1 /*
2 * Copyright (c) 1998, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "runtime/atomic.hpp"
27 #include "runtime/interfaceSupport.hpp"
28 #include "runtime/mutex.hpp"
29 #include "runtime/orderAccess.inline.hpp"
30 #include "runtime/osThread.hpp"
31 #include "runtime/thread.inline.hpp"
32 #include "utilities/events.hpp"
33 #include "utilities/macros.hpp"
34
35 // 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
36 //
37 // Native Monitor-Mutex locking - theory of operations
38 //
39 // * Native Monitors are completely unrelated to Java-level monitors,
40 // although the "back-end" slow-path implementations share a common lineage.
41 // See objectMonitor:: in synchronizer.cpp.
42 // Native Monitors do *not* support nesting or recursion but otherwise
43 // they're basically Hoare-flavor monitors.
44 //
45 // * A thread acquires ownership of a Monitor/Mutex by CASing the LockByte
46 // in the _LockWord from zero to non-zero. Note that the _Owner field
47 // is advisory and is used only to verify that the thread calling unlock()
48 // is indeed the last thread to have acquired the lock.
49 //
50 // * Contending threads "push" themselves onto the front of the contention
51 // queue -- called the cxq -- with CAS and then spin/park.
52 // The _LockWord contains the LockByte as well as the pointer to the head
53 // of the cxq. Colocating the LockByte with the cxq precludes certain races.
54 //
55 // * Using a separately addressable LockByte allows for CAS:MEMBAR or CAS:0
56 // idioms. We currently use MEMBAR in the uncontended unlock() path, as
57 // MEMBAR often has less latency than CAS. If warranted, we could switch to
58 // a CAS:0 mode, using timers to close the resultant race, as is done
59 // with Java Monitors in synchronizer.cpp.
60 //
61 // See the following for a discussion of the relative cost of atomics (CAS)
62 // MEMBAR, and ways to eliminate such instructions from the common-case paths:
63 // -- http://blogs.sun.com/dave/entry/biased_locking_in_hotspot
64 // -- http://blogs.sun.com/dave/resource/MustangSync.pdf
65 // -- http://blogs.sun.com/dave/resource/synchronization-public2.pdf
66 // -- synchronizer.cpp
67 //
68 // * Overall goals - desiderata
69 // 1. Minimize context switching
70 // 2. Minimize lock migration
71 // 3. Minimize CPI -- affinity and locality
72 // 4. Minimize the execution of high-latency instructions such as CAS or MEMBAR
73 // 5. Minimize outer lock hold times
74 // 6. Behave gracefully on a loaded system
75 //
76 // * Thread flow and list residency:
77 //
78 // Contention queue --> EntryList --> OnDeck --> Owner --> !Owner
79 // [..resident on monitor list..]
80 // [...........contending..................]
81 //
82 // -- The contention queue (cxq) contains recently-arrived threads (RATs).
83 // Threads on the cxq eventually drain into the EntryList.
84 // -- Invariant: a thread appears on at most one list -- cxq, EntryList
85 // or WaitSet -- at any one time.
86 // -- For a given monitor there can be at most one "OnDeck" thread at any
87 // given time but if needbe this particular invariant could be relaxed.
88 //
89 // * The WaitSet and EntryList linked lists are composed of ParkEvents.
90 // I use ParkEvent instead of threads as ParkEvents are immortal and
91 // type-stable, meaning we can safely unpark() a possibly stale
92 // list element in the unlock()-path. (That's benign).
93 //
94 // * Succession policy - providing for progress:
95 //
96 // As necessary, the unlock()ing thread identifies, unlinks, and unparks
97 // an "heir presumptive" tentative successor thread from the EntryList.
98 // This becomes the so-called "OnDeck" thread, of which there can be only
99 // one at any given time for a given monitor. The wakee will recontend
100 // for ownership of monitor.
101 //
102 // Succession is provided for by a policy of competitive handoff.
103 // The exiting thread does _not_ grant or pass ownership to the
104 // successor thread. (This is also referred to as "handoff" succession").
105 // Instead the exiting thread releases ownership and possibly wakes
106 // a successor, so the successor can (re)compete for ownership of the lock.
107 //
108 // Competitive handoff provides excellent overall throughput at the expense
109 // of short-term fairness. If fairness is a concern then one remedy might
110 // be to add an AcquireCounter field to the monitor. After a thread acquires
111 // the lock it will decrement the AcquireCounter field. When the count
112 // reaches 0 the thread would reset the AcquireCounter variable, abdicate
113 // the lock directly to some thread on the EntryList, and then move itself to the
114 // tail of the EntryList.
115 //
116 // But in practice most threads engage or otherwise participate in resource
117 // bounded producer-consumer relationships, so lock domination is not usually
118 // a practical concern. Recall too, that in general it's easier to construct
119 // a fair lock from a fast lock, but not vice-versa.
120 //
121 // * The cxq can have multiple concurrent "pushers" but only one concurrent
122 // detaching thread. This mechanism is immune from the ABA corruption.
123 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
124 // We use OnDeck as a pseudo-lock to enforce the at-most-one detaching
125 // thread constraint.
126 //
127 // * Taken together, the cxq and the EntryList constitute or form a
128 // single logical queue of threads stalled trying to acquire the lock.
129 // We use two distinct lists to reduce heat on the list ends.
130 // Threads in lock() enqueue onto cxq while threads in unlock() will
131 // dequeue from the EntryList. (c.f. Michael Scott's "2Q" algorithm).
132 // A key desideratum is to minimize queue & monitor metadata manipulation
133 // that occurs while holding the "outer" monitor lock -- that is, we want to
134 // minimize monitor lock holds times.
135 //
136 // The EntryList is ordered by the prevailing queue discipline and
137 // can be organized in any convenient fashion, such as a doubly-linked list or
138 // a circular doubly-linked list. If we need a priority queue then something akin
139 // to Solaris' sleepq would work nicely. Viz.,
140 // -- http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
141 // -- http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/uts/common/os/sleepq.c
142 // Queue discipline is enforced at ::unlock() time, when the unlocking thread
143 // drains the cxq into the EntryList, and orders or reorders the threads on the
144 // EntryList accordingly.
145 //
146 // Barring "lock barging", this mechanism provides fair cyclic ordering,
147 // somewhat similar to an elevator-scan.
148 //
149 // * OnDeck
150 // -- For a given monitor there can be at most one OnDeck thread at any given
151 // instant. The OnDeck thread is contending for the lock, but has been
152 // unlinked from the EntryList and cxq by some previous unlock() operations.
153 // Once a thread has been designated the OnDeck thread it will remain so
154 // until it manages to acquire the lock -- being OnDeck is a stable property.
155 // -- Threads on the EntryList or cxq are _not allowed to attempt lock acquisition.
156 // -- OnDeck also serves as an "inner lock" as follows. Threads in unlock() will, after
157 // having cleared the LockByte and dropped the outer lock, attempt to "trylock"
158 // OnDeck by CASing the field from null to non-null. If successful, that thread
159 // is then responsible for progress and succession and can use CAS to detach and
160 // drain the cxq into the EntryList. By convention, only this thread, the holder of
161 // the OnDeck inner lock, can manipulate the EntryList or detach and drain the
162 // RATs on the cxq into the EntryList. This avoids ABA corruption on the cxq as
163 // we allow multiple concurrent "push" operations but restrict detach concurrency
164 // to at most one thread. Having selected and detached a successor, the thread then
165 // changes the OnDeck to refer to that successor, and then unparks the successor.
166 // That successor will eventually acquire the lock and clear OnDeck. Beware
167 // that the OnDeck usage as a lock is asymmetric. A thread in unlock() transiently
168 // "acquires" OnDeck, performs queue manipulations, passes OnDeck to some successor,
169 // and then the successor eventually "drops" OnDeck. Note that there's never
170 // any sense of contention on the inner lock, however. Threads never contend
171 // or wait for the inner lock.
