyee would simply impale itself on the lock held
210 // by the notifier.
211 //
212 // * An interesting alternative is to encode cxq as (List,LockByte) where
213 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary
214 // variable, like _recursions, in the scheme. The threads or Events that form
215 // the list would have to be aligned in 256-byte addresses. A thread would
216 // try to acquire the lock or enqueue itself with CAS, but exiting threads
217 // could use a 1-0 protocol and simply STB to set the LockByte to 0.
218 // Note that is is *not* word-tearing, but it does presume that full-word
219 // CAS operations are coherent with intermix with STB operations. That's true
220 // on most common processors.
221 //
222 // * See also http://blogs.sun.com/dave
223
224
225 // Check that object() and set_object() are called from the right context:
226 static void check_object_context() {
227 #ifdef ASSERT
228 Thread* self = Thread::current();
229 if (self->is_Java_thread()) {
230 // Mostly called from JavaThreads so sanity check the thread state.
231 JavaThread* jt = JavaThread::cast(self);
232 switch (jt->thread_state()) {
233 case _thread_in_vm: // the usual case
234 case _thread_in_Java: // during deopt
235 break;
236 default:
237 fatal("called from an unsafe thread state");
238 }
239 assert(jt->is_active_Java_thread(), "must be active JavaThread");
240 } else {
241 // However, ThreadService::get_current_contended_monitor()
242 // can call here via the VMThread so sanity check it.
243 assert(self->is_VM_thread(), "must be");
244 }
245 #endif // ASSERT
246 }
247
248 ObjectMonitor::ObjectMonitor(oop object) :
249 _header(markWord::zero()),
250 _object(_oop_storage, object),
251 _owner(nullptr),
252 _previous_owner_tid(0),
253 _next_om(nullptr),
254 _recursions(0),
255 _EntryList(nullptr),
256 _cxq(nullptr),
257 _succ(nullptr),
258 _Responsible(nullptr),
259 _SpinDuration(ObjectMonitor::Knob_SpinLimit),
260 _contentions(0),
261 _WaitSet(nullptr),
262 _waiters(0),
263 _WaitSetLock(0)
264 { }
265
266 ObjectMonitor::~ObjectMonitor() {
267 _object.release(_oop_storage);
268 }
269
270 oop ObjectMonitor::object() const {
271 check_object_context();
272 return _object.resolve();
273 }
274
275 oop ObjectMonitor::object_peek() const {
276 return _object.peek();
277 }
278
279 void ObjectMonitor::ExitOnSuspend::operator()(JavaThread* current) {
280 if (current->is_suspended()) {
281 _om->_recursions = 0;
282 _om->_succ = nullptr;
283 // Don't need a full fence after clearing successor here because of the call to exit().
284 _om->exit(current, false /* not_suspended */);
285 _om_exited = true;
286
287 current->set_current_pending_monitor(_om);
288 }
289 }
290
291 void ObjectMonitor::ClearSuccOnSuspend::operator()(JavaThread* current) {
292 if (current->is_suspended()) {
293 if (_om->_succ == current) {
294 _om->_succ = nullptr;
295 OrderAccess::fence(); // always do a full fence when successor is cleared
296 }
297 }
298 }
299
300 // -----------------------------------------------------------------------------
301 // Enter support
302
303 bool ObjectMonitor::enter_for(JavaThread* locking_thread) {
304 // Used by ObjectSynchronizer::enter_for to enter for another thread.
305 // The monitor is private to or already owned by locking_thread which must be suspended.
306 // So this code may only contend with deflation.
307 assert(locking_thread == Thread::current() || locking_thread->is_obj_deopt_suspend(), "must be");
308
309 // Block out deflation as soon as possible.
310 add_to_contentions(1);
311
312 bool success = false;
313 if (!is_being_async_deflated()) {
314 void* prev_owner = try_set_owner_from(nullptr, locking_thread);
315
316 if (prev_owner == nullptr) {
317 assert(_recursions == 0, "invariant");
318 success = true;
319 } else if (prev_owner == locking_thread) {
320 _recursions++;
321 success = true;
322 } else if (prev_owner == DEFLATER_MARKER) {
323 // Racing with deflation.
324 prev_owner = try_set_owner_from(DEFLATER_MARKER, locking_thread);
325 if (prev_owner == DEFLATER_MARKER) {
326 // Cancelled deflation. Increment contentions as part of the deflation protocol.
327 add_to_contentions(1);
328 success = true;
329 } else if (prev_owner == nullptr) {
330 // At this point we cannot race with deflation as we have both incremented
331 // contentions, seen contention > 0 and seen a DEFLATER_MARKER.
332 // success will only be false if this races with something other than
333 // deflation.
334 prev_owner = try_set_owner_from(nullptr, locking_thread);
335 success = prev_owner == nullptr;
336 }
337 } else if (LockingMode == LM_LEGACY && locking_thread->is_lock_owned((address)prev_owner)) {
338 assert(_recursions == 0, "must be");
339 _recursions = 1;
340 set_owner_from_BasicLock(prev_owner, locking_thread);
341 success = true;
342 }
343 assert(success, "Failed to enter_for: locking_thread=" INTPTR_FORMAT
344 ", this=" INTPTR_FORMAT "{owner=" INTPTR_FORMAT "}, observed owner: " INTPTR_FORMAT,
345 p2i(locking_thread), p2i(this), p2i(owner_raw()), p2i(prev_owner));
346 } else {
347 // Async deflation is in progress and our contentions increment
348 // above lost the race to async deflation. Undo the work and
349 // force the caller to retry.
350 const oop l_object = object();
351 if (l_object != nullptr) {
352 // Attempt to restore the header/dmw to the object's header so that
353 // we only retry once if the deflater thread happens to be slow.
354 install_displaced_markword_in_object(l_object);
355 }
356 }
357
358 add_to_contentions(-1);
359
360 assert(!success || owner_raw() == locking_thread, "must be");
361
362 return success;
363 }
364
365 bool ObjectMonitor::enter(JavaThread* current) {
366 assert(current == JavaThread::current(), "must be");
367 // The following code is ordered to check the most common cases first
368 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
369
370 void* cur = try_set_owner_from(nullptr, current);
371 if (cur == nullptr) {
372 assert(_recursions == 0, "invariant");
373 return true;
374 }
375
376 if (cur == current) {
377 // TODO-FIXME: check for integer overflow! BUGID 6557169.
378 _recursions++;
379 return true;
380 }
381
382 if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
383 assert(_recursions == 0, "internal state error");
384 _recursions = 1;
385 set_owner_from_BasicLock(cur, current); // Convert from BasicLock* to Thread*.
386 return true;
387 }
388
389 // We've encountered genuine contention.
390
391 // Try one round of spinning *before* enqueueing current
392 // and before going through the awkward and expensive state
393 // transitions. The following spin is strictly optional ...
394 // Note that if we acquire the monitor from an initial spin
395 // we forgo posting JVMTI events and firing DTRACE probes.
396 if (TrySpin(current)) {
397 assert(owner_raw() == current, "must be current: owner=" INTPTR_FORMAT, p2i(owner_raw()));
398 assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions);
399 assert(object()->mark() == markWord::encode(this),
400 "object mark must match encoded this: mark=" INTPTR_FORMAT
401 ", encoded this=" INTPTR_FORMAT, object()->mark().value(),
402 markWord::encode(this).value());
403 return true;
404 }
405
406 assert(owner_raw() != current, "invariant");
407 assert(_succ != current, "invariant");
408 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
409 assert(current->thread_state() != _thread_blocked, "invariant");
410
411 // Keep track of contention for JVM/TI and M&M queries.
412 add_to_contentions(1);
413 if (is_being_async_deflated()) {
414 // Async deflation is in progress and our contentions increment
415 // above lost the race to async deflation. Undo the work and
416 // force the caller to retry.
417 const oop l_object = object();
418 if (l_object != nullptr) {
419 // Attempt to restore the header/dmw to the object's header so that
420 // we only retry once if the deflater thread happens to be slow.
421 install_displaced_markword_in_object(l_object);
422 }
423 add_to_contentions(-1);
424 return false;
425 }
426
427 JFR_ONLY(JfrConditionalFlush<EventJavaMonitorEnter> flush(current);)
428 EventJavaMonitorEnter event;
429 if (event.is_started()) {
430 event.set_monitorClass(object()->klass());
431 // Set an address that is 'unique enough', such that events close in
432 // time and with the same address are likely (but not guaranteed) to
433 // belong to the same object.
434 event.set_address((uintptr_t)this);
435 }
436
437 { // Change java thread status to indicate blocked on monitor enter.
438 JavaThreadBlockedOnMonitorEnterState jtbmes(current, this);
439
440 assert(current->current_pending_monitor() == nullptr, "invariant");
441 current->set_current_pending_monitor(this);
442
443 DTRACE_MONITOR_PROBE(contended__enter, this, object(), current);
444 if (JvmtiExport::should_post_monitor_contended_enter()) {
445 JvmtiExport::post_monitor_contended_enter(current, this);
446
447 // The current thread does not yet own the monitor and does not
448 // yet appear on any queues that would get it made the successor.
449 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
450 // handler cannot accidentally consume an unpark() meant for the
451 // ParkEvent associated with this ObjectMonitor.
452 }
453
454 OSThreadContendState osts(current->osthread());
455
456 assert(current->thread_state() == _thread_in_vm, "invariant");
457
458 for (;;) {
459 ExitOnSuspend eos(this);
460 {
461 ThreadBlockInVMPreprocess<ExitOnSuspend> tbivs(current, eos, true /* allow_suspend */);
462 EnterI(current);
463 current->set_current_pending_monitor(nullptr);
464 // We can go to a safepoint at the end of this block. If we
465 // do a thread dump during that safepoint, then this thread will show
466 // as having "-locked" the monitor, but the OS and java.lang.Thread
467 // states will still report that the thread is blocked trying to
468 // acquire it.
469 // If there is a suspend request, ExitOnSuspend will exit the OM
470 // and set the OM as pending.
