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 #ifdef TARGET_OS_FAMILY_linux
47 # include "os_linux.inline.hpp"
48 #endif
49 #ifdef TARGET_OS_FAMILY_solaris
50 # include "os_solaris.inline.hpp"
51 #endif
52 #ifdef TARGET_OS_FAMILY_windows
53 # include "os_windows.inline.hpp"
54 #endif
55 #ifdef TARGET_OS_FAMILY_bsd
56 # include "os_bsd.inline.hpp"
57 #endif
58
59 #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64)
60 // Need to inhibit inlining for older versions of GCC to avoid build-time failures
61 #define ATTR __attribute__((noinline))
62 #else
63 #define ATTR
64 #endif
65
66
67 #ifdef DTRACE_ENABLED
68
69 // Only bother with this argument setup if dtrace is available
70 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
71
72
73 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
74 char* bytes = NULL; \
75 int len = 0; \
76 jlong jtid = SharedRuntime::get_java_tid(thread); \
77 Symbol* klassname = ((oop)obj)->klass()->name(); \
78 if (klassname != NULL) { \
79 bytes = (char*)klassname->bytes(); \
80 len = klassname->utf8_length(); \
81 }
82
83 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
84 { \
85 if (DTraceMonitorProbes) { \
86 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
87 HOTSPOT_MONITOR_WAIT(jtid, \
88 (monitor), bytes, len, (millis)); \
89 } \
90 }
91
92 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
93 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
94 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
95 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
96 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
97
98 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
99 { \
100 if (DTraceMonitorProbes) { \
101 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
102 HOTSPOT_MONITOR_##probe(jtid, \
103 (uintptr_t)(monitor), bytes, len); \
104 } \
105 }
106
107 #else // ndef DTRACE_ENABLED
108
109 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
110 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
111
112 #endif // ndef DTRACE_ENABLED
113
114 // Tunables ...
115 // The knob* variables are effectively final. Once set they should
116 // never be modified hence. Consider using __read_mostly with GCC.
117
118 int ObjectMonitor::Knob_Verbose = 0 ;
119 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool -
120 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins
121 static int Knob_HandOff = 0 ;
122 static int Knob_ReportSettings = 0 ;
123
124 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin
125 static int Knob_SpinBackOff = 0 ; // spin-loop backoff
126 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS
127 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change
128 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field
129 static int Knob_SpinEarly = 1 ;
130 static int Knob_SuccEnabled = 1 ; // futile wake throttling
131 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one
132 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs
133 static int Knob_Bonus = 100 ; // spin success bonus
134 static int Knob_BonusB = 100 ; // spin success bonus
135 static int Knob_Penalty = 200 ; // spin failure penalty
136 static int Knob_Poverty = 1000 ;
137 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park()
138 static int Knob_FixedSpin = 0 ;
139 static int Knob_OState = 3 ; // Spinner checks thread state of _owner
140 static int Knob_UsePause = 1 ;
141 static int Knob_ExitPolicy = 0 ;
142 static int Knob_PreSpin = 10 ; // 20-100 likely better
143 static int Knob_ResetEvent = 0 ;
144 static int BackOffMask = 0 ;
145
146 static int Knob_FastHSSEC = 0 ;
147 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee
148 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline
149 static volatile int InitDone = 0 ;
150
151 #define TrySpin TrySpin_VaryDuration
152
153 // -----------------------------------------------------------------------------
154 // Theory of operations -- Monitors lists, thread residency, etc:
155 //
156 // * A thread acquires ownership of a monitor by successfully
157 // CAS()ing the _owner field from null to non-null.
158 //
159 // * Invariant: A thread appears on at most one monitor list --
160 // cxq, EntryList or WaitSet -- at any one time.
161 //
162 // * Contending threads "push" themselves onto the cxq with CAS
163 // and then spin/park.
164 //
165 // * After a contending thread eventually acquires the lock it must
166 // dequeue itself from either the EntryList or the cxq.
167 //
168 // * The exiting thread identifies and unparks an "heir presumptive"
169 // tentative successor thread on the EntryList. Critically, the
170 // exiting thread doesn't unlink the successor thread from the EntryList.
171 // After having been unparked, the wakee will recontend for ownership of
172 // the monitor. The successor (wakee) will either acquire the lock or
173 // re-park itself.
174 //
175 // Succession is provided for by a policy of competitive handoff.
176 // The exiting thread does _not_ grant or pass ownership to the
177 // successor thread. (This is also referred to as "handoff" succession").
178 // Instead the exiting thread releases ownership and possibly wakes
179 // a successor, so the successor can (re)compete for ownership of the lock.
180 // If the EntryList is empty but the cxq is populated the exiting
181 // thread will drain the cxq into the EntryList. It does so by
182 // by detaching the cxq (installing null with CAS) and folding
183 // the threads from the cxq into the EntryList. The EntryList is
184 // doubly linked, while the cxq is singly linked because of the
185 // CAS-based "push" used to enqueue recently arrived threads (RATs).
186 //
187 // * Concurrency invariants:
188 //
189 // -- only the monitor owner may access or mutate the EntryList.
190 // The mutex property of the monitor itself protects the EntryList
191 // from concurrent interference.
192 // -- Only the monitor owner may detach the cxq.
193 //
194 // * The monitor entry list operations avoid locks, but strictly speaking
195 // they're not lock-free. Enter is lock-free, exit is not.
196 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html
197 //
198 // * The cxq can have multiple concurrent "pushers" but only one concurrent
199 // detaching thread. This mechanism is immune from the ABA corruption.
200 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
201 //
202 // * Taken together, the cxq and the EntryList constitute or form a
203 // single logical queue of threads stalled trying to acquire the lock.
204 // We use two distinct lists to improve the odds of a constant-time
205 // dequeue operation after acquisition (in the ::enter() epilogue) and
206 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm).
207 // A key desideratum is to minimize queue & monitor metadata manipulation
208 // that occurs while holding the monitor lock -- that is, we want to
209 // minimize monitor lock holds times. Note that even a small amount of
210 // fixed spinning will greatly reduce the # of enqueue-dequeue operations
211 // on EntryList|cxq. That is, spinning relieves contention on the "inner"
212 // locks and monitor metadata.
213 //
214 // Cxq points to the the set of Recently Arrived Threads attempting entry.
215 // Because we push threads onto _cxq with CAS, the RATs must take the form of
216 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when
217 // the unlocking thread notices that EntryList is null but _cxq is != null.
218 //
219 // The EntryList is ordered by the prevailing queue discipline and
220 // can be organized in any convenient fashion, such as a doubly-linked list or
221 // a circular doubly-linked list. Critically, we want insert and delete operations
222 // to operate in constant-time. If we need a priority queue then something akin
223 // to Solaris' sleepq would work nicely. Viz.,
224 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
225 // Queue discipline is enforced at ::exit() time, when the unlocking thread
226 // drains the cxq into the EntryList, and orders or reorders the threads on the
227 // EntryList accordingly.
228 //
229 // Barring "lock barging", this mechanism provides fair cyclic ordering,
230 // somewhat similar to an elevator-scan.
231 //
232 // * The monitor synchronization subsystem avoids the use of native
233 // synchronization primitives except for the narrow platform-specific
234 // park-unpark abstraction. See the comments in os_solaris.cpp regarding
235 // the semantics of park-unpark. Put another way, this monitor implementation
236 // depends only on atomic operations and park-unpark. The monitor subsystem
237 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
238 // underlying OS manages the READY<->RUN transitions.
239 //
240 // * Waiting threads reside on the WaitSet list -- wait() puts
241 // the caller onto the WaitSet.
242 //
243 // * notify() or notifyAll() simply transfers threads from the WaitSet to
244 // either the EntryList or cxq. Subsequent exit() operations will
245 // unpark the notifyee. Unparking a notifee in notify() is inefficient -
246 // it's likely the notifyee would simply impale itself on the lock held
247 // by the notifier.
248 //
249 // * An interesting alternative is to encode cxq as (List,LockByte) where
250 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary
251 // variable, like _recursions, in the scheme. The threads or Events that form
252 // the list would have to be aligned in 256-byte addresses. A thread would
253 // try to acquire the lock or enqueue itself with CAS, but exiting threads
254 // could use a 1-0 protocol and simply STB to set the LockByte to 0.
255 // Note that is is *not* word-tearing, but it does presume that full-word
256 // CAS operations are coherent with intermix with STB operations. That's true
257 // on most common processors.
258 //
259 // * See also http://blogs.sun.com/dave
260
261
262 // -----------------------------------------------------------------------------
263 // Enter support
264
265 bool ObjectMonitor::try_enter(Thread* THREAD) {
266 if (THREAD != _owner) {
267 if (THREAD->is_lock_owned ((address)_owner)) {
268 assert(_recursions == 0, "internal state error");
269 _owner = THREAD ;
270 _recursions = 1 ;
271 OwnerIsThread = 1 ;
272 return true;
273 }
274 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
275 return false;
276 }
277 return true;
278 } else {
279 _recursions++;
280 return true;
281 }
282 }
283
284 void ATTR ObjectMonitor::enter(TRAPS) {
285 // The following code is ordered to check the most common cases first
286 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
287 Thread * const Self = THREAD ;
288 void * cur ;
289
290 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ;
291 if (cur == NULL) {
292 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
293 assert (_recursions == 0 , "invariant") ;
294 assert (_owner == Self, "invariant") ;
295 // CONSIDER: set or assert OwnerIsThread == 1
296 return ;
297 }
298
299 if (cur == Self) {
300 // TODO-FIXME: check for integer overflow! BUGID 6557169.
301 _recursions ++ ;
302 return ;
303 }
304
305 if (Self->is_lock_owned ((address)cur)) {
306 assert (_recursions == 0, "internal state error");
307 _recursions = 1 ;
308 // Commute owner from a thread-specific on-stack BasicLockObject address to
309 // a full-fledged "Thread *".
310 _owner = Self ;
311 OwnerIsThread = 1 ;
312 return ;
313 }
314
315 // We've encountered genuine contention.
316 assert (Self->_Stalled == 0, "invariant") ;
317 Self->_Stalled = intptr_t(this) ;
318
319 // Try one round of spinning *before* enqueueing Self
320 // and before going through the awkward and expensive state
321 // transitions. The following spin is strictly optional ...
322 // Note that if we acquire the monitor from an initial spin
323 // we forgo posting JVMTI events and firing DTRACE probes.
324 if (Knob_SpinEarly && TrySpin (Self) > 0) {
325 assert (_owner == Self , "invariant") ;
326 assert (_recursions == 0 , "invariant") ;
327 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
328 Self->_Stalled = 0 ;
329 return ;
330 }
331
332 assert (_owner != Self , "invariant") ;
333 assert (_succ != Self , "invariant") ;
334 assert (Self->is_Java_thread() , "invariant") ;
335 JavaThread * jt = (JavaThread *) Self ;
336 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ;
337 assert (jt->thread_state() != _thread_blocked , "invariant") ;
338 assert (this->object() != NULL , "invariant") ;
339 assert (_count >= 0, "invariant") ;
340
341 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy().
342 // Ensure the object-monitor relationship remains stable while there's contention.
343 Atomic::inc_ptr(&_count);
344
345 EventJavaMonitorEnter event;
346
347 { // Change java thread status to indicate blocked on monitor enter.
348 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
349
350 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
351 if (JvmtiExport::should_post_monitor_contended_enter()) {
352 JvmtiExport::post_monitor_contended_enter(jt, this);
353
354 // The current thread does not yet own the monitor and does not
355 // yet appear on any queues that would get it made the successor.