172 // -- OnDeck provides for futile wakeup throttling a described in section 3.3 of
173 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
174 // In a sense, OnDeck subsumes the ObjectMonitor _Succ and ObjectWaiter
175 // TState fields found in Java-level objectMonitors. (See synchronizer.cpp).
176 //
177 // * Waiting threads reside on the WaitSet list -- wait() puts
178 // the caller onto the WaitSet. Notify() or notifyAll() simply
179 // transfers threads from the WaitSet to either the EntryList or cxq.
180 // Subsequent unlock() operations will eventually unpark the notifyee.
181 // Unparking a notifee in notify() proper is inefficient - if we were to do so
182 // it's likely the notifyee would simply impale itself on the lock held
183 // by the notifier.
184 //
185 // * The mechanism is obstruction-free in that if the holder of the transient
186 // OnDeck lock in unlock() is preempted or otherwise stalls, other threads
187 // can still acquire and release the outer lock and continue to make progress.
188 // At worst, waking of already blocked contending threads may be delayed,
189 // but nothing worse. (We only use "trylock" operations on the inner OnDeck
190 // lock).
191 //
192 // * Note that thread-local storage must be initialized before a thread
193 // uses Native monitors or mutexes. The native monitor-mutex subsystem
194 // depends on Thread::current().
195 //
196 // * The monitor synchronization subsystem avoids the use of native
197 // synchronization primitives except for the narrow platform-specific
198 // park-unpark abstraction. See the comments in os_solaris.cpp regarding
199 // the semantics of park-unpark. Put another way, this monitor implementation
200 // depends only on atomic operations and park-unpark. The monitor subsystem
201 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
202 // underlying OS manages the READY<->RUN transitions.
203 //
204 // * The memory consistency model provide by lock()-unlock() is at least as
205 // strong or stronger than the Java Memory model defined by JSR-133.
206 // That is, we guarantee at least entry consistency, if not stronger.
207 // See http://g.oswego.edu/dl/jmm/cookbook.html.
208 //
209 // * Thread:: currently contains a set of purpose-specific ParkEvents:
210 // _MutexEvent, _ParkEvent, etc. A better approach might be to do away with
211 // the purpose-specific ParkEvents and instead implement a general per-thread
212 // stack of available ParkEvents which we could provision on-demand. The
213 // stack acts as a local cache to avoid excessive calls to ParkEvent::Allocate()
214 // and ::Release(). A thread would simply pop an element from the local stack before it
215 // enqueued or park()ed. When the contention was over the thread would
216 // push the no-longer-needed ParkEvent back onto its stack.
217 //
218 // * A slightly reduced form of ILock() and IUnlock() have been partially
219 // model-checked (Murphi) for safety and progress at T=1,2,3 and 4.
220 // It'd be interesting to see if TLA/TLC could be useful as well.
221 //
222 // * Mutex-Monitor is a low-level "leaf" subsystem. That is, the monitor
223 // code should never call other code in the JVM that might itself need to
224 // acquire monitors or mutexes. That's true *except* in the case of the
225 // ThreadBlockInVM state transition wrappers. The ThreadBlockInVM DTOR handles
226 // mutator reentry (ingress) by checking for a pending safepoint in which case it will
227 // call SafepointSynchronize::block(), which in turn may call Safepoint_lock->lock(), etc.
228 // In that particular case a call to lock() for a given Monitor can end up recursively
229 // calling lock() on another monitor. While distasteful, this is largely benign
230 // as the calls come from jacket that wraps lock(), and not from deep within lock() itself.
231 //
232 // It's unfortunate that native mutexes and thread state transitions were convolved.
233 // They're really separate concerns and should have remained that way. Melding
234 // them together was facile -- a bit too facile. The current implementation badly
235 // conflates the two concerns.
236 //
237 // * TODO-FIXME:
238 //
239 // -- Add DTRACE probes for contended acquire, contended acquired, contended unlock
240 // We should also add DTRACE probes in the ParkEvent subsystem for
241 // Park-entry, Park-exit, and Unpark.
242 //
243 // -- We have an excess of mutex-like constructs in the JVM, namely:
244 // 1. objectMonitors for Java-level synchronization (synchronizer.cpp)
245 // 2. low-level muxAcquire and muxRelease
246 // 3. low-level spinAcquire and spinRelease
247 // 4. native Mutex:: and Monitor::
248 // 5. jvm_raw_lock() and _unlock()
249 // 6. JVMTI raw monitors -- distinct from (5) despite having a confusingly
250 // similar name.
251 //
252 // 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
253
254 #define UNS(x) (uintptr_t(x))
255 #define TRACE(m) \
256 { \
257 static volatile int ctr = 0; \
258 int x = ++ctr; \
259 if ((x & (x - 1)) == 0) { \
260 ::printf("%d:%s\n", x, #m); \
261 ::fflush(stdout); \
262 } \
263 }
264
265 const intptr_t _LBIT = 1;
266
267 // Endian-ness ... index of least-significant byte in SplitWord.Bytes[]
268 #ifdef VM_LITTLE_ENDIAN
269 #define _LSBINDEX 0
270 #else
271 #define _LSBINDEX (sizeof(intptr_t)-1)
272 #endif
273
274 // Simplistic low-quality Marsaglia SHIFT-XOR RNG.
275 // Bijective except for the trailing mask operation.
276 // Useful for spin loops as the compiler can't optimize it away.
277
278 static inline jint MarsagliaXORV(jint x) {
279 if (x == 0) x = 1|os::random();
280 x ^= x << 6;
281 x ^= ((unsigned)x) >> 21;
282 x ^= x << 7;
283 return x & 0x7FFFFFFF;
284 }
285
286 static int Stall(int its) {
287 static volatile jint rv = 1;
288 volatile int OnFrame = 0;
289 jint v = rv ^ UNS(OnFrame);
290 while (--its >= 0) {
291 v = MarsagliaXORV(v);
292 }
293 // Make this impossible for the compiler to optimize away,
294 // but (mostly) avoid W coherency sharing on MP systems.
295 if (v == 0x12345) rv = v;
296 return v;
297 }
298
299 int Monitor::TryLock() {
300 intptr_t v = _LockWord.FullWord;
301 for (;;) {
302 if ((v & _LBIT) != 0) return 0;
303 const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v);
304 if (v == u) return 1;
305 v = u;
306 }
307 }
308
309 int Monitor::TryFast() {
310 // Optimistic fast-path form ...
311 // Fast-path attempt for the common uncontended case.
312 // Avoid RTS->RTO $ coherence upgrade on typical SMP systems.
313 intptr_t v = Atomic::cmpxchg(_LBIT, &_LockWord.FullWord, (intptr_t)0); // agro ...
314 if (v == 0) return 1;
315
316 for (;;) {
317 if ((v & _LBIT) != 0) return 0;
318 const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v);
319 if (v == u) return 1;
320 v = u;
321 }
322 }
323
324 int Monitor::ILocked() {
325 const intptr_t w = _LockWord.FullWord & 0xFF;
326 assert(w == 0 || w == _LBIT, "invariant");
327 return w == _LBIT;
328 }
329
330 // Polite TATAS spinlock with exponential backoff - bounded spin.
331 // Ideally we'd use processor cycles, time or vtime to control
332 // the loop, but we currently use iterations.
333 // All the constants within were derived empirically but work over
334 // over the spectrum of J2SE reference platforms.
335 // On Niagara-class systems the back-off is unnecessary but
336 // is relatively harmless. (At worst it'll slightly retard
337 // acquisition times). The back-off is critical for older SMP systems
338 // where constant fetching of the LockWord would otherwise impair
339 // scalability.
340 //
341 // Clamp spinning at approximately 1/2 of a context-switch round-trip.
342 // See synchronizer.cpp for details and rationale.
