1 2 /* 3 * Copyright (c) 1998, 2010, Oracle and/or its affiliates. All rights reserved. 4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 5 * 6 * This code is free software; you can redistribute it and/or modify it 7 * under the terms of the GNU General Public License version 2 only, as 8 * published by the Free Software Foundation. 9 * 10 * This code is distributed in the hope that it will be useful, but WITHOUT 11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 13 * version 2 for more details (a copy is included in the LICENSE file that 14 * accompanied this code). 15 * 16 * You should have received a copy of the GNU General Public License version 17 * 2 along with this work; if not, write to the Free Software Foundation, 18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 19 * 20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 21 * or visit www.oracle.com if you need additional information or have any 22 * questions. 23 * 24 */ 25 26 #include "precompiled.hpp" 27 #include "runtime/mutex.hpp" 28 #include "runtime/osThread.hpp" 29 #include "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 42 43 // 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 44 // 45 // Native Monitor-Mutex locking - theory of operations 46 // 47 // * Native Monitors are completely unrelated to Java-level monitors, 48 // although the "back-end" slow-path implementations share a common lineage. 49 // See objectMonitor:: in synchronizer.cpp. 50 // Native Monitors do *not* support nesting or recursion but otherwise 51 // they're basically Hoare-flavor monitors. 52 // 53 // * A thread acquires ownership of a Monitor/Mutex by CASing the LockByte 54 // in the _LockWord from zero to non-zero. Note that the _Owner field 55 // is advisory and is used only to verify that the thread calling unlock() 56 // is indeed the last thread to have acquired the lock. 57 // 58 // * Contending threads "push" themselves onto the front of the contention 59 // queue -- called the cxq -- with CAS and then spin/park. 60 // The _LockWord contains the LockByte as well as the pointer to the head 61 // of the cxq. Colocating the LockByte with the cxq precludes certain races. 62 // 63 // * Using a separately addressable LockByte allows for CAS:MEMBAR or CAS:0 64 // idioms. We currently use MEMBAR in the uncontended unlock() path, as 65 // MEMBAR often has less latency than CAS. If warranted, we could switch to 66 // a CAS:0 mode, using timers to close the resultant race, as is done 67 // with Java Monitors in synchronizer.cpp. 68 // 69 // See the following for a discussion of the relative cost of atomics (CAS) 70 // MEMBAR, and ways to eliminate such instructions from the common-case paths: 71 // -- http://blogs.sun.com/dave/entry/biased_locking_in_hotspot 72 // -- http://blogs.sun.com/dave/resource/MustangSync.pdf 73 // -- http://blogs.sun.com/dave/resource/synchronization-public2.pdf 74 // -- synchronizer.cpp 75 // 76 // * Overall goals - desiderata 77 // 1. Minimize context switching 78 // 2. Minimize lock migration 79 // 3. Minimize CPI -- affinity and locality 80 // 4. Minimize the execution of high-latency instructions such as CAS or MEMBAR 81 // 5. Minimize outer lock hold times 82 // 6. Behave gracefully on a loaded system 83 // 84 // * Thread flow and list residency: 85 // 86 // Contention queue --> EntryList --> OnDeck --> Owner --> !Owner 87 // [..resident on monitor list..] 88 // [...........contending..................] 89 // 90 // -- The contention queue (cxq) contains recently-arrived threads (RATs). 91 // Threads on the cxq eventually drain into the EntryList. 92 // -- Invariant: a thread appears on at most one list -- cxq, EntryList 93 // or WaitSet -- at any one time. 94 // -- For a given monitor there can be at most one "OnDeck" thread at any 95 // given time but if needbe this particular invariant could be relaxed. 96 // 97 // * The WaitSet and EntryList linked lists are composed of ParkEvents. 98 // I use ParkEvent instead of threads as ParkEvents are immortal and 99 // type-stable, meaning we can safely unpark() a possibly stale 100 // list element in the unlock()-path. (That's benign). 101 // 102 // * Succession policy - providing for progress: 103 // 104 // As necessary, the unlock()ing thread identifies, unlinks, and unparks 105 // an "heir presumptive" tentative successor thread from the EntryList. 106 // This becomes the so-called "OnDeck" thread, of which there can be only 107 // one at any given time for a given monitor. The wakee will recontend 108 // for ownership of monitor. 109 // 110 // Succession is provided for by a policy of competitive handoff. 111 // The exiting thread does _not_ grant or pass ownership to the 112 // successor thread. (This is also referred to as "handoff" succession"). 113 // Instead the exiting thread releases ownership and possibly wakes 114 // a successor, so the successor can (re)compete for ownership of the lock. 115 // 116 // Competitive handoff provides excellent overall throughput at the expense 117 // of short-term fairness. If fairness is a concern then one remedy might 118 // be to add an AcquireCounter field to the monitor. After a thread acquires 119 // the lock it will decrement the AcquireCounter field. When the count 120 // reaches 0 the thread would reset the AcquireCounter variable, abdicate 121 // the lock directly to some thread on the EntryList, and then move itself to the 122 // tail of the EntryList. 123 // 124 // But in practice most threads engage or otherwise participate in resource 125 // bounded producer-consumer relationships, so lock domination is not usually 126 // a practical concern. Recall too, that in general it's easier to construct 127 // a fair lock from a fast lock, but not vice-versa. 128 // 129 // * The cxq can have multiple concurrent "pushers" but only one concurrent 130 // detaching thread. This mechanism is immune from the ABA corruption. 131 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. 132 // We use OnDeck as a pseudo-lock to enforce the at-most-one detaching 133 // thread constraint. 134 // 135 // * Taken together, the cxq and the EntryList constitute or form a 136 // single logical queue of threads stalled trying to acquire the lock. 137 // We use two distinct lists to reduce heat on the list ends. 138 // Threads in lock() enqueue onto cxq while threads in unlock() will 139 // dequeue from the EntryList. (c.f. Michael Scott's "2Q" algorithm). 140 // A key desideratum is to minimize queue & monitor metadata manipulation 141 // that occurs while holding the "outer" monitor lock -- that is, we want to 142 // minimize monitor lock holds times. 143 // 144 // The EntryList is ordered by the prevailing queue discipline and 145 // can be organized in any convenient fashion, such as a doubly-linked list or 146 // a circular doubly-linked list. If we need a priority queue then something akin 147 // to Solaris' sleepq would work nicely. Viz., 148 // -- http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. 149 // -- http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/uts/common/os/sleepq.c 150 // Queue discipline is enforced at ::unlock() time, when the unlocking thread 151 // drains the cxq into the EntryList, and orders or reorders the threads on the 152 // EntryList accordingly. 