471 }
472 if (!eos.exited()) {
473 // ExitOnSuspend did not exit the OM
474 assert(owner_raw() == current, "invariant");
475 break;
476 }
477 }
478
479 // We've just gotten past the enter-check-for-suspend dance and we now own
480 // the monitor free and clear.
481 }
482
483 add_to_contentions(-1);
484 assert(contentions() >= 0, "must not be negative: contentions=%d", contentions());
485
486 // Must either set _recursions = 0 or ASSERT _recursions == 0.
487 assert(_recursions == 0, "invariant");
488 assert(owner_raw() == current, "invariant");
489 assert(_succ != current, "invariant");
490 assert(object()->mark() == markWord::encode(this), "invariant");
491
492 // The thread -- now the owner -- is back in vm mode.
493 // Report the glorious news via TI,DTrace and jvmstat.
494 // The probe effect is non-trivial. All the reportage occurs
495 // while we hold the monitor, increasing the length of the critical
496 // section. Amdahl's parallel speedup law comes vividly into play.
497 //
498 // Another option might be to aggregate the events (thread local or
499 // per-monitor aggregation) and defer reporting until a more opportune
500 // time -- such as next time some thread encounters contention but has
501 // yet to acquire the lock. While spinning that thread could
502 // spinning we could increment JVMStat counters, etc.
503
504 DTRACE_MONITOR_PROBE(contended__entered, this, object(), current);
505 if (JvmtiExport::should_post_monitor_contended_entered()) {
506 JvmtiExport::post_monitor_contended_entered(current, this);
507
508 // The current thread already owns the monitor and is not going to
509 // call park() for the remainder of the monitor enter protocol. So
510 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
511 // event handler consumed an unpark() issued by the thread that
512 // just exited the monitor.
513 }
514 if (event.should_commit()) {
515 event.set_previousOwner(_previous_owner_tid);
516 event.commit();
517 }
518 OM_PERFDATA_OP(ContendedLockAttempts, inc());
519 return true;
520 }
521
522 // Caveat: TryLock() is not necessarily serializing if it returns failure.
523 // Callers must compensate as needed.
524
525 ObjectMonitor::TryLockResult ObjectMonitor::TryLock(JavaThread* current) {
526 void* own = owner_raw();
527 if (own != nullptr) return TryLockResult::HasOwner;
528 if (try_set_owner_from(nullptr, current) == nullptr) {
529 assert(_recursions == 0, "invariant");
530 return TryLockResult::Success;
531 }
532 // The lock had been free momentarily, but we lost the race to the lock.
533 // Interference -- the CAS failed.
534 // We can either return -1 or retry.
535 // Retry doesn't make as much sense because the lock was just acquired.
536 return TryLockResult::Interference;
537 }
538
539 // Deflate the specified ObjectMonitor if not in-use. Returns true if it
540 // was deflated and false otherwise.
541 //
542 // The async deflation protocol sets owner to DEFLATER_MARKER and
543 // makes contentions negative as signals to contending threads that
544 // an async deflation is in progress. There are a number of checks
545 // as part of the protocol to make sure that the calling thread has
546 // not lost the race to a contending thread.
547 //
548 // The ObjectMonitor has been successfully async deflated when:
549 // (contentions < 0)
550 // Contending threads that see that condition know to retry their operation.
551 //
552 bool ObjectMonitor::deflate_monitor() {
553 if (is_busy()) {
554 // Easy checks are first - the ObjectMonitor is busy so no deflation.
555 return false;
556 }
557
558 const oop obj = object_peek();
559
560 if (obj == nullptr) {
561 // If the object died, we can recycle the monitor without racing with
562 // Java threads. The GC already broke the association with the object.
563 set_owner_from(nullptr, DEFLATER_MARKER);
564 assert(contentions() >= 0, "must be non-negative: contentions=%d", contentions());
565 _contentions = INT_MIN; // minimum negative int
566 } else {
567 // Attempt async deflation protocol.
568
569 // Set a null owner to DEFLATER_MARKER to force any contending thread
570 // through the slow path. This is just the first part of the async
571 // deflation dance.
572 if (try_set_owner_from(nullptr, DEFLATER_MARKER) != nullptr) {
573 // The owner field is no longer null so we lost the race since the
574 // ObjectMonitor is now busy.
575 return false;
576 }
577
578 if (contentions() > 0 || _waiters != 0) {
579 // Another thread has raced to enter the ObjectMonitor after
580 // is_busy() above or has already entered and waited on
581 // it which makes it busy so no deflation. Restore owner to
582 // null if it is still DEFLATER_MARKER.
583 if (try_set_owner_from(DEFLATER_MARKER, nullptr) != DEFLATER_MARKER) {
584 // Deferred decrement for the JT EnterI() that cancelled the async deflation.
585 add_to_contentions(-1);
586 }
587 return false;
588 }
589
590 // Make a zero contentions field negative to force any contending threads
591 // to retry. This is the second part of the async deflation dance.
592 if (Atomic::cmpxchg(&_contentions, 0, INT_MIN) != 0) {
593 // Contentions was no longer 0 so we lost the race since the
594 // ObjectMonitor is now busy. Restore owner to null if it is
595 // 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
604 // Sanity checks for the races:
605 guarantee(owner_is_DEFLATER_MARKER(), "must be deflater marker");
606 guarantee(contentions() < 0, "must be negative: contentions=%d",
607 contentions());
608 guarantee(_waiters == 0, "must be 0: waiters=%d", _waiters);
609 guarantee(_cxq == nullptr, "must be no contending threads: cxq="
610 INTPTR_FORMAT, p2i(_cxq));
611 guarantee(_EntryList == nullptr,
612 "must be no entering threads: EntryList=" INTPTR_FORMAT,
613 p2i(_EntryList));
614
615 if (obj != nullptr) {
616 if (log_is_enabled(Trace, monitorinflation)) {
617 ResourceMark rm;
618 log_trace(monitorinflation)("deflate_monitor: object=" INTPTR_FORMAT
619 ", mark=" INTPTR_FORMAT ", type='%s'",
620 p2i(obj), obj->mark().value(),
621 obj->klass()->external_name());
622 }
623
624 // Install the old mark word if nobody else has already done it.
625 install_displaced_markword_in_object(obj);
626 }
627
628 // We leave owner == DEFLATER_MARKER and contentions < 0
629 // to force any racing threads to retry.
630 return true; // Success, ObjectMonitor has been deflated.
631 }
632
633 // Install the displaced mark word (dmw) of a deflating ObjectMonitor
634 // into the header of the object associated with the monitor. This
635 // idempotent method is called by a thread that is deflating a
636 // monitor and by other threads that have detected a race with the
637 // deflation process.
638 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) {
639 // This function must only be called when (owner == DEFLATER_MARKER
640 // && contentions <= 0), but we can't guarantee that here because
641 // those values could change when the ObjectMonitor gets moved from
642 // the global free list to a per-thread free list.
643
644 guarantee(obj != nullptr, "must be non-null");
645
646 // Separate loads in is_being_async_deflated(), which is almost always
647 // called before this function, from the load of dmw/header below.
648
649 // _contentions and dmw/header may get written by different threads.
650 // Make sure to observe them in the same order when having several observers.
651 OrderAccess::loadload_for_IRIW();
652
653 const oop l_object = object_peek();
654 if (l_object == nullptr) {
655 // ObjectMonitor's object ref has already been cleared by async
656 // deflation or GC so we're done here.
657 return;
658 }
659 assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj="
660 INTPTR_FORMAT, p2i(l_object), p2i(obj));
661
662 markWord dmw = header();
663 // The dmw has to be neutral (not null, not locked and not marked).
664 assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value());
665
666 // Install displaced mark word if the object's header still points
667 // to this ObjectMonitor. More than one racing caller to this function
668 // can rarely reach this point, but only one can win.
669 markWord res = obj->cas_set_mark(dmw, markWord::encode(this));
670 if (res != markWord::encode(this)) {
671 // This should be rare so log at the Info level when it happens.
672 log_info(monitorinflation)("install_displaced_markword_in_object: "
673 "failed cas_set_mark: new_mark=" INTPTR_FORMAT
674 ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT,
675 dmw.value(), markWord::encode(this).value(),
676 res.value());
677 }
678
679 // Note: It does not matter which thread restored the header/dmw
680 // into the object's header. The thread deflating the monitor just
681 // wanted the object's header restored and it is. The threads that
682 // detected a race with the deflation process also wanted the
683 // object's header restored before they retry their operation and
684 // because it is restored they will only retry once.
685 }
686
687 // Convert the fields used by is_busy() to a string that can be
688 // used for diagnostic output.
689 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) {
690 ss->print("is_busy: waiters=%d"
691 ", contentions=%d"
692 ", owner=" PTR_FORMAT
693 ", cxq=" PTR_FORMAT
694 ", EntryList=" PTR_FORMAT,
695 _waiters,
696 (contentions() > 0 ? contentions() : 0),
697 owner_is_DEFLATER_MARKER()
698 // We report null instead of DEFLATER_MARKER here because is_busy()
699 // ignores DEFLATER_MARKER values.
700 ? p2i(nullptr)
701 : p2i(owner_raw()),
702 p2i(_cxq),
703 p2i(_EntryList));
704 return ss->base();
705 }
706
707 #define MAX_RECHECK_INTERVAL 1000
708
709 void ObjectMonitor::EnterI(JavaThread* current) {
710 assert(current->thread_state() == _thread_blocked, "invariant");
711
712 // Try the lock - TATAS
713 if (TryLock(current) == TryLockResult::Success) {
714 assert(_succ != current, "invariant");
715 assert(owner_raw() == current, "invariant");
716 assert(_Responsible != current, "invariant");
717 return;
718 }
719
720 if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
721 // Cancelled the in-progress async deflation by changing owner from
722 // DEFLATER_MARKER to current. As part of the contended enter protocol,
723 // contentions was incremented to a positive value before EnterI()
724 // was called and that prevents the deflater thread from winning the
725 // last part of the 2-part async deflation protocol. After EnterI()
726 // returns to enter(), contentions is decremented because the caller
727 // now owns the monitor. We bump contentions an extra time here to
728 // prevent the deflater thread from winning the last part of the
729 // 2-part async deflation protocol after the regular decrement
730 // occurs in enter(). The deflater thread will decrement contentions
731 // after it recognizes that the async deflation was cancelled.
