Tue, 08 Aug 2017 15:57:29 +0800
merge
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 */
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/handles.inline.hpp"
31 #include "runtime/interfaceSupport.hpp"
32 #include "runtime/mutexLocker.hpp"
33 #include "runtime/objectMonitor.hpp"
34 #include "runtime/objectMonitor.inline.hpp"
35 #include "runtime/osThread.hpp"
36 #include "runtime/stubRoutines.hpp"
37 #include "runtime/thread.inline.hpp"
38 #include "services/threadService.hpp"
39 #include "trace/tracing.hpp"
40 #include "trace/traceMacros.hpp"
41 #include "utilities/dtrace.hpp"
42 #include "utilities/macros.hpp"
43 #include "utilities/preserveException.hpp"
44 #ifdef TARGET_OS_FAMILY_linux
45 # include "os_linux.inline.hpp"
46 #endif
47 #ifdef TARGET_OS_FAMILY_solaris
48 # include "os_solaris.inline.hpp"
49 #endif
50 #ifdef TARGET_OS_FAMILY_windows
51 # include "os_windows.inline.hpp"
52 #endif
53 #ifdef TARGET_OS_FAMILY_bsd
54 # include "os_bsd.inline.hpp"
55 #endif
57 #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64)
58 // Need to inhibit inlining for older versions of GCC to avoid build-time failures
59 #define ATTR __attribute__((noinline))
60 #else
61 #define ATTR
62 #endif
65 #ifdef DTRACE_ENABLED
67 // Only bother with this argument setup if dtrace is available
68 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
71 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
72 char* bytes = NULL; \
73 int len = 0; \
74 jlong jtid = SharedRuntime::get_java_tid(thread); \
75 Symbol* klassname = ((oop)obj)->klass()->name(); \
76 if (klassname != NULL) { \
77 bytes = (char*)klassname->bytes(); \
78 len = klassname->utf8_length(); \
79 }
81 #ifndef USDT2
83 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify,
84 jlong, uintptr_t, char*, int);
85 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll,
86 jlong, uintptr_t, char*, int);
87 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter,
88 jlong, uintptr_t, char*, int);
89 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered,
90 jlong, uintptr_t, char*, int);
91 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit,
92 jlong, uintptr_t, char*, int);
94 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
95 { \
96 if (DTraceMonitorProbes) { \
97 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
98 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \
99 (monitor), bytes, len, (millis)); \
100 } \
101 }
103 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
104 { \
105 if (DTraceMonitorProbes) { \
106 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
107 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \
108 (uintptr_t)(monitor), bytes, len); \
109 } \
110 }
112 #else /* USDT2 */
114 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
115 { \
116 if (DTraceMonitorProbes) { \
117 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
118 HOTSPOT_MONITOR_WAIT(jtid, \
119 (monitor), bytes, len, (millis)); \
120 } \
121 }
123 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
124 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
125 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
126 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
127 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
129 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
130 { \
131 if (DTraceMonitorProbes) { \
132 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
133 HOTSPOT_MONITOR_##probe(jtid, \
134 (uintptr_t)(monitor), bytes, len); \
135 } \
136 }
138 #endif /* USDT2 */
139 #else // ndef DTRACE_ENABLED
141 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
142 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
144 #endif // ndef DTRACE_ENABLED
146 // Tunables ...
147 // The knob* variables are effectively final. Once set they should
148 // never be modified hence. Consider using __read_mostly with GCC.
150 int ObjectMonitor::Knob_Verbose = 0 ;
151 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool -
152 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins
153 static int Knob_HandOff = 0 ;
154 static int Knob_ReportSettings = 0 ;
156 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin
157 static int Knob_SpinBackOff = 0 ; // spin-loop backoff
158 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS
159 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change
160 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field
161 static int Knob_SpinEarly = 1 ;
162 static int Knob_SuccEnabled = 1 ; // futile wake throttling
163 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one
164 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs
165 static int Knob_Bonus = 100 ; // spin success bonus
166 static int Knob_BonusB = 100 ; // spin success bonus
167 static int Knob_Penalty = 200 ; // spin failure penalty
168 static int Knob_Poverty = 1000 ;
169 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park()
170 static int Knob_FixedSpin = 0 ;
171 static int Knob_OState = 3 ; // Spinner checks thread state of _owner
172 static int Knob_UsePause = 1 ;
173 static int Knob_ExitPolicy = 0 ;
174 static int Knob_PreSpin = 10 ; // 20-100 likely better
175 static int Knob_ResetEvent = 0 ;
176 static int BackOffMask = 0 ;
178 static int Knob_FastHSSEC = 0 ;
179 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee
180 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline
181 static volatile int InitDone = 0 ;
183 #define TrySpin TrySpin_VaryDuration
185 // -----------------------------------------------------------------------------
186 // Theory of operations -- Monitors lists, thread residency, etc:
187 //
188 // * A thread acquires ownership of a monitor by successfully
189 // CAS()ing the _owner field from null to non-null.
190 //
191 // * Invariant: A thread appears on at most one monitor list --
192 // cxq, EntryList or WaitSet -- at any one time.
193 //
194 // * Contending threads "push" themselves onto the cxq with CAS
195 // and then spin/park.
196 //
197 // * After a contending thread eventually acquires the lock it must
198 // dequeue itself from either the EntryList or the cxq.
199 //
200 // * The exiting thread identifies and unparks an "heir presumptive"
201 // tentative successor thread on the EntryList. Critically, the
202 // exiting thread doesn't unlink the successor thread from the EntryList.
203 // After having been unparked, the wakee will recontend for ownership of
204 // the monitor. The successor (wakee) will either acquire the lock or
205 // re-park itself.
206 //
207 // Succession is provided for by a policy of competitive handoff.
208 // The exiting thread does _not_ grant or pass ownership to the
209 // successor thread. (This is also referred to as "handoff" succession").
210 // Instead the exiting thread releases ownership and possibly wakes
211 // a successor, so the successor can (re)compete for ownership of the lock.
212 // If the EntryList is empty but the cxq is populated the exiting
213 // thread will drain the cxq into the EntryList. It does so by
214 // by detaching the cxq (installing null with CAS) and folding
215 // the threads from the cxq into the EntryList. The EntryList is
216 // doubly linked, while the cxq is singly linked because of the
217 // CAS-based "push" used to enqueue recently arrived threads (RATs).
218 //
219 // * Concurrency invariants:
220 //
221 // -- only the monitor owner may access or mutate the EntryList.
222 // The mutex property of the monitor itself protects the EntryList
223 // from concurrent interference.
224 // -- Only the monitor owner may detach the cxq.
225 //
226 // * The monitor entry list operations avoid locks, but strictly speaking
227 // they're not lock-free. Enter is lock-free, exit is not.
228 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html
229 //
230 // * The cxq can have multiple concurrent "pushers" but only one concurrent
231 // detaching thread. This mechanism is immune from the ABA corruption.
232 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
233 //
234 // * Taken together, the cxq and the EntryList constitute or form a
235 // single logical queue of threads stalled trying to acquire the lock.
236 // We use two distinct lists to improve the odds of a constant-time
237 // dequeue operation after acquisition (in the ::enter() epilog) and
238 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm).
239 // A key desideratum is to minimize queue & monitor metadata manipulation
240 // that occurs while holding the monitor lock -- that is, we want to
241 // minimize monitor lock holds times. Note that even a small amount of
242 // fixed spinning will greatly reduce the # of enqueue-dequeue operations
243 // on EntryList|cxq. That is, spinning relieves contention on the "inner"
244 // locks and monitor metadata.
245 //
246 // Cxq points to the the set of Recently Arrived Threads attempting entry.
247 // Because we push threads onto _cxq with CAS, the RATs must take the form of
248 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when
249 // the unlocking thread notices that EntryList is null but _cxq is != null.
250 //
251 // The EntryList is ordered by the prevailing queue discipline and
252 // can be organized in any convenient fashion, such as a doubly-linked list or
253 // a circular doubly-linked list. Critically, we want insert and delete operations
254 // to operate in constant-time. If we need a priority queue then something akin
255 // to Solaris' sleepq would work nicely. Viz.,
256 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
257 // Queue discipline is enforced at ::exit() time, when the unlocking thread
258 // drains the cxq into the EntryList, and orders or reorders the threads on the
259 // EntryList accordingly.
260 //
261 // Barring "lock barging", this mechanism provides fair cyclic ordering,
262 // somewhat similar to an elevator-scan.
263 //
264 // * The monitor synchronization subsystem avoids the use of native
265 // synchronization primitives except for the narrow platform-specific
266 // park-unpark abstraction. See the comments in os_solaris.cpp regarding
267 // the semantics of park-unpark. Put another way, this monitor implementation
268 // depends only on atomic operations and park-unpark. The monitor subsystem
269 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
270 // underlying OS manages the READY<->RUN transitions.
271 //
272 // * Waiting threads reside on the WaitSet list -- wait() puts
273 // the caller onto the WaitSet.
274 //
275 // * notify() or notifyAll() simply transfers threads from the WaitSet to
276 // either the EntryList or cxq. Subsequent exit() operations will
277 // unpark the notifyee. Unparking a notifee in notify() is inefficient -
278 // it's likely the notifyee would simply impale itself on the lock held
279 // by the notifier.
280 //
281 // * An interesting alternative is to encode cxq as (List,LockByte) where
282 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary
283 // variable, like _recursions, in the scheme. The threads or Events that form
284 // the list would have to be aligned in 256-byte addresses. A thread would
285 // try to acquire the lock or enqueue itself with CAS, but exiting threads
286 // could use a 1-0 protocol and simply STB to set the LockByte to 0.
287 // Note that is is *not* word-tearing, but it does presume that full-word
288 // CAS operations are coherent with intermix with STB operations. That's true
289 // on most common processors.
290 //
291 // * See also http://blogs.sun.com/dave
294 // -----------------------------------------------------------------------------
295 // Enter support
297 bool ObjectMonitor::try_enter(Thread* THREAD) {
298 if (THREAD != _owner) {
299 if (THREAD->is_lock_owned ((address)_owner)) {
300 assert(_recursions == 0, "internal state error");
301 _owner = THREAD ;
302 _recursions = 1 ;
303 OwnerIsThread = 1 ;
304 return true;
305 }
306 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
307 return false;
308 }
309 return true;
310 } else {
311 _recursions++;
312 return true;
313 }
314 }
316 void ATTR ObjectMonitor::enter(TRAPS) {
317 // The following code is ordered to check the most common cases first
318 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
319 Thread * const Self = THREAD ;
320 void * cur ;
322 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ;
323 if (cur == NULL) {
324 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0.
325 assert (_recursions == 0 , "invariant") ;
326 assert (_owner == Self, "invariant") ;
327 // CONSIDER: set or assert OwnerIsThread == 1
328 return ;
329 }
331 if (cur == Self) {
332 // TODO-FIXME: check for integer overflow! BUGID 6557169.
333 _recursions ++ ;
334 return ;
335 }
337 if (Self->is_lock_owned ((address)cur)) {
338 assert (_recursions == 0, "internal state error");
339 _recursions = 1 ;
340 // Commute owner from a thread-specific on-stack BasicLockObject address to
341 // a full-fledged "Thread *".
342 _owner = Self ;
343 OwnerIsThread = 1 ;
344 return ;
345 }
347 // We've encountered genuine contention.
348 assert (Self->_Stalled == 0, "invariant") ;
349 Self->_Stalled = intptr_t(this) ;
351 // Try one round of spinning *before* enqueueing Self
352 // and before going through the awkward and expensive state
353 // transitions. The following spin is strictly optional ...
354 // Note that if we acquire the monitor from an initial spin
355 // we forgo posting JVMTI events and firing DTRACE probes.