153 // 154 // Barring "lock barging", this mechanism provides fair cyclic ordering, 155 // somewhat similar to an elevator-scan. 156 // 157 // * OnDeck 158 // -- For a given monitor there can be at most one OnDeck thread at any given 159 // instant. The OnDeck thread is contending for the lock, but has been 160 // unlinked from the EntryList and cxq by some previous unlock() operations. 161 // Once a thread has been designated the OnDeck thread it will remain so 162 // until it manages to acquire the lock -- being OnDeck is a stable property. 163 // -- Threads on the EntryList or cxq are _not allowed to attempt lock acquisition. 164 // -- OnDeck also serves as an "inner lock" as follows. Threads in unlock() will, after 165 // having cleared the LockByte and dropped the outer lock, attempt to "trylock" 166 // OnDeck by CASing the field from null to non-null. If successful, that thread 167 // is then responsible for progress and succession and can use CAS to detach and 168 // drain the cxq into the EntryList. By convention, only this thread, the holder of 169 // the OnDeck inner lock, can manipulate the EntryList or detach and drain the 170 // RATs on the cxq into the EntryList. This avoids ABA corruption on the cxq as 171 // we allow multiple concurrent "push" operations but restrict detach concurrency 172 // to at most one thread. Having selected and detached a successor, the thread then 173 // changes the OnDeck to refer to that successor, and then unparks the successor. 174 // That successor will eventually acquire the lock and clear OnDeck. Beware 175 // that the OnDeck usage as a lock is asymmetric. A thread in unlock() transiently 176 // "acquires" OnDeck, performs queue manipulations, passes OnDeck to some successor, 177 // and then the successor eventually "drops" OnDeck. Note that there's never 178 // any sense of contention on the inner lock, however. Threads never contend 179 // or wait for the inner lock. 180 // -- OnDeck provides for futile wakeup throttling a described in section 3.3 of 181 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf 182 // In a sense, OnDeck subsumes the ObjectMonitor _Succ and ObjectWaiter 183 // TState fields found in Java-level objectMonitors. (See synchronizer.cpp). 184 // 185 // * Waiting threads reside on the WaitSet list -- wait() puts 186 // the caller onto the WaitSet. Notify() or notifyAll() simply 187 // transfers threads from the WaitSet to either the EntryList or cxq. 188 // Subsequent unlock() operations will eventually unpark the notifyee. 189 // Unparking a notifee in notify() proper is inefficient - if we were to do so 190 // it's likely the notifyee would simply impale itself on the lock held 191 // by the notifier. 192 // 193 // * The mechanism is obstruction-free in that if the holder of the transient 194 // OnDeck lock in unlock() is preempted or otherwise stalls, other threads 195 // can still acquire and release the outer lock and continue to make progress. 196 // At worst, waking of already blocked contending threads may be delayed, 197 // but nothing worse. (We only use "trylock" operations on the inner OnDeck 198 // lock). 199 // 200 // * Note that thread-local storage must be initialized before a thread 201 // uses Native monitors or mutexes. The native monitor-mutex subsystem 202 // depends on Thread::current(). 203 // 204 // * The monitor synchronization subsystem avoids the use of native 205 // synchronization primitives except for the narrow platform-specific 206 // park-unpark abstraction. See the comments in os_solaris.cpp regarding 207 // the semantics of park-unpark. Put another way, this monitor implementation 208 // depends only on atomic operations and park-unpark. The monitor subsystem 209 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the 210 // underlying OS manages the READY<->RUN transitions. 211 // 212 // * The memory consistency model provide by lock()-unlock() is at least as 213 // strong or stronger than the Java Memory model defined by JSR-133. 214 // That is, we guarantee at least entry consistency, if not stronger. 215 // See http://g.oswego.edu/dl/jmm/cookbook.html. 216 // 217 // * Thread:: currently contains a set of purpose-specific ParkEvents: 218 // _MutexEvent, _ParkEvent, etc. A better approach might be to do away with 219 // the purpose-specific ParkEvents and instead implement a general per-thread 220 // stack of available ParkEvents which we could provision on-demand. The 221 // stack acts as a local cache to avoid excessive calls to ParkEvent::Allocate() 222 // and ::Release(). A thread would simply pop an element from the local stack before it 223 // enqueued or park()ed. When the contention was over the thread would 224 // push the no-longer-needed ParkEvent back onto its stack. 225 // 226 // * A slightly reduced form of ILock() and IUnlock() have been partially 227 // model-checked (Murphi) for safety and progress at T=1,2,3 and 4. 228 // It'd be interesting to see if TLA/TLC could be useful as well. 229 // 230 // * Mutex-Monitor is a low-level "leaf" subsystem. That is, the monitor 231 // code should never call other code in the JVM that might itself need to 232 // acquire monitors or mutexes. That's true *except* in the case of the 233 // ThreadBlockInVM state transition wrappers. The ThreadBlockInVM DTOR handles 234 // mutator reentry (ingress) by checking for a pending safepoint in which case it will 235 // call SafepointSynchronize::block(), which in turn may call Safepoint_lock->lock(), etc. 236 // In that particular case a call to lock() for a given Monitor can end up recursively 237 // calling lock() on another monitor. While distasteful, this is largely benign 238 // as the calls come from jacket that wraps lock(), and not from deep within lock() itself. 239 // 240 // It's unfortunate that native mutexes and thread state transitions were convolved. 241 // They're really separate concerns and should have remained that way. Melding 242 // them together was facile -- a bit too facile. The current implementation badly 243 // conflates the two concerns. 244 // 245 // * TODO-FIXME: 246 // 247 // -- Add DTRACE probes for contended acquire, contended acquired, contended unlock 248 // We should also add DTRACE probes in the ParkEvent subsystem for 249 // Park-entry, Park-exit, and Unpark. 250 // 251 // -- We have an excess of mutex-like constructs in the JVM, namely: 252 // 1. objectMonitors for Java-level synchronization (synchronizer.cpp) 253 // 2. low-level muxAcquire and muxRelease 254 // 3. low-level spinAcquire and spinRelease 255 // 4. native Mutex:: and Monitor:: 256 // 5. jvm_raw_lock() and _unlock() 257 // 6. JVMTI raw monitors -- distinct from (5) despite having a confusingly 258 // similar name. 259 // 260 // 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 261 262 263 // CASPTR() uses the canonical argument order that dominates in the literature. 264 // Our internal cmpxchg_ptr() uses a bastardized ordering to accommodate Sun .il templates. 265 266 #define CASPTR(a,c,s) intptr_t(Atomic::cmpxchg_ptr ((void *)(s),(void *)(a),(void *)(c))) 267 #define UNS(x) (uintptr_t(x)) 268 #define TRACE(m) { static volatile int ctr = 0 ; int x = ++ctr ; if ((x & (x-1))==0) { ::printf ("%d:%s\n", x, #m); ::fflush(stdout); }} 269 270 // Simplistic low-quality Marsaglia SHIFT-XOR RNG. 271 // Bijective except for the trailing mask operation. 