356 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
357 // handler cannot accidentally consume an unpark() meant for the
358 // ParkEvent associated with this ObjectMonitor.
359 }
360
361 OSThreadContendState osts(Self->osthread());
362 ThreadBlockInVM tbivm(jt);
363
364 Self->set_current_pending_monitor(this);
365
366 // TODO-FIXME: change the following for(;;) loop to straight-line code.
367 for (;;) {
368 jt->set_suspend_equivalent();
369 // cleared by handle_special_suspend_equivalent_condition()
370 // or java_suspend_self()
371
372 EnterI (THREAD) ;
373
374 if (!ExitSuspendEquivalent(jt)) break ;
375
376 //
377 // We have acquired the contended monitor, but while we were
378 // waiting another thread suspended us. We don't want to enter
379 // the monitor while suspended because that would surprise the
380 // thread that suspended us.
381 //
382 _recursions = 0 ;
383 _succ = NULL ;
384 exit (false, Self) ;
385
386 jt->java_suspend_self();
387 }
388 Self->set_current_pending_monitor(NULL);
389 }
390
391 Atomic::dec_ptr(&_count);
392 assert (_count >= 0, "invariant") ;
393 Self->_Stalled = 0 ;
394
395 // Must either set _recursions = 0 or ASSERT _recursions == 0.
396 assert (_recursions == 0 , "invariant") ;
397 assert (_owner == Self , "invariant") ;
398 assert (_succ != Self , "invariant") ;
399 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
400
401 // The thread -- now the owner -- is back in vm mode.
402 // Report the glorious news via TI,DTrace and jvmstat.
403 // The probe effect is non-trivial. All the reportage occurs
404 // while we hold the monitor, increasing the length of the critical
405 // section. Amdahl's parallel speedup law comes vividly into play.
406 //
407 // Another option might be to aggregate the events (thread local or
408 // per-monitor aggregation) and defer reporting until a more opportune
409 // time -- such as next time some thread encounters contention but has
410 // yet to acquire the lock. While spinning that thread could
411 // spinning we could increment JVMStat counters, etc.
412
413 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
414 if (JvmtiExport::should_post_monitor_contended_entered()) {
415 JvmtiExport::post_monitor_contended_entered(jt, this);
416
417 // The current thread already owns the monitor and is not going to
418 // call park() for the remainder of the monitor enter protocol. So
419 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
420 // event handler consumed an unpark() issued by the thread that
421 // just exited the monitor.
422 }
423
424 if (event.should_commit()) {
425 event.set_klass(((oop)this->object())->klass());
426 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid);
427 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr()));
428 event.commit();
429 }
430
431 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) {
432 ObjectMonitor::_sync_ContendedLockAttempts->inc() ;
433 }
434 }
435
436
437 // Caveat: TryLock() is not necessarily serializing if it returns failure.
438 // Callers must compensate as needed.
439
440 int ObjectMonitor::TryLock (Thread * Self) {
441 for (;;) {
442 void * own = _owner ;
443 if (own != NULL) return 0 ;
444 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) {
445 // Either guarantee _recursions == 0 or set _recursions = 0.
446 assert (_recursions == 0, "invariant") ;
447 assert (_owner == Self, "invariant") ;
448 // CONSIDER: set or assert that OwnerIsThread == 1
449 return 1 ;
450 }
451 // The lock had been free momentarily, but we lost the race to the lock.
452 // Interference -- the CAS failed.
453 // We can either return -1 or retry.
454 // Retry doesn't make as much sense because the lock was just acquired.
455 if (true) return -1 ;
456 }
457 }
458
459 void ATTR ObjectMonitor::EnterI (TRAPS) {
460 Thread * Self = THREAD ;
461 assert (Self->is_Java_thread(), "invariant") ;
462 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ;
463
464 // Try the lock - TATAS
465 if (TryLock (Self) > 0) {
466 assert (_succ != Self , "invariant") ;
467 assert (_owner == Self , "invariant") ;
468 assert (_Responsible != Self , "invariant") ;
469 return ;
470 }
471
472 DeferredInitialize () ;
473
474 // We try one round of spinning *before* enqueueing Self.
475 //
476 // If the _owner is ready but OFFPROC we could use a YieldTo()
477 // operation to donate the remainder of this thread's quantum
478 // to the owner. This has subtle but beneficial affinity
479 // effects.
480
481 if (TrySpin (Self) > 0) {
482 assert (_owner == Self , "invariant") ;
483 assert (_succ != Self , "invariant") ;
484 assert (_Responsible != Self , "invariant") ;
485 return ;
486 }
487
488 // The Spin failed -- Enqueue and park the thread ...
489 assert (_succ != Self , "invariant") ;
490 assert (_owner != Self , "invariant") ;
491 assert (_Responsible != Self , "invariant") ;
492
493 // Enqueue "Self" on ObjectMonitor's _cxq.
494 //
495 // Node acts as a proxy for Self.
496 // As an aside, if were to ever rewrite the synchronization code mostly
497 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
498 // Java objects. This would avoid awkward lifecycle and liveness issues,
499 // as well as eliminate a subset of ABA issues.
500 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
501 //
502
503 ObjectWaiter node(Self) ;
504 Self->_ParkEvent->reset() ;
505 node._prev = (ObjectWaiter *) 0xBAD ;
506 node.TState = ObjectWaiter::TS_CXQ ;
507
508 // Push "Self" onto the front of the _cxq.
509 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
510 // Note that spinning tends to reduce the rate at which threads
511 // enqueue and dequeue on EntryList|cxq.
512 ObjectWaiter * nxt ;
513 for (;;) {
514 node._next = nxt = _cxq ;
515 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ;
516
517 // Interference - the CAS failed because _cxq changed. Just retry.
518 // As an optional optimization we retry the lock.
519 if (TryLock (Self) > 0) {
520 assert (_succ != Self , "invariant") ;
521 assert (_owner == Self , "invariant") ;
522 assert (_Responsible != Self , "invariant") ;
523 return ;
524 }
525 }
526
527 // Check for cxq|EntryList edge transition to non-null. This indicates
528 // the onset of contention. While contention persists exiting threads
529 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit
530 // operations revert to the faster 1-0 mode. This enter operation may interleave
531 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
532 // arrange for one of the contending thread to use a timed park() operations
533 // to detect and recover from the race. (Stranding is form of progress failure
534 // where the monitor is unlocked but all the contending threads remain parked).
535 // That is, at least one of the contended threads will periodically poll _owner.
536 // One of the contending threads will become the designated "Responsible" thread.
537 // The Responsible thread uses a timed park instead of a normal indefinite park
538 // operation -- it periodically wakes and checks for and recovers from potential
539 // strandings admitted by 1-0 exit operations. We need at most one Responsible
540 // thread per-monitor at any given moment. Only threads on cxq|EntryList may
541 // be responsible for a monitor.
542 //
543 // Currently, one of the contended threads takes on the added role of "Responsible".
544 // A viable alternative would be to use a dedicated "stranding checker" thread
545 // that periodically iterated over all the threads (or active monitors) and unparked
546 // successors where there was risk of stranding. This would help eliminate the
547 // timer scalability issues we see on some platforms as we'd only have one thread
548 // -- the checker -- parked on a timer.
549
550 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) {
551 // Try to assume the role of responsible thread for the monitor.
552 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self }
553 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ;
554 }
555
556 // The lock have been released while this thread was occupied queueing
557 // itself onto _cxq. To close the race and avoid "stranding" and
558 // progress-liveness failure we must resample-retry _owner before parking.
559 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
560 // In this case the ST-MEMBAR is accomplished with CAS().
561 //
562 // TODO: Defer all thread state transitions until park-time.
563 // Since state transitions are heavy and inefficient we'd like
564 // to defer the state transitions until absolutely necessary,
565 // and in doing so avoid some transitions ...
566
567 TEVENT (Inflated enter - Contention) ;
568 int nWakeups = 0 ;
569 int RecheckInterval = 1 ;
570
571 for (;;) {
572
573 if (TryLock (Self) > 0) break ;
574 assert (_owner != Self, "invariant") ;
575
576 if ((SyncFlags & 2) && _Responsible == NULL) {
577 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ;
578 }
579
580 // park self
581 if (_Responsible == Self || (SyncFlags & 1)) {
582 TEVENT (Inflated enter - park TIMED) ;
583 Self->_ParkEvent->park ((jlong) RecheckInterval) ;
584 // Increase the RecheckInterval, but clamp the value.
585 RecheckInterval *= 8 ;
586 if (RecheckInterval > 1000) RecheckInterval = 1000 ;
587 } else {
588 TEVENT (Inflated enter - park UNTIMED) ;
589 Self->_ParkEvent->park() ;
590 }
591
592 if (TryLock(Self) > 0) break ;
593
594 // The lock is still contested.
595 // Keep a tally of the # of futile wakeups.
596 // Note that the counter is not protected by a lock or updated by atomics.
597 // That is by design - we trade "lossy" counters which are exposed to
598 // races during updates for a lower probe effect.
599 TEVENT (Inflated enter - Futile wakeup) ;
600 if (ObjectMonitor::_sync_FutileWakeups != NULL) {
601 ObjectMonitor::_sync_FutileWakeups->inc() ;
602 }
603 ++ nWakeups ;
604
605 // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
606 // We can defer clearing _succ until after the spin completes
607 // TrySpin() must tolerate being called with _succ == Self.
608 // Try yet another round of adaptive spinning.
609 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ;
610
611 // We can find that we were unpark()ed and redesignated _succ while
612 // we were spinning. That's harmless. If we iterate and call park(),
613 // park() will consume the event and return immediately and we'll
614 // just spin again. This pattern can repeat, leaving _succ to simply
615 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks().
616 // Alternately, we can sample fired() here, and if set, forgo spinning
617 // in the next iteration.
618
619 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
620 Self->_ParkEvent->reset() ;
621 OrderAccess::fence() ;
622 }
623 if (_succ == Self) _succ = NULL ;
624
625 // Invariant: after clearing _succ a thread *must* retry _owner before parking.
626 OrderAccess::fence() ;
627 }
628
629 // Egress :
630 // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
631 // Normally we'll find Self on the EntryList .
632 // From the perspective of the lock owner (this thread), the
633 // EntryList is stable and cxq is prepend-only.
634 // The head of cxq is volatile but the interior is stable.
635 // In addition, Self.TState is stable.
636
637 assert (_owner == Self , "invariant") ;
638 assert (object() != NULL , "invariant") ;
639 // I'd like to write:
640 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
641 // but as we're at a safepoint that's not safe.
642
643 UnlinkAfterAcquire (Self, &node) ;
644 if (_succ == Self) _succ = NULL ;
645
646 assert (_succ != Self, "invariant") ;
647 if (_Responsible == Self) {
648 _Responsible = NULL ;
649 OrderAccess::fence(); // Dekker pivot-point
650
651 // We may leave threads on cxq|EntryList without a designated
652 // "Responsible" thread. This is benign. When this thread subsequently
653 // exits the monitor it can "see" such preexisting "old" threads --
654 // threads that arrived on the cxq|EntryList before the fence, above --
655 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads
656 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
657 // non-null and elect a new "Responsible" timer thread.
658 //
659 // This thread executes:
660 // ST Responsible=null; MEMBAR (in enter epilogue - here)
661 // LD cxq|EntryList (in subsequent exit)
662 //
663 // Entering threads in the slow/contended path execute:
664 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
665 // The (ST cxq; MEMBAR) is accomplished with CAS().
666 //
667 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
668 // exit operation from floating above the ST Responsible=null.
669 }
670
671 // We've acquired ownership with CAS().