356 if (Knob_SpinEarly && TrySpin (Self) > 0) {
357 assert (_owner == Self , "invariant") ;
358 assert (_recursions == 0 , "invariant") ;
359 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
360 Self->_Stalled = 0 ;
361 return ;
362 }
364 assert (_owner != Self , "invariant") ;
365 assert (_succ != Self , "invariant") ;
366 assert (Self->is_Java_thread() , "invariant") ;
367 JavaThread * jt = (JavaThread *) Self ;
368 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ;
369 assert (jt->thread_state() != _thread_blocked , "invariant") ;
370 assert (this->object() != NULL , "invariant") ;
371 assert (_count >= 0, "invariant") ;
373 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy().
374 // Ensure the object-monitor relationship remains stable while there's contention.
375 Atomic::inc_ptr(&_count);
377 EventJavaMonitorEnter event;
379 { // Change java thread status to indicate blocked on monitor enter.
380 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
382 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
383 if (JvmtiExport::should_post_monitor_contended_enter()) {
384 JvmtiExport::post_monitor_contended_enter(jt, this);
386 // The current thread does not yet own the monitor and does not
387 // yet appear on any queues that would get it made the successor.
388 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
389 // handler cannot accidentally consume an unpark() meant for the
390 // ParkEvent associated with this ObjectMonitor.
391 }
393 OSThreadContendState osts(Self->osthread());
394 ThreadBlockInVM tbivm(jt);
396 Self->set_current_pending_monitor(this);
398 // TODO-FIXME: change the following for(;;) loop to straight-line code.
399 for (;;) {
400 jt->set_suspend_equivalent();
401 // cleared by handle_special_suspend_equivalent_condition()
402 // or java_suspend_self()
404 EnterI (THREAD) ;
406 if (!ExitSuspendEquivalent(jt)) break ;
408 //
409 // We have acquired the contended monitor, but while we were
410 // waiting another thread suspended us. We don't want to enter
411 // the monitor while suspended because that would surprise the
412 // thread that suspended us.
413 //
414 _recursions = 0 ;
415 _succ = NULL ;
416 exit (false, Self) ;
418 jt->java_suspend_self();
419 }
420 Self->set_current_pending_monitor(NULL);
422 // We cleared the pending monitor info since we've just gotten past
423 // the enter-check-for-suspend dance and we now own the monitor free
424 // and clear, i.e., it is no longer pending. The ThreadBlockInVM
425 // destructor can go to a safepoint at the end of this block. If we
426 // do a thread dump during that safepoint, then this thread will show
427 // as having "-locked" the monitor, but the OS and java.lang.Thread
428 // states will still report that the thread is blocked trying to
429 // acquire it.
430 }
432 Atomic::dec_ptr(&_count);
433 assert (_count >= 0, "invariant") ;
434 Self->_Stalled = 0 ;
436 // Must either set _recursions = 0 or ASSERT _recursions == 0.
437 assert (_recursions == 0 , "invariant") ;
438 assert (_owner == Self , "invariant") ;
439 assert (_succ != Self , "invariant") ;
440 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
442 // The thread -- now the owner -- is back in vm mode.
443 // Report the glorious news via TI,DTrace and jvmstat.
444 // The probe effect is non-trivial. All the reportage occurs
445 // while we hold the monitor, increasing the length of the critical
446 // section. Amdahl's parallel speedup law comes vividly into play.
447 //
448 // Another option might be to aggregate the events (thread local or
449 // per-monitor aggregation) and defer reporting until a more opportune
450 // time -- such as next time some thread encounters contention but has
451 // yet to acquire the lock. While spinning that thread could
452 // spinning we could increment JVMStat counters, etc.
454 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
455 if (JvmtiExport::should_post_monitor_contended_entered()) {
456 JvmtiExport::post_monitor_contended_entered(jt, this);
458 // The current thread already owns the monitor and is not going to
459 // call park() for the remainder of the monitor enter protocol. So
460 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
461 // event handler consumed an unpark() issued by the thread that
462 // just exited the monitor.
463 }
465 if (event.should_commit()) {
466 event.set_klass(((oop)this->object())->klass());
467 event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid);
468 event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr()));
469 event.commit();
470 }
472 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) {
473 ObjectMonitor::_sync_ContendedLockAttempts->inc() ;
474 }
475 }
478 // Caveat: TryLock() is not necessarily serializing if it returns failure.
479 // Callers must compensate as needed.
481 int ObjectMonitor::TryLock (Thread * Self) {
482 for (;;) {
483 void * own = _owner ;
484 if (own != NULL) return 0 ;
485 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) {
486 // Either guarantee _recursions == 0 or set _recursions = 0.
487 assert (_recursions == 0, "invariant") ;
488 assert (_owner == Self, "invariant") ;
489 // CONSIDER: set or assert that OwnerIsThread == 1
490 return 1 ;
491 }
492 // The lock had been free momentarily, but we lost the race to the lock.
493 // Interference -- the CAS failed.
494 // We can either return -1 or retry.
495 // Retry doesn't make as much sense because the lock was just acquired.
496 if (true) return -1 ;
497 }
498 }
500 void ATTR ObjectMonitor::EnterI (TRAPS) {
501 Thread * Self = THREAD ;
502 assert (Self->is_Java_thread(), "invariant") ;
503 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ;
505 // Try the lock - TATAS
506 if (TryLock (Self) > 0) {
507 assert (_succ != Self , "invariant") ;
508 assert (_owner == Self , "invariant") ;
509 assert (_Responsible != Self , "invariant") ;
510 return ;
511 }
513 DeferredInitialize () ;
515 // We try one round of spinning *before* enqueueing Self.
516 //
517 // If the _owner is ready but OFFPROC we could use a YieldTo()
518 // operation to donate the remainder of this thread's quantum
519 // to the owner. This has subtle but beneficial affinity
520 // effects.
522 if (TrySpin (Self) > 0) {
523 assert (_owner == Self , "invariant") ;
524 assert (_succ != Self , "invariant") ;
525 assert (_Responsible != Self , "invariant") ;
526 return ;
527 }
529 // The Spin failed -- Enqueue and park the thread ...
530 assert (_succ != Self , "invariant") ;
531 assert (_owner != Self , "invariant") ;
532 assert (_Responsible != Self , "invariant") ;
534 // Enqueue "Self" on ObjectMonitor's _cxq.
535 //
536 // Node acts as a proxy for Self.
537 // As an aside, if were to ever rewrite the synchronization code mostly
538 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
539 // Java objects. This would avoid awkward lifecycle and liveness issues,
540 // as well as eliminate a subset of ABA issues.
541 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
542 //
544 ObjectWaiter node(Self) ;
545 Self->_ParkEvent->reset() ;
546 node._prev = (ObjectWaiter *) 0xBAD ;
547 node.TState = ObjectWaiter::TS_CXQ ;
549 // Push "Self" onto the front of the _cxq.
550 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
551 // Note that spinning tends to reduce the rate at which threads
552 // enqueue and dequeue on EntryList|cxq.
553 ObjectWaiter * nxt ;
554 for (;;) {
555 node._next = nxt = _cxq ;
556 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ;
558 // Interference - the CAS failed because _cxq changed. Just retry.
559 // As an optional optimization we retry the lock.
560 if (TryLock (Self) > 0) {
561 assert (_succ != Self , "invariant") ;
562 assert (_owner == Self , "invariant") ;
563 assert (_Responsible != Self , "invariant") ;
564 return ;
565 }
566 }
568 // Check for cxq|EntryList edge transition to non-null. This indicates
569 // the onset of contention. While contention persists exiting threads
570 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit
571 // operations revert to the faster 1-0 mode. This enter operation may interleave
572 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
573 // arrange for one of the contending thread to use a timed park() operations
574 // to detect and recover from the race. (Stranding is form of progress failure
575 // where the monitor is unlocked but all the contending threads remain parked).
576 // That is, at least one of the contended threads will periodically poll _owner.
577 // One of the contending threads will become the designated "Responsible" thread.
578 // The Responsible thread uses a timed park instead of a normal indefinite park
579 // operation -- it periodically wakes and checks for and recovers from potential
580 // strandings admitted by 1-0 exit operations. We need at most one Responsible
581 // thread per-monitor at any given moment. Only threads on cxq|EntryList may
582 // be responsible for a monitor.
583 //
584 // Currently, one of the contended threads takes on the added role of "Responsible".
585 // A viable alternative would be to use a dedicated "stranding checker" thread
586 // that periodically iterated over all the threads (or active monitors) and unparked
587 // successors where there was risk of stranding. This would help eliminate the
588 // timer scalability issues we see on some platforms as we'd only have one thread
589 // -- the checker -- parked on a timer.
591 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) {
592 // Try to assume the role of responsible thread for the monitor.
593 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self }
594 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ;
595 }
597 // The lock have been released while this thread was occupied queueing
598 // itself onto _cxq. To close the race and avoid "stranding" and
599 // progress-liveness failure we must resample-retry _owner before parking.
600 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
601 // In this case the ST-MEMBAR is accomplished with CAS().
602 //
603 // TODO: Defer all thread state transitions until park-time.
604 // Since state transitions are heavy and inefficient we'd like
605 // to defer the state transitions until absolutely necessary,
606 // and in doing so avoid some transitions ...
608 TEVENT (Inflated enter - Contention) ;
609 int nWakeups = 0 ;
610 int RecheckInterval = 1 ;
612 for (;;) {
614 if (TryLock (Self) > 0) break ;
615 assert (_owner != Self, "invariant") ;
617 if ((SyncFlags & 2) && _Responsible == NULL) {
618 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ;
619 }
621 // park self
622 if (_Responsible == Self || (SyncFlags & 1)) {
623 TEVENT (Inflated enter - park TIMED) ;
624 Self->_ParkEvent->park ((jlong) RecheckInterval) ;
625 // Increase the RecheckInterval, but clamp the value.
626 RecheckInterval *= 8 ;
627 if (RecheckInterval > 1000) RecheckInterval = 1000 ;
628 } else {
629 TEVENT (Inflated enter - park UNTIMED) ;
630 Self->_ParkEvent->park() ;
631 }
633 if (TryLock(Self) > 0) break ;
635 // The lock is still contested.
636 // Keep a tally of the # of futile wakeups.
637 // Note that the counter is not protected by a lock or updated by atomics.
638 // That is by design - we trade "lossy" counters which are exposed to
639 // races during updates for a lower probe effect.
640 TEVENT (Inflated enter - Futile wakeup) ;
641 if (ObjectMonitor::_sync_FutileWakeups != NULL) {
642 ObjectMonitor::_sync_FutileWakeups->inc() ;
643 }
644 ++ nWakeups ;
646 // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
647 // We can defer clearing _succ until after the spin completes
648 // TrySpin() must tolerate being called with _succ == Self.
649 // Try yet another round of adaptive spinning.
650 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ;
652 // We can find that we were unpark()ed and redesignated _succ while
653 // we were spinning. That's harmless. If we iterate and call park(),
654 // park() will consume the event and return immediately and we'll
655 // just spin again. This pattern can repeat, leaving _succ to simply
656 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks().
657 // Alternately, we can sample fired() here, and if set, forgo spinning
658 // in the next iteration.
660 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) {
661 Self->_ParkEvent->reset() ;
662 OrderAccess::fence() ;
663 }
664 if (_succ == Self) _succ = NULL ;
666 // Invariant: after clearing _succ a thread *must* retry _owner before parking.
667 OrderAccess::fence() ;
668 }
670 // Egress :
671 // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
672 // Normally we'll find Self on the EntryList .
673 // From the perspective of the lock owner (this thread), the
674 // EntryList is stable and cxq is prepend-only.
675 // The head of cxq is volatile but the interior is stable.
676 // In addition, Self.TState is stable.