732 add_to_contentions(1);
733 assert(_succ != current, "invariant");
734 assert(_Responsible != current, "invariant");
735 return;
736 }
737
738 assert(InitDone, "Unexpectedly not initialized");
739
740 // We try one round of spinning *before* enqueueing current.
741 //
742 // If the _owner is ready but OFFPROC we could use a YieldTo()
743 // operation to donate the remainder of this thread's quantum
744 // to the owner. This has subtle but beneficial affinity
745 // effects.
746
747 if (TrySpin(current)) {
748 assert(owner_raw() == current, "invariant");
749 assert(_succ != current, "invariant");
750 assert(_Responsible != current, "invariant");
751 return;
752 }
753
754 // The Spin failed -- Enqueue and park the thread ...
755 assert(_succ != current, "invariant");
756 assert(owner_raw() != current, "invariant");
757 assert(_Responsible != current, "invariant");
758
759 // Enqueue "current" on ObjectMonitor's _cxq.
760 //
761 // Node acts as a proxy for current.
762 // As an aside, if were to ever rewrite the synchronization code mostly
763 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
764 // Java objects. This would avoid awkward lifecycle and liveness issues,
765 // as well as eliminate a subset of ABA issues.
766 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
767
768 ObjectWaiter node(current);
769 current->_ParkEvent->reset();
770 node._prev = (ObjectWaiter*) 0xBAD;
771 node.TState = ObjectWaiter::TS_CXQ;
772
773 // Push "current" onto the front of the _cxq.
774 // Once on cxq/EntryList, current stays on-queue until it acquires the lock.
775 // Note that spinning tends to reduce the rate at which threads
776 // enqueue and dequeue on EntryList|cxq.
777 ObjectWaiter* nxt;
778 for (;;) {
779 node._next = nxt = _cxq;
780 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break;
781
782 // Interference - the CAS failed because _cxq changed. Just retry.
783 // As an optional optimization we retry the lock.
784 if (TryLock(current) == TryLockResult::Success) {
785 assert(_succ != current, "invariant");
786 assert(owner_raw() == current, "invariant");
787 assert(_Responsible != current, "invariant");
788 return;
789 }
790 }
791
792 // Check for cxq|EntryList edge transition to non-null. This indicates
793 // the onset of contention. While contention persists exiting threads
794 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit
795 // operations revert to the faster 1-0 mode. This enter operation may interleave
796 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
797 // arrange for one of the contending thread to use a timed park() operations
798 // to detect and recover from the race. (Stranding is form of progress failure
799 // where the monitor is unlocked but all the contending threads remain parked).
800 // That is, at least one of the contended threads will periodically poll _owner.
801 // One of the contending threads will become the designated "Responsible" thread.
802 // The Responsible thread uses a timed park instead of a normal indefinite park
803 // operation -- it periodically wakes and checks for and recovers from potential
804 // strandings admitted by 1-0 exit operations. We need at most one Responsible
805 // thread per-monitor at any given moment. Only threads on cxq|EntryList may
806 // be responsible for a monitor.
807 //
808 // Currently, one of the contended threads takes on the added role of "Responsible".
809 // A viable alternative would be to use a dedicated "stranding checker" thread
810 // that periodically iterated over all the threads (or active monitors) and unparked
811 // successors where there was risk of stranding. This would help eliminate the
812 // timer scalability issues we see on some platforms as we'd only have one thread
813 // -- the checker -- parked on a timer.
814
815 if (nxt == nullptr && _EntryList == nullptr) {
816 // Try to assume the role of responsible thread for the monitor.
817 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=current }
818 Atomic::replace_if_null(&_Responsible, current);
819 }
820
821 // The lock might have been released while this thread was occupied queueing
822 // itself onto _cxq. To close the race and avoid "stranding" and
823 // progress-liveness failure we must resample-retry _owner before parking.
824 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
825 // In this case the ST-MEMBAR is accomplished with CAS().
826 //
827 // TODO: Defer all thread state transitions until park-time.
828 // Since state transitions are heavy and inefficient we'd like
829 // to defer the state transitions until absolutely necessary,
830 // and in doing so avoid some transitions ...
831
832 int nWakeups = 0;
833 int recheckInterval = 1;
834
835 for (;;) {
836
837 if (TryLock(current) == TryLockResult::Success) {
838 break;
839 }
840 assert(owner_raw() != current, "invariant");
841
842 // park self
843 if (_Responsible == current) {
844 current->_ParkEvent->park((jlong) recheckInterval);
845 // Increase the recheckInterval, but clamp the value.
846 recheckInterval *= 8;
847 if (recheckInterval > MAX_RECHECK_INTERVAL) {
848 recheckInterval = MAX_RECHECK_INTERVAL;
849 }
850 } else {
851 current->_ParkEvent->park();
852 }
853
854 if (TryLock(current) == TryLockResult::Success) {
855 break;
856 }
857
858 if (try_set_owner_from(DEFLATER_MARKER, current) == DEFLATER_MARKER) {
859 // Cancelled the in-progress async deflation by changing owner from
860 // DEFLATER_MARKER to current. As part of the contended enter protocol,
861 // contentions was incremented to a positive value before EnterI()
862 // was called and that prevents the deflater thread from winning the
863 // last part of the 2-part async deflation protocol. After EnterI()
864 // returns to enter(), contentions is decremented because the caller
865 // now owns the monitor. We bump contentions an extra time here to
866 // prevent the deflater thread from winning the last part of the
867 // 2-part async deflation protocol after the regular decrement
868 // occurs in enter(). The deflater thread will decrement contentions
869 // after it recognizes that the async deflation was cancelled.
870 add_to_contentions(1);
871 break;
872 }
873
874 // The lock is still contested.
875 // Keep a tally of the # of futile wakeups.
876 // Note that the counter is not protected by a lock or updated by atomics.
877 // That is by design - we trade "lossy" counters which are exposed to
878 // races during updates for a lower probe effect.
879
880 // This PerfData object can be used in parallel with a safepoint.
881 // See the work around in PerfDataManager::destroy().
882 OM_PERFDATA_OP(FutileWakeups, inc());
883 ++nWakeups;
884
885 // Assuming this is not a spurious wakeup we'll normally find _succ == current.
886 // We can defer clearing _succ until after the spin completes
887 // TrySpin() must tolerate being called with _succ == current.
888 // Try yet another round of adaptive spinning.
889 if (TrySpin(current)) {
890 break;
891 }
892
893 // We can find that we were unpark()ed and redesignated _succ while
894 // we were spinning. That's harmless. If we iterate and call park(),
895 // park() will consume the event and return immediately and we'll
896 // just spin again. This pattern can repeat, leaving _succ to simply
897 // spin on a CPU.
898
899 if (_succ == current) _succ = nullptr;
900
901 // Invariant: after clearing _succ a thread *must* retry _owner before parking.
902 OrderAccess::fence();
903 }
904
905 // Egress :
906 // current has acquired the lock -- Unlink current from the cxq or EntryList.
907 // Normally we'll find current on the EntryList .
908 // From the perspective of the lock owner (this thread), the
909 // EntryList is stable and cxq is prepend-only.
910 // The head of cxq is volatile but the interior is stable.
911 // In addition, current.TState is stable.
912
913 assert(owner_raw() == current, "invariant");
914
915 UnlinkAfterAcquire(current, &node);
916 if (_succ == current) _succ = nullptr;
917
918 assert(_succ != current, "invariant");
919 if (_Responsible == current) {
920 _Responsible = nullptr;
921 OrderAccess::fence(); // Dekker pivot-point
922
923 // We may leave threads on cxq|EntryList without a designated
924 // "Responsible" thread. This is benign. When this thread subsequently
925 // exits the monitor it can "see" such preexisting "old" threads --
926 // threads that arrived on the cxq|EntryList before the fence, above --
927 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads
928 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
929 // non-null and elect a new "Responsible" timer thread.
930 //
931 // This thread executes:
932 // ST Responsible=null; MEMBAR (in enter epilogue - here)
933 // LD cxq|EntryList (in subsequent exit)
934 //
935 // Entering threads in the slow/contended path execute:
936 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
937 // The (ST cxq; MEMBAR) is accomplished with CAS().
938 //
939 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
940 // exit operation from floating above the ST Responsible=null.
941 }
942
943 // We've acquired ownership with CAS().
944 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
945 // But since the CAS() this thread may have also stored into _succ,
946 // EntryList, cxq or Responsible. These meta-data updates must be
947 // visible __before this thread subsequently drops the lock.
948 // Consider what could occur if we didn't enforce this constraint --
949 // STs to monitor meta-data and user-data could reorder with (become
950 // visible after) the ST in exit that drops ownership of the lock.
951 // Some other thread could then acquire the lock, but observe inconsistent
952 // or old monitor meta-data and heap data. That violates the JMM.
953 // To that end, the 1-0 exit() operation must have at least STST|LDST
954 // "release" barrier semantics. Specifically, there must be at least a
955 // STST|LDST barrier in exit() before the ST of null into _owner that drops
956 // the lock. The barrier ensures that changes to monitor meta-data and data
957 // protected by the lock will be visible before we release the lock, and
958 // therefore before some other thread (CPU) has a chance to acquire the lock.
959 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
960 //
961 // Critically, any prior STs to _succ or EntryList must be visible before
962 // the ST of null into _owner in the *subsequent* (following) corresponding
963 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily
964 // execute a serializing instruction.
965
966 return;
967 }
968
969 // ReenterI() is a specialized inline form of the latter half of the
970 // contended slow-path from EnterI(). We use ReenterI() only for
971 // monitor reentry in wait().
972 //
973 // In the future we should reconcile EnterI() and ReenterI().
