Mercurial > hg > truffle
annotate src/share/vm/runtime/mutex.cpp @ 4142:bc8527f3071c
Adjust code base to new level of warnings.
author | Thomas Wuerthinger <thomas.wuerthinger@oracle.com> |
---|---|
date | Sun, 18 Dec 2011 05:24:06 +0100 |
parents | f08d439fab8c |
children | eccc4b1f8945 |
rev | line source |
---|---|
0 | 1 |
2 /* | |
1972 | 3 * Copyright (c) 1998, 2010, Oracle and/or its affiliates. All rights reserved. |
0 | 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 * | |
1552
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1490
diff
changeset
|
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1490
diff
changeset
|
21 * or visit www.oracle.com if you need additional information or have any |
c18cbe5936b8
6941466: Oracle rebranding changes for Hotspot repositories
trims
parents:
1490
diff
changeset
|
22 * questions. |
0 | 23 * |
24 */ | |
25 | |
1972 | 26 #include "precompiled.hpp" |
27 #include "runtime/mutex.hpp" | |
28 #include "runtime/osThread.hpp" | |
29 #include "utilities/events.hpp" | |
30 #ifdef TARGET_OS_FAMILY_linux | |
31 # include "mutex_linux.inline.hpp" | |
32 # include "thread_linux.inline.hpp" | |
33 #endif | |
34 #ifdef TARGET_OS_FAMILY_solaris | |
35 # include "mutex_solaris.inline.hpp" | |
36 # include "thread_solaris.inline.hpp" | |
37 #endif | |
38 #ifdef TARGET_OS_FAMILY_windows | |
39 # include "mutex_windows.inline.hpp" | |
40 # include "thread_windows.inline.hpp" | |
41 #endif | |
3960 | 42 #ifdef TARGET_OS_FAMILY_bsd |
43 # include "mutex_bsd.inline.hpp" | |
44 # include "thread_bsd.inline.hpp" | |
45 #endif | |
0 | 46 |
47 // 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 | |
48 // | |
49 // Native Monitor-Mutex locking - theory of operations | |
50 // | |
51 // * Native Monitors are completely unrelated to Java-level monitors, | |
52 // although the "back-end" slow-path implementations share a common lineage. | |
53 // See objectMonitor:: in synchronizer.cpp. | |
54 // Native Monitors do *not* support nesting or recursion but otherwise | |
55 // they're basically Hoare-flavor monitors. | |
56 // | |
57 // * A thread acquires ownership of a Monitor/Mutex by CASing the LockByte | |
58 // in the _LockWord from zero to non-zero. Note that the _Owner field | |
59 // is advisory and is used only to verify that the thread calling unlock() | |
60 // is indeed the last thread to have acquired the lock. | |
61 // | |
62 // * Contending threads "push" themselves onto the front of the contention | |
63 // queue -- called the cxq -- with CAS and then spin/park. | |
64 // The _LockWord contains the LockByte as well as the pointer to the head | |
65 // of the cxq. Colocating the LockByte with the cxq precludes certain races. | |
66 // | |
67 // * Using a separately addressable LockByte allows for CAS:MEMBAR or CAS:0 | |
68 // idioms. We currently use MEMBAR in the uncontended unlock() path, as | |
69 // MEMBAR often has less latency than CAS. If warranted, we could switch to | |
70 // a CAS:0 mode, using timers to close the resultant race, as is done | |
71 // with Java Monitors in synchronizer.cpp. | |
72 // | |
73 // See the following for a discussion of the relative cost of atomics (CAS) | |
74 // MEMBAR, and ways to eliminate such instructions from the common-case paths: | |
75 // -- http://blogs.sun.com/dave/entry/biased_locking_in_hotspot | |
76 // -- http://blogs.sun.com/dave/resource/MustangSync.pdf | |
77 // -- http://blogs.sun.com/dave/resource/synchronization-public2.pdf | |
78 // -- synchronizer.cpp | |
79 // | |
80 // * Overall goals - desiderata | |
81 // 1. Minimize context switching | |
82 // 2. Minimize lock migration | |
83 // 3. Minimize CPI -- affinity and locality | |
84 // 4. Minimize the execution of high-latency instructions such as CAS or MEMBAR | |
85 // 5. Minimize outer lock hold times | |
86 // 6. Behave gracefully on a loaded system | |
87 // | |
88 // * Thread flow and list residency: | |
89 // | |
90 // Contention queue --> EntryList --> OnDeck --> Owner --> !Owner | |
91 // [..resident on monitor list..] | |
92 // [...........contending..................] | |
93 // | |
94 // -- The contention queue (cxq) contains recently-arrived threads (RATs). | |
95 // Threads on the cxq eventually drain into the EntryList. | |
96 // -- Invariant: a thread appears on at most one list -- cxq, EntryList | |
97 // or WaitSet -- at any one time. | |
98 // -- For a given monitor there can be at most one "OnDeck" thread at any | |
99 // given time but if needbe this particular invariant could be relaxed. | |
100 // | |
101 // * The WaitSet and EntryList linked lists are composed of ParkEvents. | |
102 // I use ParkEvent instead of threads as ParkEvents are immortal and | |
103 // type-stable, meaning we can safely unpark() a possibly stale | |
104 // list element in the unlock()-path. (That's benign). | |
105 // | |
106 // * Succession policy - providing for progress: | |
107 // | |
108 // As necessary, the unlock()ing thread identifies, unlinks, and unparks | |
109 // an "heir presumptive" tentative successor thread from the EntryList. | |
110 // This becomes the so-called "OnDeck" thread, of which there can be only | |
111 // one at any given time for a given monitor. The wakee will recontend | |
112 // for ownership of monitor. | |
113 // | |
114 // Succession is provided for by a policy of competitive handoff. | |
115 // The exiting thread does _not_ grant or pass ownership to the | |
116 // successor thread. (This is also referred to as "handoff" succession"). | |
117 // Instead the exiting thread releases ownership and possibly wakes | |
118 // a successor, so the successor can (re)compete for ownership of the lock. | |
119 // | |
120 // Competitive handoff provides excellent overall throughput at the expense | |
121 // of short-term fairness. If fairness is a concern then one remedy might | |
122 // be to add an AcquireCounter field to the monitor. After a thread acquires | |
123 // the lock it will decrement the AcquireCounter field. When the count | |
124 // reaches 0 the thread would reset the AcquireCounter variable, abdicate | |
125 // the lock directly to some thread on the EntryList, and then move itself to the | |
126 // tail of the EntryList. | |
127 // | |
128 // But in practice most threads engage or otherwise participate in resource | |
129 // bounded producer-consumer relationships, so lock domination is not usually | |
130 // a practical concern. Recall too, that in general it's easier to construct | |
131 // a fair lock from a fast lock, but not vice-versa. | |
132 // | |
133 // * The cxq can have multiple concurrent "pushers" but only one concurrent | |
134 // detaching thread. This mechanism is immune from the ABA corruption. | |
135 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. | |
136 // We use OnDeck as a pseudo-lock to enforce the at-most-one detaching | |
137 // thread constraint. | |
138 // | |
139 // * Taken together, the cxq and the EntryList constitute or form a | |
140 // single logical queue of threads stalled trying to acquire the lock. | |
141 // We use two distinct lists to reduce heat on the list ends. | |
142 // Threads in lock() enqueue onto cxq while threads in unlock() will | |
143 // dequeue from the EntryList. (c.f. Michael Scott's "2Q" algorithm). | |
144 // A key desideratum is to minimize queue & monitor metadata manipulation | |
145 // that occurs while holding the "outer" monitor lock -- that is, we want to | |
146 // minimize monitor lock holds times. | |
147 // | |
148 // The EntryList is ordered by the prevailing queue discipline and | |
149 // can be organized in any convenient fashion, such as a doubly-linked list or | |
150 // a circular doubly-linked list. If we need a priority queue then something akin | |
151 // to Solaris' sleepq would work nicely. Viz., | |
152 // -- http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. | |
153 // -- http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/uts/common/os/sleepq.c | |
154 // Queue discipline is enforced at ::unlock() time, when the unlocking thread | |
155 // drains the cxq into the EntryList, and orders or reorders the threads on the | |
156 // EntryList accordingly. | |