672 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
673 // But since the CAS() this thread may have also stored into _succ,
674 // EntryList, cxq or Responsible. These meta-data updates must be
675 // visible __before this thread subsequently drops the lock.
676 // Consider what could occur if we didn't enforce this constraint --
677 // STs to monitor meta-data and user-data could reorder with (become
678 // visible after) the ST in exit that drops ownership of the lock.
679 // Some other thread could then acquire the lock, but observe inconsistent
680 // or old monitor meta-data and heap data. That violates the JMM.
681 // To that end, the 1-0 exit() operation must have at least STST|LDST
682 // "release" barrier semantics. Specifically, there must be at least a
683 // STST|LDST barrier in exit() before the ST of null into _owner that drops
684 // the lock. The barrier ensures that changes to monitor meta-data and data
685 // protected by the lock will be visible before we release the lock, and
686 // therefore before some other thread (CPU) has a chance to acquire the lock.
687 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
688 //
689 // Critically, any prior STs to _succ or EntryList must be visible before
690 // the ST of null into _owner in the *subsequent* (following) corresponding
691 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily
692 // execute a serializing instruction.
693
694 if (SyncFlags & 8) {
695 OrderAccess::fence() ;
696 }
697 return ;
698 }
699
700 // ReenterI() is a specialized inline form of the latter half of the
701 // contended slow-path from EnterI(). We use ReenterI() only for
702 // monitor reentry in wait().
703 //
704 // In the future we should reconcile EnterI() and ReenterI(), adding
705 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the
706 // loop accordingly.
707
708 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) {
709 assert (Self != NULL , "invariant") ;
710 assert (SelfNode != NULL , "invariant") ;
711 assert (SelfNode->_thread == Self , "invariant") ;
712 assert (_waiters > 0 , "invariant") ;
713 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ;
714 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ;
715 JavaThread * jt = (JavaThread *) Self ;
716
717 int nWakeups = 0 ;
718 for (;;) {
719 ObjectWaiter::TStates v = SelfNode->TState ;
720 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ;
721 assert (_owner != Self, "invariant") ;
722
723 if (TryLock (Self) > 0) break ;
724 if (TrySpin (Self) > 0) break ;
725
726 TEVENT (Wait Reentry - parking) ;
727
728 // State transition wrappers around park() ...
729 // ReenterI() wisely defers state transitions until
730 // it's clear we must park the thread.
731 {
732 OSThreadContendState osts(Self->osthread());
733 ThreadBlockInVM tbivm(jt);
734
735 // cleared by handle_special_suspend_equivalent_condition()
736 // or java_suspend_self()
737 jt->set_suspend_equivalent();
738 if (SyncFlags & 1) {
739 Self->_ParkEvent->park ((jlong)1000) ;
740 } else {
741 Self->_ParkEvent->park () ;
742 }
743
744 // were we externally suspended while we were waiting?
745 for (;;) {
746 if (!ExitSuspendEquivalent (jt)) break ;
747 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
748 jt->java_suspend_self();
749 jt->set_suspend_equivalent();
750 }
751 }
752
753 // Try again, but just so we distinguish between futile wakeups and
754 // successful wakeups. The following test isn't algorithmically
755 // necessary, but it helps us maintain sensible statistics.
756 if (TryLock(Self) > 0) break ;
757
758 // The lock is still contested.
759 // Keep a tally of the # of futile wakeups.
760 // Note that the counter is not protected by a lock or updated by atomics.
761 // That is by design - we trade "lossy" counters which are exposed to
762 // races during updates for a lower probe effect.
763 TEVENT (Wait Reentry - futile wakeup) ;
764 ++ nWakeups ;
765
766 // Assuming this is not a spurious wakeup we'll normally
767 // find that _succ == Self.
768 if (_succ == Self) _succ = NULL ;
769
770 // Invariant: after clearing _succ a contending thread
771 // *must* retry _owner before parking.
772 OrderAccess::fence() ;
773
774 if (ObjectMonitor::_sync_FutileWakeups != NULL) {
775 ObjectMonitor::_sync_FutileWakeups->inc() ;
776 }
777 }
778
779 // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
780 // Normally we'll find Self on the EntryList.
781 // Unlinking from the EntryList is constant-time and atomic-free.
782 // From the perspective of the lock owner (this thread), the
783 // EntryList is stable and cxq is prepend-only.
784 // The head of cxq is volatile but the interior is stable.
785 // In addition, Self.TState is stable.
786
787 assert (_owner == Self, "invariant") ;
788 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
789 UnlinkAfterAcquire (Self, SelfNode) ;
790 if (_succ == Self) _succ = NULL ;
791 assert (_succ != Self, "invariant") ;
792 SelfNode->TState = ObjectWaiter::TS_RUN ;
793 OrderAccess::fence() ; // see comments at the end of EnterI()
794 }
795
796 // after the thread acquires the lock in ::enter(). Equally, we could defer
797 // unlinking the thread until ::exit()-time.
798
799 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode)
800 {
801 assert (_owner == Self, "invariant") ;
802 assert (SelfNode->_thread == Self, "invariant") ;
803
804 if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
805 // Normal case: remove Self from the DLL EntryList .
806 // This is a constant-time operation.
807 ObjectWaiter * nxt = SelfNode->_next ;
808 ObjectWaiter * prv = SelfNode->_prev ;
809 if (nxt != NULL) nxt->_prev = prv ;
810 if (prv != NULL) prv->_next = nxt ;
811 if (SelfNode == _EntryList ) _EntryList = nxt ;
812 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ;
813 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ;
814 TEVENT (Unlink from EntryList) ;
815 } else {
816 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ;
817 // Inopportune interleaving -- Self is still on the cxq.
818 // This usually means the enqueue of self raced an exiting thread.
819 // Normally we'll find Self near the front of the cxq, so
820 // dequeueing is typically fast. If needbe we can accelerate
821 // this with some MCS/CHL-like bidirectional list hints and advisory
822 // back-links so dequeueing from the interior will normally operate
823 // in constant-time.
824 // Dequeue Self from either the head (with CAS) or from the interior
825 // with a linear-time scan and normal non-atomic memory operations.
826 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
827 // and then unlink Self from EntryList. We have to drain eventually,
828 // so it might as well be now.
829
830 ObjectWaiter * v = _cxq ;
831 assert (v != NULL, "invariant") ;
832 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) {
833 // The CAS above can fail from interference IFF a "RAT" arrived.
834 // In that case Self must be in the interior and can no longer be
835 // at the head of cxq.
836 if (v == SelfNode) {
837 assert (_cxq != v, "invariant") ;
838 v = _cxq ; // CAS above failed - start scan at head of list
839 }
840 ObjectWaiter * p ;
841 ObjectWaiter * q = NULL ;
842 for (p = v ; p != NULL && p != SelfNode; p = p->_next) {
843 q = p ;
844 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ;
845 }
846 assert (v != SelfNode, "invariant") ;
847 assert (p == SelfNode, "Node not found on cxq") ;
848 assert (p != _cxq, "invariant") ;
849 assert (q != NULL, "invariant") ;
850 assert (q->_next == p, "invariant") ;
851 q->_next = p->_next ;
852 }
853 TEVENT (Unlink from cxq) ;
854 }
855
856 // Diagnostic hygiene ...
857 SelfNode->_prev = (ObjectWaiter *) 0xBAD ;
858 SelfNode->_next = (ObjectWaiter *) 0xBAD ;
859 SelfNode->TState = ObjectWaiter::TS_RUN ;
860 }
861
862 // -----------------------------------------------------------------------------
863 // Exit support
864 //
865 // exit()
866 // ~~~~~~
867 // Note that the collector can't reclaim the objectMonitor or deflate
868 // the object out from underneath the thread calling ::exit() as the
869 // thread calling ::exit() never transitions to a stable state.
870 // This inhibits GC, which in turn inhibits asynchronous (and
871 // inopportune) reclamation of "this".
872 //
873 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
874 // There's one exception to the claim above, however. EnterI() can call
875 // exit() to drop a lock if the acquirer has been externally suspended.
876 // In that case exit() is called with _thread_state as _thread_blocked,
877 // but the monitor's _count field is > 0, which inhibits reclamation.
878 //
879 // 1-0 exit
880 // ~~~~~~~~
881 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
882 // the fast-path operators have been optimized so the common ::exit()
883 // operation is 1-0. See i486.ad fast_unlock(), for instance.
884 // The code emitted by fast_unlock() elides the usual MEMBAR. This
885 // greatly improves latency -- MEMBAR and CAS having considerable local
886 // latency on modern processors -- but at the cost of "stranding". Absent the
887 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
888 // ::enter() path, resulting in the entering thread being stranding
889 // and a progress-liveness failure. Stranding is extremely rare.
890 // We use timers (timed park operations) & periodic polling to detect
891 // and recover from stranding. Potentially stranded threads periodically
892 // wake up and poll the lock. See the usage of the _Responsible variable.
893 //
894 // The CAS() in enter provides for safety and exclusion, while the CAS or
895 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking
896 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding.
897 // We detect and recover from stranding with timers.
898 //
899 // If a thread transiently strands it'll park until (a) another
900 // thread acquires the lock and then drops the lock, at which time the
901 // exiting thread will notice and unpark the stranded thread, or, (b)
902 // the timer expires. If the lock is high traffic then the stranding latency
903 // will be low due to (a). If the lock is low traffic then the odds of
904 // stranding are lower, although the worst-case stranding latency
905 // is longer. Critically, we don't want to put excessive load in the
906 // platform's timer subsystem. We want to minimize both the timer injection
907 // rate (timers created/sec) as well as the number of timers active at
908 // any one time. (more precisely, we want to minimize timer-seconds, which is
909 // the integral of the # of active timers at any instant over time).
910 // Both impinge on OS scalability. Given that, at most one thread parked on
911 // a monitor will use a timer.
912
913 void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) {
914 Thread * Self = THREAD ;
915 if (THREAD != _owner) {
916 if (THREAD->is_lock_owned((address) _owner)) {
917 // Transmute _owner from a BasicLock pointer to a Thread address.
918 // We don't need to hold _mutex for this transition.
919 // Non-null to Non-null is safe as long as all readers can
920 // tolerate either flavor.
921 assert (_recursions == 0, "invariant") ;
922 _owner = THREAD ;
923 _recursions = 0 ;
924 OwnerIsThread = 1 ;
925 } else {
926 // NOTE: we need to handle unbalanced monitor enter/exit
927 // in native code by throwing an exception.
928 // TODO: Throw an IllegalMonitorStateException ?
929 TEVENT (Exit - Throw IMSX) ;
930 assert(false, "Non-balanced monitor enter/exit!");
931 if (false) {
932 THROW(vmSymbols::java_lang_IllegalMonitorStateException());
933 }
934 return;
935 }
936 }
937
938 if (_recursions != 0) {
939 _recursions--; // this is simple recursive enter
940 TEVENT (Inflated exit - recursive) ;
941 return ;
942 }
943
944 // Invariant: after setting Responsible=null an thread must execute
945 // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
946 if ((SyncFlags & 4) == 0) {
947 _Responsible = NULL ;
948 }
949
950 #if INCLUDE_TRACE
951 // get the owner's thread id for the MonitorEnter event
952 // if it is enabled and the thread isn't suspended
953 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) {
954 _previous_owner_tid = SharedRuntime::get_java_tid(Self);
955 }
956 #endif
957
958 for (;;) {
959 assert (THREAD == _owner, "invariant") ;
960
961
962 if (Knob_ExitPolicy == 0) {
963 // release semantics: prior loads and stores from within the critical section
964 // must not float (reorder) past the following store that drops the lock.