974
975 void ObjectMonitor::ReenterI(JavaThread* current, ObjectWaiter* currentNode) {
976 assert(current != nullptr, "invariant");
977 assert(currentNode != nullptr, "invariant");
978 assert(currentNode->_thread == current, "invariant");
979 assert(_waiters > 0, "invariant");
980 assert(object()->mark() == markWord::encode(this), "invariant");
981
982 assert(current->thread_state() != _thread_blocked, "invariant");
983
984 int nWakeups = 0;
985 for (;;) {
986 ObjectWaiter::TStates v = currentNode->TState;
987 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
988 assert(owner_raw() != current, "invariant");
989
990 if (TrySpin(current)) {
991 break;
992 }
993
994 {
995 OSThreadContendState osts(current->osthread());
996
997 assert(current->thread_state() == _thread_in_vm, "invariant");
998
999 {
1000 ClearSuccOnSuspend csos(this);
1001 ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1002 current->_ParkEvent->park();
1003 }
1004 }
1005
1006 // Try again, but just so we distinguish between futile wakeups and
1007 // successful wakeups. The following test isn't algorithmically
1008 // necessary, but it helps us maintain sensible statistics.
1009 if (TryLock(current) == TryLockResult::Success) {
1010 break;
1011 }
1012
1013 // The lock is still contested.
1014 // Keep a tally of the # of futile wakeups.
1015 // Note that the counter is not protected by a lock or updated by atomics.
1016 // That is by design - we trade "lossy" counters which are exposed to
1017 // races during updates for a lower probe effect.
1018 ++nWakeups;
1019
1020 // Assuming this is not a spurious wakeup we'll normally
1021 // find that _succ == current.
1022 if (_succ == current) _succ = nullptr;
1023
1024 // Invariant: after clearing _succ a contending thread
1025 // *must* retry _owner before parking.
1026 OrderAccess::fence();
1027
1028 // This PerfData object can be used in parallel with a safepoint.
1029 // See the work around in PerfDataManager::destroy().
1030 OM_PERFDATA_OP(FutileWakeups, inc());
1031 }
1032
1033 // current has acquired the lock -- Unlink current from the cxq or EntryList .
1034 // Normally we'll find current on the EntryList.
1035 // Unlinking from the EntryList is constant-time and atomic-free.
1036 // From the perspective of the lock owner (this thread), the
1037 // EntryList is stable and cxq is prepend-only.
1038 // The head of cxq is volatile but the interior is stable.
1039 // In addition, current.TState is stable.
1040
1041 assert(owner_raw() == current, "invariant");
1042 assert(object()->mark() == markWord::encode(this), "invariant");
1043 UnlinkAfterAcquire(current, currentNode);
1044 if (_succ == current) _succ = nullptr;
1045 assert(_succ != current, "invariant");
1046 currentNode->TState = ObjectWaiter::TS_RUN;
1047 OrderAccess::fence(); // see comments at the end of EnterI()
1048 }
1049
1050 // By convention we unlink a contending thread from EntryList|cxq immediately
1051 // after the thread acquires the lock in ::enter(). Equally, we could defer
1052 // unlinking the thread until ::exit()-time.
1053
1054 void ObjectMonitor::UnlinkAfterAcquire(JavaThread* current, ObjectWaiter* currentNode) {
1055 assert(owner_raw() == current, "invariant");
1056 assert(currentNode->_thread == current, "invariant");
1057
1058 if (currentNode->TState == ObjectWaiter::TS_ENTER) {
1059 // Normal case: remove current from the DLL EntryList .
1060 // This is a constant-time operation.
1061 ObjectWaiter* nxt = currentNode->_next;
1062 ObjectWaiter* prv = currentNode->_prev;
1063 if (nxt != nullptr) nxt->_prev = prv;
1064 if (prv != nullptr) prv->_next = nxt;
1065 if (currentNode == _EntryList) _EntryList = nxt;
1066 assert(nxt == nullptr || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
1067 assert(prv == nullptr || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
1068 } else {
1069 assert(currentNode->TState == ObjectWaiter::TS_CXQ, "invariant");
1070 // Inopportune interleaving -- current is still on the cxq.
1071 // This usually means the enqueue of self raced an exiting thread.
1072 // Normally we'll find current near the front of the cxq, so
1073 // dequeueing is typically fast. If needbe we can accelerate
1074 // this with some MCS/CHL-like bidirectional list hints and advisory
1075 // back-links so dequeueing from the interior will normally operate
1076 // in constant-time.
1077 // Dequeue current from either the head (with CAS) or from the interior
1078 // with a linear-time scan and normal non-atomic memory operations.
1079 // CONSIDER: if current is on the cxq then simply drain cxq into EntryList
1080 // and then unlink current from EntryList. We have to drain eventually,
1081 // so it might as well be now.
1082
1083 ObjectWaiter* v = _cxq;
1084 assert(v != nullptr, "invariant");
1085 if (v != currentNode || Atomic::cmpxchg(&_cxq, v, currentNode->_next) != v) {
1086 // The CAS above can fail from interference IFF a "RAT" arrived.
1087 // In that case current must be in the interior and can no longer be
1088 // at the head of cxq.
1089 if (v == currentNode) {
1090 assert(_cxq != v, "invariant");
1091 v = _cxq; // CAS above failed - start scan at head of list
1092 }
1093 ObjectWaiter* p;
1094 ObjectWaiter* q = nullptr;
1095 for (p = v; p != nullptr && p != currentNode; p = p->_next) {
1096 q = p;
1097 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
1098 }
1099 assert(v != currentNode, "invariant");
1100 assert(p == currentNode, "Node not found on cxq");
1101 assert(p != _cxq, "invariant");
1102 assert(q != nullptr, "invariant");
1103 assert(q->_next == p, "invariant");
1104 q->_next = p->_next;
1105 }
1106 }
1107
1108 #ifdef ASSERT
1109 // Diagnostic hygiene ...
1110 currentNode->_prev = (ObjectWaiter*) 0xBAD;
1111 currentNode->_next = (ObjectWaiter*) 0xBAD;
1112 currentNode->TState = ObjectWaiter::TS_RUN;
1113 #endif
1114 }
1115
1116 // -----------------------------------------------------------------------------
1117 // Exit support
1118 //
1119 // exit()
1120 // ~~~~~~
1121 // Note that the collector can't reclaim the objectMonitor or deflate
1122 // the object out from underneath the thread calling ::exit() as the
1123 // thread calling ::exit() never transitions to a stable state.
1124 // This inhibits GC, which in turn inhibits asynchronous (and
1125 // inopportune) reclamation of "this".
1126 //
1127 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
1128 // There's one exception to the claim above, however. EnterI() can call
1129 // exit() to drop a lock if the acquirer has been externally suspended.
1130 // In that case exit() is called with _thread_state == _thread_blocked,
1131 // but the monitor's _contentions field is > 0, which inhibits reclamation.
1132 //
1133 // 1-0 exit
1134 // ~~~~~~~~
1135 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
1136 // the fast-path operators have been optimized so the common ::exit()
1137 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
1138 // The code emitted by fast_unlock() elides the usual MEMBAR. This
1139 // greatly improves latency -- MEMBAR and CAS having considerable local
1140 // latency on modern processors -- but at the cost of "stranding". Absent the
1141 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
1142 // ::enter() path, resulting in the entering thread being stranding
1143 // and a progress-liveness failure. Stranding is extremely rare.
1144 // We use timers (timed park operations) & periodic polling to detect
1145 // and recover from stranding. Potentially stranded threads periodically
1146 // wake up and poll the lock. See the usage of the _Responsible variable.
1147 //
1148 // The CAS() in enter provides for safety and exclusion, while the CAS or
1149 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking
1150 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
1151 // We detect and recover from stranding with timers.
1152 //
1153 // If a thread transiently strands it'll park until (a) another
1154 // thread acquires the lock and then drops the lock, at which time the
1155 // exiting thread will notice and unpark the stranded thread, or, (b)
1156 // the timer expires. If the lock is high traffic then the stranding latency
1157 // will be low due to (a). If the lock is low traffic then the odds of
1158 // stranding are lower, although the worst-case stranding latency
1159 // is longer. Critically, we don't want to put excessive load in the
1160 // platform's timer subsystem. We want to minimize both the timer injection
1161 // rate (timers created/sec) as well as the number of timers active at
1162 // any one time. (more precisely, we want to minimize timer-seconds, which is
1163 // the integral of the # of active timers at any instant over time).
1164 // Both impinge on OS scalability. Given that, at most one thread parked on
1165 // a monitor will use a timer.
1166 //
1167 // There is also the risk of a futile wake-up. If we drop the lock
1168 // another thread can reacquire the lock immediately, and we can
1169 // then wake a thread unnecessarily. This is benign, and we've
1170 // structured the code so the windows are short and the frequency
1171 // of such futile wakups is low.
1172
1173 void ObjectMonitor::exit(JavaThread* current, bool not_suspended) {
1174 void* cur = owner_raw();
1175 if (current != cur) {
1176 if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1177 assert(_recursions == 0, "invariant");
1178 set_owner_from_BasicLock(cur, current); // Convert from BasicLock* to Thread*.
1179 _recursions = 0;
1180 } else {
1181 // Apparent unbalanced locking ...
1182 // Naively we'd like to throw IllegalMonitorStateException.
1183 // As a practical matter we can neither allocate nor throw an
1184 // exception as ::exit() can be called from leaf routines.
1185 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
1186 // Upon deeper reflection, however, in a properly run JVM the only
1187 // way we should encounter this situation is in the presence of
1188 // unbalanced JNI locking. TODO: CheckJNICalls.
1189 // See also: CR4414101
1190 #ifdef ASSERT
1191 LogStreamHandle(Error, monitorinflation) lsh;
1192 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT
1193 " is exiting an ObjectMonitor it does not own.", p2i(current));
1194 lsh.print_cr("The imbalance is possibly caused by JNI locking.");
1195 print_debug_style_on(&lsh);
1196 assert(false, "Non-balanced monitor enter/exit!");
1197 #endif
1198 return;
1199 }
1200 }
1201
1202 if (_recursions != 0) {
1203 _recursions--; // this is simple recursive enter
1204 return;
1205 }
1206
1207 // Invariant: after setting Responsible=null an thread must execute
1208 // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
1209 _Responsible = nullptr;
1210
1211 #if INCLUDE_JFR
1212 // get the owner's thread id for the MonitorEnter event
1213 // if it is enabled and the thread isn't suspended
1214 if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
1215 _previous_owner_tid = JFR_THREAD_ID(current);
1216 }
1217 #endif
1218
1219 for (;;) {
1220 assert(current == owner_raw(), "invariant");
1221
1222 // Drop the lock.
