1 /*
2 * Copyright (c) 1998, 2018, 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 "jfr/jfrEvents.hpp"
28 #include "jfr/support/jfrThreadId.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "memory/resourceArea.hpp"
31 #include "oops/markOop.hpp"
32 #include "oops/oop.inline.hpp"
33 #include "runtime/atomic.hpp"
34 #include "runtime/handles.inline.hpp"
35 #include "runtime/interfaceSupport.inline.hpp"
36 #include "runtime/mutexLocker.hpp"
37 #include "runtime/objectMonitor.hpp"
38 #include "runtime/objectMonitor.inline.hpp"
39 #include "runtime/orderAccess.hpp"
40 #include "runtime/osThread.hpp"
41 #include "runtime/safepointMechanism.inline.hpp"
42 #include "runtime/sharedRuntime.hpp"
43 #include "runtime/stubRoutines.hpp"
44 #include "runtime/thread.inline.hpp"
45 #include "services/threadService.hpp"
46 #include "utilities/dtrace.hpp"
47 #include "utilities/macros.hpp"
48 #include "utilities/preserveException.hpp"
49 #if INCLUDE_JFR
50 #include "jfr/support/jfrFlush.hpp"
51 #endif
52
53 #ifdef DTRACE_ENABLED
54
55 // Only bother with this argument setup if dtrace is available
56 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
57
58
59 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
60 char* bytes = NULL; \
61 int len = 0; \
62 jlong jtid = SharedRuntime::get_java_tid(thread); \
63 Symbol* klassname = ((oop)obj)->klass()->name(); \
64 if (klassname != NULL) { \
65 bytes = (char*)klassname->bytes(); \
66 len = klassname->utf8_length(); \
67 }
68
69 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
70 { \
71 if (DTraceMonitorProbes) { \
72 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
73 HOTSPOT_MONITOR_WAIT(jtid, \
74 (monitor), bytes, len, (millis)); \
75 } \
76 }
77
78 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
79 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
80 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
81 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
82 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
83
84 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
85 { \
86 if (DTraceMonitorProbes) { \
87 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
88 HOTSPOT_MONITOR_##probe(jtid, \
89 (uintptr_t)(monitor), bytes, len); \
90 } \
91 }
92
93 #else // ndef DTRACE_ENABLED
94
95 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
96 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
97
98 #endif // ndef DTRACE_ENABLED
99
100 // Tunables ...
101 // The knob* variables are effectively final. Once set they should
102 // never be modified hence. Consider using __read_mostly with GCC.
103
104 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool -
105
106 static int Knob_Bonus = 100; // spin success bonus
107 static int Knob_BonusB = 100; // spin success bonus
108 static int Knob_Penalty = 200; // spin failure penalty
109 static int Knob_Poverty = 1000;
110 static int Knob_FixedSpin = 0;
111 static int Knob_OState = 3; // Spinner checks thread state of _owner
112 static int Knob_UsePause = 1;
113 static int Knob_ExitPolicy = 0;
114 static int Knob_PreSpin = 10; // 20-100 likely better
115 static int Knob_ResetEvent = 0;
116
117 static int Knob_FastHSSEC = 0;
118 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee
119 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline
120 static volatile int InitDone = 0;
121
122 // -----------------------------------------------------------------------------
123 // Theory of operations -- Monitors lists, thread residency, etc:
124 //
125 // * A thread acquires ownership of a monitor by successfully
126 // CAS()ing the _owner field from null to non-null.
127 //
128 // * Invariant: A thread appears on at most one monitor list --
129 // cxq, EntryList or WaitSet -- at any one time.
130 //
131 // * Contending threads "push" themselves onto the cxq with CAS
132 // and then spin/park.
133 //
134 // * After a contending thread eventually acquires the lock it must
135 // dequeue itself from either the EntryList or the cxq.
136 //
137 // * The exiting thread identifies and unparks an "heir presumptive"
138 // tentative successor thread on the EntryList. Critically, the
139 // exiting thread doesn't unlink the successor thread from the EntryList.
140 // After having been unparked, the wakee will recontend for ownership of
141 // the monitor. The successor (wakee) will either acquire the lock or
142 // re-park itself.
143 //
144 // Succession is provided for by a policy of competitive handoff.
145 // The exiting thread does _not_ grant or pass ownership to the
146 // successor thread. (This is also referred to as "handoff" succession").
147 // Instead the exiting thread releases ownership and possibly wakes
148 // a successor, so the successor can (re)compete for ownership of the lock.
149 // If the EntryList is empty but the cxq is populated the exiting
150 // thread will drain the cxq into the EntryList. It does so by
151 // by detaching the cxq (installing null with CAS) and folding
152 // the threads from the cxq into the EntryList. The EntryList is
153 // doubly linked, while the cxq is singly linked because of the
154 // CAS-based "push" used to enqueue recently arrived threads (RATs).
155 //
156 // * Concurrency invariants:
157 //
158 // -- only the monitor owner may access or mutate the EntryList.
159 // The mutex property of the monitor itself protects the EntryList
160 // from concurrent interference.
161 // -- Only the monitor owner may detach the cxq.
162 //
163 // * The monitor entry list operations avoid locks, but strictly speaking
164 // they're not lock-free. Enter is lock-free, exit is not.
165 // For a description of 'Methods and apparatus providing non-blocking access
166 // to a resource,' see U.S. Pat. No. 7844973.
167 //
168 // * The cxq can have multiple concurrent "pushers" but only one concurrent
169 // detaching thread. This mechanism is immune from the ABA corruption.
170 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
171 //
172 // * Taken together, the cxq and the EntryList constitute or form a
173 // single logical queue of threads stalled trying to acquire the lock.
174 // We use two distinct lists to improve the odds of a constant-time
175 // dequeue operation after acquisition (in the ::enter() epilogue) and
176 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm).
177 // A key desideratum is to minimize queue & monitor metadata manipulation
178 // that occurs while holding the monitor lock -- that is, we want to
179 // minimize monitor lock holds times. Note that even a small amount of
180 // fixed spinning will greatly reduce the # of enqueue-dequeue operations
181 // on EntryList|cxq. That is, spinning relieves contention on the "inner"
182 // locks and monitor metadata.
183 //
184 // Cxq points to the set of Recently Arrived Threads attempting entry.
185 // Because we push threads onto _cxq with CAS, the RATs must take the form of
186 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when
187 // the unlocking thread notices that EntryList is null but _cxq is != null.
188 //
189 // The EntryList is ordered by the prevailing queue discipline and
190 // can be organized in any convenient fashion, such as a doubly-linked list or
191 // a circular doubly-linked list. Critically, we want insert and delete operations
192 // to operate in constant-time. If we need a priority queue then something akin
193 // to Solaris' sleepq would work nicely. Viz.,
194 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
195 // Queue discipline is enforced at ::exit() time, when the unlocking thread
196 // drains the cxq into the EntryList, and orders or reorders the threads on the
197 // EntryList accordingly.
198 //
199 // Barring "lock barging", this mechanism provides fair cyclic ordering,
200 // somewhat similar to an elevator-scan.
201 //
202 // * The monitor synchronization subsystem avoids the use of native
203 // synchronization primitives except for the narrow platform-specific
204 // park-unpark abstraction. See the comments in os_solaris.cpp regarding
205 // the semantics of park-unpark. Put another way, this monitor implementation
206 // depends only on atomic operations and park-unpark. The monitor subsystem
207 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
208 // underlying OS manages the READY<->RUN transitions.
209 //
210 // * Waiting threads reside on the WaitSet list -- wait() puts
211 // the caller onto the WaitSet.
212 //
213 // * notify() or notifyAll() simply transfers threads from the WaitSet to
214 // either the EntryList or cxq. Subsequent exit() operations will
215 // unpark the notifyee. Unparking a notifee in notify() is inefficient -
216 // it's likely the notifyee would simply impale itself on the lock held
217 // by the notifier.
218 //
219 // * An interesting alternative is to encode cxq as (List,LockByte) where
220 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary
221 // variable, like _recursions, in the scheme. The threads or Events that form
222 // the list would have to be aligned in 256-byte addresses. A thread would
223 // try to acquire the lock or enqueue itself with CAS, but exiting threads
224 // could use a 1-0 protocol and simply STB to set the LockByte to 0.
225 // Note that is is *not* word-tearing, but it does presume that full-word
226 // CAS operations are coherent with intermix with STB operations. That's true
227 // on most common processors.
228 //
229 // * See also http://blogs.sun.com/dave
230
231
232 void* ObjectMonitor::operator new (size_t size) throw() {
233 return AllocateHeap(size, mtInternal);
234 }
235 void* ObjectMonitor::operator new[] (size_t size) throw() {
236 return operator new (size);
237 }
238 void ObjectMonitor::operator delete(void* p) {
239 FreeHeap(p);
240 }
241 void ObjectMonitor::operator delete[] (void *p) {
242 operator delete(p);
243 }
244
245 // -----------------------------------------------------------------------------
246 // Enter support
247
248 void ObjectMonitor::enter(TRAPS) {
249 // The following code is ordered to check the most common cases first
250 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
251 Thread * const Self = THREAD;
252
253 void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL);
254 if (cur == NULL) {
255 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
256 assert(_recursions == 0, "invariant");
257 assert(_owner == Self, "invariant");
258 return;
259 }
260
261 if (cur == Self) {
262 // TODO-FIXME: check for integer overflow! BUGID 6557169.
263 _recursions++;
264 return;
265 }
266
267 if (Self->is_lock_owned ((address)cur)) {
268 assert(_recursions == 0, "internal state error");
269 _recursions = 1;
270 // Commute owner from a thread-specific on-stack BasicLockObject address to
271 // a full-fledged "Thread *".
272 _owner = Self;
273 return;
274 }
275
276 // We've encountered genuine contention.
277 assert(Self->_Stalled == 0, "invariant");
278 Self->_Stalled = intptr_t(this);
279
280 // Try one round of spinning *before* enqueueing Self
281 // and before going through the awkward and expensive state
282 // transitions. The following spin is strictly optional ...