965 // On SPARC that requires MEMBAR #loadstore|#storestore.
966 // But of course in TSO #loadstore|#storestore is not required.
967 // I'd like to write one of the following:
968 // A. OrderAccess::release() ; _owner = NULL
969 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
970 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
971 // store into a _dummy variable. That store is not needed, but can result
972 // in massive wasteful coherency traffic on classic SMP systems.
973 // Instead, I use release_store(), which is implemented as just a simple
974 // ST on x64, x86 and SPARC.
975 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock
976 OrderAccess::storeload() ; // See if we need to wake a successor
977 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
978 TEVENT (Inflated exit - simple egress) ;
979 return ;
980 }
981 TEVENT (Inflated exit - complex egress) ;
982
983 // Normally the exiting thread is responsible for ensuring succession,
984 // but if other successors are ready or other entering threads are spinning
985 // then this thread can simply store NULL into _owner and exit without
986 // waking a successor. The existence of spinners or ready successors
987 // guarantees proper succession (liveness). Responsibility passes to the
988 // ready or running successors. The exiting thread delegates the duty.
989 // More precisely, if a successor already exists this thread is absolved
990 // of the responsibility of waking (unparking) one.
991 //
992 // The _succ variable is critical to reducing futile wakeup frequency.
993 // _succ identifies the "heir presumptive" thread that has been made
994 // ready (unparked) but that has not yet run. We need only one such
995 // successor thread to guarantee progress.
996 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
997 // section 3.3 "Futile Wakeup Throttling" for details.
998 //
999 // Note that spinners in Enter() also set _succ non-null.
1000 // In the current implementation spinners opportunistically set
1001 // _succ so that exiting threads might avoid waking a successor.
1002 // Another less appealing alternative would be for the exiting thread
1003 // to drop the lock and then spin briefly to see if a spinner managed
1004 // to acquire the lock. If so, the exiting thread could exit
1005 // immediately without waking a successor, otherwise the exiting
1006 // thread would need to dequeue and wake a successor.
1007 // (Note that we'd need to make the post-drop spin short, but no
1008 // shorter than the worst-case round-trip cache-line migration time.
1009 // The dropped lock needs to become visible to the spinner, and then
1010 // the acquisition of the lock by the spinner must become visible to
1011 // the exiting thread).
1012 //
1013
1014 // It appears that an heir-presumptive (successor) must be made ready.
1015 // Only the current lock owner can manipulate the EntryList or
1016 // drain _cxq, so we need to reacquire the lock. If we fail
1017 // to reacquire the lock the responsibility for ensuring succession
1018 // falls to the new owner.
1019 //
1020 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
1021 return ;
1022 }
1023 TEVENT (Exit - Reacquired) ;
1024 } else {
1025 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
1026 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock
1027 OrderAccess::storeload() ;
1028 // Ratify the previously observed values.
1029 if (_cxq == NULL || _succ != NULL) {
1030 TEVENT (Inflated exit - simple egress) ;
1031 return ;
1032 }
1033
1034 // inopportune interleaving -- the exiting thread (this thread)
1035 // in the fast-exit path raced an entering thread in the slow-enter
1036 // path.
1037 // We have two choices:
1038 // A. Try to reacquire the lock.
1039 // If the CAS() fails return immediately, otherwise
1040 // we either restart/rerun the exit operation, or simply
1041 // fall-through into the code below which wakes a successor.
1042 // B. If the elements forming the EntryList|cxq are TSM
1043 // we could simply unpark() the lead thread and return
1044 // without having set _succ.
1045 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
1046 TEVENT (Inflated exit - reacquired succeeded) ;
1047 return ;
1048 }
1049 TEVENT (Inflated exit - reacquired failed) ;
1050 } else {
1051 TEVENT (Inflated exit - complex egress) ;
1052 }
1053 }
1054
1055 guarantee (_owner == THREAD, "invariant") ;
1056
1057 ObjectWaiter * w = NULL ;
1058 int QMode = Knob_QMode ;
1059
1060 if (QMode == 2 && _cxq != NULL) {
1061 // QMode == 2 : cxq has precedence over EntryList.
1062 // Try to directly wake a successor from the cxq.
1063 // If successful, the successor will need to unlink itself from cxq.
1064 w = _cxq ;
1065 assert (w != NULL, "invariant") ;
1066 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1067 ExitEpilog (Self, w) ;
1068 return ;
1069 }
1070
1071 if (QMode == 3 && _cxq != NULL) {
1072 // Aggressively drain cxq into EntryList at the first opportunity.
1073 // This policy ensure that recently-run threads live at the head of EntryList.
1074 // Drain _cxq into EntryList - bulk transfer.
1075 // First, detach _cxq.
1076 // The following loop is tantamount to: w = swap (&cxq, NULL)
1077 w = _cxq ;
1078 for (;;) {
1079 assert (w != NULL, "Invariant") ;
1080 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ;
1081 if (u == w) break ;
1082 w = u ;
1083 }
1084 assert (w != NULL , "invariant") ;
1085
1086 ObjectWaiter * q = NULL ;
1087 ObjectWaiter * p ;
1088 for (p = w ; p != NULL ; p = p->_next) {
1089 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1090 p->TState = ObjectWaiter::TS_ENTER ;
1091 p->_prev = q ;
1092 q = p ;
1093 }
1094
1095 // Append the RATs to the EntryList
1096 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1097 ObjectWaiter * Tail ;
1098 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ;
1099 if (Tail == NULL) {
1100 _EntryList = w ;
1101 } else {
1102 Tail->_next = w ;
1103 w->_prev = Tail ;
1104 }
1105
1106 // Fall thru into code that tries to wake a successor from EntryList
1107 }
1108
1109 if (QMode == 4 && _cxq != NULL) {
1110 // Aggressively drain cxq into EntryList at the first opportunity.
1111 // This policy ensure that recently-run threads live at the head of EntryList.
1112
1113 // Drain _cxq into EntryList - bulk transfer.
1114 // First, detach _cxq.
1115 // The following loop is tantamount to: w = swap (&cxq, NULL)
1116 w = _cxq ;
1117 for (;;) {
1118 assert (w != NULL, "Invariant") ;
1119 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ;
1120 if (u == w) break ;
1121 w = u ;
1122 }
1123 assert (w != NULL , "invariant") ;
1124
1125 ObjectWaiter * q = NULL ;
1126 ObjectWaiter * p ;
1127 for (p = w ; p != NULL ; p = p->_next) {
1128 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1129 p->TState = ObjectWaiter::TS_ENTER ;
1130 p->_prev = q ;
1131 q = p ;
1132 }
1133
1134 // Prepend the RATs to the EntryList
1135 if (_EntryList != NULL) {
1136 q->_next = _EntryList ;
1137 _EntryList->_prev = q ;
1138 }
1139 _EntryList = w ;
1140
1141 // Fall thru into code that tries to wake a successor from EntryList
1142 }
1143
1144 w = _EntryList ;
1145 if (w != NULL) {
1146 // I'd like to write: guarantee (w->_thread != Self).
1147 // But in practice an exiting thread may find itself on the EntryList.
1148 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and
1149 // then calls exit(). Exit release the lock by setting O._owner to NULL.
1150 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The
1151 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1152 // release the lock "O". T2 resumes immediately after the ST of null into
1153 // _owner, above. T2 notices that the EntryList is populated, so it
1154 // reacquires the lock and then finds itself on the EntryList.
1155 // Given all that, we have to tolerate the circumstance where "w" is
1156 // associated with Self.
1157 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1158 ExitEpilog (Self, w) ;
1159 return ;
1160 }
1161
1162 // If we find that both _cxq and EntryList are null then just
1163 // re-run the exit protocol from the top.
1164 w = _cxq ;
1165 if (w == NULL) continue ;
1166
1167 // Drain _cxq into EntryList - bulk transfer.
1168 // First, detach _cxq.
1169 // The following loop is tantamount to: w = swap (&cxq, NULL)
1170 for (;;) {
1171 assert (w != NULL, "Invariant") ;
1172 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ;
1173 if (u == w) break ;
1174 w = u ;
1175 }
1176 TEVENT (Inflated exit - drain cxq into EntryList) ;
1177
1178 assert (w != NULL , "invariant") ;
1179 assert (_EntryList == NULL , "invariant") ;
1180
1181 // Convert the LIFO SLL anchored by _cxq into a DLL.
1182 // The list reorganization step operates in O(LENGTH(w)) time.
1183 // It's critical that this step operate quickly as
1184 // "Self" still holds the outer-lock, restricting parallelism
1185 // and effectively lengthening the critical section.
1186 // Invariant: s chases t chases u.
1187 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1188 // we have faster access to the tail.
1189
1190 if (QMode == 1) {
1191 // QMode == 1 : drain cxq to EntryList, reversing order
1192 // We also reverse the order of the list.
1193 ObjectWaiter * s = NULL ;
1194 ObjectWaiter * t = w ;
1195 ObjectWaiter * u = NULL ;
1196 while (t != NULL) {
1197 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ;
1198 t->TState = ObjectWaiter::TS_ENTER ;
1199 u = t->_next ;
1200 t->_prev = u ;
1201 t->_next = s ;
1202 s = t;
1203 t = u ;
1204 }
1205 _EntryList = s ;
1206 assert (s != NULL, "invariant") ;
1207 } else {
1208 // QMode == 0 or QMode == 2
1209 _EntryList = w ;
1210 ObjectWaiter * q = NULL ;
1211 ObjectWaiter * p ;
1212 for (p = w ; p != NULL ; p = p->_next) {
1213 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1214 p->TState = ObjectWaiter::TS_ENTER ;
1215 p->_prev = q ;
1216 q = p ;
1217 }
1218 }
1219
1220 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1221 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1222
1223 // See if we can abdicate to a spinner instead of waking a thread.
1224 // A primary goal of the implementation is to reduce the
1225 // context-switch rate.
1226 if (_succ != NULL) continue;
1227
1228 w = _EntryList ;
1229 if (w != NULL) {
1230 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1231 ExitEpilog (Self, w) ;
1232 return ;
1233 }
1234 }
1235 }
1236
1237 // ExitSuspendEquivalent:
1238 // A faster alternate to handle_special_suspend_equivalent_condition()
1239 //
1240 // handle_special_suspend_equivalent_condition() unconditionally
1241 // acquires the SR_lock. On some platforms uncontended MutexLocker()
1242 // operations have high latency. Note that in ::enter() we call HSSEC
1243 // while holding the monitor, so we effectively lengthen the critical sections.
1244 //
1245 // There are a number of possible solutions:
1246 //
1247 // A. To ameliorate the problem we might also defer state transitions
1248 // to as late as possible -- just prior to parking.
1249 // Given that, we'd call HSSEC after having returned from park(),
1250 // but before attempting to acquire the monitor. This is only a
1251 // partial solution. It avoids calling HSSEC while holding the
1252 // monitor (good), but it still increases successor reacquisition latency --
1253 // the interval between unparking a successor and the time the successor
1254 // resumes and retries the lock. See ReenterI(), which defers state transitions.
1255 // If we use this technique we can also avoid EnterI()-exit() loop
1256 // in ::enter() where we iteratively drop the lock and then attempt
1257 // to reacquire it after suspending.
1258 //
1259 // B. In the future we might fold all the suspend bits into a
1260 // composite per-thread suspend flag and then update it with CAS().