157 // | |
158 // Barring "lock barging", this mechanism provides fair cyclic ordering, | |
159 // somewhat similar to an elevator-scan. | |
160 // | |
161 // * OnDeck | |
162 // -- For a given monitor there can be at most one OnDeck thread at any given | |
163 // instant. The OnDeck thread is contending for the lock, but has been | |
164 // unlinked from the EntryList and cxq by some previous unlock() operations. | |
165 // Once a thread has been designated the OnDeck thread it will remain so | |
166 // until it manages to acquire the lock -- being OnDeck is a stable property. | |
167 // -- Threads on the EntryList or cxq are _not allowed to attempt lock acquisition. | |
168 // -- OnDeck also serves as an "inner lock" as follows. Threads in unlock() will, after | |
169 // having cleared the LockByte and dropped the outer lock, attempt to "trylock" | |
170 // OnDeck by CASing the field from null to non-null. If successful, that thread | |
171 // is then responsible for progress and succession and can use CAS to detach and | |
172 // drain the cxq into the EntryList. By convention, only this thread, the holder of | |
173 // the OnDeck inner lock, can manipulate the EntryList or detach and drain the | |
174 // RATs on the cxq into the EntryList. This avoids ABA corruption on the cxq as | |
175 // we allow multiple concurrent "push" operations but restrict detach concurrency | |
176 // to at most one thread. Having selected and detached a successor, the thread then | |
177 // changes the OnDeck to refer to that successor, and then unparks the successor. | |
178 // That successor will eventually acquire the lock and clear OnDeck. Beware | |
179 // that the OnDeck usage as a lock is asymmetric. A thread in unlock() transiently | |
180 // "acquires" OnDeck, performs queue manipulations, passes OnDeck to some successor, | |
181 // and then the successor eventually "drops" OnDeck. Note that there's never | |
182 // any sense of contention on the inner lock, however. Threads never contend | |
183 // or wait for the inner lock. | |
184 // -- OnDeck provides for futile wakeup throttling a described in section 3.3 of | |
185 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf | |
186 // In a sense, OnDeck subsumes the ObjectMonitor _Succ and ObjectWaiter | |
187 // TState fields found in Java-level objectMonitors. (See synchronizer.cpp). | |
188 // | |
189 // * Waiting threads reside on the WaitSet list -- wait() puts | |
190 // the caller onto the WaitSet. Notify() or notifyAll() simply | |
191 // transfers threads from the WaitSet to either the EntryList or cxq. | |
192 // Subsequent unlock() operations will eventually unpark the notifyee. | |
193 // Unparking a notifee in notify() proper is inefficient - if we were to do so | |
194 // it's likely the notifyee would simply impale itself on the lock held | |
195 // by the notifier. | |
196 // | |
197 // * The mechanism is obstruction-free in that if the holder of the transient | |
198 // OnDeck lock in unlock() is preempted or otherwise stalls, other threads | |
199 // can still acquire and release the outer lock and continue to make progress. | |
200 // At worst, waking of already blocked contending threads may be delayed, | |
201 // but nothing worse. (We only use "trylock" operations on the inner OnDeck | |
202 // lock). | |
203 // | |
204 // * Note that thread-local storage must be initialized before a thread | |
205 // uses Native monitors or mutexes. The native monitor-mutex subsystem | |
206 // depends on Thread::current(). | |
207 // | |
208 // * The monitor synchronization subsystem avoids the use of native | |
209 // synchronization primitives except for the narrow platform-specific | |
210 // park-unpark abstraction. See the comments in os_solaris.cpp regarding | |
211 // the semantics of park-unpark. Put another way, this monitor implementation | |
212 // depends only on atomic operations and park-unpark. The monitor subsystem | |
213 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the | |
214 // underlying OS manages the READY<->RUN transitions. | |
215 // | |
216 // * The memory consistency model provide by lock()-unlock() is at least as | |
217 // strong or stronger than the Java Memory model defined by JSR-133. | |
218 // That is, we guarantee at least entry consistency, if not stronger. | |
219 // See http://g.oswego.edu/dl/jmm/cookbook.html. | |
220 // | |
221 // * Thread:: currently contains a set of purpose-specific ParkEvents: | |
222 // _MutexEvent, _ParkEvent, etc. A better approach might be to do away with | |
223 // the purpose-specific ParkEvents and instead implement a general per-thread | |
224 // stack of available ParkEvents which we could provision on-demand. The | |
225 // stack acts as a local cache to avoid excessive calls to ParkEvent::Allocate() | |
226 // and ::Release(). A thread would simply pop an element from the local stack before it | |
227 // enqueued or park()ed. When the contention was over the thread would | |
228 // push the no-longer-needed ParkEvent back onto its stack. | |
229 // | |
230 // * A slightly reduced form of ILock() and IUnlock() have been partially | |
231 // model-checked (Murphi) for safety and progress at T=1,2,3 and 4. | |
232 // It'd be interesting to see if TLA/TLC could be useful as well. | |
233 // | |
234 // * Mutex-Monitor is a low-level "leaf" subsystem. That is, the monitor | |
235 // code should never call other code in the JVM that might itself need to | |
236 // acquire monitors or mutexes. That's true *except* in the case of the | |
237 // ThreadBlockInVM state transition wrappers. The ThreadBlockInVM DTOR handles | |
238 // mutator reentry (ingress) by checking for a pending safepoint in which case it will | |
239 // call SafepointSynchronize::block(), which in turn may call Safepoint_lock->lock(), etc. | |
240 // In that particular case a call to lock() for a given Monitor can end up recursively | |
241 // calling lock() on another monitor. While distasteful, this is largely benign | |
242 // as the calls come from jacket that wraps lock(), and not from deep within lock() itself. | |
243 // | |
244 // It's unfortunate that native mutexes and thread state transitions were convolved. | |
245 // They're really separate concerns and should have remained that way. Melding | |
246 // them together was facile -- a bit too facile. The current implementation badly | |
247 // conflates the two concerns. | |
248 // | |
249 // * TODO-FIXME: | |
250 // | |
251 // -- Add DTRACE probes for contended acquire, contended acquired, contended unlock | |
252 // We should also add DTRACE probes in the ParkEvent subsystem for | |
253 // Park-entry, Park-exit, and Unpark. | |
254 // | |
255 // -- We have an excess of mutex-like constructs in the JVM, namely: | |
256 // 1. objectMonitors for Java-level synchronization (synchronizer.cpp) | |
257 // 2. low-level muxAcquire and muxRelease | |
258 // 3. low-level spinAcquire and spinRelease | |
259 // 4. native Mutex:: and Monitor:: | |
260 // 5. jvm_raw_lock() and _unlock() | |
261 // 6. JVMTI raw monitors -- distinct from (5) despite having a confusingly | |
262 // similar name. | |
263 // | |
264 // 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 | |
265 | |
266 | |
267 // CASPTR() uses the canonical argument order that dominates in the literature. | |
268 // Our internal cmpxchg_ptr() uses a bastardized ordering to accommodate Sun .il templates. | |
269 | |
270 #define CASPTR(a,c,s) intptr_t(Atomic::cmpxchg_ptr ((void *)(s),(void *)(a),(void *)(c))) | |
271 #define UNS(x) (uintptr_t(x)) | |
272 #define TRACE(m) { static volatile int ctr = 0 ; int x = ++ctr ; if ((x & (x-1))==0) { ::printf ("%d:%s\n", x, #m); ::fflush(stdout); }} | |
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 inline jint MarsagliaXOR (jint * const a) { | |
287 jint x = *a ; | |
288 if (x == 0) x = UNS(a)|1 ; | |
289 x ^= x << 6; | |
290 x ^= ((unsigned)x) >> 21; | |
291 x ^= x << 7 ; | |
292 *a = x ; | |
293 return x & 0x7FFFFFFF ; | |
294 } | |
295 | |
296 static int Stall (int its) { | |
297 static volatile jint rv = 1 ; | |
298 volatile int OnFrame = 0 ; | |
299 jint v = rv ^ UNS(OnFrame) ; | |
300 while (--its >= 0) { | |