678 assert (_owner == Self , "invariant") ;
679 assert (object() != NULL , "invariant") ;
680 // I'd like to write:
681 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
682 // but as we're at a safepoint that's not safe.
684 UnlinkAfterAcquire (Self, &node) ;
685 if (_succ == Self) _succ = NULL ;
687 assert (_succ != Self, "invariant") ;
688 if (_Responsible == Self) {
689 _Responsible = NULL ;
690 OrderAccess::fence(); // Dekker pivot-point
692 // We may leave threads on cxq|EntryList without a designated
693 // "Responsible" thread. This is benign. When this thread subsequently
694 // exits the monitor it can "see" such preexisting "old" threads --
695 // threads that arrived on the cxq|EntryList before the fence, above --
696 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads
697 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
698 // non-null and elect a new "Responsible" timer thread.
699 //
700 // This thread executes:
701 // ST Responsible=null; MEMBAR (in enter epilog - here)
702 // LD cxq|EntryList (in subsequent exit)
703 //
704 // Entering threads in the slow/contended path execute:
705 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
706 // The (ST cxq; MEMBAR) is accomplished with CAS().
707 //
708 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
709 // exit operation from floating above the ST Responsible=null.
710 }
712 // We've acquired ownership with CAS().
713 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
714 // But since the CAS() this thread may have also stored into _succ,
715 // EntryList, cxq or Responsible. These meta-data updates must be
716 // visible __before this thread subsequently drops the lock.
717 // Consider what could occur if we didn't enforce this constraint --
718 // STs to monitor meta-data and user-data could reorder with (become
719 // visible after) the ST in exit that drops ownership of the lock.
720 // Some other thread could then acquire the lock, but observe inconsistent
721 // or old monitor meta-data and heap data. That violates the JMM.
722 // To that end, the 1-0 exit() operation must have at least STST|LDST
723 // "release" barrier semantics. Specifically, there must be at least a
724 // STST|LDST barrier in exit() before the ST of null into _owner that drops
725 // the lock. The barrier ensures that changes to monitor meta-data and data
726 // protected by the lock will be visible before we release the lock, and
727 // therefore before some other thread (CPU) has a chance to acquire the lock.
728 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
729 //
730 // Critically, any prior STs to _succ or EntryList must be visible before
731 // the ST of null into _owner in the *subsequent* (following) corresponding
732 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily
733 // execute a serializing instruction.
735 if (SyncFlags & 8) {
736 OrderAccess::fence() ;
737 }
738 return ;
739 }
741 // ReenterI() is a specialized inline form of the latter half of the
742 // contended slow-path from EnterI(). We use ReenterI() only for
743 // monitor reentry in wait().
744 //
745 // In the future we should reconcile EnterI() and ReenterI(), adding
746 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the
747 // loop accordingly.
749 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) {
750 assert (Self != NULL , "invariant") ;
751 assert (SelfNode != NULL , "invariant") ;
752 assert (SelfNode->_thread == Self , "invariant") ;
753 assert (_waiters > 0 , "invariant") ;
754 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ;
755 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ;
756 JavaThread * jt = (JavaThread *) Self ;
758 int nWakeups = 0 ;
759 for (;;) {
760 ObjectWaiter::TStates v = SelfNode->TState ;
761 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ;
762 assert (_owner != Self, "invariant") ;
764 if (TryLock (Self) > 0) break ;
765 if (TrySpin (Self) > 0) break ;
767 TEVENT (Wait Reentry - parking) ;
769 // State transition wrappers around park() ...
770 // ReenterI() wisely defers state transitions until
771 // it's clear we must park the thread.
772 {
773 OSThreadContendState osts(Self->osthread());
774 ThreadBlockInVM tbivm(jt);
776 // cleared by handle_special_suspend_equivalent_condition()
777 // or java_suspend_self()
778 jt->set_suspend_equivalent();
779 if (SyncFlags & 1) {
780 Self->_ParkEvent->park ((jlong)1000) ;
781 } else {
782 Self->_ParkEvent->park () ;
783 }
785 // were we externally suspended while we were waiting?
786 for (;;) {
787 if (!ExitSuspendEquivalent (jt)) break ;
788 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
789 jt->java_suspend_self();
790 jt->set_suspend_equivalent();
791 }
792 }
794 // Try again, but just so we distinguish between futile wakeups and
795 // successful wakeups. The following test isn't algorithmically
796 // necessary, but it helps us maintain sensible statistics.
797 if (TryLock(Self) > 0) break ;
799 // The lock is still contested.
800 // Keep a tally of the # of futile wakeups.
801 // Note that the counter is not protected by a lock or updated by atomics.
802 // That is by design - we trade "lossy" counters which are exposed to
803 // races during updates for a lower probe effect.
804 TEVENT (Wait Reentry - futile wakeup) ;
805 ++ nWakeups ;
807 // Assuming this is not a spurious wakeup we'll normally
808 // find that _succ == Self.
809 if (_succ == Self) _succ = NULL ;
811 // Invariant: after clearing _succ a contending thread
812 // *must* retry _owner before parking.
813 OrderAccess::fence() ;
815 if (ObjectMonitor::_sync_FutileWakeups != NULL) {
816 ObjectMonitor::_sync_FutileWakeups->inc() ;
817 }
818 }
820 // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
821 // Normally we'll find Self on the EntryList.
822 // Unlinking from the EntryList is constant-time and atomic-free.
823 // From the perspective of the lock owner (this thread), the
824 // EntryList is stable and cxq is prepend-only.
825 // The head of cxq is volatile but the interior is stable.
826 // In addition, Self.TState is stable.
828 assert (_owner == Self, "invariant") ;
829 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
830 UnlinkAfterAcquire (Self, SelfNode) ;
831 if (_succ == Self) _succ = NULL ;
832 assert (_succ != Self, "invariant") ;
833 SelfNode->TState = ObjectWaiter::TS_RUN ;
834 OrderAccess::fence() ; // see comments at the end of EnterI()
835 }
837 // after the thread acquires the lock in ::enter(). Equally, we could defer
838 // unlinking the thread until ::exit()-time.
840 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode)
841 {
842 assert (_owner == Self, "invariant") ;
843 assert (SelfNode->_thread == Self, "invariant") ;
845 if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
846 // Normal case: remove Self from the DLL EntryList .
847 // This is a constant-time operation.
848 ObjectWaiter * nxt = SelfNode->_next ;
849 ObjectWaiter * prv = SelfNode->_prev ;
850 if (nxt != NULL) nxt->_prev = prv ;
851 if (prv != NULL) prv->_next = nxt ;
852 if (SelfNode == _EntryList ) _EntryList = nxt ;
853 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ;
854 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ;
855 TEVENT (Unlink from EntryList) ;
856 } else {
857 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ;
858 // Inopportune interleaving -- Self is still on the cxq.
859 // This usually means the enqueue of self raced an exiting thread.
860 // Normally we'll find Self near the front of the cxq, so
861 // dequeueing is typically fast. If needbe we can accelerate
862 // this with some MCS/CHL-like bidirectional list hints and advisory
863 // back-links so dequeueing from the interior will normally operate
864 // in constant-time.
865 // Dequeue Self from either the head (with CAS) or from the interior
866 // with a linear-time scan and normal non-atomic memory operations.
867 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
868 // and then unlink Self from EntryList. We have to drain eventually,
869 // so it might as well be now.
871 ObjectWaiter * v = _cxq ;
872 assert (v != NULL, "invariant") ;
873 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) {
874 // The CAS above can fail from interference IFF a "RAT" arrived.
875 // In that case Self must be in the interior and can no longer be
876 // at the head of cxq.
877 if (v == SelfNode) {
878 assert (_cxq != v, "invariant") ;
879 v = _cxq ; // CAS above failed - start scan at head of list
880 }
881 ObjectWaiter * p ;
882 ObjectWaiter * q = NULL ;
883 for (p = v ; p != NULL && p != SelfNode; p = p->_next) {
884 q = p ;
885 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ;
886 }
887 assert (v != SelfNode, "invariant") ;
888 assert (p == SelfNode, "Node not found on cxq") ;
889 assert (p != _cxq, "invariant") ;
890 assert (q != NULL, "invariant") ;
891 assert (q->_next == p, "invariant") ;
892 q->_next = p->_next ;
893 }
894 TEVENT (Unlink from cxq) ;
895 }
897 // Diagnostic hygiene ...
898 SelfNode->_prev = (ObjectWaiter *) 0xBAD ;
899 SelfNode->_next = (ObjectWaiter *) 0xBAD ;
900 SelfNode->TState = ObjectWaiter::TS_RUN ;
901 }
903 // -----------------------------------------------------------------------------
904 // Exit support
905 //
906 // exit()
907 // ~~~~~~
908 // Note that the collector can't reclaim the objectMonitor or deflate
909 // the object out from underneath the thread calling ::exit() as the
910 // thread calling ::exit() never transitions to a stable state.
911 // This inhibits GC, which in turn inhibits asynchronous (and
912 // inopportune) reclamation of "this".
913 //
914 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
915 // There's one exception to the claim above, however. EnterI() can call
916 // exit() to drop a lock if the acquirer has been externally suspended.
917 // In that case exit() is called with _thread_state as _thread_blocked,
918 // but the monitor's _count field is > 0, which inhibits reclamation.
919 //
920 // 1-0 exit
921 // ~~~~~~~~
922 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
923 // the fast-path operators have been optimized so the common ::exit()
924 // operation is 1-0. See i486.ad fast_unlock(), for instance.
925 // The code emitted by fast_unlock() elides the usual MEMBAR. This
926 // greatly improves latency -- MEMBAR and CAS having considerable local
927 // latency on modern processors -- but at the cost of "stranding". Absent the
928 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
929 // ::enter() path, resulting in the entering thread being stranding
930 // and a progress-liveness failure. Stranding is extremely rare.
931 // We use timers (timed park operations) & periodic polling to detect
932 // and recover from stranding. Potentially stranded threads periodically
933 // wake up and poll the lock. See the usage of the _Responsible variable.
934 //
935 // The CAS() in enter provides for safety and exclusion, while the CAS or
936 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking
937 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding.
938 // We detect and recover from stranding with timers.
939 //
940 // If a thread transiently strands it'll park until (a) another
941 // thread acquires the lock and then drops the lock, at which time the
942 // exiting thread will notice and unpark the stranded thread, or, (b)
943 // the timer expires. If the lock is high traffic then the stranding latency
944 // will be low due to (a). If the lock is low traffic then the odds of
945 // stranding are lower, although the worst-case stranding latency
946 // is longer. Critically, we don't want to put excessive load in the
947 // platform's timer subsystem. We want to minimize both the timer injection
948 // rate (timers created/sec) as well as the number of timers active at
949 // any one time. (more precisely, we want to minimize timer-seconds, which is
950 // the integral of the # of active timers at any instant over time).
951 // Both impinge on OS scalability. Given that, at most one thread parked on
952 // a monitor will use a timer.
954 void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) {
955 Thread * Self = THREAD ;
956 if (THREAD != _owner) {
957 if (THREAD->is_lock_owned((address) _owner)) {
958 // Transmute _owner from a BasicLock pointer to a Thread address.
959 // We don't need to hold _mutex for this transition.
960 // Non-null to Non-null is safe as long as all readers can
961 // tolerate either flavor.
962 assert (_recursions == 0, "invariant") ;
963 _owner = THREAD ;
964 _recursions = 0 ;
965 OwnerIsThread = 1 ;
966 } else {
967 // NOTE: we need to handle unbalanced monitor enter/exit
968 // in native code by throwing an exception.
969 // TODO: Throw an IllegalMonitorStateException ?
