1 /* 2 * Copyright (c) 1998, 2018, 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.inline.hpp" 28 #include "runtime/mutex.hpp" 29 #include "runtime/orderAccess.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 #define UNS(x) (uintptr_t(x)) 256 #define TRACE(m) \ 257 { \ 258 static volatile int ctr = 0; \ 259 int x = ++ctr; \ 260 if ((x & (x - 1)) == 0) { \ 261 ::printf("%d:%s\n", x, #m); \ 262 ::fflush(stdout); \ 263 } \ 264 } 265 266 const intptr_t _LBIT = 1; 267 268 // Endian-ness ... index of least-significant byte in SplitWord.Bytes[] 269 #ifdef VM_LITTLE_ENDIAN 270 #define _LSBINDEX 0 271 #else 272 #define _LSBINDEX (sizeof(intptr_t)-1) 273 #endif 274 275 // Simplistic low-quality Marsaglia SHIFT-XOR RNG. 276 // Bijective except for the trailing mask operation. 277 // Useful for spin loops as the compiler can't optimize it away. 278 279 static inline jint MarsagliaXORV(jint x) { 280 if (x == 0) x = 1|os::random(); 281 x ^= x << 6; 282 x ^= ((unsigned)x) >> 21; 283 x ^= x << 7; 284 return x & 0x7FFFFFFF; 285 } 286 287 static int Stall(int its) { 288 static volatile jint rv = 1; 289 volatile int OnFrame = 0; 290 jint v = rv ^ UNS(OnFrame); 291 while (--its >= 0) { 292 v = MarsagliaXORV(v); 293 } 294 // Make this impossible for the compiler to optimize away, 295 // but (mostly) avoid W coherency sharing on MP systems. 296 if (v == 0x12345) rv = v; 297 return v; 298 } 299 300 int Monitor::TryLock() { 301 intptr_t v = _LockWord.FullWord; 302 for (;;) { 303 if ((v & _LBIT) != 0) return 0; 304 const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v); 305 if (v == u) return 1; 306 v = u; 307 } 308 } 309 310 int Monitor::TryFast() { 311 // Optimistic fast-path form ... 312 // Fast-path attempt for the common uncontended case. 313 // Avoid RTS->RTO $ coherence upgrade on typical SMP systems. 314 intptr_t v = Atomic::cmpxchg(_LBIT, &_LockWord.FullWord, (intptr_t)0); // agro ... 315 if (v == 0) return 1; 316 317 for (;;) { 318 if ((v & _LBIT) != 0) return 0; 319 const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v); 320 if (v == u) return 1; 321 v = u; 322 } 323 } 324 325 int Monitor::ILocked() { 326 const intptr_t w = _LockWord.FullWord & 0xFF; 327 assert(w == 0 || w == _LBIT, "invariant"); 328 return w == _LBIT; 329 } 330 331 // Polite TATAS spinlock with exponential backoff - bounded spin. 332 // Ideally we'd use processor cycles, time or vtime to control 333 // the loop, but we currently use iterations. 334 // All the constants within were derived empirically but work over 335 // over the spectrum of J2SE reference platforms. 336 // On Niagara-class systems the back-off is unnecessary but 337 // is relatively harmless. (At worst it'll slightly retard 338 // acquisition times). The back-off is critical for older SMP systems 339 // where constant fetching of the LockWord would otherwise impair 340 // scalability. 341 // 342 // Clamp spinning at approximately 1/2 of a context-switch round-trip. 343 // See synchronizer.cpp for details and rationale. 344 345 int Monitor::TrySpin(Thread * const Self) { 346 if (TryLock()) return 1; 347 if (!os::is_MP()) return 0; 348 349 int Probes = 0; 350 int Delay = 0; 351 int SpinMax = 20; 352 for (;;) { 353 intptr_t v = _LockWord.FullWord; 354 if ((v & _LBIT) == 0) { 355 if (Atomic::cmpxchg (v|_LBIT, &_LockWord.FullWord, v) == v) { 356 return 1; 357 } 358 continue; 359 } 360 361 SpinPause(); 362 363 // Periodically increase Delay -- variable Delay form 364 // conceptually: delay *= 1 + 1/Exponent 365 ++Probes; 366 if (Probes > SpinMax) return 0; 367 368 if ((Probes & 0x7) == 0) { 369 Delay = ((Delay << 1)|1) & 0x7FF; 370 // CONSIDER: Delay += 1 + (Delay/4); Delay &= 0x7FF ; 371 } 372 373 // Consider checking _owner's schedctl state, if OFFPROC abort spin. 374 // If the owner is OFFPROC then it's unlike that the lock will be dropped 375 // in a timely fashion, which suggests that spinning would not be fruitful 376 // or profitable. 377 378 // Stall for "Delay" time units - iterations in the current implementation. 379 // Avoid generating coherency traffic while stalled. 