301 v = MarsagliaXORV (v) ; | |
302 } | |
303 // Make this impossible for the compiler to optimize away, | |
304 // but (mostly) avoid W coherency sharing on MP systems. | |
305 if (v == 0x12345) rv = v ; | |
306 return v ; | |
307 } | |
308 | |
309 int Monitor::TryLock () { | |
310 intptr_t v = _LockWord.FullWord ; | |
311 for (;;) { | |
312 if ((v & _LBIT) != 0) return 0 ; | |
313 const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ; | |
314 if (v == u) return 1 ; | |
315 v = u ; | |
316 } | |
317 } | |
318 | |
319 int Monitor::TryFast () { | |
320 // Optimistic fast-path form ... | |
321 // Fast-path attempt for the common uncontended case. | |
322 // Avoid RTS->RTO $ coherence upgrade on typical SMP systems. | |
323 intptr_t v = CASPTR (&_LockWord, 0, _LBIT) ; // agro ... | |
324 if (v == 0) return 1 ; | |
325 | |
326 for (;;) { | |
327 if ((v & _LBIT) != 0) return 0 ; | |
328 const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ; | |
329 if (v == u) return 1 ; | |
330 v = u ; | |
331 } | |
332 } | |
333 | |
334 int Monitor::ILocked () { | |
335 const intptr_t w = _LockWord.FullWord & 0xFF ; | |
336 assert (w == 0 || w == _LBIT, "invariant") ; | |
337 return w == _LBIT ; | |
338 } | |
339 | |
340 // Polite TATAS spinlock with exponential backoff - bounded spin. | |
341 // Ideally we'd use processor cycles, time or vtime to control | |
342 // the loop, but we currently use iterations. | |
343 // All the constants within were derived empirically but work over | |
344 // over the spectrum of J2SE reference platforms. | |
345 // On Niagara-class systems the back-off is unnecessary but | |
346 // is relatively harmless. (At worst it'll slightly retard | |
347 // acquisition times). The back-off is critical for older SMP systems | |
348 // where constant fetching of the LockWord would otherwise impair | |
349 // scalability. | |
350 // | |
351 // Clamp spinning at approximately 1/2 of a context-switch round-trip. | |
352 // See synchronizer.cpp for details and rationale. | |
353 | |
354 int Monitor::TrySpin (Thread * const Self) { | |
355 if (TryLock()) return 1 ; | |
356 if (!os::is_MP()) return 0 ; | |
357 | |
358 int Probes = 0 ; | |
359 int Delay = 0 ; | |
360 int Steps = 0 ; | |
361 int SpinMax = NativeMonitorSpinLimit ; | |
362 int flgs = NativeMonitorFlags ; | |
363 for (;;) { | |
364 intptr_t v = _LockWord.FullWord; | |
365 if ((v & _LBIT) == 0) { | |
366 if (CASPTR (&_LockWord, v, v|_LBIT) == v) { | |
367 return 1 ; | |
368 } | |
369 continue ; | |
370 } | |
371 | |
372 if ((flgs & 8) == 0) { | |
373 SpinPause () ; | |
374 } | |
375 | |
376 // Periodically increase Delay -- variable Delay form | |
377 // conceptually: delay *= 1 + 1/Exponent | |
378 ++ Probes; | |
379 if (Probes > SpinMax) return 0 ; | |
380 | |
381 if ((Probes & 0x7) == 0) { | |
382 Delay = ((Delay << 1)|1) & 0x7FF ; | |
383 // CONSIDER: Delay += 1 + (Delay/4); Delay &= 0x7FF ; | |
384 } | |
385 | |
386 if (flgs & 2) continue ; | |
387 | |
388 // Consider checking _owner's schedctl state, if OFFPROC abort spin. | |
389 // If the owner is OFFPROC then it's unlike that the lock will be dropped | |
390 // in a timely fashion, which suggests that spinning would not be fruitful | |
391 // or profitable. | |
392 | |
393 // Stall for "Delay" time units - iterations in the current implementation. | |
394 // Avoid generating coherency traffic while stalled. | |
395 // Possible ways to delay: | |
396 // PAUSE, SLEEP, MEMBAR #sync, MEMBAR #halt, | |
397 // wr %g0,%asi, gethrtime, rdstick, rdtick, rdtsc, etc. ... | |
398 // Note that on Niagara-class systems we want to minimize STs in the | |
399 // spin loop. N1 and brethren write-around the L1$ over the xbar into the L2$. | |
400 // Furthermore, they don't have a W$ like traditional SPARC processors. | |
401 // We currently use a Marsaglia Shift-Xor RNG loop. | |
402 Steps += Delay ; | |
403 if (Self != NULL) { | |
404 jint rv = Self->rng[0] ; | |
405 for (int k = Delay ; --k >= 0; ) { | |
406 rv = MarsagliaXORV (rv) ; | |
407 if ((flgs & 4) == 0 && SafepointSynchronize::do_call_back()) return 0 ; | |
408 } | |
409 Self->rng[0] = rv ; | |
410 } else { | |
411 Stall (Delay) ; | |
412 } | |
413 } | |
414 } | |
415 | |
416 static int ParkCommon (ParkEvent * ev, jlong timo) { | |
417 // Diagnostic support - periodically unwedge blocked threads | |
418 intx nmt = NativeMonitorTimeout ; | |
419 if (nmt > 0 && (nmt < timo || timo <= 0)) { | |
420 timo = nmt ; | |
421 } | |
422 int err = OS_OK ; | |
423 if (0 == timo) { | |
424 ev->park() ; | |
425 } else { | |
426 err = ev->park(timo) ; | |
427 } | |
428 return err ; | |
429 } | |
430 | |
431 inline int Monitor::AcquireOrPush (ParkEvent * ESelf) { | |
432 intptr_t v = _LockWord.FullWord ; | |
433 for (;;) { | |
434 if ((v & _LBIT) == 0) { | |
435 const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ; | |
436 if (u == v) return 1 ; // indicate acquired | |
437 v = u ; | |
438 } else { | |
439 // Anticipate success ... | |
440 ESelf->ListNext = (ParkEvent *) (v & ~_LBIT) ; | |
441 const intptr_t u = CASPTR (&_LockWord, v, intptr_t(ESelf)|_LBIT) ; | |
442 if (u == v) return 0 ; // indicate pushed onto cxq | |
443 v = u ; | |
444 } | |
445 // Interference - LockWord change - just retry | |
446 } | |
447 } | |
448 | |
449 // ILock and IWait are the lowest level primitive internal blocking | |
450 // synchronization functions. The callers of IWait and ILock must have | |
451 // performed any needed state transitions beforehand. | |
452 // IWait and ILock may directly call park() without any concern for thread state. | |
453 // Note that ILock and IWait do *not* access _owner. | |
454 // _owner is a higher-level logical concept. | |
455 | |
456 void Monitor::ILock (Thread * Self) { | |
457 assert (_OnDeck != Self->_MutexEvent, "invariant") ; | |
458 | |
459 if (TryFast()) { | |
460 Exeunt: | |
461 assert (ILocked(), "invariant") ; | |
462 return ; | |
463 } | |
464 | |
465 ParkEvent * const ESelf = Self->_MutexEvent ; | |
466 assert (_OnDeck != ESelf, "invariant") ; | |
467 | |
468 // As an optimization, spinners could conditionally try to set ONDECK to _LBIT | |
469 // Synchronizer.cpp uses a similar optimization. | |
470 if (TrySpin (Self)) goto Exeunt ; | |
471 | |
472 // Slow-path - the lock is contended. | |
473 // Either Enqueue Self on cxq or acquire the outer lock. | |
474 // LockWord encoding = (cxq,LOCKBYTE) | |
475 ESelf->reset() ; | |
476 OrderAccess::fence() ; | |
477 | |
478 // Optional optimization ... try barging on the inner lock | |
479 if ((NativeMonitorFlags & 32) && CASPTR (&_OnDeck, NULL, UNS(Self)) == 0) { | |
480 goto OnDeck_LOOP ; | |
481 } | |
482 | |
483 if (AcquireOrPush (ESelf)) goto Exeunt ; | |
484 | |
485 // At any given time there is at most one ondeck thread. | |
486 // ondeck implies not resident on cxq and not resident on EntryList | |
487 // Only the OnDeck thread can try to acquire -- contended for -- the lock. | |
488 // CONSIDER: use Self->OnDeck instead of m->OnDeck. | |
489 // Deschedule Self so that others may run. | |
490 while (_OnDeck != ESelf) { | |
491 ParkCommon (ESelf, 0) ; | |
492 } | |
493 | |
494 // Self is now in the ONDECK position and will remain so until it | |
495 // manages to acquire the lock. | |
496 OnDeck_LOOP: | |
497 for (;;) { | |
498 assert (_OnDeck == ESelf, "invariant") ; | |
499 if (TrySpin (Self)) break ; | |
500 // CONSIDER: if ESelf->TryPark() && TryLock() break ... | |
501 // It's probably wise to spin only if we *actually* blocked | |
502 // CONSIDER: check the lockbyte, if it remains set then | |
503 // preemptively drain the cxq into the EntryList. | |
504 // The best place and time to perform queue operations -- lock metadata -- | |
505 // is _before having acquired the outer lock, while waiting for the lock to drop. | |
506 ParkCommon (ESelf, 0) ; | |
507 } | |
508 | |
509 assert (_OnDeck == ESelf, "invariant") ; | |
510 _OnDeck = NULL ; | |
511 | |
512 // Note that we current drop the inner lock (clear OnDeck) in the slow-path | |
513 // epilog immediately after having acquired the outer lock. | |
514 // But instead we could consider the following optimizations: | |