272 // Useful for spin loops as the compiler can't optimize it away. 273 274 static inline jint MarsagliaXORV (jint x) { 275 if (x == 0) x = 1|os::random() ; 276 x ^= x << 6; 277 x ^= ((unsigned)x) >> 21; 278 x ^= x << 7 ; 279 return x & 0x7FFFFFFF ; 280 } 281 282 static inline jint MarsagliaXOR (jint * const a) { 283 jint x = *a ; 284 if (x == 0) x = UNS(a)|1 ; 285 x ^= x << 6; 286 x ^= ((unsigned)x) >> 21; 287 x ^= x << 7 ; 288 *a = x ; 289 return x & 0x7FFFFFFF ; 290 } 291 292 static int Stall (int its) { 293 static volatile jint rv = 1 ; 294 volatile int OnFrame = 0 ; 295 jint v = rv ^ UNS(OnFrame) ; 296 while (--its >= 0) { 297 v = MarsagliaXORV (v) ; 298 } 299 // Make this impossible for the compiler to optimize away, 300 // but (mostly) avoid W coherency sharing on MP systems. 301 if (v == 0x12345) rv = v ; 302 return v ; 303 } 304 305 int Monitor::TryLock () { 306 intptr_t v = _LockWord.FullWord ; 307 for (;;) { 308 if ((v & _LBIT) != 0) return 0 ; 309 const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ; 310 if (v == u) return 1 ; 311 v = u ; 312 } 313 } 314 315 int Monitor::TryFast () { 316 // Optimistic fast-path form ... 317 // Fast-path attempt for the common uncontended case. 318 // Avoid RTS->RTO $ coherence upgrade on typical SMP systems. 319 intptr_t v = CASPTR (&_LockWord, 0, _LBIT) ; // agro ... 320 if (v == 0) return 1 ; 321 322 for (;;) { 323 if ((v & _LBIT) != 0) return 0 ; 324 const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ; 325 if (v == u) return 1 ; 326 v = u ; 327 } 328 } 329 330 int Monitor::ILocked () { 331 const intptr_t w = _LockWord.FullWord & 0xFF ; 332 assert (w == 0 || w == _LBIT, "invariant") ; 333 return w == _LBIT ; 334 } 335 336 // Polite TATAS spinlock with exponential backoff - bounded spin. 337 // Ideally we'd use processor cycles, time or vtime to control 338 // the loop, but we currently use iterations. 339 // All the constants within were derived empirically but work over 340 // over the spectrum of J2SE reference platforms. 341 // On Niagara-class systems the back-off is unnecessary but 342 // is relatively harmless. (At worst it'll slightly retard 343 // acquisition times). The back-off is critical for older SMP systems 344 // where constant fetching of the LockWord would otherwise impair 345 // scalability. 346 // 347 // Clamp spinning at approximately 1/2 of a context-switch round-trip. 348 // See synchronizer.cpp for details and rationale. 349 350 int Monitor::TrySpin (Thread * const Self) { 351 if (TryLock()) return 1 ; 352 if (!os::is_MP()) return 0 ; 353 354 int Probes = 0 ; 355 int Delay = 0 ; 356 int Steps = 0 ; 357 int SpinMax = NativeMonitorSpinLimit ; 358 int flgs = NativeMonitorFlags ; 359 for (;;) { 360 intptr_t v = _LockWord.FullWord; 361 if ((v & _LBIT) == 0) { 362 if (CASPTR (&_LockWord, v, v|_LBIT) == v) { 363 return 1 ; 364 } 365 continue ; 366 } 367 368 if ((flgs & 8) == 0) { 369 SpinPause () ; 370 } 371 372 // Periodically increase Delay -- variable Delay form 373 // conceptually: delay *= 1 + 1/Exponent 374 ++ Probes; 375 if (Probes > SpinMax) return 0 ; 376 377 if ((Probes & 0x7) == 0) { 378 Delay = ((Delay << 1)|1) & 0x7FF ; 379 // CONSIDER: Delay += 1 + (Delay/4); Delay &= 0x7FF ; 380 } 381 382 if (flgs & 2) continue ; 383 384 // Consider checking _owner's schedctl state, if OFFPROC abort spin. 385 // If the owner is OFFPROC then it's unlike that the lock will be dropped 386 // in a timely fashion, which suggests that spinning would not be fruitful 387 // or profitable. 388 389 // Stall for "Delay" time units - iterations in the current implementation. 390 // Avoid generating coherency traffic while stalled. 391 // Possible ways to delay: 392 // PAUSE, SLEEP, MEMBAR #sync, MEMBAR #halt, 393 // wr %g0,%asi, gethrtime, rdstick, rdtick, rdtsc, etc. ... 394 // Note that on Niagara-class systems we want to minimize STs in the 395 // spin loop. N1 and brethren write-around the L1$ over the xbar into the L2$. 396 // Furthermore, they don't have a W$ like traditional SPARC processors. 397 // We currently use a Marsaglia Shift-Xor RNG loop. 398 Steps += Delay ; 399 if (Self != NULL) { 400 jint rv = Self->rng[0] ; 401 for (int k = Delay ; --k >= 0; ) { 402 rv = MarsagliaXORV (rv) ; 403 if ((flgs & 4) == 0 && SafepointSynchronize::do_call_back()) return 0 ; 404 } 405 Self->rng[0] = rv ; 406 } else { 407 Stall (Delay) ; 408 } 409 } 410 } 411 412 static int ParkCommon (ParkEvent * ev, jlong timo) { 413 // Diagnostic support - periodically unwedge blocked threads 414 intx nmt = NativeMonitorTimeout ; 415 if (nmt > 0 && (nmt < timo || timo <= 0)) { 416 timo = nmt ; 417 } 418 int err = OS_OK ; 419 if (0 == timo) { 420 ev->park() ; 421 } else { 422 err = ev->park(timo) ; 423 } 424 return err ; 425 } 426 427 inline int Monitor::AcquireOrPush (ParkEvent * ESelf) { 428 intptr_t v = _LockWord.FullWord ; 429 for (;;) { 430 if ((v & _LBIT) == 0) { 431 const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ; 432 if (u == v) return 1 ; // indicate acquired 433 v = u ; 434 } else { 435 // Anticipate success ... 436 ESelf->ListNext = (ParkEvent *) (v & ~_LBIT) ; 437 const intptr_t u = CASPTR (&_LockWord, v, intptr_t(ESelf)|_LBIT) ; 438 if (u == v) return 0 ; // indicate pushed onto cxq 439 v = u ; 440 } 441 // Interference - LockWord change - just retry 442 } 443 } 444 445 // ILock and IWait are the lowest level primitive internal blocking 446 // synchronization functions. The callers of IWait and ILock must have 447 // performed any needed state transitions beforehand. 448 // IWait and ILock may directly call park() without any concern for thread state. 449 // Note that ILock and IWait do *not* access _owner. 450 // _owner is a higher-level logical concept. 451 452 void Monitor::ILock (Thread * Self) { 453 assert (_OnDeck != Self->_MutexEvent, "invariant") ; 454 455 if (TryFast()) { 456 Exeunt: 457 assert (ILocked(), "invariant") ; 458 return ; 459 } 460 461 ParkEvent * const ESelf = Self->_MutexEvent ; 462 assert (_OnDeck != ESelf, "invariant") ; 463 464 // As an optimization, spinners could conditionally try to set ONDECK to _LBIT 465 // Synchronizer.cpp uses a similar optimization. 466 if (TrySpin (Self)) goto Exeunt ; 467 468 // Slow-path - the lock is contended. 469 // Either Enqueue Self on cxq or acquire the outer lock. 470 // LockWord encoding = (cxq,LOCKBYTE) 471 ESelf->reset() ; 472 OrderAccess::fence() ; 473 474 // Optional optimization ... try barging on the inner lock 475 if ((NativeMonitorFlags & 32) && CASPTR (&_OnDeck, NULL, UNS(Self)) == 0) { 476 goto OnDeck_LOOP ; 477 } 478 479 if (AcquireOrPush (ESelf)) goto Exeunt ; 480 481 // At any given time there is at most one ondeck thread. 482 // ondeck implies not resident on cxq and not resident on EntryList 483 // Only the OnDeck thread can try to acquire -- contended for -- the lock. 484 // CONSIDER: use Self->OnDeck instead of m->OnDeck. 485 // Deschedule Self so that others may run. 486 while (_OnDeck != ESelf) { 487 ParkCommon (ESelf, 0) ; 488 } 489 490 // Self is now in the ONDECK position and will remain so until it 491 // manages to acquire the lock. 492 OnDeck_LOOP: 493 for (;;) { 494 assert (_OnDeck == ESelf, "invariant") ; 495 if (TrySpin (Self)) break ; 496 // CONSIDER: if ESelf->TryPark() && TryLock() break ... 