1
2 /*
3 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "runtime/mutex.hpp"
28 #include "runtime/osThread.hpp"
29 #include "runtime/thread.inline.hpp"
30 #include "utilities/events.hpp"
31 #ifdef TARGET_OS_FAMILY_linux
32 # include "mutex_linux.inline.hpp"
33 #endif
34 #ifdef TARGET_OS_FAMILY_solaris
35 # include "mutex_solaris.inline.hpp"
36 #endif
37 #ifdef TARGET_OS_FAMILY_windows
38 # include "mutex_windows.inline.hpp"
39 #endif
40 #ifdef TARGET_OS_FAMILY_bsd
41 # include "mutex_bsd.inline.hpp"
42 #endif
43
44 // o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o
45 //
46 // Native Monitor-Mutex locking - theory of operations
47 //
48 // * Native Monitors are completely unrelated to Java-level monitors,
49 // although the "back-end" slow-path implementations share a common lineage.
50 // See objectMonitor:: in synchronizer.cpp.
51 // Native Monitors do *not* support nesting or recursion but otherwise
52 // they're basically Hoare-flavor monitors.
53 //
54 // * A thread acquires ownership of a Monitor/Mutex by CASing the LockByte
55 // in the _LockWord from zero to non-zero. Note that the _Owner field
56 // is advisory and is used only to verify that the thread calling unlock()
57 // is indeed the last thread to have acquired the lock.
58 //
59 // * Contending threads "push" themselves onto the front of the contention
60 // queue -- called the cxq -- with CAS and then spin/park.
61 // The _LockWord contains the LockByte as well as the pointer to the head
62 // of the cxq. Colocating the LockByte with the cxq precludes certain races.
63 //
64 // * Using a separately addressable LockByte allows for CAS:MEMBAR or CAS:0
65 // idioms. We currently use MEMBAR in the uncontended unlock() path, as
66 // MEMBAR often has less latency than CAS. If warranted, we could switch to
67 // a CAS:0 mode, using timers to close the resultant race, as is done
68 // with Java Monitors in synchronizer.cpp.
69 //
70 // See the following for a discussion of the relative cost of atomics (CAS)
71 // MEMBAR, and ways to eliminate such instructions from the common-case paths:
72 // -- http://blogs.sun.com/dave/entry/biased_locking_in_hotspot
73 // -- http://blogs.sun.com/dave/resource/MustangSync.pdf
74 // -- http://blogs.sun.com/dave/resource/synchronization-public2.pdf
75 // -- synchronizer.cpp
76 //
77 // * Overall goals - desiderata
78 // 1. Minimize context switching
79 // 2. Minimize lock migration
80 // 3. Minimize CPI -- affinity and locality
81 // 4. Minimize the execution of high-latency instructions such as CAS or MEMBAR
82 // 5. Minimize outer lock hold times
83 // 6. Behave gracefully on a loaded system
84 //
85 // * Thread flow and list residency:
86 //
87 // Contention queue --> EntryList --> OnDeck --> Owner --> !Owner
88 // [..resident on monitor list..]
89 // [...........contending..................]
90 //
91 // -- The contention queue (cxq) contains recently-arrived threads (RATs).
92 // Threads on the cxq eventually drain into the EntryList.
93 // -- Invariant: a thread appears on at most one list -- cxq, EntryList
94 // or WaitSet -- at any one time.
95 // -- For a given monitor there can be at most one "OnDeck" thread at any
96 // given time but if needbe this particular invariant could be relaxed.
97 //
98 // * The WaitSet and EntryList linked lists are composed of ParkEvents.
99 // I use ParkEvent instead of threads as ParkEvents are immortal and
100 // type-stable, meaning we can safely unpark() a possibly stale
101 // list element in the unlock()-path. (That's benign).
102 //
103 // * Succession policy - providing for progress:
104 //
105 // As necessary, the unlock()ing thread identifies, unlinks, and unparks
106 // an "heir presumptive" tentative successor thread from the EntryList.
107 // This becomes the so-called "OnDeck" thread, of which there can be only
108 // one at any given time for a given monitor. The wakee will recontend
109 // for ownership of monitor.
110 //
111 // Succession is provided for by a policy of competitive handoff.
112 // The exiting thread does _not_ grant or pass ownership to the
113 // successor thread. (This is also referred to as "handoff" succession").
114 // Instead the exiting thread releases ownership and possibly wakes
115 // a successor, so the successor can (re)compete for ownership of the lock.
116 //
117 // Competitive handoff provides excellent overall throughput at the expense
118 // of short-term fairness. If fairness is a concern then one remedy might
119 // be to add an AcquireCounter field to the monitor. After a thread acquires
120 // the lock it will decrement the AcquireCounter field. When the count
121 // reaches 0 the thread would reset the AcquireCounter variable, abdicate
122 // the lock directly to some thread on the EntryList, and then move itself to the
123 // tail of the EntryList.
124 //
125 // But in practice most threads engage or otherwise participate in resource
126 // bounded producer-consumer relationships, so lock domination is not usually
127 // a practical concern. Recall too, that in general it's easier to construct
128 // a fair lock from a fast lock, but not vice-versa.
129 //
130 // * The cxq can have multiple concurrent "pushers" but only one concurrent
131 // detaching thread. This mechanism is immune from the ABA corruption.
132 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
133 // We use OnDeck as a pseudo-lock to enforce the at-most-one detaching
134 // thread constraint.
135 //
136 // * Taken together, the cxq and the EntryList constitute or form a
137 // single logical queue of threads stalled trying to acquire the lock.
138 // We use two distinct lists to reduce heat on the list ends.
139 // Threads in lock() enqueue onto cxq while threads in unlock() will
140 // dequeue from the EntryList. (c.f. Michael Scott's "2Q" algorithm).
141 // A key desideratum is to minimize queue & monitor metadata manipulation
142 // that occurs while holding the "outer" monitor lock -- that is, we want to
143 // minimize monitor lock holds times.
144 //
145 // The EntryList is ordered by the prevailing queue discipline and
146 // can be organized in any convenient fashion, such as a doubly-linked list or
147 // a circular doubly-linked list. If we need a priority queue then something akin
148 // to Solaris' sleepq would work nicely. Viz.,
149 // -- http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
150 // -- http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/uts/common/os/sleepq.c
151 // Queue discipline is enforced at ::unlock() time, when the unlocking thread
152 // drains the cxq into the EntryList, and orders or reorders the threads on the
153 // EntryList accordingly.
154 //
155 // Barring "lock barging", this mechanism provides fair cyclic ordering,
156 // somewhat similar to an elevator-scan.
157 //
158 // * OnDeck
159 // -- For a given monitor there can be at most one OnDeck thread at any given
160 // instant. The OnDeck thread is contending for the lock, but has been
161 // unlinked from the EntryList and cxq by some previous unlock() operations.
162 // Once a thread has been designated the OnDeck thread it will remain so
163 // until it manages to acquire the lock -- being OnDeck is a stable property.
164 // -- Threads on the EntryList or cxq are _not allowed to attempt lock acquisition.
165 // -- OnDeck also serves as an "inner lock" as follows. Threads in unlock() will, after
166 // having cleared the LockByte and dropped the outer lock, attempt to "trylock"
167 // OnDeck by CASing the field from null to non-null. If successful, that thread
168 // is then responsible for progress and succession and can use CAS to detach and
169 // drain the cxq into the EntryList. By convention, only this thread, the holder of
170 // the OnDeck inner lock, can manipulate the EntryList or detach and drain the
171 // RATs on the cxq into the EntryList. This avoids ABA corruption on the cxq as
172 // we allow multiple concurrent "push" operations but restrict detach concurrency
173 // to at most one thread. Having selected and detached a successor, the thread then
174 // changes the OnDeck to refer to that successor, and then unparks the successor.
