1 /*
2 * Copyright (c) 1998, 2014, 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 "memory/resourceArea.hpp"
28 #include "oops/markOop.hpp"
29 #include "oops/oop.inline.hpp"
30 #include "runtime/atomic.inline.hpp"
31 #include "runtime/handles.inline.hpp"
32 #include "runtime/interfaceSupport.hpp"
33 #include "runtime/mutexLocker.hpp"
34 #include "runtime/objectMonitor.hpp"
35 #include "runtime/objectMonitor.inline.hpp"
36 #include "runtime/orderAccess.inline.hpp"
37 #include "runtime/osThread.hpp"
38 #include "runtime/stubRoutines.hpp"
39 #include "runtime/thread.inline.hpp"
40 #include "services/threadService.hpp"
41 #include "trace/tracing.hpp"
42 #include "trace/traceMacros.hpp"
43 #include "utilities/dtrace.hpp"
44 #include "utilities/macros.hpp"
45 #include "utilities/preserveException.hpp"
46
47 #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64)
48 // Need to inhibit inlining for older versions of GCC to avoid build-time failures
49 #define NOINLINE __attribute__((noinline))
50 #else
51 #define NOINLINE
52 #endif
53
54
55 #ifdef DTRACE_ENABLED
56
57 // Only bother with this argument setup if dtrace is available
58 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
59
60
61 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
62 char* bytes = NULL; \
63 int len = 0; \
64 jlong jtid = SharedRuntime::get_java_tid(thread); \
65 Symbol* klassname = ((oop)obj)->klass()->name(); \
66 if (klassname != NULL) { \
67 bytes = (char*)klassname->bytes(); \
68 len = klassname->utf8_length(); \
69 }
70
71 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
72 { \
73 if (DTraceMonitorProbes) { \
74 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
75 HOTSPOT_MONITOR_WAIT(jtid, \
76 (monitor), bytes, len, (millis)); \
77 } \
78 }
79
80 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
81 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
82 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
83 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
84 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
85
86 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
87 { \
88 if (DTraceMonitorProbes) { \
89 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
90 HOTSPOT_MONITOR_##probe(jtid, \
91 (uintptr_t)(monitor), bytes, len); \
92 } \
93 }
94
95 #else // ndef DTRACE_ENABLED
96
97 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
98 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
99
100 #endif // ndef DTRACE_ENABLED
101
102 // Tunables ...
103 // The knob* variables are effectively final. Once set they should
104 // never be modified hence. Consider using __read_mostly with GCC.
105
106 int ObjectMonitor::Knob_Verbose = 0;
107 int ObjectMonitor::Knob_VerifyInUse = 0;
108 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool -
109 static int Knob_LogSpins = 0; // enable jvmstat tally for spins
110 static int Knob_HandOff = 0;
111 static int Knob_ReportSettings = 0;
112
113 static int Knob_SpinBase = 0; // Floor AKA SpinMin
114 static int Knob_SpinBackOff = 0; // spin-loop backoff
115 static int Knob_CASPenalty = -1; // Penalty for failed CAS
116 static int Knob_OXPenalty = -1; // Penalty for observed _owner change
117 static int Knob_SpinSetSucc = 1; // spinners set the _succ field
118 static int Knob_SpinEarly = 1;
119 static int Knob_SuccEnabled = 1; // futile wake throttling
120 static int Knob_SuccRestrict = 0; // Limit successors + spinners to at-most-one
121 static int Knob_MaxSpinners = -1; // Should be a function of # CPUs
122 static int Knob_Bonus = 100; // spin success bonus
123 static int Knob_BonusB = 100; // spin success bonus
124 static int Knob_Penalty = 200; // spin failure penalty
125 static int Knob_Poverty = 1000;
126 static int Knob_SpinAfterFutile = 1; // Spin after returning from park()
127 static int Knob_FixedSpin = 0;
128 static int Knob_OState = 3; // Spinner checks thread state of _owner
129 static int Knob_UsePause = 1;
130 static int Knob_ExitPolicy = 0;
131 static int Knob_PreSpin = 10; // 20-100 likely better
132 static int Knob_ResetEvent = 0;
133 static int BackOffMask = 0;
134
135 static int Knob_FastHSSEC = 0;
136 static int Knob_MoveNotifyee = 2; // notify() - disposition of notifyee
137 static int Knob_QMode = 0; // EntryList-cxq policy - queue discipline
138 static volatile int InitDone = 0;
139
140 #define TrySpin TrySpin_VaryDuration
141
142 // -----------------------------------------------------------------------------
143 // Theory of operations -- Monitors lists, thread residency, etc:
144 //
145 // * A thread acquires ownership of a monitor by successfully
146 // CAS()ing the _owner field from null to non-null.
147 //
148 // * Invariant: A thread appears on at most one monitor list --
149 // cxq, EntryList or WaitSet -- at any one time.
150 //
151 // * Contending threads "push" themselves onto the cxq with CAS
152 // and then spin/park.
153 //
154 // * After a contending thread eventually acquires the lock it must
155 // dequeue itself from either the EntryList or the cxq.
156 //
157 // * The exiting thread identifies and unparks an "heir presumptive"
158 // tentative successor thread on the EntryList. Critically, the
159 // exiting thread doesn't unlink the successor thread from the EntryList.
160 // After having been unparked, the wakee will recontend for ownership of
161 // the monitor. The successor (wakee) will either acquire the lock or
162 // re-park itself.
163 //
164 // Succession is provided for by a policy of competitive handoff.
165 // The exiting thread does _not_ grant or pass ownership to the
166 // successor thread. (This is also referred to as "handoff" succession").
167 // Instead the exiting thread releases ownership and possibly wakes
168 // a successor, so the successor can (re)compete for ownership of the lock.
169 // If the EntryList is empty but the cxq is populated the exiting
170 // thread will drain the cxq into the EntryList. It does so by
171 // by detaching the cxq (installing null with CAS) and folding
172 // the threads from the cxq into the EntryList. The EntryList is
173 // doubly linked, while the cxq is singly linked because of the
174 // CAS-based "push" used to enqueue recently arrived threads (RATs).
175 //
176 // * Concurrency invariants:
177 //
178 // -- only the monitor owner may access or mutate the EntryList.
179 // The mutex property of the monitor itself protects the EntryList
180 // from concurrent interference.
181 // -- Only the monitor owner may detach the cxq.
182 //
183 // * The monitor entry list operations avoid locks, but strictly speaking
184 // they're not lock-free. Enter is lock-free, exit is not.
185 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html
186 //
187 // * The cxq can have multiple concurrent "pushers" but only one concurrent
188 // detaching thread. This mechanism is immune from the ABA corruption.
189 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
190 //
191 // * Taken together, the cxq and the EntryList constitute or form a
192 // single logical queue of threads stalled trying to acquire the lock.
193 // We use two distinct lists to improve the odds of a constant-time
194 // dequeue operation after acquisition (in the ::enter() epilogue) and
195 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm).
196 // A key desideratum is to minimize queue & monitor metadata manipulation
197 // that occurs while holding the monitor lock -- that is, we want to
198 // minimize monitor lock holds times. Note that even a small amount of
199 // fixed spinning will greatly reduce the # of enqueue-dequeue operations
200 // on EntryList|cxq. That is, spinning relieves contention on the "inner"
201 // locks and monitor metadata.
202 //
203 // Cxq points to the set of Recently Arrived Threads attempting entry.
204 // Because we push threads onto _cxq with CAS, the RATs must take the form of
205 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when
206 // the unlocking thread notices that EntryList is null but _cxq is != null.
207 //
208 // The EntryList is ordered by the prevailing queue discipline and
209 // can be organized in any convenient fashion, such as a doubly-linked list or
210 // a circular doubly-linked list. Critically, we want insert and delete operations
211 // to operate in constant-time. If we need a priority queue then something akin
212 // to Solaris' sleepq would work nicely. Viz.,
213 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
214 // Queue discipline is enforced at ::exit() time, when the unlocking thread
215 // drains the cxq into the EntryList, and orders or reorders the threads on the
216 // EntryList accordingly.
217 //
218 // Barring "lock barging", this mechanism provides fair cyclic ordering,
219 // somewhat similar to an elevator-scan.
220 //
221 // * The monitor synchronization subsystem avoids the use of native
222 // synchronization primitives except for the narrow platform-specific
223 // park-unpark abstraction. See the comments in os_solaris.cpp regarding
224 // the semantics of park-unpark. Put another way, this monitor implementation
225 // depends only on atomic operations and park-unpark. The monitor subsystem
226 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
227 // underlying OS manages the READY<->RUN transitions.
228 //
229 // * Waiting threads reside on the WaitSet list -- wait() puts
230 // the caller onto the WaitSet.
231 //
232 // * notify() or notifyAll() simply transfers threads from the WaitSet to
233 // either the EntryList or cxq. Subsequent exit() operations will
234 // unpark the notifyee. Unparking a notifee in notify() is inefficient -
235 // it's likely the notifyee would simply impale itself on the lock held
236 // by the notifier.
237 //
238 // * An interesting alternative is to encode cxq as (List,LockByte) where
239 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary
240 // variable, like _recursions, in the scheme. The threads or Events that form
241 // the list would have to be aligned in 256-byte addresses. A thread would
242 // try to acquire the lock or enqueue itself with CAS, but exiting threads
243 // could use a 1-0 protocol and simply STB to set the LockByte to 0.
244 // Note that is is *not* word-tearing, but it does presume that full-word
245 // CAS operations are coherent with intermix with STB operations. That's true
246 // on most common processors.
247 //
248 // * See also http://blogs.sun.com/dave
249
250
251 // -----------------------------------------------------------------------------
252 // Enter support
253
254 bool ObjectMonitor::try_enter(Thread* THREAD) {
255 if (THREAD != _owner) {
256 if (THREAD->is_lock_owned ((address)_owner)) {
257 assert(_recursions == 0, "internal state error");
258 _owner = THREAD;
259 _recursions = 1;
260 return true;
261 }
262 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
263 return false;
264 }
265 return true;
266 } else {
267 _recursions++;
268 return true;
269 }
270 }
271
272 void NOINLINE ObjectMonitor::enter(TRAPS) {
273 // The following code is ordered to check the most common cases first
274 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
275 Thread * const Self = THREAD;
276
277 void * cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL);
278 if (cur == NULL) {
279 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
280 assert(_recursions == 0, "invariant");
281 assert(_owner == Self, "invariant");
282 return;
283 }
284
285 if (cur == Self) {
286 // TODO-FIXME: check for integer overflow! BUGID 6557169.
287 _recursions++;
288 return;
289 }
290
291 if (Self->is_lock_owned ((address)cur)) {
292 assert(_recursions == 0, "internal state error");
293 _recursions = 1;
294 // Commute owner from a thread-specific on-stack BasicLockObject address to
295 // a full-fledged "Thread *".
296 _owner = Self;
297 return;
298 }
299
300 // We've encountered genuine contention.
301 assert(Self->_Stalled == 0, "invariant");
302 Self->_Stalled = intptr_t(this);
303
304 // Try one round of spinning *before* enqueueing Self
305 // and before going through the awkward and expensive state
306 // transitions. The following spin is strictly optional ...
307 // Note that if we acquire the monitor from an initial spin
308 // we forgo posting JVMTI events and firing DTRACE probes.
309 if (Knob_SpinEarly && TrySpin (Self) > 0) {
310 assert(_owner == Self, "invariant");
311 assert(_recursions == 0, "invariant");
312 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
313 Self->_Stalled = 0;
314 return;
315 }
316
317 assert(_owner != Self, "invariant");
318 assert(_succ != Self, "invariant");
319 assert(Self->is_Java_thread(), "invariant");
320 JavaThread * jt = (JavaThread *) Self;
321 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
322 assert(jt->thread_state() != _thread_blocked, "invariant");
323 assert(this->object() != NULL, "invariant");
324 assert(_count >= 0, "invariant");
325
326 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy().