1223 // release semantics: prior loads and stores from within the critical section
1224 // must not float (reorder) past the following store that drops the lock.
1225 // Uses a storeload to separate release_store(owner) from the
1226 // successor check. The try_set_owner() below uses cmpxchg() so
1227 // we get the fence down there.
1228 release_clear_owner(current);
1229 OrderAccess::storeload();
1230
1231 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != nullptr) {
1232 return;
1233 }
1234 // Other threads are blocked trying to acquire the lock.
1235
1236 // Normally the exiting thread is responsible for ensuring succession,
1237 // but if other successors are ready or other entering threads are spinning
1238 // then this thread can simply store null into _owner and exit without
1239 // waking a successor. The existence of spinners or ready successors
1240 // guarantees proper succession (liveness). Responsibility passes to the
1241 // ready or running successors. The exiting thread delegates the duty.
1242 // More precisely, if a successor already exists this thread is absolved
1243 // of the responsibility of waking (unparking) one.
1244 //
1245 // The _succ variable is critical to reducing futile wakeup frequency.
1246 // _succ identifies the "heir presumptive" thread that has been made
1247 // ready (unparked) but that has not yet run. We need only one such
1248 // successor thread to guarantee progress.
1249 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1250 // section 3.3 "Futile Wakeup Throttling" for details.
1251 //
1252 // Note that spinners in Enter() also set _succ non-null.
1253 // In the current implementation spinners opportunistically set
1254 // _succ so that exiting threads might avoid waking a successor.
1255 // Another less appealing alternative would be for the exiting thread
1256 // to drop the lock and then spin briefly to see if a spinner managed
1257 // to acquire the lock. If so, the exiting thread could exit
1258 // immediately without waking a successor, otherwise the exiting
1259 // thread would need to dequeue and wake a successor.
1260 // (Note that we'd need to make the post-drop spin short, but no
1261 // shorter than the worst-case round-trip cache-line migration time.
1262 // The dropped lock needs to become visible to the spinner, and then
1263 // the acquisition of the lock by the spinner must become visible to
1264 // the exiting thread).
1265
1266 // It appears that an heir-presumptive (successor) must be made ready.
1267 // Only the current lock owner can manipulate the EntryList or
1268 // drain _cxq, so we need to reacquire the lock. If we fail
1269 // to reacquire the lock the responsibility for ensuring succession
1270 // falls to the new owner.
1271 //
1272 if (try_set_owner_from(nullptr, current) != nullptr) {
1273 return;
1274 }
1275
1276 guarantee(owner_raw() == current, "invariant");
1277
1278 ObjectWaiter* w = nullptr;
1279
1280 w = _EntryList;
1281 if (w != nullptr) {
1282 // I'd like to write: guarantee (w->_thread != current).
1283 // But in practice an exiting thread may find itself on the EntryList.
1284 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and
1285 // then calls exit(). Exit release the lock by setting O._owner to null.
1286 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The
1287 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1288 // release the lock "O". T2 resumes immediately after the ST of null into
1289 // _owner, above. T2 notices that the EntryList is populated, so it
1290 // reacquires the lock and then finds itself on the EntryList.
1291 // Given all that, we have to tolerate the circumstance where "w" is
1292 // associated with current.
1293 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1294 ExitEpilog(current, w);
1295 return;
1296 }
1297
1298 // If we find that both _cxq and EntryList are null then just
1299 // re-run the exit protocol from the top.
1300 w = _cxq;
1301 if (w == nullptr) continue;
1302
1303 // Drain _cxq into EntryList - bulk transfer.
1304 // First, detach _cxq.
1305 // The following loop is tantamount to: w = swap(&cxq, nullptr)
1306 for (;;) {
1307 assert(w != nullptr, "Invariant");
1308 ObjectWaiter* u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)nullptr);
1309 if (u == w) break;
1310 w = u;
1311 }
1312
1313 assert(w != nullptr, "invariant");
1314 assert(_EntryList == nullptr, "invariant");
1315
1316 // Convert the LIFO SLL anchored by _cxq into a DLL.
1317 // The list reorganization step operates in O(LENGTH(w)) time.
1318 // It's critical that this step operate quickly as
1319 // "current" still holds the outer-lock, restricting parallelism
1320 // and effectively lengthening the critical section.
1321 // Invariant: s chases t chases u.
1322 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1323 // we have faster access to the tail.
1324
1325 _EntryList = w;
1326 ObjectWaiter* q = nullptr;
1327 ObjectWaiter* p;
1328 for (p = w; p != nullptr; p = p->_next) {
1329 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1330 p->TState = ObjectWaiter::TS_ENTER;
1331 p->_prev = q;
1332 q = p;
1333 }
1334
1335 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = nullptr
1336 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1337
1338 // See if we can abdicate to a spinner instead of waking a thread.
1339 // A primary goal of the implementation is to reduce the
1340 // context-switch rate.
1341 if (_succ != nullptr) continue;
1342
1343 w = _EntryList;
1344 if (w != nullptr) {
1345 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1346 ExitEpilog(current, w);
1347 return;
1348 }
1349 }
1350 }
1351
1352 void ObjectMonitor::ExitEpilog(JavaThread* current, ObjectWaiter* Wakee) {
1353 assert(owner_raw() == current, "invariant");
1354
1355 // Exit protocol:
1356 // 1. ST _succ = wakee
1357 // 2. membar #loadstore|#storestore;
1358 // 2. ST _owner = nullptr
1359 // 3. unpark(wakee)
1360
1361 _succ = Wakee->_thread;
1362 ParkEvent * Trigger = Wakee->_event;
1363
1364 // Hygiene -- once we've set _owner = nullptr we can't safely dereference Wakee again.
1365 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1366 // out-of-scope (non-extant).
1367 Wakee = nullptr;
1368
1369 // Drop the lock.
1370 // Uses a fence to separate release_store(owner) from the LD in unpark().
1371 release_clear_owner(current);
1372 OrderAccess::fence();
1373
1374 DTRACE_MONITOR_PROBE(contended__exit, this, object(), current);
1375 Trigger->unpark();
1376
1377 // Maintain stats and report events to JVMTI
1378 OM_PERFDATA_OP(Parks, inc());
1379 }
1380
1381 // complete_exit exits a lock returning recursion count
1382 // complete_exit requires an inflated monitor
1383 // The _owner field is not always the Thread addr even with an
1384 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1385 // thread due to contention.
1386 intx ObjectMonitor::complete_exit(JavaThread* current) {
1387 assert(InitDone, "Unexpectedly not initialized");
1388
1389 void* cur = owner_raw();
1390 if (current != cur) {
1391 if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1392 assert(_recursions == 0, "internal state error");
1393 set_owner_from_BasicLock(cur, current); // Convert from BasicLock* to Thread*.
1394 _recursions = 0;
1395 }
1396 }
1397
1398 guarantee(current == owner_raw(), "complete_exit not owner");
1399 intx save = _recursions; // record the old recursion count
1400 _recursions = 0; // set the recursion level to be 0
1401 exit(current); // exit the monitor
1402 guarantee(owner_raw() != current, "invariant");
1403 return save;
1404 }
1405
1406 // Checks that the current THREAD owns this monitor and causes an
1407 // immediate return if it doesn't. We don't use the CHECK macro
1408 // because we want the IMSE to be the only exception that is thrown
1409 // from the call site when false is returned. Any other pending
1410 // exception is ignored.
1411 #define CHECK_OWNER() \
1412 do { \
1413 if (!check_owner(THREAD)) { \
1414 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \
1415 return; \
1416 } \
1417 } while (false)
1418
1419 // Returns true if the specified thread owns the ObjectMonitor.
1420 // Otherwise returns false and throws IllegalMonitorStateException
1421 // (IMSE). If there is a pending exception and the specified thread
1422 // is not the owner, that exception will be replaced by the IMSE.
1423 bool ObjectMonitor::check_owner(TRAPS) {
1424 JavaThread* current = THREAD;
1425 void* cur = owner_raw();
1426 assert(cur != anon_owner_ptr(), "no anon owner here");
1427 if (cur == current) {
1428 return true;
1429 }
1430 if (LockingMode != LM_LIGHTWEIGHT && current->is_lock_owned((address)cur)) {
1431 set_owner_from_BasicLock(cur, current); // Convert from BasicLock* to Thread*.
1432 _recursions = 0;
1433 return true;
1434 }
1435 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(),
1436 "current thread is not owner", false);
1437 }
1438
1439 static inline bool is_excluded(const Klass* monitor_klass) {
1440 assert(monitor_klass != nullptr, "invariant");
1441 NOT_JFR_RETURN_(false);
1442 JFR_ONLY(return vmSymbols::jfr_chunk_rotation_monitor() == monitor_klass->name();)
1443 }
1444
1445 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1446 ObjectMonitor* monitor,
1447 uint64_t notifier_tid,
1448 jlong timeout,
1449 bool timedout) {
1450 assert(event != nullptr, "invariant");
1451 assert(monitor != nullptr, "invariant");
1452 const Klass* monitor_klass = monitor->object()->klass();
1453 if (is_excluded(monitor_klass)) {
1454 return;
1455 }
1456 event->set_monitorClass(monitor_klass);
1457 event->set_timeout(timeout);
1458 // Set an address that is 'unique enough', such that events close in
1459 // time and with the same address are likely (but not guaranteed) to
1460 // belong to the same object.