175 // That successor will eventually acquire the lock and clear OnDeck. Beware
176 // that the OnDeck usage as a lock is asymmetric. A thread in unlock() transiently
177 // "acquires" OnDeck, performs queue manipulations, passes OnDeck to some successor,
178 // and then the successor eventually "drops" OnDeck. Note that there's never
179 // any sense of contention on the inner lock, however. Threads never contend
180 // or wait for the inner lock.
181 // -- OnDeck provides for futile wakeup throttling a described in section 3.3 of
182 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
183 // In a sense, OnDeck subsumes the ObjectMonitor _Succ and ObjectWaiter
184 // TState fields found in Java-level objectMonitors. (See synchronizer.cpp).
185 //
186 // * Waiting threads reside on the WaitSet list -- wait() puts
187 // the caller onto the WaitSet. Notify() or notifyAll() simply
188 // transfers threads from the WaitSet to either the EntryList or cxq.
189 // Subsequent unlock() operations will eventually unpark the notifyee.
190 // Unparking a notifee in notify() proper is inefficient - if we were to do so
191 // it's likely the notifyee would simply impale itself on the lock held
192 // by the notifier.
193 //
194 // * The mechanism is obstruction-free in that if the holder of the transient
195 // OnDeck lock in unlock() is preempted or otherwise stalls, other threads
196 // can still acquire and release the outer lock and continue to make progress.
197 // At worst, waking of already blocked contending threads may be delayed,
198 // but nothing worse. (We only use "trylock" operations on the inner OnDeck
199 // lock).
200 //
201 // * Note that thread-local storage must be initialized before a thread
202 // uses Native monitors or mutexes. The native monitor-mutex subsystem
203 // depends on Thread::current().
204 //
205 // * The monitor synchronization subsystem avoids the use of native
206 // synchronization primitives except for the narrow platform-specific
207 // park-unpark abstraction. See the comments in os_solaris.cpp regarding
208 // the semantics of park-unpark. Put another way, this monitor implementation
209 // depends only on atomic operations and park-unpark. The monitor subsystem
210 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
211 // underlying OS manages the READY<->RUN transitions.
212 //
213 // * The memory consistency model provide by lock()-unlock() is at least as
214 // strong or stronger than the Java Memory model defined by JSR-133.
215 // That is, we guarantee at least entry consistency, if not stronger.
216 // See http://g.oswego.edu/dl/jmm/cookbook.html.
217 //
218 // * Thread:: currently contains a set of purpose-specific ParkEvents:
219 // _MutexEvent, _ParkEvent, etc. A better approach might be to do away with
220 // the purpose-specific ParkEvents and instead implement a general per-thread
221 // stack of available ParkEvents which we could provision on-demand. The
222 // stack acts as a local cache to avoid excessive calls to ParkEvent::Allocate()
223 // and ::Release(). A thread would simply pop an element from the local stack before it
224 // enqueued or park()ed. When the contention was over the thread would
225 // push the no-longer-needed ParkEvent back onto its stack.
226 //
227 // * A slightly reduced form of ILock() and IUnlock() have been partially
228 // model-checked (Murphi) for safety and progress at T=1,2,3 and 4.
229 // It'd be interesting to see if TLA/TLC could be useful as well.
230 //
231 // * Mutex-Monitor is a low-level "leaf" subsystem. That is, the monitor
232 // code should never call other code in the JVM that might itself need to
233 // acquire monitors or mutexes. That's true *except* in the case of the
234 // ThreadBlockInVM state transition wrappers. The ThreadBlockInVM DTOR handles
235 // mutator reentry (ingress) by checking for a pending safepoint in which case it will
236 // call SafepointSynchronize::block(), which in turn may call Safepoint_lock->lock(), etc.
237 // In that particular case a call to lock() for a given Monitor can end up recursively
238 // calling lock() on another monitor. While distasteful, this is largely benign
239 // as the calls come from jacket that wraps lock(), and not from deep within lock() itself.
240 //
241 // It's unfortunate that native mutexes and thread state transitions were convolved.
242 // They're really separate concerns and should have remained that way. Melding
243 // them together was facile -- a bit too facile. The current implementation badly
244 // conflates the two concerns.
245 //
246 // * TODO-FIXME:
247 //
248 // -- Add DTRACE probes for contended acquire, contended acquired, contended unlock
249 // We should also add DTRACE probes in the ParkEvent subsystem for
250 // Park-entry, Park-exit, and Unpark.
251 //
252 // -- We have an excess of mutex-like constructs in the JVM, namely:
253 // 1. objectMonitors for Java-level synchronization (synchronizer.cpp)
254 // 2. low-level muxAcquire and muxRelease
255 // 3. low-level spinAcquire and spinRelease
256 // 4. native Mutex:: and Monitor::
257 // 5. jvm_raw_lock() and _unlock()
258 // 6. JVMTI raw monitors -- distinct from (5) despite having a confusingly
259 // similar name.
260 //
261 // o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o-o
262
263
264 // CASPTR() uses the canonical argument order that dominates in the literature.
265 // Our internal cmpxchg_ptr() uses a bastardized ordering to accommodate Sun .il templates.
266
267 #define CASPTR(a,c,s) intptr_t(Atomic::cmpxchg_ptr ((void *)(s),(void *)(a),(void *)(c)))
268 #define UNS(x) (uintptr_t(x))
269 #define TRACE(m) { static volatile int ctr = 0 ; int x = ++ctr ; if ((x & (x-1))==0) { ::printf ("%d:%s\n", x, #m); ::fflush(stdout); }}
270
271 // Simplistic low-quality Marsaglia SHIFT-XOR RNG.
272 // Bijective except for the trailing mask operation.
273 // Useful for spin loops as the compiler can't optimize it away.
274
275 static inline jint MarsagliaXORV (jint x) {
276 if (x == 0) x = 1|os::random() ;
277 x ^= x << 6;
278 x ^= ((unsigned)x) >> 21;
279 x ^= x << 7 ;
280 return x & 0x7FFFFFFF ;
281 }
282
283 static inline jint MarsagliaXOR (jint * const a) {
284 jint x = *a ;
285 if (x == 0) x = UNS(a)|1 ;
286 x ^= x << 6;
287 x ^= ((unsigned)x) >> 21;
288 x ^= x << 7 ;
289 *a = x ;
290 return x & 0x7FFFFFFF ;
291 }
292
293 static int Stall (int its) {
294 static volatile jint rv = 1 ;
295 volatile int OnFrame = 0 ;
296 jint v = rv ^ UNS(OnFrame) ;
297 while (--its >= 0) {
298 v = MarsagliaXORV (v) ;
299 }
300 // Make this impossible for the compiler to optimize away,
301 // but (mostly) avoid W coherency sharing on MP systems.
302 if (v == 0x12345) rv = v ;
303 return v ;
304 }
305
306 int Monitor::TryLock () {
307 intptr_t v = _LockWord.FullWord ;
308 for (;;) {
309 if ((v & _LBIT) != 0) return 0 ;
310 const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ;
311 if (v == u) return 1 ;
312 v = u ;
313 }
314 }
315
316 int Monitor::TryFast () {
317 // Optimistic fast-path form ...
318 // Fast-path attempt for the common uncontended case.
319 // Avoid RTS->RTO $ coherence upgrade on typical SMP systems.
320 intptr_t v = CASPTR (&_LockWord, 0, _LBIT) ; // agro ...
321 if (v == 0) return 1 ;
322
323 for (;;) {
324 if ((v & _LBIT) != 0) return 0 ;
325 const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ;
326 if (v == u) return 1 ;
327 v = u ;
328 }
329 }
330
331 int Monitor::ILocked () {
332 const intptr_t w = _LockWord.FullWord & 0xFF ;
333 assert (w == 0 || w == _LBIT, "invariant") ;
334 return w == _LBIT ;
335 }
336
337 // Polite TATAS spinlock with exponential backoff - bounded spin.