497 // It's probably wise to spin only if we *actually* blocked 498 // CONSIDER: check the lockbyte, if it remains set then 499 // preemptively drain the cxq into the EntryList. 500 // The best place and time to perform queue operations -- lock metadata -- 501 // is _before having acquired the outer lock, while waiting for the lock to drop. 502 ParkCommon (ESelf, 0) ; 503 } 504 505 assert (_OnDeck == ESelf, "invariant") ; 506 _OnDeck = NULL ; 507 508 // Note that we current drop the inner lock (clear OnDeck) in the slow-path 509 // epilog immediately after having acquired the outer lock. 510 // But instead we could consider the following optimizations: 511 // A. Shift or defer dropping the inner lock until the subsequent IUnlock() operation. 512 // This might avoid potential reacquisition of the inner lock in IUlock(). 513 // B. While still holding the inner lock, attempt to opportunistically select 514 // and unlink the next ONDECK thread from the EntryList. 515 // If successful, set ONDECK to refer to that thread, otherwise clear ONDECK. 516 // It's critical that the select-and-unlink operation run in constant-time as 517 // it executes when holding the outer lock and may artificially increase the 518 // effective length of the critical section. 519 // Note that (A) and (B) are tantamount to succession by direct handoff for 520 // the inner lock. 521 goto Exeunt ; 522 } 523 524 void Monitor::IUnlock (bool RelaxAssert) { 525 assert (ILocked(), "invariant") ; 526 _LockWord.Bytes[_LSBINDEX] = 0 ; // drop outer lock 527 OrderAccess::storeload (); 528 ParkEvent * const w = _OnDeck ; 529 assert (RelaxAssert || w != Thread::current()->_MutexEvent, "invariant") ; 530 if (w != NULL) { 531 // Either we have a valid ondeck thread or ondeck is transiently "locked" 532 // by some exiting thread as it arranges for succession. The LSBit of 533 // OnDeck allows us to discriminate two cases. If the latter, the 534 // responsibility for progress and succession lies with that other thread. 535 // For good performance, we also depend on the fact that redundant unpark() 536 // operations are cheap. That is, repeated Unpark()ing of the ONDECK thread 537 // is inexpensive. This approach provides implicit futile wakeup throttling. 538 // Note that the referent "w" might be stale with respect to the lock. 539 // In that case the following unpark() is harmless and the worst that'll happen 540 // is a spurious return from a park() operation. Critically, if "w" _is stale, 541 // then progress is known to have occurred as that means the thread associated 542 // with "w" acquired the lock. In that case this thread need take no further 543 // action to guarantee progress. 544 if ((UNS(w) & _LBIT) == 0) w->unpark() ; 545 return ; 546 } 547 548 intptr_t cxq = _LockWord.FullWord ; 549 if (((cxq & ~_LBIT)|UNS(_EntryList)) == 0) { 550 return ; // normal fast-path exit - cxq and EntryList both empty 551 } 552 if (cxq & _LBIT) { 553 // Optional optimization ... 554 // Some other thread acquired the lock in the window since this 555 // thread released it. Succession is now that thread's responsibility. 556 return ; 557 } 558 559 Succession: 560 // Slow-path exit - this thread must ensure succession and progress. 561 // OnDeck serves as lock to protect cxq and EntryList. 562 // Only the holder of OnDeck can manipulate EntryList or detach the RATs from cxq. 563 // Avoid ABA - allow multiple concurrent producers (enqueue via push-CAS) 564 // but only one concurrent consumer (detacher of RATs). 565 // Consider protecting this critical section with schedctl on Solaris. 566 // Unlike a normal lock, however, the exiting thread "locks" OnDeck, 567 // picks a successor and marks that thread as OnDeck. That successor 568 // thread will then clear OnDeck once it eventually acquires the outer lock. 569 if (CASPTR (&_OnDeck, NULL, _LBIT) != UNS(NULL)) { 570 return ; 571 } 572 573 ParkEvent * List = _EntryList ; 574 if (List != NULL) { 575 // Transfer the head of the EntryList to the OnDeck position. 576 // Once OnDeck, a thread stays OnDeck until it acquires the lock. 577 // For a given lock there is at most OnDeck thread at any one instant. 578 WakeOne: 579 assert (List == _EntryList, "invariant") ; 580 ParkEvent * const w = List ; 581 assert (RelaxAssert || w != Thread::current()->_MutexEvent, "invariant") ; 582 _EntryList = w->ListNext ; 583 // as a diagnostic measure consider setting w->_ListNext = BAD 584 assert (UNS(_OnDeck) == _LBIT, "invariant") ; 585 _OnDeck = w ; // pass OnDeck to w. 586 // w will clear OnDeck once it acquires the outer lock 587 588 // Another optional optimization ... 589 // For heavily contended locks it's not uncommon that some other 590 // thread acquired the lock while this thread was arranging succession. 591 // Try to defer the unpark() operation - Delegate the responsibility 592 // for unpark()ing the OnDeck thread to the current or subsequent owners 593 // That is, the new owner is responsible for unparking the OnDeck thread. 594 OrderAccess::storeload() ; 595 cxq = _LockWord.FullWord ; 596 if (cxq & _LBIT) return ; 597 598 w->unpark() ; 599 return ; 600 } 601 602 cxq = _LockWord.FullWord ; 603 if ((cxq & ~_LBIT) != 0) { 604 // The EntryList is empty but the cxq is populated. 605 // drain RATs from cxq into EntryList 606 // Detach RATs segment with CAS and then merge into EntryList 607 for (;;) { 608 // optional optimization - if locked, the owner is responsible for succession 609 if (cxq & _LBIT) goto Punt ; 610 const intptr_t vfy = CASPTR (&_LockWord, cxq, cxq & _LBIT) ; 611 if (vfy == cxq) break ; 612 cxq = vfy ; 613 // Interference - LockWord changed - Just retry 614 // We can see concurrent interference from contending threads 615 // pushing themselves onto the cxq or from lock-unlock operations. 616 // From the perspective of this thread, EntryList is stable and 617 // the cxq is prepend-only -- the head is volatile but the interior 618 // of the cxq is stable. In theory if we encounter interference from threads 619 // pushing onto cxq we could simply break off the original cxq suffix and 620 // move that segment to the EntryList, avoiding a 2nd or multiple CAS attempts 621 // on the high-traffic LockWord variable. For instance lets say the cxq is "ABCD" 622 // when we first fetch cxq above. Between the fetch -- where we observed "A" 623 // -- and CAS -- where we attempt to CAS null over A -- "PQR" arrive, 624 // yielding cxq = "PQRABCD". In this case we could simply set A.ListNext 625 // null, leaving cxq = "PQRA" and transfer the "BCD" segment to the EntryList. 626 // Note too, that it's safe for this thread to traverse the cxq 627 // without taking any special concurrency precautions. 628 } 629 630 // We don't currently reorder the cxq segment as we move it onto 631 // the EntryList, but it might make sense to reverse the order 632 // or perhaps sort by thread priority. See the comments in 633 // synchronizer.cpp objectMonitor::exit(). 634 assert (_EntryList == NULL, "invariant") ; 635 _EntryList = List = (ParkEvent *)(cxq & ~_LBIT) ; 636 assert (List != NULL, "invariant") ; 637 goto WakeOne ; 638 } 639 640 // cxq|EntryList is empty. 641 // w == NULL implies that cxq|EntryList == NULL in the past. 642 // Possible race - rare inopportune interleaving. 643 // A thread could have added itself to cxq since this thread previously checked. 