1461 event->set_address((uintptr_t)monitor);
1462 event->set_notifier(notifier_tid);
1463 event->set_timedOut(timedout);
1464 event->commit();
1465 }
1466
1467 // -----------------------------------------------------------------------------
1468 // Wait/Notify/NotifyAll
1469 //
1470 // Note: a subset of changes to ObjectMonitor::wait()
1471 // will need to be replicated in complete_exit
1472 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1473 JavaThread* current = THREAD;
1474
1475 assert(InitDone, "Unexpectedly not initialized");
1476
1477 CHECK_OWNER(); // Throws IMSE if not owner.
1478
1479 EventJavaMonitorWait event;
1480
1481 // check for a pending interrupt
1482 if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1483 // post monitor waited event. Note that this is past-tense, we are done waiting.
1484 if (JvmtiExport::should_post_monitor_waited()) {
1485 // Note: 'false' parameter is passed here because the
1486 // wait was not timed out due to thread interrupt.
1487 JvmtiExport::post_monitor_waited(current, this, false);
1488
1489 // In this short circuit of the monitor wait protocol, the
1490 // current thread never drops ownership of the monitor and
1491 // never gets added to the wait queue so the current thread
1492 // cannot be made the successor. This means that the
1493 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1494 // consume an unpark() meant for the ParkEvent associated with
1495 // this ObjectMonitor.
1496 }
1497 if (event.should_commit()) {
1498 post_monitor_wait_event(&event, this, 0, millis, false);
1499 }
1500 THROW(vmSymbols::java_lang_InterruptedException());
1501 return;
1502 }
1503
1504 current->set_current_waiting_monitor(this);
1505
1506 // create a node to be put into the queue
1507 // Critically, after we reset() the event but prior to park(), we must check
1508 // for a pending interrupt.
1509 ObjectWaiter node(current);
1510 node.TState = ObjectWaiter::TS_WAIT;
1511 current->_ParkEvent->reset();
1512 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag
1513
1514 // Enter the waiting queue, which is a circular doubly linked list in this case
1515 // but it could be a priority queue or any data structure.
1516 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only
1517 // by the owner of the monitor *except* in the case where park()
1518 // returns because of a timeout of interrupt. Contention is exceptionally rare
1519 // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1520
1521 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1522 AddWaiter(&node);
1523 Thread::SpinRelease(&_WaitSetLock);
1524
1525 _Responsible = nullptr;
1526
1527 intx save = _recursions; // record the old recursion count
1528 _waiters++; // increment the number of waiters
1529 _recursions = 0; // set the recursion level to be 1
1530 exit(current); // exit the monitor
1531 guarantee(owner_raw() != current, "invariant");
1532
1533 // The thread is on the WaitSet list - now park() it.
1534 // On MP systems it's conceivable that a brief spin before we park
1535 // could be profitable.
1536 //
1537 // TODO-FIXME: change the following logic to a loop of the form
1538 // while (!timeout && !interrupted && _notified == 0) park()
1539
1540 int ret = OS_OK;
1541 int WasNotified = 0;
1542
1543 // Need to check interrupt state whilst still _thread_in_vm
1544 bool interrupted = interruptible && current->is_interrupted(false);
1545
1546 { // State transition wrappers
1547 OSThread* osthread = current->osthread();
1548 OSThreadWaitState osts(osthread, true);
1549
1550 assert(current->thread_state() == _thread_in_vm, "invariant");
1551
1552 {
1553 ClearSuccOnSuspend csos(this);
1554 ThreadBlockInVMPreprocess<ClearSuccOnSuspend> tbivs(current, csos, true /* allow_suspend */);
1555 if (interrupted || HAS_PENDING_EXCEPTION) {
1556 // Intentionally empty
1557 } else if (node._notified == 0) {
1558 if (millis <= 0) {
1559 current->_ParkEvent->park();
1560 } else {
1561 ret = current->_ParkEvent->park(millis);
1562 }
1563 }
1564 }
1565
1566 // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1567 // from the WaitSet to the EntryList.
1568 // See if we need to remove Node from the WaitSet.
1569 // We use double-checked locking to avoid grabbing _WaitSetLock
1570 // if the thread is not on the wait queue.
1571 //
1572 // Note that we don't need a fence before the fetch of TState.
1573 // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1574 // written by the is thread. (perhaps the fetch might even be satisfied
1575 // by a look-aside into the processor's own store buffer, although given
1576 // the length of the code path between the prior ST and this load that's
1577 // highly unlikely). If the following LD fetches a stale TS_WAIT value
1578 // then we'll acquire the lock and then re-fetch a fresh TState value.
1579 // That is, we fail toward safety.
1580
1581 if (node.TState == ObjectWaiter::TS_WAIT) {
1582 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1583 if (node.TState == ObjectWaiter::TS_WAIT) {
1584 DequeueSpecificWaiter(&node); // unlink from WaitSet
1585 assert(node._notified == 0, "invariant");
1586 node.TState = ObjectWaiter::TS_RUN;
1587 }
1588 Thread::SpinRelease(&_WaitSetLock);
1589 }
1590
1591 // The thread is now either on off-list (TS_RUN),
1592 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1593 // The Node's TState variable is stable from the perspective of this thread.
1594 // No other threads will asynchronously modify TState.
1595 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1596 OrderAccess::loadload();
1597 if (_succ == current) _succ = nullptr;
1598 WasNotified = node._notified;
1599
1600 // Reentry phase -- reacquire the monitor.
1601 // re-enter contended monitor after object.wait().
1602 // retain OBJECT_WAIT state until re-enter successfully completes
1603 // Thread state is thread_in_vm and oop access is again safe,
1604 // although the raw address of the object may have changed.
1605 // (Don't cache naked oops over safepoints, of course).
1606
1607 // post monitor waited event. Note that this is past-tense, we are done waiting.
1608 if (JvmtiExport::should_post_monitor_waited()) {
1609 JvmtiExport::post_monitor_waited(current, this, ret == OS_TIMEOUT);
1610
1611 if (node._notified != 0 && _succ == current) {
1612 // In this part of the monitor wait-notify-reenter protocol it
1613 // is possible (and normal) for another thread to do a fastpath
1614 // monitor enter-exit while this thread is still trying to get
1615 // to the reenter portion of the protocol.
1616 //
1617 // The ObjectMonitor was notified and the current thread is
1618 // the successor which also means that an unpark() has already
1619 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1620 // consume the unpark() that was done when the successor was
1621 // set because the same ParkEvent is shared between Java
1622 // monitors and JVM/TI RawMonitors (for now).
1623 //
1624 // We redo the unpark() to ensure forward progress, i.e., we
1625 // don't want all pending threads hanging (parked) with none
1626 // entering the unlocked monitor.
1627 node._event->unpark();
1628 }
1629 }
1630
1631 if (event.should_commit()) {
1632 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1633 }
1634
1635 OrderAccess::fence();
1636
1637 assert(owner_raw() != current, "invariant");
1638 ObjectWaiter::TStates v = node.TState;
1639 if (v == ObjectWaiter::TS_RUN) {
1640 enter(current);
1641 } else {
1642 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1643 ReenterI(current, &node);
1644 node.wait_reenter_end(this);
1645 }
1646
1647 // current has reacquired the lock.
1648 // Lifecycle - the node representing current must not appear on any queues.
1649 // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1650 // want residual elements associated with this thread left on any lists.
1651 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1652 assert(owner_raw() == current, "invariant");
1653 assert(_succ != current, "invariant");
1654 } // OSThreadWaitState()
1655
1656 current->set_current_waiting_monitor(nullptr);
1657
1658 guarantee(_recursions == 0, "invariant");
1659 int relock_count = JvmtiDeferredUpdates::get_and_reset_relock_count_after_wait(current);
1660 _recursions = save // restore the old recursion count
1661 + relock_count; // increased by the deferred relock count
1662 current->inc_held_monitor_count(relock_count); // Deopt never entered these counts.
1663 _waiters--; // decrement the number of waiters
1664
1665 // Verify a few postconditions
1666 assert(owner_raw() == current, "invariant");
1667 assert(_succ != current, "invariant");
1668 assert(object()->mark() == markWord::encode(this), "invariant");
1669
1670 // check if the notification happened
1671 if (!WasNotified) {
1672 // no, it could be timeout or Thread.interrupt() or both
1673 // check for interrupt event, otherwise it is timeout
1674 if (interruptible && current->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1675 THROW(vmSymbols::java_lang_InterruptedException());
1676 }
1677 }
1678
1679 // NOTE: Spurious wake up will be consider as timeout.
1680 // Monitor notify has precedence over thread interrupt.
1681 }
1682
1683
1684 // Consider:
1685 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1686 // then instead of transferring a thread from the WaitSet to the EntryList
1687 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1688
1689 void ObjectMonitor::INotify(JavaThread* current) {
1690 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1691 ObjectWaiter* iterator = DequeueWaiter();
1692 if (iterator != nullptr) {
1693 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1694 guarantee(iterator->_notified == 0, "invariant");
1695 // Disposition - what might we do with iterator ?
1696 // a. add it directly to the EntryList - either tail (policy == 1)
1697 // or head (policy == 0).
1698 // b. push it onto the front of the _cxq (policy == 2).
1699 // For now we use (b).
1700
1701 iterator->TState = ObjectWaiter::TS_ENTER;
1702
1703 iterator->_notified = 1;
1704 iterator->_notifier_tid = JFR_THREAD_ID(current);
1705
1706 ObjectWaiter* list = _EntryList;
1707 if (list != nullptr) {
1708 assert(list->_prev == nullptr, "invariant");
1709 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1710 assert(list != iterator, "invariant");
1711 }
1712
1713 // prepend to cxq
1714 if (list == nullptr) {
1715 iterator->_next = iterator->_prev = nullptr;
1716 _EntryList = iterator;
1717 } else {
1718 iterator->TState = ObjectWaiter::TS_CXQ;
1719 for (;;) {
1720 ObjectWaiter* front = _cxq;
1721 iterator->_next = front;
1722 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) {
1723 break;
1724 }
1725 }
1726 }
1727
1728 // _WaitSetLock protects the wait queue, not the EntryList. We could
1729 // move the add-to-EntryList operation, above, outside the critical section
1730 // protected by _WaitSetLock. In practice that's not useful. With the
1731 // exception of wait() timeouts and interrupts the monitor owner
1732 // is the only thread that grabs _WaitSetLock. There's almost no contention
1733 // on _WaitSetLock so it's not profitable to reduce the length of the
1734 // critical section.