380 // Possible ways to delay: 381 // PAUSE, SLEEP, MEMBAR #sync, MEMBAR #halt, 382 // wr %g0,%asi, gethrtime, rdstick, rdtick, rdtsc, etc. ... 383 // Note that on Niagara-class systems we want to minimize STs in the 384 // spin loop. N1 and brethren write-around the L1$ over the xbar into the L2$. 385 // Furthermore, they don't have a W$ like traditional SPARC processors. 386 // We currently use a Marsaglia Shift-Xor RNG loop. 387 if (Self != NULL) { 388 jint rv = Self->rng[0]; 389 for (int k = Delay; --k >= 0;) { 390 rv = MarsagliaXORV(rv); 391 if (SafepointMechanism::poll(Self)) return 0; 392 } 393 Self->rng[0] = rv; 394 } else { 395 Stall(Delay); 396 } 397 } 398 } 399 400 static int ParkCommon(ParkEvent * ev, jlong timo) { 401 // Diagnostic support - periodically unwedge blocked threads 402 int err = OS_OK; 403 if (0 == timo) { 404 ev->park(); 405 } else { 406 err = ev->park(timo); 407 } 408 return err; 409 } 410 411 inline int Monitor::AcquireOrPush(ParkEvent * ESelf) { 412 intptr_t v = _LockWord.FullWord; 413 for (;;) { 414 if ((v & _LBIT) == 0) { 415 const intptr_t u = Atomic::cmpxchg(v|_LBIT, &_LockWord.FullWord, v); 416 if (u == v) return 1; // indicate acquired 417 v = u; 418 } else { 419 // Anticipate success ... 420 ESelf->ListNext = (ParkEvent *)(v & ~_LBIT); 421 const intptr_t u = Atomic::cmpxchg(intptr_t(ESelf)|_LBIT, &_LockWord.FullWord, v); 422 if (u == v) return 0; // indicate pushed onto cxq 423 v = u; 424 } 425 // Interference - LockWord change - just retry 426 } 427 } 428 429 // ILock and IWait are the lowest level primitive internal blocking 430 // synchronization functions. The callers of IWait and ILock must have 431 // performed any needed state transitions beforehand. 432 // IWait and ILock may directly call park() without any concern for thread state. 433 // Note that ILock and IWait do *not* access _owner. 434 // _owner is a higher-level logical concept. 435 436 void Monitor::ILock(Thread * Self) { 437 assert(_OnDeck != Self->_MutexEvent, "invariant"); 438 439 if (TryFast()) { 440 Exeunt: 441 assert(ILocked(), "invariant"); 442 return; 443 } 444 445 ParkEvent * const ESelf = Self->_MutexEvent; 446 assert(_OnDeck != ESelf, "invariant"); 447 448 // As an optimization, spinners could conditionally try to set _OnDeck to _LBIT 449 // Synchronizer.cpp uses a similar optimization. 450 if (TrySpin(Self)) goto Exeunt; 451 452 // Slow-path - the lock is contended. 453 // Either Enqueue Self on cxq or acquire the outer lock. 454 // LockWord encoding = (cxq,LOCKBYTE) 455 ESelf->reset(); 456 OrderAccess::fence(); 457 458 if (AcquireOrPush(ESelf)) goto Exeunt; 459 460 // At any given time there is at most one ondeck thread. 461 // ondeck implies not resident on cxq and not resident on EntryList 462 // Only the OnDeck thread can try to acquire -- contend for -- the lock. 463 // CONSIDER: use Self->OnDeck instead of m->OnDeck. 464 // Deschedule Self so that others may run. 465 while (OrderAccess::load_acquire(&_OnDeck) != ESelf) { 466 ParkCommon(ESelf, 0); 467 } 468 469 // Self is now in the OnDeck position and will remain so until it 470 // manages to acquire the lock. 471 for (;;) { 472 assert(_OnDeck == ESelf, "invariant"); 473 if (TrySpin(Self)) break; 474 // It's probably wise to spin only if we *actually* blocked 475 // CONSIDER: check the lockbyte, if it remains set then 476 // preemptively drain the cxq into the EntryList. 477 // The best place and time to perform queue operations -- lock metadata -- 478 // is _before having acquired the outer lock, while waiting for the lock to drop. 479 ParkCommon(ESelf, 0); 480 } 481 482 assert(_OnDeck == ESelf, "invariant"); 483 _OnDeck = NULL; 484 485 // Note that we current drop the inner lock (clear OnDeck) in the slow-path 486 // epilogue immediately after having acquired the outer lock. 487 // But instead we could consider the following optimizations: 488 // A. Shift or defer dropping the inner lock until the subsequent IUnlock() operation. 489 // This might avoid potential reacquisition of the inner lock in IUlock(). 