644 // Detect and recover by refetching cxq. 645 Punt: 646 assert (UNS(_OnDeck) == _LBIT, "invariant") ; 647 _OnDeck = NULL ; // Release inner lock. 648 OrderAccess::storeload(); // Dekker duality - pivot point 649 650 // Resample LockWord/cxq to recover from possible race. 651 // For instance, while this thread T1 held OnDeck, some other thread T2 might 652 // acquire the outer lock. Another thread T3 might try to acquire the outer 653 // lock, but encounter contention and enqueue itself on cxq. T2 then drops the 654 // outer lock, but skips succession as this thread T1 still holds OnDeck. 655 // T1 is and remains responsible for ensuring succession of T3. 656 // 657 // Note that we don't need to recheck EntryList, just cxq. 658 // If threads moved onto EntryList since we dropped OnDeck 659 // that implies some other thread forced succession. 660 cxq = _LockWord.FullWord ; 661 if ((cxq & ~_LBIT) != 0 && (cxq & _LBIT) == 0) { 662 goto Succession ; // potential race -- re-run succession 663 } 664 return ; 665 } 666 667 bool Monitor::notify() { 668 assert (_owner == Thread::current(), "invariant") ; 669 assert (ILocked(), "invariant") ; 670 if (_WaitSet == NULL) return true ; 671 NotifyCount ++ ; 672 673 // Transfer one thread from the WaitSet to the EntryList or cxq. 674 // Currently we just unlink the head of the WaitSet and prepend to the cxq. 675 // And of course we could just unlink it and unpark it, too, but 676 // in that case it'd likely impale itself on the reentry. 677 Thread::muxAcquire (_WaitLock, "notify:WaitLock") ; 678 ParkEvent * nfy = _WaitSet ; 679 if (nfy != NULL) { // DCL idiom 680 _WaitSet = nfy->ListNext ; 681 assert (nfy->Notified == 0, "invariant") ; 682 // push nfy onto the cxq 683 for (;;) { 684 const intptr_t v = _LockWord.FullWord ; 685 assert ((v & 0xFF) == _LBIT, "invariant") ; 686 nfy->ListNext = (ParkEvent *)(v & ~_LBIT); 687 if (CASPTR (&_LockWord, v, UNS(nfy)|_LBIT) == v) break; 688 // interference - _LockWord changed -- just retry 689 } 690 // Note that setting Notified before pushing nfy onto the cxq is 691 // also legal and safe, but the safety properties are much more 692 // subtle, so for the sake of code stewardship ... 693 OrderAccess::fence() ; 694 nfy->Notified = 1; 695 } 696 Thread::muxRelease (_WaitLock) ; 697 if (nfy != NULL && (NativeMonitorFlags & 16)) { 698 // Experimental code ... light up the wakee in the hope that this thread (the owner) 699 // will drop the lock just about the time the wakee comes ONPROC. 700 nfy->unpark() ; 701 } 702 assert (ILocked(), "invariant") ; 703 return true ; 704 } 705 706 // Currently notifyAll() transfers the waiters one-at-a-time from the waitset 707 // to the cxq. This could be done more efficiently with a single bulk en-mass transfer, 708 // but in practice notifyAll() for large #s of threads is rare and not time-critical. 709 // Beware too, that we invert the order of the waiters. Lets say that the 710 // waitset is "ABCD" and the cxq is "XYZ". After a notifyAll() the waitset 711 // will be empty and the cxq will be "DCBAXYZ". This is benign, of course. 712 713 bool Monitor::notify_all() { 714 assert (_owner == Thread::current(), "invariant") ; 715 assert (ILocked(), "invariant") ; 716 while (_WaitSet != NULL) notify() ; 717 return true ; 718 } 719 720 int Monitor::IWait (Thread * Self, jlong timo) { 721 assert (ILocked(), "invariant") ; 722 723 // Phases: 724 // 1. Enqueue Self on WaitSet - currently prepend 725 // 2. unlock - drop the outer lock 726 // 3. wait for either notification or timeout 727 // 4. lock - reentry - reacquire the outer lock 728 729 ParkEvent * const ESelf = Self->_MutexEvent ; 730 ESelf->Notified = 0 ; 731 ESelf->reset() ; 732 OrderAccess::fence() ; 733 734 // Add Self to WaitSet 735 // Ideally only the holder of the outer lock would manipulate the WaitSet - 736 // That is, the outer lock would implicitly protect the WaitSet. 737 // But if a thread in wait() encounters a timeout it will need to dequeue itself 738 // from the WaitSet _before it becomes the owner of the lock. We need to dequeue 739 // as the ParkEvent -- which serves as a proxy for the thread -- can't reside 740 // on both the WaitSet and the EntryList|cxq at the same time.. That is, a thread 741 // on the WaitSet can't be allowed to compete for the lock until it has managed to 742 // unlink its ParkEvent from WaitSet. Thus the need for WaitLock. 743 // Contention on the WaitLock is minimal. 744 // 745 // Another viable approach would be add another ParkEvent, "WaitEvent" to the 746 // thread class. The WaitSet would be composed of WaitEvents. Only the 747 // owner of the outer lock would manipulate the WaitSet. A thread in wait() 748 // could then compete for the outer lock, and then, if necessary, unlink itself 749 // from the WaitSet only after having acquired the outer lock. More precisely, 750 // there would be no WaitLock. A thread in in wait() would enqueue its WaitEvent 751 // on the WaitSet; release the outer lock; wait for either notification or timeout; 752 // reacquire the inner lock; and then, if needed, unlink itself from the WaitSet. 753 // 754 // Alternatively, a 2nd set of list link fields in the ParkEvent might suffice. 755 // One set would be for the WaitSet and one for the EntryList. 756 // We could also deconstruct the ParkEvent into a "pure" event and add a 757 // new immortal/TSM "ListElement" class that referred to ParkEvents. 758 // In that case we could have one ListElement on the WaitSet and another 759 // on the EntryList, with both referring to the same pure Event. 760 761 Thread::muxAcquire (_WaitLock, "wait:WaitLock:Add") ; 762 ESelf->ListNext = _WaitSet ; 763 _WaitSet = ESelf ; 764 Thread::muxRelease (_WaitLock) ; 765 766 // Release the outer lock 767 // We call IUnlock (RelaxAssert=true) as a thread T1 might 768 // enqueue itself on the WaitSet, call IUnlock(), drop the lock, 769 // and then stall before it can attempt to wake a successor. 770 // Some other thread T2 acquires the lock, and calls notify(), moving 771 // T1 from the WaitSet to the cxq. T2 then drops the lock. T1 resumes, 772 // and then finds *itself* on the cxq. During the course of a normal 773 // IUnlock() call a thread should _never find itself on the EntryList 774 // or cxq, but in the case of wait() it's possible. 775 // See synchronizer.cpp objectMonitor::wait(). 776 IUnlock (true) ; 777 778 // Wait for either notification or timeout 779 // Beware that in some circumstances we might propagate 780 // spurious wakeups back to the caller. 781 782 for (;;) { 783 if (ESelf->Notified) break ; 784 int err = ParkCommon (ESelf, timo) ; 785 if (err == OS_TIMEOUT || (NativeMonitorFlags & 1)) break ; 786 } 787 788 // Prepare for reentry - if necessary, remove ESelf from WaitSet 789 // ESelf can be: 790 // 1. Still on the WaitSet. This can happen if we exited the loop by timeout. 791 // 2. On the cxq or EntryList 792 // 3. Not resident on cxq, EntryList or WaitSet, but in the OnDeck position. 