343
344 int Monitor::TrySpin(Thread * const Self) {
345 if (TryLock()) return 1;
346 if (!os::is_MP()) return 0;
347
348 int Probes = 0;
349 int Delay = 0;
350 int Steps = 0;
351 int SpinMax = NativeMonitorSpinLimit;
352 int flgs = NativeMonitorFlags;
353 for (;;) {
354 intptr_t v = _LockWord.FullWord;
355 if ((v & _LBIT) == 0) {
356 if (Atomic::cmpxchg (v|_LBIT, &_LockWord.FullWord, v) == v) {
357 return 1;
358 }
359 continue;
360 }
361
362 if ((flgs & 8) == 0) {
363 SpinPause();
364 }
365
366 // Periodically increase Delay -- variable Delay form
367 // conceptually: delay *= 1 + 1/Exponent
368 ++Probes;
369 if (Probes > SpinMax) return 0;
370
371 if ((Probes & 0x7) == 0) {
372 Delay = ((Delay << 1)|1) & 0x7FF;
373 // CONSIDER: Delay += 1 + (Delay/4); Delay &= 0x7FF ;
374 }
375
376 if (flgs & 2) continue;
377
378 // Consider checking _owner's schedctl state, if OFFPROC abort spin.
379 // If the owner is OFFPROC then it's unlike that the lock will be dropped
380 // in a timely fashion, which suggests that spinning would not be fruitful
381 // or profitable.
382
383 // Stall for "Delay" time units - iterations in the current implementation.
384 // Avoid generating coherency traffic while stalled.
385 // Possible ways to delay:
386 // PAUSE, SLEEP, MEMBAR #sync, MEMBAR #halt,
387 // wr %g0,%asi, gethrtime, rdstick, rdtick, rdtsc, etc. ...
388 // Note that on Niagara-class systems we want to minimize STs in the
389 // spin loop. N1 and brethren write-around the L1$ over the xbar into the L2$.
390 // Furthermore, they don't have a W$ like traditional SPARC processors.
391 // We currently use a Marsaglia Shift-Xor RNG loop.
392 Steps += Delay;
393 if (Self != NULL) {
394 jint rv = Self->rng[0];
395 for (int k = Delay; --k >= 0;) {
396 rv = MarsagliaXORV(rv);
397 if ((flgs & 4) == 0 && SafepointSynchronize::do_call_back()) return 0;
398 }
399 Self->rng[0] = rv;
400 } else {
401 Stall(Delay);
402 }
403 }
404 }
405
406 static int ParkCommon(ParkEvent * ev, jlong timo) {
407 // Diagnostic support - periodically unwedge blocked threads
408 intx nmt = NativeMonitorTimeout;
409 if (nmt > 0 && (nmt < timo || timo <= 0)) {
410 timo = nmt;
411 }
412 int err = OS_OK;
413 if (0 == timo) {
414 ev->park();
415 } else {
416 err = ev->park(timo);
417 }
418 return err;
419 }
420
421 inline int Monitor::AcquireOrPush(ParkEvent * ESelf) {
422 intptr_t v = _LockWord.FullWord;
423 for (;;) {
424 if ((v & _LBIT) == 0) {
425 const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v);
426 if (u == v) return 1; // indicate acquired
427 v = u;
428 } else {
429 // Anticipate success ...
430 ESelf->ListNext = (ParkEvent *)(v & ~_LBIT);
431 const intptr_t u = Atomic::cmpxchg(intptr_t(ESelf)|_LBIT, &_LockWord.FullWord, v);
432 if (u == v) return 0; // indicate pushed onto cxq
433 v = u;
434 }
435 // Interference - LockWord change - just retry
436 }
437 }
438
439 // ILock and IWait are the lowest level primitive internal blocking
440 // synchronization functions. The callers of IWait and ILock must have
441 // performed any needed state transitions beforehand.
442 // IWait and ILock may directly call park() without any concern for thread state.
443 // Note that ILock and IWait do *not* access _owner.
444 // _owner is a higher-level logical concept.
445
446 void Monitor::ILock(Thread * Self) {
447 assert(_OnDeck != Self->_MutexEvent, "invariant");
448
449 if (TryFast()) {
450 Exeunt:
451 assert(ILocked(), "invariant");
452 return;
453 }
454
455 ParkEvent * const ESelf = Self->_MutexEvent;
456 assert(_OnDeck != ESelf, "invariant");
457
458 // As an optimization, spinners could conditionally try to set _OnDeck to _LBIT
459 // Synchronizer.cpp uses a similar optimization.
460 if (TrySpin(Self)) goto Exeunt;
461
462 // Slow-path - the lock is contended.
463 // Either Enqueue Self on cxq or acquire the outer lock.
464 // LockWord encoding = (cxq,LOCKBYTE)
465 ESelf->reset();
466 OrderAccess::fence();
467
468 // Optional optimization ... try barging on the inner lock
469 if ((NativeMonitorFlags & 32) && Atomic::cmpxchg(ESelf, &_OnDeck, (ParkEvent*)NULL) == NULL) {
470 goto OnDeck_LOOP;
471 }
472
473 if (AcquireOrPush(ESelf)) goto Exeunt;
474
475 // At any given time there is at most one ondeck thread.
476 // ondeck implies not resident on cxq and not resident on EntryList
477 // Only the OnDeck thread can try to acquire -- contend for -- the lock.
478 // CONSIDER: use Self->OnDeck instead of m->OnDeck.
479 // Deschedule Self so that others may run.
480 while (OrderAccess::load_acquire(&_OnDeck) != ESelf) {
481 ParkCommon(ESelf, 0);
482 }
483
484 // Self is now in the OnDeck position and will remain so until it
485 // manages to acquire the lock.
486 OnDeck_LOOP:
487 for (;;) {
488 assert(_OnDeck == ESelf, "invariant");
489 if (TrySpin(Self)) break;
490 // It's probably wise to spin only if we *actually* blocked
491 // CONSIDER: check the lockbyte, if it remains set then
492 // preemptively drain the cxq into the EntryList.
493 // The best place and time to perform queue operations -- lock metadata --
494 // is _before having acquired the outer lock, while waiting for the lock to drop.
495 ParkCommon(ESelf, 0);
496 }
497
498 assert(_OnDeck == ESelf, "invariant");
499 _OnDeck = NULL;
500
501 // Note that we current drop the inner lock (clear OnDeck) in the slow-path
502 // epilogue immediately after having acquired the outer lock.
503 // But instead we could consider the following optimizations:
504 // A. Shift or defer dropping the inner lock until the subsequent IUnlock() operation.
505 // This might avoid potential reacquisition of the inner lock in IUlock().
506 // B. While still holding the inner lock, attempt to opportunistically select
507 // and unlink the next OnDeck thread from the EntryList.
508 // If successful, set OnDeck to refer to that thread, otherwise clear OnDeck.
509 // It's critical that the select-and-unlink operation run in constant-time as
510 // it executes when holding the outer lock and may artificially increase the
511 // effective length of the critical section.
512 // Note that (A) and (B) are tantamount to succession by direct handoff for
513 // the inner lock.
514 goto Exeunt;
515 }
516
517 void Monitor::IUnlock(bool RelaxAssert) {
518 assert(ILocked(), "invariant");
519 // Conceptually we need a MEMBAR #storestore|#loadstore barrier or fence immediately
520 // before the store that releases the lock. Crucially, all the stores and loads in the
521 // critical section must be globally visible before the store of 0 into the lock-word
522 // that releases the lock becomes globally visible. That is, memory accesses in the
523 // critical section should not be allowed to bypass or overtake the following ST that
524 // releases the lock. As such, to prevent accesses within the critical section
525 // from "leaking" out, we need a release fence between the critical section and the
526 // store that releases the lock. In practice that release barrier is elided on
527 // platforms with strong memory models such as TSO.
528 //
529 // Note that the OrderAccess::storeload() fence that appears after unlock store
530 // provides for progress conditions and succession and is _not related to exclusion
531 // safety or lock release consistency.