515 // A. Shift or defer dropping the inner lock until the subsequent IUnlock() operation. | |
516 // This might avoid potential reacquisition of the inner lock in IUlock(). | |
517 // B. While still holding the inner lock, attempt to opportunistically select | |
518 // and unlink the next ONDECK thread from the EntryList. | |
519 // If successful, set ONDECK to refer to that thread, otherwise clear ONDECK. | |
520 // It's critical that the select-and-unlink operation run in constant-time as | |
521 // it executes when holding the outer lock and may artificially increase the | |
522 // effective length of the critical section. | |
523 // Note that (A) and (B) are tantamount to succession by direct handoff for | |
524 // the inner lock. | |
525 goto Exeunt ; | |
526 } | |
527 | |
528 void Monitor::IUnlock (bool RelaxAssert) { | |
529 assert (ILocked(), "invariant") ; | |
530 _LockWord.Bytes[_LSBINDEX] = 0 ; // drop outer lock | |
531 OrderAccess::storeload (); | |
532 ParkEvent * const w = _OnDeck ; | |
533 assert (RelaxAssert || w != Thread::current()->_MutexEvent, "invariant") ; | |
534 if (w != NULL) { | |
535 // Either we have a valid ondeck thread or ondeck is transiently "locked" | |
536 // by some exiting thread as it arranges for succession. The LSBit of | |
537 // OnDeck allows us to discriminate two cases. If the latter, the | |
538 // responsibility for progress and succession lies with that other thread. | |
539 // For good performance, we also depend on the fact that redundant unpark() | |
540 // operations are cheap. That is, repeated Unpark()ing of the ONDECK thread | |
541 // is inexpensive. This approach provides implicit futile wakeup throttling. | |
542 // Note that the referent "w" might be stale with respect to the lock. | |
543 // In that case the following unpark() is harmless and the worst that'll happen | |
544 // is a spurious return from a park() operation. Critically, if "w" _is stale, | |
545 // then progress is known to have occurred as that means the thread associated | |
546 // with "w" acquired the lock. In that case this thread need take no further | |
547 // action to guarantee progress. | |
548 if ((UNS(w) & _LBIT) == 0) w->unpark() ; | |
549 return ; | |
550 } | |
551 | |
552 intptr_t cxq = _LockWord.FullWord ; | |
553 if (((cxq & ~_LBIT)|UNS(_EntryList)) == 0) { | |
554 return ; // normal fast-path exit - cxq and EntryList both empty | |
555 } | |
556 if (cxq & _LBIT) { | |
557 // Optional optimization ... | |
558 // Some other thread acquired the lock in the window since this | |
559 // thread released it. Succession is now that thread's responsibility. | |
560 return ; | |
561 } | |
562 | |
563 Succession: | |
564 // Slow-path exit - this thread must ensure succession and progress. | |
565 // OnDeck serves as lock to protect cxq and EntryList. | |
566 // Only the holder of OnDeck can manipulate EntryList or detach the RATs from cxq. | |
567 // Avoid ABA - allow multiple concurrent producers (enqueue via push-CAS) | |
568 // but only one concurrent consumer (detacher of RATs). | |
569 // Consider protecting this critical section with schedctl on Solaris. | |
570 // Unlike a normal lock, however, the exiting thread "locks" OnDeck, | |
571 // picks a successor and marks that thread as OnDeck. That successor | |
572 // thread will then clear OnDeck once it eventually acquires the outer lock. | |
573 if (CASPTR (&_OnDeck, NULL, _LBIT) != UNS(NULL)) { | |
574 return ; | |
575 } | |
576 | |
577 ParkEvent * List = _EntryList ; | |
578 if (List != NULL) { | |
579 // Transfer the head of the EntryList to the OnDeck position. | |
580 // Once OnDeck, a thread stays OnDeck until it acquires the lock. | |
581 // For a given lock there is at most OnDeck thread at any one instant. | |
582 WakeOne: | |
583 assert (List == _EntryList, "invariant") ; | |
584 ParkEvent * const w = List ; | |
585 assert (RelaxAssert || w != Thread::current()->_MutexEvent, "invariant") ; | |
586 _EntryList = w->ListNext ; | |
587 // as a diagnostic measure consider setting w->_ListNext = BAD | |
588 assert (UNS(_OnDeck) == _LBIT, "invariant") ; | |
589 _OnDeck = w ; // pass OnDeck to w. | |
590 // w will clear OnDeck once it acquires the outer lock | |
591 | |
592 // Another optional optimization ... | |
593 // For heavily contended locks it's not uncommon that some other | |
594 // thread acquired the lock while this thread was arranging succession. | |
595 // Try to defer the unpark() operation - Delegate the responsibility | |
596 // for unpark()ing the OnDeck thread to the current or subsequent owners | |
597 // That is, the new owner is responsible for unparking the OnDeck thread. | |
598 OrderAccess::storeload() ; | |
599 cxq = _LockWord.FullWord ; | |
600 if (cxq & _LBIT) return ; | |
601 | |
602 w->unpark() ; | |
603 return ; | |
604 } | |
605 | |
606 cxq = _LockWord.FullWord ; | |
607 if ((cxq & ~_LBIT) != 0) { | |
608 // The EntryList is empty but the cxq is populated. | |
609 // drain RATs from cxq into EntryList | |
610 // Detach RATs segment with CAS and then merge into EntryList | |
611 for (;;) { | |
612 // optional optimization - if locked, the owner is responsible for succession | |
613 if (cxq & _LBIT) goto Punt ; | |
614 const intptr_t vfy = CASPTR (&_LockWord, cxq, cxq & _LBIT) ; | |
615 if (vfy == cxq) break ; | |
616 cxq = vfy ; | |
617 // Interference - LockWord changed - Just retry | |
618 // We can see concurrent interference from contending threads | |
619 // pushing themselves onto the cxq or from lock-unlock operations. | |
620 // From the perspective of this thread, EntryList is stable and | |
621 // the cxq is prepend-only -- the head is volatile but the interior | |
622 // of the cxq is stable. In theory if we encounter interference from threads | |
623 // pushing onto cxq we could simply break off the original cxq suffix and | |
624 // move that segment to the EntryList, avoiding a 2nd or multiple CAS attempts | |
625 // on the high-traffic LockWord variable. For instance lets say the cxq is "ABCD" | |
626 // when we first fetch cxq above. Between the fetch -- where we observed "A" | |
627 // -- and CAS -- where we attempt to CAS null over A -- "PQR" arrive, | |
628 // yielding cxq = "PQRABCD". In this case we could simply set A.ListNext | |
629 // null, leaving cxq = "PQRA" and transfer the "BCD" segment to the EntryList. | |
630 // Note too, that it's safe for this thread to traverse the cxq | |
631 // without taking any special concurrency precautions. | |
632 } | |
633 | |
634 // We don't currently reorder the cxq segment as we move it onto | |
635 // the EntryList, but it might make sense to reverse the order | |
636 // or perhaps sort by thread priority. See the comments in | |
637 // synchronizer.cpp objectMonitor::exit(). | |
638 assert (_EntryList == NULL, "invariant") ; | |
639 _EntryList = List = (ParkEvent *)(cxq & ~_LBIT) ; | |
640 assert (List != NULL, "invariant") ; | |
641 goto WakeOne ; | |
642 } | |
643 | |
644 // cxq|EntryList is empty. | |
645 // w == NULL implies that cxq|EntryList == NULL in the past. | |
646 // Possible race - rare inopportune interleaving. | |
647 // A thread could have added itself to cxq since this thread previously checked. | |
648 // Detect and recover by refetching cxq. | |
649 Punt: | |
650 assert (UNS(_OnDeck) == _LBIT, "invariant") ; | |
651 _OnDeck = NULL ; // Release inner lock. | |
652 OrderAccess::storeload(); // Dekker duality - pivot point | |
653 | |
654 // Resample LockWord/cxq to recover from possible race. | |
655 // For instance, while this thread T1 held OnDeck, some other thread T2 might | |
656 // acquire the outer lock. Another thread T3 might try to acquire the outer | |
657 // lock, but encounter contention and enqueue itself on cxq. T2 then drops the | |
658 // outer lock, but skips succession as this thread T1 still holds OnDeck. | |
659 // T1 is and remains responsible for ensuring succession of T3. | |
660 // | |
661 // Note that we don't need to recheck EntryList, just cxq. | |
662 // If threads moved onto EntryList since we dropped OnDeck | |
663 // that implies some other thread forced succession. | |
664 cxq = _LockWord.FullWord ; | |
665 if ((cxq & ~_LBIT) != 0 && (cxq & _LBIT) == 0) { | |