283 // Note that if we acquire the monitor from an initial spin
284 // we forgo posting JVMTI events and firing DTRACE probes.
285 if (TrySpin(Self) > 0) {
286 assert(_owner == Self, "invariant");
287 assert(_recursions == 0, "invariant");
288 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
289 Self->_Stalled = 0;
290 return;
291 }
292
293 assert(_owner != Self, "invariant");
294 assert(_succ != Self, "invariant");
295 assert(Self->is_Java_thread(), "invariant");
296 JavaThread * jt = (JavaThread *) Self;
297 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
298 assert(jt->thread_state() != _thread_blocked, "invariant");
299 assert(this->object() != NULL, "invariant");
300 assert(_count >= 0, "invariant");
301
302 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy().
303 // Ensure the object-monitor relationship remains stable while there's contention.
304 Atomic::inc(&_count);
305
306 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);)
307 EventJavaMonitorEnter event;
308 if (event.should_commit()) {
309 event.set_monitorClass(((oop)this->object())->klass());
310 event.set_address((uintptr_t)(this->object_addr()));
311 }
312
313 { // Change java thread status to indicate blocked on monitor enter.
314 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
315
316 Self->set_current_pending_monitor(this);
317
318 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
319 if (JvmtiExport::should_post_monitor_contended_enter()) {
320 JvmtiExport::post_monitor_contended_enter(jt, this);
321
322 // The current thread does not yet own the monitor and does not
323 // yet appear on any queues that would get it made the successor.
324 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
325 // handler cannot accidentally consume an unpark() meant for the
326 // ParkEvent associated with this ObjectMonitor.
327 }
328
329 OSThreadContendState osts(Self->osthread());
330 ThreadBlockInVM tbivm(jt);
331
332 // TODO-FIXME: change the following for(;;) loop to straight-line code.
333 for (;;) {
334 jt->set_suspend_equivalent();
335 // cleared by handle_special_suspend_equivalent_condition()
336 // or java_suspend_self()
337
338 EnterI(THREAD);
339
340 if (!ExitSuspendEquivalent(jt)) break;
341
342 // We have acquired the contended monitor, but while we were
343 // waiting another thread suspended us. We don't want to enter
344 // the monitor while suspended because that would surprise the
345 // thread that suspended us.
346 //
347 _recursions = 0;
348 _succ = NULL;
349 exit(false, Self);
350
351 jt->java_suspend_self();
352 }
353 Self->set_current_pending_monitor(NULL);
354
355 // We cleared the pending monitor info since we've just gotten past
356 // the enter-check-for-suspend dance and we now own the monitor free
357 // and clear, i.e., it is no longer pending. The ThreadBlockInVM
358 // destructor can go to a safepoint at the end of this block. If we
359 // do a thread dump during that safepoint, then this thread will show
360 // as having "-locked" the monitor, but the OS and java.lang.Thread
361 // states will still report that the thread is blocked trying to
362 // acquire it.
363 }
364
365 Atomic::dec(&_count);
366 assert(_count >= 0, "invariant");
367 Self->_Stalled = 0;
368
369 // Must either set _recursions = 0 or ASSERT _recursions == 0.
370 assert(_recursions == 0, "invariant");
371 assert(_owner == Self, "invariant");
372 assert(_succ != Self, "invariant");
373 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
374
375 // The thread -- now the owner -- is back in vm mode.
376 // Report the glorious news via TI,DTrace and jvmstat.
377 // The probe effect is non-trivial. All the reportage occurs
378 // while we hold the monitor, increasing the length of the critical
379 // section. Amdahl's parallel speedup law comes vividly into play.
380 //
381 // Another option might be to aggregate the events (thread local or
382 // per-monitor aggregation) and defer reporting until a more opportune
383 // time -- such as next time some thread encounters contention but has
384 // yet to acquire the lock. While spinning that thread could
385 // spinning we could increment JVMStat counters, etc.
386
387 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
388 if (JvmtiExport::should_post_monitor_contended_entered()) {
389 JvmtiExport::post_monitor_contended_entered(jt, this);
390
391 // The current thread already owns the monitor and is not going to
392 // call park() for the remainder of the monitor enter protocol. So
393 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
394 // event handler consumed an unpark() issued by the thread that
395 // just exited the monitor.
396 }
397 if (event.should_commit()) {
398 event.set_previousOwner((uintptr_t)_previous_owner_tid);
399 event.commit();
400 }
401 OM_PERFDATA_OP(ContendedLockAttempts, inc());
402 }
403
404 // Caveat: TryLock() is not necessarily serializing if it returns failure.
405 // Callers must compensate as needed.
406
407 int ObjectMonitor::TryLock(Thread * Self) {
408 void * own = _owner;
409 if (own != NULL) return 0;
410 if (Atomic::replace_if_null(Self, &_owner)) {
411 // Either guarantee _recursions == 0 or set _recursions = 0.
412 assert(_recursions == 0, "invariant");
413 assert(_owner == Self, "invariant");
414 return 1;
415 }
416 // The lock had been free momentarily, but we lost the race to the lock.
417 // Interference -- the CAS failed.
418 // We can either return -1 or retry.
419 // Retry doesn't make as much sense because the lock was just acquired.
420 return -1;
421 }
422
423 #define MAX_RECHECK_INTERVAL 1000
424
425 void ObjectMonitor::EnterI(TRAPS) {
426 Thread * const Self = THREAD;
427 assert(Self->is_Java_thread(), "invariant");
428 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant");
429
430 // Try the lock - TATAS
431 if (TryLock (Self) > 0) {
432 assert(_succ != Self, "invariant");
433 assert(_owner == Self, "invariant");
434 assert(_Responsible != Self, "invariant");
435 return;
436 }
437
438 DeferredInitialize();
439
440 // We try one round of spinning *before* enqueueing Self.
441 //
442 // If the _owner is ready but OFFPROC we could use a YieldTo()
443 // operation to donate the remainder of this thread's quantum
444 // to the owner. This has subtle but beneficial affinity
445 // effects.
446
447 if (TrySpin(Self) > 0) {
448 assert(_owner == Self, "invariant");
449 assert(_succ != Self, "invariant");
450 assert(_Responsible != Self, "invariant");
451 return;
452 }
453
454 // The Spin failed -- Enqueue and park the thread ...
455 assert(_succ != Self, "invariant");
456 assert(_owner != Self, "invariant");
457 assert(_Responsible != Self, "invariant");
458
459 // Enqueue "Self" on ObjectMonitor's _cxq.
460 //
461 // Node acts as a proxy for Self.
462 // As an aside, if were to ever rewrite the synchronization code mostly
463 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
464 // Java objects. This would avoid awkward lifecycle and liveness issues,
465 // as well as eliminate a subset of ABA issues.
466 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
467
468 ObjectWaiter node(Self);
469 Self->_ParkEvent->reset();
470 node._prev = (ObjectWaiter *) 0xBAD;
471 node.TState = ObjectWaiter::TS_CXQ;
472
473 // Push "Self" onto the front of the _cxq.
474 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
475 // Note that spinning tends to reduce the rate at which threads
476 // enqueue and dequeue on EntryList|cxq.
477 ObjectWaiter * nxt;
478 for (;;) {
479 node._next = nxt = _cxq;
480 if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break;
481
482 // Interference - the CAS failed because _cxq changed. Just retry.
483 // As an optional optimization we retry the lock.
484 if (TryLock (Self) > 0) {
485 assert(_succ != Self, "invariant");
486 assert(_owner == Self, "invariant");
487 assert(_Responsible != Self, "invariant");
488 return;
489 }
490 }
491
492 // Check for cxq|EntryList edge transition to non-null. This indicates
493 // the onset of contention. While contention persists exiting threads
494 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit
495 // operations revert to the faster 1-0 mode. This enter operation may interleave
496 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
497 // arrange for one of the contending thread to use a timed park() operations
498 // to detect and recover from the race. (Stranding is form of progress failure
499 // where the monitor is unlocked but all the contending threads remain parked).
500 // That is, at least one of the contended threads will periodically poll _owner.
501 // One of the contending threads will become the designated "Responsible" thread.
502 // The Responsible thread uses a timed park instead of a normal indefinite park
503 // operation -- it periodically wakes and checks for and recovers from potential
504 // strandings admitted by 1-0 exit operations. We need at most one Responsible
505 // thread per-monitor at any given moment. Only threads on cxq|EntryList may
506 // be responsible for a monitor.
507 //
508 // Currently, one of the contended threads takes on the added role of "Responsible".
509 // A viable alternative would be to use a dedicated "stranding checker" thread
510 // that periodically iterated over all the threads (or active monitors) and unparked
511 // successors where there was risk of stranding. This would help eliminate the
512 // timer scalability issues we see on some platforms as we'd only have one thread
513 // -- the checker -- parked on a timer.
514
515 if (nxt == NULL && _EntryList == NULL) {
516 // Try to assume the role of responsible thread for the monitor.
517 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self }
518 Atomic::replace_if_null(Self, &_Responsible);
519 }
520
521 // The lock might have been released while this thread was occupied queueing
522 // itself onto _cxq. To close the race and avoid "stranding" and
523 // progress-liveness failure we must resample-retry _owner before parking.
524 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
525 // In this case the ST-MEMBAR is accomplished with CAS().
526 //
527 // TODO: Defer all thread state transitions until park-time.
528 // Since state transitions are heavy and inefficient we'd like
529 // to defer the state transitions until absolutely necessary,
530 // and in doing so avoid some transitions ...
531
532 int nWakeups = 0;
533 int recheckInterval = 1;
534
535 for (;;) {
536
537 if (TryLock(Self) > 0) break;
538 assert(_owner != Self, "invariant");
539
540 // park self
541 if (_Responsible == Self) {
542 Self->_ParkEvent->park((jlong) recheckInterval);
543 // Increase the recheckInterval, but clamp the value.
544 recheckInterval *= 8;
545 if (recheckInterval > MAX_RECHECK_INTERVAL) {
546 recheckInterval = MAX_RECHECK_INTERVAL;
547 }
548 } else {
549 Self->_ParkEvent->park();
550 }
551
552 if (TryLock(Self) > 0) break;
553
554 // The lock is still contested.