490 // B. While still holding the inner lock, attempt to opportunistically select 491 // and unlink the next OnDeck thread from the EntryList. 492 // If successful, set OnDeck to refer to that thread, otherwise clear OnDeck. 493 // It's critical that the select-and-unlink operation run in constant-time as 494 // it executes when holding the outer lock and may artificially increase the 495 // effective length of the critical section. 496 // Note that (A) and (B) are tantamount to succession by direct handoff for 497 // the inner lock. 498 goto Exeunt; 499 } 500 501 void Monitor::IUnlock(bool RelaxAssert) { 502 assert(ILocked(), "invariant"); 503 // Conceptually we need a MEMBAR #storestore|#loadstore barrier or fence immediately 504 // before the store that releases the lock. Crucially, all the stores and loads in the 505 // critical section must be globally visible before the store of 0 into the lock-word 506 // that releases the lock becomes globally visible. That is, memory accesses in the 507 // critical section should not be allowed to bypass or overtake the following ST that 508 // releases the lock. As such, to prevent accesses within the critical section 509 // from "leaking" out, we need a release fence between the critical section and the 510 // store that releases the lock. In practice that release barrier is elided on 511 // platforms with strong memory models such as TSO. 512 // 513 // Note that the OrderAccess::storeload() fence that appears after unlock store 514 // provides for progress conditions and succession and is _not related to exclusion 515 // safety or lock release consistency. 516 OrderAccess::release_store(&_LockWord.Bytes[_LSBINDEX], jbyte(0)); // drop outer lock 517 518 OrderAccess::storeload(); 519 ParkEvent * const w = _OnDeck; // raw load as we will just return if non-NULL 520 assert(RelaxAssert || w != Thread::current()->_MutexEvent, "invariant"); 521 if (w != NULL) { 522 // Either we have a valid ondeck thread or ondeck is transiently "locked" 523 // by some exiting thread as it arranges for succession. The LSBit of 524 // OnDeck allows us to discriminate two cases. If the latter, the 525 // responsibility for progress and succession lies with that other thread. 526 // For good performance, we also depend on the fact that redundant unpark() 527 // operations are cheap. That is, repeated Unpark()ing of the OnDeck thread 528 // is inexpensive. This approach provides implicit futile wakeup throttling. 529 // Note that the referent "w" might be stale with respect to the lock. 530 // In that case the following unpark() is harmless and the worst that'll happen 531 // is a spurious return from a park() operation. Critically, if "w" _is stale, 532 // then progress is known to have occurred as that means the thread associated 533 // with "w" acquired the lock. In that case this thread need take no further 534 // action to guarantee progress. 535 if ((UNS(w) & _LBIT) == 0) w->unpark(); 536 return; 537 } 538 539 intptr_t cxq = _LockWord.FullWord; 540 if (((cxq & ~_LBIT)|UNS(_EntryList)) == 0) { 541 return; // normal fast-path exit - cxq and EntryList both empty 542 } 543 if (cxq & _LBIT) { 544 // Optional optimization ... 545 // Some other thread acquired the lock in the window since this 546 // thread released it. Succession is now that thread's responsibility. 547 return; 548 } 549 550 Succession: 551 // Slow-path exit - this thread must ensure succession and progress. 552 // OnDeck serves as lock to protect cxq and EntryList. 553 // Only the holder of OnDeck can manipulate EntryList or detach the RATs from cxq. 554 // Avoid ABA - allow multiple concurrent producers (enqueue via push-CAS) 555 // but only one concurrent consumer (detacher of RATs). 556 // Consider protecting this critical section with schedctl on Solaris. 557 // Unlike a normal lock, however, the exiting thread "locks" OnDeck, 558 // picks a successor and marks that thread as OnDeck. That successor 559 // thread will then clear OnDeck once it eventually acquires the outer lock. 560 if (!Atomic::replace_if_null((ParkEvent*)_LBIT, &_OnDeck)) { 561 return; 562 } 563 564 ParkEvent * List = _EntryList; 565 if (List != NULL) { 566 // Transfer the head of the EntryList to the OnDeck position. 