338 // Ideally we'd use processor cycles, time or vtime to control
339 // the loop, but we currently use iterations.
340 // All the constants within were derived empirically but work over
341 // over the spectrum of J2SE reference platforms.
342 // On Niagara-class systems the back-off is unnecessary but
343 // is relatively harmless. (At worst it'll slightly retard
344 // acquisition times). The back-off is critical for older SMP systems
345 // where constant fetching of the LockWord would otherwise impair
346 // scalability.
347 //
348 // Clamp spinning at approximately 1/2 of a context-switch round-trip.
349 // See synchronizer.cpp for details and rationale.
350
351 int Monitor::TrySpin (Thread * const Self) {
352 if (TryLock()) return 1 ;
353 if (!os::is_MP()) return 0 ;
354
355 int Probes = 0 ;
356 int Delay = 0 ;
357 int Steps = 0 ;
358 int SpinMax = NativeMonitorSpinLimit ;
359 int flgs = NativeMonitorFlags ;
360 for (;;) {
361 intptr_t v = _LockWord.FullWord;
362 if ((v & _LBIT) == 0) {
363 if (CASPTR (&_LockWord, v, v|_LBIT) == v) {
364 return 1 ;
365 }
366 continue ;
367 }
368
369 if ((flgs & 8) == 0) {
370 SpinPause () ;
371 }
372
373 // Periodically increase Delay -- variable Delay form
374 // conceptually: delay *= 1 + 1/Exponent
375 ++ Probes;
376 if (Probes > SpinMax) return 0 ;
377
378 if ((Probes & 0x7) == 0) {
379 Delay = ((Delay << 1)|1) & 0x7FF ;
380 // CONSIDER: Delay += 1 + (Delay/4); Delay &= 0x7FF ;
381 }
382
383 if (flgs & 2) continue ;
384
385 // Consider checking _owner's schedctl state, if OFFPROC abort spin.
386 // If the owner is OFFPROC then it's unlike that the lock will be dropped
387 // in a timely fashion, which suggests that spinning would not be fruitful
388 // or profitable.
389
390 // Stall for "Delay" time units - iterations in the current implementation.
391 // Avoid generating coherency traffic while stalled.
392 // Possible ways to delay:
393 // PAUSE, SLEEP, MEMBAR #sync, MEMBAR #halt,
394 // wr %g0,%asi, gethrtime, rdstick, rdtick, rdtsc, etc. ...
395 // Note that on Niagara-class systems we want to minimize STs in the
396 // spin loop. N1 and brethren write-around the L1$ over the xbar into the L2$.
397 // Furthermore, they don't have a W$ like traditional SPARC processors.
398 // We currently use a Marsaglia Shift-Xor RNG loop.
399 Steps += Delay ;
400 if (Self != NULL) {
401 jint rv = Self->rng[0] ;
402 for (int k = Delay ; --k >= 0; ) {
403 rv = MarsagliaXORV (rv) ;
404 if ((flgs & 4) == 0 && SafepointSynchronize::do_call_back()) return 0 ;
405 }
406 Self->rng[0] = rv ;
407 } else {
408 Stall (Delay) ;
409 }
410 }
411 }
412
413 static int ParkCommon (ParkEvent * ev, jlong timo) {
414 // Diagnostic support - periodically unwedge blocked threads
415 intx nmt = NativeMonitorTimeout ;
416 if (nmt > 0 && (nmt < timo || timo <= 0)) {
417 timo = nmt ;
418 }
419 int err = OS_OK ;
420 if (0 == timo) {
421 ev->park() ;
422 } else {
423 err = ev->park(timo) ;
424 }
425 return err ;
426 }
427
428 inline int Monitor::AcquireOrPush (ParkEvent * ESelf) {
429 intptr_t v = _LockWord.FullWord ;
430 for (;;) {
431 if ((v & _LBIT) == 0) {
432 const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ;
433 if (u == v) return 1 ; // indicate acquired
434 v = u ;
435 } else {
436 // Anticipate success ...
437 ESelf->ListNext = (ParkEvent *) (v & ~_LBIT) ;
438 const intptr_t u = CASPTR (&_LockWord, v, intptr_t(ESelf)|_LBIT) ;
439 if (u == v) return 0 ; // indicate pushed onto cxq
440 v = u ;
441 }
442 // Interference - LockWord change - just retry
443 }
444 }
445
446 // ILock and IWait are the lowest level primitive internal blocking
447 // synchronization functions. The callers of IWait and ILock must have
448 // performed any needed state transitions beforehand.
449 // IWait and ILock may directly call park() without any concern for thread state.
450 // Note that ILock and IWait do *not* access _owner.
451 // _owner is a higher-level logical concept.
452
453 void Monitor::ILock (Thread * Self) {
454 assert (_OnDeck != Self->_MutexEvent, "invariant") ;
455
456 if (TryFast()) {
457 Exeunt:
458 assert (ILocked(), "invariant") ;
459 return ;
460 }
461
462 ParkEvent * const ESelf = Self->_MutexEvent ;
463 assert (_OnDeck != ESelf, "invariant") ;
464
465 // As an optimization, spinners could conditionally try to set ONDECK to _LBIT
466 // Synchronizer.cpp uses a similar optimization.
467 if (TrySpin (Self)) goto Exeunt ;
468
469 // Slow-path - the lock is contended.
470 // Either Enqueue Self on cxq or acquire the outer lock.
471 // LockWord encoding = (cxq,LOCKBYTE)
472 ESelf->reset() ;
473 OrderAccess::fence() ;
474
475 // Optional optimization ... try barging on the inner lock
476 if ((NativeMonitorFlags & 32) && CASPTR (&_OnDeck, NULL, UNS(Self)) == 0) {
477 goto OnDeck_LOOP ;
478 }
479
480 if (AcquireOrPush (ESelf)) goto Exeunt ;
481
482 // At any given time there is at most one ondeck thread.
483 // ondeck implies not resident on cxq and not resident on EntryList
484 // Only the OnDeck thread can try to acquire -- contended for -- the lock.
485 // CONSIDER: use Self->OnDeck instead of m->OnDeck.
486 // Deschedule Self so that others may run.
487 while (_OnDeck != ESelf) {
488 ParkCommon (ESelf, 0) ;
489 }
490
491 // Self is now in the ONDECK position and will remain so until it
492 // manages to acquire the lock.
493 OnDeck_LOOP:
494 for (;;) {
495 assert (_OnDeck == ESelf, "invariant") ;
496 if (TrySpin (Self)) break ;
497 // CONSIDER: if ESelf->TryPark() && TryLock() break ...
498 // It's probably wise to spin only if we *actually* blocked
499 // CONSIDER: check the lockbyte, if it remains set then
500 // preemptively drain the cxq into the EntryList.
501 // The best place and time to perform queue operations -- lock metadata --
502 // is _before having acquired the outer lock, while waiting for the lock to drop.
503 ParkCommon (ESelf, 0) ;
504 }
505
506 assert (_OnDeck == ESelf, "invariant") ;
507 _OnDeck = NULL ;
508
509 // Note that we current drop the inner lock (clear OnDeck) in the slow-path
510 // epilog immediately after having acquired the outer lock.
511 // But instead we could consider the following optimizations:
512 // A. Shift or defer dropping the inner lock until the subsequent IUnlock() operation.
513 // This might avoid potential reacquisition of the inner lock in IUlock().
514 // B. While still holding the inner lock, attempt to opportunistically select
515 // and unlink the next ONDECK thread from the EntryList.
516 // If successful, set ONDECK to refer to that thread, otherwise clear ONDECK.
517 // It's critical that the select-and-unlink operation run in constant-time as
518 // it executes when holding the outer lock and may artificially increase the
519 // effective length of the critical section.
520 // Note that (A) and (B) are tantamount to succession by direct handoff for
521 // the inner lock.