555 // Keep a tally of the # of futile wakeups.
556 // Note that the counter is not protected by a lock or updated by atomics.
557 // That is by design - we trade "lossy" counters which are exposed to
558 // races during updates for a lower probe effect.
559
560 // This PerfData object can be used in parallel with a safepoint.
561 // See the work around in PerfDataManager::destroy().
562 OM_PERFDATA_OP(FutileWakeups, inc());
563 ++nWakeups;
564
565 // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
566 // We can defer clearing _succ until after the spin completes
567 // TrySpin() must tolerate being called with _succ == Self.
568 // Try yet another round of adaptive spinning.
569 if (TrySpin(Self) > 0) break;
570
571 // We can find that we were unpark()ed and redesignated _succ while
572 // we were spinning. That's harmless. If we iterate and call park(),
573 // park() will consume the event and return immediately and we'll
574 // just spin again. This pattern can repeat, leaving _succ to simply
575 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks().
576 // Alternately, we can sample fired() here, and if set, forgo spinning
577 // in the next iteration.
578
579 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
580 Self->_ParkEvent->reset();
581 OrderAccess::fence();
582 }
583 if (_succ == Self) _succ = NULL;
584
585 // Invariant: after clearing _succ a thread *must* retry _owner before parking.
586 OrderAccess::fence();
587 }
588
589 // Egress :
590 // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
591 // Normally we'll find Self on the EntryList .
592 // From the perspective of the lock owner (this thread), the
593 // EntryList is stable and cxq is prepend-only.
594 // The head of cxq is volatile but the interior is stable.
595 // In addition, Self.TState is stable.
596
597 assert(_owner == Self, "invariant");
598 assert(object() != NULL, "invariant");
599 // I'd like to write:
600 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
601 // but as we're at a safepoint that's not safe.
602
603 UnlinkAfterAcquire(Self, &node);
604 if (_succ == Self) _succ = NULL;
605
606 assert(_succ != Self, "invariant");
607 if (_Responsible == Self) {
608 _Responsible = NULL;
609 OrderAccess::fence(); // Dekker pivot-point
610
611 // We may leave threads on cxq|EntryList without a designated
612 // "Responsible" thread. This is benign. When this thread subsequently
613 // exits the monitor it can "see" such preexisting "old" threads --
614 // threads that arrived on the cxq|EntryList before the fence, above --
615 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads
616 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
617 // non-null and elect a new "Responsible" timer thread.
618 //
619 // This thread executes:
620 // ST Responsible=null; MEMBAR (in enter epilogue - here)
621 // LD cxq|EntryList (in subsequent exit)
622 //
623 // Entering threads in the slow/contended path execute:
624 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
625 // The (ST cxq; MEMBAR) is accomplished with CAS().
626 //
627 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
628 // exit operation from floating above the ST Responsible=null.
629 }
630
631 // We've acquired ownership with CAS().
632 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
633 // But since the CAS() this thread may have also stored into _succ,
634 // EntryList, cxq or Responsible. These meta-data updates must be
635 // visible __before this thread subsequently drops the lock.
636 // Consider what could occur if we didn't enforce this constraint --
637 // STs to monitor meta-data and user-data could reorder with (become
638 // visible after) the ST in exit that drops ownership of the lock.
639 // Some other thread could then acquire the lock, but observe inconsistent
640 // or old monitor meta-data and heap data. That violates the JMM.
641 // To that end, the 1-0 exit() operation must have at least STST|LDST
642 // "release" barrier semantics. Specifically, there must be at least a
643 // STST|LDST barrier in exit() before the ST of null into _owner that drops
644 // the lock. The barrier ensures that changes to monitor meta-data and data
645 // protected by the lock will be visible before we release the lock, and
646 // therefore before some other thread (CPU) has a chance to acquire the lock.
647 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
648 //
649 // Critically, any prior STs to _succ or EntryList must be visible before
650 // the ST of null into _owner in the *subsequent* (following) corresponding
651 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily
652 // execute a serializing instruction.
653
654 return;
655 }
656
657 // ReenterI() is a specialized inline form of the latter half of the
658 // contended slow-path from EnterI(). We use ReenterI() only for
659 // monitor reentry in wait().
660 //
661 // In the future we should reconcile EnterI() and ReenterI().
662
663 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) {
664 assert(Self != NULL, "invariant");
665 assert(SelfNode != NULL, "invariant");
666 assert(SelfNode->_thread == Self, "invariant");
667 assert(_waiters > 0, "invariant");
668 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
669 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
670 JavaThread * jt = (JavaThread *) Self;
671
672 int nWakeups = 0;
673 for (;;) {
674 ObjectWaiter::TStates v = SelfNode->TState;
675 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
676 assert(_owner != Self, "invariant");
677
678 if (TryLock(Self) > 0) break;
679 if (TrySpin(Self) > 0) break;
680
681 // State transition wrappers around park() ...
682 // ReenterI() wisely defers state transitions until
683 // it's clear we must park the thread.
684 {
685 OSThreadContendState osts(Self->osthread());
686 ThreadBlockInVM tbivm(jt);
687
688 // cleared by handle_special_suspend_equivalent_condition()
689 // or java_suspend_self()
690 jt->set_suspend_equivalent();
691 Self->_ParkEvent->park();
692
693 // were we externally suspended while we were waiting?
694 for (;;) {
695 if (!ExitSuspendEquivalent(jt)) break;
696 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
697 jt->java_suspend_self();
698 jt->set_suspend_equivalent();
699 }
700 }
701
702 // Try again, but just so we distinguish between futile wakeups and
703 // successful wakeups. The following test isn't algorithmically
704 // necessary, but it helps us maintain sensible statistics.
705 if (TryLock(Self) > 0) break;
706
707 // The lock is still contested.
708 // Keep a tally of the # of futile wakeups.
709 // Note that the counter is not protected by a lock or updated by atomics.
710 // That is by design - we trade "lossy" counters which are exposed to
711 // races during updates for a lower probe effect.
712 ++nWakeups;
713
714 // Assuming this is not a spurious wakeup we'll normally
715 // find that _succ == Self.
716 if (_succ == Self) _succ = NULL;
717
718 // Invariant: after clearing _succ a contending thread
719 // *must* retry _owner before parking.
720 OrderAccess::fence();
721
722 // This PerfData object can be used in parallel with a safepoint.
723 // See the work around in PerfDataManager::destroy().
724 OM_PERFDATA_OP(FutileWakeups, inc());
725 }
726
727 // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
728 // Normally we'll find Self on the EntryList.
729 // Unlinking from the EntryList is constant-time and atomic-free.
730 // From the perspective of the lock owner (this thread), the
731 // EntryList is stable and cxq is prepend-only.
732 // The head of cxq is volatile but the interior is stable.
733 // In addition, Self.TState is stable.
734
735 assert(_owner == Self, "invariant");
736 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
737 UnlinkAfterAcquire(Self, SelfNode);
738 if (_succ == Self) _succ = NULL;
739 assert(_succ != Self, "invariant");
740 SelfNode->TState = ObjectWaiter::TS_RUN;
741 OrderAccess::fence(); // see comments at the end of EnterI()
742 }
743
744 // By convention we unlink a contending thread from EntryList|cxq immediately
745 // after the thread acquires the lock in ::enter(). Equally, we could defer
746 // unlinking the thread until ::exit()-time.
747
748 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) {
749 assert(_owner == Self, "invariant");
750 assert(SelfNode->_thread == Self, "invariant");
751
752 if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
753 // Normal case: remove Self from the DLL EntryList .
754 // This is a constant-time operation.
755 ObjectWaiter * nxt = SelfNode->_next;
756 ObjectWaiter * prv = SelfNode->_prev;
757 if (nxt != NULL) nxt->_prev = prv;
758 if (prv != NULL) prv->_next = nxt;
759 if (SelfNode == _EntryList) _EntryList = nxt;
760 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
761 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
762 } else {
763 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant");
764 // Inopportune interleaving -- Self is still on the cxq.
765 // This usually means the enqueue of self raced an exiting thread.
766 // Normally we'll find Self near the front of the cxq, so
767 // dequeueing is typically fast. If needbe we can accelerate
768 // this with some MCS/CHL-like bidirectional list hints and advisory
769 // back-links so dequeueing from the interior will normally operate
770 // in constant-time.
771 // Dequeue Self from either the head (with CAS) or from the interior
772 // with a linear-time scan and normal non-atomic memory operations.
773 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
774 // and then unlink Self from EntryList. We have to drain eventually,
775 // so it might as well be now.
776
777 ObjectWaiter * v = _cxq;
778 assert(v != NULL, "invariant");
779 if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) {
780 // The CAS above can fail from interference IFF a "RAT" arrived.
781 // In that case Self must be in the interior and can no longer be
782 // at the head of cxq.
783 if (v == SelfNode) {
784 assert(_cxq != v, "invariant");
785 v = _cxq; // CAS above failed - start scan at head of list
786 }
787 ObjectWaiter * p;
788 ObjectWaiter * q = NULL;
789 for (p = v; p != NULL && p != SelfNode; p = p->_next) {
790 q = p;
791 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
792 }
793 assert(v != SelfNode, "invariant");
794 assert(p == SelfNode, "Node not found on cxq");
795 assert(p != _cxq, "invariant");
796 assert(q != NULL, "invariant");
797 assert(q->_next == p, "invariant");
798 q->_next = p->_next;
799 }
800 }
801
802 #ifdef ASSERT
803 // Diagnostic hygiene ...
804 SelfNode->_prev = (ObjectWaiter *) 0xBAD;
805 SelfNode->_next = (ObjectWaiter *) 0xBAD;
806 SelfNode->TState = ObjectWaiter::TS_RUN;
807 #endif
808 }
809
810 // -----------------------------------------------------------------------------
811 // Exit support
812 //
813 // exit()
814 // ~~~~~~
815 // Note that the collector can't reclaim the objectMonitor or deflate
816 // the object out from underneath the thread calling ::exit() as the
817 // thread calling ::exit() never transitions to a stable state.