970 TEVENT (Exit - Throw IMSX) ;
971 assert(false, "Non-balanced monitor enter/exit!");
972 if (false) {
973 THROW(vmSymbols::java_lang_IllegalMonitorStateException());
974 }
975 return;
976 }
977 }
979 if (_recursions != 0) {
980 _recursions--; // this is simple recursive enter
981 TEVENT (Inflated exit - recursive) ;
982 return ;
983 }
985 // Invariant: after setting Responsible=null an thread must execute
986 // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
987 if ((SyncFlags & 4) == 0) {
988 _Responsible = NULL ;
989 }
991 #if INCLUDE_TRACE
992 // get the owner's thread id for the MonitorEnter event
993 // if it is enabled and the thread isn't suspended
994 if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) {
995 _previous_owner_tid = SharedRuntime::get_java_tid(Self);
996 }
997 #endif
999 for (;;) {
1000 assert (THREAD == _owner, "invariant") ;
1003 if (Knob_ExitPolicy == 0) {
1004 // release semantics: prior loads and stores from within the critical section
1005 // must not float (reorder) past the following store that drops the lock.
1006 // On SPARC that requires MEMBAR #loadstore|#storestore.
1007 // But of course in TSO #loadstore|#storestore is not required.
1008 // I'd like to write one of the following:
1009 // A. OrderAccess::release() ; _owner = NULL
1010 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL;
1011 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both
1012 // store into a _dummy variable. That store is not needed, but can result
1013 // in massive wasteful coherency traffic on classic SMP systems.
1014 // Instead, I use release_store(), which is implemented as just a simple
1015 // ST on x64, x86 and SPARC.
1016 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock
1017 OrderAccess::storeload() ; // See if we need to wake a successor
1018 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
1019 TEVENT (Inflated exit - simple egress) ;
1020 return ;
1021 }
1022 TEVENT (Inflated exit - complex egress) ;
1024 // Normally the exiting thread is responsible for ensuring succession,
1025 // but if other successors are ready or other entering threads are spinning
1026 // then this thread can simply store NULL into _owner and exit without
1027 // waking a successor. The existence of spinners or ready successors
1028 // guarantees proper succession (liveness). Responsibility passes to the
1029 // ready or running successors. The exiting thread delegates the duty.
1030 // More precisely, if a successor already exists this thread is absolved
1031 // of the responsibility of waking (unparking) one.
1032 //
1033 // The _succ variable is critical to reducing futile wakeup frequency.
1034 // _succ identifies the "heir presumptive" thread that has been made
1035 // ready (unparked) but that has not yet run. We need only one such
1036 // successor thread to guarantee progress.
1037 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1038 // section 3.3 "Futile Wakeup Throttling" for details.
1039 //
1040 // Note that spinners in Enter() also set _succ non-null.
1041 // In the current implementation spinners opportunistically set
1042 // _succ so that exiting threads might avoid waking a successor.
1043 // Another less appealing alternative would be for the exiting thread
1044 // to drop the lock and then spin briefly to see if a spinner managed
1045 // to acquire the lock. If so, the exiting thread could exit
1046 // immediately without waking a successor, otherwise the exiting
1047 // thread would need to dequeue and wake a successor.
1048 // (Note that we'd need to make the post-drop spin short, but no
1049 // shorter than the worst-case round-trip cache-line migration time.
1050 // The dropped lock needs to become visible to the spinner, and then
1051 // the acquisition of the lock by the spinner must become visible to
1052 // the exiting thread).
1053 //
1055 // It appears that an heir-presumptive (successor) must be made ready.
1056 // Only the current lock owner can manipulate the EntryList or
1057 // drain _cxq, so we need to reacquire the lock. If we fail
1058 // to reacquire the lock the responsibility for ensuring succession
1059 // falls to the new owner.
1060 //
1061 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
1062 return ;
1063 }
1064 TEVENT (Exit - Reacquired) ;
1065 } else {
1066 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
1067 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock
1068 OrderAccess::storeload() ;
1069 // Ratify the previously observed values.
1070 if (_cxq == NULL || _succ != NULL) {
1071 TEVENT (Inflated exit - simple egress) ;
1072 return ;
1073 }
1075 // inopportune interleaving -- the exiting thread (this thread)
1076 // in the fast-exit path raced an entering thread in the slow-enter
1077 // path.
1078 // We have two choices:
1079 // A. Try to reacquire the lock.
1080 // If the CAS() fails return immediately, otherwise
1081 // we either restart/rerun the exit operation, or simply
1082 // fall-through into the code below which wakes a successor.
1083 // B. If the elements forming the EntryList|cxq are TSM
1084 // we could simply unpark() the lead thread and return
1085 // without having set _succ.
1086 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) {
1087 TEVENT (Inflated exit - reacquired succeeded) ;
1088 return ;
1089 }
1090 TEVENT (Inflated exit - reacquired failed) ;
1091 } else {
1092 TEVENT (Inflated exit - complex egress) ;
1093 }
1094 }
1096 guarantee (_owner == THREAD, "invariant") ;
1098 ObjectWaiter * w = NULL ;
1099 int QMode = Knob_QMode ;
1101 if (QMode == 2 && _cxq != NULL) {
1102 // QMode == 2 : cxq has precedence over EntryList.
1103 // Try to directly wake a successor from the cxq.
1104 // If successful, the successor will need to unlink itself from cxq.
1105 w = _cxq ;
1106 assert (w != NULL, "invariant") ;
1107 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1108 ExitEpilog (Self, w) ;
1109 return ;
1110 }
1112 if (QMode == 3 && _cxq != NULL) {
1113 // Aggressively drain cxq into EntryList at the first opportunity.
1114 // This policy ensure that recently-run threads live at the head of EntryList.
1115 // Drain _cxq into EntryList - bulk transfer.
1116 // First, detach _cxq.
1117 // The following loop is tantamount to: w = swap (&cxq, NULL)
1118 w = _cxq ;
1119 for (;;) {
1120 assert (w != NULL, "Invariant") ;
1121 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ;
1122 if (u == w) break ;
1123 w = u ;
1124 }
1125 assert (w != NULL , "invariant") ;
1127 ObjectWaiter * q = NULL ;
1128 ObjectWaiter * p ;
1129 for (p = w ; p != NULL ; p = p->_next) {
1130 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1131 p->TState = ObjectWaiter::TS_ENTER ;
1132 p->_prev = q ;
1133 q = p ;
1134 }
1136 // Append the RATs to the EntryList
1137 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time.
1138 ObjectWaiter * Tail ;
1139 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ;
1140 if (Tail == NULL) {
1141 _EntryList = w ;
1142 } else {
1143 Tail->_next = w ;
1144 w->_prev = Tail ;
1145 }
1147 // Fall thru into code that tries to wake a successor from EntryList
1148 }
1150 if (QMode == 4 && _cxq != NULL) {
1151 // Aggressively drain cxq into EntryList at the first opportunity.
1152 // This policy ensure that recently-run threads live at the head of EntryList.
1154 // Drain _cxq into EntryList - bulk transfer.
1155 // First, detach _cxq.
1156 // The following loop is tantamount to: w = swap (&cxq, NULL)
1157 w = _cxq ;
1158 for (;;) {
1159 assert (w != NULL, "Invariant") ;
1160 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ;
1161 if (u == w) break ;
1162 w = u ;
1163 }
1164 assert (w != NULL , "invariant") ;
1166 ObjectWaiter * q = NULL ;
1167 ObjectWaiter * p ;
1168 for (p = w ; p != NULL ; p = p->_next) {
1169 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1170 p->TState = ObjectWaiter::TS_ENTER ;
1171 p->_prev = q ;
1172 q = p ;
1173 }
1175 // Prepend the RATs to the EntryList
1176 if (_EntryList != NULL) {
1177 q->_next = _EntryList ;
1178 _EntryList->_prev = q ;
1179 }
1180 _EntryList = w ;
1182 // Fall thru into code that tries to wake a successor from EntryList
1183 }
1185 w = _EntryList ;
1186 if (w != NULL) {
1187 // I'd like to write: guarantee (w->_thread != Self).
1188 // But in practice an exiting thread may find itself on the EntryList.
1189 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and
1190 // then calls exit(). Exit release the lock by setting O._owner to NULL.
1191 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The
1192 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1193 // release the lock "O". T2 resumes immediately after the ST of null into
1194 // _owner, above. T2 notices that the EntryList is populated, so it
1195 // reacquires the lock and then finds itself on the EntryList.
1196 // Given all that, we have to tolerate the circumstance where "w" is
1197 // associated with Self.
1198 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1199 ExitEpilog (Self, w) ;
1200 return ;
1201 }
1203 // If we find that both _cxq and EntryList are null then just
1204 // re-run the exit protocol from the top.
1205 w = _cxq ;
1206 if (w == NULL) continue ;
1208 // Drain _cxq into EntryList - bulk transfer.
1209 // First, detach _cxq.
1210 // The following loop is tantamount to: w = swap (&cxq, NULL)
1211 for (;;) {
1212 assert (w != NULL, "Invariant") ;
1213 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ;
1214 if (u == w) break ;
1215 w = u ;
1216 }
1217 TEVENT (Inflated exit - drain cxq into EntryList) ;
1219 assert (w != NULL , "invariant") ;
1220 assert (_EntryList == NULL , "invariant") ;
1222 // Convert the LIFO SLL anchored by _cxq into a DLL.
1223 // The list reorganization step operates in O(LENGTH(w)) time.
1224 // It's critical that this step operate quickly as
1225 // "Self" still holds the outer-lock, restricting parallelism
1226 // and effectively lengthening the critical section.
1227 // Invariant: s chases t chases u.
1228 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1229 // we have faster access to the tail.
1231 if (QMode == 1) {
1232 // QMode == 1 : drain cxq to EntryList, reversing order
1233 // We also reverse the order of the list.
1234 ObjectWaiter * s = NULL ;
1235 ObjectWaiter * t = w ;
1236 ObjectWaiter * u = NULL ;
1237 while (t != NULL) {
1238 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ;
1239 t->TState = ObjectWaiter::TS_ENTER ;
1240 u = t->_next ;
1241 t->_prev = u ;
1242 t->_next = s ;
1243 s = t;
1244 t = u ;
1245 }
1246 _EntryList = s ;
1247 assert (s != NULL, "invariant") ;
1248 } else {
1249 // QMode == 0 or QMode == 2
1250 _EntryList = w ;
1251 ObjectWaiter * q = NULL ;
1252 ObjectWaiter * p ;
1253 for (p = w ; p != NULL ; p = p->_next) {
1254 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ;
1255 p->TState = ObjectWaiter::TS_ENTER ;
1256 p->_prev = q ;
1257 q = p ;
1258 }
1259 }
1261 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1262 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1264 // See if we can abdicate to a spinner instead of waking a thread.
1265 // A primary goal of the implementation is to reduce the
1266 // context-switch rate.
1267 if (_succ != NULL) continue;
1269 w = _EntryList ;
1270 if (w != NULL) {
1271 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1272 ExitEpilog (Self, w) ;
1273 return ;
1274 }
1275 }
1276 }
1278 // ExitSuspendEquivalent:
1279 // A faster alternate to handle_special_suspend_equivalent_condition()
1280 //
1281 // handle_special_suspend_equivalent_condition() unconditionally
1282 // acquires the SR_lock. On some platforms uncontended MutexLocker()
1283 // operations have high latency. Note that in ::enter() we call HSSEC
1284 // while holding the monitor, so we effectively lengthen the critical sections.
1285 //
1286 // There are a number of possible solutions:
1287 //
1288 // A. To ameliorate the problem we might also defer state transitions
1289 // to as late as possible -- just prior to parking.
1290 // Given that, we'd call HSSEC after having returned from park(),
1291 // but before attempting to acquire the monitor. This is only a
1292 // partial solution. It avoids calling HSSEC while holding the
1293 // monitor (good), but it still increases successor reacquisition latency --
1294 // the interval between unparking a successor and the time the successor
1295 // resumes and retries the lock. See ReenterI(), which defers state transitions.
