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_ExitRelease = 0;
105 int ObjectMonitor::Knob_InlineNotify = 1;
106 int ObjectMonitor::Knob_Verbose = 0;
107 int ObjectMonitor::Knob_VerifyInUse = 0;
108 int ObjectMonitor::Knob_VerifyMatch = 0;
109 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool -
110
111 static int Knob_SpinBase = 0; // Floor AKA SpinMin
112 static int Knob_SpinBackOff = 0; // spin-loop backoff
113 static int Knob_CASPenalty = -1; // Penalty for failed CAS
114 static int Knob_OXPenalty = -1; // Penalty for observed _owner change
115 static int Knob_SpinSetSucc = 1; // spinners set the _succ field
116 static int Knob_SpinEarly = 1;
117 static int Knob_SuccEnabled = 1; // futile wake throttling
118 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one
119 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs
120 static int Knob_Bonus = 100; // spin success bonus
121 static int Knob_BonusB = 100; // spin success bonus
122 static int Knob_Penalty = 200; // spin failure penalty
123 static int Knob_Poverty = 1000;
124 static int Knob_SpinAfterFutile = 1; // Spin after returning from park()
125 static int Knob_FixedSpin = 0;
126 static int Knob_OState = 3; // Spinner checks thread state of _owner
127 static int Knob_UsePause = 1;
128 static int Knob_ExitPolicy = 0;
129 static int Knob_PreSpin = 10; // 20-100 likely better
130 static int Knob_ResetEvent = 0;
131 static int BackOffMask = 0;
132
133 static int Knob_FastHSSEC = 0;
134 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee
135 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline
136 static volatile int InitDone = 0;
137
138 // -----------------------------------------------------------------------------
139 // Theory of operations -- Monitors lists, thread residency, etc:
140 //
141 // * A thread acquires ownership of a monitor by successfully
142 // CAS()ing the _owner field from null to non-null.
143 //
144 // * Invariant: A thread appears on at most one monitor list --
145 // cxq, EntryList or WaitSet -- at any one time.
146 //
147 // * Contending threads "push" themselves onto the cxq with CAS
148 // and then spin/park.
149 //
150 // * After a contending thread eventually acquires the lock it must
151 // dequeue itself from either the EntryList or the cxq.
152 //
153 // * The exiting thread identifies and unparks an "heir presumptive"
154 // tentative successor thread on the EntryList. Critically, the
155 // exiting thread doesn't unlink the successor thread from the EntryList.
156 // After having been unparked, the wakee will recontend for ownership of
157 // the monitor. The successor (wakee) will either acquire the lock or
158 // re-park itself.
159 //
160 // Succession is provided for by a policy of competitive handoff.
161 // The exiting thread does _not_ grant or pass ownership to the
162 // successor thread. (This is also referred to as "handoff" succession").
163 // Instead the exiting thread releases ownership and possibly wakes
164 // a successor, so the successor can (re)compete for ownership of the lock.
165 // If the EntryList is empty but the cxq is populated the exiting
166 // thread will drain the cxq into the EntryList. It does so by
167 // by detaching the cxq (installing null with CAS) and folding
168 // the threads from the cxq into the EntryList. The EntryList is
169 // doubly linked, while the cxq is singly linked because of the
170 // CAS-based "push" used to enqueue recently arrived threads (RATs).
171 //
172 // * Concurrency invariants:
173 //
174 // -- only the monitor owner may access or mutate the EntryList.
175 // The mutex property of the monitor itself protects the EntryList
176 // from concurrent interference.
177 // -- Only the monitor owner may detach the cxq.
178 //
179 // * The monitor entry list operations avoid locks, but strictly speaking
180 // they're not lock-free. Enter is lock-free, exit is not.
181 // For a description of 'Methods and apparatus providing non-blocking access
182 // to a resource,' see U.S. Pat. No. 7844973.
183 //
184 // * The cxq can have multiple concurrent "pushers" but only one concurrent
185 // detaching thread. This mechanism is immune from the ABA corruption.
186 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
187 //
188 // * Taken together, the cxq and the EntryList constitute or form a
189 // single logical queue of threads stalled trying to acquire the lock.
190 // We use two distinct lists to improve the odds of a constant-time
191 // dequeue operation after acquisition (in the ::enter() epilogue) and
192 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm).
193 // A key desideratum is to minimize queue & monitor metadata manipulation
194 // that occurs while holding the monitor lock -- that is, we want to
195 // minimize monitor lock holds times. Note that even a small amount of
196 // fixed spinning will greatly reduce the # of enqueue-dequeue operations
197 // on EntryList|cxq. That is, spinning relieves contention on the "inner"
198 // locks and monitor metadata.
199 //
200 // Cxq points to the set of Recently Arrived Threads attempting entry.
201 // Because we push threads onto _cxq with CAS, the RATs must take the form of
202 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when
203 // the unlocking thread notices that EntryList is null but _cxq is != null.
204 //
205 // The EntryList is ordered by the prevailing queue discipline and
206 // can be organized in any convenient fashion, such as a doubly-linked list or
207 // a circular doubly-linked list. Critically, we want insert and delete operations
208 // to operate in constant-time. If we need a priority queue then something akin
209 // to Solaris' sleepq would work nicely. Viz.,
210 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
211 // Queue discipline is enforced at ::exit() time, when the unlocking thread
212 // drains the cxq into the EntryList, and orders or reorders the threads on the
213 // EntryList accordingly.
214 //
215 // Barring "lock barging", this mechanism provides fair cyclic ordering,
216 // somewhat similar to an elevator-scan.
217 //
218 // * The monitor synchronization subsystem avoids the use of native
219 // synchronization primitives except for the narrow platform-specific
220 // park-unpark abstraction. See the comments in os_solaris.cpp regarding
221 // the semantics of park-unpark. Put another way, this monitor implementation
222 // depends only on atomic operations and park-unpark. The monitor subsystem
223 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
224 // underlying OS manages the READY<->RUN transitions.
225 //
226 // * Waiting threads reside on the WaitSet list -- wait() puts
227 // the caller onto the WaitSet.
228 //
229 // * notify() or notifyAll() simply transfers threads from the WaitSet to
230 // either the EntryList or cxq. Subsequent exit() operations will
231 // unpark the notifyee. Unparking a notifee in notify() is inefficient -
232 // it's likely the notifyee would simply impale itself on the lock held
233 // by the notifier.
234 //
235 // * An interesting alternative is to encode cxq as (List,LockByte) where
236 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary
237 // variable, like _recursions, in the scheme. The threads or Events that form
238 // the list would have to be aligned in 256-byte addresses. A thread would
239 // try to acquire the lock or enqueue itself with CAS, but exiting threads
240 // could use a 1-0 protocol and simply STB to set the LockByte to 0.
241 // Note that is is *not* word-tearing, but it does presume that full-word
242 // CAS operations are coherent with intermix with STB operations. That's true
243 // on most common processors.
244 //
245 // * See also http://blogs.sun.com/dave
246
247
248 void* ObjectMonitor::operator new (size_t size) throw() {
249 return AllocateHeap(size, mtInternal);
250 }
251 void* ObjectMonitor::operator new[] (size_t size) throw() {
252 return operator new (size);
253 }
254 void ObjectMonitor::operator delete(void* p) {
255 FreeHeap(p);
256 }
257 void ObjectMonitor::operator delete[] (void *p) {
258 operator delete(p);
259 }
260
261 // -----------------------------------------------------------------------------
262 // Enter support
263
264 void ObjectMonitor::enter(TRAPS) {
265 // The following code is ordered to check the most common cases first
266 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
267 Thread * const Self = THREAD;
268
269 void * cur = Atomic::cmpxchg(Self, &_owner, (void*)NULL);
270 if (cur == NULL) {
271 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
272 assert(_recursions == 0, "invariant");
273 assert(_owner == Self, "invariant");
274 return;
275 }
276
277 if (cur == Self) {
278 // TODO-FIXME: check for integer overflow! BUGID 6557169.
279 _recursions++;
280 return;
281 }
282
283 if (Self->is_lock_owned ((address)cur)) {
284 assert(_recursions == 0, "internal state error");
285 _recursions = 1;
286 // Commute owner from a thread-specific on-stack BasicLockObject address to
287 // a full-fledged "Thread *".
288 _owner = Self;
289 return;
290 }
291
292 // We've encountered genuine contention.
293 assert(Self->_Stalled == 0, "invariant");
294 Self->_Stalled = intptr_t(this);
295
296 // Try one round of spinning *before* enqueueing Self
297 // and before going through the awkward and expensive state
298 // transitions. The following spin is strictly optional ...
299 // Note that if we acquire the monitor from an initial spin
300 // we forgo posting JVMTI events and firing DTRACE probes.
301 if (Knob_SpinEarly && TrySpin (Self) > 0) {
302 assert(_owner == Self, "invariant");
303 assert(_recursions == 0, "invariant");
304 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
305 Self->_Stalled = 0;
306 return;
307 }
308
309 assert(_owner != Self, "invariant");
310 assert(_succ != Self, "invariant");
311 assert(Self->is_Java_thread(), "invariant");
312 JavaThread * jt = (JavaThread *) Self;
313 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
314 assert(jt->thread_state() != _thread_blocked, "invariant");
315 assert(this->object() != NULL, "invariant");
316 assert(_count >= 0, "invariant");
317
318 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy().
319 // Ensure the object-monitor relationship remains stable while there's contention.
320 Atomic::inc(&_count);
321
322 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);)
323 EventJavaMonitorEnter event;
324 if (event.should_commit()) {
325 event.set_monitorClass(((oop)this->object())->klass());
326 event.set_address((uintptr_t)(this->object_addr()));
327 }
328
329 { // Change java thread status to indicate blocked on monitor enter.
330 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
331
332 Self->set_current_pending_monitor(this);
333
334 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
335 if (JvmtiExport::should_post_monitor_contended_enter()) {
336 JvmtiExport::post_monitor_contended_enter(jt, this);
337
338 // The current thread does not yet own the monitor and does not
339 // yet appear on any queues that would get it made the successor.
340 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
341 // handler cannot accidentally consume an unpark() meant for the
342 // ParkEvent associated with this ObjectMonitor.