666 goto Succession ; // potential race -- re-run succession | |
667 } | |
668 return ; | |
669 } | |
670 | |
671 bool Monitor::notify() { | |
672 assert (_owner == Thread::current(), "invariant") ; | |
673 assert (ILocked(), "invariant") ; | |
674 if (_WaitSet == NULL) return true ; | |
675 NotifyCount ++ ; | |
676 | |
677 // Transfer one thread from the WaitSet to the EntryList or cxq. | |
678 // Currently we just unlink the head of the WaitSet and prepend to the cxq. | |
679 // And of course we could just unlink it and unpark it, too, but | |
680 // in that case it'd likely impale itself on the reentry. | |
681 Thread::muxAcquire (_WaitLock, "notify:WaitLock") ; | |
682 ParkEvent * nfy = _WaitSet ; | |
683 if (nfy != NULL) { // DCL idiom | |
684 _WaitSet = nfy->ListNext ; | |
685 assert (nfy->Notified == 0, "invariant") ; | |
686 // push nfy onto the cxq | |
687 for (;;) { | |
688 const intptr_t v = _LockWord.FullWord ; | |
689 assert ((v & 0xFF) == _LBIT, "invariant") ; | |
690 nfy->ListNext = (ParkEvent *)(v & ~_LBIT); | |
691 if (CASPTR (&_LockWord, v, UNS(nfy)|_LBIT) == v) break; | |
692 // interference - _LockWord changed -- just retry | |
693 } | |
694 // Note that setting Notified before pushing nfy onto the cxq is | |
695 // also legal and safe, but the safety properties are much more | |
696 // subtle, so for the sake of code stewardship ... | |
697 OrderAccess::fence() ; | |
698 nfy->Notified = 1; | |
699 } | |
700 Thread::muxRelease (_WaitLock) ; | |
701 if (nfy != NULL && (NativeMonitorFlags & 16)) { | |
702 // Experimental code ... light up the wakee in the hope that this thread (the owner) | |
703 // will drop the lock just about the time the wakee comes ONPROC. | |
704 nfy->unpark() ; | |
705 } | |
706 assert (ILocked(), "invariant") ; | |
707 return true ; | |
708 } | |
709 | |
710 // Currently notifyAll() transfers the waiters one-at-a-time from the waitset | |
711 // to the cxq. This could be done more efficiently with a single bulk en-mass transfer, | |
712 // but in practice notifyAll() for large #s of threads is rare and not time-critical. | |
713 // Beware too, that we invert the order of the waiters. Lets say that the | |
714 // waitset is "ABCD" and the cxq is "XYZ". After a notifyAll() the waitset | |
715 // will be empty and the cxq will be "DCBAXYZ". This is benign, of course. | |
716 | |
717 bool Monitor::notify_all() { | |
718 assert (_owner == Thread::current(), "invariant") ; | |
719 assert (ILocked(), "invariant") ; | |
720 while (_WaitSet != NULL) notify() ; | |
721 return true ; | |
722 } | |
723 | |
724 int Monitor::IWait (Thread * Self, jlong timo) { | |
725 assert (ILocked(), "invariant") ; | |
726 | |
727 // Phases: | |
728 // 1. Enqueue Self on WaitSet - currently prepend | |
729 // 2. unlock - drop the outer lock | |
730 // 3. wait for either notification or timeout | |
731 // 4. lock - reentry - reacquire the outer lock | |
732 | |
733 ParkEvent * const ESelf = Self->_MutexEvent ; | |
734 ESelf->Notified = 0 ; | |
735 ESelf->reset() ; | |
736 OrderAccess::fence() ; | |
737 | |
738 // Add Self to WaitSet | |
739 // Ideally only the holder of the outer lock would manipulate the WaitSet - | |
740 // That is, the outer lock would implicitly protect the WaitSet. | |
741 // But if a thread in wait() encounters a timeout it will need to dequeue itself | |
742 // from the WaitSet _before it becomes the owner of the lock. We need to dequeue | |
743 // as the ParkEvent -- which serves as a proxy for the thread -- can't reside | |
744 // on both the WaitSet and the EntryList|cxq at the same time.. That is, a thread | |
745 // on the WaitSet can't be allowed to compete for the lock until it has managed to | |
746 // unlink its ParkEvent from WaitSet. Thus the need for WaitLock. | |
747 // Contention on the WaitLock is minimal. | |
748 // | |
749 // Another viable approach would be add another ParkEvent, "WaitEvent" to the | |
750 // thread class. The WaitSet would be composed of WaitEvents. Only the | |
751 // owner of the outer lock would manipulate the WaitSet. A thread in wait() | |
752 // could then compete for the outer lock, and then, if necessary, unlink itself | |
753 // from the WaitSet only after having acquired the outer lock. More precisely, | |
754 // there would be no WaitLock. A thread in in wait() would enqueue its WaitEvent | |
755 // on the WaitSet; release the outer lock; wait for either notification or timeout; | |
756 // reacquire the inner lock; and then, if needed, unlink itself from the WaitSet. | |
757 // | |
758 // Alternatively, a 2nd set of list link fields in the ParkEvent might suffice. | |
759 // One set would be for the WaitSet and one for the EntryList. | |
760 // We could also deconstruct the ParkEvent into a "pure" event and add a | |
761 // new immortal/TSM "ListElement" class that referred to ParkEvents. | |
762 // In that case we could have one ListElement on the WaitSet and another | |
763 // on the EntryList, with both referring to the same pure Event. | |
764 | |
765 Thread::muxAcquire (_WaitLock, "wait:WaitLock:Add") ; | |
766 ESelf->ListNext = _WaitSet ; | |
767 _WaitSet = ESelf ; | |
768 Thread::muxRelease (_WaitLock) ; | |
769 | |
770 // Release the outer lock | |
771 // We call IUnlock (RelaxAssert=true) as a thread T1 might | |
772 // enqueue itself on the WaitSet, call IUnlock(), drop the lock, | |
773 // and then stall before it can attempt to wake a successor. | |
774 // Some other thread T2 acquires the lock, and calls notify(), moving | |
775 // T1 from the WaitSet to the cxq. T2 then drops the lock. T1 resumes, | |
776 // and then finds *itself* on the cxq. During the course of a normal | |
777 // IUnlock() call a thread should _never find itself on the EntryList | |
778 // or cxq, but in the case of wait() it's possible. | |
779 // See synchronizer.cpp objectMonitor::wait(). | |
780 IUnlock (true) ; | |
781 | |
782 // Wait for either notification or timeout | |
783 // Beware that in some circumstances we might propagate | |
784 // spurious wakeups back to the caller. | |
785 | |
786 for (;;) { | |
787 if (ESelf->Notified) break ; | |
788 int err = ParkCommon (ESelf, timo) ; | |
789 if (err == OS_TIMEOUT || (NativeMonitorFlags & 1)) break ; | |
790 } | |
791 | |
792 // Prepare for reentry - if necessary, remove ESelf from WaitSet | |
793 // ESelf can be: | |
794 // 1. Still on the WaitSet. This can happen if we exited the loop by timeout. | |
795 // 2. On the cxq or EntryList | |
796 // 3. Not resident on cxq, EntryList or WaitSet, but in the OnDeck position. | |
797 | |
798 OrderAccess::fence() ; | |
799 int WasOnWaitSet = 0 ; | |
800 if (ESelf->Notified == 0) { | |
801 Thread::muxAcquire (_WaitLock, "wait:WaitLock:remove") ; | |
802 if (ESelf->Notified == 0) { // DCL idiom | |
803 assert (_OnDeck != ESelf, "invariant") ; // can't be both OnDeck and on WaitSet | |
804 // ESelf is resident on the WaitSet -- unlink it. | |
805 // A doubly-linked list would be better here so we can unlink in constant-time. | |
806 // We have to unlink before we potentially recontend as ESelf might otherwise | |
807 // end up on the cxq|EntryList -- it can't be on two lists at once. | |
808 ParkEvent * p = _WaitSet ; | |
809 ParkEvent * q = NULL ; // classic q chases p | |
810 while (p != NULL && p != ESelf) { | |
811 q = p ; | |
812 p = p->ListNext ; | |
813 } | |
814 assert (p == ESelf, "invariant") ; | |
815 if (p == _WaitSet) { // found at head | |
816 assert (q == NULL, "invariant") ; | |
817 _WaitSet = p->ListNext ; | |
818 } else { // found in interior | |
819 assert (q->ListNext == p, "invariant") ; | |
820 q->ListNext = p->ListNext ; | |
821 } | |
822 WasOnWaitSet = 1 ; // We were *not* notified but instead encountered timeout | |
823 } | |
824 Thread::muxRelease (_WaitLock) ; | |
825 } | |
826 | |
827 // Reentry phase - reacquire the lock | |
828 if (WasOnWaitSet) { | |
829 // ESelf was previously on the WaitSet but we just unlinked it above | |
830 // because of a timeout. ESelf is not resident on any list and is not OnDeck | |
831 assert (_OnDeck != ESelf, "invariant") ; | |
832 ILock (Self) ; | |
833 } else { | |
834 // A prior notify() operation moved ESelf from the WaitSet to the cxq. | |