1296 // If we use this technique we can also avoid EnterI()-exit() loop
1297 // in ::enter() where we iteratively drop the lock and then attempt
1298 // to reacquire it after suspending.
1299 //
1300 // B. In the future we might fold all the suspend bits into a
1301 // composite per-thread suspend flag and then update it with CAS().
1302 // Alternately, a Dekker-like mechanism with multiple variables
1303 // would suffice:
1304 // ST Self->_suspend_equivalent = false
1305 // MEMBAR
1306 // LD Self_>_suspend_flags
1307 //
1310 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) {
1311 int Mode = Knob_FastHSSEC ;
1312 if (Mode && !jSelf->is_external_suspend()) {
1313 assert (jSelf->is_suspend_equivalent(), "invariant") ;
1314 jSelf->clear_suspend_equivalent() ;
1315 if (2 == Mode) OrderAccess::storeload() ;
1316 if (!jSelf->is_external_suspend()) return false ;
1317 // We raced a suspension -- fall thru into the slow path
1318 TEVENT (ExitSuspendEquivalent - raced) ;
1319 jSelf->set_suspend_equivalent() ;
1320 }
1321 return jSelf->handle_special_suspend_equivalent_condition() ;
1322 }
1325 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) {
1326 assert (_owner == Self, "invariant") ;
1328 // Exit protocol:
1329 // 1. ST _succ = wakee
1330 // 2. membar #loadstore|#storestore;
1331 // 2. ST _owner = NULL
1332 // 3. unpark(wakee)
1334 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ;
1335 ParkEvent * Trigger = Wakee->_event ;
1337 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1338 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1339 // out-of-scope (non-extant).
1340 Wakee = NULL ;
1342 // Drop the lock
1343 OrderAccess::release_store_ptr (&_owner, NULL) ;
1344 OrderAccess::fence() ; // ST _owner vs LD in unpark()
1346 if (SafepointSynchronize::do_call_back()) {
1347 TEVENT (unpark before SAFEPOINT) ;
1348 }
1350 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1351 Trigger->unpark() ;
1353 // Maintain stats and report events to JVMTI
1354 if (ObjectMonitor::_sync_Parks != NULL) {
1355 ObjectMonitor::_sync_Parks->inc() ;
1356 }
1357 }
1360 // -----------------------------------------------------------------------------
1361 // Class Loader deadlock handling.
1362 //
1363 // complete_exit exits a lock returning recursion count
1364 // complete_exit/reenter operate as a wait without waiting
1365 // complete_exit requires an inflated monitor
1366 // The _owner field is not always the Thread addr even with an
1367 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1368 // thread due to contention.
1369 intptr_t ObjectMonitor::complete_exit(TRAPS) {
1370 Thread * const Self = THREAD;
1371 assert(Self->is_Java_thread(), "Must be Java thread!");
1372 JavaThread *jt = (JavaThread *)THREAD;
1374 DeferredInitialize();
1376 if (THREAD != _owner) {
1377 if (THREAD->is_lock_owned ((address)_owner)) {
1378 assert(_recursions == 0, "internal state error");
1379 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */
1380 _recursions = 0 ;
1381 OwnerIsThread = 1 ;
1382 }
1383 }
1385 guarantee(Self == _owner, "complete_exit not owner");
1386 intptr_t save = _recursions; // record the old recursion count
1387 _recursions = 0; // set the recursion level to be 0
1388 exit (true, Self) ; // exit the monitor
1389 guarantee (_owner != Self, "invariant");
1390 return save;
1391 }
1393 // reenter() enters a lock and sets recursion count
1394 // complete_exit/reenter operate as a wait without waiting
1395 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) {
1396 Thread * const Self = THREAD;
1397 assert(Self->is_Java_thread(), "Must be Java thread!");
1398 JavaThread *jt = (JavaThread *)THREAD;
1400 guarantee(_owner != Self, "reenter already owner");
1401 enter (THREAD); // enter the monitor
1402 guarantee (_recursions == 0, "reenter recursion");
1403 _recursions = recursions;
1404 return;
1405 }
1408 // -----------------------------------------------------------------------------
1409 // A macro is used below because there may already be a pending
1410 // exception which should not abort the execution of the routines
1411 // which use this (which is why we don't put this into check_slow and
1412 // call it with a CHECK argument).
1414 #define CHECK_OWNER() \
1415 do { \
1416 if (THREAD != _owner) { \
1417 if (THREAD->is_lock_owned((address) _owner)) { \
1418 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \
1419 _recursions = 0; \
1420 OwnerIsThread = 1 ; \
1421 } else { \
1422 TEVENT (Throw IMSX) ; \
1423 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \
1424 } \
1425 } \
1426 } while (false)
1428 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception.
1429 // TODO-FIXME: remove check_slow() -- it's likely dead.
1431 void ObjectMonitor::check_slow(TRAPS) {
1432 TEVENT (check_slow - throw IMSX) ;
1433 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner");
1434 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner");
1435 }
1437 static int Adjust (volatile int * adr, int dx) {
1438 int v ;
1439 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ;
1440 return v ;
1441 }
1443 // helper method for posting a monitor wait event
1444 void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event,
1445 jlong notifier_tid,
1446 jlong timeout,
1447 bool timedout) {
1448 event->set_klass(((oop)this->object())->klass());
1449 event->set_timeout((TYPE_ULONG)timeout);
1450 event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr()));
1451 event->set_notifier((TYPE_OSTHREAD)notifier_tid);
1452 event->set_timedOut((TYPE_BOOLEAN)timedout);
1453 event->commit();
1454 }
1456 // -----------------------------------------------------------------------------
1457 // Wait/Notify/NotifyAll
1458 //
1459 // Note: a subset of changes to ObjectMonitor::wait()
1460 // will need to be replicated in complete_exit above
1461 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1462 Thread * const Self = THREAD ;
1463 assert(Self->is_Java_thread(), "Must be Java thread!");
1464 JavaThread *jt = (JavaThread *)THREAD;
1466 DeferredInitialize () ;
1468 // Throw IMSX or IEX.
1469 CHECK_OWNER();
1471 EventJavaMonitorWait event;
1473 // check for a pending interrupt
1474 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1475 // post monitor waited event. Note that this is past-tense, we are done waiting.
1476 if (JvmtiExport::should_post_monitor_waited()) {
1477 // Note: 'false' parameter is passed here because the
1478 // wait was not timed out due to thread interrupt.
1479 JvmtiExport::post_monitor_waited(jt, this, false);
1481 // In this short circuit of the monitor wait protocol, the
1482 // current thread never drops ownership of the monitor and
1483 // never gets added to the wait queue so the current thread
1484 // cannot be made the successor. This means that the
1485 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1486 // consume an unpark() meant for the ParkEvent associated with
1487 // this ObjectMonitor.
1488 }
1489 if (event.should_commit()) {
1490 post_monitor_wait_event(&event, 0, millis, false);
1491 }
1492 TEVENT (Wait - Throw IEX) ;
1493 THROW(vmSymbols::java_lang_InterruptedException());
1494 return ;
1495 }
1497 TEVENT (Wait) ;
1499 assert (Self->_Stalled == 0, "invariant") ;
1500 Self->_Stalled = intptr_t(this) ;
1501 jt->set_current_waiting_monitor(this);
1503 // create a node to be put into the queue
1504 // Critically, after we reset() the event but prior to park(), we must check
1505 // for a pending interrupt.
1506 ObjectWaiter node(Self);
1507 node.TState = ObjectWaiter::TS_WAIT ;
1508 Self->_ParkEvent->reset() ;
1509 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag
1511 // Enter the waiting queue, which is a circular doubly linked list in this case
1512 // but it could be a priority queue or any data structure.
1513 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only
1514 // by the the owner of the monitor *except* in the case where park()
1515 // returns because of a timeout of interrupt. Contention is exceptionally rare
1516 // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1518 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ;
1519 AddWaiter (&node) ;
1520 Thread::SpinRelease (&_WaitSetLock) ;
1522 if ((SyncFlags & 4) == 0) {
1523 _Responsible = NULL ;
1524 }
1525 intptr_t save = _recursions; // record the old recursion count
1526 _waiters++; // increment the number of waiters
1527 _recursions = 0; // set the recursion level to be 1
1528 exit (true, Self) ; // exit the monitor
1529 guarantee (_owner != Self, "invariant") ;
1531 // The thread is on the WaitSet list - now park() it.
1532 // On MP systems it's conceivable that a brief spin before we park
1533 // could be profitable.
1534 //
1535 // TODO-FIXME: change the following logic to a loop of the form
1536 // while (!timeout && !interrupted && _notified == 0) park()
1538 int ret = OS_OK ;
1539 int WasNotified = 0 ;
1540 { // State transition wrappers
1541 OSThread* osthread = Self->osthread();
1542 OSThreadWaitState osts(osthread, true);
1543 {
1544 ThreadBlockInVM tbivm(jt);
1545 // Thread is in thread_blocked state and oop access is unsafe.
1546 jt->set_suspend_equivalent();
1548 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) {
1549 // Intentionally empty
1550 } else
1551 if (node._notified == 0) {
1552 if (millis <= 0) {
1553 Self->_ParkEvent->park () ;
1554 } else {
1555 ret = Self->_ParkEvent->park (millis) ;
1556 }
1557 }
1559 // were we externally suspended while we were waiting?
1560 if (ExitSuspendEquivalent (jt)) {
1561 // TODO-FIXME: add -- if succ == Self then succ = null.
1562 jt->java_suspend_self();
1563 }
1565 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1568 // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1569 // from the WaitSet to the EntryList.
1570 // See if we need to remove Node from the WaitSet.
1571 // We use double-checked locking to avoid grabbing _WaitSetLock
1572 // if the thread is not on the wait queue.
1573 //
1574 // Note that we don't need a fence before the fetch of TState.
1575 // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1576 // written by the is thread. (perhaps the fetch might even be satisfied
1577 // by a look-aside into the processor's own store buffer, although given
1578 // the length of the code path between the prior ST and this load that's
1579 // highly unlikely). If the following LD fetches a stale TS_WAIT value
1580 // then we'll acquire the lock and then re-fetch a fresh TState value.
1581 // That is, we fail toward safety.
1583 if (node.TState == ObjectWaiter::TS_WAIT) {
1584 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ;
1585 if (node.TState == ObjectWaiter::TS_WAIT) {
1586 DequeueSpecificWaiter (&node) ; // unlink from WaitSet
1587 assert(node._notified == 0, "invariant");
1588 node.TState = ObjectWaiter::TS_RUN ;
1589 }
1590 Thread::SpinRelease (&_WaitSetLock) ;
1591 }
1593 // The thread is now either on off-list (TS_RUN),
1594 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1595 // The Node's TState variable is stable from the perspective of this thread.
1596 // No other threads will asynchronously modify TState.
1597 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ;
1598 OrderAccess::loadload() ;
1599 if (_succ == Self) _succ = NULL ;
1600 WasNotified = node._notified ;
1602 // Reentry phase -- reacquire the monitor.
1603 // re-enter contended monitor after object.wait().
1604 // retain OBJECT_WAIT state until re-enter successfully completes
1605 // Thread state is thread_in_vm and oop access is again safe,
1606 // although the raw address of the object may have changed.
1607 // (Don't cache naked oops over safepoints, of course).
1609 // post monitor waited event. Note that this is past-tense, we are done waiting.
1610 if (JvmtiExport::should_post_monitor_waited()) {
1611 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1613 if (node._notified != 0 && _succ == Self) {
1614 // In this part of the monitor wait-notify-reenter protocol it
1615 // is possible (and normal) for another thread to do a fastpath
1616 // monitor enter-exit while this thread is still trying to get
1617 // to the reenter portion of the protocol.