522 goto Exeunt ;
523 }
524
525 void Monitor::IUnlock (bool RelaxAssert) {
526 assert (ILocked(), "invariant") ;
527 // Conceptually we need a MEMBAR #storestore|#loadstore barrier or fence immediately
528 // before the store that releases the lock. Crucially, all the stores and loads in the
529 // critical section must be globally visible before the store of 0 into the lock-word
530 // that releases the lock becomes globally visible. That is, memory accesses in the
531 // critical section should not be allowed to bypass or overtake the following ST that
532 // releases the lock. As such, to prevent accesses within the critical section
533 // from "leaking" out, we need a release fence between the critical section and the
534 // store that releases the lock. In practice that release barrier is elided on
535 // platforms with strong memory models such as TSO.
536 //
537 // Note that the OrderAccess::storeload() fence that appears after unlock store
538 // provides for progress conditions and succession and is _not related to exclusion
539 // safety or lock release consistency.
540 OrderAccess::release_store(&_LockWord.Bytes[_LSBINDEX], 0); // drop outer lock
541
542 OrderAccess::storeload ();
543 ParkEvent * const w = _OnDeck ;
544 assert (RelaxAssert || w != Thread::current()->_MutexEvent, "invariant") ;
545 if (w != NULL) {
546 // Either we have a valid ondeck thread or ondeck is transiently "locked"
547 // by some exiting thread as it arranges for succession. The LSBit of
548 // OnDeck allows us to discriminate two cases. If the latter, the
549 // responsibility for progress and succession lies with that other thread.
550 // For good performance, we also depend on the fact that redundant unpark()
551 // operations are cheap. That is, repeated Unpark()ing of the ONDECK thread
552 // is inexpensive. This approach provides implicit futile wakeup throttling.
553 // Note that the referent "w" might be stale with respect to the lock.
554 // In that case the following unpark() is harmless and the worst that'll happen
555 // is a spurious return from a park() operation. Critically, if "w" _is stale,
556 // then progress is known to have occurred as that means the thread associated
557 // with "w" acquired the lock. In that case this thread need take no further
558 // action to guarantee progress.
559 if ((UNS(w) & _LBIT) == 0) w->unpark() ;
560 return ;
561 }
562
563 intptr_t cxq = _LockWord.FullWord ;
564 if (((cxq & ~_LBIT)|UNS(_EntryList)) == 0) {
565 return ; // normal fast-path exit - cxq and EntryList both empty
566 }
567 if (cxq & _LBIT) {
568 // Optional optimization ...
569 // Some other thread acquired the lock in the window since this
570 // thread released it. Succession is now that thread's responsibility.
571 return ;
572 }
573
574 Succession:
575 // Slow-path exit - this thread must ensure succession and progress.
576 // OnDeck serves as lock to protect cxq and EntryList.
577 // Only the holder of OnDeck can manipulate EntryList or detach the RATs from cxq.
578 // Avoid ABA - allow multiple concurrent producers (enqueue via push-CAS)
579 // but only one concurrent consumer (detacher of RATs).
580 // Consider protecting this critical section with schedctl on Solaris.
581 // Unlike a normal lock, however, the exiting thread "locks" OnDeck,
582 // picks a successor and marks that thread as OnDeck. That successor
583 // thread will then clear OnDeck once it eventually acquires the outer lock.
584 if (CASPTR (&_OnDeck, NULL, _LBIT) != UNS(NULL)) {
585 return ;
586 }
587
588 ParkEvent * List = _EntryList ;
589 if (List != NULL) {
590 // Transfer the head of the EntryList to the OnDeck position.
591 // Once OnDeck, a thread stays OnDeck until it acquires the lock.
592 // For a given lock there is at most OnDeck thread at any one instant.
593 WakeOne:
594 assert (List == _EntryList, "invariant") ;
595 ParkEvent * const w = List ;
596 assert (RelaxAssert || w != Thread::current()->_MutexEvent, "invariant") ;
597 _EntryList = w->ListNext ;
598 // as a diagnostic measure consider setting w->_ListNext = BAD
599 assert (UNS(_OnDeck) == _LBIT, "invariant") ;
600 _OnDeck = w ; // pass OnDeck to w.
601 // w will clear OnDeck once it acquires the outer lock
602
603 // Another optional optimization ...
604 // For heavily contended locks it's not uncommon that some other
605 // thread acquired the lock while this thread was arranging succession.
606 // Try to defer the unpark() operation - Delegate the responsibility
607 // for unpark()ing the OnDeck thread to the current or subsequent owners
608 // That is, the new owner is responsible for unparking the OnDeck thread.
609 OrderAccess::storeload() ;
610 cxq = _LockWord.FullWord ;
611 if (cxq & _LBIT) return ;
612
613 w->unpark() ;
614 return ;
615 }
616
617 cxq = _LockWord.FullWord ;
618 if ((cxq & ~_LBIT) != 0) {
619 // The EntryList is empty but the cxq is populated.
620 // drain RATs from cxq into EntryList
621 // Detach RATs segment with CAS and then merge into EntryList
622 for (;;) {
623 // optional optimization - if locked, the owner is responsible for succession
624 if (cxq & _LBIT) goto Punt ;
625 const intptr_t vfy = CASPTR (&_LockWord, cxq, cxq & _LBIT) ;
626 if (vfy == cxq) break ;
627 cxq = vfy ;
628 // Interference - LockWord changed - Just retry
629 // We can see concurrent interference from contending threads
630 // pushing themselves onto the cxq or from lock-unlock operations.
631 // From the perspective of this thread, EntryList is stable and
632 // the cxq is prepend-only -- the head is volatile but the interior
633 // of the cxq is stable. In theory if we encounter interference from threads
634 // pushing onto cxq we could simply break off the original cxq suffix and
635 // move that segment to the EntryList, avoiding a 2nd or multiple CAS attempts
636 // on the high-traffic LockWord variable. For instance lets say the cxq is "ABCD"
637 // when we first fetch cxq above. Between the fetch -- where we observed "A"
638 // -- and CAS -- where we attempt to CAS null over A -- "PQR" arrive,
639 // yielding cxq = "PQRABCD". In this case we could simply set A.ListNext
640 // null, leaving cxq = "PQRA" and transfer the "BCD" segment to the EntryList.
641 // Note too, that it's safe for this thread to traverse the cxq
642 // without taking any special concurrency precautions.
643 }
644
645 // We don't currently reorder the cxq segment as we move it onto
646 // the EntryList, but it might make sense to reverse the order
647 // or perhaps sort by thread priority. See the comments in
648 // synchronizer.cpp objectMonitor::exit().
649 assert (_EntryList == NULL, "invariant") ;
650 _EntryList = List = (ParkEvent *)(cxq & ~_LBIT) ;
651 assert (List != NULL, "invariant") ;
652 goto WakeOne ;
653 }
654
655 // cxq|EntryList is empty.
656 // w == NULL implies that cxq|EntryList == NULL in the past.
657 // Possible race - rare inopportune interleaving.
658 // A thread could have added itself to cxq since this thread previously checked.
659 // Detect and recover by refetching cxq.
660 Punt:
661 assert (UNS(_OnDeck) == _LBIT, "invariant") ;
662 _OnDeck = NULL ; // Release inner lock.
663 OrderAccess::storeload(); // Dekker duality - pivot point
664
665 // Resample LockWord/cxq to recover from possible race.
666 // For instance, while this thread T1 held OnDeck, some other thread T2 might
667 // acquire the outer lock. Another thread T3 might try to acquire the outer
668 // lock, but encounter contention and enqueue itself on cxq. T2 then drops the
669 // outer lock, but skips succession as this thread T1 still holds OnDeck.
670 // T1 is and remains responsible for ensuring succession of T3.
671 //
672 // Note that we don't need to recheck EntryList, just cxq.