1261 // Alternately, a Dekker-like mechanism with multiple variables
1262 // would suffice:
1263 // ST Self->_suspend_equivalent = false
1264 // MEMBAR
1265 // LD Self_>_suspend_flags
1266 //
1267
1268
1269 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) {
1270 int Mode = Knob_FastHSSEC ;
1271 if (Mode && !jSelf->is_external_suspend()) {
1272 assert (jSelf->is_suspend_equivalent(), "invariant") ;
1273 jSelf->clear_suspend_equivalent() ;
1274 if (2 == Mode) OrderAccess::storeload() ;
1275 if (!jSelf->is_external_suspend()) return false ;
1276 // We raced a suspension -- fall thru into the slow path
1277 TEVENT (ExitSuspendEquivalent - raced) ;
1278 jSelf->set_suspend_equivalent() ;
1279 }
1280 return jSelf->handle_special_suspend_equivalent_condition() ;
1281 }
1282
1283
1284 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) {
1285 assert (_owner == Self, "invariant") ;
1286
1287 // Exit protocol:
1288 // 1. ST _succ = wakee
1289 // 2. membar #loadstore|#storestore;
1290 // 2. ST _owner = NULL
1291 // 3. unpark(wakee)
1292
1293 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ;
1294 ParkEvent * Trigger = Wakee->_event ;
1295
1296 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1297 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1298 // out-of-scope (non-extant).
1299 Wakee = NULL ;
1300
1301 // Drop the lock
1302 OrderAccess::release_store_ptr (&_owner, NULL) ;
1303 OrderAccess::fence() ; // ST _owner vs LD in unpark()
1304
1305 if (SafepointSynchronize::do_call_back()) {
1306 TEVENT (unpark before SAFEPOINT) ;
1307 }
1308
1309 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1310 Trigger->unpark() ;
1311
1312 // Maintain stats and report events to JVMTI
1313 if (ObjectMonitor::_sync_Parks != NULL) {
1314 ObjectMonitor::_sync_Parks->inc() ;
1315 }
1316 }
1317
1318
1319 // -----------------------------------------------------------------------------
1320 // Class Loader deadlock handling.
1321 //
1322 // complete_exit exits a lock returning recursion count
1323 // complete_exit/reenter operate as a wait without waiting
1324 // complete_exit requires an inflated monitor
1325 // The _owner field is not always the Thread addr even with an
1326 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1327 // thread due to contention.
1328 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1329 Thread * const Self = THREAD;
1330 assert(Self->is_Java_thread(), "Must be Java thread!");
1331 JavaThread *jt = (JavaThread *)THREAD;
1332
1333 DeferredInitialize();
1334
1335 if (THREAD != _owner) {
1336 if (THREAD->is_lock_owned ((address)_owner)) {
1337 assert(_recursions == 0, "internal state error");
1338 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */
1339 _recursions = 0 ;
1340 OwnerIsThread = 1 ;
1341 }
1342 }
1343
1344 guarantee(Self == _owner, "complete_exit not owner");
1345 intptr_t save = _recursions; // record the old recursion count
1346 _recursions = 0; // set the recursion level to be 0
1347 exit (true, Self) ; // exit the monitor
1348 guarantee (_owner != Self, "invariant");
1349 return save;
1350 }
1351
1352 // reenter() enters a lock and sets recursion count
1353 // complete_exit/reenter operate as a wait without waiting
1354 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1355 Thread * const Self = THREAD;
1356 assert(Self->is_Java_thread(), "Must be Java thread!");
1357 JavaThread *jt = (JavaThread *)THREAD;
1358
1359 guarantee(_owner != Self, "reenter already owner");
1360 enter (THREAD); // enter the monitor
1361 guarantee (_recursions == 0, "reenter recursion");
1362 _recursions = recursions;
1363 return;
1364 }
1365
1366
1367 // -----------------------------------------------------------------------------
1368 // A macro is used below because there may already be a pending
1369 // exception which should not abort the execution of the routines
1370 // which use this (which is why we don't put this into check_slow and
1371 // call it with a CHECK argument).
1372
1373 #define CHECK_OWNER() \
1374 do { \
1375 if (THREAD != _owner) { \
1376 if (THREAD->is_lock_owned((address) _owner)) { \
1377 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \
1378 _recursions = 0; \
1379 OwnerIsThread = 1 ; \
1380 } else { \
1381 TEVENT (Throw IMSX) ; \
1382 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \
1383 } \
1384 } \
1385 } while (false)
1386
1387 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception.
1388 // TODO-FIXME: remove check_slow() -- it's likely dead.
1389
1390 void ObjectMonitor::check_slow(TRAPS) {
1391 TEVENT (check_slow - throw IMSX) ;
1392 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1393 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1394 }
1395
1396 static int Adjust (volatile int * adr, int dx) {
1397 int v ;
1398 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ;
1399 return v ;
1400 }
1401
1402 // helper method for posting a monitor wait event
1403 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event,
1404 jlong notifier_tid,
1405 jlong timeout,
1406 bool timedout) {
1407 event->set_klass(((oop)this->object())->klass());
1408 event->set_timeout((TYPE_ULONG)timeout);
1409 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr()));
1410 event->set_notifier((TYPE_OSTHREAD)notifier_tid);
1411 event->set_timedOut((TYPE_BOOLEAN)timedout);
1412 event->commit();
1413 }
1414
1415 // -----------------------------------------------------------------------------
1416 // Wait/Notify/NotifyAll
1417 //
1418 // Note: a subset of changes to ObjectMonitor::wait()
1419 // will need to be replicated in complete_exit above
1420 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1421 Thread * const Self = THREAD ;
1422 assert(Self->is_Java_thread(), "Must be Java thread!");
1423 JavaThread *jt = (JavaThread *)THREAD;
1424
1425 DeferredInitialize () ;
1426
1427 // Throw IMSX or IEX.
1428 CHECK_OWNER();
1429
1430 EventJavaMonitorWait event;
1431
1432 // check for a pending interrupt
1433 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1434 // post monitor waited event. Note that this is past-tense, we are done waiting.
1435 if (JvmtiExport::should_post_monitor_waited()) {
1436 // Note: 'false' parameter is passed here because the
1437 // wait was not timed out due to thread interrupt.
1438 JvmtiExport::post_monitor_waited(jt, this, false);
1439
1440 // In this short circuit of the monitor wait protocol, the
1441 // current thread never drops ownership of the monitor and
1442 // never gets added to the wait queue so the current thread
1443 // cannot be made the successor. This means that the
1444 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1445 // consume an unpark() meant for the ParkEvent associated with
1446 // this ObjectMonitor.
1447 }
1448 if (event.should_commit()) {
1449 post_monitor_wait_event(&event, 0, millis, false);
1450 }
1451 TEVENT (Wait - Throw IEX) ;
1452 THROW(vmSymbols::java_lang_InterruptedException());
1453 return ;
1454 }
1455
1456 TEVENT (Wait) ;
1457
1458 assert (Self->_Stalled == 0, "invariant") ;
1459 Self->_Stalled = intptr_t(this) ;
1460 jt->set_current_waiting_monitor(this);
1461
1462 // create a node to be put into the queue
1463 // Critically, after we reset() the event but prior to park(), we must check
1464 // for a pending interrupt.
1465 ObjectWaiter node(Self);
1466 node.TState = ObjectWaiter::TS_WAIT ;
1467 Self->_ParkEvent->reset() ;
1468 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag
1469
1470 // Enter the waiting queue, which is a circular doubly linked list in this case
1471 // but it could be a priority queue or any data structure.
1472 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only
1473 // by the the owner of the monitor *except* in the case where park()
1474 // returns because of a timeout of interrupt. Contention is exceptionally rare
1475 // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1476
1477 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ;
1478 AddWaiter (&node) ;
1479 Thread::SpinRelease (&_WaitSetLock) ;
1480
1481 if ((SyncFlags & 4) == 0) {
1482 _Responsible = NULL ;
1483 }
1484 intptr_t save = _recursions; // record the old recursion count
1485 _waiters++; // increment the number of waiters
1486 _recursions = 0; // set the recursion level to be 1
1487 exit (true, Self) ; // exit the monitor
1488 guarantee (_owner != Self, "invariant") ;
1489
1490 // The thread is on the WaitSet list - now park() it.
1491 // On MP systems it's conceivable that a brief spin before we park
1492 // could be profitable.
1493 //
1494 // TODO-FIXME: change the following logic to a loop of the form
1495 // while (!timeout && !interrupted && _notified == 0) park()
1496
1497 int ret = OS_OK ;
1498 int WasNotified = 0 ;
1499 { // State transition wrappers
1500 OSThread* osthread = Self->osthread();
1501 OSThreadWaitState osts(osthread, true);
1502 {
1503 ThreadBlockInVM tbivm(jt);
1504 // Thread is in thread_blocked state and oop access is unsafe.
1505 jt->set_suspend_equivalent();
1506
1507 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1508 // Intentionally empty
1509 } else
1510 if (node._notified == 0) {
1511 if (millis <= 0) {
1512 Self->_ParkEvent->park () ;
1513 } else {
1514 ret = Self->_ParkEvent->park (millis) ;
1515 }
1516 }
1517
1518 // were we externally suspended while we were waiting?
1519 if (ExitSuspendEquivalent (jt)) {
1520 // TODO-FIXME: add -- if succ == Self then succ = null.
1521 jt->java_suspend_self();
1522 }
1523
1524 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1525
1526
1527 // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1528 // from the WaitSet to the EntryList.
1529 // See if we need to remove Node from the WaitSet.
1530 // We use double-checked locking to avoid grabbing _WaitSetLock
1531 // if the thread is not on the wait queue.
1532 //
1533 // Note that we don't need a fence before the fetch of TState.
1534 // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1535 // written by the is thread. (perhaps the fetch might even be satisfied
1536 // by a look-aside into the processor's own store buffer, although given
1537 // the length of the code path between the prior ST and this load that's
1538 // highly unlikely). If the following LD fetches a stale TS_WAIT value
1539 // then we'll acquire the lock and then re-fetch a fresh TState value.
1540 // That is, we fail toward safety.
1541
1542 if (node.TState == ObjectWaiter::TS_WAIT) {
1543 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ;
1544 if (node.TState == ObjectWaiter::TS_WAIT) {
1545 DequeueSpecificWaiter (&node) ; // unlink from WaitSet
1546 assert(node._notified == 0, "invariant");
1547 node.TState = ObjectWaiter::TS_RUN ;
1548 }
1549 Thread::SpinRelease (&_WaitSetLock) ;
1550 }
1551
1552 // The thread is now either on off-list (TS_RUN),
1553 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1554 // The Node's TState variable is stable from the perspective of this thread.
1555 // No other threads will asynchronously modify TState.
1556 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ;
1557 OrderAccess::loadload() ;
1558 if (_succ == Self) _succ = NULL ;
1559 WasNotified = node._notified ;
1560
1561 // Reentry phase -- reacquire the monitor.
1562 // re-enter contended monitor after object.wait().
1563 // retain OBJECT_WAIT state until re-enter successfully completes
1564 // Thread state is thread_in_vm and oop access is again safe,
1565 // although the raw address of the object may have changed.
1566 // (Don't cache naked oops over safepoints, of course).
1567
1568 // post monitor waited event. Note that this is past-tense, we are done waiting.
1569 if (JvmtiExport::should_post_monitor_waited()) {
1570 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1571
1572 if (node._notified != 0 && _succ == Self) {
1573 // In this part of the monitor wait-notify-reenter protocol it
1574 // is possible (and normal) for another thread to do a fastpath
1575 // monitor enter-exit while this thread is still trying to get
1576 // to the reenter portion of the protocol.