818 // This inhibits GC, which in turn inhibits asynchronous (and
819 // inopportune) reclamation of "this".
820 //
821 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
822 // There's one exception to the claim above, however. EnterI() can call
823 // exit() to drop a lock if the acquirer has been externally suspended.
824 // In that case exit() is called with _thread_state as _thread_blocked,
825 // but the monitor's _count field is > 0, which inhibits reclamation.
826 //
827 // 1-0 exit
828 // ~~~~~~~~
829 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
830 // the fast-path operators have been optimized so the common ::exit()
831 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
832 // The code emitted by fast_unlock() elides the usual MEMBAR. This
833 // greatly improves latency -- MEMBAR and CAS having considerable local
834 // latency on modern processors -- but at the cost of "stranding". Absent the
835 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
836 // ::enter() path, resulting in the entering thread being stranding
837 // and a progress-liveness failure. Stranding is extremely rare.
838 // We use timers (timed park operations) & periodic polling to detect
839 // and recover from stranding. Potentially stranded threads periodically
840 // wake up and poll the lock. See the usage of the _Responsible variable.
841 //
842 // The CAS() in enter provides for safety and exclusion, while the CAS or
843 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking
844 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
845 // We detect and recover from stranding with timers.
846 //
847 // If a thread transiently strands it'll park until (a) another
848 // thread acquires the lock and then drops the lock, at which time the
849 // exiting thread will notice and unpark the stranded thread, or, (b)
850 // the timer expires. If the lock is high traffic then the stranding latency
851 // will be low due to (a). If the lock is low traffic then the odds of
852 // stranding are lower, although the worst-case stranding latency
853 // is longer. Critically, we don't want to put excessive load in the
854 // platform's timer subsystem. We want to minimize both the timer injection
855 // rate (timers created/sec) as well as the number of timers active at
856 // any one time. (more precisely, we want to minimize timer-seconds, which is
857 // the integral of the # of active timers at any instant over time).
858 // Both impinge on OS scalability. Given that, at most one thread parked on
859 // a monitor will use a timer.
860 //
861 // There is also the risk of a futile wake-up. If we drop the lock
862 // another thread can reacquire the lock immediately, and we can
863 // then wake a thread unnecessarily. This is benign, and we've
864 // structured the code so the windows are short and the frequency
865 // of such futile wakups is low.
866
867 void ObjectMonitor::exit(bool not_suspended, TRAPS) {
868 Thread * const Self = THREAD;
869 if (THREAD != _owner) {
870 if (THREAD->is_lock_owned((address) _owner)) {
871 // Transmute _owner from a BasicLock pointer to a Thread address.
872 // We don't need to hold _mutex for this transition.
873 // Non-null to Non-null is safe as long as all readers can
874 // tolerate either flavor.
875 assert(_recursions == 0, "invariant");
876 _owner = THREAD;
877 _recursions = 0;
878 } else {
879 // Apparent unbalanced locking ...
880 // Naively we'd like to throw IllegalMonitorStateException.
881 // As a practical matter we can neither allocate nor throw an
882 // exception as ::exit() can be called from leaf routines.
883 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
884 // Upon deeper reflection, however, in a properly run JVM the only
885 // way we should encounter this situation is in the presence of
886 // unbalanced JNI locking. TODO: CheckJNICalls.
887 // See also: CR4414101
888 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking");
889 return;
890 }
891 }
892
893 if (_recursions != 0) {
894 _recursions--; // this is simple recursive enter
895 return;
896 }
897
898 // Invariant: after setting Responsible=null an thread must execute
899 // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
900 _Responsible = NULL;
901
902 #if INCLUDE_JFR
903 // get the owner's thread id for the MonitorEnter event
904 // if it is enabled and the thread isn't suspended
905 if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
906 _previous_owner_tid = JFR_THREAD_ID(Self);
907 }
908 #endif
909
910 for (;;) {
911 assert(THREAD == _owner, "invariant");
912
913 if (Knob_ExitPolicy == 0) {
914 // release semantics: prior loads and stores from within the critical section
915 // must not float (reorder) past the following store that drops the lock.
916 // On SPARC that requires MEMBAR #loadstore|#storestore.
917 // But of course in TSO #loadstore|#storestore is not required.
918 // I'd like to write one of the following:
919 // A. OrderAccess::release() ; _owner = NULL
920 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
921 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
922 // store into a _dummy variable. That store is not needed, but can result
923 // in massive wasteful coherency traffic on classic SMP systems.
924 // Instead, I use release_store(), which is implemented as just a simple
925 // ST on x64, x86 and SPARC.
926 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock
927 OrderAccess::storeload(); // See if we need to wake a successor
928 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
929 return;
930 }
931 // Other threads are blocked trying to acquire the lock.
932
933 // Normally the exiting thread is responsible for ensuring succession,
934 // but if other successors are ready or other entering threads are spinning
935 // then this thread can simply store NULL into _owner and exit without
936 // waking a successor. The existence of spinners or ready successors
937 // guarantees proper succession (liveness). Responsibility passes to the
938 // ready or running successors. The exiting thread delegates the duty.
939 // More precisely, if a successor already exists this thread is absolved
940 // of the responsibility of waking (unparking) one.
941 //
942 // The _succ variable is critical to reducing futile wakeup frequency.
943 // _succ identifies the "heir presumptive" thread that has been made
944 // ready (unparked) but that has not yet run. We need only one such
945 // successor thread to guarantee progress.
946 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
947 // section 3.3 "Futile Wakeup Throttling" for details.
948 //
949 // Note that spinners in Enter() also set _succ non-null.
950 // In the current implementation spinners opportunistically set
951 // _succ so that exiting threads might avoid waking a successor.
952 // Another less appealing alternative would be for the exiting thread
953 // to drop the lock and then spin briefly to see if a spinner managed
954 // to acquire the lock. If so, the exiting thread could exit
955 // immediately without waking a successor, otherwise the exiting
956 // thread would need to dequeue and wake a successor.
957 // (Note that we'd need to make the post-drop spin short, but no
958 // shorter than the worst-case round-trip cache-line migration time.
959 // The dropped lock needs to become visible to the spinner, and then
960 // the acquisition of the lock by the spinner must become visible to
961 // the exiting thread).
962
963 // It appears that an heir-presumptive (successor) must be made ready.
964 // Only the current lock owner can manipulate the EntryList or
965 // drain _cxq, so we need to reacquire the lock. If we fail
966 // to reacquire the lock the responsibility for ensuring succession
967 // falls to the new owner.
968 //
969 if (!Atomic::replace_if_null(THREAD, &_owner)) {
970 return;
971 }
972 } else {
973 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
974 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock
975 OrderAccess::storeload();
976 // Ratify the previously observed values.
977 if (_cxq == NULL || _succ != NULL) {
978 return;
979 }
980
981 // inopportune interleaving -- the exiting thread (this thread)
982 // in the fast-exit path raced an entering thread in the slow-enter
983 // path.
984 // We have two choices:
985 // A. Try to reacquire the lock.
986 // If the CAS() fails return immediately, otherwise
987 // we either restart/rerun the exit operation, or simply
988 // fall-through into the code below which wakes a successor.
989 // B. If the elements forming the EntryList|cxq are TSM
990 // we could simply unpark() the lead thread and return
991 // without having set _succ.
992 if (!Atomic::replace_if_null(THREAD, &_owner)) {
993 return;
994 }
995 }
996 }
997
998 guarantee(_owner == THREAD, "invariant");
999
1000 ObjectWaiter * w = NULL;
1001 int QMode = Knob_QMode;
1002
1003 if (QMode == 2 && _cxq != NULL) {
1004 // QMode == 2 : cxq has precedence over EntryList.
1005 // Try to directly wake a successor from the cxq.
1006 // If successful, the successor will need to unlink itself from cxq.
1007 w = _cxq;
1008 assert(w != NULL, "invariant");
1009 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant");
1010 ExitEpilog(Self, w);
1011 return;
1012 }
1013
1014 if (QMode == 3 && _cxq != NULL) {
1015 // Aggressively drain cxq into EntryList at the first opportunity.
1016 // This policy ensure that recently-run threads live at the head of EntryList.
1017 // Drain _cxq into EntryList - bulk transfer.
1018 // First, detach _cxq.
1019 // The following loop is tantamount to: w = swap(&cxq, NULL)
1020 w = _cxq;
1021 for (;;) {
1022 assert(w != NULL, "Invariant");
1023 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1024 if (u == w) break;
1025 w = u;
1026 }
1027 assert(w != NULL, "invariant");
1028
1029 ObjectWaiter * q = NULL;
1030 ObjectWaiter * p;
1031 for (p = w; p != NULL; p = p->_next) {
1032 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1033 p->TState = ObjectWaiter::TS_ENTER;
1034 p->_prev = q;
1035 q = p;
1036 }
1037
1038 // Append the RATs to the EntryList
1039 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1040 ObjectWaiter * Tail;
1041 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL;
1042 Tail = Tail->_next)
1043 /* empty */;
1044 if (Tail == NULL) {
1045 _EntryList = w;
1046 } else {
1047 Tail->_next = w;
1048 w->_prev = Tail;
1049 }
1050
1051 // Fall thru into code that tries to wake a successor from EntryList
1052 }
1053
1054 if (QMode == 4 && _cxq != NULL) {
1055 // Aggressively drain cxq into EntryList at the first opportunity.
1056 // This policy ensure that recently-run threads live at the head of EntryList.
1057
1058 // Drain _cxq into EntryList - bulk transfer.
1059 // First, detach _cxq.