327 // Ensure the object-monitor relationship remains stable while there's contention.
328 Atomic::inc(&_count);
329
330 EventJavaMonitorEnter event;
331
332 { // Change java thread status to indicate blocked on monitor enter.
333 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
334
335 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
336 if (JvmtiExport::should_post_monitor_contended_enter()) {
337 JvmtiExport::post_monitor_contended_enter(jt, this);
338
339 // The current thread does not yet own the monitor and does not
340 // yet appear on any queues that would get it made the successor.
341 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
342 // handler cannot accidentally consume an unpark() meant for the
343 // ParkEvent associated with this ObjectMonitor.
344 }
345
346 OSThreadContendState osts(Self->osthread());
347 ThreadBlockInVM tbivm(jt);
348
349 Self->set_current_pending_monitor(this);
350
351 // TODO-FIXME: change the following for(;;) loop to straight-line code.
352 for (;;) {
353 jt->set_suspend_equivalent();
354 // cleared by handle_special_suspend_equivalent_condition()
355 // or java_suspend_self()
356
357 EnterI(THREAD);
358
359 if (!ExitSuspendEquivalent(jt)) break;
360
361 // We have acquired the contended monitor, but while we were
362 // waiting another thread suspended us. We don't want to enter
363 // the monitor while suspended because that would surprise the
364 // thread that suspended us.
365 //
366 _recursions = 0;
367 _succ = NULL;
368 exit(false, Self);
369
370 jt->java_suspend_self();
371 }
372 Self->set_current_pending_monitor(NULL);
373
374 // We cleared the pending monitor info since we've just gotten past
375 // the enter-check-for-suspend dance and we now own the monitor free
376 // and clear, i.e., it is no longer pending. The ThreadBlockInVM
377 // destructor can go to a safepoint at the end of this block. If we
378 // do a thread dump during that safepoint, then this thread will show
379 // as having "-locked" the monitor, but the OS and java.lang.Thread
380 // states will still report that the thread is blocked trying to
381 // acquire it.
382 }
383
384 Atomic::dec(&_count);
385 assert(_count >= 0, "invariant");
386 Self->_Stalled = 0;
387
388 // Must either set _recursions = 0 or ASSERT _recursions == 0.
389 assert(_recursions == 0, "invariant");
390 assert(_owner == Self, "invariant");
391 assert(_succ != Self, "invariant");
392 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
393
394 // The thread -- now the owner -- is back in vm mode.
395 // Report the glorious news via TI,DTrace and jvmstat.
396 // The probe effect is non-trivial. All the reportage occurs
397 // while we hold the monitor, increasing the length of the critical
398 // section. Amdahl's parallel speedup law comes vividly into play.
399 //
400 // Another option might be to aggregate the events (thread local or
401 // per-monitor aggregation) and defer reporting until a more opportune
402 // time -- such as next time some thread encounters contention but has
403 // yet to acquire the lock. While spinning that thread could
404 // spinning we could increment JVMStat counters, etc.
405
406 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
407 if (JvmtiExport::should_post_monitor_contended_entered()) {
408 JvmtiExport::post_monitor_contended_entered(jt, this);
409
410 // The current thread already owns the monitor and is not going to
411 // call park() for the remainder of the monitor enter protocol. So
412 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
413 // event handler consumed an unpark() issued by the thread that
414 // just exited the monitor.
415 }
416
417 if (event.should_commit()) {
418 event.set_klass(((oop)this->object())->klass());
419 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid);
420 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr()));
421 event.commit();
422 }
423
424 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) {
425 ObjectMonitor::_sync_ContendedLockAttempts->inc();
426 }
427 }
428
429
430 // Caveat: TryLock() is not necessarily serializing if it returns failure.
431 // Callers must compensate as needed.
432
433 int ObjectMonitor::TryLock(Thread * Self) {
434 void * own = _owner;
435 if (own != NULL) return 0;
436 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) {
437 // Either guarantee _recursions == 0 or set _recursions = 0.
438 assert(_recursions == 0, "invariant");
439 assert(_owner == Self, "invariant");
440 return 1;
441 }
442 // The lock had been free momentarily, but we lost the race to the lock.
443 // Interference -- the CAS failed.
444 // We can either return -1 or retry.
445 // Retry doesn't make as much sense because the lock was just acquired.
446 return -1;
447 }
448
449 #define MAX_RECHECK_INTERVAL 1000
450
451 void NOINLINE ObjectMonitor::EnterI(TRAPS) {
452 Thread * const Self = THREAD;
453 assert(Self->is_Java_thread(), "invariant");
454 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant");
455
456 // Try the lock - TATAS
457 if (TryLock (Self) > 0) {
458 assert(_succ != Self, "invariant");
459 assert(_owner == Self, "invariant");
460 assert(_Responsible != Self, "invariant");
461 return;
462 }
463
464 DeferredInitialize();
465
466 // We try one round of spinning *before* enqueueing Self.
467 //
468 // If the _owner is ready but OFFPROC we could use a YieldTo()
469 // operation to donate the remainder of this thread's quantum
470 // to the owner. This has subtle but beneficial affinity
471 // effects.
472
473 if (TrySpin (Self) > 0) {
474 assert(_owner == Self, "invariant");
475 assert(_succ != Self, "invariant");
476 assert(_Responsible != Self, "invariant");
477 return;
478 }
479
480 // The Spin failed -- Enqueue and park the thread ...
481 assert(_succ != Self, "invariant");
482 assert(_owner != Self, "invariant");
483 assert(_Responsible != Self, "invariant");
484
485 // Enqueue "Self" on ObjectMonitor's _cxq.
486 //
487 // Node acts as a proxy for Self.
488 // As an aside, if were to ever rewrite the synchronization code mostly
489 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
490 // Java objects. This would avoid awkward lifecycle and liveness issues,
491 // as well as eliminate a subset of ABA issues.
492 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
493
494 ObjectWaiter node(Self);
495 Self->_ParkEvent->reset();
496 node._prev = (ObjectWaiter *) 0xBAD;
497 node.TState = ObjectWaiter::TS_CXQ;
498
499 // Push "Self" onto the front of the _cxq.
500 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
501 // Note that spinning tends to reduce the rate at which threads
502 // enqueue and dequeue on EntryList|cxq.
503 ObjectWaiter * nxt;
504 for (;;) {
505 node._next = nxt = _cxq;
506 if (Atomic::cmpxchg_ptr(&node, &_cxq, nxt) == nxt) break;
507
508 // Interference - the CAS failed because _cxq changed. Just retry.
509 // As an optional optimization we retry the lock.
510 if (TryLock (Self) > 0) {
511 assert(_succ != Self, "invariant");
512 assert(_owner == Self, "invariant");
513 assert(_Responsible != Self, "invariant");
514 return;
515 }
516 }
517
518 // Check for cxq|EntryList edge transition to non-null. This indicates
519 // the onset of contention. While contention persists exiting threads
520 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit
521 // operations revert to the faster 1-0 mode. This enter operation may interleave
522 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
523 // arrange for one of the contending thread to use a timed park() operations
524 // to detect and recover from the race. (Stranding is form of progress failure
525 // where the monitor is unlocked but all the contending threads remain parked).
526 // That is, at least one of the contended threads will periodically poll _owner.
527 // One of the contending threads will become the designated "Responsible" thread.
528 // The Responsible thread uses a timed park instead of a normal indefinite park
529 // operation -- it periodically wakes and checks for and recovers from potential
530 // strandings admitted by 1-0 exit operations. We need at most one Responsible
531 // thread per-monitor at any given moment. Only threads on cxq|EntryList may
532 // be responsible for a monitor.
533 //
534 // Currently, one of the contended threads takes on the added role of "Responsible".
535 // A viable alternative would be to use a dedicated "stranding checker" thread
536 // that periodically iterated over all the threads (or active monitors) and unparked
537 // successors where there was risk of stranding. This would help eliminate the
538 // timer scalability issues we see on some platforms as we'd only have one thread
539 // -- the checker -- parked on a timer.
540
541 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) {
542 // Try to assume the role of responsible thread for the monitor.
543 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self }
544 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL);
545 }
546
547 // The lock might have been released while this thread was occupied queueing
548 // itself onto _cxq. To close the race and avoid "stranding" and
549 // progress-liveness failure we must resample-retry _owner before parking.
550 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
551 // In this case the ST-MEMBAR is accomplished with CAS().
552 //
553 // TODO: Defer all thread state transitions until park-time.
554 // Since state transitions are heavy and inefficient we'd like
555 // to defer the state transitions until absolutely necessary,
556 // and in doing so avoid some transitions ...
557
558 TEVENT(Inflated enter - Contention);
559 int nWakeups = 0;
560 int recheckInterval = 1;
561
562 for (;;) {
563
564 if (TryLock(Self) > 0) break;
565 assert(_owner != Self, "invariant");
566
567 if ((SyncFlags & 2) && _Responsible == NULL) {
568 Atomic::cmpxchg_ptr(Self, &_Responsible, NULL);
569 }
570
571 // park self
572 if (_Responsible == Self || (SyncFlags & 1)) {
573 TEVENT(Inflated enter - park TIMED);
574 Self->_ParkEvent->park((jlong) recheckInterval);
575 // Increase the recheckInterval, but clamp the value.
576 recheckInterval *= 8;
577 if (recheckInterval > MAX_RECHECK_INTERVAL) {
578 recheckInterval = MAX_RECHECK_INTERVAL;
579 }
580 } else {
581 TEVENT(Inflated enter - park UNTIMED);
582 Self->_ParkEvent->park();
583 }
584
585 if (TryLock(Self) > 0) break;
586
587 // The lock is still contested.
588 // Keep a tally of the # of futile wakeups.
589 // Note that the counter is not protected by a lock or updated by atomics.
590 // That is by design - we trade "lossy" counters which are exposed to
591 // races during updates for a lower probe effect.
592 TEVENT(Inflated enter - Futile wakeup);
593 if (ObjectMonitor::_sync_FutileWakeups != NULL) {
594 ObjectMonitor::_sync_FutileWakeups->inc();
595 }
596 ++nWakeups;
597
598 // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
599 // We can defer clearing _succ until after the spin completes
600 // TrySpin() must tolerate being called with _succ == Self.
601 // Try yet another round of adaptive spinning.
602 if ((Knob_SpinAfterFutile & 1) && TrySpin(Self) > 0) break;
603
604 // We can find that we were unpark()ed and redesignated _succ while
605 // we were spinning. That's harmless. If we iterate and call park(),
606 // park() will consume the event and return immediately and we'll
607 // just spin again. This pattern can repeat, leaving _succ to simply
608 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks().
609 // Alternately, we can sample fired() here, and if set, forgo spinning
610 // in the next iteration.
611
612 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
613 Self->_ParkEvent->reset();
614 OrderAccess::fence();
615 }
616 if (_succ == Self) _succ = NULL;
617
618 // Invariant: after clearing _succ a thread *must* retry _owner before parking.
619 OrderAccess::fence();
620 }
621
622 // Egress :
623 // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
624 // Normally we'll find Self on the EntryList .
625 // From the perspective of the lock owner (this thread), the
626 // EntryList is stable and cxq is prepend-only.
627 // The head of cxq is volatile but the interior is stable.
628 // In addition, Self.TState is stable.
629
630 assert(_owner == Self, "invariant");
631 assert(object() != NULL, "invariant");
632 // I'd like to write:
633 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
634 // but as we're at a safepoint that's not safe.
635
636 UnlinkAfterAcquire(Self, &node);
637 if (_succ == Self) _succ = NULL;
638
639 assert(_succ != Self, "invariant");
640 if (_Responsible == Self) {
641 _Responsible = NULL;
642 OrderAccess::fence(); // Dekker pivot-point
643
644 // We may leave threads on cxq|EntryList without a designated
645 // "Responsible" thread. This is benign. When this thread subsequently
646 // exits the monitor it can "see" such preexisting "old" threads --
647 // threads that arrived on the cxq|EntryList before the fence, above --
648 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads
649 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
650 // non-null and elect a new "Responsible" timer thread.