1577 //
1578 // The ObjectMonitor was notified and the current thread is
1579 // the successor which also means that an unpark() has already
1580 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1581 // consume the unpark() that was done when the successor was
1582 // set because the same ParkEvent is shared between Java
1583 // monitors and JVM/TI RawMonitors (for now).
1584 //
1585 // We redo the unpark() to ensure forward progress, i.e., we
1586 // don't want all pending threads hanging (parked) with none
1587 // entering the unlocked monitor.
1588 node._event->unpark();
1589 }
1590 }
1591
1592 if (event.should_commit()) {
1593 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT);
1594 }
1595
1596 OrderAccess::fence() ;
1597
1598 assert (Self->_Stalled != 0, "invariant") ;
1599 Self->_Stalled = 0 ;
1600
1601 assert (_owner != Self, "invariant") ;
1602 ObjectWaiter::TStates v = node.TState ;
1603 if (v == ObjectWaiter::TS_RUN) {
1604 enter (Self) ;
1605 } else {
1606 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ;
1607 ReenterI (Self, &node) ;
1608 node.wait_reenter_end(this);
1609 }
1610
1611 // Self has reacquired the lock.
1612 // Lifecycle - the node representing Self must not appear on any queues.
1613 // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1614 // want residual elements associated with this thread left on any lists.
1615 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ;
1616 assert (_owner == Self, "invariant") ;
1617 assert (_succ != Self , "invariant") ;
1618 } // OSThreadWaitState()
1619
1620 jt->set_current_waiting_monitor(NULL);
1621
1622 guarantee (_recursions == 0, "invariant") ;
1623 _recursions = save; // restore the old recursion count
1624 _waiters--; // decrement the number of waiters
1625
1626 // Verify a few postconditions
1627 assert (_owner == Self , "invariant") ;
1628 assert (_succ != Self , "invariant") ;
1629 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
1630
1631 if (SyncFlags & 32) {
1632 OrderAccess::fence() ;
1633 }
1634
1635 // check if the notification happened
1636 if (!WasNotified) {
1637 // no, it could be timeout or Thread.interrupt() or both
1638 // check for interrupt event, otherwise it is timeout
1639 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1640 TEVENT (Wait - throw IEX from epilog) ;
1641 THROW(vmSymbols::java_lang_InterruptedException());
1642 }
1643 }
1644
1645 // NOTE: Spurious wake up will be consider as timeout.
1646 // Monitor notify has precedence over thread interrupt.
1647 }
1648
1649
1650 // Consider:
1651 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1652 // then instead of transferring a thread from the WaitSet to the EntryList
1653 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1654
1655 void ObjectMonitor::notify(TRAPS) {
1656 CHECK_OWNER();
1657 if (_WaitSet == NULL) {
1658 TEVENT (Empty-Notify) ;
1659 return ;
1660 }
1661 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1662
1663 int Policy = Knob_MoveNotifyee ;
1664
1665 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ;
1666 ObjectWaiter * iterator = DequeueWaiter() ;
1667 if (iterator != NULL) {
1668 TEVENT (Notify1 - Transfer) ;
1669 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ;
1670 guarantee (iterator->_notified == 0, "invariant") ;
1671 if (Policy != 4) {
1672 iterator->TState = ObjectWaiter::TS_ENTER ;
1673 }
1674 iterator->_notified = 1 ;
1675 Thread * Self = THREAD;
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
1696 if (Policy == 1) { // append to EntryList
1697 if (List == NULL) {
1698 iterator->_next = iterator->_prev = NULL ;
1699 _EntryList = iterator ;
1700 } else {
1701 // CONSIDER: finding the tail currently requires a linear-time walk of
1702 // the EntryList. We can make tail access constant-time by converting to
1703 // a CDLL instead of using our current DLL.
1704 ObjectWaiter * Tail ;
1705 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ;
1706 assert (Tail != NULL && Tail->_next == NULL, "invariant") ;
1707 Tail->_next = iterator ;
1708 iterator->_prev = Tail ;
1709 iterator->_next = NULL ;
1710 }
1711 } else
1712 if (Policy == 2) { // prepend to cxq
1713 // prepend to cxq
1714 if (List == NULL) {
1715 iterator->_next = iterator->_prev = NULL ;
1716 _EntryList = iterator ;
1717 } else {
1718 iterator->TState = ObjectWaiter::TS_CXQ ;
1719 for (;;) {
1720 ObjectWaiter * Front = _cxq ;
1721 iterator->_next = Front ;
1722 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) {
1723 break ;
1724 }
1725 }
1726 }
1727 } else
1728 if (Policy == 3) { // append to cxq
1729 iterator->TState = ObjectWaiter::TS_CXQ ;
1730 for (;;) {
1731 ObjectWaiter * Tail ;
1732 Tail = _cxq ;
1733 if (Tail == NULL) {
1734 iterator->_next = NULL ;
1735 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) {
1736 break ;
1737 }
1738 } else {
1739 while (Tail->_next != NULL) Tail = Tail->_next ;
1740 Tail->_next = iterator ;
1741 iterator->_prev = Tail ;
1742 iterator->_next = NULL ;
1743 break ;
1744 }
1745 }
1746 } else {
1747 ParkEvent * ev = iterator->_event ;
1748 iterator->TState = ObjectWaiter::TS_RUN ;
1749 OrderAccess::fence() ;
1750 ev->unpark() ;
1751 }
1752
1753 if (Policy < 4) {
1754 iterator->wait_reenter_begin(this);
1755 }
1756
1757 // _WaitSetLock protects the wait queue, not the EntryList. We could
1758 // move the add-to-EntryList operation, above, outside the critical section
1759 // protected by _WaitSetLock. In practice that's not useful. With the
1760 // exception of wait() timeouts and interrupts the monitor owner
1761 // is the only thread that grabs _WaitSetLock. There's almost no contention
1762 // on _WaitSetLock so it's not profitable to reduce the length of the
1763 // critical section.
1764 }
1765
1766 Thread::SpinRelease (&_WaitSetLock) ;
1767
1768 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) {
1769 ObjectMonitor::_sync_Notifications->inc() ;
1770 }
1771 }
1772
1773
1774 void ObjectMonitor::notifyAll(TRAPS) {
1775 CHECK_OWNER();
1776 ObjectWaiter* iterator;
1777 if (_WaitSet == NULL) {
1778 TEVENT (Empty-NotifyAll) ;
1779 return ;
1780 }
1781 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1782
1783 int Policy = Knob_MoveNotifyee ;
1784 int Tally = 0 ;
1785 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ;
1786
1787 for (;;) {
1788 iterator = DequeueWaiter () ;
1789 if (iterator == NULL) break ;
1790 TEVENT (NotifyAll - Transfer1) ;
1791 ++Tally ;
1792
1793 // Disposition - what might we do with iterator ?
1794 // a. add it directly to the EntryList - either tail or head.
1795 // b. push it onto the front of the _cxq.
1796 // For now we use (a).
1797
1798 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ;
1799 guarantee (iterator->_notified == 0, "invariant") ;
1800 iterator->_notified = 1 ;
1801 Thread * Self = THREAD;
1802 iterator->_notifier_tid = Self->osthread()->thread_id();
1803 if (Policy != 4) {
1804 iterator->TState = ObjectWaiter::TS_ENTER ;
1805 }
1806
1807 ObjectWaiter * List = _EntryList ;
1808 if (List != NULL) {
1809 assert (List->_prev == NULL, "invariant") ;
1810 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1811 assert (List != iterator, "invariant") ;
1812 }
1813
1814 if (Policy == 0) { // prepend to EntryList
1815 if (List == NULL) {
1816 iterator->_next = iterator->_prev = NULL ;
1817 _EntryList = iterator ;
1818 } else {
1819 List->_prev = iterator ;
1820 iterator->_next = List ;
1821 iterator->_prev = NULL ;
1822 _EntryList = iterator ;
1823 }
1824 } else
1825 if (Policy == 1) { // append to EntryList
1826 if (List == NULL) {
1827 iterator->_next = iterator->_prev = NULL ;
1828 _EntryList = iterator ;
1829 } else {
1830 // CONSIDER: finding the tail currently requires a linear-time walk of
1831 // the EntryList. We can make tail access constant-time by converting to
1832 // a CDLL instead of using our current DLL.
1833 ObjectWaiter * Tail ;
1834 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ;
1835 assert (Tail != NULL && Tail->_next == NULL, "invariant") ;
1836 Tail->_next = iterator ;
1837 iterator->_prev = Tail ;
1838 iterator->_next = NULL ;
1839 }
1840 } else
1841 if (Policy == 2) { // prepend to cxq
1842 // prepend to cxq
1843 iterator->TState = ObjectWaiter::TS_CXQ ;
1844 for (;;) {
1845 ObjectWaiter * Front = _cxq ;
1846 iterator->_next = Front ;
1847 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) {
1848 break ;
1849 }
1850 }
1851 } else
1852 if (Policy == 3) { // append to cxq
1853 iterator->TState = ObjectWaiter::TS_CXQ ;
1854 for (;;) {
1855 ObjectWaiter * Tail ;
1856 Tail = _cxq ;
1857 if (Tail == NULL) {
1858 iterator->_next = NULL ;
1859 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) {
1860 break ;
1861 }
1862 } else {
1863 while (Tail->_next != NULL) Tail = Tail->_next ;
1864 Tail->_next = iterator ;
1865 iterator->_prev = Tail ;
1866 iterator->_next = NULL ;
1867 break ;
1868 }
1869 }
1870 } else {
1871 ParkEvent * ev = iterator->_event ;
1872 iterator->TState = ObjectWaiter::TS_RUN ;
1873 OrderAccess::fence() ;
1874 ev->unpark() ;
1875 }
1876
1877 if (Policy < 4) {
1878 iterator->wait_reenter_begin(this);
1879 }
1880
1881 // _WaitSetLock protects the wait queue, not the EntryList. We could
1882 // move the add-to-EntryList operation, above, outside the critical section
1883 // protected by _WaitSetLock. In practice that's not useful. With the
1884 // exception of wait() timeouts and interrupts the monitor owner
1885 // is the only thread that grabs _WaitSetLock. There's almost no contention
1886 // on _WaitSetLock so it's not profitable to reduce the length of the
1887 // critical section.
1888 }
1889
1890 Thread::SpinRelease (&_WaitSetLock) ;
1891
1892 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) {
1893 ObjectMonitor::_sync_Notifications->inc(Tally) ;
1894 }
1895 }
1896
1897 // -----------------------------------------------------------------------------
1898 // Adaptive Spinning Support
1899 //
1900 // Adaptive spin-then-block - rational spinning
1901 //
1902 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1903 // algorithm. On high order SMP systems it would be better to start with
1904 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH,
1905 // a contending thread could enqueue itself on the cxq and then spin locally
1906 // on a thread-specific variable such as its ParkEvent._Event flag.
1907 // That's left as an exercise for the reader. Note that global spinning is
1908 // not problematic on Niagara, as the L2$ serves the interconnect and has both
1909 // low latency and massive bandwidth.
1910 //
1911 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1912 // acquisition attempts where we opt to spin -- at 100% and vary the spin count
1913 // (duration) or we can fix the count at approximately the duration of
1914 // a context switch and vary the frequency. Of course we could also
1915 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1916 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html.
1917 //
1918 // This implementation varies the duration "D", where D varies with
1919 // the success rate of recent spin attempts. (D is capped at approximately
1920 // length of a round-trip context switch). The success rate for recent
1921 // spin attempts is a good predictor of the success rate of future spin
1922 // attempts. The mechanism adapts automatically to varying critical
1923 // section length (lock modality), system load and degree of parallelism.