343 }
344
345 OSThreadContendState osts(Self->osthread());
346 ThreadBlockInVM tbivm(jt);
347
348 // TODO-FIXME: change the following for(;;) loop to straight-line code.
349 for (;;) {
350 jt->set_suspend_equivalent();
351 // cleared by handle_special_suspend_equivalent_condition()
352 // or java_suspend_self()
353
354 EnterI(THREAD);
355
356 if (!ExitSuspendEquivalent(jt)) break;
357
358 // We have acquired the contended monitor, but while we were
359 // waiting another thread suspended us. We don't want to enter
360 // the monitor while suspended because that would surprise the
361 // thread that suspended us.
362 //
363 _recursions = 0;
364 _succ = NULL;
365 exit(false, Self);
366
367 jt->java_suspend_self();
368 }
369 Self->set_current_pending_monitor(NULL);
370
371 // We cleared the pending monitor info since we've just gotten past
372 // the enter-check-for-suspend dance and we now own the monitor free
373 // and clear, i.e., it is no longer pending. The ThreadBlockInVM
374 // destructor can go to a safepoint at the end of this block. If we
375 // do a thread dump during that safepoint, then this thread will show
376 // as having "-locked" the monitor, but the OS and java.lang.Thread
377 // states will still report that the thread is blocked trying to
378 // acquire it.
379 }
380
381 Atomic::dec(&_count);
382 assert(_count >= 0, "invariant");
383 Self->_Stalled = 0;
384
385 // Must either set _recursions = 0 or ASSERT _recursions == 0.
386 assert(_recursions == 0, "invariant");
387 assert(_owner == Self, "invariant");
388 assert(_succ != Self, "invariant");
389 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
390
391 // The thread -- now the owner -- is back in vm mode.
392 // Report the glorious news via TI,DTrace and jvmstat.
393 // The probe effect is non-trivial. All the reportage occurs
394 // while we hold the monitor, increasing the length of the critical
395 // section. Amdahl's parallel speedup law comes vividly into play.
396 //
397 // Another option might be to aggregate the events (thread local or
398 // per-monitor aggregation) and defer reporting until a more opportune
399 // time -- such as next time some thread encounters contention but has
400 // yet to acquire the lock. While spinning that thread could
401 // spinning we could increment JVMStat counters, etc.
402
403 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
404 if (JvmtiExport::should_post_monitor_contended_entered()) {
405 JvmtiExport::post_monitor_contended_entered(jt, this);
406
407 // The current thread already owns the monitor and is not going to
408 // call park() for the remainder of the monitor enter protocol. So
409 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
410 // event handler consumed an unpark() issued by the thread that
411 // just exited the monitor.
412 }
413 if (event.should_commit()) {
414 event.set_previousOwner((uintptr_t)_previous_owner_tid);
415 event.commit();
416 }
417 OM_PERFDATA_OP(ContendedLockAttempts, inc());
418 }
419
420 // Caveat: TryLock() is not necessarily serializing if it returns failure.
421 // Callers must compensate as needed.
422
423 int ObjectMonitor::TryLock(Thread * Self) {
424 void * own = _owner;
425 if (own != NULL) return 0;
426 if (Atomic::replace_if_null(Self, &_owner)) {
427 // Either guarantee _recursions == 0 or set _recursions = 0.
428 assert(_recursions == 0, "invariant");
429 assert(_owner == Self, "invariant");
430 return 1;
431 }
432 // The lock had been free momentarily, but we lost the race to the lock.
433 // Interference -- the CAS failed.
434 // We can either return -1 or retry.
435 // Retry doesn't make as much sense because the lock was just acquired.
436 return -1;
437 }
438
439 #define MAX_RECHECK_INTERVAL 1000
440
441 void ObjectMonitor::EnterI(TRAPS) {
442 Thread * const Self = THREAD;
443 assert(Self->is_Java_thread(), "invariant");
444 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant");
445
446 // Try the lock - TATAS
447 if (TryLock (Self) > 0) {
448 assert(_succ != Self, "invariant");
449 assert(_owner == Self, "invariant");
450 assert(_Responsible != Self, "invariant");
451 return;
452 }
453
454 DeferredInitialize();
455
456 // We try one round of spinning *before* enqueueing Self.
457 //
458 // If the _owner is ready but OFFPROC we could use a YieldTo()
459 // operation to donate the remainder of this thread's quantum
460 // to the owner. This has subtle but beneficial affinity
461 // effects.
462
463 if (TrySpin (Self) > 0) {
464 assert(_owner == Self, "invariant");
465 assert(_succ != Self, "invariant");
466 assert(_Responsible != Self, "invariant");
467 return;
468 }
469
470 // The Spin failed -- Enqueue and park the thread ...
471 assert(_succ != Self, "invariant");
472 assert(_owner != Self, "invariant");
473 assert(_Responsible != Self, "invariant");
474
475 // Enqueue "Self" on ObjectMonitor's _cxq.
476 //
477 // Node acts as a proxy for Self.
478 // As an aside, if were to ever rewrite the synchronization code mostly
479 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
480 // Java objects. This would avoid awkward lifecycle and liveness issues,
481 // as well as eliminate a subset of ABA issues.
482 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
483
484 ObjectWaiter node(Self);
485 Self->_ParkEvent->reset();
486 node._prev = (ObjectWaiter *) 0xBAD;
487 node.TState = ObjectWaiter::TS_CXQ;
488
489 // Push "Self" onto the front of the _cxq.
490 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
491 // Note that spinning tends to reduce the rate at which threads
492 // enqueue and dequeue on EntryList|cxq.
493 ObjectWaiter * nxt;
494 for (;;) {
495 node._next = nxt = _cxq;
496 if (Atomic::cmpxchg(&node, &_cxq, nxt) == nxt) break;
497
498 // Interference - the CAS failed because _cxq changed. Just retry.
499 // As an optional optimization we retry the lock.
500 if (TryLock (Self) > 0) {
501 assert(_succ != Self, "invariant");
502 assert(_owner == Self, "invariant");
503 assert(_Responsible != Self, "invariant");
504 return;
505 }
506 }
507
508 // Check for cxq|EntryList edge transition to non-null. This indicates
509 // the onset of contention. While contention persists exiting threads
510 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit
511 // operations revert to the faster 1-0 mode. This enter operation may interleave
512 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
513 // arrange for one of the contending thread to use a timed park() operations
514 // to detect and recover from the race. (Stranding is form of progress failure
515 // where the monitor is unlocked but all the contending threads remain parked).
516 // That is, at least one of the contended threads will periodically poll _owner.
517 // One of the contending threads will become the designated "Responsible" thread.
518 // The Responsible thread uses a timed park instead of a normal indefinite park
519 // operation -- it periodically wakes and checks for and recovers from potential
520 // strandings admitted by 1-0 exit operations. We need at most one Responsible
521 // thread per-monitor at any given moment. Only threads on cxq|EntryList may
522 // be responsible for a monitor.
523 //
524 // Currently, one of the contended threads takes on the added role of "Responsible".
525 // A viable alternative would be to use a dedicated "stranding checker" thread
526 // that periodically iterated over all the threads (or active monitors) and unparked
527 // successors where there was risk of stranding. This would help eliminate the
528 // timer scalability issues we see on some platforms as we'd only have one thread
529 // -- the checker -- parked on a timer.
530
531 if (nxt == NULL && _EntryList == NULL) {
532 // Try to assume the role of responsible thread for the monitor.
533 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self }
534 Atomic::replace_if_null(Self, &_Responsible);
535 }
536
537 // The lock might have been released while this thread was occupied queueing
538 // itself onto _cxq. To close the race and avoid "stranding" and
539 // progress-liveness failure we must resample-retry _owner before parking.
540 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
541 // In this case the ST-MEMBAR is accomplished with CAS().
542 //
543 // TODO: Defer all thread state transitions until park-time.
544 // Since state transitions are heavy and inefficient we'd like
545 // to defer the state transitions until absolutely necessary,
546 // and in doing so avoid some transitions ...
547
548 int nWakeups = 0;
549 int recheckInterval = 1;
550
551 for (;;) {
552
553 if (TryLock(Self) > 0) break;
554 assert(_owner != Self, "invariant");
555
556 // park self
557 if (_Responsible == Self) {
558 Self->_ParkEvent->park((jlong) recheckInterval);
559 // Increase the recheckInterval, but clamp the value.
560 recheckInterval *= 8;
561 if (recheckInterval > MAX_RECHECK_INTERVAL) {
562 recheckInterval = MAX_RECHECK_INTERVAL;
563 }
564 } else {
565 Self->_ParkEvent->park();
566 }
567
568 if (TryLock(Self) > 0) break;
569
570 // The lock is still contested.
571 // Keep a tally of the # of futile wakeups.
572 // Note that the counter is not protected by a lock or updated by atomics.
573 // That is by design - we trade "lossy" counters which are exposed to
574 // races during updates for a lower probe effect.
575
576 // This PerfData object can be used in parallel with a safepoint.
577 // See the work around in PerfDataManager::destroy().
578 OM_PERFDATA_OP(FutileWakeups, inc());
579 ++nWakeups;
580
581 // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
582 // We can defer clearing _succ until after the spin completes
583 // TrySpin() must tolerate being called with _succ == Self.
584 // Try yet another round of adaptive spinning.
585 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break;
586
587 // We can find that we were unpark()ed and redesignated _succ while
588 // we were spinning. That's harmless. If we iterate and call park(),
589 // park() will consume the event and return immediately and we'll
590 // just spin again. This pattern can repeat, leaving _succ to simply
591 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks().
592 // Alternately, we can sample fired() here, and if set, forgo spinning
593 // in the next iteration.
594
595 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
596 Self->_ParkEvent->reset();
597 OrderAccess::fence();
598 }
599 if (_succ == Self) _succ = NULL;
600
601 // Invariant: after clearing _succ a thread *must* retry _owner before parking.
602 OrderAccess::fence();
603 }
604
605 // Egress :
606 // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
607 // Normally we'll find Self on the EntryList .
608 // From the perspective of the lock owner (this thread), the
609 // EntryList is stable and cxq is prepend-only.
610 // The head of cxq is volatile but the interior is stable.
611 // In addition, Self.TState is stable.