651 //
652 // This thread executes:
653 // ST Responsible=null; MEMBAR (in enter epilogue - here)
654 // LD cxq|EntryList (in subsequent exit)
655 //
656 // Entering threads in the slow/contended path execute:
657 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
658 // The (ST cxq; MEMBAR) is accomplished with CAS().
659 //
660 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
661 // exit operation from floating above the ST Responsible=null.
662 }
663
664 // We've acquired ownership with CAS().
665 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
666 // But since the CAS() this thread may have also stored into _succ,
667 // EntryList, cxq or Responsible. These meta-data updates must be
668 // visible __before this thread subsequently drops the lock.
669 // Consider what could occur if we didn't enforce this constraint --
670 // STs to monitor meta-data and user-data could reorder with (become
671 // visible after) the ST in exit that drops ownership of the lock.
672 // Some other thread could then acquire the lock, but observe inconsistent
673 // or old monitor meta-data and heap data. That violates the JMM.
674 // To that end, the 1-0 exit() operation must have at least STST|LDST
675 // "release" barrier semantics. Specifically, there must be at least a
676 // STST|LDST barrier in exit() before the ST of null into _owner that drops
677 // the lock. The barrier ensures that changes to monitor meta-data and data
678 // protected by the lock will be visible before we release the lock, and
679 // therefore before some other thread (CPU) has a chance to acquire the lock.
680 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
681 //
682 // Critically, any prior STs to _succ or EntryList must be visible before
683 // the ST of null into _owner in the *subsequent* (following) corresponding
684 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily
685 // execute a serializing instruction.
686
687 if (SyncFlags & 8) {
688 OrderAccess::fence();
689 }
690 return;
691 }
692
693 // ReenterI() is a specialized inline form of the latter half of the
694 // contended slow-path from EnterI(). We use ReenterI() only for
695 // monitor reentry in wait().
696 //
697 // In the future we should reconcile EnterI() and ReenterI(), adding
698 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the
699 // loop accordingly.
700
701 void NOINLINE ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) {
702 assert(Self != NULL, "invariant");
703 assert(SelfNode != NULL, "invariant");
704 assert(SelfNode->_thread == Self, "invariant");
705 assert(_waiters > 0, "invariant");
706 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
707 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
708 JavaThread * jt = (JavaThread *) Self;
709
710 int nWakeups = 0;
711 for (;;) {
712 ObjectWaiter::TStates v = SelfNode->TState;
713 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
714 assert(_owner != Self, "invariant");
715
716 if (TryLock(Self) > 0) break;
717 if (TrySpin(Self) > 0) break;
718
719 TEVENT(Wait Reentry - parking);
720
721 // State transition wrappers around park() ...
722 // ReenterI() wisely defers state transitions until
723 // it's clear we must park the thread.
724 {
725 OSThreadContendState osts(Self->osthread());
726 ThreadBlockInVM tbivm(jt);
727
728 // cleared by handle_special_suspend_equivalent_condition()
729 // or java_suspend_self()
730 jt->set_suspend_equivalent();
731 if (SyncFlags & 1) {
732 Self->_ParkEvent->park((jlong)MAX_RECHECK_INTERVAL);
733 } else {
734 Self->_ParkEvent->park();
735 }
736
737 // were we externally suspended while we were waiting?
738 for (;;) {
739 if (!ExitSuspendEquivalent(jt)) break;
740 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
741 jt->java_suspend_self();
742 jt->set_suspend_equivalent();
743 }
744 }
745
746 // Try again, but just so we distinguish between futile wakeups and
747 // successful wakeups. The following test isn't algorithmically
748 // necessary, but it helps us maintain sensible statistics.
749 if (TryLock(Self) > 0) break;
750
751 // The lock is still contested.
752 // Keep a tally of the # of futile wakeups.
753 // Note that the counter is not protected by a lock or updated by atomics.
754 // That is by design - we trade "lossy" counters which are exposed to
755 // races during updates for a lower probe effect.
756 TEVENT(Wait Reentry - futile wakeup);
757 ++nWakeups;
758
759 // Assuming this is not a spurious wakeup we'll normally
760 // find that _succ == Self.
761 if (_succ == Self) _succ = NULL;
762
763 // Invariant: after clearing _succ a contending thread
764 // *must* retry _owner before parking.
765 OrderAccess::fence();
766
767 if (ObjectMonitor::_sync_FutileWakeups != NULL) {
768 ObjectMonitor::_sync_FutileWakeups->inc();
769 }
770 }
771
772 // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
773 // Normally we'll find Self on the EntryList.
774 // Unlinking from the EntryList is constant-time and atomic-free.
775 // From the perspective of the lock owner (this thread), the
776 // EntryList is stable and cxq is prepend-only.
777 // The head of cxq is volatile but the interior is stable.
778 // In addition, Self.TState is stable.
779
780 assert(_owner == Self, "invariant");
781 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
782 UnlinkAfterAcquire(Self, SelfNode);
783 if (_succ == Self) _succ = NULL;
784 assert(_succ != Self, "invariant");
785 SelfNode->TState = ObjectWaiter::TS_RUN;
786 OrderAccess::fence(); // see comments at the end of EnterI()
787 }
788
789 // By convention we unlink a contending thread from EntryList|cxq immediately
790 // after the thread acquires the lock in ::enter(). Equally, we could defer
791 // unlinking the thread until ::exit()-time.
792
793 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) {
794 assert(_owner == Self, "invariant");
795 assert(SelfNode->_thread == Self, "invariant");
796
797 if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
798 // Normal case: remove Self from the DLL EntryList .
799 // This is a constant-time operation.
800 ObjectWaiter * nxt = SelfNode->_next;
801 ObjectWaiter * prv = SelfNode->_prev;
802 if (nxt != NULL) nxt->_prev = prv;
803 if (prv != NULL) prv->_next = nxt;
804 if (SelfNode == _EntryList) _EntryList = nxt;
805 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
806 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
807 TEVENT(Unlink from EntryList);
808 } else {
809 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant");
810 // Inopportune interleaving -- Self is still on the cxq.
811 // This usually means the enqueue of self raced an exiting thread.
812 // Normally we'll find Self near the front of the cxq, so
813 // dequeueing is typically fast. If needbe we can accelerate
814 // this with some MCS/CHL-like bidirectional list hints and advisory
815 // back-links so dequeueing from the interior will normally operate
816 // in constant-time.
817 // Dequeue Self from either the head (with CAS) or from the interior
818 // with a linear-time scan and normal non-atomic memory operations.
819 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
820 // and then unlink Self from EntryList. We have to drain eventually,
821 // so it might as well be now.
822
823 ObjectWaiter * v = _cxq;
824 assert(v != NULL, "invariant");
825 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) {
826 // The CAS above can fail from interference IFF a "RAT" arrived.
827 // In that case Self must be in the interior and can no longer be
828 // at the head of cxq.
829 if (v == SelfNode) {
830 assert(_cxq != v, "invariant");
831 v = _cxq; // CAS above failed - start scan at head of list
832 }
833 ObjectWaiter * p;
834 ObjectWaiter * q = NULL;
835 for (p = v; p != NULL && p != SelfNode; p = p->_next) {
836 q = p;
837 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
838 }
839 assert(v != SelfNode, "invariant");
840 assert(p == SelfNode, "Node not found on cxq");
841 assert(p != _cxq, "invariant");
842 assert(q != NULL, "invariant");
843 assert(q->_next == p, "invariant");
844 q->_next = p->_next;
845 }
846 TEVENT(Unlink from cxq);
847 }
848
849 #ifdef ASSERT
850 // Diagnostic hygiene ...
851 SelfNode->_prev = (ObjectWaiter *) 0xBAD;
852 SelfNode->_next = (ObjectWaiter *) 0xBAD;
853 SelfNode->TState = ObjectWaiter::TS_RUN;
854 #endif
855 }
856
857 // -----------------------------------------------------------------------------
858 // Exit support
859 //
860 // exit()
861 // ~~~~~~
862 // Note that the collector can't reclaim the objectMonitor or deflate
863 // the object out from underneath the thread calling ::exit() as the
864 // thread calling ::exit() never transitions to a stable state.
865 // This inhibits GC, which in turn inhibits asynchronous (and
866 // inopportune) reclamation of "this".
867 //
868 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
869 // There's one exception to the claim above, however. EnterI() can call
870 // exit() to drop a lock if the acquirer has been externally suspended.
871 // In that case exit() is called with _thread_state as _thread_blocked,
872 // but the monitor's _count field is > 0, which inhibits reclamation.
873 //
874 // 1-0 exit
875 // ~~~~~~~~
876 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
877 // the fast-path operators have been optimized so the common ::exit()
878 // operation is 1-0. See i486.ad fast_unlock(), for instance.
879 // The code emitted by fast_unlock() elides the usual MEMBAR. This
880 // greatly improves latency -- MEMBAR and CAS having considerable local
881 // latency on modern processors -- but at the cost of "stranding". Absent the
882 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
883 // ::enter() path, resulting in the entering thread being stranding
884 // and a progress-liveness failure. Stranding is extremely rare.
885 // We use timers (timed park operations) & periodic polling to detect
886 // and recover from stranding. Potentially stranded threads periodically
887 // wake up and poll the lock. See the usage of the _Responsible variable.
888 //
889 // The CAS() in enter provides for safety and exclusion, while the CAS or
890 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking
891 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding.
892 // We detect and recover from stranding with timers.
893 //
894 // If a thread transiently strands it'll park until (a) another
895 // thread acquires the lock and then drops the lock, at which time the
896 // exiting thread will notice and unpark the stranded thread, or, (b)
897 // the timer expires. If the lock is high traffic then the stranding latency
898 // will be low due to (a). If the lock is low traffic then the odds of
899 // stranding are lower, although the worst-case stranding latency
900 // is longer. Critically, we don't want to put excessive load in the
901 // platform's timer subsystem. We want to minimize both the timer injection
902 // rate (timers created/sec) as well as the number of timers active at
903 // any one time. (more precisely, we want to minimize timer-seconds, which is
904 // the integral of the # of active timers at any instant over time).
905 // Both impinge on OS scalability. Given that, at most one thread parked on
906 // a monitor will use a timer.
907 //
908 // There is also the risk of a futile wake-up. If we drop the lock
909 // another thread can reacquire the lock immediately, and we can
910 // then wake a thread unnecessarily. This is benign, and we've
911 // structured the code so the windows are short and the frequency
912 // of such futile wakups is low.
913
914 void NOINLINE ObjectMonitor::exit(bool not_suspended, TRAPS) {
915 Thread * const Self = THREAD;
916 if (THREAD != _owner) {
917 if (THREAD->is_lock_owned((address) _owner)) {
918 // Transmute _owner from a BasicLock pointer to a Thread address.
919 // We don't need to hold _mutex for this transition.
920 // Non-null to Non-null is safe as long as all readers can
921 // tolerate either flavor.
922 assert(_recursions == 0, "invariant");
923 _owner = THREAD;
924 _recursions = 0;
925 } else {
926 // Apparent unbalanced locking ...
927 // Naively we'd like to throw IllegalMonitorStateException.
928 // As a practical matter we can neither allocate nor throw an
929 // exception as ::exit() can be called from leaf routines.
930 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
931 // Upon deeper reflection, however, in a properly run JVM the only
932 // way we should encounter this situation is in the presence of
933 // unbalanced JNI locking. TODO: CheckJNICalls.