1060 // The following loop is tantamount to: w = swap(&cxq, NULL)
1061 w = _cxq;
1062 for (;;) {
1063 assert(w != NULL, "Invariant");
1064 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1065 if (u == w) break;
1066 w = u;
1067 }
1068 assert(w != NULL, "invariant");
1069
1070 ObjectWaiter * q = NULL;
1071 ObjectWaiter * p;
1072 for (p = w; p != NULL; p = p->_next) {
1073 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1074 p->TState = ObjectWaiter::TS_ENTER;
1075 p->_prev = q;
1076 q = p;
1077 }
1078
1079 // Prepend the RATs to the EntryList
1080 if (_EntryList != NULL) {
1081 q->_next = _EntryList;
1082 _EntryList->_prev = q;
1083 }
1084 _EntryList = w;
1085
1086 // Fall thru into code that tries to wake a successor from EntryList
1087 }
1088
1089 w = _EntryList;
1090 if (w != NULL) {
1091 // I'd like to write: guarantee (w->_thread != Self).
1092 // But in practice an exiting thread may find itself on the EntryList.
1093 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and
1094 // then calls exit(). Exit release the lock by setting O._owner to NULL.
1095 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The
1096 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1097 // release the lock "O". T2 resumes immediately after the ST of null into
1098 // _owner, above. T2 notices that the EntryList is populated, so it
1099 // reacquires the lock and then finds itself on the EntryList.
1100 // Given all that, we have to tolerate the circumstance where "w" is
1101 // associated with Self.
1102 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1103 ExitEpilog(Self, w);
1104 return;
1105 }
1106
1107 // If we find that both _cxq and EntryList are null then just
1108 // re-run the exit protocol from the top.
1109 w = _cxq;
1110 if (w == NULL) continue;
1111
1112 // Drain _cxq into EntryList - bulk transfer.
1113 // First, detach _cxq.
1114 // The following loop is tantamount to: w = swap(&cxq, NULL)
1115 for (;;) {
1116 assert(w != NULL, "Invariant");
1117 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1118 if (u == w) break;
1119 w = u;
1120 }
1121
1122 assert(w != NULL, "invariant");
1123 assert(_EntryList == NULL, "invariant");
1124
1125 // Convert the LIFO SLL anchored by _cxq into a DLL.
1126 // The list reorganization step operates in O(LENGTH(w)) time.
1127 // It's critical that this step operate quickly as
1128 // "Self" still holds the outer-lock, restricting parallelism
1129 // and effectively lengthening the critical section.
1130 // Invariant: s chases t chases u.
1131 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1132 // we have faster access to the tail.
1133
1134 if (QMode == 1) {
1135 // QMode == 1 : drain cxq to EntryList, reversing order
1136 // We also reverse the order of the list.
1137 ObjectWaiter * s = NULL;
1138 ObjectWaiter * t = w;
1139 ObjectWaiter * u = NULL;
1140 while (t != NULL) {
1141 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant");
1142 t->TState = ObjectWaiter::TS_ENTER;
1143 u = t->_next;
1144 t->_prev = u;
1145 t->_next = s;
1146 s = t;
1147 t = u;
1148 }
1149 _EntryList = s;
1150 assert(s != NULL, "invariant");
1151 } else {
1152 // QMode == 0 or QMode == 2
1153 _EntryList = w;
1154 ObjectWaiter * q = NULL;
1155 ObjectWaiter * p;
1156 for (p = w; p != NULL; p = p->_next) {
1157 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1158 p->TState = ObjectWaiter::TS_ENTER;
1159 p->_prev = q;
1160 q = p;
1161 }
1162 }
1163
1164 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1165 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1166
1167 // See if we can abdicate to a spinner instead of waking a thread.
1168 // A primary goal of the implementation is to reduce the
1169 // context-switch rate.
1170 if (_succ != NULL) continue;
1171
1172 w = _EntryList;
1173 if (w != NULL) {
1174 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1175 ExitEpilog(Self, w);
1176 return;
1177 }
1178 }
1179 }
1180
1181 // ExitSuspendEquivalent:
1182 // A faster alternate to handle_special_suspend_equivalent_condition()
1183 //
1184 // handle_special_suspend_equivalent_condition() unconditionally
1185 // acquires the SR_lock. On some platforms uncontended MutexLocker()
1186 // operations have high latency. Note that in ::enter() we call HSSEC
1187 // while holding the monitor, so we effectively lengthen the critical sections.
1188 //
1189 // There are a number of possible solutions:
1190 //
1191 // A. To ameliorate the problem we might also defer state transitions
1192 // to as late as possible -- just prior to parking.
1193 // Given that, we'd call HSSEC after having returned from park(),
1194 // but before attempting to acquire the monitor. This is only a
1195 // partial solution. It avoids calling HSSEC while holding the
1196 // monitor (good), but it still increases successor reacquisition latency --
1197 // the interval between unparking a successor and the time the successor
1198 // resumes and retries the lock. See ReenterI(), which defers state transitions.
1199 // If we use this technique we can also avoid EnterI()-exit() loop
1200 // in ::enter() where we iteratively drop the lock and then attempt
1201 // to reacquire it after suspending.
1202 //
1203 // B. In the future we might fold all the suspend bits into a
1204 // composite per-thread suspend flag and then update it with CAS().
1205 // Alternately, a Dekker-like mechanism with multiple variables
1206 // would suffice:
1207 // ST Self->_suspend_equivalent = false
1208 // MEMBAR
1209 // LD Self_>_suspend_flags
1210 //
1211 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem
1212 // with a more efficient implementation, the need to use "FastHSSEC" has
1213 // decreased. - Dave
1214
1215
1216 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) {
1217 const int Mode = Knob_FastHSSEC;
1218 if (Mode && !jSelf->is_external_suspend()) {
1219 assert(jSelf->is_suspend_equivalent(), "invariant");
1220 jSelf->clear_suspend_equivalent();
1221 if (2 == Mode) OrderAccess::storeload();
1222 if (!jSelf->is_external_suspend()) return false;
1223 // We raced a suspension -- fall thru into the slow path
1224 jSelf->set_suspend_equivalent();
1225 }
1226 return jSelf->handle_special_suspend_equivalent_condition();
1227 }
1228
1229
1230 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) {
1231 assert(_owner == Self, "invariant");
1232
1233 // Exit protocol:
1234 // 1. ST _succ = wakee
1235 // 2. membar #loadstore|#storestore;
1236 // 2. ST _owner = NULL
1237 // 3. unpark(wakee)
1238
1239 _succ = Wakee->_thread;
1240 ParkEvent * Trigger = Wakee->_event;
1241
1242 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1243 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1244 // out-of-scope (non-extant).
1245 Wakee = NULL;
1246
1247 // Drop the lock
1248 OrderAccess::release_store(&_owner, (void*)NULL);
1249 OrderAccess::fence(); // ST _owner vs LD in unpark()
1250
1251 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1252 Trigger->unpark();
1253
1254 // Maintain stats and report events to JVMTI
1255 OM_PERFDATA_OP(Parks, inc());
1256 }
1257
1258
1259 // -----------------------------------------------------------------------------
1260 // Class Loader deadlock handling.
1261 //
1262 // complete_exit exits a lock returning recursion count
1263 // complete_exit/reenter operate as a wait without waiting
1264 // complete_exit requires an inflated monitor
1265 // The _owner field is not always the Thread addr even with an
1266 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1267 // thread due to contention.
1268 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1269 Thread * const Self = THREAD;
1270 assert(Self->is_Java_thread(), "Must be Java thread!");
1271 JavaThread *jt = (JavaThread *)THREAD;
1272
1273 DeferredInitialize();
1274
1275 if (THREAD != _owner) {
1276 if (THREAD->is_lock_owned ((address)_owner)) {
1277 assert(_recursions == 0, "internal state error");
1278 _owner = THREAD; // Convert from basiclock addr to Thread addr
1279 _recursions = 0;
1280 }
1281 }
1282
1283 guarantee(Self == _owner, "complete_exit not owner");
1284 intptr_t save = _recursions; // record the old recursion count
1285 _recursions = 0; // set the recursion level to be 0
1286 exit(true, Self); // exit the monitor
1287 guarantee(_owner != Self, "invariant");
1288 return save;
1289 }
1290
1291 // reenter() enters a lock and sets recursion count
1292 // complete_exit/reenter operate as a wait without waiting
1293 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1294 Thread * const Self = THREAD;
1295 assert(Self->is_Java_thread(), "Must be Java thread!");
1296 JavaThread *jt = (JavaThread *)THREAD;
1297
1298 guarantee(_owner != Self, "reenter already owner");
1299 enter(THREAD); // enter the monitor
1300 guarantee(_recursions == 0, "reenter recursion");
1301 _recursions = recursions;
1302 return;
1303 }
1304
1305
1306 // -----------------------------------------------------------------------------
1307 // A macro is used below because there may already be a pending
1308 // exception which should not abort the execution of the routines
1309 // which use this (which is why we don't put this into check_slow and
1310 // call it with a CHECK argument).
1311
1312 #define CHECK_OWNER() \
1313 do { \
1314 if (THREAD != _owner) { \
1315 if (THREAD->is_lock_owned((address) _owner)) { \
1316 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \
1317 _recursions = 0; \
1318 } else { \
1319 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \
1320 } \
1321 } \
1322 } while (false)
1323
1324 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception.
1325 // TODO-FIXME: remove check_slow() -- it's likely dead.
1326
1327 void ObjectMonitor::check_slow(TRAPS) {
1328 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1329 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1330 }
1331
1332 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1333 ObjectMonitor* monitor,
1334 jlong notifier_tid,
1335 jlong timeout,
1336 bool timedout) {
1337 assert(event != NULL, "invariant");
1338 assert(monitor != NULL, "invariant");
1339 event->set_monitorClass(((oop)monitor->object())->klass());
1340 event->set_timeout(timeout);
1341 event->set_address((uintptr_t)monitor->object_addr());
1342 event->set_notifier(notifier_tid);
1343 event->set_timedOut(timedout);
1344 event->commit();
1345 }
1346
1347 // -----------------------------------------------------------------------------
1348 // Wait/Notify/NotifyAll
1349 //
1350 // Note: a subset of changes to ObjectMonitor::wait()
1351 // will need to be replicated in complete_exit
1352 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1353 Thread * const Self = THREAD;
1354 assert(Self->is_Java_thread(), "Must be Java thread!");
1355 JavaThread *jt = (JavaThread *)THREAD;
1356
1357 DeferredInitialize();
1358
1359 // Throw IMSX or IEX.