532 OrderAccess::release_store(&_LockWord.Bytes[_LSBINDEX], jbyte(0)); // drop outer lock
533
534 OrderAccess::storeload();
535 ParkEvent * const w = _OnDeck; // raw load as we will just return if non-NULL
536 assert(RelaxAssert || w != Thread::current()->_MutexEvent, "invariant");
537 if (w != NULL) {
538 // Either we have a valid ondeck thread or ondeck is transiently "locked"
539 // by some exiting thread as it arranges for succession. The LSBit of
540 // OnDeck allows us to discriminate two cases. If the latter, the
541 // responsibility for progress and succession lies with that other thread.
542 // For good performance, we also depend on the fact that redundant unpark()
543 // operations are cheap. That is, repeated Unpark()ing of the OnDeck thread
544 // is inexpensive. This approach provides implicit futile wakeup throttling.
545 // Note that the referent "w" might be stale with respect to the lock.
546 // In that case the following unpark() is harmless and the worst that'll happen
547 // is a spurious return from a park() operation. Critically, if "w" _is stale,
548 // then progress is known to have occurred as that means the thread associated
549 // with "w" acquired the lock. In that case this thread need take no further
550 // action to guarantee progress.
551 if ((UNS(w) & _LBIT) == 0) w->unpark();
552 return;
553 }
554
555 intptr_t cxq = _LockWord.FullWord;
556 if (((cxq & ~_LBIT)|UNS(_EntryList)) == 0) {
557 return; // normal fast-path exit - cxq and EntryList both empty
558 }
559 if (cxq & _LBIT) {
560 // Optional optimization ...
561 // Some other thread acquired the lock in the window since this
562 // thread released it. Succession is now that thread's responsibility.
563 return;
564 }
565
566 Succession:
567 // Slow-path exit - this thread must ensure succession and progress.
568 // OnDeck serves as lock to protect cxq and EntryList.
569 // Only the holder of OnDeck can manipulate EntryList or detach the RATs from cxq.
570 // Avoid ABA - allow multiple concurrent producers (enqueue via push-CAS)
571 // but only one concurrent consumer (detacher of RATs).
572 // Consider protecting this critical section with schedctl on Solaris.
573 // Unlike a normal lock, however, the exiting thread "locks" OnDeck,
574 // picks a successor and marks that thread as OnDeck. That successor
575 // thread will then clear OnDeck once it eventually acquires the outer lock.
576 if (Atomic::cmpxchg((ParkEvent*)_LBIT, &_OnDeck, (ParkEvent*)NULL) != NULL) {
577 return;
578 }
579
580 ParkEvent * List = _EntryList;
581 if (List != NULL) {
582 // Transfer the head of the EntryList to the OnDeck position.
583 // Once OnDeck, a thread stays OnDeck until it acquires the lock.
584 // For a given lock there is at most OnDeck thread at any one instant.
585 WakeOne:
586 assert(List == _EntryList, "invariant");
587 ParkEvent * const w = List;
588 assert(RelaxAssert || w != Thread::current()->_MutexEvent, "invariant");
589 _EntryList = w->ListNext;
590 // as a diagnostic measure consider setting w->_ListNext = BAD
591 assert(intptr_t(_OnDeck) == _LBIT, "invariant");
592
593 // Pass OnDeck role to w, ensuring that _EntryList has been set first.
594 // w will clear _OnDeck once it acquires the outer lock.
595 // Note that once we set _OnDeck that thread can acquire the mutex, proceed
596 // with its critical section and then enter this code to unlock the mutex. So
597 // you can have multiple threads active in IUnlock at the same time.
598 OrderAccess::release_store(&_OnDeck, w);
599
600 // Another optional optimization ...
601 // For heavily contended locks it's not uncommon that some other
602 // thread acquired the lock while this thread was arranging succession.
603 // Try to defer the unpark() operation - Delegate the responsibility
604 // for unpark()ing the OnDeck thread to the current or subsequent owners
605 // That is, the new owner is responsible for unparking the OnDeck thread.
606 OrderAccess::storeload();
607 cxq = _LockWord.FullWord;
608 if (cxq & _LBIT) return;
609
610 w->unpark();
611 return;
612 }
613
614 cxq = _LockWord.FullWord;
615 if ((cxq & ~_LBIT) != 0) {
616 // The EntryList is empty but the cxq is populated.
617 // drain RATs from cxq into EntryList
618 // Detach RATs segment with CAS and then merge into EntryList
619 for (;;) {
620 // optional optimization - if locked, the owner is responsible for succession
621 if (cxq & _LBIT) goto Punt;
622 const intptr_t vfy = Atomic::cmpxchg(cxq & _LBIT, &_LockWord.FullWord, cxq);
623 if (vfy == cxq) break;
624 cxq = vfy;
625 // Interference - LockWord changed - Just retry
626 // We can see concurrent interference from contending threads
627 // pushing themselves onto the cxq or from lock-unlock operations.
628 // From the perspective of this thread, EntryList is stable and
629 // the cxq is prepend-only -- the head is volatile but the interior
630 // of the cxq is stable. In theory if we encounter interference from threads
631 // pushing onto cxq we could simply break off the original cxq suffix and
632 // move that segment to the EntryList, avoiding a 2nd or multiple CAS attempts
633 // on the high-traffic LockWord variable. For instance lets say the cxq is "ABCD"
634 // when we first fetch cxq above. Between the fetch -- where we observed "A"
635 // -- and CAS -- where we attempt to CAS null over A -- "PQR" arrive,
636 // yielding cxq = "PQRABCD". In this case we could simply set A.ListNext
637 // null, leaving cxq = "PQRA" and transfer the "BCD" segment to the EntryList.
638 // Note too, that it's safe for this thread to traverse the cxq
639 // without taking any special concurrency precautions.
640 }
641
642 // We don't currently reorder the cxq segment as we move it onto
643 // the EntryList, but it might make sense to reverse the order
644 // or perhaps sort by thread priority. See the comments in
645 // synchronizer.cpp objectMonitor::exit().
646 assert(_EntryList == NULL, "invariant");
647 _EntryList = List = (ParkEvent *)(cxq & ~_LBIT);
648 assert(List != NULL, "invariant");
649 goto WakeOne;
650 }
651
652 // cxq|EntryList is empty.
653 // w == NULL implies that cxq|EntryList == NULL in the past.
654 // Possible race - rare inopportune interleaving.
655 // A thread could have added itself to cxq since this thread previously checked.
656 // Detect and recover by refetching cxq.
657 Punt:
658 assert(intptr_t(_OnDeck) == _LBIT, "invariant");
659 _OnDeck = NULL; // Release inner lock.
660 OrderAccess::storeload(); // Dekker duality - pivot point
661
662 // Resample LockWord/cxq to recover from possible race.
663 // For instance, while this thread T1 held OnDeck, some other thread T2 might
664 // acquire the outer lock. Another thread T3 might try to acquire the outer
665 // lock, but encounter contention and enqueue itself on cxq. T2 then drops the
666 // outer lock, but skips succession as this thread T1 still holds OnDeck.
667 // T1 is and remains responsible for ensuring succession of T3.
668 //
669 // Note that we don't need to recheck EntryList, just cxq.
670 // If threads moved onto EntryList since we dropped OnDeck
671 // that implies some other thread forced succession.
672 cxq = _LockWord.FullWord;
673 if ((cxq & ~_LBIT) != 0 && (cxq & _LBIT) == 0) {
674 goto Succession; // potential race -- re-run succession
675 }
676 return;
677 }
678
679 bool Monitor::notify() {
680 assert(_owner == Thread::current(), "invariant");
681 assert(ILocked(), "invariant");
682 if (_WaitSet == NULL) return true;
683 NotifyCount++;
684
685 // Transfer one thread from the WaitSet to the EntryList or cxq.
686 // Currently we just unlink the head of the WaitSet and prepend to the cxq.
687 // And of course we could just unlink it and unpark it, too, but
688 // in that case it'd likely impale itself on the reentry.