934 // See also: CR4414101
935 TEVENT(Exit - Throw IMSX);
936 assert(false, "Non-balanced monitor enter/exit! Likely JNI locking");
937 return;
938 }
939 }
940
941 if (_recursions != 0) {
942 _recursions--; // this is simple recursive enter
943 TEVENT(Inflated exit - recursive);
944 return;
945 }
946
947 // Invariant: after setting Responsible=null an thread must execute
948 // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
949 if ((SyncFlags & 4) == 0) {
950 _Responsible = NULL;
951 }
952
953 #if INCLUDE_TRACE
954 // get the owner's thread id for the MonitorEnter event
955 // if it is enabled and the thread isn't suspended
956 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) {
957 _previous_owner_tid = SharedRuntime::get_java_tid(Self);
958 }
959 #endif
960
961 for (;;) {
962 assert(THREAD == _owner, "invariant");
963
964
965 if (Knob_ExitPolicy == 0) {
966 // release semantics: prior loads and stores from within the critical section
967 // must not float (reorder) past the following store that drops the lock.
968 // On SPARC that requires MEMBAR #loadstore|#storestore.
969 // But of course in TSO #loadstore|#storestore is not required.
970 // I'd like to write one of the following:
971 // A. OrderAccess::release() ; _owner = NULL
972 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
973 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
974 // store into a _dummy variable. That store is not needed, but can result
975 // in massive wasteful coherency traffic on classic SMP systems.
976 // Instead, I use release_store(), which is implemented as just a simple
977 // ST on x64, x86 and SPARC.
978 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock
979 OrderAccess::storeload(); // See if we need to wake a successor
980 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
981 TEVENT(Inflated exit - simple egress);
982 return;
983 }
984 TEVENT(Inflated exit - complex egress);
985 // Other threads are blocked trying to acquire the lock.
986
987 // Normally the exiting thread is responsible for ensuring succession,
988 // but if other successors are ready or other entering threads are spinning
989 // then this thread can simply store NULL into _owner and exit without
990 // waking a successor. The existence of spinners or ready successors
991 // guarantees proper succession (liveness). Responsibility passes to the
992 // ready or running successors. The exiting thread delegates the duty.
993 // More precisely, if a successor already exists this thread is absolved
994 // of the responsibility of waking (unparking) one.
995 //
996 // The _succ variable is critical to reducing futile wakeup frequency.
997 // _succ identifies the "heir presumptive" thread that has been made
998 // ready (unparked) but that has not yet run. We need only one such
999 // successor thread to guarantee progress.
1000 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1001 // section 3.3 "Futile Wakeup Throttling" for details.
1002 //
1003 // Note that spinners in Enter() also set _succ non-null.
1004 // In the current implementation spinners opportunistically set
1005 // _succ so that exiting threads might avoid waking a successor.
1006 // Another less appealing alternative would be for the exiting thread
1007 // to drop the lock and then spin briefly to see if a spinner managed
1008 // to acquire the lock. If so, the exiting thread could exit
1009 // immediately without waking a successor, otherwise the exiting
1010 // thread would need to dequeue and wake a successor.
1011 // (Note that we'd need to make the post-drop spin short, but no
1012 // shorter than the worst-case round-trip cache-line migration time.
1013 // The dropped lock needs to become visible to the spinner, and then
1014 // the acquisition of the lock by the spinner must become visible to
1015 // the exiting thread).
1016
1017 // It appears that an heir-presumptive (successor) must be made ready.
1018 // Only the current lock owner can manipulate the EntryList or
1019 // drain _cxq, so we need to reacquire the lock. If we fail
1020 // to reacquire the lock the responsibility for ensuring succession
1021 // falls to the new owner.
1022 //
1023 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
1024 return;
1025 }
1026 TEVENT(Exit - Reacquired);
1027 } else {
1028 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
1029 OrderAccess::release_store_ptr(&_owner, NULL); // drop the lock
1030 OrderAccess::storeload();
1031 // Ratify the previously observed values.
1032 if (_cxq == NULL || _succ != NULL) {
1033 TEVENT(Inflated exit - simple egress);
1034 return;
1035 }
1036
1037 // inopportune interleaving -- the exiting thread (this thread)
1038 // in the fast-exit path raced an entering thread in the slow-enter
1039 // path.
1040 // We have two choices:
1041 // A. Try to reacquire the lock.
1042 // If the CAS() fails return immediately, otherwise
1043 // we either restart/rerun the exit operation, or simply
1044 // fall-through into the code below which wakes a successor.
1045 // B. If the elements forming the EntryList|cxq are TSM
1046 // we could simply unpark() the lead thread and return
1047 // without having set _succ.
1048 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
1049 TEVENT(Inflated exit - reacquired succeeded);
1050 return;
1051 }
1052 TEVENT(Inflated exit - reacquired failed);
1053 } else {
1054 TEVENT(Inflated exit - complex egress);
1055 }
1056 }
1057
1058 guarantee(_owner == THREAD, "invariant");
1059
1060 ObjectWaiter * w = NULL;
1061 int QMode = Knob_QMode;
1062
1063 if (QMode == 2 && _cxq != NULL) {
1064 // QMode == 2 : cxq has precedence over EntryList.
1065 // Try to directly wake a successor from the cxq.
1066 // If successful, the successor will need to unlink itself from cxq.
1067 w = _cxq;
1068 assert(w != NULL, "invariant");
1069 assert(w->TState == ObjectWaiter::TS_CXQ, "Invariant");
1070 ExitEpilog(Self, w);
1071 return;
1072 }
1073
1074 if (QMode == 3 && _cxq != NULL) {
1075 // Aggressively drain cxq into EntryList at the first opportunity.
1076 // This policy ensure that recently-run threads live at the head of EntryList.
1077 // Drain _cxq into EntryList - bulk transfer.
1078 // First, detach _cxq.
1079 // The following loop is tantamount to: w = swap(&cxq, NULL)
1080 w = _cxq;
1081 for (;;) {
1082 assert(w != NULL, "Invariant");
1083 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w);
1084 if (u == w) break;
1085 w = u;
1086 }
1087 assert(w != NULL, "invariant");
1088
1089 ObjectWaiter * q = NULL;
1090 ObjectWaiter * p;
1091 for (p = w; p != NULL; p = p->_next) {
1092 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1093 p->TState = ObjectWaiter::TS_ENTER;
1094 p->_prev = q;
1095 q = p;
1096 }
1097
1098 // Append the RATs to the EntryList
1099 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1100 ObjectWaiter * Tail;
1101 for (Tail = _EntryList; Tail != NULL && Tail->_next != NULL;
1102 Tail = Tail->_next)
1103 /* empty */;
1104 if (Tail == NULL) {
1105 _EntryList = w;
1106 } else {
1107 Tail->_next = w;
1108 w->_prev = Tail;
1109 }
1110
1111 // Fall thru into code that tries to wake a successor from EntryList
1112 }
1113
1114 if (QMode == 4 && _cxq != NULL) {
1115 // Aggressively drain cxq into EntryList at the first opportunity.
1116 // This policy ensure that recently-run threads live at the head of EntryList.
1117
1118 // Drain _cxq into EntryList - bulk transfer.
1119 // First, detach _cxq.
1120 // The following loop is tantamount to: w = swap(&cxq, NULL)
1121 w = _cxq;
1122 for (;;) {
1123 assert(w != NULL, "Invariant");
1124 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w);
1125 if (u == w) break;
1126 w = u;
1127 }
1128 assert(w != NULL, "invariant");
1129
1130 ObjectWaiter * q = NULL;
1131 ObjectWaiter * p;
1132 for (p = w; p != NULL; p = p->_next) {
1133 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1134 p->TState = ObjectWaiter::TS_ENTER;
1135 p->_prev = q;
1136 q = p;
1137 }
1138
1139 // Prepend the RATs to the EntryList
1140 if (_EntryList != NULL) {
1141 q->_next = _EntryList;
1142 _EntryList->_prev = q;
1143 }
1144 _EntryList = w;
1145
1146 // Fall thru into code that tries to wake a successor from EntryList
1147 }
1148
1149 w = _EntryList;
1150 if (w != NULL) {
1151 // I'd like to write: guarantee (w->_thread != Self).
1152 // But in practice an exiting thread may find itself on the EntryList.
1153 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and
1154 // then calls exit(). Exit release the lock by setting O._owner to NULL.
1155 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The
1156 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1157 // release the lock "O". T2 resumes immediately after the ST of null into
1158 // _owner, above. T2 notices that the EntryList is populated, so it
1159 // reacquires the lock and then finds itself on the EntryList.
1160 // Given all that, we have to tolerate the circumstance where "w" is
1161 // associated with Self.
1162 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1163 ExitEpilog(Self, w);
1164 return;
1165 }
1166
1167 // If we find that both _cxq and EntryList are null then just
1168 // re-run the exit protocol from the top.
1169 w = _cxq;
1170 if (w == NULL) continue;
1171
1172 // Drain _cxq into EntryList - bulk transfer.
1173 // First, detach _cxq.
1174 // The following loop is tantamount to: w = swap(&cxq, NULL)
1175 for (;;) {
1176 assert(w != NULL, "Invariant");
1177 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr(NULL, &_cxq, w);
1178 if (u == w) break;
1179 w = u;
1180 }
1181 TEVENT(Inflated exit - drain cxq into EntryList);
1182
1183 assert(w != NULL, "invariant");
1184 assert(_EntryList == NULL, "invariant");
1185
1186 // Convert the LIFO SLL anchored by _cxq into a DLL.
1187 // The list reorganization step operates in O(LENGTH(w)) time.
1188 // It's critical that this step operate quickly as
1189 // "Self" still holds the outer-lock, restricting parallelism
1190 // and effectively lengthening the critical section.
1191 // Invariant: s chases t chases u.
1192 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1193 // we have faster access to the tail.
1194
1195 if (QMode == 1) {
1196 // QMode == 1 : drain cxq to EntryList, reversing order
1197 // We also reverse the order of the list.
1198 ObjectWaiter * s = NULL;
1199 ObjectWaiter * t = w;
1200 ObjectWaiter * u = NULL;
1201 while (t != NULL) {
1202 guarantee(t->TState == ObjectWaiter::TS_CXQ, "invariant");
1203 t->TState = ObjectWaiter::TS_ENTER;
1204 u = t->_next;
1205 t->_prev = u;
1206 t->_next = s;
1207 s = t;
1208 t = u;
1209 }
1210 _EntryList = s;
1211 assert(s != NULL, "invariant");
1212 } else {
1213 // QMode == 0 or QMode == 2
1214 _EntryList = w;
1215 ObjectWaiter * q = NULL;
1216 ObjectWaiter * p;
1217 for (p = w; p != NULL; p = p->_next) {
1218 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1219 p->TState = ObjectWaiter::TS_ENTER;
1220 p->_prev = q;
1221 q = p;
1222 }
1223 }
1224
1225 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1226 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1227
1228 // See if we can abdicate to a spinner instead of waking a thread.
1229 // A primary goal of the implementation is to reduce the
1230 // context-switch rate.
1231 if (_succ != NULL) continue;
1232
1233 w = _EntryList;
1234 if (w != NULL) {
1235 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1236 ExitEpilog(Self, w);
1237 return;
1238 }
1239 }
1240 }
1241
1242 // ExitSuspendEquivalent:
1243 // A faster alternate to handle_special_suspend_equivalent_condition()
1244 //
1245 // handle_special_suspend_equivalent_condition() unconditionally
1246 // acquires the SR_lock. On some platforms uncontended MutexLocker()
1247 // operations have high latency. Note that in ::enter() we call HSSEC
1248 // while holding the monitor, so we effectively lengthen the critical sections.
1249 //
1250 // There are a number of possible solutions:
1251 //
1252 // A. To ameliorate the problem we might also defer state transitions
1253 // to as late as possible -- just prior to parking.
1254 // Given that, we'd call HSSEC after having returned from park(),
1255 // but before attempting to acquire the monitor. This is only a
1256 // partial solution. It avoids calling HSSEC while holding the
1257 // monitor (good), but it still increases successor reacquisition latency --
1258 // the interval between unparking a successor and the time the successor
1259 // resumes and retries the lock. See ReenterI(), which defers state transitions.