835 // ESelf is now on the cxq, EntryList or at the OnDeck position. | |
836 // The following fragment is extracted from Monitor::ILock() | |
837 for (;;) { | |
838 if (_OnDeck == ESelf && TrySpin(Self)) break ; | |
839 ParkCommon (ESelf, 0) ; | |
840 } | |
841 assert (_OnDeck == ESelf, "invariant") ; | |
842 _OnDeck = NULL ; | |
843 } | |
844 | |
845 assert (ILocked(), "invariant") ; | |
846 return WasOnWaitSet != 0 ; // return true IFF timeout | |
847 } | |
848 | |
849 | |
850 // ON THE VMTHREAD SNEAKING PAST HELD LOCKS: | |
851 // In particular, there are certain types of global lock that may be held | |
852 // by a Java thread while it is blocked at a safepoint but before it has | |
853 // written the _owner field. These locks may be sneakily acquired by the | |
854 // VM thread during a safepoint to avoid deadlocks. Alternatively, one should | |
855 // identify all such locks, and ensure that Java threads never block at | |
856 // safepoints while holding them (_no_safepoint_check_flag). While it | |
857 // seems as though this could increase the time to reach a safepoint | |
858 // (or at least increase the mean, if not the variance), the latter | |
859 // approach might make for a cleaner, more maintainable JVM design. | |
860 // | |
861 // Sneaking is vile and reprehensible and should be excised at the 1st | |
862 // opportunity. It's possible that the need for sneaking could be obviated | |
863 // as follows. Currently, a thread might (a) while TBIVM, call pthread_mutex_lock | |
864 // or ILock() thus acquiring the "physical" lock underlying Monitor/Mutex. | |
865 // (b) stall at the TBIVM exit point as a safepoint is in effect. Critically, | |
866 // it'll stall at the TBIVM reentry state transition after having acquired the | |
867 // underlying lock, but before having set _owner and having entered the actual | |
868 // critical section. The lock-sneaking facility leverages that fact and allowed the | |
869 // VM thread to logically acquire locks that had already be physically locked by mutators | |
870 // but where mutators were known blocked by the reentry thread state transition. | |
871 // | |
872 // If we were to modify the Monitor-Mutex so that TBIVM state transitions tightly | |
873 // wrapped calls to park(), then we could likely do away with sneaking. We'd | |
874 // decouple lock acquisition and parking. The critical invariant to eliminating | |
875 // sneaking is to ensure that we never "physically" acquire the lock while TBIVM. | |
876 // An easy way to accomplish this is to wrap the park calls in a narrow TBIVM jacket. | |
877 // One difficulty with this approach is that the TBIVM wrapper could recurse and | |
878 // call lock() deep from within a lock() call, while the MutexEvent was already enqueued. | |
879 // Using a stack (N=2 at minimum) of ParkEvents would take care of that problem. | |
880 // | |
881 // But of course the proper ultimate approach is to avoid schemes that require explicit | |
882 // sneaking or dependence on any any clever invariants or subtle implementation properties | |
883 // of Mutex-Monitor and instead directly address the underlying design flaw. | |
884 | |
885 void Monitor::lock (Thread * Self) { | |
886 #ifdef CHECK_UNHANDLED_OOPS | |
887 // Clear unhandled oops so we get a crash right away. Only clear for non-vm | |
888 // or GC threads. | |
889 if (Self->is_Java_thread()) { | |
890 Self->clear_unhandled_oops(); | |
891 } | |
892 #endif // CHECK_UNHANDLED_OOPS | |
893 | |
894 debug_only(check_prelock_state(Self)); | |
895 assert (_owner != Self , "invariant") ; | |
896 assert (_OnDeck != Self->_MutexEvent, "invariant") ; | |
897 | |
898 if (TryFast()) { | |
899 Exeunt: | |
900 assert (ILocked(), "invariant") ; | |
901 assert (owner() == NULL, "invariant"); | |
902 set_owner (Self); | |
903 return ; | |
904 } | |
905 | |
906 // The lock is contended ... | |
907 | |
908 bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint(); | |
909 if (can_sneak && _owner == NULL) { | |
910 // a java thread has locked the lock but has not entered the | |
911 // critical region -- let's just pretend we've locked the lock | |
912 // and go on. we note this with _snuck so we can also | |
913 // pretend to unlock when the time comes. | |
914 _snuck = true; | |
915 goto Exeunt ; | |
916 } | |
917 | |
918 // Try a brief spin to avoid passing thru thread state transition ... | |
919 if (TrySpin (Self)) goto Exeunt ; | |
920 | |
921 check_block_state(Self); | |
922 if (Self->is_Java_thread()) { | |
923 // Horribile dictu - we suffer through a state transition | |
924 assert(rank() > Mutex::special, "Potential deadlock with special or lesser rank mutex"); | |
925 ThreadBlockInVM tbivm ((JavaThread *) Self) ; | |
926 ILock (Self) ; | |
927 } else { | |
928 // Mirabile dictu | |
929 ILock (Self) ; | |
930 } | |
931 goto Exeunt ; | |
932 } | |
933 | |
934 void Monitor::lock() { | |
935 this->lock(Thread::current()); | |
936 } | |
937 | |
938 // Lock without safepoint check - a degenerate variant of lock(). | |
939 // Should ONLY be used by safepoint code and other code | |
940 // that is guaranteed not to block while running inside the VM. If this is called with | |
941 // thread state set to be in VM, the safepoint synchronization code will deadlock! | |
942 | |
943 void Monitor::lock_without_safepoint_check (Thread * Self) { | |
944 assert (_owner != Self, "invariant") ; | |
945 ILock (Self) ; | |
946 assert (_owner == NULL, "invariant"); | |
947 set_owner (Self); | |
948 } | |
949 | |
950 void Monitor::lock_without_safepoint_check () { | |
951 lock_without_safepoint_check (Thread::current()) ; | |
952 } | |
953 | |
954 | |
955 // Returns true if thread succeceed [sic] in grabbing the lock, otherwise false. | |
956 | |
957 bool Monitor::try_lock() { | |
958 Thread * const Self = Thread::current(); | |
959 debug_only(check_prelock_state(Self)); | |
960 // assert(!thread->is_inside_signal_handler(), "don't lock inside signal handler"); | |
961 | |
962 // Special case, where all Java threads are stopped. | |
963 // The lock may have been acquired but _owner is not yet set. | |
964 // In that case the VM thread can safely grab the lock. | |
965 // It strikes me this should appear _after the TryLock() fails, below. | |
966 bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint(); | |
967 if (can_sneak && _owner == NULL) { | |
968 set_owner(Self); // Do not need to be atomic, since we are at a safepoint | |
969 _snuck = true; | |
970 return true; | |
971 } | |
972 | |
973 if (TryLock()) { | |
974 // We got the lock | |
975 assert (_owner == NULL, "invariant"); | |
976 set_owner (Self); | |
977 return true; | |
978 } | |
979 return false; | |
980 } | |
981 | |
982 void Monitor::unlock() { | |
983 assert (_owner == Thread::current(), "invariant") ; | |
984 assert (_OnDeck != Thread::current()->_MutexEvent , "invariant") ; | |
985 set_owner (NULL) ; | |
986 if (_snuck) { | |
987 assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak"); | |
988 _snuck = false; | |
989 return ; | |
990 } | |
991 IUnlock (false) ; | |
992 } | |
993 | |
994 // Yet another degenerate version of Monitor::lock() or lock_without_safepoint_check() | |
995 // jvm_raw_lock() and _unlock() can be called by non-Java threads via JVM_RawMonitorEnter. | |
996 // | |
997 // There's no expectation that JVM_RawMonitors will interoperate properly with the native | |
998 // Mutex-Monitor constructs. We happen to implement JVM_RawMonitors in terms of | |
999 // native Mutex-Monitors simply as a matter of convenience. A simple abstraction layer | |
1000 // over a pthread_mutex_t would work equally as well, but require more platform-specific | |
1001 // code -- a "PlatformMutex". Alternatively, a simply layer over muxAcquire-muxRelease | |
1002 // would work too. | |
1003 // | |
1004 // Since the caller might be a foreign thread, we don't necessarily have a Thread.MutexEvent | |
1005 // instance available. Instead, we transiently allocate a ParkEvent on-demand if | |
1006 // we encounter contention. That ParkEvent remains associated with the thread | |
1007 // until it manages to acquire the lock, at which time we return the ParkEvent | |