612
613 assert(_owner == Self, "invariant");
614 assert(object() != NULL, "invariant");
615 // I'd like to write:
616 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
617 // but as we're at a safepoint that's not safe.
618
619 UnlinkAfterAcquire(Self, &node);
620 if (_succ == Self) _succ = NULL;
621
622 assert(_succ != Self, "invariant");
623 if (_Responsible == Self) {
624 _Responsible = NULL;
625 OrderAccess::fence(); // Dekker pivot-point
626
627 // We may leave threads on cxq|EntryList without a designated
628 // "Responsible" thread. This is benign. When this thread subsequently
629 // exits the monitor it can "see" such preexisting "old" threads --
630 // threads that arrived on the cxq|EntryList before the fence, above --
631 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads
632 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
633 // non-null and elect a new "Responsible" timer thread.
634 //
635 // This thread executes:
636 // ST Responsible=null; MEMBAR (in enter epilogue - here)
637 // LD cxq|EntryList (in subsequent exit)
638 //
639 // Entering threads in the slow/contended path execute:
640 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
641 // The (ST cxq; MEMBAR) is accomplished with CAS().
642 //
643 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
644 // exit operation from floating above the ST Responsible=null.
645 }
646
647 // We've acquired ownership with CAS().
648 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
649 // But since the CAS() this thread may have also stored into _succ,
650 // EntryList, cxq or Responsible. These meta-data updates must be
651 // visible __before this thread subsequently drops the lock.
652 // Consider what could occur if we didn't enforce this constraint --
653 // STs to monitor meta-data and user-data could reorder with (become
654 // visible after) the ST in exit that drops ownership of the lock.
655 // Some other thread could then acquire the lock, but observe inconsistent
656 // or old monitor meta-data and heap data. That violates the JMM.
657 // To that end, the 1-0 exit() operation must have at least STST|LDST
658 // "release" barrier semantics. Specifically, there must be at least a
659 // STST|LDST barrier in exit() before the ST of null into _owner that drops
660 // the lock. The barrier ensures that changes to monitor meta-data and data
661 // protected by the lock will be visible before we release the lock, and
662 // therefore before some other thread (CPU) has a chance to acquire the lock.
663 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
664 //
665 // Critically, any prior STs to _succ or EntryList must be visible before
666 // the ST of null into _owner in the *subsequent* (following) corresponding
667 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily
668 // execute a serializing instruction.
669
670 return;
671 }
672
673 // ReenterI() is a specialized inline form of the latter half of the
674 // contended slow-path from EnterI(). We use ReenterI() only for
675 // monitor reentry in wait().
676 //
677 // In the future we should reconcile EnterI() and ReenterI(), adding
678 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the
679 // loop accordingly.
680
681 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) {
682 assert(Self != NULL, "invariant");
683 assert(SelfNode != NULL, "invariant");
684 assert(SelfNode->_thread == Self, "invariant");
685 assert(_waiters > 0, "invariant");
686 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
687 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
688 JavaThread * jt = (JavaThread *) Self;
689
690 int nWakeups = 0;
691 for (;;) {
692 ObjectWaiter::TStates v = SelfNode->TState;
693 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
694 assert(_owner != Self, "invariant");
695
696 if (TryLock(Self) > 0) break;
697 if (TrySpin(Self) > 0) break;
698
699 // State transition wrappers around park() ...
700 // ReenterI() wisely defers state transitions until
701 // it's clear we must park the thread.
702 {
703 OSThreadContendState osts(Self->osthread());
704 ThreadBlockInVM tbivm(jt);
705
706 // cleared by handle_special_suspend_equivalent_condition()
707 // or java_suspend_self()
708 jt->set_suspend_equivalent();
709 Self->_ParkEvent->park();
710
711 // were we externally suspended while we were waiting?
712 for (;;) {
713 if (!ExitSuspendEquivalent(jt)) break;
714 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
715 jt->java_suspend_self();
716 jt->set_suspend_equivalent();
717 }
718 }
719
720 // Try again, but just so we distinguish between futile wakeups and
721 // successful wakeups. The following test isn't algorithmically
722 // necessary, but it helps us maintain sensible statistics.
723 if (TryLock(Self) > 0) break;
724
725 // The lock is still contested.
726 // Keep a tally of the # of futile wakeups.
727 // Note that the counter is not protected by a lock or updated by atomics.
728 // That is by design - we trade "lossy" counters which are exposed to
729 // races during updates for a lower probe effect.
730 ++nWakeups;
731
732 // Assuming this is not a spurious wakeup we'll normally
733 // find that _succ == Self.
734 if (_succ == Self) _succ = NULL;
735
736 // Invariant: after clearing _succ a contending thread
737 // *must* retry _owner before parking.
738 OrderAccess::fence();
739
740 // This PerfData object can be used in parallel with a safepoint.
741 // See the work around in PerfDataManager::destroy().
742 OM_PERFDATA_OP(FutileWakeups, inc());
743 }
744
745 // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
746 // Normally we'll find Self on the EntryList.
747 // Unlinking from the EntryList is constant-time and atomic-free.
748 // From the perspective of the lock owner (this thread), the
749 // EntryList is stable and cxq is prepend-only.
750 // The head of cxq is volatile but the interior is stable.
751 // In addition, Self.TState is stable.
752
753 assert(_owner == Self, "invariant");
754 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
755 UnlinkAfterAcquire(Self, SelfNode);
756 if (_succ == Self) _succ = NULL;
757 assert(_succ != Self, "invariant");
758 SelfNode->TState = ObjectWaiter::TS_RUN;
759 OrderAccess::fence(); // see comments at the end of EnterI()
760 }
761
762 // By convention we unlink a contending thread from EntryList|cxq immediately
763 // after the thread acquires the lock in ::enter(). Equally, we could defer
764 // unlinking the thread until ::exit()-time.
765
766 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) {
767 assert(_owner == Self, "invariant");
768 assert(SelfNode->_thread == Self, "invariant");
769
770 if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
771 // Normal case: remove Self from the DLL EntryList .
772 // This is a constant-time operation.
773 ObjectWaiter * nxt = SelfNode->_next;
774 ObjectWaiter * prv = SelfNode->_prev;
775 if (nxt != NULL) nxt->_prev = prv;
776 if (prv != NULL) prv->_next = nxt;
777 if (SelfNode == _EntryList) _EntryList = nxt;
778 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
779 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
780 } else {
781 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant");
782 // Inopportune interleaving -- Self is still on the cxq.
783 // This usually means the enqueue of self raced an exiting thread.
784 // Normally we'll find Self near the front of the cxq, so
785 // dequeueing is typically fast. If needbe we can accelerate
786 // this with some MCS/CHL-like bidirectional list hints and advisory
787 // back-links so dequeueing from the interior will normally operate
788 // in constant-time.
789 // Dequeue Self from either the head (with CAS) or from the interior
790 // with a linear-time scan and normal non-atomic memory operations.
791 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
792 // and then unlink Self from EntryList. We have to drain eventually,
793 // so it might as well be now.
794
795 ObjectWaiter * v = _cxq;
796 assert(v != NULL, "invariant");
797 if (v != SelfNode || Atomic::cmpxchg(SelfNode->_next, &_cxq, v) != v) {
798 // The CAS above can fail from interference IFF a "RAT" arrived.
799 // In that case Self must be in the interior and can no longer be
800 // at the head of cxq.
801 if (v == SelfNode) {
802 assert(_cxq != v, "invariant");
803 v = _cxq; // CAS above failed - start scan at head of list
804 }
805 ObjectWaiter * p;
806 ObjectWaiter * q = NULL;
807 for (p = v; p != NULL && p != SelfNode; p = p->_next) {
808 q = p;
809 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
810 }
811 assert(v != SelfNode, "invariant");
812 assert(p == SelfNode, "Node not found on cxq");
813 assert(p != _cxq, "invariant");
814 assert(q != NULL, "invariant");
815 assert(q->_next == p, "invariant");
816 q->_next = p->_next;
817 }
818 }
819
820 #ifdef ASSERT
821 // Diagnostic hygiene ...
822 SelfNode->_prev = (ObjectWaiter *) 0xBAD;
823 SelfNode->_next = (ObjectWaiter *) 0xBAD;
824 SelfNode->TState = ObjectWaiter::TS_RUN;
825 #endif
826 }
827
828 // -----------------------------------------------------------------------------
829 // Exit support
830 //
831 // exit()
832 // ~~~~~~
833 // Note that the collector can't reclaim the objectMonitor or deflate
834 // the object out from underneath the thread calling ::exit() as the
835 // thread calling ::exit() never transitions to a stable state.
836 // This inhibits GC, which in turn inhibits asynchronous (and
837 // inopportune) reclamation of "this".
838 //
839 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
840 // There's one exception to the claim above, however. EnterI() can call
841 // exit() to drop a lock if the acquirer has been externally suspended.
842 // In that case exit() is called with _thread_state as _thread_blocked,
843 // but the monitor's _count field is > 0, which inhibits reclamation.
844 //
845 // 1-0 exit
846 // ~~~~~~~~
847 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
848 // the fast-path operators have been optimized so the common ::exit()
849 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
850 // The code emitted by fast_unlock() elides the usual MEMBAR. This
851 // greatly improves latency -- MEMBAR and CAS having considerable local
852 // latency on modern processors -- but at the cost of "stranding". Absent the
853 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
854 // ::enter() path, resulting in the entering thread being stranding
855 // and a progress-liveness failure. Stranding is extremely rare.
856 // We use timers (timed park operations) & periodic polling to detect
857 // and recover from stranding. Potentially stranded threads periodically
858 // wake up and poll the lock. See the usage of the _Responsible variable.
859 //
860 // The CAS() in enter provides for safety and exclusion, while the CAS or
861 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking
862 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
863 // We detect and recover from stranding with timers.
864 //
865 // If a thread transiently strands it'll park until (a) another
866 // thread acquires the lock and then drops the lock, at which time the
867 // exiting thread will notice and unpark the stranded thread, or, (b)
868 // the timer expires. If the lock is high traffic then the stranding latency
869 // will be low due to (a). If the lock is low traffic then the odds of
870 // stranding are lower, although the worst-case stranding latency
871 // is longer. Critically, we don't want to put excessive load in the
872 // platform's timer subsystem. We want to minimize both the timer injection
873 // rate (timers created/sec) as well as the number of timers active at
874 // any one time. (more precisely, we want to minimize timer-seconds, which is
875 // the integral of the # of active timers at any instant over time).