793 794 OrderAccess::fence() ; 795 int WasOnWaitSet = 0 ; 796 if (ESelf->Notified == 0) { 797 Thread::muxAcquire (_WaitLock, "wait:WaitLock:remove") ; 798 if (ESelf->Notified == 0) { // DCL idiom 799 assert (_OnDeck != ESelf, "invariant") ; // can't be both OnDeck and on WaitSet 800 // ESelf is resident on the WaitSet -- unlink it. 801 // A doubly-linked list would be better here so we can unlink in constant-time. 802 // We have to unlink before we potentially recontend as ESelf might otherwise 803 // end up on the cxq|EntryList -- it can't be on two lists at once. 804 ParkEvent * p = _WaitSet ; 805 ParkEvent * q = NULL ; // classic q chases p 806 while (p != NULL && p != ESelf) { 807 q = p ; 808 p = p->ListNext ; 809 } 810 assert (p == ESelf, "invariant") ; 811 if (p == _WaitSet) { // found at head 812 assert (q == NULL, "invariant") ; 813 _WaitSet = p->ListNext ; 814 } else { // found in interior 815 assert (q->ListNext == p, "invariant") ; 816 q->ListNext = p->ListNext ; 817 } 818 WasOnWaitSet = 1 ; // We were *not* notified but instead encountered timeout 819 } 820 Thread::muxRelease (_WaitLock) ; 821 } 822 823 // Reentry phase - reacquire the lock 824 if (WasOnWaitSet) { 825 // ESelf was previously on the WaitSet but we just unlinked it above 826 // because of a timeout. ESelf is not resident on any list and is not OnDeck 827 assert (_OnDeck != ESelf, "invariant") ; 828 ILock (Self) ; 829 } else { 830 // A prior notify() operation moved ESelf from the WaitSet to the cxq. 831 // ESelf is now on the cxq, EntryList or at the OnDeck position. 832 // The following fragment is extracted from Monitor::ILock() 833 for (;;) { 834 if (_OnDeck == ESelf && TrySpin(Self)) break ; 835 ParkCommon (ESelf, 0) ; 836 } 837 assert (_OnDeck == ESelf, "invariant") ; 838 _OnDeck = NULL ; 839 } 840 841 assert (ILocked(), "invariant") ; 842 return WasOnWaitSet != 0 ; // return true IFF timeout 843 } 844 845 846 // ON THE VMTHREAD SNEAKING PAST HELD LOCKS: 847 // In particular, there are certain types of global lock that may be held 848 // by a Java thread while it is blocked at a safepoint but before it has 849 // written the _owner field. These locks may be sneakily acquired by the 850 // VM thread during a safepoint to avoid deadlocks. Alternatively, one should 851 // identify all such locks, and ensure that Java threads never block at 852 // safepoints while holding them (_no_safepoint_check_flag). While it 853 // seems as though this could increase the time to reach a safepoint 854 // (or at least increase the mean, if not the variance), the latter 855 // approach might make for a cleaner, more maintainable JVM design. 856 // 857 // Sneaking is vile and reprehensible and should be excised at the 1st 858 // opportunity. It's possible that the need for sneaking could be obviated 859 // as follows. Currently, a thread might (a) while TBIVM, call pthread_mutex_lock 860 // or ILock() thus acquiring the "physical" lock underlying Monitor/Mutex. 861 // (b) stall at the TBIVM exit point as a safepoint is in effect. Critically, 862 // it'll stall at the TBIVM reentry state transition after having acquired the 863 // underlying lock, but before having set _owner and having entered the actual 864 // critical section. The lock-sneaking facility leverages that fact and allowed the 865 // VM thread to logically acquire locks that had already be physically locked by mutators 866 // but where mutators were known blocked by the reentry thread state transition. 867 // 868 // If we were to modify the Monitor-Mutex so that TBIVM state transitions tightly 869 // wrapped calls to park(), then we could likely do away with sneaking. We'd 870 // decouple lock acquisition and parking. The critical invariant to eliminating 871 // sneaking is to ensure that we never "physically" acquire the lock while TBIVM. 872 // An easy way to accomplish this is to wrap the park calls in a narrow TBIVM jacket. 873 // One difficulty with this approach is that the TBIVM wrapper could recurse and 874 // call lock() deep from within a lock() call, while the MutexEvent was already enqueued. 875 // Using a stack (N=2 at minimum) of ParkEvents would take care of that problem. 876 // 877 // But of course the proper ultimate approach is to avoid schemes that require explicit 878 // sneaking or dependence on any any clever invariants or subtle implementation properties 879 // of Mutex-Monitor and instead directly address the underlying design flaw. 880 881 void Monitor::lock (Thread * Self) { 882 #ifdef CHECK_UNHANDLED_OOPS 883 // Clear unhandled oops so we get a crash right away. Only clear for non-vm 884 // or GC threads. 885 if (Self->is_Java_thread()) { 886 Self->clear_unhandled_oops(); 887 } 888 #endif // CHECK_UNHANDLED_OOPS 889 890 debug_only(check_prelock_state(Self)); 891 assert (_owner != Self , "invariant") ; 892 assert (_OnDeck != Self->_MutexEvent, "invariant") ; 893 894 if (TryFast()) { 895 Exeunt: 896 assert (ILocked(), "invariant") ; 897 assert (owner() == NULL, "invariant"); 898 set_owner (Self); 899 return ; 900 } 901 902 // The lock is contended ... 903 904 bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint(); 905 if (can_sneak && _owner == NULL) { 906 // a java thread has locked the lock but has not entered the 907 // critical region -- let's just pretend we've locked the lock 908 // and go on. we note this with _snuck so we can also 909 // pretend to unlock when the time comes. 910 _snuck = true; 911 goto Exeunt ; 912 } 913 914 // Try a brief spin to avoid passing thru thread state transition ... 915 if (TrySpin (Self)) goto Exeunt ; 916 917 check_block_state(Self); 918 if (Self->is_Java_thread()) { 919 // Horribile dictu - we suffer through a state transition 920 assert(rank() > Mutex::special, "Potential deadlock with special or lesser rank mutex"); 921 ThreadBlockInVM tbivm ((JavaThread *) Self) ; 922 ILock (Self) ; 923 } else { 924 // Mirabile dictu 925 ILock (Self) ; 926 } 927 goto Exeunt ; 928 } 929 930 void Monitor::lock() { 931 this->lock(Thread::current()); 932 } 933 934 // Lock without safepoint check - a degenerate variant of lock(). 935 // Should ONLY be used by safepoint code and other code 936 // that is guaranteed not to block while running inside the VM. If this is called with 937 // thread state set to be in VM, the safepoint synchronization code will deadlock! 938 939 void Monitor::lock_without_safepoint_check (Thread * Self) { 940 assert (_owner != Self, "invariant") ; 941 ILock (Self) ; 942 assert (_owner == NULL, "invariant"); 943 set_owner (Self); 944 } 945 946 void Monitor::lock_without_safepoint_check () { 947 lock_without_safepoint_check (Thread::current()) ; 948 } 949 950 951 // Returns true if thread succeceed [sic] in grabbing the lock, otherwise false. 952 953 bool Monitor::try_lock() { 954 Thread * const Self = Thread::current(); 955 debug_only(check_prelock_state(Self)); 956 // assert(!thread->is_inside_signal_handler(), "don't lock inside signal handler"); 957 958 // Special case, where all Java threads are stopped. 959 // The lock may have been acquired but _owner is not yet set. 960 // In that case the VM thread can safely grab the lock. 961 // It strikes me this should appear _after the TryLock() fails, below. 962 bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint(); 963 if (can_sneak && _owner == NULL) { 964 set_owner(Self); // Do not need to be atomic, since we are at a safepoint 965 _snuck = true; 966 return true; 967 } 968 969 if (TryLock()) { 970 // We got the lock 971 assert (_owner == NULL, "invariant"); 972 set_owner (Self); 973 return true; 974 } 975 return false; 976 } 977 978 void Monitor::unlock() { 979 assert (_owner == Thread::current(), "invariant") ; 980 assert (_OnDeck != Thread::current()->_MutexEvent , "invariant") ; 981 set_owner (NULL) ; 982 if (_snuck) { 983 assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak"); 984 _snuck = false; 985 return ; 986 } 987 IUnlock (false) ; 988 } 989 990 // Yet another degenerate version of Monitor::lock() or lock_without_safepoint_check() 991 // jvm_raw_lock() and _unlock() can be called by non-Java threads via JVM_RawMonitorEnter. 