689 Thread::muxAcquire(_WaitLock, "notify:WaitLock");
690 ParkEvent * nfy = _WaitSet;
691 if (nfy != NULL) { // DCL idiom
692 _WaitSet = nfy->ListNext;
693 assert(nfy->Notified == 0, "invariant");
694 // push nfy onto the cxq
695 for (;;) {
696 const intptr_t v = _LockWord.FullWord;
697 assert((v & 0xFF) == _LBIT, "invariant");
698 nfy->ListNext = (ParkEvent *)(v & ~_LBIT);
699 if (Atomic::cmpxchg(intptr_t(nfy)|_LBIT, &_LockWord.FullWord, v) == v) break;
700 // interference - _LockWord changed -- just retry
701 }
702 // Note that setting Notified before pushing nfy onto the cxq is
703 // also legal and safe, but the safety properties are much more
704 // subtle, so for the sake of code stewardship ...
705 OrderAccess::fence();
706 nfy->Notified = 1;
707 }
708 Thread::muxRelease(_WaitLock);
709 if (nfy != NULL && (NativeMonitorFlags & 16)) {
710 // Experimental code ... light up the wakee in the hope that this thread (the owner)
711 // will drop the lock just about the time the wakee comes ONPROC.
712 nfy->unpark();
713 }
714 assert(ILocked(), "invariant");
715 return true;
716 }
717
718 // Currently notifyAll() transfers the waiters one-at-a-time from the waitset
719 // to the cxq. This could be done more efficiently with a single bulk en-mass transfer,
720 // but in practice notifyAll() for large #s of threads is rare and not time-critical.
721 // Beware too, that we invert the order of the waiters. Lets say that the
722 // waitset is "ABCD" and the cxq is "XYZ". After a notifyAll() the waitset
723 // will be empty and the cxq will be "DCBAXYZ". This is benign, of course.
724
725 bool Monitor::notify_all() {
726 assert(_owner == Thread::current(), "invariant");
727 assert(ILocked(), "invariant");
728 while (_WaitSet != NULL) notify();
729 return true;
730 }
731
732 int Monitor::IWait(Thread * Self, jlong timo) {
733 assert(ILocked(), "invariant");
734
735 // Phases:
736 // 1. Enqueue Self on WaitSet - currently prepend
737 // 2. unlock - drop the outer lock
738 // 3. wait for either notification or timeout
739 // 4. lock - reentry - reacquire the outer lock
740
741 ParkEvent * const ESelf = Self->_MutexEvent;
742 ESelf->Notified = 0;
743 ESelf->reset();
744 OrderAccess::fence();
745
746 // Add Self to WaitSet
747 // Ideally only the holder of the outer lock would manipulate the WaitSet -
748 // That is, the outer lock would implicitly protect the WaitSet.
749 // But if a thread in wait() encounters a timeout it will need to dequeue itself
750 // from the WaitSet _before it becomes the owner of the lock. We need to dequeue
751 // as the ParkEvent -- which serves as a proxy for the thread -- can't reside
752 // on both the WaitSet and the EntryList|cxq at the same time.. That is, a thread
753 // on the WaitSet can't be allowed to compete for the lock until it has managed to
754 // unlink its ParkEvent from WaitSet. Thus the need for WaitLock.
755 // Contention on the WaitLock is minimal.
756 //
757 // Another viable approach would be add another ParkEvent, "WaitEvent" to the
758 // thread class. The WaitSet would be composed of WaitEvents. Only the
759 // owner of the outer lock would manipulate the WaitSet. A thread in wait()
760 // could then compete for the outer lock, and then, if necessary, unlink itself
761 // from the WaitSet only after having acquired the outer lock. More precisely,
762 // there would be no WaitLock. A thread in in wait() would enqueue its WaitEvent
763 // on the WaitSet; release the outer lock; wait for either notification or timeout;
764 // reacquire the inner lock; and then, if needed, unlink itself from the WaitSet.
765 //
766 // Alternatively, a 2nd set of list link fields in the ParkEvent might suffice.
767 // One set would be for the WaitSet and one for the EntryList.
768 // We could also deconstruct the ParkEvent into a "pure" event and add a
769 // new immortal/TSM "ListElement" class that referred to ParkEvents.
770 // In that case we could have one ListElement on the WaitSet and another
771 // on the EntryList, with both referring to the same pure Event.
772
773 Thread::muxAcquire(_WaitLock, "wait:WaitLock:Add");
774 ESelf->ListNext = _WaitSet;
775 _WaitSet = ESelf;
776 Thread::muxRelease(_WaitLock);
777
778 // Release the outer lock
779 // We call IUnlock (RelaxAssert=true) as a thread T1 might
780 // enqueue itself on the WaitSet, call IUnlock(), drop the lock,
781 // and then stall before it can attempt to wake a successor.
782 // Some other thread T2 acquires the lock, and calls notify(), moving
783 // T1 from the WaitSet to the cxq. T2 then drops the lock. T1 resumes,
784 // and then finds *itself* on the cxq. During the course of a normal
785 // IUnlock() call a thread should _never find itself on the EntryList
786 // or cxq, but in the case of wait() it's possible.
787 // See synchronizer.cpp objectMonitor::wait().
788 IUnlock(true);
789
790 // Wait for either notification or timeout
791 // Beware that in some circumstances we might propagate
792 // spurious wakeups back to the caller.
793
794 for (;;) {
795 if (ESelf->Notified) break;
796 int err = ParkCommon(ESelf, timo);
797 if (err == OS_TIMEOUT || (NativeMonitorFlags & 1)) break;
798 }
799
800 // Prepare for reentry - if necessary, remove ESelf from WaitSet
801 // ESelf can be:
802 // 1. Still on the WaitSet. This can happen if we exited the loop by timeout.
803 // 2. On the cxq or EntryList
804 // 3. Not resident on cxq, EntryList or WaitSet, but in the OnDeck position.
805
806 OrderAccess::fence();
807 int WasOnWaitSet = 0;
808 if (ESelf->Notified == 0) {
809 Thread::muxAcquire(_WaitLock, "wait:WaitLock:remove");
810 if (ESelf->Notified == 0) { // DCL idiom
811 assert(_OnDeck != ESelf, "invariant"); // can't be both OnDeck and on WaitSet
812 // ESelf is resident on the WaitSet -- unlink it.
813 // A doubly-linked list would be better here so we can unlink in constant-time.
814 // We have to unlink before we potentially recontend as ESelf might otherwise
815 // end up on the cxq|EntryList -- it can't be on two lists at once.
816 ParkEvent * p = _WaitSet;
817 ParkEvent * q = NULL; // classic q chases p
818 while (p != NULL && p != ESelf) {
819 q = p;
820 p = p->ListNext;
821 }
822 assert(p == ESelf, "invariant");
823 if (p == _WaitSet) { // found at head
824 assert(q == NULL, "invariant");
825 _WaitSet = p->ListNext;
826 } else { // found in interior
827 assert(q->ListNext == p, "invariant");
828 q->ListNext = p->ListNext;
829 }
830 WasOnWaitSet = 1; // We were *not* notified but instead encountered timeout
831 }
832 Thread::muxRelease(_WaitLock);
833 }
834
835 // Reentry phase - reacquire the lock
836 if (WasOnWaitSet) {
837 // ESelf was previously on the WaitSet but we just unlinked it above
838 // because of a timeout. ESelf is not resident on any list and is not OnDeck
839 assert(_OnDeck != ESelf, "invariant");
840 ILock(Self);
841 } else {
842 // A prior notify() operation moved ESelf from the WaitSet to the cxq.
843 // ESelf is now on the cxq, EntryList or at the OnDeck position.