673 // If threads moved onto EntryList since we dropped OnDeck
674 // that implies some other thread forced succession.
675 cxq = _LockWord.FullWord ;
676 if ((cxq & ~_LBIT) != 0 && (cxq & _LBIT) == 0) {
677 goto Succession ; // potential race -- re-run succession
678 }
679 return ;
680 }
681
682 bool Monitor::notify() {
683 assert (_owner == Thread::current(), "invariant") ;
684 assert (ILocked(), "invariant") ;
685 if (_WaitSet == NULL) return true ;
686 NotifyCount ++ ;
687
688 // Transfer one thread from the WaitSet to the EntryList or cxq.
689 // Currently we just unlink the head of the WaitSet and prepend to the cxq.
690 // And of course we could just unlink it and unpark it, too, but
691 // in that case it'd likely impale itself on the reentry.
692 Thread::muxAcquire (_WaitLock, "notify:WaitLock") ;
693 ParkEvent * nfy = _WaitSet ;
694 if (nfy != NULL) { // DCL idiom
695 _WaitSet = nfy->ListNext ;
696 assert (nfy->Notified == 0, "invariant") ;
697 // push nfy onto the cxq
698 for (;;) {
699 const intptr_t v = _LockWord.FullWord ;
700 assert ((v & 0xFF) == _LBIT, "invariant") ;
701 nfy->ListNext = (ParkEvent *)(v & ~_LBIT);
702 if (CASPTR (&_LockWord, v, UNS(nfy)|_LBIT) == v) break;
703 // interference - _LockWord changed -- just retry
704 }
705 // Note that setting Notified before pushing nfy onto the cxq is
706 // also legal and safe, but the safety properties are much more
707 // subtle, so for the sake of code stewardship ...
708 OrderAccess::fence() ;
709 nfy->Notified = 1;
710 }
711 Thread::muxRelease (_WaitLock) ;
712 if (nfy != NULL && (NativeMonitorFlags & 16)) {
713 // Experimental code ... light up the wakee in the hope that this thread (the owner)
714 // will drop the lock just about the time the wakee comes ONPROC.
715 nfy->unpark() ;
716 }
717 assert (ILocked(), "invariant") ;
718 return true ;
719 }
720
721 // Currently notifyAll() transfers the waiters one-at-a-time from the waitset
722 // to the cxq. This could be done more efficiently with a single bulk en-mass transfer,
723 // but in practice notifyAll() for large #s of threads is rare and not time-critical.
724 // Beware too, that we invert the order of the waiters. Lets say that the
725 // waitset is "ABCD" and the cxq is "XYZ". After a notifyAll() the waitset
726 // will be empty and the cxq will be "DCBAXYZ". This is benign, of course.
727
728 bool Monitor::notify_all() {
729 assert (_owner == Thread::current(), "invariant") ;
730 assert (ILocked(), "invariant") ;
731 while (_WaitSet != NULL) notify() ;
732 return true ;
733 }
734
735 int Monitor::IWait (Thread * Self, jlong timo) {
736 assert (ILocked(), "invariant") ;
737
738 // Phases:
739 // 1. Enqueue Self on WaitSet - currently prepend
740 // 2. unlock - drop the outer lock
741 // 3. wait for either notification or timeout
742 // 4. lock - reentry - reacquire the outer lock
743
744 ParkEvent * const ESelf = Self->_MutexEvent ;
745 ESelf->Notified = 0 ;
746 ESelf->reset() ;
747 OrderAccess::fence() ;
748
749 // Add Self to WaitSet
750 // Ideally only the holder of the outer lock would manipulate the WaitSet -
751 // That is, the outer lock would implicitly protect the WaitSet.
752 // But if a thread in wait() encounters a timeout it will need to dequeue itself
753 // from the WaitSet _before it becomes the owner of the lock. We need to dequeue
754 // as the ParkEvent -- which serves as a proxy for the thread -- can't reside
755 // on both the WaitSet and the EntryList|cxq at the same time.. That is, a thread
756 // on the WaitSet can't be allowed to compete for the lock until it has managed to
757 // unlink its ParkEvent from WaitSet. Thus the need for WaitLock.
758 // Contention on the WaitLock is minimal.
759 //
760 // Another viable approach would be add another ParkEvent, "WaitEvent" to the
761 // thread class. The WaitSet would be composed of WaitEvents. Only the
762 // owner of the outer lock would manipulate the WaitSet. A thread in wait()
763 // could then compete for the outer lock, and then, if necessary, unlink itself
764 // from the WaitSet only after having acquired the outer lock. More precisely,
765 // there would be no WaitLock. A thread in in wait() would enqueue its WaitEvent
766 // on the WaitSet; release the outer lock; wait for either notification or timeout;
767 // reacquire the inner lock; and then, if needed, unlink itself from the WaitSet.
768 //
769 // Alternatively, a 2nd set of list link fields in the ParkEvent might suffice.
770 // One set would be for the WaitSet and one for the EntryList.
771 // We could also deconstruct the ParkEvent into a "pure" event and add a
772 // new immortal/TSM "ListElement" class that referred to ParkEvents.
773 // In that case we could have one ListElement on the WaitSet and another
774 // on the EntryList, with both referring to the same pure Event.
775
776 Thread::muxAcquire (_WaitLock, "wait:WaitLock:Add") ;
777 ESelf->ListNext = _WaitSet ;
778 _WaitSet = ESelf ;
779 Thread::muxRelease (_WaitLock) ;
780
781 // Release the outer lock
782 // We call IUnlock (RelaxAssert=true) as a thread T1 might
783 // enqueue itself on the WaitSet, call IUnlock(), drop the lock,
784 // and then stall before it can attempt to wake a successor.
785 // Some other thread T2 acquires the lock, and calls notify(), moving
786 // T1 from the WaitSet to the cxq. T2 then drops the lock. T1 resumes,
787 // and then finds *itself* on the cxq. During the course of a normal
788 // IUnlock() call a thread should _never find itself on the EntryList
789 // or cxq, but in the case of wait() it's possible.
790 // See synchronizer.cpp objectMonitor::wait().
791 IUnlock (true) ;
792
793 // Wait for either notification or timeout
794 // Beware that in some circumstances we might propagate
795 // spurious wakeups back to the caller.
796
797 for (;;) {
798 if (ESelf->Notified) break ;
799 int err = ParkCommon (ESelf, timo) ;
800 if (err == OS_TIMEOUT || (NativeMonitorFlags & 1)) break ;
801 }
802
803 // Prepare for reentry - if necessary, remove ESelf from WaitSet
804 // ESelf can be:
805 // 1. Still on the WaitSet. This can happen if we exited the loop by timeout.
806 // 2. On the cxq or EntryList
807 // 3. Not resident on cxq, EntryList or WaitSet, but in the OnDeck position.
808
809 OrderAccess::fence() ;
810 int WasOnWaitSet = 0 ;
811 if (ESelf->Notified == 0) {
812 Thread::muxAcquire (_WaitLock, "wait:WaitLock:remove") ;
813 if (ESelf->Notified == 0) { // DCL idiom
814 assert (_OnDeck != ESelf, "invariant") ; // can't be both OnDeck and on WaitSet
815 // ESelf is resident on the WaitSet -- unlink it.
816 // A doubly-linked list would be better here so we can unlink in constant-time.
817 // We have to unlink before we potentially recontend as ESelf might otherwise
818 // end up on the cxq|EntryList -- it can't be on two lists at once.