1924 // D is maintained per-monitor in _SpinDuration and is initialized
1925 // optimistically. Spin frequency is fixed at 100%.
1926 //
1927 // Note that _SpinDuration is volatile, but we update it without locks
1928 // or atomics. The code is designed so that _SpinDuration stays within
1929 // a reasonable range even in the presence of races. The arithmetic
1930 // operations on _SpinDuration are closed over the domain of legal values,
1931 // so at worst a race will install and older but still legal value.
1932 // At the very worst this introduces some apparent non-determinism.
1933 // We might spin when we shouldn't or vice-versa, but since the spin
1934 // count are relatively short, even in the worst case, the effect is harmless.
1935 //
1936 // Care must be taken that a low "D" value does not become an
1937 // an absorbing state. Transient spinning failures -- when spinning
1938 // is overall profitable -- should not cause the system to converge
1939 // on low "D" values. We want spinning to be stable and predictable
1940 // and fairly responsive to change and at the same time we don't want
1941 // it to oscillate, become metastable, be "too" non-deterministic,
1942 // or converge on or enter undesirable stable absorbing states.
1943 //
1944 // We implement a feedback-based control system -- using past behavior
1945 // to predict future behavior. We face two issues: (a) if the
1946 // input signal is random then the spin predictor won't provide optimal
1947 // results, and (b) if the signal frequency is too high then the control
1948 // system, which has some natural response lag, will "chase" the signal.
1949 // (b) can arise from multimodal lock hold times. Transient preemption
1950 // can also result in apparent bimodal lock hold times.
1951 // Although sub-optimal, neither condition is particularly harmful, as
1952 // in the worst-case we'll spin when we shouldn't or vice-versa.
1953 // The maximum spin duration is rather short so the failure modes aren't bad.
1954 // To be conservative, I've tuned the gain in system to bias toward
1955 // _not spinning. Relatedly, the system can sometimes enter a mode where it
1956 // "rings" or oscillates between spinning and not spinning. This happens
1957 // when spinning is just on the cusp of profitability, however, so the
1958 // situation is not dire. The state is benign -- there's no need to add
1959 // hysteresis control to damp the transition rate between spinning and
1960 // not spinning.
1961 //
1962
1963 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ;
1964 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ;
1965
1966 // Spinning: Fixed frequency (100%), vary duration
1967
1968
1969 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) {
1970
1971 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning.
1972 int ctr = Knob_FixedSpin ;
1973 if (ctr != 0) {
1974 while (--ctr >= 0) {
1975 if (TryLock (Self) > 0) return 1 ;
1976 SpinPause () ;
1977 }
1978 return 0 ;
1979 }
1980
1981 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) {
1982 if (TryLock(Self) > 0) {
1983 // Increase _SpinDuration ...
1984 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1985 // Raising _SpurDuration to the poverty line is key.
1986 int x = _SpinDuration ;
1987 if (x < Knob_SpinLimit) {
1988 if (x < Knob_Poverty) x = Knob_Poverty ;
1989 _SpinDuration = x + Knob_BonusB ;
1990 }
1991 return 1 ;
1992 }
1993 SpinPause () ;
1994 }
1995
1996 // Admission control - verify preconditions for spinning
1997 //
1998 // We always spin a little bit, just to prevent _SpinDuration == 0 from
1999 // becoming an absorbing state. Put another way, we spin briefly to
2000 // sample, just in case the system load, parallelism, contention, or lock
2001 // modality changed.
2002 //
2003 // Consider the following alternative:
2004 // Periodically set _SpinDuration = _SpinLimit and try a long/full
2005 // spin attempt. "Periodically" might mean after a tally of
2006 // the # of failed spin attempts (or iterations) reaches some threshold.
2007 // This takes us into the realm of 1-out-of-N spinning, where we
2008 // hold the duration constant but vary the frequency.
2009
2010 ctr = _SpinDuration ;
2011 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ;
2012 if (ctr <= 0) return 0 ;
2013
2014 if (Knob_SuccRestrict && _succ != NULL) return 0 ;
2015 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
2016 TEVENT (Spin abort - notrunnable [TOP]);
2017 return 0 ;
2018 }
2019
2020 int MaxSpin = Knob_MaxSpinners ;
2021 if (MaxSpin >= 0) {
2022 if (_Spinner > MaxSpin) {
2023 TEVENT (Spin abort -- too many spinners) ;
2024 return 0 ;
2025 }
2026 // Slightly racy, but benign ...
2027 Adjust (&_Spinner, 1) ;
2028 }
2029
2030 // We're good to spin ... spin ingress.
2031 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
2032 // when preparing to LD...CAS _owner, etc and the CAS is likely
2033 // to succeed.
2034 int hits = 0 ;
2035 int msk = 0 ;
2036 int caspty = Knob_CASPenalty ;
2037 int oxpty = Knob_OXPenalty ;
2038 int sss = Knob_SpinSetSucc ;
2039 if (sss && _succ == NULL ) _succ = Self ;
2040 Thread * prv = NULL ;
2041
2042 // There are three ways to exit the following loop:
2043 // 1. A successful spin where this thread has acquired the lock.
2044 // 2. Spin failure with prejudice
2045 // 3. Spin failure without prejudice
2046
2047 while (--ctr >= 0) {
2048
2049 // Periodic polling -- Check for pending GC
2050 // Threads may spin while they're unsafe.
2051 // We don't want spinning threads to delay the JVM from reaching
2052 // a stop-the-world safepoint or to steal cycles from GC.
2053 // If we detect a pending safepoint we abort in order that
2054 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
2055 // this thread, if safe, doesn't steal cycles from GC.
2056 // This is in keeping with the "no loitering in runtime" rule.
2057 // We periodically check to see if there's a safepoint pending.
2058 if ((ctr & 0xFF) == 0) {
2059 if (SafepointSynchronize::do_call_back()) {
2060 TEVENT (Spin: safepoint) ;
2061 goto Abort ; // abrupt spin egress
2062 }
2063 if (Knob_UsePause & 1) SpinPause () ;
2064
2065 int (*scb)(intptr_t,int) = SpinCallbackFunction ;
2066 if (hits > 50 && scb != NULL) {
2067 int abend = (*scb)(SpinCallbackArgument, 0) ;
2068 }
2069 }
2070
2071 if (Knob_UsePause & 2) SpinPause() ;
2072
2073 // Exponential back-off ... Stay off the bus to reduce coherency traffic.
2074 // This is useful on classic SMP systems, but is of less utility on
2075 // N1-style CMT platforms.
2076 //
2077 // Trade-off: lock acquisition latency vs coherency bandwidth.
2078 // Lock hold times are typically short. A histogram
2079 // of successful spin attempts shows that we usually acquire
2080 // the lock early in the spin. That suggests we want to
2081 // sample _owner frequently in the early phase of the spin,
2082 // but then back-off and sample less frequently as the spin
2083 // progresses. The back-off makes a good citizen on SMP big
2084 // SMP systems. Oversampling _owner can consume excessive
2085 // coherency bandwidth. Relatedly, if we _oversample _owner we
2086 // can inadvertently interfere with the the ST m->owner=null.
2087 // executed by the lock owner.
2088 if (ctr & msk) continue ;
2089 ++hits ;
2090 if ((hits & 0xF) == 0) {
2091 // The 0xF, above, corresponds to the exponent.
2092 // Consider: (msk+1)|msk
2093 msk = ((msk << 2)|3) & BackOffMask ;
2094 }
2095
2096 // Probe _owner with TATAS
2097 // If this thread observes the monitor transition or flicker
2098 // from locked to unlocked to locked, then the odds that this
2099 // thread will acquire the lock in this spin attempt go down
2100 // considerably. The same argument applies if the CAS fails
2101 // or if we observe _owner change from one non-null value to
2102 // another non-null value. In such cases we might abort
2103 // the spin without prejudice or apply a "penalty" to the
2104 // spin count-down variable "ctr", reducing it by 100, say.
2105
2106 Thread * ox = (Thread *) _owner ;
2107 if (ox == NULL) {
2108 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ;
2109 if (ox == NULL) {
2110 // The CAS succeeded -- this thread acquired ownership
2111 // Take care of some bookkeeping to exit spin state.
2112 if (sss && _succ == Self) {
2113 _succ = NULL ;
2114 }
2115 if (MaxSpin > 0) Adjust (&_Spinner, -1) ;
2116
2117 // Increase _SpinDuration :
2118 // The spin was successful (profitable) so we tend toward
2119 // longer spin attempts in the future.
2120 // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
2121 // If we acquired the lock early in the spin cycle it
2122 // makes sense to increase _SpinDuration proportionally.
2123 // Note that we don't clamp SpinDuration precisely at SpinLimit.
2124 int x = _SpinDuration ;
2125 if (x < Knob_SpinLimit) {
2126 if (x < Knob_Poverty) x = Knob_Poverty ;
2127 _SpinDuration = x + Knob_Bonus ;
2128 }
2129 return 1 ;
2130 }
2131
2132 // The CAS failed ... we can take any of the following actions:
2133 // * penalize: ctr -= Knob_CASPenalty
2134 // * exit spin with prejudice -- goto Abort;
2135 // * exit spin without prejudice.
2136 // * Since CAS is high-latency, retry again immediately.
2137 prv = ox ;
2138 TEVENT (Spin: cas failed) ;
2139 if (caspty == -2) break ;
2140 if (caspty == -1) goto Abort ;
2141 ctr -= caspty ;
2142 continue ;
2143 }
2144
2145 // Did lock ownership change hands ?
2146 if (ox != prv && prv != NULL ) {
2147 TEVENT (spin: Owner changed)
2148 if (oxpty == -2) break ;
2149 if (oxpty == -1) goto Abort ;
2150 ctr -= oxpty ;
2151 }
2152 prv = ox ;
2153
2154 // Abort the spin if the owner is not executing.
2155 // The owner must be executing in order to drop the lock.
2156 // Spinning while the owner is OFFPROC is idiocy.
2157 // Consider: ctr -= RunnablePenalty ;
2158 if (Knob_OState && NotRunnable (Self, ox)) {
2159 TEVENT (Spin abort - notrunnable);
2160 goto Abort ;
2161 }
2162 if (sss && _succ == NULL ) _succ = Self ;
2163 }
2164
2165 // Spin failed with prejudice -- reduce _SpinDuration.
2166 // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2167 // AIMD is globally stable.
2168 TEVENT (Spin failure) ;
2169 {
2170 int x = _SpinDuration ;
2171 if (x > 0) {
2172 // Consider an AIMD scheme like: x -= (x >> 3) + 100
2173 // This is globally sample and tends to damp the response.
2174 x -= Knob_Penalty ;
2175 if (x < 0) x = 0 ;
2176 _SpinDuration = x ;
2177 }
2178 }
2179
2180 Abort:
2181 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ;
2182 if (sss && _succ == Self) {
2183 _succ = NULL ;
2184 // Invariant: after setting succ=null a contending thread
2185 // must recheck-retry _owner before parking. This usually happens
2186 // in the normal usage of TrySpin(), but it's safest
2187 // to make TrySpin() as foolproof as possible.
2188 OrderAccess::fence() ;
2189 if (TryLock(Self) > 0) return 1 ;
2190 }
2191 return 0 ;
2192 }
2193
2194 // NotRunnable() -- informed spinning
2195 //
2196 // Don't bother spinning if the owner is not eligible to drop the lock.
2197 // Peek at the owner's schedctl.sc_state and Thread._thread_values and
2198 // spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2199 // The thread must be runnable in order to drop the lock in timely fashion.
2200 // If the _owner is not runnable then spinning will not likely be
2201 // successful (profitable).