844 // The following fragment is extracted from Monitor::ILock()
845 for (;;) {
846 if (OrderAccess::load_acquire(&_OnDeck) == ESelf && TrySpin(Self)) break;
847 ParkCommon(ESelf, 0);
848 }
849 assert(_OnDeck == ESelf, "invariant");
850 _OnDeck = NULL;
851 }
852
853 assert(ILocked(), "invariant");
854 return WasOnWaitSet != 0; // return true IFF timeout
855 }
856
857
858 // ON THE VMTHREAD SNEAKING PAST HELD LOCKS:
859 // In particular, there are certain types of global lock that may be held
860 // by a Java thread while it is blocked at a safepoint but before it has
861 // written the _owner field. These locks may be sneakily acquired by the
862 // VM thread during a safepoint to avoid deadlocks. Alternatively, one should
863 // identify all such locks, and ensure that Java threads never block at
864 // safepoints while holding them (_no_safepoint_check_flag). While it
865 // seems as though this could increase the time to reach a safepoint
866 // (or at least increase the mean, if not the variance), the latter
867 // approach might make for a cleaner, more maintainable JVM design.
868 //
869 // Sneaking is vile and reprehensible and should be excised at the 1st
870 // opportunity. It's possible that the need for sneaking could be obviated
871 // as follows. Currently, a thread might (a) while TBIVM, call pthread_mutex_lock
872 // or ILock() thus acquiring the "physical" lock underlying Monitor/Mutex.
873 // (b) stall at the TBIVM exit point as a safepoint is in effect. Critically,
874 // it'll stall at the TBIVM reentry state transition after having acquired the
875 // underlying lock, but before having set _owner and having entered the actual
876 // critical section. The lock-sneaking facility leverages that fact and allowed the
877 // VM thread to logically acquire locks that had already be physically locked by mutators
878 // but where mutators were known blocked by the reentry thread state transition.
879 //
880 // If we were to modify the Monitor-Mutex so that TBIVM state transitions tightly
881 // wrapped calls to park(), then we could likely do away with sneaking. We'd
882 // decouple lock acquisition and parking. The critical invariant to eliminating
883 // sneaking is to ensure that we never "physically" acquire the lock while TBIVM.
884 // An easy way to accomplish this is to wrap the park calls in a narrow TBIVM jacket.
885 // One difficulty with this approach is that the TBIVM wrapper could recurse and
886 // call lock() deep from within a lock() call, while the MutexEvent was already enqueued.
887 // Using a stack (N=2 at minimum) of ParkEvents would take care of that problem.
888 //
889 // But of course the proper ultimate approach is to avoid schemes that require explicit
890 // sneaking or dependence on any any clever invariants or subtle implementation properties
891 // of Mutex-Monitor and instead directly address the underlying design flaw.
892
893 void Monitor::lock(Thread * Self) {
894 // Ensure that the Monitor requires/allows safepoint checks.
895 assert(_safepoint_check_required != Monitor::_safepoint_check_never,
896 "This lock should never have a safepoint check: %s", name());
897
898 #ifdef CHECK_UNHANDLED_OOPS
899 // Clear unhandled oops so we get a crash right away. Only clear for non-vm
900 // or GC threads.
901 if (Self->is_Java_thread()) {
902 Self->clear_unhandled_oops();
903 }
904 #endif // CHECK_UNHANDLED_OOPS
905
906 debug_only(check_prelock_state(Self));
907 assert(_owner != Self, "invariant");
908 assert(_OnDeck != Self->_MutexEvent, "invariant");
909
910 if (TryFast()) {
911 Exeunt:
912 assert(ILocked(), "invariant");
913 assert(owner() == NULL, "invariant");
914 set_owner(Self);
915 return;
916 }
917
918 // The lock is contended ...
919
920 bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint();
921 if (can_sneak && _owner == NULL) {
922 // a java thread has locked the lock but has not entered the
923 // critical region -- let's just pretend we've locked the lock
924 // and go on. we note this with _snuck so we can also
925 // pretend to unlock when the time comes.
926 _snuck = true;
927 goto Exeunt;
928 }
929
930 // Try a brief spin to avoid passing thru thread state transition ...
931 if (TrySpin(Self)) goto Exeunt;
932
933 check_block_state(Self);
934 if (Self->is_Java_thread()) {
935 // Horrible dictu - we suffer through a state transition
936 assert(rank() > Mutex::special, "Potential deadlock with special or lesser rank mutex");
937 ThreadBlockInVM tbivm((JavaThread *) Self);
938 ILock(Self);
939 } else {
940 // Mirabile dictu
941 ILock(Self);
942 }
943 goto Exeunt;
944 }
945
946 void Monitor::lock() {
947 this->lock(Thread::current());
948 }
949
950 // Lock without safepoint check - a degenerate variant of lock().
951 // Should ONLY be used by safepoint code and other code
952 // that is guaranteed not to block while running inside the VM. If this is called with
953 // thread state set to be in VM, the safepoint synchronization code will deadlock!
954
955 void Monitor::lock_without_safepoint_check(Thread * Self) {
956 // Ensure that the Monitor does not require or allow safepoint checks.
957 assert(_safepoint_check_required != Monitor::_safepoint_check_always,
958 "This lock should always have a safepoint check: %s", name());
959 assert(_owner != Self, "invariant");
960 ILock(Self);
961 assert(_owner == NULL, "invariant");
962 set_owner(Self);
963 }
964
965 void Monitor::lock_without_safepoint_check() {
966 lock_without_safepoint_check(Thread::current());
967 }
968
969
970 // Returns true if thread succeeds in grabbing the lock, otherwise false.
971
972 bool Monitor::try_lock() {
973 Thread * const Self = Thread::current();
974 debug_only(check_prelock_state(Self));
975 // assert(!thread->is_inside_signal_handler(), "don't lock inside signal handler");
976
977 // Special case, where all Java threads are stopped.
978 // The lock may have been acquired but _owner is not yet set.
979 // In that case the VM thread can safely grab the lock.
980 // It strikes me this should appear _after the TryLock() fails, below.
981 bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint();
982 if (can_sneak && _owner == NULL) {
983 set_owner(Self); // Do not need to be atomic, since we are at a safepoint
984 _snuck = true;
985 return true;
986 }
987
988 if (TryLock()) {
989 // We got the lock
990 assert(_owner == NULL, "invariant");
991 set_owner(Self);
992 return true;
993 }
994 return false;
995 }
996
997 void Monitor::unlock() {
998 assert(_owner == Thread::current(), "invariant");
999 assert(_OnDeck != Thread::current()->_MutexEvent, "invariant");
1000 set_owner(NULL);
1001 if (_snuck) {
1002 assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak");
1003 _snuck = false;
1004 return;
1005 }
1006 IUnlock(false);
1007 }
1008
1009 // Yet another degenerate version of Monitor::lock() or lock_without_safepoint_check()
1010 // jvm_raw_lock() and _unlock() can be called by non-Java threads via JVM_RawMonitorEnter.
1011 //
1012 // There's no expectation that JVM_RawMonitors will interoperate properly with the native
1013 // Mutex-Monitor constructs. We happen to implement JVM_RawMonitors in terms of
1014 // native Mutex-Monitors simply as a matter of convenience. A simple abstraction layer
1015 // over a pthread_mutex_t would work equally as well, but require more platform-specific
1016 // code -- a "PlatformMutex". Alternatively, a simply layer over muxAcquire-muxRelease
1017 // would work too.
1018 //
1019 // Since the caller might be a foreign thread, we don't necessarily have a Thread.MutexEvent
1020 // instance available. Instead, we transiently allocate a ParkEvent on-demand if
1021 // we encounter contention. That ParkEvent remains associated with the thread
1022 // until it manages to acquire the lock, at which time we return the ParkEvent
1023 // to the global ParkEvent free list. This is correct and suffices for our purposes.
1024 //
1025 // Beware that the original jvm_raw_unlock() had a "_snuck" test but that
1026 // jvm_raw_lock() didn't have the corresponding test. I suspect that's an
1027 // oversight, but I've replicated the original suspect logic in the new code ...