876 // Both impinge on OS scalability. Given that, at most one thread parked on
877 // a monitor will use a timer.
878 //
879 // There is also the risk of a futile wake-up. If we drop the lock
880 // another thread can reacquire the lock immediately, and we can
881 // then wake a thread unnecessarily. This is benign, and we've
882 // structured the code so the windows are short and the frequency
883 // of such futile wakups is low.
884
885 void ObjectMonitor::exit(bool not_suspended, TRAPS) {
886 Thread * const Self = THREAD;
887 if (THREAD != _owner) {
888 if (THREAD->is_lock_owned((address) _owner)) {
889 // Transmute _owner from a BasicLock pointer to a Thread address.
890 // We don't need to hold _mutex for this transition.
891 // Non-null to Non-null is safe as long as all readers can
892 // tolerate either flavor.
893 assert(_recursions == 0, "invariant");
894 _owner = THREAD;
895 _recursions = 0;
896 } else {
897 // Apparent unbalanced locking ...
898 // Naively we'd like to throw IllegalMonitorStateException.
899 // As a practical matter we can neither allocate nor throw an
900 // exception as ::exit() can be called from leaf routines.
901 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
902 // Upon deeper reflection, however, in a properly run JVM the only
903 // way we should encounter this situation is in the presence of
904 // unbalanced JNI locking. TODO: CheckJNICalls.
905 // See also: CR4414101
906 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking");
907 return;
908 }
909 }
910
911 if (_recursions != 0) {
912 _recursions--; // this is simple recursive enter
913 return;
914 }
915
916 // Invariant: after setting Responsible=null an thread must execute
917 // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
918 _Responsible = NULL;
919
920 #if INCLUDE_JFR
921 // get the owner's thread id for the MonitorEnter event
922 // if it is enabled and the thread isn't suspended
923 if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
924 _previous_owner_tid = JFR_THREAD_ID(Self);
925 }
926 #endif
927
928 for (;;) {
929 assert(THREAD == _owner, "invariant");
930
931 if (Knob_ExitPolicy == 0) {
932 // release semantics: prior loads and stores from within the critical section
933 // must not float (reorder) past the following store that drops the lock.
934 // On SPARC that requires MEMBAR #loadstore|#storestore.
935 // But of course in TSO #loadstore|#storestore is not required.
936 // I'd like to write one of the following:
937 // A. OrderAccess::release() ; _owner = NULL
938 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
939 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
940 // store into a _dummy variable. That store is not needed, but can result
941 // in massive wasteful coherency traffic on classic SMP systems.
942 // Instead, I use release_store(), which is implemented as just a simple
943 // ST on x64, x86 and SPARC.
944 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock
945 OrderAccess::storeload(); // See if we need to wake a successor
946 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
947 return;
948 }
949 // Other threads are blocked trying to acquire the lock.
950
951 // Normally the exiting thread is responsible for ensuring succession,
952 // but if other successors are ready or other entering threads are spinning
953 // then this thread can simply store NULL into _owner and exit without
954 // waking a successor. The existence of spinners or ready successors
955 // guarantees proper succession (liveness). Responsibility passes to the
956 // ready or running successors. The exiting thread delegates the duty.
957 // More precisely, if a successor already exists this thread is absolved
958 // of the responsibility of waking (unparking) one.
959 //
960 // The _succ variable is critical to reducing futile wakeup frequency.
961 // _succ identifies the "heir presumptive" thread that has been made
962 // ready (unparked) but that has not yet run. We need only one such
963 // successor thread to guarantee progress.
964 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
965 // section 3.3 "Futile Wakeup Throttling" for details.
966 //
967 // Note that spinners in Enter() also set _succ non-null.
968 // In the current implementation spinners opportunistically set
969 // _succ so that exiting threads might avoid waking a successor.
970 // Another less appealing alternative would be for the exiting thread
971 // to drop the lock and then spin briefly to see if a spinner managed
972 // to acquire the lock. If so, the exiting thread could exit
973 // immediately without waking a successor, otherwise the exiting
974 // thread would need to dequeue and wake a successor.
975 // (Note that we'd need to make the post-drop spin short, but no
976 // shorter than the worst-case round-trip cache-line migration time.
977 // The dropped lock needs to become visible to the spinner, and then
978 // the acquisition of the lock by the spinner must become visible to
979 // the exiting thread).
980
981 // It appears that an heir-presumptive (successor) must be made ready.
982 // Only the current lock owner can manipulate the EntryList or
983 // drain _cxq, so we need to reacquire the lock. If we fail
984 // to reacquire the lock the responsibility for ensuring succession
985 // falls to the new owner.
986 //
987 if (!Atomic::replace_if_null(THREAD, &_owner)) {
988 return;
989 }
990 } else {
991 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
992 OrderAccess::release_store(&_owner, (void*)NULL); // drop the lock
993 OrderAccess::storeload();
994 // Ratify the previously observed values.
995 if (_cxq == NULL || _succ != NULL) {
996 return;
997 }
998
999 // inopportune interleaving -- the exiting thread (this thread)
1000 // in the fast-exit path raced an entering thread in the slow-enter
1001 // path.
1002 // We have two choices:
1003 // A. Try to reacquire the lock.
1004 // If the CAS() fails return immediately, otherwise
1005 // we either restart/rerun the exit operation, or simply
1006 // fall-through into the code below which wakes a successor.
1007 // B. If the elements forming the EntryList|cxq are TSM
1008 // we could simply unpark() the lead thread and return
1009 // without having set _succ.
1010 if (!Atomic::replace_if_null(THREAD, &_owner)) {
1011 return;
1012 }
1013 }
1014 }
1015
1016 guarantee(_owner == THREAD, "invariant");
1017
1018 ObjectWaiter * w = NULL;
1019 int QMode = Knob_QMode;
1020
1021 if (QMode == 2 && _cxq != NULL) {
1022 // QMode == 2 : cxq has precedence over EntryList.
1023 // Try to directly wake a successor from the cxq.
1024 // If successful, the successor will need to unlink itself from cxq.
1025 w = _cxq;
1026 assert(w != NULL, "invariant");
1027 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant");
1028 ExitEpilog(Self, w);
1029 return;
1030 }
1031
1032 if (QMode == 3 && _cxq != NULL) {
1033 // Aggressively drain cxq into EntryList at the first opportunity.
1034 // This policy ensure that recently-run threads live at the head of EntryList.
1035 // Drain _cxq into EntryList - bulk transfer.
1036 // First, detach _cxq.
1037 // The following loop is tantamount to: w = swap(&cxq, NULL)
1038 w = _cxq;
1039 for (;;) {
1040 assert(w != NULL, "Invariant");
1041 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1042 if (u == w) break;
1043 w = u;
1044 }
1045 assert(w != NULL, "invariant");
1046
1047 ObjectWaiter * q = NULL;
1048 ObjectWaiter * p;
1049 for (p = w; p != NULL; p = p->_next) {
1050 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1051 p->TState = ObjectWaiter::TS_ENTER;
1052 p->_prev = q;
1053 q = p;
1054 }
1055
1056 // Append the RATs to the EntryList
1057 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1058 ObjectWaiter * Tail;
1059 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL;
1060 Tail = Tail->_next)
1061 /* empty */;
1062 if (Tail == NULL) {
1063 _EntryList = w;
1064 } else {
1065 Tail->_next = w;
1066 w->_prev = Tail;
1067 }
1068
1069 // Fall thru into code that tries to wake a successor from EntryList
1070 }
1071
1072 if (QMode == 4 && _cxq != NULL) {
1073 // Aggressively drain cxq into EntryList at the first opportunity.
1074 // This policy ensure that recently-run threads live at the head of EntryList.
1075
1076 // Drain _cxq into EntryList - bulk transfer.
1077 // First, detach _cxq.
1078 // The following loop is tantamount to: w = swap(&cxq, NULL)
1079 w = _cxq;
1080 for (;;) {
1081 assert(w != NULL, "Invariant");
1082 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1083 if (u == w) break;
1084 w = u;
1085 }
1086 assert(w != NULL, "invariant");
1087
1088 ObjectWaiter * q = NULL;
1089 ObjectWaiter * p;
1090 for (p = w; p != NULL; p = p->_next) {
1091 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1092 p->TState = ObjectWaiter::TS_ENTER;
1093 p->_prev = q;
1094 q = p;
1095 }
1096
1097 // Prepend the RATs to the EntryList
1098 if (_EntryList != NULL) {
1099 q->_next = _EntryList;
1100 _EntryList->_prev = q;
1101 }
1102 _EntryList = w;
1103
1104 // Fall thru into code that tries to wake a successor from EntryList
1105 }
1106
1107 w = _EntryList;
1108 if (w != NULL) {
1109 // I'd like to write: guarantee (w->_thread != Self).
1110 // But in practice an exiting thread may find itself on the EntryList.
1111 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and
1112 // then calls exit(). Exit release the lock by setting O._owner to NULL.
1113 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The
1114 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1115 // release the lock "O". T2 resumes immediately after the ST of null into
1116 // _owner, above. T2 notices that the EntryList is populated, so it
1117 // reacquires the lock and then finds itself on the EntryList.
1118 // Given all that, we have to tolerate the circumstance where "w" is
1119 // associated with Self.
1120 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1121 ExitEpilog(Self, w);
1122 return;
1123 }
1124
1125 // If we find that both _cxq and EntryList are null then just
1126 // re-run the exit protocol from the top.
1127 w = _cxq;
1128 if (w == NULL) continue;
1129
1130 // Drain _cxq into EntryList - bulk transfer.
1131 // First, detach _cxq.
1132 // The following loop is tantamount to: w = swap(&cxq, NULL)
1133 for (;;) {
1134 assert(w != NULL, "Invariant");
1135 ObjectWaiter * u = Atomic::cmpxchg((ObjectWaiter*)NULL, &_cxq, w);
1136 if (u == w) break;
1137 w = u;
1138 }
1139
1140 assert(w != NULL, "invariant");
1141 assert(_EntryList == NULL, "invariant");
1142
1143 // Convert the LIFO SLL anchored by _cxq into a DLL.
