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