1735
1736 iterator->wait_reenter_begin(this);
1737 }
1738 Thread::SpinRelease(&_WaitSetLock);
1739 }
1740
1741 // Consider: a not-uncommon synchronization bug is to use notify() when
1742 // notifyAll() is more appropriate, potentially resulting in stranded
1743 // threads; this is one example of a lost wakeup. A useful diagnostic
1744 // option is to force all notify() operations to behave as notifyAll().
1745 //
1746 // Note: We can also detect many such problems with a "minimum wait".
1747 // When the "minimum wait" is set to a small non-zero timeout value
1748 // and the program does not hang whereas it did absent "minimum wait",
1749 // that suggests a lost wakeup bug.
1750
1751 void ObjectMonitor::notify(TRAPS) {
1752 JavaThread* current = THREAD;
1753 CHECK_OWNER(); // Throws IMSE if not owner.
1754 if (_WaitSet == nullptr) {
1755 return;
1756 }
1757 DTRACE_MONITOR_PROBE(notify, this, object(), current);
1758 INotify(current);
1759 OM_PERFDATA_OP(Notifications, inc(1));
1760 }
1761
1762
1763 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1764 // from the waitset to the EntryList. This could be done more efficiently with a
1765 // single bulk transfer but in practice it's not time-critical. Beware too,
1766 // that in prepend-mode we invert the order of the waiters. Let's say that the
1767 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1768 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1769
1770 void ObjectMonitor::notifyAll(TRAPS) {
1771 JavaThread* current = THREAD;
1772 CHECK_OWNER(); // Throws IMSE if not owner.
1773 if (_WaitSet == nullptr) {
1774 return;
1775 }
1776
1777 DTRACE_MONITOR_PROBE(notifyAll, this, object(), current);
1778 int tally = 0;
1779 while (_WaitSet != nullptr) {
1780 tally++;
1781 INotify(current);
1782 }
1783
1784 OM_PERFDATA_OP(Notifications, inc(tally));
1785 }
1786
1787 // -----------------------------------------------------------------------------
1788 // Adaptive Spinning Support
1789 //
1790 // Adaptive spin-then-block - rational spinning
1791 //
1792 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1793 // algorithm. On high order SMP systems it would be better to start with
1794 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH,
1795 // a contending thread could enqueue itself on the cxq and then spin locally
1796 // on a thread-specific variable such as its ParkEvent._Event flag.
1797 // That's left as an exercise for the reader. Note that global spinning is
1798 // not problematic on Niagara, as the L2 cache serves the interconnect and
1799 // has both low latency and massive bandwidth.
1800 //
1801 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1802 // acquisition attempts where we opt to spin -- at 100% and vary the spin count
1803 // (duration) or we can fix the count at approximately the duration of
1804 // a context switch and vary the frequency. Of course we could also
1805 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1806 // For a description of 'Adaptive spin-then-block mutual exclusion in
1807 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1808 //
1809 // This implementation varies the duration "D", where D varies with
1810 // the success rate of recent spin attempts. (D is capped at approximately
1811 // length of a round-trip context switch). The success rate for recent
1812 // spin attempts is a good predictor of the success rate of future spin
1813 // attempts. The mechanism adapts automatically to varying critical
1814 // section length (lock modality), system load and degree of parallelism.
1815 // D is maintained per-monitor in _SpinDuration and is initialized
1816 // optimistically. Spin frequency is fixed at 100%.
1817 //
1818 // Note that _SpinDuration is volatile, but we update it without locks
1819 // or atomics. The code is designed so that _SpinDuration stays within
1820 // a reasonable range even in the presence of races. The arithmetic
1821 // operations on _SpinDuration are closed over the domain of legal values,
1822 // so at worst a race will install and older but still legal value.
1823 // At the very worst this introduces some apparent non-determinism.
1824 // We might spin when we shouldn't or vice-versa, but since the spin
1825 // count are relatively short, even in the worst case, the effect is harmless.
1826 //
1827 // Care must be taken that a low "D" value does not become an
1828 // an absorbing state. Transient spinning failures -- when spinning
1829 // is overall profitable -- should not cause the system to converge
1830 // on low "D" values. We want spinning to be stable and predictable
1831 // and fairly responsive to change and at the same time we don't want
1832 // it to oscillate, become metastable, be "too" non-deterministic,
1833 // or converge on or enter undesirable stable absorbing states.
1834 //
1835 // We implement a feedback-based control system -- using past behavior
1836 // to predict future behavior. We face two issues: (a) if the
1837 // input signal is random then the spin predictor won't provide optimal
1838 // results, and (b) if the signal frequency is too high then the control
1839 // system, which has some natural response lag, will "chase" the signal.
1840 // (b) can arise from multimodal lock hold times. Transient preemption
1841 // can also result in apparent bimodal lock hold times.
1842 // Although sub-optimal, neither condition is particularly harmful, as
1843 // in the worst-case we'll spin when we shouldn't or vice-versa.
1844 // The maximum spin duration is rather short so the failure modes aren't bad.
1845 // To be conservative, I've tuned the gain in system to bias toward
1846 // _not spinning. Relatedly, the system can sometimes enter a mode where it
1847 // "rings" or oscillates between spinning and not spinning. This happens
1848 // when spinning is just on the cusp of profitability, however, so the
1849 // situation is not dire. The state is benign -- there's no need to add
1850 // hysteresis control to damp the transition rate between spinning and
1851 // not spinning.
1852
1853 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool
1854
1855 static int Knob_Bonus = 100; // spin success bonus
1856 static int Knob_Penalty = 200; // spin failure penalty
1857 static int Knob_Poverty = 1000;
1858 static int Knob_FixedSpin = 0;
1859 static int Knob_PreSpin = 10; // 20-100 likely better, but it's not better in my testing.
1860
1861 inline static int adjust_up(int spin_duration) {
1862 int x = spin_duration;
1863 if (x < ObjectMonitor::Knob_SpinLimit) {
1864 if (x < Knob_Poverty) {
1865 x = Knob_Poverty;
1866 }
1867 return x + Knob_Bonus;
1868 } else {
1869 return spin_duration;
1870 }
1871 }
1872
1873 inline static int adjust_down(int spin_duration) {
1874 // TODO: Use an AIMD-like policy to adjust _SpinDuration.
1875 // AIMD is globally stable.
1876 int x = spin_duration;
1877 if (x > 0) {
1878 // Consider an AIMD scheme like: x -= (x >> 3) + 100
1879 // This is globally sample and tends to damp the response.
1880 x -= Knob_Penalty;
1881 if (x < 0) { x = 0; }
1882 return x;
1883 } else {
1884 return spin_duration;
1885 }
1886 }
1887
1888 bool ObjectMonitor::short_fixed_spin(JavaThread* current, int spin_count, bool adapt) {
1889 for (int ctr = 0; ctr < spin_count; ctr++) {
1890 TryLockResult status = TryLock(current);
1891 if (status == TryLockResult::Success) {
1892 if (adapt) {
1893 _SpinDuration = adjust_up(_SpinDuration);
1894 }
1895 return true;
1896 } else if (status == TryLockResult::Interference) {
1897 break;
1898 }
1899 SpinPause();
1900 }
1901 return false;
1902 }
1903
1904 // Spinning: Fixed frequency (100%), vary duration
1905 bool ObjectMonitor::TrySpin(JavaThread* current) {
1906
1907 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning.
1908 int knob_fixed_spin = Knob_FixedSpin; // 0 (don't spin: default), 2000 good test
1909 if (knob_fixed_spin > 0) {
1910 return short_fixed_spin(current, knob_fixed_spin, false);
1911 }
1912
1913 // Admission control - verify preconditions for spinning
1914 //
1915 // We always spin a little bit, just to prevent _SpinDuration == 0 from
1916 // becoming an absorbing state. Put another way, we spin briefly to
1917 // sample, just in case the system load, parallelism, contention, or lock
1918 // modality changed.
1919
1920 int knob_pre_spin = Knob_PreSpin; // 10 (default), 100, 1000 or 2000
1921 if (short_fixed_spin(current, knob_pre_spin, true)) {
1922 return true;
1923 }
1924
1925 //
1926 // Consider the following alternative:
1927 // Periodically set _SpinDuration = _SpinLimit and try a long/full
1928 // spin attempt. "Periodically" might mean after a tally of
1929 // the # of failed spin attempts (or iterations) reaches some threshold.
1930 // This takes us into the realm of 1-out-of-N spinning, where we
1931 // hold the duration constant but vary the frequency.
1932
1933 int ctr = _SpinDuration;
1934 if (ctr <= 0) return false;
1935
1936 // We're good to spin ... spin ingress.
1937 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1938 // when preparing to LD...CAS _owner, etc and the CAS is likely
1939 // to succeed.
1940 if (_succ == nullptr) {
1941 _succ = current;
1942 }
1943 Thread* prv = nullptr;
1944
1945 // There are three ways to exit the following loop:
1946 // 1. A successful spin where this thread has acquired the lock.
1947 // 2. Spin failure with prejudice
1948 // 3. Spin failure without prejudice
1949
1950 while (--ctr >= 0) {
1951
1952 // Periodic polling -- Check for pending GC
1953 // Threads may spin while they're unsafe.
1954 // We don't want spinning threads to delay the JVM from reaching
1955 // a stop-the-world safepoint or to steal cycles from GC.
1956 // If we detect a pending safepoint we abort in order that
1957 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1958 // this thread, if safe, doesn't steal cycles from GC.
1959 // This is in keeping with the "no loitering in runtime" rule.
1960 // We periodically check to see if there's a safepoint pending.
1961 if ((ctr & 0xFF) == 0) {
1962 // Can't call SafepointMechanism::should_process() since that
1963 // might update the poll values and we could be in a thread_blocked
1964 // state here which is not allowed so just check the poll.