1260 // If we use this technique we can also avoid EnterI()-exit() loop
1261 // in ::enter() where we iteratively drop the lock and then attempt
1262 // to reacquire it after suspending.
1263 //
1264 // B. In the future we might fold all the suspend bits into a
1265 // composite per-thread suspend flag and then update it with CAS().
1266 // Alternately, a Dekker-like mechanism with multiple variables
1267 // would suffice:
1268 // ST Self->_suspend_equivalent = false
1269 // MEMBAR
1270 // LD Self_>_suspend_flags
1271 //
1272 // UPDATE 2007-10-6: since I've replaced the native Mutex/Monitor subsystem
1273 // with a more efficient implementation, the need to use "FastHSSEC" has
1274 // decreased. - Dave
1275
1276
1277 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) {
1278 const int Mode = Knob_FastHSSEC;
1279 if (Mode && !jSelf->is_external_suspend()) {
1280 assert(jSelf->is_suspend_equivalent(), "invariant");
1281 jSelf->clear_suspend_equivalent();
1282 if (2 == Mode) OrderAccess::storeload();
1283 if (!jSelf->is_external_suspend()) return false;
1284 // We raced a suspension -- fall thru into the slow path
1285 TEVENT(ExitSuspendEquivalent - raced);
1286 jSelf->set_suspend_equivalent();
1287 }
1288 return jSelf->handle_special_suspend_equivalent_condition();
1289 }
1290
1291
1292 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) {
1293 assert(_owner == Self, "invariant");
1294
1295 // Exit protocol:
1296 // 1. ST _succ = wakee
1297 // 2. membar #loadstore|#storestore;
1298 // 2. ST _owner = NULL
1299 // 3. unpark(wakee)
1300
1301 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL;
1302 ParkEvent * Trigger = Wakee->_event;
1303
1304 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1305 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1306 // out-of-scope (non-extant).
1307 Wakee = NULL;
1308
1309 // Drop the lock
1310 OrderAccess::release_store_ptr(&_owner, NULL);
1311 OrderAccess::fence(); // ST _owner vs LD in unpark()
1312
1313 if (SafepointSynchronize::do_call_back()) {
1314 TEVENT(unpark before SAFEPOINT);
1315 }
1316
1317 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1318 Trigger->unpark();
1319
1320 // Maintain stats and report events to JVMTI
1321 if (ObjectMonitor::_sync_Parks != NULL) {
1322 ObjectMonitor::_sync_Parks->inc();
1323 }
1324 }
1325
1326
1327 // -----------------------------------------------------------------------------
1328 // Class Loader deadlock handling.
1329 //
1330 // complete_exit exits a lock returning recursion count
1331 // complete_exit/reenter operate as a wait without waiting
1332 // complete_exit requires an inflated monitor
1333 // The _owner field is not always the Thread addr even with an
1334 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1335 // thread due to contention.
1336 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1337 Thread * const Self = THREAD;
1338 assert(Self->is_Java_thread(), "Must be Java thread!");
1339 JavaThread *jt = (JavaThread *)THREAD;
1340
1341 DeferredInitialize();
1342
1343 if (THREAD != _owner) {
1344 if (THREAD->is_lock_owned ((address)_owner)) {
1345 assert(_recursions == 0, "internal state error");
1346 _owner = THREAD; // Convert from basiclock addr to Thread addr
1347 _recursions = 0;
1348 }
1349 }
1350
1351 guarantee(Self == _owner, "complete_exit not owner");
1352 intptr_t save = _recursions; // record the old recursion count
1353 _recursions = 0; // set the recursion level to be 0
1354 exit(true, Self); // exit the monitor
1355 guarantee(_owner != Self, "invariant");
1356 return save;
1357 }
1358
1359 // reenter() enters a lock and sets recursion count
1360 // complete_exit/reenter operate as a wait without waiting
1361 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1362 Thread * const Self = THREAD;
1363 assert(Self->is_Java_thread(), "Must be Java thread!");
1364 JavaThread *jt = (JavaThread *)THREAD;
1365
1366 guarantee(_owner != Self, "reenter already owner");
1367 enter(THREAD); // enter the monitor
1368 guarantee(_recursions == 0, "reenter recursion");
1369 _recursions = recursions;
1370 return;
1371 }
1372
1373
1374 // -----------------------------------------------------------------------------
1375 // A macro is used below because there may already be a pending
1376 // exception which should not abort the execution of the routines
1377 // which use this (which is why we don't put this into check_slow and
1378 // call it with a CHECK argument).
1379
1380 #define CHECK_OWNER() \
1381 do { \
1382 if (THREAD != _owner) { \
1383 if (THREAD->is_lock_owned((address) _owner)) { \
1384 _owner = THREAD; /* Convert from basiclock addr to Thread addr */ \
1385 _recursions = 0; \
1386 } else { \
1387 TEVENT(Throw IMSX); \
1388 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \
1389 } \
1390 } \
1391 } while (false)
1392
1393 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception.
1394 // TODO-FIXME: remove check_slow() -- it's likely dead.
1395
1396 void ObjectMonitor::check_slow(TRAPS) {
1397 TEVENT(check_slow - throw IMSX);
1398 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1399 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1400 }
1401
1402 static int Adjust(volatile int * adr, int dx) {
1403 int v;
1404 for (v = *adr; Atomic::cmpxchg(v + dx, adr, v) != v; v = *adr) /* empty */;
1405 return v;
1406 }
1407
1408 // helper method for posting a monitor wait event
1409 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event,
1410 jlong notifier_tid,
1411 jlong timeout,
1412 bool timedout) {
1413 event->set_klass(((oop)this->object())->klass());
1414 event->set_timeout((TYPE_ULONG)timeout);
1415 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr()));
1416 event->set_notifier((TYPE_OSTHREAD)notifier_tid);
1417 event->set_timedOut((TYPE_BOOLEAN)timedout);
1418 event->commit();
1419 }
1420
1421 // -----------------------------------------------------------------------------
1422 // Wait/Notify/NotifyAll
1423 //
1424 // Note: a subset of changes to ObjectMonitor::wait()
1425 // will need to be replicated in complete_exit
1426 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1427 Thread * const Self = THREAD;
1428 assert(Self->is_Java_thread(), "Must be Java thread!");
1429 JavaThread *jt = (JavaThread *)THREAD;
1430
1431 DeferredInitialize();
1432
1433 // Throw IMSX or IEX.
1434 CHECK_OWNER();
1435
1436 EventJavaMonitorWait event;
1437
1438 // check for a pending interrupt
1439 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1440 // post monitor waited event. Note that this is past-tense, we are done waiting.
1441 if (JvmtiExport::should_post_monitor_waited()) {
1442 // Note: 'false' parameter is passed here because the
1443 // wait was not timed out due to thread interrupt.
1444 JvmtiExport::post_monitor_waited(jt, this, false);
1445
1446 // In this short circuit of the monitor wait protocol, the
1447 // current thread never drops ownership of the monitor and
1448 // never gets added to the wait queue so the current thread
1449 // cannot be made the successor. This means that the
1450 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1451 // consume an unpark() meant for the ParkEvent associated with
1452 // this ObjectMonitor.
1453 }
1454 if (event.should_commit()) {
1455 post_monitor_wait_event(&event, 0, millis, false);
1456 }
1457 TEVENT(Wait - Throw IEX);
1458 THROW(vmSymbols::java_lang_InterruptedException());
1459 return;
1460 }
1461
1462 TEVENT(Wait);
1463
1464 assert(Self->_Stalled == 0, "invariant");
1465 Self->_Stalled = intptr_t(this);
1466 jt->set_current_waiting_monitor(this);
1467
1468 // create a node to be put into the queue
1469 // Critically, after we reset() the event but prior to park(), we must check
1470 // for a pending interrupt.
1471 ObjectWaiter node(Self);
1472 node.TState = ObjectWaiter::TS_WAIT;
1473 Self->_ParkEvent->reset();
1474 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag
1475
1476 // Enter the waiting queue, which is a circular doubly linked list in this case
1477 // but it could be a priority queue or any data structure.
1478 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only
1479 // by the the owner of the monitor *except* in the case where park()
1480 // returns because of a timeout of interrupt. Contention is exceptionally rare
1481 // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1482
1483 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1484 AddWaiter(&node);
1485 Thread::SpinRelease(&_WaitSetLock);
1486
1487 if ((SyncFlags & 4) == 0) {
1488 _Responsible = NULL;
1489 }
1490 intptr_t save = _recursions; // record the old recursion count
1491 _waiters++; // increment the number of waiters
1492 _recursions = 0; // set the recursion level to be 1
1493 exit(true, Self); // exit the monitor
1494 guarantee(_owner != Self, "invariant");
1495
1496 // The thread is on the WaitSet list - now park() it.
1497 // On MP systems it's conceivable that a brief spin before we park
1498 // could be profitable.
1499 //
1500 // TODO-FIXME: change the following logic to a loop of the form
1501 // while (!timeout && !interrupted && _notified == 0) park()
1502
1503 int ret = OS_OK;
1504 int WasNotified = 0;
1505 { // State transition wrappers
1506 OSThread* osthread = Self->osthread();
1507 OSThreadWaitState osts(osthread, true);
1508 {
1509 ThreadBlockInVM tbivm(jt);
1510 // Thread is in thread_blocked state and oop access is unsafe.
1511 jt->set_suspend_equivalent();
1512
1513 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1514 // Intentionally empty
1515 } else if (node._notified == 0) {
1516 if (millis <= 0) {
1517 Self->_ParkEvent->park();
1518 } else {
1519 ret = Self->_ParkEvent->park(millis);
1520 }
1521 }
1522
1523 // were we externally suspended while we were waiting?
1524 if (ExitSuspendEquivalent (jt)) {
1525 // TODO-FIXME: add -- if succ == Self then succ = null.
1526 jt->java_suspend_self();
1527 }
1528
1529 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1530
1531 // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1532 // from the WaitSet to the EntryList.
1533 // See if we need to remove Node from the WaitSet.
1534 // We use double-checked locking to avoid grabbing _WaitSetLock
1535 // if the thread is not on the wait queue.
1536 //
1537 // Note that we don't need a fence before the fetch of TState.
1538 // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1539 // written by the is thread. (perhaps the fetch might even be satisfied
1540 // by a look-aside into the processor's own store buffer, although given
1541 // the length of the code path between the prior ST and this load that's
1542 // highly unlikely). If the following LD fetches a stale TS_WAIT value
1543 // then we'll acquire the lock and then re-fetch a fresh TState value.
1544 // That is, we fail toward safety.
1545
1546 if (node.TState == ObjectWaiter::TS_WAIT) {
1547 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1548 if (node.TState == ObjectWaiter::TS_WAIT) {
1549 DequeueSpecificWaiter(&node); // unlink from WaitSet
1550 assert(node._notified == 0, "invariant");
1551 node.TState = ObjectWaiter::TS_RUN;
1552 }
1553 Thread::SpinRelease(&_WaitSetLock);
1554 }
1555
1556 // The thread is now either on off-list (TS_RUN),
1557 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1558 // The Node's TState variable is stable from the perspective of this thread.
1559 // No other threads will asynchronously modify TState.
1560 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1561 OrderAccess::loadload();
1562 if (_succ == Self) _succ = NULL;
1563 WasNotified = node._notified;
1564
1565 // Reentry phase -- reacquire the monitor.
1566 // re-enter contended monitor after object.wait().
1567 // retain OBJECT_WAIT state until re-enter successfully completes
1568 // Thread state is thread_in_vm and oop access is again safe,
1569 // although the raw address of the object may have changed.
1570 // (Don't cache naked oops over safepoints, of course).
1571
1572 // post monitor waited event. Note that this is past-tense, we are done waiting.