1618 //
1619 // The ObjectMonitor was notified and the current thread is
1620 // the successor which also means that an unpark() has already
1621 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1622 // consume the unpark() that was done when the successor was
1623 // set because the same ParkEvent is shared between Java
1624 // monitors and JVM/TI RawMonitors (for now).
1625 //
1626 // We redo the unpark() to ensure forward progress, i.e., we
1627 // don't want all pending threads hanging (parked) with none
1628 // entering the unlocked monitor.
1629 node._event->unpark();
1630 }
1631 }
1633 if (event.should_commit()) {
1634 post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT);
1635 }
1637 OrderAccess::fence() ;
1639 assert (Self->_Stalled != 0, "invariant") ;
1640 Self->_Stalled = 0 ;
1642 assert (_owner != Self, "invariant") ;
1643 ObjectWaiter::TStates v = node.TState ;
1644 if (v == ObjectWaiter::TS_RUN) {
1645 enter (Self) ;
1646 } else {
1647 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ;
1648 ReenterI (Self, &node) ;
1649 node.wait_reenter_end(this);
1650 }
1652 // Self has reacquired the lock.
1653 // Lifecycle - the node representing Self must not appear on any queues.
1654 // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1655 // want residual elements associated with this thread left on any lists.
1656 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ;
1657 assert (_owner == Self, "invariant") ;
1658 assert (_succ != Self , "invariant") ;
1659 } // OSThreadWaitState()
1661 jt->set_current_waiting_monitor(NULL);
1663 guarantee (_recursions == 0, "invariant") ;
1664 _recursions = save; // restore the old recursion count
1665 _waiters--; // decrement the number of waiters
1667 // Verify a few postconditions
1668 assert (_owner == Self , "invariant") ;
1669 assert (_succ != Self , "invariant") ;
1670 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ;
1672 if (SyncFlags & 32) {
1673 OrderAccess::fence() ;
1674 }
1676 // check if the notification happened
1677 if (!WasNotified) {
1678 // no, it could be timeout or Thread.interrupt() or both
1679 // check for interrupt event, otherwise it is timeout
1680 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) {
1681 TEVENT (Wait - throw IEX from epilog) ;
1682 THROW(vmSymbols::java_lang_InterruptedException());
1683 }
1684 }
1686 // NOTE: Spurious wake up will be consider as timeout.
1687 // Monitor notify has precedence over thread interrupt.
1688 }
1691 // Consider:
1692 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1693 // then instead of transferring a thread from the WaitSet to the EntryList
1694 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1696 void ObjectMonitor::notify(TRAPS) {
1697 CHECK_OWNER();
1698 if (_WaitSet == NULL) {
1699 TEVENT (Empty-Notify) ;
1700 return ;
1701 }
1702 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1704 int Policy = Knob_MoveNotifyee ;
1706 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ;
1707 ObjectWaiter * iterator = DequeueWaiter() ;
1708 if (iterator != NULL) {
1709 TEVENT (Notify1 - Transfer) ;
1710 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ;
1711 guarantee (iterator->_notified == 0, "invariant") ;
1712 if (Policy != 4) {
1713 iterator->TState = ObjectWaiter::TS_ENTER ;
1714 }
1715 iterator->_notified = 1 ;
1716 Thread * Self = THREAD;
1717 iterator->_notifier_tid = Self->osthread()->thread_id();
1719 ObjectWaiter * List = _EntryList ;
1720 if (List != NULL) {
1721 assert (List->_prev == NULL, "invariant") ;
1722 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1723 assert (List != iterator, "invariant") ;
1724 }
1726 if (Policy == 0) { // prepend to EntryList
1727 if (List == NULL) {
1728 iterator->_next = iterator->_prev = NULL ;
1729 _EntryList = iterator ;
1730 } else {
1731 List->_prev = iterator ;
1732 iterator->_next = List ;
1733 iterator->_prev = NULL ;
1734 _EntryList = iterator ;
1735 }
1736 } else
1737 if (Policy == 1) { // append to EntryList
1738 if (List == NULL) {
1739 iterator->_next = iterator->_prev = NULL ;
1740 _EntryList = iterator ;
1741 } else {
1742 // CONSIDER: finding the tail currently requires a linear-time walk of
1743 // the EntryList. We can make tail access constant-time by converting to
1744 // a CDLL instead of using our current DLL.
1745 ObjectWaiter * Tail ;
1746 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ;
1747 assert (Tail != NULL && Tail->_next == NULL, "invariant") ;
1748 Tail->_next = iterator ;
1749 iterator->_prev = Tail ;
1750 iterator->_next = NULL ;
1751 }
1752 } else
1753 if (Policy == 2) { // prepend to cxq
1754 // prepend to cxq
1755 if (List == NULL) {
1756 iterator->_next = iterator->_prev = NULL ;
1757 _EntryList = iterator ;
1758 } else {
1759 iterator->TState = ObjectWaiter::TS_CXQ ;
1760 for (;;) {
1761 ObjectWaiter * Front = _cxq ;
1762 iterator->_next = Front ;
1763 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) {
1764 break ;
1765 }
1766 }
1767 }
1768 } else
1769 if (Policy == 3) { // append to cxq
1770 iterator->TState = ObjectWaiter::TS_CXQ ;
1771 for (;;) {
1772 ObjectWaiter * Tail ;
1773 Tail = _cxq ;
1774 if (Tail == NULL) {
1775 iterator->_next = NULL ;
1776 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) {
1777 break ;
1778 }
1779 } else {
1780 while (Tail->_next != NULL) Tail = Tail->_next ;
1781 Tail->_next = iterator ;
1782 iterator->_prev = Tail ;
1783 iterator->_next = NULL ;
1784 break ;
1785 }
1786 }
1787 } else {
1788 ParkEvent * ev = iterator->_event ;
1789 iterator->TState = ObjectWaiter::TS_RUN ;
1790 OrderAccess::fence() ;
1791 ev->unpark() ;
1792 }
1794 if (Policy < 4) {
1795 iterator->wait_reenter_begin(this);
1796 }
1798 // _WaitSetLock protects the wait queue, not the EntryList. We could
1799 // move the add-to-EntryList operation, above, outside the critical section
1800 // protected by _WaitSetLock. In practice that's not useful. With the
1801 // exception of wait() timeouts and interrupts the monitor owner
1802 // is the only thread that grabs _WaitSetLock. There's almost no contention
1803 // on _WaitSetLock so it's not profitable to reduce the length of the
1804 // critical section.
1805 }
1807 Thread::SpinRelease (&_WaitSetLock) ;
1809 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) {
1810 ObjectMonitor::_sync_Notifications->inc() ;
1811 }
1812 }
1815 void ObjectMonitor::notifyAll(TRAPS) {
1816 CHECK_OWNER();
1817 ObjectWaiter* iterator;
1818 if (_WaitSet == NULL) {
1819 TEVENT (Empty-NotifyAll) ;
1820 return ;
1821 }
1822 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1824 int Policy = Knob_MoveNotifyee ;
1825 int Tally = 0 ;
1826 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ;
1828 for (;;) {
1829 iterator = DequeueWaiter () ;
1830 if (iterator == NULL) break ;
1831 TEVENT (NotifyAll - Transfer1) ;
1832 ++Tally ;
1834 // Disposition - what might we do with iterator ?
1835 // a. add it directly to the EntryList - either tail or head.
1836 // b. push it onto the front of the _cxq.
1837 // For now we use (a).
1839 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ;
1840 guarantee (iterator->_notified == 0, "invariant") ;
1841 iterator->_notified = 1 ;
1842 Thread * Self = THREAD;
1843 iterator->_notifier_tid = Self->osthread()->thread_id();
1844 if (Policy != 4) {
1845 iterator->TState = ObjectWaiter::TS_ENTER ;
1846 }
1848 ObjectWaiter * List = _EntryList ;
1849 if (List != NULL) {
1850 assert (List->_prev == NULL, "invariant") ;
1851 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ;
1852 assert (List != iterator, "invariant") ;
1853 }
1855 if (Policy == 0) { // prepend to EntryList
1856 if (List == NULL) {
1857 iterator->_next = iterator->_prev = NULL ;
1858 _EntryList = iterator ;
1859 } else {
1860 List->_prev = iterator ;
1861 iterator->_next = List ;
1862 iterator->_prev = NULL ;
1863 _EntryList = iterator ;
1864 }
1865 } else
1866 if (Policy == 1) { // append to EntryList
1867 if (List == NULL) {
1868 iterator->_next = iterator->_prev = NULL ;
1869 _EntryList = iterator ;
1870 } else {
1871 // CONSIDER: finding the tail currently requires a linear-time walk of
1872 // the EntryList. We can make tail access constant-time by converting to
1873 // a CDLL instead of using our current DLL.
1874 ObjectWaiter * Tail ;
1875 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ;
1876 assert (Tail != NULL && Tail->_next == NULL, "invariant") ;
1877 Tail->_next = iterator ;
1878 iterator->_prev = Tail ;
1879 iterator->_next = NULL ;
1880 }
1881 } else
1882 if (Policy == 2) { // prepend to cxq
1883 // prepend to cxq
1884 iterator->TState = ObjectWaiter::TS_CXQ ;
1885 for (;;) {
1886 ObjectWaiter * Front = _cxq ;
1887 iterator->_next = Front ;
1888 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) {
1889 break ;
1890 }
1891 }
1892 } else
1893 if (Policy == 3) { // append to cxq
1894 iterator->TState = ObjectWaiter::TS_CXQ ;
1895 for (;;) {
1896 ObjectWaiter * Tail ;
1897 Tail = _cxq ;
1898 if (Tail == NULL) {
1899 iterator->_next = NULL ;
1900 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) {
1901 break ;
1902 }
1903 } else {
1904 while (Tail->_next != NULL) Tail = Tail->_next ;
1905 Tail->_next = iterator ;
1906 iterator->_prev = Tail ;
1907 iterator->_next = NULL ;
1908 break ;
1909 }
1910 }
1911 } else {
1912 ParkEvent * ev = iterator->_event ;
1913 iterator->TState = ObjectWaiter::TS_RUN ;
1914 OrderAccess::fence() ;
1915 ev->unpark() ;
1916 }
1918 if (Policy < 4) {
1919 iterator->wait_reenter_begin(this);
1920 }
1922 // _WaitSetLock protects the wait queue, not the EntryList. We could
1923 // move the add-to-EntryList operation, above, outside the critical section
1924 // protected by _WaitSetLock. In practice that's not useful. With the
1925 // exception of wait() timeouts and interrupts the monitor owner
1926 // is the only thread that grabs _WaitSetLock. There's almost no contention
1927 // on _WaitSetLock so it's not profitable to reduce the length of the
1928 // critical section.
1929 }
1931 Thread::SpinRelease (&_WaitSetLock) ;
1933 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) {
1934 ObjectMonitor::_sync_Notifications->inc(Tally) ;
1935 }
1936 }
1938 // -----------------------------------------------------------------------------
1939 // Adaptive Spinning Support
1940 //
1941 // Adaptive spin-then-block - rational spinning
1942 //
1943 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1944 // algorithm. On high order SMP systems it would be better to start with
1945 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH,
1946 // a contending thread could enqueue itself on the cxq and then spin locally
1947 // on a thread-specific variable such as its ParkEvent._Event flag.
1948 // That's left as an exercise for the reader. Note that global spinning is
1949 // not problematic on Niagara, as the L2$ serves the interconnect and has both
1950 // low latency and massive bandwidth.
1951 //
1952 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1953 // acquisition attempts where we opt to spin -- at 100% and vary the spin count
1954 // (duration) or we can fix the count at approximately the duration of
1955 // a context switch and vary the frequency. Of course we could also
1956 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1957 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html.
