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