1028
1029 void Monitor::jvm_raw_lock() {
1030 assert(rank() == native, "invariant");
1031
1032 if (TryLock()) {
1033 Exeunt:
1034 assert(ILocked(), "invariant");
1035 assert(_owner == NULL, "invariant");
1036 // This can potentially be called by non-java Threads. Thus, the Thread::current_or_null()
1037 // might return NULL. Don't call set_owner since it will break on an NULL owner
1038 // Consider installing a non-null "ANON" distinguished value instead of just NULL.
1039 _owner = Thread::current_or_null();
1040 return;
1041 }
1042
1043 if (TrySpin(NULL)) goto Exeunt;
1044
1045 // slow-path - apparent contention
1046 // Allocate a ParkEvent for transient use.
1047 // The ParkEvent remains associated with this thread until
1048 // the time the thread manages to acquire the lock.
1049 ParkEvent * const ESelf = ParkEvent::Allocate(NULL);
1050 ESelf->reset();
1051 OrderAccess::storeload();
1052
1053 // Either Enqueue Self on cxq or acquire the outer lock.
1054 if (AcquireOrPush (ESelf)) {
1055 ParkEvent::Release(ESelf); // surrender the ParkEvent
1056 goto Exeunt;
1057 }
1058
1059 // At any given time there is at most one ondeck thread.
1060 // ondeck implies not resident on cxq and not resident on EntryList
1061 // Only the OnDeck thread can try to acquire -- contend for -- the lock.
1062 // CONSIDER: use Self->OnDeck instead of m->OnDeck.
1063 for (;;) {
1064 if (OrderAccess::load_acquire(&_OnDeck) == ESelf && TrySpin(NULL)) break;
1065 ParkCommon(ESelf, 0);
1066 }
1067
1068 assert(_OnDeck == ESelf, "invariant");
1069 _OnDeck = NULL;
1070 ParkEvent::Release(ESelf); // surrender the ParkEvent
1071 goto Exeunt;
1072 }
1073
1074 void Monitor::jvm_raw_unlock() {
1075 // Nearly the same as Monitor::unlock() ...
1076 // directly set _owner instead of using set_owner(null)
1077 _owner = NULL;
1078 if (_snuck) { // ???
1079 assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak");
1080 _snuck = false;
1081 return;
1082 }
1083 IUnlock(false);
1084 }
1085
1086 bool Monitor::wait(bool no_safepoint_check, long timeout,
1087 bool as_suspend_equivalent) {
1088 // Make sure safepoint checking is used properly.
1089 assert(!(_safepoint_check_required == Monitor::_safepoint_check_never && no_safepoint_check == false),
1090 "This lock should never have a safepoint check: %s", name());
1091 assert(!(_safepoint_check_required == Monitor::_safepoint_check_always && no_safepoint_check == true),
1092 "This lock should always have a safepoint check: %s", name());
1093
1094 Thread * const Self = Thread::current();
1095 assert(_owner == Self, "invariant");
1096 assert(ILocked(), "invariant");
1097
1098 // as_suspend_equivalent logically implies !no_safepoint_check
1099 guarantee(!as_suspend_equivalent || !no_safepoint_check, "invariant");
1100 // !no_safepoint_check logically implies java_thread
1101 guarantee(no_safepoint_check || Self->is_Java_thread(), "invariant");
1102
1103 #ifdef ASSERT
1104 Monitor * least = get_least_ranked_lock_besides_this(Self->owned_locks());
1105 assert(least != this, "Specification of get_least_... call above");
1106 if (least != NULL && least->rank() <= special) {
1107 tty->print("Attempting to wait on monitor %s/%d while holding"
1108 " lock %s/%d -- possible deadlock",
1109 name(), rank(), least->name(), least->rank());
1110 assert(false, "Shouldn't block(wait) while holding a lock of rank special");
1111 }
1112 #endif // ASSERT
1113
1114 int wait_status;
1115 // conceptually set the owner to NULL in anticipation of
1116 // abdicating the lock in wait
1117 set_owner(NULL);
1118 if (no_safepoint_check) {
1119 wait_status = IWait(Self, timeout);
1120 } else {
1121 assert(Self->is_Java_thread(), "invariant");
1122 JavaThread *jt = (JavaThread *)Self;
1123
1124 // Enter safepoint region - ornate and Rococo ...
1125 ThreadBlockInVM tbivm(jt);
1126 OSThreadWaitState osts(Self->osthread(), false /* not Object.wait() */);
1127
1128 if (as_suspend_equivalent) {
1129 jt->set_suspend_equivalent();
1130 // cleared by handle_special_suspend_equivalent_condition() or
1131 // java_suspend_self()
1132 }
1133
1134 wait_status = IWait(Self, timeout);
1135
1136 // were we externally suspended while we were waiting?
1137 if (as_suspend_equivalent && jt->handle_special_suspend_equivalent_condition()) {
1138 // Our event wait has finished and we own the lock, but
1139 // while we were waiting another thread suspended us. We don't
1140 // want to hold the lock while suspended because that
1141 // would surprise the thread that suspended us.
1142 assert(ILocked(), "invariant");
1143 IUnlock(true);
1144 jt->java_suspend_self();
1145 ILock(Self);
1146 assert(ILocked(), "invariant");
1147 }
1148 }
1149
1150 // Conceptually reestablish ownership of the lock.
1151 // The "real" lock -- the LockByte -- was reacquired by IWait().
1152 assert(ILocked(), "invariant");
1153 assert(_owner == NULL, "invariant");
1154 set_owner(Self);
1155 return wait_status != 0; // return true IFF timeout
1156 }
1157
1158 Monitor::~Monitor() {
1159 #ifdef ASSERT
1160 uintptr_t owner = UNS(_owner);
1161 uintptr_t lockword = UNS(_LockWord.FullWord);
1162 uintptr_t entrylist = UNS(_EntryList);
1163 uintptr_t waitset = UNS(_WaitSet);
1164 uintptr_t ondeck = UNS(_OnDeck);
1165 // Print _name with precision limit, in case failure is due to memory
1166 // corruption that also trashed _name.