1144 // The list reorganization step operates in O(LENGTH(w)) time.
1145 // It's critical that this step operate quickly as
1146 // "Self" still holds the outer-lock, restricting parallelism
1147 // and effectively lengthening the critical section.
1148 // Invariant: s chases t chases u.
1149 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1150 // we have faster access to the tail.
1151
1152 if (QMode == 1) {
1153 // QMode == 1 : drain cxq to EntryList, reversing order
1154 // We also reverse the order of the list.
1155 ObjectWaiter * s = NULL;
1156 ObjectWaiter * t = w;
1157 ObjectWaiter * u = NULL;
1158 while (t != NULL) {
1159 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant");
1160 t->TState = ObjectWaiter::TS_ENTER;
1161 u = t->_next;
1162 t->_prev = u;
1163 t->_next = s;
1164 s = t;
1165 t = u;
1166 }
1167 _EntryList = s;
1168 assert(s != NULL, "invariant");
1169 } else {
1170 // QMode == 0 or QMode == 2
1171 _EntryList = w;
1172 ObjectWaiter * q = NULL;
1173 ObjectWaiter * p;
1174 for (p = w; p != NULL; p = p->_next) {
1175 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1176 p->TState = ObjectWaiter::TS_ENTER;
1177 p->_prev = q;
1178 q = p;
1179 }
1180 }
1181
1182 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1183 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1184
1185 // See if we can abdicate to a spinner instead of waking a thread.
1186 // A primary goal of the implementation is to reduce the
1187 // context-switch rate.
1188 if (_succ != NULL) continue;
1189
1190 w = _EntryList;
1191 if (w != NULL) {
1192 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1193 ExitEpilog(Self, w);
1194 return;
1195 }
1196 }
1197 }
1198
1199 // ExitSuspendEquivalent:
1200 // A faster alternate to handle_special_suspend_equivalent_condition()
1201 //
1202 // handle_special_suspend_equivalent_condition() unconditionally
1203 // acquires the SR_lock. On some platforms uncontended MutexLocker()
1204 // operations have high latency. Note that in ::enter() we call HSSEC
1205 // while holding the monitor, so we effectively lengthen the critical sections.
1206 //
1207 // There are a number of possible solutions:
1208 //
1209 // A. To ameliorate the problem we might also defer state transitions
1210 // to as late as possible -- just prior to parking.
1211 // Given that, we'd call HSSEC after having returned from park(),
1212 // but before attempting to acquire the monitor. This is only a
1213 // partial solution. It avoids calling HSSEC while holding the
1214 // monitor (good), but it still increases successor reacquisition latency --
1215 // the interval between unparking a successor and the time the successor
1216 // resumes and retries the lock. See ReenterI(), which defers state transitions.
1217 // If we use this technique we can also avoid EnterI()-exit() loop
1218 // in ::enter() where we iteratively drop the lock and then attempt
1219 // to reacquire it after suspending.
1220 //
1221 // B. In the future we might fold all the suspend bits into a
1222 // composite per-thread suspend flag and then update it with CAS().
1223 // Alternately, a Dekker-like mechanism with multiple variables
1224 // would suffice:
1225 // ST Self->_suspend_equivalent = false
1226 // MEMBAR
1227 // LD Self_>_suspend_flags
1228 //
1229 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem
1230 // with a more efficient implementation, the need to use "FastHSSEC" has
1231 // decreased. - Dave
1232
1233
1234 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) {
1235 const int Mode = Knob_FastHSSEC;
1236 if (Mode && !jSelf->is_external_suspend()) {
1237 assert(jSelf->is_suspend_equivalent(), "invariant");
1238 jSelf->clear_suspend_equivalent();
1239 if (2 == Mode) OrderAccess::storeload();
1240 if (!jSelf->is_external_suspend()) return false;
1241 // We raced a suspension -- fall thru into the slow path
1242 jSelf->set_suspend_equivalent();
1243 }
1244 return jSelf->handle_special_suspend_equivalent_condition();
1245 }
1246
1247
1248 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) {
1249 assert(_owner == Self, "invariant");
1250
1251 // Exit protocol:
1252 // 1. ST _succ = wakee
1253 // 2. membar #loadstore|#storestore;
1254 // 2. ST _owner = NULL
1255 // 3. unpark(wakee)
1256
1257 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL;
1258 ParkEvent * Trigger = Wakee->_event;
1259
1260 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1261 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1262 // out-of-scope (non-extant).
1263 Wakee = NULL;
1264
1265 // Drop the lock
1266 OrderAccess::release_store(&_owner, (void*)NULL);
1267 OrderAccess::fence(); // ST _owner vs LD in unpark()
1268
1269 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1270 Trigger->unpark();
1271
1272 // Maintain stats and report events to JVMTI
1273 OM_PERFDATA_OP(Parks, inc());
1274 }
1275
1276
1277 // -----------------------------------------------------------------------------
1278 // Class Loader deadlock handling.
1279 //
1280 // complete_exit exits a lock returning recursion count
1281 // complete_exit/reenter operate as a wait without waiting
1282 // complete_exit requires an inflated monitor
1283 // The _owner field is not always the Thread addr even with an
1284 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1285 // thread due to contention.
1286 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1287 Thread * const Self = THREAD;
1288 assert(Self->is_Java_thread(), "Must be Java thread!");
1289 JavaThread *jt = (JavaThread *)THREAD;
1290
1291 DeferredInitialize();
1292
1293 if (THREAD != _owner) {
1294 if (THREAD->is_lock_owned ((address)_owner)) {
1295 assert(_recursions == 0, "internal state error");
1296 _owner = THREAD; // Convert from basiclock addr to Thread addr
1297 _recursions = 0;
1298 }
1299 }
1300
1301 guarantee(Self == _owner, "complete_exit not owner");
1302 intptr_t save = _recursions; // record the old recursion count
1303 _recursions = 0; // set the recursion level to be 0
1304 exit(true, Self); // exit the monitor
1305 guarantee(_owner != Self, "invariant");
1306 return save;
1307 }
1308
1309 // reenter() enters a lock and sets recursion count
1310 // complete_exit/reenter operate as a wait without waiting
1311 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1312 Thread * const Self = THREAD;
1313 assert(Self->is_Java_thread(), "Must be Java thread!");
1314 JavaThread *jt = (JavaThread *)THREAD;
1315
1316 guarantee(_owner != Self, "reenter already owner");
1317 enter(THREAD); // enter the monitor
1318 guarantee(_recursions == 0, "reenter recursion");
1319 _recursions = recursions;
1320 return;
1321 }
1322
1323
1324 // -----------------------------------------------------------------------------
1325 // A macro is used below because there may already be a pending
1326 // exception which should not abort the execution of the routines
1327 // which use this (which is why we don't put this into check_slow and
1328 // call it with a CHECK argument).
1329
1330 #define CHECK_OWNER() \
1331 do { \
1332 if (THREAD != _owner) { \
1333 if (THREAD->is_lock_owned((address) _owner)) { \
1334 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \
1335 _recursions = 0; \
1336 } else { \
1337 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \
1338 } \
1339 } \
1340 } while (false)
1341
1342 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception.
1343 // TODO-FIXME: remove check_slow() -- it's likely dead.
1344
1345 void ObjectMonitor::check_slow(TRAPS) {
1346 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1347 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1348 }
1349
1350 static int Adjust(volatile int * adr, int dx) {
1351 int v;
1352 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */;
1353 return v;
1354 }
1355
1356 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1357 ObjectMonitor* monitor,
1358 jlong notifier_tid,
1359 jlong timeout,
1360 bool timedout) {
1361 assert(event != NULL, "invariant");
1362 assert(monitor != NULL, "invariant");
1363 event->set_monitorClass(((oop)monitor->object())->klass());
1364 event->set_timeout(timeout);
1365 event->set_address((uintptr_t)monitor->object_addr());
1366 event->set_notifier(notifier_tid);
1367 event->set_timedOut(timedout);
1368 event->commit();
1369 }
1370
1371 // -----------------------------------------------------------------------------
1372 // Wait/Notify/NotifyAll
1373 //
1374 // Note: a subset of changes to ObjectMonitor::wait()
1375 // will need to be replicated in complete_exit
1376 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1377 Thread * const Self = THREAD;
1378 assert(Self->is_Java_thread(), "Must be Java thread!");
1379 JavaThread *jt = (JavaThread *)THREAD;
1380
1381 DeferredInitialize();
1382
1383 // Throw IMSX or IEX.
1384 CHECK_OWNER();
1385
1386 EventJavaMonitorWait event;
1387
1388 // check for a pending interrupt
1389 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1390 // post monitor waited event. Note that this is past-tense, we are done waiting.
1391 if (JvmtiExport::should_post_monitor_waited()) {
1392 // Note: 'false' parameter is passed here because the
1393 // wait was not timed out due to thread interrupt.
1394 JvmtiExport::post_monitor_waited(jt, this, false);
1395
1396 // In this short circuit of the monitor wait protocol, the
1397 // current thread never drops ownership of the monitor and
1398 // never gets added to the wait queue so the current thread
1399 // cannot be made the successor. This means that the
1400 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1401 // consume an unpark() meant for the ParkEvent associated with
1402 // this ObjectMonitor.
1403 }
1404 if (event.should_commit()) {
1405 post_monitor_wait_event(&event, this, 0, millis, false);
1406 }
1407 THROW(vmSymbols::java_lang_InterruptedException());
1408 return;
1409 }
1410
1411 assert(Self->_Stalled == 0, "invariant");
1412 Self->_Stalled = intptr_t(this);
1413 jt->set_current_waiting_monitor(this);
1414
1415 // create a node to be put into the queue
1416 // Critically, after we reset() the event but prior to park(), we must check
1417 // for a pending interrupt.
1418 ObjectWaiter node(Self);
1419 node.TState = ObjectWaiter::TS_WAIT;
1420 Self->_ParkEvent->reset();
1421 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag
1422
1423 // Enter the waiting queue, which is a circular doubly linked list in this case
1424 // but it could be a priority queue or any data structure.