819 ParkEvent * p = _WaitSet ;
820 ParkEvent * q = NULL ; // classic q chases p
821 while (p != NULL && p != ESelf) {
822 q = p ;
823 p = p->ListNext ;
824 }
825 assert (p == ESelf, "invariant") ;
826 if (p == _WaitSet) { // found at head
827 assert (q == NULL, "invariant") ;
828 _WaitSet = p->ListNext ;
829 } else { // found in interior
830 assert (q->ListNext == p, "invariant") ;
831 q->ListNext = p->ListNext ;
832 }
833 WasOnWaitSet = 1 ; // We were *not* notified but instead encountered timeout
834 }
835 Thread::muxRelease (_WaitLock) ;
836 }
837
838 // Reentry phase - reacquire the lock
839 if (WasOnWaitSet) {
840 // ESelf was previously on the WaitSet but we just unlinked it above
841 // because of a timeout. ESelf is not resident on any list and is not OnDeck
842 assert (_OnDeck != ESelf, "invariant") ;
843 ILock (Self) ;
844 } else {
845 // A prior notify() operation moved ESelf from the WaitSet to the cxq.
846 // ESelf is now on the cxq, EntryList or at the OnDeck position.
847 // The following fragment is extracted from Monitor::ILock()
848 for (;;) {
849 if (_OnDeck == ESelf && TrySpin(Self)) break ;
850 ParkCommon (ESelf, 0) ;
851 }
852 assert (_OnDeck == ESelf, "invariant") ;
853 _OnDeck = NULL ;
854 }
855
856 assert (ILocked(), "invariant") ;
857 return WasOnWaitSet != 0 ; // return true IFF timeout
858 }
859
860
861 // ON THE VMTHREAD SNEAKING PAST HELD LOCKS:
862 // In particular, there are certain types of global lock that may be held
863 // by a Java thread while it is blocked at a safepoint but before it has
864 // written the _owner field. These locks may be sneakily acquired by the
865 // VM thread during a safepoint to avoid deadlocks. Alternatively, one should
866 // identify all such locks, and ensure that Java threads never block at
867 // safepoints while holding them (_no_safepoint_check_flag). While it
868 // seems as though this could increase the time to reach a safepoint
869 // (or at least increase the mean, if not the variance), the latter
870 // approach might make for a cleaner, more maintainable JVM design.
871 //
872 // Sneaking is vile and reprehensible and should be excised at the 1st
873 // opportunity. It's possible that the need for sneaking could be obviated
874 // as follows. Currently, a thread might (a) while TBIVM, call pthread_mutex_lock
875 // or ILock() thus acquiring the "physical" lock underlying Monitor/Mutex.
876 // (b) stall at the TBIVM exit point as a safepoint is in effect. Critically,
877 // it'll stall at the TBIVM reentry state transition after having acquired the
878 // underlying lock, but before having set _owner and having entered the actual
879 // critical section. The lock-sneaking facility leverages that fact and allowed the
880 // VM thread to logically acquire locks that had already be physically locked by mutators
881 // but where mutators were known blocked by the reentry thread state transition.
882 //
883 // If we were to modify the Monitor-Mutex so that TBIVM state transitions tightly
884 // wrapped calls to park(), then we could likely do away with sneaking. We'd
885 // decouple lock acquisition and parking. The critical invariant to eliminating
886 // sneaking is to ensure that we never "physically" acquire the lock while TBIVM.
887 // An easy way to accomplish this is to wrap the park calls in a narrow TBIVM jacket.
888 // One difficulty with this approach is that the TBIVM wrapper could recurse and
889 // call lock() deep from within a lock() call, while the MutexEvent was already enqueued.
890 // Using a stack (N=2 at minimum) of ParkEvents would take care of that problem.
891 //
892 // But of course the proper ultimate approach is to avoid schemes that require explicit
893 // sneaking or dependence on any any clever invariants or subtle implementation properties
894 // of Mutex-Monitor and instead directly address the underlying design flaw.
895
896 void Monitor::lock (Thread * Self) {
897 #ifdef CHECK_UNHANDLED_OOPS
898 // Clear unhandled oops so we get a crash right away. Only clear for non-vm
899 // or GC threads.
900 if (Self->is_Java_thread()) {
901 Self->clear_unhandled_oops();
902 }
903 #endif // CHECK_UNHANDLED_OOPS
904
905 debug_only(check_prelock_state(Self));
906 assert (_owner != Self , "invariant") ;
907 assert (_OnDeck != Self->_MutexEvent, "invariant") ;
908
909 if (TryFast()) {
910 Exeunt:
911 assert (ILocked(), "invariant") ;
912 assert (owner() == NULL, "invariant");
913 set_owner (Self);
914 return ;
915 }
916
917 // The lock is contended ...
918
919 bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint();
920 if (can_sneak && _owner == NULL) {
921 // a java thread has locked the lock but has not entered the
922 // critical region -- let's just pretend we've locked the lock
923 // and go on. we note this with _snuck so we can also
924 // pretend to unlock when the time comes.
925 _snuck = true;
926 goto Exeunt ;
927 }
928
929 // Try a brief spin to avoid passing thru thread state transition ...
930 if (TrySpin (Self)) goto Exeunt ;
931
932 check_block_state(Self);
933 if (Self->is_Java_thread()) {
934 // Horribile dictu - we suffer through a state transition
935 assert(rank() > Mutex::special, "Potential deadlock with special or lesser rank mutex");
936 ThreadBlockInVM tbivm ((JavaThread *) Self) ;
937 ILock (Self) ;
938 } else {
939 // Mirabile dictu
940 ILock (Self) ;
941 }
942 goto Exeunt ;
943 }
944
945 void Monitor::lock() {
946 this->lock(Thread::current());
947 }
948
949 // Lock without safepoint check - a degenerate variant of lock().
950 // Should ONLY be used by safepoint code and other code
951 // that is guaranteed not to block while running inside the VM. If this is called with
952 // thread state set to be in VM, the safepoint synchronization code will deadlock!
953
954 void Monitor::lock_without_safepoint_check (Thread * Self) {
955 assert (_owner != Self, "invariant") ;
956 ILock (Self) ;
957 assert (_owner == NULL, "invariant");
958 set_owner (Self);
959 }
960
961 void Monitor::lock_without_safepoint_check () {
962 lock_without_safepoint_check (Thread::current()) ;
963 }
964
965
966 // Returns true if thread succeceed [sic] in grabbing the lock, otherwise false.
967
968 bool Monitor::try_lock() {
969 Thread * const Self = Thread::current();
970 debug_only(check_prelock_state(Self));
971 // assert(!thread->is_inside_signal_handler(), "don't lock inside signal handler");
972
973 // Special case, where all Java threads are stopped.
974 // The lock may have been acquired but _owner is not yet set.
975 // In that case the VM thread can safely grab the lock.
976 // It strikes me this should appear _after the TryLock() fails, below.
977 bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint();
978 if (can_sneak && _owner == NULL) {
979 set_owner(Self); // Do not need to be atomic, since we are at a safepoint
980 _snuck = true;
981 return true;
982 }
983
984 if (TryLock()) {
985 // We got the lock
986 assert (_owner == NULL, "invariant");
987 set_owner (Self);
988 return true;
989 }
990 return false;
991 }
992
993 void Monitor::unlock() {
994 assert (_owner == Thread::current(), "invariant") ;
995 assert (_OnDeck != Thread::current()->_MutexEvent , "invariant") ;
996 set_owner (NULL) ;
997 if (_snuck) {
998 assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak");
999 _snuck = false;
1000 return ;
1001 }
1002 IUnlock (false) ;
1003 }
1004
1005 // Yet another degenerate version of Monitor::lock() or lock_without_safepoint_check()
1006 // jvm_raw_lock() and _unlock() can be called by non-Java threads via JVM_RawMonitorEnter.
1007 //
1008 // There's no expectation that JVM_RawMonitors will interoperate properly with the native
1009 // Mutex-Monitor constructs. We happen to implement JVM_RawMonitors in terms of
1010 // native Mutex-Monitors simply as a matter of convenience. A simple abstraction layer
1011 // over a pthread_mutex_t would work equally as well, but require more platform-specific
1012 // code -- a "PlatformMutex". Alternatively, a simply layer over muxAcquire-muxRelease
1013 // would work too.