1167 assert((owner|lockword|entrylist|waitset|ondeck) == 0,
1168 "%.*s: _owner(" INTPTR_FORMAT ")|_LockWord(" INTPTR_FORMAT ")|_EntryList(" INTPTR_FORMAT ")|_WaitSet("
1169 INTPTR_FORMAT ")|_OnDeck(" INTPTR_FORMAT ") != 0",
1170 MONITOR_NAME_LEN, _name, owner, lockword, entrylist, waitset, ondeck);
1171 #endif
1172 }
1173
1174 void Monitor::ClearMonitor(Monitor * m, const char *name) {
1175 m->_owner = NULL;
1176 m->_snuck = false;
1177 if (name == NULL) {
1178 strcpy(m->_name, "UNKNOWN");
1179 } else {
1180 strncpy(m->_name, name, MONITOR_NAME_LEN - 1);
1181 m->_name[MONITOR_NAME_LEN - 1] = '\0';
1182 }
1183 m->_LockWord.FullWord = 0;
1184 m->_EntryList = NULL;
1185 m->_OnDeck = NULL;
1186 m->_WaitSet = NULL;
1187 m->_WaitLock[0] = 0;
1188 }
1189
1190 Monitor::Monitor() { ClearMonitor(this); }
1191
1192 Monitor::Monitor(int Rank, const char * name, bool allow_vm_block,
1193 SafepointCheckRequired safepoint_check_required) {
1194 ClearMonitor(this, name);
1195 #ifdef ASSERT
1196 _allow_vm_block = allow_vm_block;
1197 _rank = Rank;
1198 NOT_PRODUCT(_safepoint_check_required = safepoint_check_required;)
1199 #endif
1200 }
1201
1202 Mutex::Mutex(int Rank, const char * name, bool allow_vm_block,
1203 SafepointCheckRequired safepoint_check_required) {
1204 ClearMonitor((Monitor *) this, name);
1205 #ifdef ASSERT
1206 _allow_vm_block = allow_vm_block;
1207 _rank = Rank;
1208 NOT_PRODUCT(_safepoint_check_required = safepoint_check_required;)
1209 #endif
1210 }
1211
1212 bool Monitor::owned_by_self() const {
1213 bool ret = _owner == Thread::current();
1214 assert(!ret || _LockWord.Bytes[_LSBINDEX] != 0, "invariant");
1215 return ret;
1216 }
1217
1218 void Monitor::print_on_error(outputStream* st) const {
1219 st->print("[" PTR_FORMAT, p2i(this));
1220 st->print("] %s", _name);
1221 st->print(" - owner thread: " PTR_FORMAT, p2i(_owner));
1222 }
1223
1224
1225
1226
1227 // ----------------------------------------------------------------------------------
1228 // Non-product code
1229
1230 #ifndef PRODUCT
1231 void Monitor::print_on(outputStream* st) const {
1232 st->print_cr("Mutex: [" PTR_FORMAT "/" PTR_FORMAT "] %s - owner: " PTR_FORMAT,
1233 p2i(this), _LockWord.FullWord, _name, p2i(_owner));
1234 }
1235 #endif
1236
1237 #ifndef PRODUCT
1238 #ifdef ASSERT
1239 Monitor * Monitor::get_least_ranked_lock(Monitor * locks) {
1240 Monitor *res, *tmp;
1241 for (res = tmp = locks; tmp != NULL; tmp = tmp->next()) {
1242 if (tmp->rank() < res->rank()) {
1243 res = tmp;
1244 }
1245 }
1246 if (!SafepointSynchronize::is_at_safepoint()) {
1247 // In this case, we expect the held locks to be
1248 // in increasing rank order (modulo any native ranks)
1249 for (tmp = locks; tmp != NULL; tmp = tmp->next()) {
1250 if (tmp->next() != NULL) {
1251 assert(tmp->rank() == Mutex::native ||
1252 tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?");
1253 }
1254 }
1255 }
1256 return res;
1257 }
1258
1259 Monitor* Monitor::get_least_ranked_lock_besides_this(Monitor* locks) {
1260 Monitor *res, *tmp;
1261 for (res = NULL, tmp = locks; tmp != NULL; tmp = tmp->next()) {
1262 if (tmp != this && (res == NULL || tmp->rank() < res->rank())) {
1263 res = tmp;
1264 }
1265 }
1266 if (!SafepointSynchronize::is_at_safepoint()) {
1267 // In this case, we expect the held locks to be
1268 // in increasing rank order (modulo any native ranks)
1269 for (tmp = locks; tmp != NULL; tmp = tmp->next()) {
1270 if (tmp->next() != NULL) {
1271 assert(tmp->rank() == Mutex::native ||
1272 tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?");
1273 }
1274 }
1275 }
1276 return res;
1277 }
1278
1279
1280 bool Monitor::contains(Monitor* locks, Monitor * lock) {
1281 for (; locks != NULL; locks = locks->next()) {
1282 if (locks == lock) {
1283 return true;
1284 }
1285 }
1286 return false;
1287 }
1288 #endif
1289
1290 // Called immediately after lock acquisition or release as a diagnostic
1291 // to track the lock-set of the thread and test for rank violations that
1292 // might indicate exposure to deadlock.
1293 // Rather like an EventListener for _owner (:>).
1294
1295 void Monitor::set_owner_implementation(Thread *new_owner) {
1296 // This function is solely responsible for maintaining
1297 // and checking the invariant that threads and locks
1298 // are in a 1/N relation, with some some locks unowned.
1299 // It uses the Mutex::_owner, Mutex::_next, and
1300 // Thread::_owned_locks fields, and no other function
1301 // changes those fields.
1302 // It is illegal to set the mutex from one non-NULL
1303 // owner to another--it must be owned by NULL as an
1304 // intermediate state.
1305
1306 if (new_owner != NULL) {
1307 // the thread is acquiring this lock
1308
1309 assert(new_owner == Thread::current(), "Should I be doing this?");
1310 assert(_owner == NULL, "setting the owner thread of an already owned mutex");
1311 _owner = new_owner; // set the owner
1312
1313 // link "this" into the owned locks list
1314
1315 #ifdef ASSERT // Thread::_owned_locks is under the same ifdef
1316 Monitor* locks = get_least_ranked_lock(new_owner->owned_locks());
1317 // Mutex::set_owner_implementation is a friend of Thread
1318
1319 assert(this->rank() >= 0, "bad lock rank");
1320
1321 // Deadlock avoidance rules require us to acquire Mutexes only in
1322 // a global total order. For example m1 is the lowest ranked mutex
1323 // that the thread holds and m2 is the mutex the thread is trying
1324 // to acquire, then deadlock avoidance rules require that the rank
1325 // of m2 be less than the rank of m1.
1326 // The rank Mutex::native is an exception in that it is not subject
1327 // to the verification rules.
1328 // Here are some further notes relating to mutex acquisition anomalies:
1329 // . it is also ok to acquire Safepoint_lock at the very end while we
1330 // already hold Terminator_lock - may happen because of periodic safepoints
1331 if (this->rank() != Mutex::native &&
1332 this->rank() != Mutex::suspend_resume &&
1333 locks != NULL && locks->rank() <= this->rank() &&
1334 !SafepointSynchronize::is_at_safepoint() &&
1335 !(this == Safepoint_lock && contains(locks, Terminator_lock) &&
1336 SafepointSynchronize::is_synchronizing())) {
1337 new_owner->print_owned_locks();
1338 fatal("acquiring lock %s/%d out of order with lock %s/%d -- "
1339 "possible deadlock", this->name(), this->rank(),
1340 locks->name(), locks->rank());
1341 }
1342
1343 this->_next = new_owner->_owned_locks;
1344 new_owner->_owned_locks = this;
1345 #endif
1346
1347 } else {
1348 // the thread is releasing this lock
1349
1350 Thread* old_owner = _owner;
1351 debug_only(_last_owner = old_owner);
1352
1353 assert(old_owner != NULL, "removing the owner thread of an unowned mutex");
1354 assert(old_owner == Thread::current(), "removing the owner thread of an unowned mutex");
1355
1356 _owner = NULL; // set the owner
1357
1358 #ifdef ASSERT
1359 Monitor *locks = old_owner->owned_locks();
1360
1361 // remove "this" from the owned locks list
1362
1363 Monitor *prev = NULL;
1364 bool found = false;
1365 for (; locks != NULL; prev = locks, locks = locks->next()) {
1366 if (locks == this) {
1367 found = true;
1368 break;
1369 }
1370 }
1371 assert(found, "Removing a lock not owned");
1372 if (prev == NULL) {
1373 old_owner->_owned_locks = _next;
1374 } else {
1375 prev->_next = _next;
1376 }
1377 _next = NULL;
1378 #endif
1379 }
1380 }
1381
1382
1383 // Factored out common sanity checks for locking mutex'es. Used by lock() and try_lock()
1384 void Monitor::check_prelock_state(Thread *thread) {
1385 assert((!thread->is_Java_thread() || ((JavaThread *)thread)->thread_state() == _thread_in_vm)
1386 || rank() == Mutex::special, "wrong thread state for using locks");
1387 if (StrictSafepointChecks) {
1388 if (thread->is_VM_thread() && !allow_vm_block()) {
1389 fatal("VM thread using lock %s (not allowed to block on)", name());
1390 }
1391 debug_only(if (rank() != Mutex::special) \
1392 thread->check_for_valid_safepoint_state(false);)
1393 }
1394 assert(!os::ThreadCrashProtection::is_crash_protected(thread),
1395 "locking not allowed when crash protection is set");
1396 }
1397
1398 void Monitor::check_block_state(Thread *thread) {
1399 if (!_allow_vm_block && thread->is_VM_thread()) {
1400 warning("VM thread blocked on lock");
1401 print();
1402 BREAKPOINT;
1403 }
1404 assert(_owner != thread, "deadlock: blocking on monitor owned by current thread");
1405 }
1406
1407 #endif // PRODUCT