1360 CHECK_OWNER();
1361
1362 EventJavaMonitorWait event;
1363
1364 // check for a pending interrupt
1365 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1366 // post monitor waited event. Note that this is past-tense, we are done waiting.
1367 if (JvmtiExport::should_post_monitor_waited()) {
1368 // Note: 'false' parameter is passed here because the
1369 // wait was not timed out due to thread interrupt.
1370 JvmtiExport::post_monitor_waited(jt, this, false);
1371
1372 // In this short circuit of the monitor wait protocol, the
1373 // current thread never drops ownership of the monitor and
1374 // never gets added to the wait queue so the current thread
1375 // cannot be made the successor. This means that the
1376 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1377 // consume an unpark() meant for the ParkEvent associated with
1378 // this ObjectMonitor.
1379 }
1380 if (event.should_commit()) {
1381 post_monitor_wait_event(&event, this, 0, millis, false);
1382 }
1383 THROW(vmSymbols::java_lang_InterruptedException());
1384 return;
1385 }
1386
1387 assert(Self->_Stalled == 0, "invariant");
1388 Self->_Stalled = intptr_t(this);
1389 jt->set_current_waiting_monitor(this);
1390
1391 // create a node to be put into the queue
1392 // Critically, after we reset() the event but prior to park(), we must check
1393 // for a pending interrupt.
1394 ObjectWaiter node(Self);
1395 node.TState = ObjectWaiter::TS_WAIT;
1396 Self->_ParkEvent->reset();
1397 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag
1398
1399 // Enter the waiting queue, which is a circular doubly linked list in this case
1400 // but it could be a priority queue or any data structure.
1401 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only
1402 // by the the owner of the monitor *except* in the case where park()
1403 // returns because of a timeout of interrupt. Contention is exceptionally rare
1404 // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1405
1406 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1407 AddWaiter(&node);
1408 Thread::SpinRelease(&_WaitSetLock);
1409
1410 _Responsible = NULL;
1411
1412 intptr_t save = _recursions; // record the old recursion count
1413 _waiters++; // increment the number of waiters
1414 _recursions = 0; // set the recursion level to be 1
1415 exit(true, Self); // exit the monitor
1416 guarantee(_owner != Self, "invariant");
1417
1418 // The thread is on the WaitSet list - now park() it.
1419 // On MP systems it's conceivable that a brief spin before we park
1420 // could be profitable.
1421 //
1422 // TODO-FIXME: change the following logic to a loop of the form
1423 // while (!timeout && !interrupted && _notified == 0) park()
1424
1425 int ret = OS_OK;
1426 int WasNotified = 0;
1427 { // State transition wrappers
1428 OSThread* osthread = Self->osthread();
1429 OSThreadWaitState osts(osthread, true);
1430 {
1431 ThreadBlockInVM tbivm(jt);
1432 // Thread is in thread_blocked state and oop access is unsafe.
1433 jt->set_suspend_equivalent();
1434
1435 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1436 // Intentionally empty
1437 } else if (node._notified == 0) {
1438 if (millis <= 0) {
1439 Self->_ParkEvent->park();
1440 } else {
1441 ret = Self->_ParkEvent->park(millis);
1442 }
1443 }
1444
1445 // were we externally suspended while we were waiting?
1446 if (ExitSuspendEquivalent (jt)) {
1447 // TODO-FIXME: add -- if succ == Self then succ = null.
1448 jt->java_suspend_self();
1449 }
1450
1451 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1452
1453 // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1454 // from the WaitSet to the EntryList.
1455 // See if we need to remove Node from the WaitSet.
1456 // We use double-checked locking to avoid grabbing _WaitSetLock
1457 // if the thread is not on the wait queue.
1458 //
1459 // Note that we don't need a fence before the fetch of TState.
1460 // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1461 // written by the is thread. (perhaps the fetch might even be satisfied
1462 // by a look-aside into the processor's own store buffer, although given
1463 // the length of the code path between the prior ST and this load that's
1464 // highly unlikely). If the following LD fetches a stale TS_WAIT value
1465 // then we'll acquire the lock and then re-fetch a fresh TState value.
1466 // That is, we fail toward safety.
1467
1468 if (node.TState == ObjectWaiter::TS_WAIT) {
1469 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1470 if (node.TState == ObjectWaiter::TS_WAIT) {
1471 DequeueSpecificWaiter(&node); // unlink from WaitSet
1472 assert(node._notified == 0, "invariant");
1473 node.TState = ObjectWaiter::TS_RUN;
1474 }
1475 Thread::SpinRelease(&_WaitSetLock);
1476 }
1477
1478 // The thread is now either on off-list (TS_RUN),
1479 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1480 // The Node's TState variable is stable from the perspective of this thread.
1481 // No other threads will asynchronously modify TState.
1482 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1483 OrderAccess::loadload();
1484 if (_succ == Self) _succ = NULL;
1485 WasNotified = node._notified;
1486
1487 // Reentry phase -- reacquire the monitor.
1488 // re-enter contended monitor after object.wait().
1489 // retain OBJECT_WAIT state until re-enter successfully completes
1490 // Thread state is thread_in_vm and oop access is again safe,
1491 // although the raw address of the object may have changed.
1492 // (Don't cache naked oops over safepoints, of course).
1493
1494 // post monitor waited event. Note that this is past-tense, we are done waiting.
1495 if (JvmtiExport::should_post_monitor_waited()) {
1496 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1497
1498 if (node._notified != 0 && _succ == Self) {
1499 // In this part of the monitor wait-notify-reenter protocol it
1500 // is possible (and normal) for another thread to do a fastpath
1501 // monitor enter-exit while this thread is still trying to get
1502 // to the reenter portion of the protocol.
1503 //
1504 // The ObjectMonitor was notified and the current thread is
1505 // the successor which also means that an unpark() has already
1506 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1507 // consume the unpark() that was done when the successor was
1508 // set because the same ParkEvent is shared between Java
1509 // monitors and JVM/TI RawMonitors (for now).
1510 //
1511 // We redo the unpark() to ensure forward progress, i.e., we
1512 // don't want all pending threads hanging (parked) with none
1513 // entering the unlocked monitor.
1514 node._event->unpark();
1515 }
1516 }
1517
1518 if (event.should_commit()) {
1519 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1520 }
1521
1522 OrderAccess::fence();
1523
1524 assert(Self->_Stalled != 0, "invariant");
1525 Self->_Stalled = 0;
1526
1527 assert(_owner != Self, "invariant");
1528 ObjectWaiter::TStates v = node.TState;
1529 if (v == ObjectWaiter::TS_RUN) {
1530 enter(Self);
1531 } else {
1532 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1533 ReenterI(Self, &node);
1534 node.wait_reenter_end(this);
1535 }
1536
1537 // Self has reacquired the lock.
1538 // Lifecycle - the node representing Self must not appear on any queues.
1539 // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1540 // want residual elements associated with this thread left on any lists.
1541 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1542 assert(_owner == Self, "invariant");
1543 assert(_succ != Self, "invariant");
1544 } // OSThreadWaitState()
1545
1546 jt->set_current_waiting_monitor(NULL);
1547
1548 guarantee(_recursions == 0, "invariant");
1549 _recursions = save; // restore the old recursion count
1550 _waiters--; // decrement the number of waiters
1551
1552 // Verify a few postconditions
1553 assert(_owner == Self, "invariant");
1554 assert(_succ != Self, "invariant");
1555 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
1556
1557 // check if the notification happened
1558 if (!WasNotified) {
1559 // no, it could be timeout or Thread.interrupt() or both
1560 // check for interrupt event, otherwise it is timeout
1561 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1562 THROW(vmSymbols::java_lang_InterruptedException());
1563 }
1564 }
1565
1566 // NOTE: Spurious wake up will be consider as timeout.
1567 // Monitor notify has precedence over thread interrupt.
1568 }
1569
1570
1571 // Consider:
1572 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1573 // then instead of transferring a thread from the WaitSet to the EntryList
1574 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1575
1576 void ObjectMonitor::INotify(Thread * Self) {
1577 const int policy = Knob_MoveNotifyee;
1578
1579 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1580 ObjectWaiter * iterator = DequeueWaiter();
1581 if (iterator != NULL) {
1582 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1583 guarantee(iterator->_notified == 0, "invariant");
1584 // Disposition - what might we do with iterator ?
1585 // a. add it directly to the EntryList - either tail (policy == 1)
1586 // or head (policy == 0).
1587 // b. push it onto the front of the _cxq (policy == 2).
1588 // For now we use (b).
1589 if (policy != 4) {
1590 iterator->TState = ObjectWaiter::TS_ENTER;
1591 }
1592 iterator->_notified = 1;
1593 iterator->_notifier_tid = JFR_THREAD_ID(Self);
1594
1595 ObjectWaiter * list = _EntryList;
1596 if (list != NULL) {
1597 assert(list->_prev == NULL, "invariant");
1598 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1599 assert(list != iterator, "invariant");
1600 }
1601
1602 if (policy == 0) { // prepend to EntryList
1603 if (list == NULL) {
1604 iterator->_next = iterator->_prev = NULL;
1605 _EntryList = iterator;
1606 } else {
1607 list->_prev = iterator;
1608 iterator->_next = list;
1609 iterator->_prev = NULL;
1610 _EntryList = iterator;
1611 }
1612 } else if (policy == 1) { // append to EntryList
1613 if (list == NULL) {
1614 iterator->_next = iterator->_prev = NULL;
1615 _EntryList = iterator;
1616 } else {
1617 // CONSIDER: finding the tail currently requires a linear-time walk of
1618 // the EntryList. We can make tail access constant-time by converting to
1619 // a CDLL instead of using our current DLL.