1425 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only
1426 // by the the owner of the monitor *except* in the case where park()
1427 // returns because of a timeout of interrupt. Contention is exceptionally rare
1428 // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1429
1430 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1431 AddWaiter(&node);
1432 Thread::SpinRelease(&_WaitSetLock);
1433
1434 _Responsible = NULL;
1435
1436 intptr_t save = _recursions; // record the old recursion count
1437 _waiters++; // increment the number of waiters
1438 _recursions = 0; // set the recursion level to be 1
1439 exit(true, Self); // exit the monitor
1440 guarantee(_owner != Self, "invariant");
1441
1442 // The thread is on the WaitSet list - now park() it.
1443 // On MP systems it's conceivable that a brief spin before we park
1444 // could be profitable.
1445 //
1446 // TODO-FIXME: change the following logic to a loop of the form
1447 // while (!timeout && !interrupted && _notified == 0) park()
1448
1449 int ret = OS_OK;
1450 int WasNotified = 0;
1451 { // State transition wrappers
1452 OSThread* osthread = Self->osthread();
1453 OSThreadWaitState osts(osthread, true);
1454 {
1455 ThreadBlockInVM tbivm(jt);
1456 // Thread is in thread_blocked state and oop access is unsafe.
1457 jt->set_suspend_equivalent();
1458
1459 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1460 // Intentionally empty
1461 } else if (node._notified == 0) {
1462 if (millis <= 0) {
1463 Self->_ParkEvent->park();
1464 } else {
1465 ret = Self->_ParkEvent->park(millis);
1466 }
1467 }
1468
1469 // were we externally suspended while we were waiting?
1470 if (ExitSuspendEquivalent (jt)) {
1471 // TODO-FIXME: add -- if succ == Self then succ = null.
1472 jt->java_suspend_self();
1473 }
1474
1475 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1476
1477 // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1478 // from the WaitSet to the EntryList.
1479 // See if we need to remove Node from the WaitSet.
1480 // We use double-checked locking to avoid grabbing _WaitSetLock
1481 // if the thread is not on the wait queue.
1482 //
1483 // Note that we don't need a fence before the fetch of TState.
1484 // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1485 // written by the is thread. (perhaps the fetch might even be satisfied
1486 // by a look-aside into the processor's own store buffer, although given
1487 // the length of the code path between the prior ST and this load that's
1488 // highly unlikely). If the following LD fetches a stale TS_WAIT value
1489 // then we'll acquire the lock and then re-fetch a fresh TState value.
1490 // That is, we fail toward safety.
1491
1492 if (node.TState == ObjectWaiter::TS_WAIT) {
1493 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1494 if (node.TState == ObjectWaiter::TS_WAIT) {
1495 DequeueSpecificWaiter(&node); // unlink from WaitSet
1496 assert(node._notified == 0, "invariant");
1497 node.TState = ObjectWaiter::TS_RUN;
1498 }
1499 Thread::SpinRelease(&_WaitSetLock);
1500 }
1501
1502 // The thread is now either on off-list (TS_RUN),
1503 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1504 // The Node's TState variable is stable from the perspective of this thread.
1505 // No other threads will asynchronously modify TState.
1506 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1507 OrderAccess::loadload();
1508 if (_succ == Self) _succ = NULL;
1509 WasNotified = node._notified;
1510
1511 // Reentry phase -- reacquire the monitor.
1512 // re-enter contended monitor after object.wait().
1513 // retain OBJECT_WAIT state until re-enter successfully completes
1514 // Thread state is thread_in_vm and oop access is again safe,
1515 // although the raw address of the object may have changed.
1516 // (Don't cache naked oops over safepoints, of course).
1517
1518 // post monitor waited event. Note that this is past-tense, we are done waiting.
1519 if (JvmtiExport::should_post_monitor_waited()) {
1520 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1521
1522 if (node._notified != 0 && _succ == Self) {
1523 // In this part of the monitor wait-notify-reenter protocol it
1524 // is possible (and normal) for another thread to do a fastpath
1525 // monitor enter-exit while this thread is still trying to get
1526 // to the reenter portion of the protocol.
1527 //
1528 // The ObjectMonitor was notified and the current thread is
1529 // the successor which also means that an unpark() has already
1530 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1531 // consume the unpark() that was done when the successor was
1532 // set because the same ParkEvent is shared between Java
1533 // monitors and JVM/TI RawMonitors (for now).
1534 //
1535 // We redo the unpark() to ensure forward progress, i.e., we
1536 // don't want all pending threads hanging (parked) with none
1537 // entering the unlocked monitor.
1538 node._event->unpark();
1539 }
1540 }
1541
1542 if (event.should_commit()) {
1543 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1544 }
1545
1546 OrderAccess::fence();
1547
1548 assert(Self->_Stalled != 0, "invariant");
1549 Self->_Stalled = 0;
1550
1551 assert(_owner != Self, "invariant");
1552 ObjectWaiter::TStates v = node.TState;
1553 if (v == ObjectWaiter::TS_RUN) {
1554 enter(Self);
1555 } else {
1556 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1557 ReenterI(Self, &node);
1558 node.wait_reenter_end(this);
1559 }
1560
1561 // Self has reacquired the lock.
1562 // Lifecycle - the node representing Self must not appear on any queues.
1563 // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1564 // want residual elements associated with this thread left on any lists.
1565 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1566 assert(_owner == Self, "invariant");
1567 assert(_succ != Self, "invariant");
1568 } // OSThreadWaitState()
1569
1570 jt->set_current_waiting_monitor(NULL);
1571
1572 guarantee(_recursions == 0, "invariant");
1573 _recursions = save; // restore the old recursion count
1574 _waiters--; // decrement the number of waiters
1575
1576 // Verify a few postconditions
1577 assert(_owner == Self, "invariant");
1578 assert(_succ != Self, "invariant");
1579 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
1580
1581 // check if the notification happened
1582 if (!WasNotified) {
1583 // no, it could be timeout or Thread.interrupt() or both
1584 // check for interrupt event, otherwise it is timeout
1585 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1586 THROW(vmSymbols::java_lang_InterruptedException());
1587 }
1588 }
1589
1590 // NOTE: Spurious wake up will be consider as timeout.
1591 // Monitor notify has precedence over thread interrupt.
1592 }
1593
1594
1595 // Consider:
1596 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1597 // then instead of transferring a thread from the WaitSet to the EntryList
1598 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1599
1600 void ObjectMonitor::INotify(Thread * Self) {
1601 const int policy = Knob_MoveNotifyee;
1602
1603 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1604 ObjectWaiter * iterator = DequeueWaiter();
1605 if (iterator != NULL) {
1606 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1607 guarantee(iterator->_notified == 0, "invariant");
1608 // Disposition - what might we do with iterator ?
1609 // a. add it directly to the EntryList - either tail (policy == 1)
1610 // or head (policy == 0).
1611 // b. push it onto the front of the _cxq (policy == 2).
1612 // For now we use (b).
1613 if (policy != 4) {
1614 iterator->TState = ObjectWaiter::TS_ENTER;
1615 }
1616 iterator->_notified = 1;
1617 iterator->_notifier_tid = JFR_THREAD_ID(Self);
1618
1619 ObjectWaiter * list = _EntryList;
1620 if (list != NULL) {
1621 assert(list->_prev == NULL, "invariant");
1622 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1623 assert(list != iterator, "invariant");
1624 }
1625
1626 if (policy == 0) { // prepend to EntryList
1627 if (list == NULL) {
1628 iterator->_next = iterator->_prev = NULL;
1629 _EntryList = iterator;
1630 } else {
1631 list->_prev = iterator;
1632 iterator->_next = list;
1633 iterator->_prev = NULL;
1634 _EntryList = iterator;
1635 }
1636 } else if (policy == 1) { // append to EntryList
1637 if (list == NULL) {
1638 iterator->_next = iterator->_prev = NULL;
1639 _EntryList = iterator;
1640 } else {
1641 // CONSIDER: finding the tail currently requires a linear-time walk of
1642 // the EntryList. We can make tail access constant-time by converting to
1643 // a CDLL instead of using our current DLL.
1644 ObjectWaiter * tail;
1645 for (tail = list; tail->_next != NULL; tail = tail->_next) {}
1646 assert(tail != NULL && tail->_next == NULL, "invariant");
1647 tail->_next = iterator;
1648 iterator->_prev = tail;
1649 iterator->_next = NULL;
1650 }
1651 } else if (policy == 2) { // prepend to cxq
1652 if (list == NULL) {
1653 iterator->_next = iterator->_prev = NULL;
1654 _EntryList = iterator;
1655 } else {
1656 iterator->TState = ObjectWaiter::TS_CXQ;
1657 for (;;) {
1658 ObjectWaiter * front = _cxq;
1659 iterator->_next = front;
1660 if (Atomic::cmpxchg(iterator, &_cxq, front) == front) {
1661 break;
1662 }
1663 }
1664 }
1665 } else if (policy == 3) { // append to cxq
1666 iterator->TState = ObjectWaiter::TS_CXQ;
1667 for (;;) {
1668 ObjectWaiter * tail = _cxq;
1669 if (tail == NULL) {
1670 iterator->_next = NULL;
1671 if (Atomic::replace_if_null(iterator, &_cxq)) {
1672 break;
1673 }
1674 } else {
1675 while (tail->_next != NULL) tail = tail->_next;
1676 tail->_next = iterator;
1677 iterator->_prev = tail;
1678 iterator->_next = NULL;
1679 break;
1680 }
1681 }
1682 } else {
1683 ParkEvent * ev = iterator->_event;
1684 iterator->TState = ObjectWaiter::TS_RUN;
1685 OrderAccess::fence();
1686 ev->unpark();
1687 }
1688
1689 // _WaitSetLock protects the wait queue, not the EntryList. We could
1690 // move the add-to-EntryList operation, above, outside the critical section
1691 // protected by _WaitSetLock. In practice that's not useful. With the
1692 // exception of wait() timeouts and interrupts the monitor owner
1693 // is the only thread that grabs _WaitSetLock. There's almost no contention
1694 // on _WaitSetLock so it's not profitable to reduce the length of the
1695 // critical section.