1014 //
1015 // Since the caller might be a foreign thread, we don't necessarily have a Thread.MutexEvent
1016 // instance available. Instead, we transiently allocate a ParkEvent on-demand if
1017 // we encounter contention. That ParkEvent remains associated with the thread
1018 // until it manages to acquire the lock, at which time we return the ParkEvent
1019 // to the global ParkEvent free list. This is correct and suffices for our purposes.
1020 //
1021 // Beware that the original jvm_raw_unlock() had a "_snuck" test but that
1022 // jvm_raw_lock() didn't have the corresponding test. I suspect that's an
1023 // oversight, but I've replicated the original suspect logic in the new code ...
1024
1025 void Monitor::jvm_raw_lock() {
1026 assert(rank() == native, "invariant");
1027
1028 if (TryLock()) {
1029 Exeunt:
1030 assert (ILocked(), "invariant") ;
1031 assert (_owner == NULL, "invariant");
1032 // This can potentially be called by non-java Threads. Thus, the ThreadLocalStorage
1033 // might return NULL. Don't call set_owner since it will break on an NULL owner
1034 // Consider installing a non-null "ANON" distinguished value instead of just NULL.
1035 _owner = ThreadLocalStorage::thread();
1036 return ;
1037 }
1038
1039 if (TrySpin(NULL)) goto Exeunt ;
1040
1041 // slow-path - apparent contention
1042 // Allocate a ParkEvent for transient use.
1043 // The ParkEvent remains associated with this thread until
1044 // the time the thread manages to acquire the lock.
1045 ParkEvent * const ESelf = ParkEvent::Allocate(NULL) ;
1046 ESelf->reset() ;
1047 OrderAccess::storeload() ;
1048
1049 // Either Enqueue Self on cxq or acquire the outer lock.
1050 if (AcquireOrPush (ESelf)) {
1051 ParkEvent::Release (ESelf) ; // surrender the ParkEvent
1052 goto Exeunt ;
1053 }
1054
1055 // At any given time there is at most one ondeck thread.
1056 // ondeck implies not resident on cxq and not resident on EntryList
1057 // Only the OnDeck thread can try to acquire -- contended for -- the lock.
1058 // CONSIDER: use Self->OnDeck instead of m->OnDeck.
1059 for (;;) {
1060 if (_OnDeck == ESelf && TrySpin(NULL)) break ;
1061 ParkCommon (ESelf, 0) ;
1062 }
1063
1064 assert (_OnDeck == ESelf, "invariant") ;
1065 _OnDeck = NULL ;
1066 ParkEvent::Release (ESelf) ; // surrender the ParkEvent
1067 goto Exeunt ;
1068 }
1069
1070 void Monitor::jvm_raw_unlock() {
1071 // Nearly the same as Monitor::unlock() ...
1072 // directly set _owner instead of using set_owner(null)
1073 _owner = NULL ;
1074 if (_snuck) { // ???
1075 assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak");
1076 _snuck = false;
1077 return ;
1078 }
1079 IUnlock(false) ;
1080 }
1081
1082 bool Monitor::wait(bool no_safepoint_check, long timeout, bool as_suspend_equivalent) {
1083 Thread * const Self = Thread::current() ;
1084 assert (_owner == Self, "invariant") ;
1085 assert (ILocked(), "invariant") ;
1086
1087 // as_suspend_equivalent logically implies !no_safepoint_check
1088 guarantee (!as_suspend_equivalent || !no_safepoint_check, "invariant") ;
1089 // !no_safepoint_check logically implies java_thread
1090 guarantee (no_safepoint_check || Self->is_Java_thread(), "invariant") ;
1091
1092 #ifdef ASSERT
1093 Monitor * least = get_least_ranked_lock_besides_this(Self->owned_locks());
1094 assert(least != this, "Specification of get_least_... call above");
1095 if (least != NULL && least->rank() <= special) {
1096 tty->print("Attempting to wait on monitor %s/%d while holding"
1097 " lock %s/%d -- possible deadlock",
1098 name(), rank(), least->name(), least->rank());
1099 assert(false, "Shouldn't block(wait) while holding a lock of rank special");
1100 }
1101 #endif // ASSERT
1102
1103 int wait_status ;
1104 // conceptually set the owner to NULL in anticipation of
1105 // abdicating the lock in wait
1106 set_owner(NULL);
1107 if (no_safepoint_check) {
1108 wait_status = IWait (Self, timeout) ;
1109 } else {
1110 assert (Self->is_Java_thread(), "invariant") ;
1111 JavaThread *jt = (JavaThread *)Self;
1112
1113 // Enter safepoint region - ornate and Rococo ...
1114 ThreadBlockInVM tbivm(jt);
1115 OSThreadWaitState osts(Self->osthread(), false /* not Object.wait() */);
1116
1117 if (as_suspend_equivalent) {
1118 jt->set_suspend_equivalent();
1119 // cleared by handle_special_suspend_equivalent_condition() or
1120 // java_suspend_self()
1121 }
1122
1123 wait_status = IWait (Self, timeout) ;
1124
1125 // were we externally suspended while we were waiting?
1126 if (as_suspend_equivalent && jt->handle_special_suspend_equivalent_condition()) {
1127 // Our event wait has finished and we own the lock, but
1128 // while we were waiting another thread suspended us. We don't
1129 // want to hold the lock while suspended because that
1130 // would surprise the thread that suspended us.
1131 assert (ILocked(), "invariant") ;
1132 IUnlock (true) ;
1133 jt->java_suspend_self();
1134 ILock (Self) ;
1135 assert (ILocked(), "invariant") ;
1136 }
1137 }
1138
1139 // Conceptually reestablish ownership of the lock.
1140 // The "real" lock -- the LockByte -- was reacquired by IWait().