1573 if (JvmtiExport::should_post_monitor_waited()) {
1574 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1575
1576 if (node._notified != 0 && _succ == Self) {
1577 // In this part of the monitor wait-notify-reenter protocol it
1578 // is possible (and normal) for another thread to do a fastpath
1579 // monitor enter-exit while this thread is still trying to get
1580 // to the reenter portion of the protocol.
1581 //
1582 // The ObjectMonitor was notified and the current thread is
1583 // the successor which also means that an unpark() has already
1584 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1585 // consume the unpark() that was done when the successor was
1586 // set because the same ParkEvent is shared between Java
1587 // monitors and JVM/TI RawMonitors (for now).
1588 //
1589 // We redo the unpark() to ensure forward progress, i.e., we
1590 // don't want all pending threads hanging (parked) with none
1591 // entering the unlocked monitor.
1592 node._event->unpark();
1593 }
1594 }
1595
1596 if (event.should_commit()) {
1597 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT);
1598 }
1599
1600 OrderAccess::fence();
1601
1602 assert(Self->_Stalled != 0, "invariant");
1603 Self->_Stalled = 0;
1604
1605 assert(_owner != Self, "invariant");
1606 ObjectWaiter::TStates v = node.TState;
1607 if (v == ObjectWaiter::TS_RUN) {
1608 enter(Self);
1609 } else {
1610 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1611 ReenterI(Self, &node);
1612 node.wait_reenter_end(this);
1613 }
1614
1615 // Self has reacquired the lock.
1616 // Lifecycle - the node representing Self must not appear on any queues.
1617 // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1618 // want residual elements associated with this thread left on any lists.
1619 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1620 assert(_owner == Self, "invariant");
1621 assert(_succ != Self, "invariant");
1622 } // OSThreadWaitState()
1623
1624 jt->set_current_waiting_monitor(NULL);
1625
1626 guarantee(_recursions == 0, "invariant");
1627 _recursions = save; // restore the old recursion count
1628 _waiters--; // decrement the number of waiters
1629
1630 // Verify a few postconditions
1631 assert(_owner == Self, "invariant");
1632 assert(_succ != Self, "invariant");
1633 assert(((oop)(object()))->mark() == markOopDesc::encode(this), "invariant");
1634
1635 if (SyncFlags & 32) {
1636 OrderAccess::fence();
1637 }
1638
1639 // check if the notification happened
1640 if (!WasNotified) {
1641 // no, it could be timeout or Thread.interrupt() or both
1642 // check for interrupt event, otherwise it is timeout
1643 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1644 TEVENT(Wait - throw IEX from epilog);
1645 THROW(vmSymbols::java_lang_InterruptedException());
1646 }
1647 }
1648
1649 // NOTE: Spurious wake up will be consider as timeout.
1650 // Monitor notify has precedence over thread interrupt.
1651 }
1652
1653
1654 // Consider:
1655 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1656 // then instead of transferring a thread from the WaitSet to the EntryList
1657 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1658
1659 void ObjectMonitor::INotify(Thread * Self) {
1660 const int policy = Knob_MoveNotifyee;
1661
1662 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1663 ObjectWaiter * iterator = DequeueWaiter();
1664 if (iterator != NULL) {
1665 TEVENT(Notify1 - Transfer);
1666 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1667 guarantee(iterator->_notified == 0, "invariant");
1668 // Disposition - what might we do with iterator ?
1669 // a. add it directly to the EntryList - either tail or head.
1670 // b. push it onto the front of the _cxq.
1671 // For now we use (a).
1672 if (policy != 4) {
1673 iterator->TState = ObjectWaiter::TS_ENTER;
1674 }
1675 iterator->_notified = 1;
1676 iterator->_notifier_tid = Self->osthread()->thread_id();
1677
1678 ObjectWaiter * list = _EntryList;
1679 if (list != NULL) {
1680 assert(list->_prev == NULL, "invariant");
1681 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1682 assert(list != iterator, "invariant");
1683 }
1684
1685 if (policy == 0) { // prepend to EntryList
1686 if (list == NULL) {
1687 iterator->_next = iterator->_prev = NULL;
1688 _EntryList = iterator;
1689 } else {
1690 list->_prev = iterator;
1691 iterator->_next = list;
1692 iterator->_prev = NULL;
1693 _EntryList = iterator;
1694 }
1695 } else if (policy == 1) { // append to EntryList
1696 if (list == NULL) {
1697 iterator->_next = iterator->_prev = NULL;
1698 _EntryList = iterator;
1699 } else {
1700 // CONSIDER: finding the tail currently requires a linear-time walk of
1701 // the EntryList. We can make tail access constant-time by converting to
1702 // a CDLL instead of using our current DLL.
1703 ObjectWaiter * tail;
1704 for (tail = list; tail->_next != NULL; tail = tail->_next) /* empty */;
1705 assert(tail != NULL && tail->_next == NULL, "invariant");
1706 tail->_next = iterator;
1707 iterator->_prev = tail;
1708 iterator->_next = NULL;
1709 }
1710 } else if (policy == 2) { // prepend to cxq
1711 if (list == NULL) {
1712 iterator->_next = iterator->_prev = NULL;
1713 _EntryList = iterator;
1714 } else {
1715 iterator->TState = ObjectWaiter::TS_CXQ;
1716 for (;;) {
1717 ObjectWaiter * front = _cxq;
1718 iterator->_next = front;
1719 if (Atomic::cmpxchg_ptr(iterator, &_cxq, front) == front) {
1720 break;
1721 }
1722 }
1723 }
1724 } else if (policy == 3) { // append to cxq
1725 iterator->TState = ObjectWaiter::TS_CXQ;
1726 for (;;) {
1727 ObjectWaiter * tail = _cxq;
1728 if (tail == NULL) {
1729 iterator->_next = NULL;
1730 if (Atomic::cmpxchg_ptr(iterator, &_cxq, NULL) == NULL) {
1731 break;
1732 }
1733 } else {
1734 while (tail->_next != NULL) tail = tail->_next;
1735 tail->_next = iterator;
1736 iterator->_prev = tail;
1737 iterator->_next = NULL;
1738 break;
1739 }
1740 }
1741 } else {
1742 ParkEvent * ev = iterator->_event;
1743 iterator->TState = ObjectWaiter::TS_RUN;
1744 OrderAccess::fence();
1745 ev->unpark();
1746 }
1747
1748 // _WaitSetLock protects the wait queue, not the EntryList. We could
1749 // move the add-to-EntryList operation, above, outside the critical section
1750 // protected by _WaitSetLock. In practice that's not useful. With the
1751 // exception of wait() timeouts and interrupts the monitor owner
1752 // is the only thread that grabs _WaitSetLock. There's almost no contention
1753 // on _WaitSetLock so it's not profitable to reduce the length of the
1754 // critical section.
1755
1756 if (policy < 4) {
1757 iterator->wait_reenter_begin(this);
1758 }
1759 }
1760 Thread::SpinRelease(&_WaitSetLock);
1761 }
1762
1763 // Consider: a not-uncommon synchronization bug is to use notify() when
1764 // notifyAll() is more appropriate, potentially resulting in stranded
1765 // threads; this is one example of a lost wakeup. A useful diagnostic
1766 // option is to force all notify() operations to behave as notifyAll().
1767 //
1768 // Note: We can also detect many such problems with a "minimum wait".
1769 // When the "minimum wait" is set to a small non-zero timeout value
1770 // and the program does not hang whereas it did absent "minimum wait",
1771 // that suggests a lost wakeup bug. The '-XX:SyncFlags=1' option uses
1772 // a "minimum wait" for all park() operations; see the recheckInterval
1773 // variable and MAX_RECHECK_INTERVAL.
1774
1775 void ObjectMonitor::notify(TRAPS) {
1776 CHECK_OWNER();
1777 if (_WaitSet == NULL) {
1778 TEVENT(Empty-Notify);
1779 return;
1780 }
1781 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1782 INotify(THREAD);
1783 if (ObjectMonitor::_sync_Notifications != NULL) {
1784 ObjectMonitor::_sync_Notifications->inc(1);
1785 }
1786 }
1787
1788
1789 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1790 // from the waitset to the EntryList. This could be done more efficiently with a
1791 // single bulk transfer but in practice it's not time-critical. Beware too,
1792 // that in prepend-mode we invert the order of the waiters. Let's say that the
1793 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1794 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1795
1796 void ObjectMonitor::notifyAll(TRAPS) {
1797 CHECK_OWNER();
1798 if (_WaitSet == NULL) {
1799 TEVENT(Empty-NotifyAll);
1800 return;
1801 }
1802
1803 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1804 int tally = 0;
1805 while (_WaitSet != NULL) {
1806 tally++;
1807 INotify(THREAD);
1808 }
1809
1810 if (tally != 0 && ObjectMonitor::_sync_Notifications != NULL) {
1811 ObjectMonitor::_sync_Notifications->inc(tally);
1812 }
1813 }
1814
1815 // -----------------------------------------------------------------------------
1816 // Adaptive Spinning Support
1817 //
1818 // Adaptive spin-then-block - rational spinning
1819 //
1820 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1821 // algorithm. On high order SMP systems it would be better to start with
1822 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH,
1823 // a contending thread could enqueue itself on the cxq and then spin locally
1824 // on a thread-specific variable such as its ParkEvent._Event flag.
1825 // That's left as an exercise for the reader. Note that global spinning is
1826 // not problematic on Niagara, as the L2 cache serves the interconnect and
1827 // has both low latency and massive bandwidth.
1828 //
1829 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1830 // acquisition attempts where we opt to spin -- at 100% and vary the spin count
1831 // (duration) or we can fix the count at approximately the duration of
1832 // a context switch and vary the frequency. Of course we could also
1833 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1834 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html.
1835 //
1836 // This implementation varies the duration "D", where D varies with
1837 // the success rate of recent spin attempts. (D is capped at approximately
1838 // length of a round-trip context switch). The success rate for recent
1839 // spin attempts is a good predictor of the success rate of future spin
1840 // attempts. The mechanism adapts automatically to varying critical
1841 // section length (lock modality), system load and degree of parallelism.
1842 // D is maintained per-monitor in _SpinDuration and is initialized
1843 // optimistically. Spin frequency is fixed at 100%.
1844 //
1845 // Note that _SpinDuration is volatile, but we update it without locks
1846 // or atomics. The code is designed so that _SpinDuration stays within
1847 // a reasonable range even in the presence of races. The arithmetic
1848 // operations on _SpinDuration are closed over the domain of legal values,
1849 // so at worst a race will install and older but still legal value.
1850 // At the very worst this introduces some apparent non-determinism.
1851 // We might spin when we shouldn't or vice-versa, but since the spin
1852 // count are relatively short, even in the worst case, the effect is harmless.
1853 //
1854 // Care must be taken that a low "D" value does not become an
1855 // an absorbing state. Transient spinning failures -- when spinning
1856 // is overall profitable -- should not cause the system to converge
1857 // on low "D" values. We want spinning to be stable and predictable
1858 // and fairly responsive to change and at the same time we don't want
1859 // it to oscillate, become metastable, be "too" non-deterministic,
1860 // or converge on or enter undesirable stable absorbing states.
1861 //
1862 // We implement a feedback-based control system -- using past behavior
1863 // to predict future behavior. We face two issues: (a) if the
1864 // input signal is random then the spin predictor won't provide optimal
1865 // results, and (b) if the signal frequency is too high then the control
1866 // system, which has some natural response lag, will "chase" the signal.
1867 // (b) can arise from multimodal lock hold times. Transient preemption
1868 // can also result in apparent bimodal lock hold times.
1869 // Although sub-optimal, neither condition is particularly harmful, as
1870 // in the worst-case we'll spin when we shouldn't or vice-versa.
1871 // The maximum spin duration is rather short so the failure modes aren't bad.