1958 //
1959 // This implementation varies the duration "D", where D varies with
1960 // the success rate of recent spin attempts. (D is capped at approximately
1961 // length of a round-trip context switch). The success rate for recent
1962 // spin attempts is a good predictor of the success rate of future spin
1963 // attempts. The mechanism adapts automatically to varying critical
1964 // section length (lock modality), system load and degree of parallelism.
1965 // D is maintained per-monitor in _SpinDuration and is initialized
1966 // optimistically. Spin frequency is fixed at 100%.
1967 //
1968 // Note that _SpinDuration is volatile, but we update it without locks
1969 // or atomics. The code is designed so that _SpinDuration stays within
1970 // a reasonable range even in the presence of races. The arithmetic
1971 // operations on _SpinDuration are closed over the domain of legal values,
1972 // so at worst a race will install and older but still legal value.
1973 // At the very worst this introduces some apparent non-determinism.
1974 // We might spin when we shouldn't or vice-versa, but since the spin
1975 // count are relatively short, even in the worst case, the effect is harmless.
1976 //
1977 // Care must be taken that a low "D" value does not become an
1978 // an absorbing state. Transient spinning failures -- when spinning
1979 // is overall profitable -- should not cause the system to converge
1980 // on low "D" values. We want spinning to be stable and predictable
1981 // and fairly responsive to change and at the same time we don't want
1982 // it to oscillate, become metastable, be "too" non-deterministic,
1983 // or converge on or enter undesirable stable absorbing states.
1984 //
1985 // We implement a feedback-based control system -- using past behavior
1986 // to predict future behavior. We face two issues: (a) if the
1987 // input signal is random then the spin predictor won't provide optimal
1988 // results, and (b) if the signal frequency is too high then the control
1989 // system, which has some natural response lag, will "chase" the signal.
1990 // (b) can arise from multimodal lock hold times. Transient preemption
1991 // can also result in apparent bimodal lock hold times.
1992 // Although sub-optimal, neither condition is particularly harmful, as
1993 // in the worst-case we'll spin when we shouldn't or vice-versa.
1994 // The maximum spin duration is rather short so the failure modes aren't bad.
1995 // To be conservative, I've tuned the gain in system to bias toward
1996 // _not spinning. Relatedly, the system can sometimes enter a mode where it
1997 // "rings" or oscillates between spinning and not spinning. This happens
1998 // when spinning is just on the cusp of profitability, however, so the
1999 // situation is not dire. The state is benign -- there's no need to add
2000 // hysteresis control to damp the transition rate between spinning and
2001 // not spinning.
2002 //
2004 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ;
2005 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ;
2007 // Spinning: Fixed frequency (100%), vary duration
2010 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) {
2012 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning.
2013 int ctr = Knob_FixedSpin ;
2014 if (ctr != 0) {
2015 while (--ctr >= 0) {
2016 if (TryLock (Self) > 0) return 1 ;
2017 SpinPause () ;
2018 }
2019 return 0 ;
2020 }
2022 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) {
2023 if (TryLock(Self) > 0) {
2024 // Increase _SpinDuration ...
2025 // Note that we don't clamp SpinDuration precisely at SpinLimit.
2026 // Raising _SpurDuration to the poverty line is key.
2027 int x = _SpinDuration ;
2028 if (x < Knob_SpinLimit) {
2029 if (x < Knob_Poverty) x = Knob_Poverty ;
2030 _SpinDuration = x + Knob_BonusB ;
2031 }
2032 return 1 ;
2033 }
2034 SpinPause () ;
2035 }
2037 // Admission control - verify preconditions for spinning
2038 //
2039 // We always spin a little bit, just to prevent _SpinDuration == 0 from
2040 // becoming an absorbing state. Put another way, we spin briefly to
2041 // sample, just in case the system load, parallelism, contention, or lock
2042 // modality changed.
2043 //
2044 // Consider the following alternative:
2045 // Periodically set _SpinDuration = _SpinLimit and try a long/full
2046 // spin attempt. "Periodically" might mean after a tally of
2047 // the # of failed spin attempts (or iterations) reaches some threshold.
2048 // This takes us into the realm of 1-out-of-N spinning, where we
2049 // hold the duration constant but vary the frequency.
2051 ctr = _SpinDuration ;
2052 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ;
2053 if (ctr <= 0) return 0 ;
2055 if (Knob_SuccRestrict && _succ != NULL) return 0 ;
2056 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) {
2057 TEVENT (Spin abort - notrunnable [TOP]);
2058 return 0 ;
2059 }
2061 int MaxSpin = Knob_MaxSpinners ;
2062 if (MaxSpin >= 0) {
2063 if (_Spinner > MaxSpin) {
2064 TEVENT (Spin abort -- too many spinners) ;
2065 return 0 ;
2066 }
2067 // Slighty racy, but benign ...
2068 Adjust (&_Spinner, 1) ;
2069 }
2071 // We're good to spin ... spin ingress.
2072 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
2073 // when preparing to LD...CAS _owner, etc and the CAS is likely
2074 // to succeed.
2075 int hits = 0 ;
2076 int msk = 0 ;
2077 int caspty = Knob_CASPenalty ;
2078 int oxpty = Knob_OXPenalty ;
2079 int sss = Knob_SpinSetSucc ;
2080 if (sss && _succ == NULL ) _succ = Self ;
2081 Thread * prv = NULL ;
2083 // There are three ways to exit the following loop:
2084 // 1. A successful spin where this thread has acquired the lock.
2085 // 2. Spin failure with prejudice
2086 // 3. Spin failure without prejudice
2088 while (--ctr >= 0) {
2090 // Periodic polling -- Check for pending GC
2091 // Threads may spin while they're unsafe.
2092 // We don't want spinning threads to delay the JVM from reaching
2093 // a stop-the-world safepoint or to steal cycles from GC.
2094 // If we detect a pending safepoint we abort in order that
2095 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
2096 // this thread, if safe, doesn't steal cycles from GC.
2097 // This is in keeping with the "no loitering in runtime" rule.
2098 // We periodically check to see if there's a safepoint pending.
2099 if ((ctr & 0xFF) == 0) {
2100 if (SafepointSynchronize::do_call_back()) {
2101 TEVENT (Spin: safepoint) ;
2102 goto Abort ; // abrupt spin egress
2103 }
2104 if (Knob_UsePause & 1) SpinPause () ;
2106 int (*scb)(intptr_t,int) = SpinCallbackFunction ;
2107 if (hits > 50 && scb != NULL) {
2108 int abend = (*scb)(SpinCallbackArgument, 0) ;
2109 }
2110 }
2112 if (Knob_UsePause & 2) SpinPause() ;
2114 // Exponential back-off ... Stay off the bus to reduce coherency traffic.
2115 // This is useful on classic SMP systems, but is of less utility on
2116 // N1-style CMT platforms.
2117 //
2118 // Trade-off: lock acquisition latency vs coherency bandwidth.
2119 // Lock hold times are typically short. A histogram
2120 // of successful spin attempts shows that we usually acquire
2121 // the lock early in the spin. That suggests we want to
2122 // sample _owner frequently in the early phase of the spin,
2123 // but then back-off and sample less frequently as the spin
2124 // progresses. The back-off makes a good citizen on SMP big
2125 // SMP systems. Oversampling _owner can consume excessive
2126 // coherency bandwidth. Relatedly, if we _oversample _owner we
2127 // can inadvertently interfere with the the ST m->owner=null.
2128 // executed by the lock owner.
2129 if (ctr & msk) continue ;
2130 ++hits ;
2131 if ((hits & 0xF) == 0) {
2132 // The 0xF, above, corresponds to the exponent.
2133 // Consider: (msk+1)|msk
2134 msk = ((msk << 2)|3) & BackOffMask ;
2135 }
2137 // Probe _owner with TATAS
2138 // If this thread observes the monitor transition or flicker
2139 // from locked to unlocked to locked, then the odds that this
2140 // thread will acquire the lock in this spin attempt go down
2141 // considerably. The same argument applies if the CAS fails
2142 // or if we observe _owner change from one non-null value to
2143 // another non-null value. In such cases we might abort
2144 // the spin without prejudice or apply a "penalty" to the
2145 // spin count-down variable "ctr", reducing it by 100, say.
2147 Thread * ox = (Thread *) _owner ;
2148 if (ox == NULL) {
2149 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ;
2150 if (ox == NULL) {
2151 // The CAS succeeded -- this thread acquired ownership
2152 // Take care of some bookkeeping to exit spin state.
2153 if (sss && _succ == Self) {
2154 _succ = NULL ;
2155 }
2156 if (MaxSpin > 0) Adjust (&_Spinner, -1) ;
2158 // Increase _SpinDuration :
2159 // The spin was successful (profitable) so we tend toward
2160 // longer spin attempts in the future.
2161 // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
2162 // If we acquired the lock early in the spin cycle it
2163 // makes sense to increase _SpinDuration proportionally.
2164 // Note that we don't clamp SpinDuration precisely at SpinLimit.
2165 int x = _SpinDuration ;
2166 if (x < Knob_SpinLimit) {
2167 if (x < Knob_Poverty) x = Knob_Poverty ;
2168 _SpinDuration = x + Knob_Bonus ;
2169 }
2170 return 1 ;
2171 }
2173 // The CAS failed ... we can take any of the following actions:
2174 // * penalize: ctr -= Knob_CASPenalty
2175 // * exit spin with prejudice -- goto Abort;
2176 // * exit spin without prejudice.
2177 // * Since CAS is high-latency, retry again immediately.
2178 prv = ox ;
2179 TEVENT (Spin: cas failed) ;
2180 if (caspty == -2) break ;
2181 if (caspty == -1) goto Abort ;
2182 ctr -= caspty ;
2183 continue ;
2184 }
2186 // Did lock ownership change hands ?
2187 if (ox != prv && prv != NULL ) {
2188 TEVENT (spin: Owner changed)
2189 if (oxpty == -2) break ;
2190 if (oxpty == -1) goto Abort ;
2191 ctr -= oxpty ;
2192 }
2193 prv = ox ;
2195 // Abort the spin if the owner is not executing.
2196 // The owner must be executing in order to drop the lock.
2197 // Spinning while the owner is OFFPROC is idiocy.
2198 // Consider: ctr -= RunnablePenalty ;
2199 if (Knob_OState && NotRunnable (Self, ox)) {
2200 TEVENT (Spin abort - notrunnable);
2201 goto Abort ;
2202 }
2203 if (sss && _succ == NULL ) _succ = Self ;
2204 }
2206 // Spin failed with prejudice -- reduce _SpinDuration.
2207 // TODO: Use an AIMD-like policy to adjust _SpinDuration.
2208 // AIMD is globally stable.
2209 TEVENT (Spin failure) ;
2210 {
2211 int x = _SpinDuration ;
2212 if (x > 0) {
2213 // Consider an AIMD scheme like: x -= (x >> 3) + 100
2214 // This is globally sample and tends to damp the response.
2215 x -= Knob_Penalty ;
2216 if (x < 0) x = 0 ;
2217 _SpinDuration = x ;
2218 }
2219 }
2221 Abort:
2222 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ;
2223 if (sss && _succ == Self) {
2224 _succ = NULL ;
2225 // Invariant: after setting succ=null a contending thread
2226 // must recheck-retry _owner before parking. This usually happens
2227 // in the normal usage of TrySpin(), but it's safest
2228 // to make TrySpin() as foolproof as possible.
2229 OrderAccess::fence() ;
2230 if (TryLock(Self) > 0) return 1 ;
2231 }
2232 return 0 ;
2233 }
2235 // NotRunnable() -- informed spinning
2236 //
2237 // Don't bother spinning if the owner is not eligible to drop the lock.