1696
1697 if (policy < 4) {
1698 iterator->wait_reenter_begin(this);
1699 }
1700 }
1701 Thread::SpinRelease(&_WaitSetLock);
1702 }
1703
1704 // Consider: a not-uncommon synchronization bug is to use notify() when
1705 // notifyAll() is more appropriate, potentially resulting in stranded
1706 // threads; this is one example of a lost wakeup. A useful diagnostic
1707 // option is to force all notify() operations to behave as notifyAll().
1708 //
1709 // Note: We can also detect many such problems with a "minimum wait".
1710 // When the "minimum wait" is set to a small non-zero timeout value
1711 // and the program does not hang whereas it did absent "minimum wait",
1712 // that suggests a lost wakeup bug.
1713
1714 void ObjectMonitor::notify(TRAPS) {
1715 CHECK_OWNER();
1716 if (_WaitSet == NULL) {
1717 return;
1718 }
1719 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1720 INotify(THREAD);
1721 OM_PERFDATA_OP(Notifications, inc(1));
1722 }
1723
1724
1725 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1726 // from the waitset to the EntryList. This could be done more efficiently with a
1727 // single bulk transfer but in practice it's not time-critical. Beware too,
1728 // that in prepend-mode we invert the order of the waiters. Let's say that the
1729 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1730 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1731
1732 void ObjectMonitor::notifyAll(TRAPS) {
1733 CHECK_OWNER();
1734 if (_WaitSet == NULL) {
1735 return;
1736 }
1737
1738 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1739 int tally = 0;
1740 while (_WaitSet != NULL) {
1741 tally++;
1742 INotify(THREAD);
1743 }
1744
1745 OM_PERFDATA_OP(Notifications, inc(tally));
1746 }
1747
1748 // -----------------------------------------------------------------------------
1749 // Adaptive Spinning Support
1750 //
1751 // Adaptive spin-then-block - rational spinning
1752 //
1753 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1754 // algorithm. On high order SMP systems it would be better to start with
1755 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH,
1756 // a contending thread could enqueue itself on the cxq and then spin locally
1757 // on a thread-specific variable such as its ParkEvent._Event flag.
1758 // That's left as an exercise for the reader. Note that global spinning is
1759 // not problematic on Niagara, as the L2 cache serves the interconnect and
1760 // has both low latency and massive bandwidth.
1761 //
1762 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1763 // acquisition attempts where we opt to spin -- at 100% and vary the spin count
1764 // (duration) or we can fix the count at approximately the duration of
1765 // a context switch and vary the frequency. Of course we could also
1766 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1767 // For a description of 'Adaptive spin-then-block mutual exclusion in
1768 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1769 //
1770 // This implementation varies the duration "D", where D varies with
1771 // the success rate of recent spin attempts. (D is capped at approximately
1772 // length of a round-trip context switch). The success rate for recent
1773 // spin attempts is a good predictor of the success rate of future spin
1774 // attempts. The mechanism adapts automatically to varying critical
1775 // section length (lock modality), system load and degree of parallelism.
1776 // D is maintained per-monitor in _SpinDuration and is initialized
1777 // optimistically. Spin frequency is fixed at 100%.
1778 //
1779 // Note that _SpinDuration is volatile, but we update it without locks
1780 // or atomics. The code is designed so that _SpinDuration stays within
1781 // a reasonable range even in the presence of races. The arithmetic
1782 // operations on _SpinDuration are closed over the domain of legal values,
1783 // so at worst a race will install and older but still legal value.
1784 // At the very worst this introduces some apparent non-determinism.
1785 // We might spin when we shouldn't or vice-versa, but since the spin
1786 // count are relatively short, even in the worst case, the effect is harmless.
1787 //
1788 // Care must be taken that a low "D" value does not become an
1789 // an absorbing state. Transient spinning failures -- when spinning
1790 // is overall profitable -- should not cause the system to converge
1791 // on low "D" values. We want spinning to be stable and predictable
1792 // and fairly responsive to change and at the same time we don't want
1793 // it to oscillate, become metastable, be "too" non-deterministic,
1794 // or converge on or enter undesirable stable absorbing states.
1795 //
1796 // We implement a feedback-based control system -- using past behavior
1797 // to predict future behavior. We face two issues: (a) if the
1798 // input signal is random then the spin predictor won't provide optimal
1799 // results, and (b) if the signal frequency is too high then the control
1800 // system, which has some natural response lag, will "chase" the signal.
1801 // (b) can arise from multimodal lock hold times. Transient preemption
1802 // can also result in apparent bimodal lock hold times.
1803 // Although sub-optimal, neither condition is particularly harmful, as
1804 // in the worst-case we'll spin when we shouldn't or vice-versa.
1805 // The maximum spin duration is rather short so the failure modes aren't bad.
1806 // To be conservative, I've tuned the gain in system to bias toward
1807 // _not spinning. Relatedly, the system can sometimes enter a mode where it
1808 // "rings" or oscillates between spinning and not spinning. This happens
1809 // when spinning is just on the cusp of profitability, however, so the
1810 // situation is not dire. The state is benign -- there's no need to add
1811 // hysteresis control to damp the transition rate between spinning and
1812 // not spinning.
1813
1814 // Spinning: Fixed frequency (100%), vary duration
1815 int ObjectMonitor::TrySpin(Thread * Self) {
1816 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning.
1817 int ctr = Knob_FixedSpin;
1818 if (ctr != 0) {
1819 while (--ctr >= 0) {
1820 if (TryLock(Self) > 0) return 1;
1821 SpinPause();
1822 }
1823 return 0;
1824 }
1825
1826 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1827 if (TryLock(Self) > 0) {
1828 // Increase _SpinDuration ...
1829 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1830 // Raising _SpurDuration to the poverty line is key.
1831 int x = _SpinDuration;
1832 if (x < Knob_SpinLimit) {
1833 if (x < Knob_Poverty) x = Knob_Poverty;
1834 _SpinDuration = x + Knob_BonusB;
1835 }
1836 return 1;
1837 }
1838 SpinPause();
1839 }
1840
1841 // Admission control - verify preconditions for spinning
1842 //
1843 // We always spin a little bit, just to prevent _SpinDuration == 0 from
1844 // becoming an absorbing state. Put another way, we spin briefly to
1845 // sample, just in case the system load, parallelism, contention, or lock
1846 // modality changed.
1847 //
1848 // Consider the following alternative:
1849 // Periodically set _SpinDuration = _SpinLimit and try a long/full
1850 // spin attempt. "Periodically" might mean after a tally of
1851 // the # of failed spin attempts (or iterations) reaches some threshold.
1852 // This takes us into the realm of 1-out-of-N spinning, where we
1853 // hold the duration constant but vary the frequency.
1854
1855 ctr = _SpinDuration;
1856 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase;
1857 if (ctr <= 0) return 0;
1858
1859 if (Knob_SuccRestrict && _succ != NULL) return 0;
1860 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
1861 return 0;
1862 }
1863
1864 int MaxSpin = Knob_MaxSpinners;
1865 if (MaxSpin >= 0) {
1866 if (_Spinner > MaxSpin) {
1867 return 0;
1868 }
1869 // Slightly racy, but benign ...
1870 Adjust(&_Spinner, 1);
1871 }
1872
1873 // We're good to spin ... spin ingress.
1874 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1875 // when preparing to LD...CAS _owner, etc and the CAS is likely
1876 // to succeed.
1877 int hits = 0;
1878 int msk = 0;
1879 int caspty = Knob_CASPenalty;
1880 int oxpty = Knob_OXPenalty;
1881 int sss = Knob_SpinSetSucc;
1882 if (sss && _succ == NULL) _succ = Self;
1883 Thread * prv = NULL;
1884
1885 // There are three ways to exit the following loop:
1886 // 1. A successful spin where this thread has acquired the lock.
1887 // 2. Spin failure with prejudice
1888 // 3. Spin failure without prejudice
1889
1890 while (--ctr >= 0) {
1891
1892 // Periodic polling -- Check for pending GC
1893 // Threads may spin while they're unsafe.
1894 // We don't want spinning threads to delay the JVM from reaching
1895 // a stop-the-world safepoint or to steal cycles from GC.
1896 // If we detect a pending safepoint we abort in order that
1897 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1898 // this thread, if safe, doesn't steal cycles from GC.
1899 // This is in keeping with the "no loitering in runtime" rule.
1900 // We periodically check to see if there's a safepoint pending.
1901 if ((ctr & 0xFF) == 0) {
1902 if (SafepointMechanism::poll(Self)) {
1903 goto Abort; // abrupt spin egress
1904 }
1905 if (Knob_UsePause & 1) SpinPause();
1906 }
1907
1908 if (Knob_UsePause & 2) SpinPause();
1909
1910 // Exponential back-off ... Stay off the bus to reduce coherency traffic.
1911 // This is useful on classic SMP systems, but is of less utility on
1912 // N1-style CMT platforms.
1913 //
1914 // Trade-off: lock acquisition latency vs coherency bandwidth.
1915 // Lock hold times are typically short. A histogram
1916 // of successful spin attempts shows that we usually acquire
1917 // the lock early in the spin. That suggests we want to
1918 // sample _owner frequently in the early phase of the spin,
1919 // but then back-off and sample less frequently as the spin
1920 // progresses. The back-off makes a good citizen on SMP big
1921 // SMP systems. Oversampling _owner can consume excessive
1922 // coherency bandwidth. Relatedly, if we _oversample _owner we
1923 // can inadvertently interfere with the the ST m->owner=null.
1924 // executed by the lock owner.
1925 if (ctr & msk) continue;
1926 ++hits;
1927 if ((hits & 0xF) == 0) {
1928 // The 0xF, above, corresponds to the exponent.
1929 // Consider: (msk+1)|msk
1930 msk = ((msk << 2)|3) & BackOffMask;
1931 }
1932
1933 // Probe _owner with TATAS
1934 // If this thread observes the monitor transition or flicker
1935 // from locked to unlocked to locked, then the odds that this
1936 // thread will acquire the lock in this spin attempt go down
1937 // considerably. The same argument applies if the CAS fails
1938 // or if we observe _owner change from one non-null value to
1939 // another non-null value. In such cases we might abort
1940 // the spin without prejudice or apply a "penalty" to the
1941 // spin count-down variable "ctr", reducing it by 100, say.