1872 // To be conservative, I've tuned the gain in system to bias toward
1873 // _not spinning. Relatedly, the system can sometimes enter a mode where it
1874 // "rings" or oscillates between spinning and not spinning. This happens
1875 // when spinning is just on the cusp of profitability, however, so the
1876 // situation is not dire. The state is benign -- there's no need to add
1877 // hysteresis control to damp the transition rate between spinning and
1878 // not spinning.
1879
1880 intptr_t ObjectMonitor::SpinCallbackArgument = 0;
1881 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL;
1882
1883 // Spinning: Fixed frequency (100%), vary duration
1884
1885
1886 int ObjectMonitor::TrySpin_VaryDuration(Thread * Self) {
1887 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning.
1888 int ctr = Knob_FixedSpin;
1889 if (ctr != 0) {
1890 while (--ctr >= 0) {
1891 if (TryLock(Self) > 0) return 1;
1892 SpinPause();
1893 }
1894 return 0;
1895 }
1896
1897 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1898 if (TryLock(Self) > 0) {
1899 // Increase _SpinDuration ...
1900 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1901 // Raising _SpurDuration to the poverty line is key.
1902 int x = _SpinDuration;
1903 if (x < Knob_SpinLimit) {
1904 if (x < Knob_Poverty) x = Knob_Poverty;
1905 _SpinDuration = x + Knob_BonusB;
1906 }
1907 return 1;
1908 }
1909 SpinPause();
1910 }
1911
1912 // Admission control - verify preconditions for spinning
1913 //
1914 // We always spin a little bit, just to prevent _SpinDuration == 0 from
1915 // becoming an absorbing state. Put another way, we spin briefly to
1916 // sample, just in case the system load, parallelism, contention, or lock
1917 // modality changed.
1918 //
1919 // Consider the following alternative:
1920 // Periodically set _SpinDuration = _SpinLimit and try a long/full
1921 // spin attempt. "Periodically" might mean after a tally of
1922 // the # of failed spin attempts (or iterations) reaches some threshold.
1923 // This takes us into the realm of 1-out-of-N spinning, where we
1924 // hold the duration constant but vary the frequency.
1925
1926 ctr = _SpinDuration;
1927 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase;
1928 if (ctr <= 0) return 0;
1929
1930 if (Knob_SuccRestrict && _succ != NULL) return 0;
1931 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
1932 TEVENT(Spin abort - notrunnable [TOP]);
1933 return 0;
1934 }
1935
1936 int MaxSpin = Knob_MaxSpinners;
1937 if (MaxSpin >= 0) {
1938 if (_Spinner > MaxSpin) {
1939 TEVENT(Spin abort -- too many spinners);
1940 return 0;
1941 }
1942 // Slightly racy, but benign ...
1943 Adjust(&_Spinner, 1);
1944 }
1945
1946 // We're good to spin ... spin ingress.
1947 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1948 // when preparing to LD...CAS _owner, etc and the CAS is likely
1949 // to succeed.
1950 int hits = 0;
1951 int msk = 0;
1952 int caspty = Knob_CASPenalty;
1953 int oxpty = Knob_OXPenalty;
1954 int sss = Knob_SpinSetSucc;
1955 if (sss && _succ == NULL) _succ = Self;
1956 Thread * prv = NULL;
1957
1958 // There are three ways to exit the following loop:
1959 // 1. A successful spin where this thread has acquired the lock.
1960 // 2. Spin failure with prejudice
1961 // 3. Spin failure without prejudice
1962
1963 while (--ctr >= 0) {
1964
1965 // Periodic polling -- Check for pending GC
1966 // Threads may spin while they're unsafe.
1967 // We don't want spinning threads to delay the JVM from reaching
1968 // a stop-the-world safepoint or to steal cycles from GC.
1969 // If we detect a pending safepoint we abort in order that
1970 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1971 // this thread, if safe, doesn't steal cycles from GC.
1972 // This is in keeping with the "no loitering in runtime" rule.
1973 // We periodically check to see if there's a safepoint pending.
1974 if ((ctr & 0xFF) == 0) {
1975 if (SafepointSynchronize::do_call_back()) {
1976 TEVENT(Spin: safepoint);
1977 goto Abort; // abrupt spin egress
1978 }
1979 if (Knob_UsePause & 1) SpinPause();
1980
1981 int (*scb)(intptr_t,int) = SpinCallbackFunction;
1982 if (hits > 50 && scb != NULL) {
1983 int abend = (*scb)(SpinCallbackArgument, 0);
1984 }
1985 }
1986
1987 if (Knob_UsePause & 2) SpinPause();
1988
1989 // Exponential back-off ... Stay off the bus to reduce coherency traffic.
1990 // This is useful on classic SMP systems, but is of less utility on
1991 // N1-style CMT platforms.
1992 //
1993 // Trade-off: lock acquisition latency vs coherency bandwidth.
1994 // Lock hold times are typically short. A histogram
1995 // of successful spin attempts shows that we usually acquire
1996 // the lock early in the spin. That suggests we want to
1997 // sample _owner frequently in the early phase of the spin,
1998 // but then back-off and sample less frequently as the spin
1999 // progresses. The back-off makes a good citizen on SMP big
2000 // SMP systems. Oversampling _owner can consume excessive
2001 // coherency bandwidth. Relatedly, if we _oversample _owner we
2002 // can inadvertently interfere with the the ST m->owner=null.
2003 // executed by the lock owner.
2004 if (ctr & msk) continue;
2005 ++hits;
2006 if ((hits & 0xF) == 0) {
2007 // The 0xF, above, corresponds to the exponent.
2008 // Consider: (msk+1)|msk
2009 msk = ((msk << 2)|3) & BackOffMask;
2010 }
2011
2012 // Probe _owner with TATAS
2013 // If this thread observes the monitor transition or flicker
2014 // from locked to unlocked to locked, then the odds that this
2015 // thread will acquire the lock in this spin attempt go down
2016 // considerably. The same argument applies if the CAS fails
2017 // or if we observe _owner change from one non-null value to
2018 // another non-null value. In such cases we might abort
2019 // the spin without prejudice or apply a "penalty" to the
2020 // spin count-down variable "ctr", reducing it by 100, say.
2021
2022 Thread * ox = (Thread *) _owner;
2023 if (ox == NULL) {
2024 ox = (Thread *) Atomic::cmpxchg_ptr(Self, &_owner, NULL);
2025 if (ox == NULL) {
2026 // The CAS succeeded -- this thread acquired ownership
2027 // Take care of some bookkeeping to exit spin state.
2028 if (sss && _succ == Self) {
2029 _succ = NULL;
2030 }
2031 if (MaxSpin > 0) Adjust(&_Spinner, -1);
2032
2033 // Increase _SpinDuration :
2034 // The spin was successful (profitable) so we tend toward
2035 // longer spin attempts in the future.
2036 // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
2037 // If we acquired the lock early in the spin cycle it
2038 // makes sense to increase _SpinDuration proportionally.
2039 // Note that we don't clamp SpinDuration precisely at SpinLimit.
2040 int x = _SpinDuration;
2041 if (x < Knob_SpinLimit) {
2042 if (x < Knob_Poverty) x = Knob_Poverty;
2043 _SpinDuration = x + Knob_Bonus;
2044 }
2045 return 1;
2046 }
2047
2048 // The CAS failed ... we can take any of the following actions:
2049 // * penalize: ctr -= Knob_CASPenalty
2050 // * exit spin with prejudice -- goto Abort;
2051 // * exit spin without prejudice.
2052 // * Since CAS is high-latency, retry again immediately.
2053 prv = ox;
2054 TEVENT(Spin: cas failed);
2055 if (caspty == -2) break;
2056 if (caspty == -1) goto Abort;
2057 ctr -= caspty;
2058 continue;
2059 }
2060
2061 // Did lock ownership change hands ?
2062 if (ox != prv && prv != NULL) {
2063 TEVENT(spin: Owner changed)
2064 if (oxpty == -2) break;
2065 if (oxpty == -1) goto Abort;
2066 ctr -= oxpty;
2067 }
2068 prv = ox;
2069
2070 // Abort the spin if the owner is not executing.
2071 // The owner must be executing in order to drop the lock.
2072 // Spinning while the owner is OFFPROC is idiocy.
2073 // Consider: ctr -= RunnablePenalty ;
2074 if (Knob_OState && NotRunnable (Self, ox)) {
2075 TEVENT(Spin abort - notrunnable);
2076 goto Abort;
2077 }
2078 if (sss && _succ == NULL) _succ = Self;
2079 }
2080
2081 // Spin failed with prejudice -- reduce _SpinDuration.
2082 // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2083 // AIMD is globally stable.
2084 TEVENT(Spin failure);
2085 {
2086 int x = _SpinDuration;
2087 if (x > 0) {
2088 // Consider an AIMD scheme like: x -= (x >> 3) + 100
2089 // This is globally sample and tends to damp the response.
2090 x -= Knob_Penalty;
2091 if (x < 0) x = 0;
2092 _SpinDuration = x;
2093 }
2094 }
2095
2096 Abort:
2097 if (MaxSpin >= 0) Adjust(&_Spinner, -1);
2098 if (sss && _succ == Self) {
2099 _succ = NULL;
2100 // Invariant: after setting succ=null a contending thread
2101 // must recheck-retry _owner before parking. This usually happens
2102 // in the normal usage of TrySpin(), but it's safest
2103 // to make TrySpin() as foolproof as possible.
2104 OrderAccess::fence();
2105 if (TryLock(Self) > 0) return 1;
2106 }
2107 return 0;
2108 }
2109
2110 // NotRunnable() -- informed spinning
2111 //
2112 // Don't bother spinning if the owner is not eligible to drop the lock.
2113 // Peek at the owner's schedctl.sc_state and Thread._thread_values and
2114 // spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2115 // The thread must be runnable in order to drop the lock in timely fashion.
2116 // If the _owner is not runnable then spinning will not likely be
2117 // successful (profitable).
2118 //
2119 // Beware -- the thread referenced by _owner could have died
2120 // so a simply fetch from _owner->_thread_state might trap.
2121 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2122 // Because of the lifecycle issues the schedctl and _thread_state values
2123 // observed by NotRunnable() might be garbage. NotRunnable must
2124 // tolerate this and consider the observed _thread_state value
2125 // as advisory.
2126 //
2127 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2128 // a thread pointer.
2129 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2130 // with the LSB of _owner. Another option would be to probablistically probe
2131 // the putative _owner->TypeTag value.
2132 //
2133 // Checking _thread_state isn't perfect. Even if the thread is
2134 // in_java it might be blocked on a page-fault or have been preempted
2135 // and sitting on a ready/dispatch queue. _thread state in conjunction
2136 // with schedctl.sc_state gives us a good picture of what the
2137 // thread is doing, however.
2138 //
2139 // TODO: check schedctl.sc_state.
2140 // We'll need to use SafeFetch32() to read from the schedctl block.
2141 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/
2142 //
2143 // The return value from NotRunnable() is *advisory* -- the
2144 // result is based on sampling and is not necessarily coherent.
2145 // The caller must tolerate false-negative and false-positive errors.
2146 // Spinning, in general, is probabilistic anyway.
2147
2148
2149 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) {
2150 // Check ox->TypeTag == 2BAD.
2151 if (ox == NULL) return 0;
2152
2153 // Avoid transitive spinning ...
2154 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L.
2155 // Immediately after T1 acquires L it's possible that T2, also
2156 // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2157 // This occurs transiently after T1 acquired L but before
2158 // T1 managed to clear T1.Stalled. T2 does not need to abort
2159 // its spin in this circumstance.
2160 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
2161
2162 if (BlockedOn == 1) return 1;
2163 if (BlockedOn != 0) {
2164 return BlockedOn != intptr_t(this) && _owner == ox;
2165 }
2166
2167 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant");
2168 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);;
2169 // consider also: jst != _thread_in_Java -- but that's overspecific.