2238 // Peek at the owner's schedctl.sc_state and Thread._thread_values and
2239 // spin only if the owner thread is _thread_in_Java or _thread_in_vm.
2240 // The thread must be runnable in order to drop the lock in timely fashion.
2241 // If the _owner is not runnable then spinning will not likely be
2242 // successful (profitable).
2243 //
2244 // Beware -- the thread referenced by _owner could have died
2245 // so a simply fetch from _owner->_thread_state might trap.
2246 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
2247 // Because of the lifecycle issues the schedctl and _thread_state values
2248 // observed by NotRunnable() might be garbage. NotRunnable must
2249 // tolerate this and consider the observed _thread_state value
2250 // as advisory.
2251 //
2252 // Beware too, that _owner is sometimes a BasicLock address and sometimes
2253 // a thread pointer. We differentiate the two cases with OwnerIsThread.
2254 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
2255 // with the LSB of _owner. Another option would be to probablistically probe
2256 // the putative _owner->TypeTag value.
2257 //
2258 // Checking _thread_state isn't perfect. Even if the thread is
2259 // in_java it might be blocked on a page-fault or have been preempted
2260 // and sitting on a ready/dispatch queue. _thread state in conjunction
2261 // with schedctl.sc_state gives us a good picture of what the
2262 // thread is doing, however.
2263 //
2264 // TODO: check schedctl.sc_state.
2265 // We'll need to use SafeFetch32() to read from the schedctl block.
2266 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/
2267 //
2268 // The return value from NotRunnable() is *advisory* -- the
2269 // result is based on sampling and is not necessarily coherent.
2270 // The caller must tolerate false-negative and false-positive errors.
2271 // Spinning, in general, is probabilistic anyway.
2274 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) {
2275 // Check either OwnerIsThread or ox->TypeTag == 2BAD.
2276 if (!OwnerIsThread) return 0 ;
2278 if (ox == NULL) return 0 ;
2280 // Avoid transitive spinning ...
2281 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L.
2282 // Immediately after T1 acquires L it's possible that T2, also
2283 // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
2284 // This occurs transiently after T1 acquired L but before
2285 // T1 managed to clear T1.Stalled. T2 does not need to abort
2286 // its spin in this circumstance.
2287 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ;
2289 if (BlockedOn == 1) return 1 ;
2290 if (BlockedOn != 0) {
2291 return BlockedOn != intptr_t(this) && _owner == ox ;
2292 }
2294 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ;
2295 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ;
2296 // consider also: jst != _thread_in_Java -- but that's overspecific.
2297 return jst == _thread_blocked || jst == _thread_in_native ;
2298 }
2301 // -----------------------------------------------------------------------------
2302 // WaitSet management ...
2304 ObjectWaiter::ObjectWaiter(Thread* thread) {
2305 _next = NULL;
2306 _prev = NULL;
2307 _notified = 0;
2308 TState = TS_RUN ;
2309 _thread = thread;
2310 _event = thread->_ParkEvent ;
2311 _active = false;
2312 assert (_event != NULL, "invariant") ;
2313 }
2315 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) {
2316 JavaThread *jt = (JavaThread *)this->_thread;
2317 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
2318 }
2320 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) {
2321 JavaThread *jt = (JavaThread *)this->_thread;
2322 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
2323 }
2325 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
2326 assert(node != NULL, "should not dequeue NULL node");
2327 assert(node->_prev == NULL, "node already in list");
2328 assert(node->_next == NULL, "node already in list");
2329 // put node at end of queue (circular doubly linked list)
2330 if (_WaitSet == NULL) {
2331 _WaitSet = node;
2332 node->_prev = node;
2333 node->_next = node;
2334 } else {
2335 ObjectWaiter* head = _WaitSet ;
2336 ObjectWaiter* tail = head->_prev;
2337 assert(tail->_next == head, "invariant check");
2338 tail->_next = node;
2339 head->_prev = node;
2340 node->_next = head;
2341 node->_prev = tail;
2342 }
2343 }
2345 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
2346 // dequeue the very first waiter
2347 ObjectWaiter* waiter = _WaitSet;
2348 if (waiter) {
2349 DequeueSpecificWaiter(waiter);
2350 }
2351 return waiter;
2352 }
2354 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
2355 assert(node != NULL, "should not dequeue NULL node");
2356 assert(node->_prev != NULL, "node already removed from list");
2357 assert(node->_next != NULL, "node already removed from list");
2358 // when the waiter has woken up because of interrupt,
2359 // timeout or other spurious wake-up, dequeue the
2360 // waiter from waiting list
2361 ObjectWaiter* next = node->_next;
2362 if (next == node) {
2363 assert(node->_prev == node, "invariant check");
2364 _WaitSet = NULL;
2365 } else {
2366 ObjectWaiter* prev = node->_prev;
2367 assert(prev->_next == node, "invariant check");
2368 assert(next->_prev == node, "invariant check");
2369 next->_prev = prev;
2370 prev->_next = next;
2371 if (_WaitSet == node) {
2372 _WaitSet = next;
2373 }
2374 }
2375 node->_next = NULL;
2376 node->_prev = NULL;
2377 }
2379 // -----------------------------------------------------------------------------
2380 // PerfData support
2381 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ;
2382 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ;
2383 PerfCounter * ObjectMonitor::_sync_Parks = NULL ;
2384 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ;
2385 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ;
2386 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ;
2387 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ;
2388 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ;
2389 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ;
2390 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ;
2391 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ;
2392 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ;
2393 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ;
2394 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ;
2395 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ;
2396 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ;
2397 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ;
2398 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ;
2400 // One-shot global initialization for the sync subsystem.
2401 // We could also defer initialization and initialize on-demand
2402 // the first time we call inflate(). Initialization would
2403 // be protected - like so many things - by the MonitorCache_lock.
2405 void ObjectMonitor::Initialize () {
2406 static int InitializationCompleted = 0 ;
2407 assert (InitializationCompleted == 0, "invariant") ;
2408 InitializationCompleted = 1 ;
2409 if (UsePerfData) {
2410 EXCEPTION_MARK ;
2411 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); }
2412 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); }
2413 NEWPERFCOUNTER(_sync_Inflations) ;
2414 NEWPERFCOUNTER(_sync_Deflations) ;
2415 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ;
2416 NEWPERFCOUNTER(_sync_FutileWakeups) ;
2417 NEWPERFCOUNTER(_sync_Parks) ;
2418 NEWPERFCOUNTER(_sync_EmptyNotifications) ;
2419 NEWPERFCOUNTER(_sync_Notifications) ;
2420 NEWPERFCOUNTER(_sync_SlowEnter) ;
2421 NEWPERFCOUNTER(_sync_SlowExit) ;
2422 NEWPERFCOUNTER(_sync_SlowNotify) ;
2423 NEWPERFCOUNTER(_sync_SlowNotifyAll) ;
2424 NEWPERFCOUNTER(_sync_FailedSpins) ;
2425 NEWPERFCOUNTER(_sync_SuccessfulSpins) ;
2426 NEWPERFCOUNTER(_sync_PrivateA) ;
2427 NEWPERFCOUNTER(_sync_PrivateB) ;
2428 NEWPERFCOUNTER(_sync_MonInCirculation) ;
2429 NEWPERFCOUNTER(_sync_MonScavenged) ;
2430 NEWPERFVARIABLE(_sync_MonExtant) ;
2431 #undef NEWPERFCOUNTER
2432 }
2433 }
2436 // Compile-time asserts
2437 // When possible, it's better to catch errors deterministically at
2438 // compile-time than at runtime. The down-side to using compile-time
2439 // asserts is that error message -- often something about negative array
2440 // indices -- is opaque.
2442 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); }
2444 void ObjectMonitor::ctAsserts() {
2445 CTASSERT(offset_of (ObjectMonitor, _header) == 0);
2446 }
2449 static char * kvGet (char * kvList, const char * Key) {
2450 if (kvList == NULL) return NULL ;
2451 size_t n = strlen (Key) ;
2452 char * Search ;
2453 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) {
2454 if (strncmp (Search, Key, n) == 0) {
2455 if (Search[n] == '=') return Search + n + 1 ;
2456 if (Search[n] == 0) return (char *) "1" ;
2457 }
2458 }
2459 return NULL ;
2460 }
2462 static int kvGetInt (char * kvList, const char * Key, int Default) {
2463 char * v = kvGet (kvList, Key) ;
2464 int rslt = v ? ::strtol (v, NULL, 0) : Default ;
2465 if (Knob_ReportSettings && v != NULL) {
2466 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ;
2467 ::fflush (stdout) ;
2468 }
2469 return rslt ;
2470 }
2472 void ObjectMonitor::DeferredInitialize () {
2473 if (InitDone > 0) return ;
2474 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) {
2475 while (InitDone != 1) ;
2476 return ;
2477 }
2479 // One-shot global initialization ...
2480 // The initialization is idempotent, so we don't need locks.
2481 // In the future consider doing this via os::init_2().
2482 // SyncKnobs consist of <Key>=<Value> pairs in the style
2483 // of environment variables. Start by converting ':' to NUL.
2485 if (SyncKnobs == NULL) SyncKnobs = "" ;
2487 size_t sz = strlen (SyncKnobs) ;
2488 char * knobs = (char *) malloc (sz + 2) ;
2489 if (knobs == NULL) {
2490 vm_exit_out_of_memory (sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs") ;
2491 guarantee (0, "invariant") ;
2492 }
2493 strcpy (knobs, SyncKnobs) ;
2494 knobs[sz+1] = 0 ;
2495 for (char * p = knobs ; *p ; p++) {
2496 if (*p == ':') *p = 0 ;
2497 }
2499 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); }
2500 SETKNOB(ReportSettings) ;
2501 SETKNOB(Verbose) ;
2502 SETKNOB(FixedSpin) ;
2503 SETKNOB(SpinLimit) ;
2504 SETKNOB(SpinBase) ;
2505 SETKNOB(SpinBackOff);
2506 SETKNOB(CASPenalty) ;
2507 SETKNOB(OXPenalty) ;
2508 SETKNOB(LogSpins) ;
2509 SETKNOB(SpinSetSucc) ;
2510 SETKNOB(SuccEnabled) ;
2511 SETKNOB(SuccRestrict) ;
2512 SETKNOB(Penalty) ;
2513 SETKNOB(Bonus) ;
2514 SETKNOB(BonusB) ;
2515 SETKNOB(Poverty) ;
2516 SETKNOB(SpinAfterFutile) ;
2517 SETKNOB(UsePause) ;
2518 SETKNOB(SpinEarly) ;
2519 SETKNOB(OState) ;
2520 SETKNOB(MaxSpinners) ;
2521 SETKNOB(PreSpin) ;
2522 SETKNOB(ExitPolicy) ;
2523 SETKNOB(QMode);
2524 SETKNOB(ResetEvent) ;
2525 SETKNOB(MoveNotifyee) ;
2526 SETKNOB(FastHSSEC) ;
2527 #undef SETKNOB
2529 if (os::is_MP()) {
2530 BackOffMask = (1 << Knob_SpinBackOff) - 1 ;
2531 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ;
2532 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1)
2533 } else {
2534 Knob_SpinLimit = 0 ;
2535 Knob_SpinBase = 0 ;
2536 Knob_PreSpin = 0 ;
2537 Knob_FixedSpin = -1 ;
2538 }
2540 if (Knob_LogSpins == 0) {
2541 ObjectMonitor::_sync_FailedSpins = NULL ;
2542 }
2544 free (knobs) ;
2545 OrderAccess::fence() ;
2546 InitDone = 1 ;
2547 }
2549 #ifndef PRODUCT
2550 void ObjectMonitor::verify() {
2551 }
2553 void ObjectMonitor::print() {
2554 }
2555 #endif