1141 assert (ILocked(), "invariant") ;
1142 assert (_owner == NULL, "invariant") ;
1143 set_owner (Self) ;
1144 return wait_status != 0 ; // return true IFF timeout
1145 }
1146
1147 Monitor::~Monitor() {
1148 assert ((UNS(_owner)|UNS(_LockWord.FullWord)|UNS(_EntryList)|UNS(_WaitSet)|UNS(_OnDeck)) == 0, "") ;
1149 }
1150
1151 void Monitor::ClearMonitor (Monitor * m, const char *name) {
1152 m->_owner = NULL ;
1153 m->_snuck = false ;
1154 if (name == NULL) {
1155 strcpy(m->_name, "UNKNOWN") ;
1156 } else {
1157 strncpy(m->_name, name, MONITOR_NAME_LEN - 1);
1158 m->_name[MONITOR_NAME_LEN - 1] = '\0';
1159 }
1160 m->_LockWord.FullWord = 0 ;
1161 m->_EntryList = NULL ;
1162 m->_OnDeck = NULL ;
1163 m->_WaitSet = NULL ;
1164 m->_WaitLock[0] = 0 ;
1165 }
1166
1167 Monitor::Monitor() { ClearMonitor(this); }
1168
1169 Monitor::Monitor (int Rank, const char * name, bool allow_vm_block) {
1170 ClearMonitor (this, name) ;
1171 #ifdef ASSERT
1172 _allow_vm_block = allow_vm_block;
1173 _rank = Rank ;
1174 #endif
1175 }
1176
1177 Mutex::~Mutex() {
1178 assert ((UNS(_owner)|UNS(_LockWord.FullWord)|UNS(_EntryList)|UNS(_WaitSet)|UNS(_OnDeck)) == 0, "") ;
1179 }
1180
1181 Mutex::Mutex (int Rank, const char * name, bool allow_vm_block) {
1182 ClearMonitor ((Monitor *) this, name) ;
1183 #ifdef ASSERT
1184 _allow_vm_block = allow_vm_block;
1185 _rank = Rank ;
1186 #endif
1187 }
1188
1189 bool Monitor::owned_by_self() const {
1190 bool ret = _owner == Thread::current();
1191 assert (!ret || _LockWord.Bytes[_LSBINDEX] != 0, "invariant") ;
1192 return ret;
1193 }
1194
1195 void Monitor::print_on_error(outputStream* st) const {
1196 st->print("[" PTR_FORMAT, this);
1197 st->print("] %s", _name);
1198 st->print(" - owner thread: " PTR_FORMAT, _owner);
1199 }
1200
1201
1202
1203
1204 // ----------------------------------------------------------------------------------
1205 // Non-product code
1206
1207 #ifndef PRODUCT
1208 void Monitor::print_on(outputStream* st) const {
1209 st->print_cr("Mutex: [0x%lx/0x%lx] %s - owner: 0x%lx", this, _LockWord.FullWord, _name, _owner);
1210 }
1211 #endif
1212
1213 #ifndef PRODUCT
1214 #ifdef ASSERT
1215 Monitor * Monitor::get_least_ranked_lock(Monitor * locks) {
1216 Monitor *res, *tmp;
1217 for (res = tmp = locks; tmp != NULL; tmp = tmp->next()) {
1218 if (tmp->rank() < res->rank()) {
1219 res = tmp;
1220 }
1221 }
1222 if (!SafepointSynchronize::is_at_safepoint()) {
1223 // In this case, we expect the held locks to be
1224 // in increasing rank order (modulo any native ranks)
1225 for (tmp = locks; tmp != NULL; tmp = tmp->next()) {
1226 if (tmp->next() != NULL) {
1227 assert(tmp->rank() == Mutex::native ||
1228 tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?");
1229 }
1230 }
1231 }
1232 return res;
1233 }
1234
1235 Monitor* Monitor::get_least_ranked_lock_besides_this(Monitor* locks) {
1236 Monitor *res, *tmp;
1237 for (res = NULL, tmp = locks; tmp != NULL; tmp = tmp->next()) {
1238 if (tmp != this && (res == NULL || tmp->rank() < res->rank())) {
1239 res = tmp;
1240 }
1241 }
1242 if (!SafepointSynchronize::is_at_safepoint()) {
1243 // In this case, we expect the held locks to be
1244 // in increasing rank order (modulo any native ranks)
1245 for (tmp = locks; tmp != NULL; tmp = tmp->next()) {
1246 if (tmp->next() != NULL) {
1247 assert(tmp->rank() == Mutex::native ||
1248 tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?");
1249 }
1250 }
1251 }
1252 return res;
1253 }
1254
1255
1256 bool Monitor::contains(Monitor* locks, Monitor * lock) {
1257 for (; locks != NULL; locks = locks->next()) {
1258 if (locks == lock)
1259 return true;
1260 }
1261 return false;
1262 }
1263 #endif
1264
1265 // Called immediately after lock acquisition or release as a diagnostic
1266 // to track the lock-set of the thread and test for rank violations that
1267 // might indicate exposure to deadlock.
1268 // Rather like an EventListener for _owner (:>).
1269
1270 void Monitor::set_owner_implementation(Thread *new_owner) {
1271 // This function is solely responsible for maintaining
1272 // and checking the invariant that threads and locks
1273 // are in a 1/N relation, with some some locks unowned.
1274 // It uses the Mutex::_owner, Mutex::_next, and
1275 // Thread::_owned_locks fields, and no other function
1276 // changes those fields.
1277 // It is illegal to set the mutex from one non-NULL
1278 // owner to another--it must be owned by NULL as an
1279 // intermediate state.
1280
1281 if (new_owner != NULL) {
1282 // the thread is acquiring this lock
1283
1284 assert(new_owner == Thread::current(), "Should I be doing this?");
1285 assert(_owner == NULL, "setting the owner thread of an already owned mutex");
1286 _owner = new_owner; // set the owner
1287
1288 // link "this" into the owned locks list
1289
1290 #ifdef ASSERT // Thread::_owned_locks is under the same ifdef
1291 Monitor* locks = get_least_ranked_lock(new_owner->owned_locks());
1292 // Mutex::set_owner_implementation is a friend of Thread
1293
1294 assert(this->rank() >= 0, "bad lock rank");
1295
1296 // Deadlock avoidance rules require us to acquire Mutexes only in
1297 // a global total order. For example m1 is the lowest ranked mutex
1298 // that the thread holds and m2 is the mutex the thread is trying
1299 // to acquire, then deadlock avoidance rules require that the rank
1300 // of m2 be less than the rank of m1.
1301 // The rank Mutex::native is an exception in that it is not subject
1302 // to the verification rules.
1303 // Here are some further notes relating to mutex acquisition anomalies:
1304 // . under Solaris, the interrupt lock gets acquired when doing
1305 // profiling, so any lock could be held.
1306 // . it is also ok to acquire Safepoint_lock at the very end while we
1307 // already hold Terminator_lock - may happen because of periodic safepoints
1308 if (this->rank() != Mutex::native &&
1309 this->rank() != Mutex::suspend_resume &&
1310 locks != NULL && locks->rank() <= this->rank() &&
1311 !SafepointSynchronize::is_at_safepoint() &&
1312 this != Interrupt_lock && this != ProfileVM_lock &&
1313 !(this == Safepoint_lock && contains(locks, Terminator_lock) &&
1314 SafepointSynchronize::is_synchronizing())) {
1315 new_owner->print_owned_locks();
1316 fatal(err_msg("acquiring lock %s/%d out of order with lock %s/%d -- "
1317 "possible deadlock", this->name(), this->rank(),
1318 locks->name(), locks->rank()));
1319 }
1320
1321 this->_next = new_owner->_owned_locks;
1322 new_owner->_owned_locks = this;
1323 #endif
1324
1325 } else {
1326 // the thread is releasing this lock
1327
1328 Thread* old_owner = _owner;
1329 debug_only(_last_owner = old_owner);
1330
1331 assert(old_owner != NULL, "removing the owner thread of an unowned mutex");
1332 assert(old_owner == Thread::current(), "removing the owner thread of an unowned mutex");
1333
1334 _owner = NULL; // set the owner
1335
1336 #ifdef ASSERT
1337 Monitor *locks = old_owner->owned_locks();
1338
1339 // remove "this" from the owned locks list
1340
1341 Monitor *prev = NULL;
1342 bool found = false;
1343 for (; locks != NULL; prev = locks, locks = locks->next()) {
1344 if (locks == this) {
1345 found = true;
1346 break;
1347 }
1348 }
1349 assert(found, "Removing a lock not owned");
1350 if (prev == NULL) {
1351 old_owner->_owned_locks = _next;
1352 } else {
1353 prev->_next = _next;
1354 }
1355 _next = NULL;
1356 #endif
1357 }
1358 }
1359
1360
1361 // Factored out common sanity checks for locking mutex'es. Used by lock() and try_lock()
1362 void Monitor::check_prelock_state(Thread *thread) {
1363 assert((!thread->is_Java_thread() || ((JavaThread *)thread)->thread_state() == _thread_in_vm)
1364 || rank() == Mutex::special, "wrong thread state for using locks");
1365 if (StrictSafepointChecks) {
1366 if (thread->is_VM_thread() && !allow_vm_block()) {
1367 fatal(err_msg("VM thread using lock %s (not allowed to block on)",
1368 name()));
1369 }
1370 debug_only(if (rank() != Mutex::special) \
1371 thread->check_for_valid_safepoint_state(false);)
1372 }
1373 }
1374
1375 void Monitor::check_block_state(Thread *thread) {
1376 if (!_allow_vm_block && thread->is_VM_thread()) {
1377 warning("VM thread blocked on lock");
1378 print();
1379 BREAKPOINT;
1380 }
1381 assert(_owner != thread, "deadlock: blocking on monitor owned by current thread");
1382 }
1383
1384 #endif // PRODUCT