1620 ObjectWaiter * tail;
1621 for (tail = list; tail->_next != NULL; tail = tail->_next) {}
1622 assert(tail != NULL && tail->_next == NULL, "invariant");
1623 tail->_next = iterator;
1624 iterator->_prev = tail;
1625 iterator->_next = NULL;
1626 }
1627 } else if (policy == 2) { // prepend to cxq
1628 if (list == NULL) {
1629 iterator->_next = iterator->_prev = NULL;
1630 _EntryList = iterator;
1631 } else {
1632 iterator->TState = ObjectWaiter::TS_CXQ;
1633 for (;;) {
1634 ObjectWaiter * front = _cxq;
1635 iterator->_next = front;
1636 if (Atomic::cmpxchg(iterator, &_cxq, front) == front) {
1637 break;
1638 }
1639 }
1640 }
1641 } else if (policy == 3) { // append to cxq
1642 iterator->TState = ObjectWaiter::TS_CXQ;
1643 for (;;) {
1644 ObjectWaiter * tail = _cxq;
1645 if (tail == NULL) {
1646 iterator->_next = NULL;
1647 if (Atomic::replace_if_null(iterator, &_cxq)) {
1648 break;
1649 }
1650 } else {
1651 while (tail->_next != NULL) tail = tail->_next;
1652 tail->_next = iterator;
1653 iterator->_prev = tail;
1654 iterator->_next = NULL;
1655 break;
1656 }
1657 }
1658 } else {
1659 ParkEvent * ev = iterator->_event;
1660 iterator->TState = ObjectWaiter::TS_RUN;
1661 OrderAccess::fence();
1662 ev->unpark();
1663 }
1664
1665 // _WaitSetLock protects the wait queue, not the EntryList. We could
1666 // move the add-to-EntryList operation, above, outside the critical section
1667 // protected by _WaitSetLock. In practice that's not useful. With the
1668 // exception of wait() timeouts and interrupts the monitor owner
1669 // is the only thread that grabs _WaitSetLock. There's almost no contention
1670 // on _WaitSetLock so it's not profitable to reduce the length of the
1671 // critical section.
1672
1673 if (policy < 4) {
1674 iterator->wait_reenter_begin(this);
1675 }
1676 }
1677 Thread::SpinRelease(&_WaitSetLock);
1678 }
1679
1680 // Consider: a not-uncommon synchronization bug is to use notify() when
1681 // notifyAll() is more appropriate, potentially resulting in stranded
1682 // threads; this is one example of a lost wakeup. A useful diagnostic
1683 // option is to force all notify() operations to behave as notifyAll().
1684 //
1685 // Note: We can also detect many such problems with a "minimum wait".
1686 // When the "minimum wait" is set to a small non-zero timeout value
1687 // and the program does not hang whereas it did absent "minimum wait",
1688 // that suggests a lost wakeup bug.
1689
1690 void ObjectMonitor::notify(TRAPS) {
1691 CHECK_OWNER();
1692 if (_WaitSet == NULL) {
1693 return;
1694 }
1695 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1696 INotify(THREAD);
1697 OM_PERFDATA_OP(Notifications, inc(1));
1698 }
1699
1700
1701 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1702 // from the waitset to the EntryList. This could be done more efficiently with a
1703 // single bulk transfer but in practice it's not time-critical. Beware too,
1704 // that in prepend-mode we invert the order of the waiters. Let's say that the
1705 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1706 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1707
1708 void ObjectMonitor::notifyAll(TRAPS) {
1709 CHECK_OWNER();
1710 if (_WaitSet == NULL) {
1711 return;
1712 }
1713
1714 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1715 int tally = 0;
1716 while (_WaitSet != NULL) {
1717 tally++;
1718 INotify(THREAD);
1719 }
1720
1721 OM_PERFDATA_OP(Notifications, inc(tally));
1722 }
1723
1724 // -----------------------------------------------------------------------------
1725 // Adaptive Spinning Support
1726 //
1727 // Adaptive spin-then-block - rational spinning
1728 //
1729 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1730 // algorithm. On high order SMP systems it would be better to start with
1731 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH,
1732 // a contending thread could enqueue itself on the cxq and then spin locally
1733 // on a thread-specific variable such as its ParkEvent._Event flag.
1734 // That's left as an exercise for the reader. Note that global spinning is
1735 // not problematic on Niagara, as the L2 cache serves the interconnect and
1736 // has both low latency and massive bandwidth.
1737 //
1738 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1739 // acquisition attempts where we opt to spin -- at 100% and vary the spin count
1740 // (duration) or we can fix the count at approximately the duration of
1741 // a context switch and vary the frequency. Of course we could also
1742 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1743 // For a description of 'Adaptive spin-then-block mutual exclusion in
1744 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1745 //
1746 // This implementation varies the duration "D", where D varies with
1747 // the success rate of recent spin attempts. (D is capped at approximately
1748 // length of a round-trip context switch). The success rate for recent
1749 // spin attempts is a good predictor of the success rate of future spin
1750 // attempts. The mechanism adapts automatically to varying critical
1751 // section length (lock modality), system load and degree of parallelism.
1752 // D is maintained per-monitor in _SpinDuration and is initialized
1753 // optimistically. Spin frequency is fixed at 100%.
1754 //
1755 // Note that _SpinDuration is volatile, but we update it without locks
1756 // or atomics. The code is designed so that _SpinDuration stays within
1757 // a reasonable range even in the presence of races. The arithmetic
1758 // operations on _SpinDuration are closed over the domain of legal values,
1759 // so at worst a race will install and older but still legal value.
1760 // At the very worst this introduces some apparent non-determinism.
1761 // We might spin when we shouldn't or vice-versa, but since the spin
1762 // count are relatively short, even in the worst case, the effect is harmless.
1763 //
1764 // Care must be taken that a low "D" value does not become an
1765 // an absorbing state. Transient spinning failures -- when spinning
1766 // is overall profitable -- should not cause the system to converge
1767 // on low "D" values. We want spinning to be stable and predictable
1768 // and fairly responsive to change and at the same time we don't want
1769 // it to oscillate, become metastable, be "too" non-deterministic,
1770 // or converge on or enter undesirable stable absorbing states.
1771 //
1772 // We implement a feedback-based control system -- using past behavior
1773 // to predict future behavior. We face two issues: (a) if the
1774 // input signal is random then the spin predictor won't provide optimal
1775 // results, and (b) if the signal frequency is too high then the control
1776 // system, which has some natural response lag, will "chase" the signal.
1777 // (b) can arise from multimodal lock hold times. Transient preemption
1778 // can also result in apparent bimodal lock hold times.
1779 // Although sub-optimal, neither condition is particularly harmful, as
1780 // in the worst-case we'll spin when we shouldn't or vice-versa.
1781 // The maximum spin duration is rather short so the failure modes aren't bad.
1782 // To be conservative, I've tuned the gain in system to bias toward
1783 // _not spinning. Relatedly, the system can sometimes enter a mode where it
1784 // "rings" or oscillates between spinning and not spinning. This happens
1785 // when spinning is just on the cusp of profitability, however, so the
1786 // situation is not dire. The state is benign -- there's no need to add
1787 // hysteresis control to damp the transition rate between spinning and
1788 // not spinning.
1789
1790 // Spinning: Fixed frequency (100%), vary duration
1791 int ObjectMonitor::TrySpin(Thread * Self) {
1792 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning.
1793 int ctr = Knob_FixedSpin;
1794 if (ctr != 0) {
1795 while (--ctr >= 0) {
1796 if (TryLock(Self) > 0) return 1;
1797 SpinPause();
1798 }
1799 return 0;
1800 }
1801
1802 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1803 if (TryLock(Self) > 0) {
1804 // Increase _SpinDuration ...
1805 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1806 // Raising _SpurDuration to the poverty line is key.
1807 int x = _SpinDuration;
1808 if (x < Knob_SpinLimit) {
1809 if (x < Knob_Poverty) x = Knob_Poverty;
1810 _SpinDuration = x + Knob_BonusB;
1811 }
1812 return 1;
1813 }
1814 SpinPause();
1815 }
1816
1817 // Admission control - verify preconditions for spinning
1818 //
1819 // We always spin a little bit, just to prevent _SpinDuration == 0 from
1820 // becoming an absorbing state. Put another way, we spin briefly to
1821 // sample, just in case the system load, parallelism, contention, or lock
1822 // modality changed.
1823 //
1824 // Consider the following alternative:
1825 // Periodically set _SpinDuration = _SpinLimit and try a long/full
1826 // spin attempt. "Periodically" might mean after a tally of
1827 // the # of failed spin attempts (or iterations) reaches some threshold.
1828 // This takes us into the realm of 1-out-of-N spinning, where we
1829 // hold the duration constant but vary the frequency.
1830
1831 ctr = _SpinDuration;
1832 if (ctr <= 0) return 0;
1833
1834 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
1835 return 0;
1836 }
1837
1838 // We're good to spin ... spin ingress.
1839 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1840 // when preparing to LD...CAS _owner, etc and the CAS is likely
1841 // to succeed.
1842 if (_succ == NULL) {
1843 _succ = Self;
1844 }
1845 Thread * prv = NULL;
1846
1847 // There are three ways to exit the following loop:
1848 // 1. A successful spin where this thread has acquired the lock.
1849 // 2. Spin failure with prejudice
1850 // 3. Spin failure without prejudice
1851
1852 while (--ctr >= 0) {
1853
1854 // Periodic polling -- Check for pending GC
1855 // Threads may spin while they're unsafe.
1856 // We don't want spinning threads to delay the JVM from reaching
1857 // a stop-the-world safepoint or to steal cycles from GC.
1858 // If we detect a pending safepoint we abort in order that
1859 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1860 // this thread, if safe, doesn't steal cycles from GC.
1861 // This is in keeping with the "no loitering in runtime" rule.
1862 // We periodically check to see if there's a safepoint pending.
1863 if ((ctr & 0xFF) == 0) {
1864 if (SafepointMechanism::poll(Self)) {
1865 goto Abort; // abrupt spin egress
1866 }
1867 if (Knob_UsePause & 1) SpinPause();
1868 }
1869
1870 if (Knob_UsePause & 2) SpinPause();
1871
1872 // Probe _owner with TATAS
1873 // If this thread observes the monitor transition or flicker
1874 // from locked to unlocked to locked, then the odds that this
1875 // thread will acquire the lock in this spin attempt go down
1876 // considerably. The same argument applies if the CAS fails
1877 // or if we observe _owner change from one non-null value to
1878 // another non-null value. In such cases we might abort
1879 // the spin without prejudice or apply a "penalty" to the
1880 // spin count-down variable "ctr", reducing it by 100, say.