992 // 993 // There's no expectation that JVM_RawMonitors will interoperate properly with the native 994 // Mutex-Monitor constructs. We happen to implement JVM_RawMonitors in terms of 995 // native Mutex-Monitors simply as a matter of convenience. A simple abstraction layer 996 // over a pthread_mutex_t would work equally as well, but require more platform-specific 997 // code -- a "PlatformMutex". Alternatively, a simply layer over muxAcquire-muxRelease 998 // would work too. 999 // 1000 // Since the caller might be a foreign thread, we don't necessarily have a Thread.MutexEvent 1001 // instance available. Instead, we transiently allocate a ParkEvent on-demand if 1002 // we encounter contention. That ParkEvent remains associated with the thread 1003 // until it manages to acquire the lock, at which time we return the ParkEvent 1004 // to the global ParkEvent free list. This is correct and suffices for our purposes. 1005 // 1006 // Beware that the original jvm_raw_unlock() had a "_snuck" test but that 1007 // jvm_raw_lock() didn't have the corresponding test. I suspect that's an 1008 // oversight, but I've replicated the original suspect logic in the new code ... 1009 1010 void Monitor::jvm_raw_lock() { 1011 assert(rank() == native, "invariant"); 1012 1013 if (TryLock()) { 1014 Exeunt: 1015 assert (ILocked(), "invariant") ; 1016 assert (_owner == NULL, "invariant"); 1017 // This can potentially be called by non-java Threads. Thus, the ThreadLocalStorage 1018 // might return NULL. Don't call set_owner since it will break on an NULL owner 1019 // Consider installing a non-null "ANON" distinguished value instead of just NULL. 1020 _owner = ThreadLocalStorage::thread(); 1021 return ; 1022 } 1023 1024 if (TrySpin(NULL)) goto Exeunt ; 1025 1026 // slow-path - apparent contention 1027 // Allocate a ParkEvent for transient use. 1028 // The ParkEvent remains associated with this thread until 1029 // the time the thread manages to acquire the lock. 1030 ParkEvent * const ESelf = ParkEvent::Allocate(NULL) ; 1031 ESelf->reset() ; 1032 OrderAccess::storeload() ; 1033 1034 // Either Enqueue Self on cxq or acquire the outer lock. 1035 if (AcquireOrPush (ESelf)) { 1036 ParkEvent::Release (ESelf) ; // surrender the ParkEvent 1037 goto Exeunt ; 1038 } 1039 1040 // At any given time there is at most one ondeck thread. 1041 // ondeck implies not resident on cxq and not resident on EntryList 1042 // Only the OnDeck thread can try to acquire -- contended for -- the lock. 1043 // CONSIDER: use Self->OnDeck instead of m->OnDeck. 1044 for (;;) { 1045 if (_OnDeck == ESelf && TrySpin(NULL)) break ; 1046 ParkCommon (ESelf, 0) ; 1047 } 1048 1049 assert (_OnDeck == ESelf, "invariant") ; 1050 _OnDeck = NULL ; 1051 ParkEvent::Release (ESelf) ; // surrender the ParkEvent 1052 goto Exeunt ; 1053 } 1054 1055 void Monitor::jvm_raw_unlock() { 1056 // Nearly the same as Monitor::unlock() ... 1057 // directly set _owner instead of using set_owner(null) 1058 _owner = NULL ; 1059 if (_snuck) { // ??? 1060 assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak"); 1061 _snuck = false; 1062 return ; 1063 } 1064 IUnlock(false) ; 1065 } 1066 1067 bool Monitor::wait(bool no_safepoint_check, long timeout, bool as_suspend_equivalent) { 1068 Thread * const Self = Thread::current() ; 1069 assert (_owner == Self, "invariant") ; 1070 assert (ILocked(), "invariant") ; 1071 1072 // as_suspend_equivalent logically implies !no_safepoint_check 1073 guarantee (!as_suspend_equivalent || !no_safepoint_check, "invariant") ; 1074 // !no_safepoint_check logically implies java_thread 1075 guarantee (no_safepoint_check || Self->is_Java_thread(), "invariant") ; 1076 1077 #ifdef ASSERT 1078 Monitor * least = get_least_ranked_lock_besides_this(Self->owned_locks()); 1079 assert(least != this, "Specification of get_least_... call above"); 1080 if (least != NULL && least->rank() <= special) { 1081 tty->print("Attempting to wait on monitor %s/%d while holding" 1082 " lock %s/%d -- possible deadlock", 1083 name(), rank(), least->name(), least->rank()); 1084 assert(false, "Shouldn't block(wait) while holding a lock of rank special"); 1085 } 1086 #endif // ASSERT 1087 1088 int wait_status ; 1089 // conceptually set the owner to NULL in anticipation of 1090 // abdicating the lock in wait 1091 set_owner(NULL); 1092 if (no_safepoint_check) { 1093 wait_status = IWait (Self, timeout) ; 1094 } else { 1095 assert (Self->is_Java_thread(), "invariant") ; 1096 JavaThread *jt = (JavaThread *)Self; 1097 1098 // Enter safepoint region - ornate and Rococo ... 1099 ThreadBlockInVM tbivm(jt); 1100 OSThreadWaitState osts(Self->osthread(), false /* not Object.wait() */); 1101 1102 if (as_suspend_equivalent) { 1103 jt->set_suspend_equivalent(); 1104 // cleared by handle_special_suspend_equivalent_condition() or 1105 // java_suspend_self() 1106 } 1107 1108 wait_status = IWait (Self, timeout) ; 1109 1110 // were we externally suspended while we were waiting? 1111 if (as_suspend_equivalent && jt->handle_special_suspend_equivalent_condition()) { 1112 // Our event wait has finished and we own the lock, but 1113 // while we were waiting another thread suspended us. We don't 1114 // want to hold the lock while suspended because that 1115 // would surprise the thread that suspended us. 1116 assert (ILocked(), "invariant") ; 1117 IUnlock (true) ; 1118 jt->java_suspend_self(); 1119 ILock (Self) ; 1120 assert (ILocked(), "invariant") ; 1121 } 1122 } 1123 1124 // Conceptually reestablish ownership of the lock. 1125 // The "real" lock -- the LockByte -- was reacquired by IWait(). 1126 assert (ILocked(), "invariant") ; 1127 assert (_owner == NULL, "invariant") ; 1128 set_owner (Self) ; 1129 return wait_status != 0 ; // return true IFF timeout 1130 } 1131 1132 Monitor::~Monitor() { 1133 assert ((UNS(_owner)|UNS(_LockWord.FullWord)|UNS(_EntryList)|UNS(_WaitSet)|UNS(_OnDeck)) == 0, "") ; 1134 } 1135 1136 void Monitor::ClearMonitor (Monitor * m, const char *name) { 1137 m->_owner = NULL ; 1138 m->_snuck = false ; 1139 if (name == NULL) { 1140 strcpy(m->_name, "UNKNOWN") ; 1141 } else { 1142 strncpy(m->_name, name, MONITOR_NAME_LEN - 1); 1143 m->_name[MONITOR_NAME_LEN - 1] = '\0'; 1144 } 1145 m->_LockWord.FullWord = 0 ; 1146 m->_EntryList = NULL ; 1147 m->_OnDeck = NULL ; 1148 m->_WaitSet = NULL ; 1149 m->_WaitLock[0] = 0 ; 1150 } 1151 1152 Monitor::Monitor() { ClearMonitor(this); } 1153 1154 Monitor::Monitor (int Rank, const char * name, bool allow_vm_block) { 1155 ClearMonitor (this, name) ; 1156 #ifdef ASSERT 1157 _allow_vm_block = allow_vm_block; 1158 _rank = Rank ; 1159 #endif 1160 } 1161 1162 Mutex::~Mutex() { 1163 assert ((UNS(_owner)|UNS(_LockWord.FullWord)|UNS(_EntryList)|UNS(_WaitSet)|UNS(_OnDeck)) == 0, "") ; 1164 } 1165 1166 Mutex::Mutex (int Rank, const char * name, bool allow_vm_block) { 1167 ClearMonitor ((Monitor *) this, name) ; 1168 #ifdef ASSERT 1169 _allow_vm_block = allow_vm_block; 1170 _rank = Rank ; 1171 #endif 1172 } 1173 1174 bool Monitor::owned_by_self() const { 1175 bool ret = _owner == Thread::current(); 1176 assert (!ret || _LockWord.Bytes[_LSBINDEX] != 0, "invariant") ; 1177 return ret; 1178 } 1179 1180 void Monitor::print_on_error(outputStream* st) const { 1181 st->print("[" PTR_FORMAT, this); 1182 st->print("] %s", _name); 1183 st->print(" - owner thread: " PTR_FORMAT, _owner); 1184 } 1185 1186 1187 1188 1189 // ---------------------------------------------------------------------------------- 1190 // Non-product code 1191 1192 #ifndef PRODUCT 1193 void Monitor::print_on(outputStream* st) const { 1194 st->print_cr("Mutex: [0x%lx/0x%lx] %s - owner: 0x%lx", this, _LockWord.FullWord, _name, _owner); 1195 } 1196 #endif 1197 1198 #ifndef PRODUCT 1199 #ifdef ASSERT 1200 Monitor * Monitor::get_least_ranked_lock(Monitor * locks) { 1201 Monitor *res, *tmp; 1202 for (res = tmp = locks; tmp != NULL; tmp = tmp->next()) { 1203 if (tmp->rank() < res->rank()) { 1204 res = tmp; 1205 } 1206 } 1207 if (!SafepointSynchronize::is_at_safepoint()) { 1208 // In this case, we expect the held locks to be 1209 // in increasing rank order (modulo any native ranks) 1210 for (tmp = locks; tmp != NULL; tmp = tmp->next()) { 1211 if (tmp->next() != NULL) { 1212 assert(tmp->rank() == Mutex::native || 1213 tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?"); 1214 } 1215 } 1216 } 1217 return res; 1218 } 1219 1220 Monitor* Monitor::get_least_ranked_lock_besides_this(Monitor* locks) { 1221 Monitor *res, *tmp; 1222 for (res = NULL, tmp = locks; tmp != NULL; tmp = tmp->next()) { 1223 if (tmp != this && (res == NULL || tmp->rank() < res->rank())) { 1224 res = tmp; 1225 } 1226 } 1227 if (!SafepointSynchronize::is_at_safepoint()) { 1228 // In this case, we expect the held locks to be 1229 // in increasing rank order (modulo any native ranks) 1230 for (tmp = locks; tmp != NULL; tmp = tmp->next()) { 1231 if (tmp->next() != NULL) { 1232 assert(tmp->rank() == Mutex::native || 1233 tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?"); 1234 } 1235 } 1236 } 1237 return res; 1238 } 1239 1240 1241 bool Monitor::contains(Monitor* locks, Monitor * lock) { 1242 for (; locks != NULL; locks = locks->next()) { 1243 if (locks == lock) 1244 return true; 1245 } 1246 return false; 1247 } 1248 #endif 1249 1250 // Called immediately after lock acquisition or release as a diagnostic 1251 // to track the lock-set of the thread and test for rank violations that 1252 // might indicate exposure to deadlock. 1253 // Rather like an EventListener for _owner (:>). 1254 1255 void Monitor::set_owner_implementation(Thread *new_owner) { 1256 // This function is solely responsible for maintaining 1257 // and checking the invariant that threads and locks 1258 // are in a 1/N relation, with some some locks unowned. 1259 // It uses the Mutex::_owner, Mutex::_next, and 1260 // Thread::_owned_locks fields, and no other function 1261 // changes those fields. 1262 // It is illegal to set the mutex from one non-NULL 1263 // owner to another--it must be owned by NULL as an 1264 // intermediate state. 1265 1266 if (new_owner != NULL) { 1267 // the thread is acquiring this lock 1268 1269 assert(new_owner == Thread::current(), "Should I be doing this?"); 1270 assert(_owner == NULL, "setting the owner thread of an already owned mutex"); 1271 _owner = new_owner; // set the owner 1272 1273 // link "this" into the owned locks list 1274 1275 #ifdef ASSERT // Thread::_owned_locks is under the same ifdef 1276 Monitor* locks = get_least_ranked_lock(new_owner->owned_locks()); 1277 // Mutex::set_owner_implementation is a friend of Thread 1278 1279 assert(this->rank() >= 0, "bad lock rank"); 1280 1281 if (LogMultipleMutexLocking && locks != NULL) { 1282 Events::log("thread " INTPTR_FORMAT " locks %s, already owns %s", new_owner, name(), locks->name()); 1283 } 1284 1285 // Deadlock avoidance rules require us to acquire Mutexes only in 1286 // a global total order. For example m1 is the lowest ranked mutex 1287 // that the thread holds and m2 is the mutex the thread is trying 1288 // to acquire, then deadlock avoidance rules require that the rank 1289 // of m2 be less than the rank of m1. 1290 // The rank Mutex::native is an exception in that it is not subject 1291 // to the verification rules. 1292 // Here are some further notes relating to mutex acquisition anomalies: 1293 // . under Solaris, the interrupt lock gets acquired when doing 1294 // profiling, so any lock could be held. 1295 // . it is also ok to acquire Safepoint_lock at the very end while we 1296 // already hold Terminator_lock - may happen because of periodic safepoints 1297 if (this->rank() != Mutex::native && 1298 this->rank() != Mutex::suspend_resume && 1299 locks != NULL && locks->rank() <= this->rank() && 1300 !SafepointSynchronize::is_at_safepoint() && 1301 this != Interrupt_lock && this != ProfileVM_lock && 1302 !(this == Safepoint_lock && contains(locks, Terminator_lock) && 1303 SafepointSynchronize::is_synchronizing())) { 1304 new_owner->print_owned_locks(); 1305 fatal(err_msg("acquiring lock %s/%d out of order with lock %s/%d -- " 1306 "possible deadlock", this->name(), this->rank(), 1307 locks->name(), locks->rank())); 1308 } 1309 1310 this->_next = new_owner->_owned_locks; 1311 new_owner->_owned_locks = this; 1312 #endif 1313 1314 } else { 1315 // the thread is releasing this lock 1316 1317 Thread* old_owner = _owner; 1318 debug_only(_last_owner = old_owner); 1319 1320 assert(old_owner != NULL, "removing the owner thread of an unowned mutex"); 1321 assert(old_owner == Thread::current(), "removing the owner thread of an unowned mutex"); 1322 1323 _owner = NULL; // set the owner 1324 1325 #ifdef ASSERT 1326 Monitor *locks = old_owner->owned_locks(); 1327 1328 if (LogMultipleMutexLocking && locks != this) { 1329 Events::log("thread " INTPTR_FORMAT " unlocks %s, still owns %s", old_owner, this->name(), locks->name()); 1330 } 1331 1332 // remove "this" from the owned locks list 1333 1334 Monitor *prev = NULL; 1335 bool found = false; 1336 for (; locks != NULL; prev = locks, locks = locks->next()) { 1337 if (locks == this) { 1338 found = true; 1339 break; 1340 } 1341 } 1342 assert(found, "Removing a lock not owned"); 1343 if (prev == NULL) { 1344 old_owner->_owned_locks = _next; 1345 } else { 1346 prev->_next = _next; 1347 } 1348 _next = NULL; 1349 #endif 1350 } 1351 } 1352 1353 1354 // Factored out common sanity checks for locking mutex'es. Used by lock() and try_lock() 1355 void Monitor::check_prelock_state(Thread *thread) { 1356 assert((!thread->is_Java_thread() || ((JavaThread *)thread)->thread_state() == _thread_in_vm) 1357 || rank() == Mutex::special, "wrong thread state for using locks"); 1358 if (StrictSafepointChecks) { 1359 if (thread->is_VM_thread() && !allow_vm_block()) { 1360 fatal(err_msg("VM thread using lock %s (not allowed to block on)", 1361 name())); 1362 } 1363 debug_only(if (rank() != Mutex::special) \ 1364 thread->check_for_valid_safepoint_state(false);) 1365 } 1366 } 1367 1368 void Monitor::check_block_state(Thread *thread) { 1369 if (!_allow_vm_block && thread->is_VM_thread()) { 1370 warning("VM thread blocked on lock"); 1371 print(); 1372 BREAKPOINT; 1373 } 1374 assert(_owner != thread, "deadlock: blocking on monitor owned by current thread"); 1375 } 1376 1377 #endif // PRODUCT