1008 // to the global ParkEvent free list. This is correct and suffices for our purposes. | |
1009 // | |
1010 // Beware that the original jvm_raw_unlock() had a "_snuck" test but that | |
1011 // jvm_raw_lock() didn't have the corresponding test. I suspect that's an | |
1012 // oversight, but I've replicated the original suspect logic in the new code ... | |
1013 | |
1014 void Monitor::jvm_raw_lock() { | |
1015 assert(rank() == native, "invariant"); | |
1016 | |
1017 if (TryLock()) { | |
1018 Exeunt: | |
1019 assert (ILocked(), "invariant") ; | |
1020 assert (_owner == NULL, "invariant"); | |
1021 // This can potentially be called by non-java Threads. Thus, the ThreadLocalStorage | |
1022 // might return NULL. Don't call set_owner since it will break on an NULL owner | |
1023 // Consider installing a non-null "ANON" distinguished value instead of just NULL. | |
1024 _owner = ThreadLocalStorage::thread(); | |
1025 return ; | |
1026 } | |
1027 | |
1028 if (TrySpin(NULL)) goto Exeunt ; | |
1029 | |
1030 // slow-path - apparent contention | |
1031 // Allocate a ParkEvent for transient use. | |
1032 // The ParkEvent remains associated with this thread until | |
1033 // the time the thread manages to acquire the lock. | |
1034 ParkEvent * const ESelf = ParkEvent::Allocate(NULL) ; | |
1035 ESelf->reset() ; | |
1036 OrderAccess::storeload() ; | |
1037 | |
1038 // Either Enqueue Self on cxq or acquire the outer lock. | |
1039 if (AcquireOrPush (ESelf)) { | |
1040 ParkEvent::Release (ESelf) ; // surrender the ParkEvent | |
1041 goto Exeunt ; | |
1042 } | |
1043 | |
1044 // At any given time there is at most one ondeck thread. | |
1045 // ondeck implies not resident on cxq and not resident on EntryList | |
1046 // Only the OnDeck thread can try to acquire -- contended for -- the lock. | |
1047 // CONSIDER: use Self->OnDeck instead of m->OnDeck. | |
1048 for (;;) { | |
1049 if (_OnDeck == ESelf && TrySpin(NULL)) break ; | |
1050 ParkCommon (ESelf, 0) ; | |
1051 } | |
1052 | |
1053 assert (_OnDeck == ESelf, "invariant") ; | |
1054 _OnDeck = NULL ; | |
1055 ParkEvent::Release (ESelf) ; // surrender the ParkEvent | |
1056 goto Exeunt ; | |
1057 } | |
1058 | |
1059 void Monitor::jvm_raw_unlock() { | |
1060 // Nearly the same as Monitor::unlock() ... | |
1061 // directly set _owner instead of using set_owner(null) | |
1062 _owner = NULL ; | |
1063 if (_snuck) { // ??? | |
1064 assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak"); | |
1065 _snuck = false; | |
1066 return ; | |
1067 } | |
1068 IUnlock(false) ; | |
1069 } | |
1070 | |
1071 bool Monitor::wait(bool no_safepoint_check, long timeout, bool as_suspend_equivalent) { | |
1072 Thread * const Self = Thread::current() ; | |
1073 assert (_owner == Self, "invariant") ; | |
1074 assert (ILocked(), "invariant") ; | |
1075 | |
1076 // as_suspend_equivalent logically implies !no_safepoint_check | |
1077 guarantee (!as_suspend_equivalent || !no_safepoint_check, "invariant") ; | |
1078 // !no_safepoint_check logically implies java_thread | |
1079 guarantee (no_safepoint_check || Self->is_Java_thread(), "invariant") ; | |
1080 | |
1081 #ifdef ASSERT | |
1082 Monitor * least = get_least_ranked_lock_besides_this(Self->owned_locks()); | |
1083 assert(least != this, "Specification of get_least_... call above"); | |
1084 if (least != NULL && least->rank() <= special) { | |
1085 tty->print("Attempting to wait on monitor %s/%d while holding" | |
1086 " lock %s/%d -- possible deadlock", | |
1087 name(), rank(), least->name(), least->rank()); | |
1088 assert(false, "Shouldn't block(wait) while holding a lock of rank special"); | |
1089 } | |
1090 #endif // ASSERT | |
1091 | |
1092 int wait_status ; | |
1093 // conceptually set the owner to NULL in anticipation of | |
1094 // abdicating the lock in wait | |
1095 set_owner(NULL); | |
1096 if (no_safepoint_check) { | |
1097 wait_status = IWait (Self, timeout) ; | |
1098 } else { | |
1099 assert (Self->is_Java_thread(), "invariant") ; | |
1100 JavaThread *jt = (JavaThread *)Self; | |
1101 | |
1102 // Enter safepoint region - ornate and Rococo ... | |
1103 ThreadBlockInVM tbivm(jt); | |
1104 OSThreadWaitState osts(Self->osthread(), false /* not Object.wait() */); | |
1105 | |
1106 if (as_suspend_equivalent) { | |
1107 jt->set_suspend_equivalent(); | |
1108 // cleared by handle_special_suspend_equivalent_condition() or | |
1109 // java_suspend_self() | |
1110 } | |
1111 | |
1112 wait_status = IWait (Self, timeout) ; | |
1113 | |
1114 // were we externally suspended while we were waiting? | |
1115 if (as_suspend_equivalent && jt->handle_special_suspend_equivalent_condition()) { | |
1116 // Our event wait has finished and we own the lock, but | |
1117 // while we were waiting another thread suspended us. We don't | |
1118 // want to hold the lock while suspended because that | |
1119 // would surprise the thread that suspended us. | |
1120 assert (ILocked(), "invariant") ; | |
1121 IUnlock (true) ; | |
1122 jt->java_suspend_self(); | |
1123 ILock (Self) ; | |
1124 assert (ILocked(), "invariant") ; | |
1125 } | |
1126 } | |
1127 | |
1128 // Conceptually reestablish ownership of the lock. | |
1129 // The "real" lock -- the LockByte -- was reacquired by IWait(). | |
1130 assert (ILocked(), "invariant") ; | |
1131 assert (_owner == NULL, "invariant") ; | |
1132 set_owner (Self) ; | |
1133 return wait_status != 0 ; // return true IFF timeout | |
1134 } | |
1135 | |
1136 Monitor::~Monitor() { | |
1137 assert ((UNS(_owner)|UNS(_LockWord.FullWord)|UNS(_EntryList)|UNS(_WaitSet)|UNS(_OnDeck)) == 0, "") ; | |
1138 } | |
1139 | |
55
2a8eb116ebbe
6610420: Debug VM crashes during monitor lock rank checking
xlu
parents:
0
diff
changeset
|
1140 void Monitor::ClearMonitor (Monitor * m, const char *name) { |
0 | 1141 m->_owner = NULL ; |
1142 m->_snuck = false ; | |
55
2a8eb116ebbe
6610420: Debug VM crashes during monitor lock rank checking
xlu
parents:
0
diff
changeset
|
1143 if (name == NULL) { |
2a8eb116ebbe
6610420: Debug VM crashes during monitor lock rank checking
xlu
parents:
0
diff
changeset
|
1144 strcpy(m->_name, "UNKNOWN") ; |
2a8eb116ebbe
6610420: Debug VM crashes during monitor lock rank checking
xlu
parents:
0
diff
changeset
|
1145 } else { |
2a8eb116ebbe
6610420: Debug VM crashes during monitor lock rank checking
xlu
parents:
0
diff
changeset
|
1146 strncpy(m->_name, name, MONITOR_NAME_LEN - 1); |
2a8eb116ebbe
6610420: Debug VM crashes during monitor lock rank checking
xlu
parents:
0
diff
changeset
|
1147 m->_name[MONITOR_NAME_LEN - 1] = '\0'; |
2a8eb116ebbe
6610420: Debug VM crashes during monitor lock rank checking
xlu
parents:
0
diff
changeset
|
1148 } |
0 | 1149 m->_LockWord.FullWord = 0 ; |
1150 m->_EntryList = NULL ; | |
1151 m->_OnDeck = NULL ; | |
1152 m->_WaitSet = NULL ; | |
1153 m->_WaitLock[0] = 0 ; | |
1154 } | |
1155 | |
1156 Monitor::Monitor() { ClearMonitor(this); } | |
1157 | |
1158 Monitor::Monitor (int Rank, const char * name, bool allow_vm_block) { | |
55
2a8eb116ebbe
6610420: Debug VM crashes during monitor lock rank checking
xlu
parents:
0
diff
changeset
|
1159 ClearMonitor (this, name) ; |
0 | 1160 #ifdef ASSERT |
1161 _allow_vm_block = allow_vm_block; | |
1162 _rank = Rank ; | |
1163 #endif | |
1164 } | |
1165 | |
1166 Mutex::~Mutex() { | |
1167 assert ((UNS(_owner)|UNS(_LockWord.FullWord)|UNS(_EntryList)|UNS(_WaitSet)|UNS(_OnDeck)) == 0, "") ; | |
1168 } | |
1169 | |
1170 Mutex::Mutex (int Rank, const char * name, bool allow_vm_block) { | |
55
2a8eb116ebbe
6610420: Debug VM crashes during monitor lock rank checking
xlu
parents:
0
diff
changeset
|
1171 ClearMonitor ((Monitor *) this, name) ; |
0 | 1172 #ifdef ASSERT |
1173 _allow_vm_block = allow_vm_block; | |
1174 _rank = Rank ; | |
1175 #endif | |
1176 } | |
1177 | |
1178 bool Monitor::owned_by_self() const { | |
1179 bool ret = _owner == Thread::current(); | |
1180 assert (!ret || _LockWord.Bytes[_LSBINDEX] != 0, "invariant") ; | |
1181 return ret; | |
1182 } | |
1183 | |
1184 void Monitor::print_on_error(outputStream* st) const { | |
1185 st->print("[" PTR_FORMAT, this); | |
1186 st->print("] %s", _name); | |
1187 st->print(" - owner thread: " PTR_FORMAT, _owner); | |
1188 } | |
1189 | |
1190 | |
1191 | |