2170 return jst == _thread_blocked || jst == _thread_in_native;
2171 }
2172
2173
2174 // -----------------------------------------------------------------------------
2175 // WaitSet management ...
2176
2177 ObjectWaiter::ObjectWaiter(Thread* thread) {
2178 _next = NULL;
2179 _prev = NULL;
2180 _notified = 0;
2181 TState = TS_RUN;
2182 _thread = thread;
2183 _event = thread->_ParkEvent;
2184 _active = false;
2185 assert(_event != NULL, "invariant");
2186 }
2187
2188 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
2189 JavaThread *jt = (JavaThread *)this->_thread;
2190 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2191 }
2192
2193 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
2194 JavaThread *jt = (JavaThread *)this->_thread;
2195 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2196 }
2197
2198 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2199 assert(node != NULL, "should not dequeue NULL node");
2200 assert(node->_prev == NULL, "node already in list");
2201 assert(node->_next == NULL, "node already in list");
2202 // put node at end of queue (circular doubly linked list)
2203 if (_WaitSet == NULL) {
2204 _WaitSet = node;
2205 node->_prev = node;
2206 node->_next = node;
2207 } else {
2208 ObjectWaiter* head = _WaitSet;
2209 ObjectWaiter* tail = head->_prev;
2210 assert(tail->_next == head, "invariant check");
2211 tail->_next = node;
2212 head->_prev = node;
2213 node->_next = head;
2214 node->_prev = tail;
2215 }
2216 }
2217
2218 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2219 // dequeue the very first waiter
2220 ObjectWaiter* waiter = _WaitSet;
2221 if (waiter) {
2222 DequeueSpecificWaiter(waiter);
2223 }
2224 return waiter;
2225 }
2226
2227 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2228 assert(node != NULL, "should not dequeue NULL node");
2229 assert(node->_prev != NULL, "node already removed from list");
2230 assert(node->_next != NULL, "node already removed from list");
2231 // when the waiter has woken up because of interrupt,
2232 // timeout or other spurious wake-up, dequeue the
2233 // waiter from waiting list
2234 ObjectWaiter* next = node->_next;
2235 if (next == node) {
2236 assert(node->_prev == node, "invariant check");
2237 _WaitSet = NULL;
2238 } else {
2239 ObjectWaiter* prev = node->_prev;
2240 assert(prev->_next == node, "invariant check");
2241 assert(next->_prev == node, "invariant check");
2242 next->_prev = prev;
2243 prev->_next = next;
2244 if (_WaitSet == node) {
2245 _WaitSet = next;
2246 }
2247 }
2248 node->_next = NULL;
2249 node->_prev = NULL;
2250 }
2251
2252 // -----------------------------------------------------------------------------
2253 // PerfData support
2254 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL;
2255 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL;
2256 PerfCounter * ObjectMonitor::_sync_Parks = NULL;
2257 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL;
2258 PerfCounter * ObjectMonitor::_sync_Notifications = NULL;
2259 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL;
2260 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL;
2261 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL;
2262 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL;
2263 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL;
2264 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL;
2265 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL;
2266 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL;
2267 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL;
2268 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL;
2269 PerfCounter * ObjectMonitor::_sync_Inflations = NULL;
2270 PerfCounter * ObjectMonitor::_sync_Deflations = NULL;
2271 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL;
2272
2273 // One-shot global initialization for the sync subsystem.
2274 // We could also defer initialization and initialize on-demand
2275 // the first time we call inflate(). Initialization would
2276 // be protected - like so many things - by the MonitorCache_lock.
2277
2278 void ObjectMonitor::Initialize() {
2279 static int InitializationCompleted = 0;
2280 assert(InitializationCompleted == 0, "invariant");
2281 InitializationCompleted = 1;
2282 if (UsePerfData) {
2283 EXCEPTION_MARK;
2284 #define NEWPERFCOUNTER(n) \
2285 { \
2286 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \
2287 CHECK); \
2288 }
2289 #define NEWPERFVARIABLE(n) \
2290 { \
2291 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \
2292 CHECK); \
2293 }
2294 NEWPERFCOUNTER(_sync_Inflations);
2295 NEWPERFCOUNTER(_sync_Deflations);
2296 NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2297 NEWPERFCOUNTER(_sync_FutileWakeups);
2298 NEWPERFCOUNTER(_sync_Parks);
2299 NEWPERFCOUNTER(_sync_EmptyNotifications);
2300 NEWPERFCOUNTER(_sync_Notifications);
2301 NEWPERFCOUNTER(_sync_SlowEnter);
2302 NEWPERFCOUNTER(_sync_SlowExit);
2303 NEWPERFCOUNTER(_sync_SlowNotify);
2304 NEWPERFCOUNTER(_sync_SlowNotifyAll);
2305 NEWPERFCOUNTER(_sync_FailedSpins);
2306 NEWPERFCOUNTER(_sync_SuccessfulSpins);
2307 NEWPERFCOUNTER(_sync_PrivateA);
2308 NEWPERFCOUNTER(_sync_PrivateB);
2309 NEWPERFCOUNTER(_sync_MonInCirculation);
2310 NEWPERFCOUNTER(_sync_MonScavenged);
2311 NEWPERFVARIABLE(_sync_MonExtant);
2312 #undef NEWPERFCOUNTER
2313 #undef NEWPERFVARIABLE
2314 }
2315 }
2316
2317 static char * kvGet(char * kvList, const char * Key) {
2318 if (kvList == NULL) return NULL;
2319 size_t n = strlen(Key);
2320 char * Search;
2321 for (Search = kvList; *Search; Search += strlen(Search) + 1) {
2322 if (strncmp (Search, Key, n) == 0) {
2323 if (Search[n] == '=') return Search + n + 1;
2324 if (Search[n] == 0) return(char *) "1";
2325 }
2326 }
2327 return NULL;
2328 }
2329
2330 static int kvGetInt(char * kvList, const char * Key, int Default) {
2331 char * v = kvGet(kvList, Key);
2332 int rslt = v ? ::strtol(v, NULL, 0) : Default;
2333 if (Knob_ReportSettings && v != NULL) {
2334 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ;
2335 ::fflush(stdout);
2336 }
2337 return rslt;
2338 }
2339
2340 void ObjectMonitor::DeferredInitialize() {
2341 if (InitDone > 0) return;
2342 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2343 while (InitDone != 1) /* empty */;
2344 return;
2345 }
2346
2347 // One-shot global initialization ...
2348 // The initialization is idempotent, so we don't need locks.
2349 // In the future consider doing this via os::init_2().
2350 // SyncKnobs consist of <Key>=<Value> pairs in the style
2351 // of environment variables. Start by converting ':' to NUL.
2352
2353 if (SyncKnobs == NULL) SyncKnobs = "";
2354
2355 size_t sz = strlen(SyncKnobs);
2356 char * knobs = (char *) malloc(sz + 2);
2357 if (knobs == NULL) {
2358 vm_exit_out_of_memory(sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs");
2359 guarantee(0, "invariant");
2360 }
2361 strcpy(knobs, SyncKnobs);
2362 knobs[sz+1] = 0;
2363 for (char * p = knobs; *p; p++) {
2364 if (*p == ':') *p = 0;
2365 }
2366
2367 #define SETKNOB(x) { Knob_##x = kvGetInt(knobs, #x, Knob_##x); }
2368 SETKNOB(ReportSettings);
2369 SETKNOB(Verbose);
2370 SETKNOB(VerifyInUse);
2371 SETKNOB(FixedSpin);
2372 SETKNOB(SpinLimit);
2373 SETKNOB(SpinBase);
2374 SETKNOB(SpinBackOff);
2375 SETKNOB(CASPenalty);
2376 SETKNOB(OXPenalty);
2377 SETKNOB(LogSpins);
2378 SETKNOB(SpinSetSucc);
2379 SETKNOB(SuccEnabled);
2380 SETKNOB(SuccRestrict);
2381 SETKNOB(Penalty);
2382 SETKNOB(Bonus);
2383 SETKNOB(BonusB);
2384 SETKNOB(Poverty);
2385 SETKNOB(SpinAfterFutile);
2386 SETKNOB(UsePause);
2387 SETKNOB(SpinEarly);
2388 SETKNOB(OState);
2389 SETKNOB(MaxSpinners);
2390 SETKNOB(PreSpin);
2391 SETKNOB(ExitPolicy);
2392 SETKNOB(QMode);
2393 SETKNOB(ResetEvent);
2394 SETKNOB(MoveNotifyee);
2395 SETKNOB(FastHSSEC);
2396 #undef SETKNOB
2397
2398 if (Knob_Verbose) {
2399 sanity_checks();
2400 }
2401
2402 if (os::is_MP()) {
2403 BackOffMask = (1 << Knob_SpinBackOff) - 1;
2404 if (Knob_ReportSettings) ::printf("BackOffMask=%X\n", BackOffMask);
2405 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1)
2406 } else {
2407 Knob_SpinLimit = 0;
2408 Knob_SpinBase = 0;
2409 Knob_PreSpin = 0;
2410 Knob_FixedSpin = -1;
2411 }
2412
2413 if (Knob_LogSpins == 0) {
2414 ObjectMonitor::_sync_FailedSpins = NULL;
2415 }
2416
2417 free(knobs);
2418 OrderAccess::fence();
2419 InitDone = 1;
2420 }
2421
2422 void ObjectMonitor::sanity_checks() {
2423 int error_cnt = 0;
2424 int warning_cnt = 0;
2425 bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests);
2426
2427 if (verbose) {
2428 tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT,
2429 sizeof(ObjectMonitor));
2430 tty->print_cr("INFO: sizeof(PaddedEnd<ObjectMonitor>)=" SIZE_FORMAT,
2431 sizeof(PaddedEnd<ObjectMonitor>));
2432 }
2433
2434 uint cache_line_size = VM_Version::L1_data_cache_line_size();
2435 if (verbose) {
2436 tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size);
2437 }
2438
2439 ObjectMonitor dummy;
2440 u_char *addr_begin = (u_char*)&dummy;
2441 u_char *addr_header = (u_char*)&dummy._header;
2442 u_char *addr_owner = (u_char*)&dummy._owner;
2443
2444 uint offset_header = (uint)(addr_header - addr_begin);
2445 if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header);
2446
2447 uint offset_owner = (uint)(addr_owner - addr_begin);
2448 if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner);
2449
2450 if ((uint)(addr_header - addr_begin) != 0) {
2451 tty->print_cr("ERROR: offset(_header) must be zero (0).");
2452 error_cnt++;
2453 }
2454
2455 if (cache_line_size != 0) {
2456 // We were able to determine the L1 data cache line size so
2457 // do some cache line specific sanity checks
2458
2459 if ((offset_owner - offset_header) < cache_line_size) {
2460 tty->print_cr("WARNING: the _header and _owner fields are closer "
2461 "than a cache line which permits false sharing.");
2462 warning_cnt++;
2463 }
2464
2465 if ((sizeof(PaddedEnd<ObjectMonitor>) % cache_line_size) != 0) {
2466 tty->print_cr("WARNING: PaddedEnd<ObjectMonitor> size is not a "
2467 "multiple of a cache line which permits false sharing.");
2468 warning_cnt++;
2469 }
2470 }
2471
2472 ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt,
2473 &warning_cnt);
2474
2475 if (verbose || error_cnt != 0 || warning_cnt != 0) {
2476 tty->print_cr("INFO: error_cnt=%d", error_cnt);
2477 tty->print_cr("INFO: warning_cnt=%d", warning_cnt);
2478 }
2479
2480 guarantee(error_cnt == 0,
2481 "Fatal error(s) found in ObjectMonitor::sanity_checks()");
2482 }
2483
2484 #ifndef PRODUCT
2485 void ObjectMonitor::verify() {
2486 }
2487
2488 void ObjectMonitor::print() {
2489 }
2490 #endif