1881
1882 Thread * ox = (Thread *) _owner;
1883 if (ox == NULL) {
1884 ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL);
1885 if (ox == NULL) {
1886 // The CAS succeeded -- this thread acquired ownership
1887 // Take care of some bookkeeping to exit spin state.
1888 if (_succ == Self) {
1889 _succ = NULL;
1890 }
1891
1892 // Increase _SpinDuration :
1893 // The spin was successful (profitable) so we tend toward
1894 // longer spin attempts in the future.
1895 // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1896 // If we acquired the lock early in the spin cycle it
1897 // makes sense to increase _SpinDuration proportionally.
1898 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1899 int x = _SpinDuration;
1900 if (x < Knob_SpinLimit) {
1901 if (x < Knob_Poverty) x = Knob_Poverty;
1902 _SpinDuration = x + Knob_Bonus;
1903 }
1904 return 1;
1905 }
1906
1907 // The CAS failed ... we can take any of the following actions:
1908 // * penalize: ctr -= CASPenalty
1909 // * exit spin with prejudice -- goto Abort;
1910 // * exit spin without prejudice.
1911 // * Since CAS is high-latency, retry again immediately.
1912 prv = ox;
1913 goto Abort;
1914 }
1915
1916 // Did lock ownership change hands ?
1917 if (ox != prv && prv != NULL) {
1918 goto Abort;
1919 }
1920 prv = ox;
1921
1922 // Abort the spin if the owner is not executing.
1923 // The owner must be executing in order to drop the lock.
1924 // Spinning while the owner is OFFPROC is idiocy.
1925 // Consider: ctr -= RunnablePenalty ;
1926 if (Knob_OState && NotRunnable (Self, ox)) {
1927 goto Abort;
1928 }
1929 if (_succ == NULL) {
1930 _succ = Self;
1931 }
1932 }
1933
1934 // Spin failed with prejudice -- reduce _SpinDuration.
1935 // TODO: Use an AIMD-like policy to adjust _SpinDuration.
1936 // AIMD is globally stable.
1937 {
1938 int x = _SpinDuration;
1939 if (x > 0) {
1940 // Consider an AIMD scheme like: x -= (x >> 3) + 100
1941 // This is globally sample and tends to damp the response.
1942 x -= Knob_Penalty;
1943 if (x < 0) x = 0;
1944 _SpinDuration = x;
1945 }
1946 }
1947
1948 Abort:
1949 if (_succ == Self) {
1950 _succ = NULL;
1951 // Invariant: after setting succ=null a contending thread
1952 // must recheck-retry _owner before parking. This usually happens
1953 // in the normal usage of TrySpin(), but it's safest
1954 // to make TrySpin() as foolproof as possible.
1955 OrderAccess::fence();
1956 if (TryLock(Self) > 0) return 1;
1957 }
1958 return 0;
1959 }
1960
1961 // NotRunnable() -- informed spinning
1962 //
1963 // Don't bother spinning if the owner is not eligible to drop the lock.
1964 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm.
1965 // The thread must be runnable in order to drop the lock in timely fashion.
1966 // If the _owner is not runnable then spinning will not likely be
1967 // successful (profitable).
1968 //
1969 // Beware -- the thread referenced by _owner could have died
1970 // so a simply fetch from _owner->_thread_state might trap.
1971 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
1972 // Because of the lifecycle issues, the _thread_state values
1973 // observed by NotRunnable() might be garbage. NotRunnable must
1974 // tolerate this and consider the observed _thread_state value
1975 // as advisory.
1976 //
1977 // Beware too, that _owner is sometimes a BasicLock address and sometimes
1978 // a thread pointer.
1979 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
1980 // with the LSB of _owner. Another option would be to probabilistically probe
1981 // the putative _owner->TypeTag value.
1982 //
1983 // Checking _thread_state isn't perfect. Even if the thread is
1984 // in_java it might be blocked on a page-fault or have been preempted
1985 // and sitting on a ready/dispatch queue.
1986 //
1987 // The return value from NotRunnable() is *advisory* -- the
1988 // result is based on sampling and is not necessarily coherent.
1989 // The caller must tolerate false-negative and false-positive errors.
1990 // Spinning, in general, is probabilistic anyway.
1991
1992
1993 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) {
1994 // Check ox->TypeTag == 2BAD.
1995 if (ox == NULL) return 0;
1996
1997 // Avoid transitive spinning ...
1998 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L.
1999 // Immediately after T1 acquires L it's possible that T2, also
2000 // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2001 // This occurs transiently after T1 acquired L but before
2002 // T1 managed to clear T1.Stalled. T2 does not need to abort
2003 // its spin in this circumstance.
2004 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2005
2006 if (BlockedOn == 1) return 1;
2007 if (BlockedOn != 0) {
2008 return BlockedOn != intptr_t(this) && _owner == ox;
2009 }
2010
2011 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant");
2012 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);;
2013 // consider also: jst != _thread_in_Java -- but that's overspecific.
2014 return jst == _thread_blocked || jst == _thread_in_native;
2015 }
2016
2017
2018 // -----------------------------------------------------------------------------
2019 // WaitSet management ...
2020
2021 ObjectWaiter::ObjectWaiter(Thread* thread) {
2022 _next = NULL;
2023 _prev = NULL;
2024 _notified = 0;
2025 _notifier_tid = 0;
2026 TState = TS_RUN;
2027 _thread = thread;
2028 _event = thread->_ParkEvent;
2029 _active = false;
2030 assert(_event != NULL, "invariant");
2031 }
2032
2033 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2034 JavaThread *jt = (JavaThread *)this->_thread;
2035 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2036 }
2037
2038 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2039 JavaThread *jt = (JavaThread *)this->_thread;
2040 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2041 }
2042
2043 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2044 assert(node != NULL, "should not add NULL node");
2045 assert(node->_prev == NULL, "node already in list");
2046 assert(node->_next == NULL, "node already in list");
2047 // put node at end of queue (circular doubly linked list)
2048 if (_WaitSet == NULL) {
2049 _WaitSet = node;
2050 node->_prev = node;
2051 node->_next = node;
2052 } else {
2053 ObjectWaiter* head = _WaitSet;
2054 ObjectWaiter* tail = head->_prev;
2055 assert(tail->_next == head, "invariant check");
2056 tail->_next = node;
2057 head->_prev = node;
2058 node->_next = head;
2059 node->_prev = tail;
2060 }
2061 }
2062
2063 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2064 // dequeue the very first waiter
2065 ObjectWaiter* waiter = _WaitSet;
2066 if (waiter) {
2067 DequeueSpecificWaiter(waiter);
2068 }
2069 return waiter;
2070 }
2071
2072 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2073 assert(node != NULL, "should not dequeue NULL node");
2074 assert(node->_prev != NULL, "node already removed from list");
2075 assert(node->_next != NULL, "node already removed from list");
2076 // when the waiter has woken up because of interrupt,
2077 // timeout or other spurious wake-up, dequeue the
2078 // waiter from waiting list
2079 ObjectWaiter* next = node->_next;
2080 if (next == node) {
2081 assert(node->_prev == node, "invariant check");
2082 _WaitSet = NULL;
2083 } else {
2084 ObjectWaiter* prev = node->_prev;
2085 assert(prev->_next == node, "invariant check");
2086 assert(next->_prev == node, "invariant check");
2087 next->_prev = prev;
2088 prev->_next = next;
2089 if (_WaitSet == node) {
2090 _WaitSet = next;
2091 }
2092 }
2093 node->_next = NULL;
2094 node->_prev = NULL;
2095 }
2096
2097 // -----------------------------------------------------------------------------
2098 // PerfData support
2099 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL;
2100 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL;
2101 PerfCounter * ObjectMonitor::_sync_Parks = NULL;
2102 PerfCounter * ObjectMonitor::_sync_Notifications = NULL;
2103 PerfCounter * ObjectMonitor::_sync_Inflations = NULL;
2104 PerfCounter * ObjectMonitor::_sync_Deflations = NULL;
2105 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL;
2106
2107 // One-shot global initialization for the sync subsystem.
2108 // We could also defer initialization and initialize on-demand
2109 // the first time we call inflate(). Initialization would
2110 // be protected - like so many things - by the MonitorCache_lock.
2111
2112 void ObjectMonitor::Initialize() {
2113 static int InitializationCompleted = 0;
2114 assert(InitializationCompleted == 0, "invariant");
2115 InitializationCompleted = 1;
2116 if (UsePerfData) {
2117 EXCEPTION_MARK;
2118 #define NEWPERFCOUNTER(n) \
2119 { \
2120 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \
2121 CHECK); \
2122 }
2123 #define NEWPERFVARIABLE(n) \
2124 { \
2125 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \
2126 CHECK); \
2127 }
2128 NEWPERFCOUNTER(_sync_Inflations);
2129 NEWPERFCOUNTER(_sync_Deflations);
2130 NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2131 NEWPERFCOUNTER(_sync_FutileWakeups);
2132 NEWPERFCOUNTER(_sync_Parks);
2133 NEWPERFCOUNTER(_sync_Notifications);
2134 NEWPERFVARIABLE(_sync_MonExtant);
2135 #undef NEWPERFCOUNTER
2136 #undef NEWPERFVARIABLE
2137 }
2138 }
2139
2140 void ObjectMonitor::DeferredInitialize() {
2141 if (InitDone > 0) return;
2142 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2143 while (InitDone != 1) /* empty */;
2144 return;
2145 }
2146
2147 // One-shot global initialization ...
2148 // The initialization is idempotent, so we don't need locks.
2149 // In the future consider doing this via os::init_2().
2150
2151 if (!os::is_MP()) {
2152 Knob_SpinLimit = 0;
2153 Knob_PreSpin = 0;
2154 Knob_FixedSpin = -1;
2155 }
2156
2157 OrderAccess::fence();
2158 InitDone = 1;
2159 }
2160