1965 if (SafepointMechanism::local_poll_armed(current)) {
1966 break;
1967 }
1968 SpinPause();
1969 }
1970
1971 // Probe _owner with TATAS
1972 // If this thread observes the monitor transition or flicker
1973 // from locked to unlocked to locked, then the odds that this
1974 // thread will acquire the lock in this spin attempt go down
1975 // considerably. The same argument applies if the CAS fails
1976 // or if we observe _owner change from one non-null value to
1977 // another non-null value. In such cases we might abort
1978 // the spin without prejudice or apply a "penalty" to the
1979 // spin count-down variable "ctr", reducing it by 100, say.
1980
1981 JavaThread* ox = static_cast<JavaThread*>(owner_raw());
1982 if (ox == nullptr) {
1983 ox = static_cast<JavaThread*>(try_set_owner_from(nullptr, current));
1984 if (ox == nullptr) {
1985 // The CAS succeeded -- this thread acquired ownership
1986 // Take care of some bookkeeping to exit spin state.
1987 if (_succ == current) {
1988 _succ = nullptr;
1989 }
1990
1991 // Increase _SpinDuration :
1992 // The spin was successful (profitable) so we tend toward
1993 // longer spin attempts in the future.
1994 // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1995 // If we acquired the lock early in the spin cycle it
1996 // makes sense to increase _SpinDuration proportionally.
1997 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1998 _SpinDuration = adjust_up(_SpinDuration);
1999 return true;
2000 }
2001
2002 // The CAS failed ... we can take any of the following actions:
2003 // * penalize: ctr -= CASPenalty
2004 // * exit spin with prejudice -- abort without adapting spinner
2005 // * exit spin without prejudice.
2006 // * Since CAS is high-latency, retry again immediately.
2007 break;
2008 }
2009
2010 // Did lock ownership change hands ?
2011 if (ox != prv && prv != nullptr) {
2012 break;
2013 }
2014 prv = ox;
2015
2016 if (_succ == nullptr) {
2017 _succ = current;
2018 }
2019 }
2020
2021 // Spin failed with prejudice -- reduce _SpinDuration.
2022 if (ctr < 0) {
2023 _SpinDuration = adjust_down(_SpinDuration);
2024 }
2025
2026 if (_succ == current) {
2027 _succ = nullptr;
2028 // Invariant: after setting succ=null a contending thread
2029 // must recheck-retry _owner before parking. This usually happens
2030 // in the normal usage of TrySpin(), but it's safest
2031 // to make TrySpin() as foolproof as possible.
2032 OrderAccess::fence();
2033 if (TryLock(current) == TryLockResult::Success) {
2034 return true;
2035 }
2036 }
2037
2038 return false;
2039 }
2040
2041
2042 // -----------------------------------------------------------------------------
2043 // WaitSet management ...
2044
2045 ObjectWaiter::ObjectWaiter(JavaThread* current) {
2046 _next = nullptr;
2047 _prev = nullptr;
2048 _notified = 0;
2049 _notifier_tid = 0;
2050 TState = TS_RUN;
2051 _thread = current;
2052 _event = _thread->_ParkEvent;
2053 _active = false;
2054 assert(_event != nullptr, "invariant");
2055 }
2056
2057 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2058 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(_thread, mon);
2059 }
2060
2061 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2062 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(_thread, _active);
2063 }
2064
2065 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2066 assert(node != nullptr, "should not add null node");
2067 assert(node->_prev == nullptr, "node already in list");
2068 assert(node->_next == nullptr, "node already in list");
2069 // put node at end of queue (circular doubly linked list)
2070 if (_WaitSet == nullptr) {
2071 _WaitSet = node;
2072 node->_prev = node;
2073 node->_next = node;
2074 } else {
2075 ObjectWaiter* head = _WaitSet;
2076 ObjectWaiter* tail = head->_prev;
2077 assert(tail->_next == head, "invariant check");
2078 tail->_next = node;
2079 head->_prev = node;
2080 node->_next = head;
2081 node->_prev = tail;
2082 }
2083 }
2084
2085 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2086 // dequeue the very first waiter
2087 ObjectWaiter* waiter = _WaitSet;
2088 if (waiter) {
2089 DequeueSpecificWaiter(waiter);
2090 }
2091 return waiter;
2092 }
2093
2094 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2095 assert(node != nullptr, "should not dequeue nullptr node");
2096 assert(node->_prev != nullptr, "node already removed from list");
2097 assert(node->_next != nullptr, "node already removed from list");
2098 // when the waiter has woken up because of interrupt,
2099 // timeout or other spurious wake-up, dequeue the
2100 // waiter from waiting list
2101 ObjectWaiter* next = node->_next;
2102 if (next == node) {
2103 assert(node->_prev == node, "invariant check");
2104 _WaitSet = nullptr;
2105 } else {
2106 ObjectWaiter* prev = node->_prev;
2107 assert(prev->_next == node, "invariant check");
2108 assert(next->_prev == node, "invariant check");
2109 next->_prev = prev;
2110 prev->_next = next;
2111 if (_WaitSet == node) {
2112 _WaitSet = next;
2113 }
2114 }
2115 node->_next = nullptr;
2116 node->_prev = nullptr;
2117 }
2118
2119 // -----------------------------------------------------------------------------
2120 // PerfData support
2121 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = nullptr;
2122 PerfCounter * ObjectMonitor::_sync_FutileWakeups = nullptr;
2123 PerfCounter * ObjectMonitor::_sync_Parks = nullptr;
2124 PerfCounter * ObjectMonitor::_sync_Notifications = nullptr;
2125 PerfCounter * ObjectMonitor::_sync_Inflations = nullptr;
2126 PerfCounter * ObjectMonitor::_sync_Deflations = nullptr;
2127 PerfLongVariable * ObjectMonitor::_sync_MonExtant = nullptr;
2128
2129 // One-shot global initialization for the sync subsystem.
2130 // We could also defer initialization and initialize on-demand
2131 // the first time we call ObjectSynchronizer::inflate().
2132 // Initialization would be protected - like so many things - by
2133 // the MonitorCache_lock.
2134
2135 void ObjectMonitor::Initialize() {
2136 assert(!InitDone, "invariant");
2137
2138 if (!os::is_MP()) {
2139 Knob_SpinLimit = 0;
2140 Knob_PreSpin = 0;
2141 Knob_FixedSpin = -1;
2142 }
2143
2144 if (UsePerfData) {
2145 EXCEPTION_MARK;
2146 #define NEWPERFCOUNTER(n) \
2147 { \
2148 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \
2149 CHECK); \
2150 }
2151 #define NEWPERFVARIABLE(n) \
2152 { \
2153 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \
2154 CHECK); \
2155 }
2156 NEWPERFCOUNTER(_sync_Inflations);
2157 NEWPERFCOUNTER(_sync_Deflations);
2158 NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2159 NEWPERFCOUNTER(_sync_FutileWakeups);
2160 NEWPERFCOUNTER(_sync_Parks);
2161 NEWPERFCOUNTER(_sync_Notifications);
2162 NEWPERFVARIABLE(_sync_MonExtant);
2163 #undef NEWPERFCOUNTER
2164 #undef NEWPERFVARIABLE
2165 }
2166
2167 _oop_storage = OopStorageSet::create_weak("ObjectSynchronizer Weak", mtSynchronizer);
2168
2169 DEBUG_ONLY(InitDone = true;)
2170 }
2171
2172 void ObjectMonitor::print_on(outputStream* st) const {
2173 // The minimal things to print for markWord printing, more can be added for debugging and logging.
2174 st->print("{contentions=0x%08x,waiters=0x%08x"
2175 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}",
2176 contentions(), waiters(), recursions(),
2177 p2i(owner()));
2178 }
2179 void ObjectMonitor::print() const { print_on(tty); }
2180
2181 #ifdef ASSERT
2182 // Print the ObjectMonitor like a debugger would:
2183 //
2184 // (ObjectMonitor) 0x00007fdfb6012e40 = {
2185 // _header = 0x0000000000000001
2186 // _object = 0x000000070ff45fd0
2187 // _pad_buf0 = {
2188 // [0] = '\0'
2189 // ...
2190 // [43] = '\0'
2191 // }
2192 // _owner = 0x0000000000000000
2193 // _previous_owner_tid = 0
2194 // _pad_buf1 = {
2195 // [0] = '\0'
2196 // ...
2197 // [47] = '\0'
2198 // }
2199 // _next_om = 0x0000000000000000
2200 // _recursions = 0
2201 // _EntryList = 0x0000000000000000
2202 // _cxq = 0x0000000000000000
2203 // _succ = 0x0000000000000000
2204 // _Responsible = 0x0000000000000000
2205 // _SpinDuration = 5000
2206 // _contentions = 0
2207 // _WaitSet = 0x0000700009756248
2208 // _waiters = 1
2209 // _WaitSetLock = 0
2210 // }
2211 //
2212 void ObjectMonitor::print_debug_style_on(outputStream* st) const {
2213 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this));
2214 st->print_cr(" _header = " INTPTR_FORMAT, header().value());
2215 st->print_cr(" _object = " INTPTR_FORMAT, p2i(object_peek()));
2216 st->print_cr(" _pad_buf0 = {");
2217 st->print_cr(" [0] = '\\0'");
2218 st->print_cr(" ...");
2219 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1);
2220 st->print_cr(" }");
2221 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(owner_raw()));
2222 st->print_cr(" _previous_owner_tid = " UINT64_FORMAT, _previous_owner_tid);
2223 st->print_cr(" _pad_buf1 = {");
2224 st->print_cr(" [0] = '\\0'");
2225 st->print_cr(" ...");
2226 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1);
2227 st->print_cr(" }");
2228 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(next_om()));
2229 st->print_cr(" _recursions = " INTX_FORMAT, _recursions);
2230 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList));
2231 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq));
2232 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ));
2233 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible));
2234 st->print_cr(" _SpinDuration = %d", _SpinDuration);
2235 st->print_cr(" _contentions = %d", contentions());
2236 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet));
2237 st->print_cr(" _waiters = %d", _waiters);
2238 st->print_cr(" _WaitSetLock = %d", _WaitSetLock);
2239 st->print_cr("}");
2240 }
2241 #endif