2202 //
2203 // Beware -- the thread referenced by _owner could have died
2204 // so a simply fetch from _owner->_thread_state might trap.
2205 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2206 // Because of the lifecycle issues the schedctl and _thread_state values
2207 // observed by NotRunnable() might be garbage. NotRunnable must
2208 // tolerate this and consider the observed _thread_state value
2209 // as advisory.
2210 //
2211 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2212 // a thread pointer. We differentiate the two cases with OwnerIsThread.
2213 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2214 // with the LSB of _owner. Another option would be to probablistically probe
2215 // the putative _owner->TypeTag value.
2216 //
2217 // Checking _thread_state isn't perfect. Even if the thread is
2218 // in_java it might be blocked on a page-fault or have been preempted
2219 // and sitting on a ready/dispatch queue. _thread state in conjunction
2220 // with schedctl.sc_state gives us a good picture of what the
2221 // thread is doing, however.
2222 //
2223 // TODO: check schedctl.sc_state.
2224 // We'll need to use SafeFetch32() to read from the schedctl block.
2225 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/
2226 //
2227 // The return value from NotRunnable() is *advisory* -- the
2228 // result is based on sampling and is not necessarily coherent.
2229 // The caller must tolerate false-negative and false-positive errors.
2230 // Spinning, in general, is probabilistic anyway.
2231
2232
2233 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) {
2234 // Check either OwnerIsThread or ox->TypeTag == 2BAD.
2235 if (!OwnerIsThread) return 0 ;
2236
2237 if (ox == NULL) return 0 ;
2238
2239 // Avoid transitive spinning ...
2240 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L.
2241 // Immediately after T1 acquires L it's possible that T2, also
2242 // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2243 // This occurs transiently after T1 acquired L but before
2244 // T1 managed to clear T1.Stalled. T2 does not need to abort
2245 // its spin in this circumstance.
2246 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ;
2247
2248 if (BlockedOn == 1) return 1 ;
2249 if (BlockedOn != 0) {
2250 return BlockedOn != intptr_t(this) && _owner == ox ;
2251 }
2252
2253 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ;
2254 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ;
2255 // consider also: jst != _thread_in_Java -- but that's overspecific.
2256 return jst == _thread_blocked || jst == _thread_in_native ;
2257 }
2258
2259
2260 // -----------------------------------------------------------------------------
2261 // WaitSet management ...
2262
2263 ObjectWaiter::ObjectWaiter(Thread* thread) {
2264 _next = NULL;
2265 _prev = NULL;
2266 _notified = 0;
2267 TState = TS_RUN ;
2268 _thread = thread;
2269 _event = thread->_ParkEvent ;
2270 _active = false;
2271 assert (_event != NULL, "invariant") ;
2272 }
2273
2274 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) {
2275 JavaThread *jt = (JavaThread *)this->_thread;
2276 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2277 }
2278
2279 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) {
2280 JavaThread *jt = (JavaThread *)this->_thread;
2281 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2282 }
2283
2284 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2285 assert(node != NULL, "should not dequeue NULL node");
2286 assert(node->_prev == NULL, "node already in list");
2287 assert(node->_next == NULL, "node already in list");
2288 // put node at end of queue (circular doubly linked list)
2289 if (_WaitSet == NULL) {
2290 _WaitSet = node;
2291 node->_prev = node;
2292 node->_next = node;
2293 } else {
2294 ObjectWaiter* head = _WaitSet ;
2295 ObjectWaiter* tail = head->_prev;
2296 assert(tail->_next == head, "invariant check");
2297 tail->_next = node;
2298 head->_prev = node;
2299 node->_next = head;
2300 node->_prev = tail;
2301 }
2302 }
2303
2304 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2305 // dequeue the very first waiter
2306 ObjectWaiter* waiter = _WaitSet;
2307 if (waiter) {
2308 DequeueSpecificWaiter(waiter);
2309 }
2310 return waiter;
2311 }
2312
2313 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2314 assert(node != NULL, "should not dequeue NULL node");
2315 assert(node->_prev != NULL, "node already removed from list");
2316 assert(node->_next != NULL, "node already removed from list");
2317 // when the waiter has woken up because of interrupt,
2318 // timeout or other spurious wake-up, dequeue the
2319 // waiter from waiting list
2320 ObjectWaiter* next = node->_next;
2321 if (next == node) {
2322 assert(node->_prev == node, "invariant check");
2323 _WaitSet = NULL;
2324 } else {
2325 ObjectWaiter* prev = node->_prev;
2326 assert(prev->_next == node, "invariant check");
2327 assert(next->_prev == node, "invariant check");
2328 next->_prev = prev;
2329 prev->_next = next;
2330 if (_WaitSet == node) {
2331 _WaitSet = next;
2332 }
2333 }
2334 node->_next = NULL;
2335 node->_prev = NULL;
2336 }
2337
2338 // -----------------------------------------------------------------------------
2339 // PerfData support
2340 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ;
2341 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ;
2342 PerfCounter * ObjectMonitor::_sync_Parks = NULL ;
2343 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ;
2344 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ;
2345 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ;
2346 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ;
2347 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ;
2348 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ;
2349 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ;
2350 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ;
2351 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ;
2352 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ;
2353 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ;
2354 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ;
2355 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ;
2356 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ;
2357 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ;
2358
2359 // One-shot global initialization for the sync subsystem.
2360 // We could also defer initialization and initialize on-demand
2361 // the first time we call inflate(). Initialization would
2362 // be protected - like so many things - by the MonitorCache_lock.
2363
2364 void ObjectMonitor::Initialize () {
2365 static int InitializationCompleted = 0 ;
2366 assert (InitializationCompleted == 0, "invariant") ;
2367 InitializationCompleted = 1 ;
2368 if (UsePerfData) {
2369 EXCEPTION_MARK ;
2370 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); }
2371 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); }
2372 NEWPERFCOUNTER(_sync_Inflations) ;
2373 NEWPERFCOUNTER(_sync_Deflations) ;
2374 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ;
2375 NEWPERFCOUNTER(_sync_FutileWakeups) ;
2376 NEWPERFCOUNTER(_sync_Parks) ;
2377 NEWPERFCOUNTER(_sync_EmptyNotifications) ;
2378 NEWPERFCOUNTER(_sync_Notifications) ;
2379 NEWPERFCOUNTER(_sync_SlowEnter) ;
2380 NEWPERFCOUNTER(_sync_SlowExit) ;
2381 NEWPERFCOUNTER(_sync_SlowNotify) ;
2382 NEWPERFCOUNTER(_sync_SlowNotifyAll) ;
2383 NEWPERFCOUNTER(_sync_FailedSpins) ;
2384 NEWPERFCOUNTER(_sync_SuccessfulSpins) ;
2385 NEWPERFCOUNTER(_sync_PrivateA) ;
2386 NEWPERFCOUNTER(_sync_PrivateB) ;
2387 NEWPERFCOUNTER(_sync_MonInCirculation) ;
2388 NEWPERFCOUNTER(_sync_MonScavenged) ;
2389 NEWPERFVARIABLE(_sync_MonExtant) ;
2390 #undef NEWPERFCOUNTER
2391 }
2392 }
2393
2394
2395 // Compile-time asserts
2396 // When possible, it's better to catch errors deterministically at
2397 // compile-time than at runtime. The down-side to using compile-time
2398 // asserts is that error message -- often something about negative array
2399 // indices -- is opaque.
2400
2401 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); }
2402
2403 void ObjectMonitor::ctAsserts() {
2404 CTASSERT(offset_of (ObjectMonitor, _header) == 0);
2405 }
2406
2407
2408 static char * kvGet (char * kvList, const char * Key) {
2409 if (kvList == NULL) return NULL ;
2410 size_t n = strlen (Key) ;
2411 char * Search ;
2412 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) {
2413 if (strncmp (Search, Key, n) == 0) {
2414 if (Search[n] == '=') return Search + n + 1 ;
2415 if (Search[n] == 0) return (char *) "1" ;
2416 }
2417 }
2418 return NULL ;
2419 }
2420
2421 static int kvGetInt (char * kvList, const char * Key, int Default) {
2422 char * v = kvGet (kvList, Key) ;
2423 int rslt = v ? ::strtol (v, NULL, 0) : Default ;
2424 if (Knob_ReportSettings && v != NULL) {
2425 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ;
2426 ::fflush (stdout) ;
2427 }
2428 return rslt ;
2429 }
2430
2431 void ObjectMonitor::DeferredInitialize () {
2432 if (InitDone > 0) return ;
2433 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2434 while (InitDone != 1) ;
2435 return ;
2436 }
2437
2438 // One-shot global initialization ...
2439 // The initialization is idempotent, so we don't need locks.
2440 // In the future consider doing this via os::init_2().
2441 // SyncKnobs consist of <Key>=<Value> pairs in the style
2442 // of environment variables. Start by converting ':' to NUL.
2443
2444 if (SyncKnobs == NULL) SyncKnobs = "" ;
2445
2446 size_t sz = strlen (SyncKnobs) ;
2447 char * knobs = (char *) malloc (sz + 2) ;
2448 if (knobs == NULL) {
2449 vm_exit_out_of_memory (sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs") ;
2450 guarantee (0, "invariant") ;
2451 }
2452 strcpy (knobs, SyncKnobs) ;
2453 knobs[sz+1] = 0 ;
2454 for (char * p = knobs ; *p ; p++) {
2455 if (*p == ':') *p = 0 ;
2456 }
2457
2458 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); }
2459 SETKNOB(ReportSettings) ;
2460 SETKNOB(Verbose) ;
2461 SETKNOB(FixedSpin) ;
2462 SETKNOB(SpinLimit) ;
2463 SETKNOB(SpinBase) ;
2464 SETKNOB(SpinBackOff);
2465 SETKNOB(CASPenalty) ;
2466 SETKNOB(OXPenalty) ;
2467 SETKNOB(LogSpins) ;
2468 SETKNOB(SpinSetSucc) ;
2469 SETKNOB(SuccEnabled) ;
2470 SETKNOB(SuccRestrict) ;
2471 SETKNOB(Penalty) ;
2472 SETKNOB(Bonus) ;
2473 SETKNOB(BonusB) ;
2474 SETKNOB(Poverty) ;
2475 SETKNOB(SpinAfterFutile) ;
2476 SETKNOB(UsePause) ;
2477 SETKNOB(SpinEarly) ;
2478 SETKNOB(OState) ;
2479 SETKNOB(MaxSpinners) ;
2480 SETKNOB(PreSpin) ;
2481 SETKNOB(ExitPolicy) ;
2482 SETKNOB(QMode);
2483 SETKNOB(ResetEvent) ;
2484 SETKNOB(MoveNotifyee) ;
2485 SETKNOB(FastHSSEC) ;
2486 #undef SETKNOB
2487
2488 if (os::is_MP()) {
2489 BackOffMask = (1 << Knob_SpinBackOff) - 1 ;
2490 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ;
2491 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1)
2492 } else {
2493 Knob_SpinLimit = 0 ;
2494 Knob_SpinBase = 0 ;
2495 Knob_PreSpin = 0 ;
2496 Knob_FixedSpin = -1 ;
2497 }
2498
2499 if (Knob_LogSpins == 0) {
2500 ObjectMonitor::_sync_FailedSpins = NULL ;
2501 }
2502
2503 free (knobs) ;
2504 OrderAccess::fence() ;
2505 InitDone = 1 ;
2506 }
2507
2508 #ifndef PRODUCT
2509 void ObjectMonitor::verify() {
2510 }
2511
2512 void ObjectMonitor::print() {
2513 }
2514 #endif