1942
1943 Thread * ox = (Thread *) _owner;
1944 if (ox == NULL) {
1945 ox = (Thread*)Atomic::cmpxchg(Self, &_owner, (void*)NULL);
1946 if (ox == NULL) {
1947 // The CAS succeeded -- this thread acquired ownership
1948 // Take care of some bookkeeping to exit spin state.
1949 if (sss && _succ == Self) {
1950 _succ = NULL;
1951 }
1952 if (MaxSpin > 0) Adjust(&_Spinner, -1);
1953
1954 // Increase _SpinDuration :
1955 // The spin was successful (profitable) so we tend toward
1956 // longer spin attempts in the future.
1957 // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1958 // If we acquired the lock early in the spin cycle it
1959 // makes sense to increase _SpinDuration proportionally.
1960 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1961 int x = _SpinDuration;
1962 if (x < Knob_SpinLimit) {
1963 if (x < Knob_Poverty) x = Knob_Poverty;
1964 _SpinDuration = x + Knob_Bonus;
1965 }
1966 return 1;
1967 }
1968
1969 // The CAS failed ... we can take any of the following actions:
1970 // * penalize: ctr -= Knob_CASPenalty
1971 // * exit spin with prejudice -- goto Abort;
1972 // * exit spin without prejudice.
1973 // * Since CAS is high-latency, retry again immediately.
1974 prv = ox;
1975 if (caspty == -2) break;
1976 if (caspty == -1) goto Abort;
1977 ctr -= caspty;
1978 continue;
1979 }
1980
1981 // Did lock ownership change hands ?
1982 if (ox != prv && prv != NULL) {
1983 if (oxpty == -2) break;
1984 if (oxpty == -1) goto Abort;
1985 ctr -= oxpty;
1986 }
1987 prv = ox;
1988
1989 // Abort the spin if the owner is not executing.
1990 // The owner must be executing in order to drop the lock.
1991 // Spinning while the owner is OFFPROC is idiocy.
1992 // Consider: ctr -= RunnablePenalty ;
1993 if (Knob_OState && NotRunnable (Self, ox)) {
1994 goto Abort;
1995 }
1996 if (sss && _succ == NULL) _succ = Self;
1997 }
1998
1999 // Spin failed with prejudice -- reduce _SpinDuration.
2000 // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2001 // AIMD is globally stable.
2002 {
2003 int x = _SpinDuration;
2004 if (x > 0) {
2005 // Consider an AIMD scheme like: x -= (x >> 3) + 100
2006 // This is globally sample and tends to damp the response.
2007 x -= Knob_Penalty;
2008 if (x < 0) x = 0;
2009 _SpinDuration = x;
2010 }
2011 }
2012
2013 Abort:
2014 if (MaxSpin >= 0) Adjust(&_Spinner, -1);
2015 if (sss && _succ == Self) {
2016 _succ = NULL;
2017 // Invariant: after setting succ=null a contending thread
2018 // must recheck-retry _owner before parking. This usually happens
2019 // in the normal usage of TrySpin(), but it's safest
2020 // to make TrySpin() as foolproof as possible.
2021 OrderAccess::fence();
2022 if (TryLock(Self) > 0) return 1;
2023 }
2024 return 0;
2025 }
2026
2027 // NotRunnable() -- informed spinning
2028 //
2029 // Don't bother spinning if the owner is not eligible to drop the lock.
2030 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2031 // The thread must be runnable in order to drop the lock in timely fashion.
2032 // If the _owner is not runnable then spinning will not likely be
2033 // successful (profitable).
2034 //
2035 // Beware -- the thread referenced by _owner could have died
2036 // so a simply fetch from _owner->_thread_state might trap.
2037 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2038 // Because of the lifecycle issues, the _thread_state values
2039 // observed by NotRunnable() might be garbage. NotRunnable must
2040 // tolerate this and consider the observed _thread_state value
2041 // as advisory.
2042 //
2043 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2044 // a thread pointer.
2045 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2046 // with the LSB of _owner. Another option would be to probabilistically probe
2047 // the putative _owner->TypeTag value.
2048 //
2049 // Checking _thread_state isn't perfect. Even if the thread is
2050 // in_java it might be blocked on a page-fault or have been preempted
2051 // and sitting on a ready/dispatch queue.
2052 //
2053 // The return value from NotRunnable() is *advisory* -- the
2054 // result is based on sampling and is not necessarily coherent.
2055 // The caller must tolerate false-negative and false-positive errors.
2056 // Spinning, in general, is probabilistic anyway.
2057
2058
2059 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) {
2060 // Check ox->TypeTag == 2BAD.
2061 if (ox == NULL) return 0;
2062
2063 // Avoid transitive spinning ...
2064 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L.
2065 // Immediately after T1 acquires L it's possible that T2, also
2066 // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2067 // This occurs transiently after T1 acquired L but before
2068 // T1 managed to clear T1.Stalled. T2 does not need to abort
2069 // its spin in this circumstance.
2070 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2071
2072 if (BlockedOn == 1) return 1;
2073 if (BlockedOn != 0) {
2074 return BlockedOn != intptr_t(this) && _owner == ox;
2075 }
2076
2077 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant");
2078 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);;
2079 // consider also: jst != _thread_in_Java -- but that's overspecific.
2080 return jst == _thread_blocked || jst == _thread_in_native;
2081 }
2082
2083
2084 // -----------------------------------------------------------------------------
2085 // WaitSet management ...
2086
2087 ObjectWaiter::ObjectWaiter(Thread* thread) {
2088 _next = NULL;
2089 _prev = NULL;
2090 _notified = 0;
2091 _notifier_tid = 0;
2092 TState = TS_RUN;
2093 _thread = thread;
2094 _event = thread->_ParkEvent;
2095 _active = false;
2096 assert(_event != NULL, "invariant");
2097 }
2098
2099 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2100 JavaThread *jt = (JavaThread *)this->_thread;
2101 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2102 }
2103
2104 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2105 JavaThread *jt = (JavaThread *)this->_thread;
2106 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2107 }
2108
2109 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2110 assert(node != NULL, "should not add NULL node");
2111 assert(node->_prev == NULL, "node already in list");
2112 assert(node->_next == NULL, "node already in list");
2113 // put node at end of queue (circular doubly linked list)
2114 if (_WaitSet == NULL) {
2115 _WaitSet = node;
2116 node->_prev = node;
2117 node->_next = node;
2118 } else {
2119 ObjectWaiter* head = _WaitSet;
2120 ObjectWaiter* tail = head->_prev;
2121 assert(tail->_next == head, "invariant check");
2122 tail->_next = node;
2123 head->_prev = node;
2124 node->_next = head;
2125 node->_prev = tail;
2126 }
2127 }
2128
2129 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2130 // dequeue the very first waiter
2131 ObjectWaiter* waiter = _WaitSet;
2132 if (waiter) {
2133 DequeueSpecificWaiter(waiter);
2134 }
2135 return waiter;
2136 }
2137
2138 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2139 assert(node != NULL, "should not dequeue NULL node");
2140 assert(node->_prev != NULL, "node already removed from list");
2141 assert(node->_next != NULL, "node already removed from list");
2142 // when the waiter has woken up because of interrupt,
2143 // timeout or other spurious wake-up, dequeue the
2144 // waiter from waiting list
2145 ObjectWaiter* next = node->_next;
2146 if (next == node) {
2147 assert(node->_prev == node, "invariant check");
2148 _WaitSet = NULL;
2149 } else {
2150 ObjectWaiter* prev = node->_prev;
2151 assert(prev->_next == node, "invariant check");
2152 assert(next->_prev == node, "invariant check");
2153 next->_prev = prev;
2154 prev->_next = next;
2155 if (_WaitSet == node) {
2156 _WaitSet = next;
2157 }
2158 }
2159 node->_next = NULL;
2160 node->_prev = NULL;
2161 }
2162
2163 // -----------------------------------------------------------------------------
2164 // PerfData support
2165 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL;
2166 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL;
2167 PerfCounter * ObjectMonitor::_sync_Parks = NULL;
2168 PerfCounter * ObjectMonitor::_sync_Notifications = NULL;
2169 PerfCounter * ObjectMonitor::_sync_Inflations = NULL;
2170 PerfCounter * ObjectMonitor::_sync_Deflations = NULL;
2171 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL;
2172
2173 // One-shot global initialization for the sync subsystem.
2174 // We could also defer initialization and initialize on-demand
2175 // the first time we call inflate(). Initialization would
2176 // be protected - like so many things - by the MonitorCache_lock.
2177
2178 void ObjectMonitor::Initialize() {
2179 static int InitializationCompleted = 0;
2180 assert(InitializationCompleted == 0, "invariant");
2181 InitializationCompleted = 1;
2182 if (UsePerfData) {
2183 EXCEPTION_MARK;
2184 #define NEWPERFCOUNTER(n) \
2185 { \
2186 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \
2187 CHECK); \
2188 }
2189 #define NEWPERFVARIABLE(n) \
2190 { \
2191 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \
2192 CHECK); \
2193 }
2194 NEWPERFCOUNTER(_sync_Inflations);
2195 NEWPERFCOUNTER(_sync_Deflations);
2196 NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2197 NEWPERFCOUNTER(_sync_FutileWakeups);
2198 NEWPERFCOUNTER(_sync_Parks);
2199 NEWPERFCOUNTER(_sync_Notifications);
2200 NEWPERFVARIABLE(_sync_MonExtant);
2201 #undef NEWPERFCOUNTER
2202 #undef NEWPERFVARIABLE
2203 }
2204 }
2205
2206 void ObjectMonitor::DeferredInitialize() {
2207 if (InitDone > 0) return;
2208 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2209 while (InitDone != 1) /* empty */;
2210 return;
2211 }
2212
2213 // One-shot global initialization ...
2214 // The initialization is idempotent, so we don't need locks.
2215 // In the future consider doing this via os::init_2().
2216
2217 if (os::is_MP()) {
2218 BackOffMask = (1 << Knob_SpinBackOff) - 1;
2219 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1)
2220 } else {
2221 Knob_SpinLimit = 0;
2222 Knob_SpinBase = 0;
2223 Knob_PreSpin = 0;
2224 Knob_FixedSpin = -1;
2225 }
2226
2227 OrderAccess::fence();
2228 InitDone = 1;
2229 }
2230