1192 | |
1193 // ---------------------------------------------------------------------------------- | |
1194 // Non-product code | |
1195 | |
1196 #ifndef PRODUCT | |
1197 void Monitor::print_on(outputStream* st) const { | |
1198 st->print_cr("Mutex: [0x%lx/0x%lx] %s - owner: 0x%lx", this, _LockWord.FullWord, _name, _owner); | |
1199 } | |
1200 #endif | |
1201 | |
1202 #ifndef PRODUCT | |
1203 #ifdef ASSERT | |
1204 Monitor * Monitor::get_least_ranked_lock(Monitor * locks) { | |
1205 Monitor *res, *tmp; | |
1206 for (res = tmp = locks; tmp != NULL; tmp = tmp->next()) { | |
1207 if (tmp->rank() < res->rank()) { | |
1208 res = tmp; | |
1209 } | |
1210 } | |
1211 if (!SafepointSynchronize::is_at_safepoint()) { | |
1212 // In this case, we expect the held locks to be | |
1213 // in increasing rank order (modulo any native ranks) | |
1214 for (tmp = locks; tmp != NULL; tmp = tmp->next()) { | |
1215 if (tmp->next() != NULL) { | |
1216 assert(tmp->rank() == Mutex::native || | |
1217 tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?"); | |
1218 } | |
1219 } | |
1220 } | |
1221 return res; | |
1222 } | |
1223 | |
1224 Monitor* Monitor::get_least_ranked_lock_besides_this(Monitor* locks) { | |
1225 Monitor *res, *tmp; | |
1226 for (res = NULL, tmp = locks; tmp != NULL; tmp = tmp->next()) { | |
1227 if (tmp != this && (res == NULL || tmp->rank() < res->rank())) { | |
1228 res = tmp; | |
1229 } | |
1230 } | |
1231 if (!SafepointSynchronize::is_at_safepoint()) { | |
1232 // In this case, we expect the held locks to be | |
1233 // in increasing rank order (modulo any native ranks) | |
1234 for (tmp = locks; tmp != NULL; tmp = tmp->next()) { | |
1235 if (tmp->next() != NULL) { | |
1236 assert(tmp->rank() == Mutex::native || | |
1237 tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?"); | |
1238 } | |
1239 } | |
1240 } | |
1241 return res; | |
1242 } | |
1243 | |
1244 | |
1245 bool Monitor::contains(Monitor* locks, Monitor * lock) { | |
1246 for (; locks != NULL; locks = locks->next()) { | |
1247 if (locks == lock) | |
1248 return true; | |
1249 } | |
1250 return false; | |
1251 } | |
1252 #endif | |
1253 | |
1254 // Called immediately after lock acquisition or release as a diagnostic | |
1255 // to track the lock-set of the thread and test for rank violations that | |
1256 // might indicate exposure to deadlock. | |
1257 // Rather like an EventListener for _owner (:>). | |
1258 | |
1259 void Monitor::set_owner_implementation(Thread *new_owner) { | |
1260 // This function is solely responsible for maintaining | |
1261 // and checking the invariant that threads and locks | |
1262 // are in a 1/N relation, with some some locks unowned. | |
1263 // It uses the Mutex::_owner, Mutex::_next, and | |
1264 // Thread::_owned_locks fields, and no other function | |
1265 // changes those fields. | |
1266 // It is illegal to set the mutex from one non-NULL | |
1267 // owner to another--it must be owned by NULL as an | |
1268 // intermediate state. | |
1269 | |
1270 if (new_owner != NULL) { | |
1271 // the thread is acquiring this lock | |
1272 | |
1273 assert(new_owner == Thread::current(), "Should I be doing this?"); | |
1274 assert(_owner == NULL, "setting the owner thread of an already owned mutex"); | |
1275 _owner = new_owner; // set the owner | |
1276 | |
1277 // link "this" into the owned locks list | |
1278 | |
1279 #ifdef ASSERT // Thread::_owned_locks is under the same ifdef | |
1280 Monitor* locks = get_least_ranked_lock(new_owner->owned_locks()); | |
1281 // Mutex::set_owner_implementation is a friend of Thread | |
1282 | |
1283 assert(this->rank() >= 0, "bad lock rank"); | |
1284 | |
1285 if (LogMultipleMutexLocking && locks != NULL) { | |
1286 Events::log("thread " INTPTR_FORMAT " locks %s, already owns %s", new_owner, name(), locks->name()); | |
1287 } | |
1288 | |
1289 // Deadlock avoidance rules require us to acquire Mutexes only in | |
1290 // a global total order. For example m1 is the lowest ranked mutex | |
1291 // that the thread holds and m2 is the mutex the thread is trying | |
1292 // to acquire, then deadlock avoidance rules require that the rank | |
1293 // of m2 be less than the rank of m1. | |
1294 // The rank Mutex::native is an exception in that it is not subject | |
1295 // to the verification rules. | |
1296 // Here are some further notes relating to mutex acquisition anomalies: | |
1297 // . under Solaris, the interrupt lock gets acquired when doing | |
1298 // profiling, so any lock could be held. | |
1299 // . it is also ok to acquire Safepoint_lock at the very end while we | |
1300 // already hold Terminator_lock - may happen because of periodic safepoints | |
1301 if (this->rank() != Mutex::native && | |
1302 this->rank() != Mutex::suspend_resume && | |
1303 locks != NULL && locks->rank() <= this->rank() && | |
1304 !SafepointSynchronize::is_at_safepoint() && | |
1305 this != Interrupt_lock && this != ProfileVM_lock && | |
1306 !(this == Safepoint_lock && contains(locks, Terminator_lock) && | |
1307 SafepointSynchronize::is_synchronizing())) { | |
1308 new_owner->print_owned_locks(); | |
1490
f03d0a26bf83
6888954: argument formatting for assert() and friends
jcoomes
parents:
196
diff
changeset
|
1309 fatal(err_msg("acquiring lock %s/%d out of order with lock %s/%d -- " |
f03d0a26bf83
6888954: argument formatting for assert() and friends
jcoomes
parents:
196
diff
changeset
|
1310 "possible deadlock", this->name(), this->rank(), |
f03d0a26bf83
6888954: argument formatting for assert() and friends
jcoomes
parents:
196
diff
changeset
|
1311 locks->name(), locks->rank())); |
0 | 1312 } |
1313 | |
1314 this->_next = new_owner->_owned_locks; | |
1315 new_owner->_owned_locks = this; | |
1316 #endif | |
1317 | |
1318 } else { | |
1319 // the thread is releasing this lock | |
1320 | |
1321 Thread* old_owner = _owner; | |
1322 debug_only(_last_owner = old_owner); | |
1323 | |
1324 assert(old_owner != NULL, "removing the owner thread of an unowned mutex"); | |
1325 assert(old_owner == Thread::current(), "removing the owner thread of an unowned mutex"); | |
1326 | |
1327 _owner = NULL; // set the owner | |
1328 | |
1329 #ifdef ASSERT | |
1330 Monitor *locks = old_owner->owned_locks(); | |
1331 | |
1332 if (LogMultipleMutexLocking && locks != this) { | |
1333 Events::log("thread " INTPTR_FORMAT " unlocks %s, still owns %s", old_owner, this->name(), locks->name()); | |
1334 } | |
1335 | |
1336 // remove "this" from the owned locks list | |
1337 | |
1338 Monitor *prev = NULL; | |
1339 bool found = false; | |
1340 for (; locks != NULL; prev = locks, locks = locks->next()) { | |
1341 if (locks == this) { | |
1342 found = true; | |
1343 break; | |
1344 } | |
1345 } | |
1346 assert(found, "Removing a lock not owned"); | |
1347 if (prev == NULL) { | |
1348 old_owner->_owned_locks = _next; | |
1349 } else { | |
1350 prev->_next = _next; | |
1351 } | |
1352 _next = NULL; | |
1353 #endif | |
1354 } | |
1355 } | |
1356 | |
1357 | |
1358 // Factored out common sanity checks for locking mutex'es. Used by lock() and try_lock() | |
1359 void Monitor::check_prelock_state(Thread *thread) { | |
1360 assert((!thread->is_Java_thread() || ((JavaThread *)thread)->thread_state() == _thread_in_vm) | |
1361 || rank() == Mutex::special, "wrong thread state for using locks"); | |
1362 if (StrictSafepointChecks) { | |
1363 if (thread->is_VM_thread() && !allow_vm_block()) { | |
1490
f03d0a26bf83
6888954: argument formatting for assert() and friends
jcoomes
parents:
196
diff
changeset
|
1364 fatal(err_msg("VM thread using lock %s (not allowed to block on)", |
f03d0a26bf83
6888954: argument formatting for assert() and friends
jcoomes
parents:
196
diff
changeset
|
1365 name())); |
0 | 1366 } |
1367 debug_only(if (rank() != Mutex::special) \ | |
1368 thread->check_for_valid_safepoint_state(false);) | |
1369 } | |
1370 } | |
1371 | |
1372 void Monitor::check_block_state(Thread *thread) { | |
1373 if (!_allow_vm_block && thread->is_VM_thread()) { | |
1374 warning("VM thread blocked on lock"); | |
1375 print(); | |
1376 BREAKPOINT; | |
1377 } | |
1378 assert(_owner != thread, "deadlock: blocking on monitor owned by current thread"); | |
1379 } | |
1380 | |
1381 #endif // PRODUCT |