567 // Once OnDeck, a thread stays OnDeck until it acquires the lock. 568 // For a given lock there is at most OnDeck thread at any one instant. 569 WakeOne: 570 assert(List == _EntryList, "invariant"); 571 ParkEvent * const w = List; 572 assert(RelaxAssert || w != Thread::current()->_MutexEvent, "invariant"); 573 _EntryList = w->ListNext; 574 // as a diagnostic measure consider setting w->_ListNext = BAD 575 assert(intptr_t(_OnDeck) == _LBIT, "invariant"); 576 577 // Pass OnDeck role to w, ensuring that _EntryList has been set first. 578 // w will clear _OnDeck once it acquires the outer lock. 579 // Note that once we set _OnDeck that thread can acquire the mutex, proceed 580 // with its critical section and then enter this code to unlock the mutex. So 581 // you can have multiple threads active in IUnlock at the same time. 582 OrderAccess::release_store(&_OnDeck, w); 583 584 // Another optional optimization ... 585 // For heavily contended locks it's not uncommon that some other 586 // thread acquired the lock while this thread was arranging succession. 587 // Try to defer the unpark() operation - Delegate the responsibility 588 // for unpark()ing the OnDeck thread to the current or subsequent owners 589 // That is, the new owner is responsible for unparking the OnDeck thread. 590 OrderAccess::storeload(); 591 cxq = _LockWord.FullWord; 592 if (cxq & _LBIT) return; 593 594 w->unpark(); 595 return; 596 } 597 598 cxq = _LockWord.FullWord; 599 if ((cxq & ~_LBIT) != 0) { 600 // The EntryList is empty but the cxq is populated. 601 // drain RATs from cxq into EntryList 602 // Detach RATs segment with CAS and then merge into EntryList 603 for (;;) { 604 // optional optimization - if locked, the owner is responsible for succession 605 if (cxq & _LBIT) goto Punt; 606 const intptr_t vfy = Atomic::cmpxchg(cxq & _LBIT, &_LockWord.FullWord, cxq); 607 if (vfy == cxq) break; 608 cxq = vfy; 609 // Interference - LockWord changed - Just retry 610 // We can see concurrent interference from contending threads 611 // pushing themselves onto the cxq or from lock-unlock operations. 612 // From the perspective of this thread, EntryList is stable and 613 // the cxq is prepend-only -- the head is volatile but the interior 614 // of the cxq is stable. In theory if we encounter interference from threads 615 // pushing onto cxq we could simply break off the original cxq suffix and 616 // move that segment to the EntryList, avoiding a 2nd or multiple CAS attempts 617 // on the high-traffic LockWord variable. For instance lets say the cxq is "ABCD" 618 // when we first fetch cxq above. Between the fetch -- where we observed "A" 619 // -- and CAS -- where we attempt to CAS null over A -- "PQR" arrive, 620 // yielding cxq = "PQRABCD". In this case we could simply set A.ListNext 621 // null, leaving cxq = "PQRA" and transfer the "BCD" segment to the EntryList. 622 // Note too, that it's safe for this thread to traverse the cxq 623 // without taking any special concurrency precautions. 624 } 625 626 // We don't currently reorder the cxq segment as we move it onto 627 // the EntryList, but it might make sense to reverse the order 628 // or perhaps sort by thread priority. See the comments in 629 // synchronizer.cpp objectMonitor::exit(). 630 assert(_EntryList == NULL, "invariant"); 631 _EntryList = List = (ParkEvent *)(cxq & ~_LBIT); 632 assert(List != NULL, "invariant"); 633 goto WakeOne; 634 } 635 636 // cxq|EntryList is empty. 637 // w == NULL implies that cxq|EntryList == NULL in the past. 638 // Possible race - rare inopportune interleaving. 639 // A thread could have added itself to cxq since this thread previously checked. 640 // Detect and recover by refetching cxq. 641 Punt: 642 assert(intptr_t(_OnDeck) == _LBIT, "invariant"); 643 _OnDeck = NULL; // Release inner lock. 644 OrderAccess::storeload(); // Dekker duality - pivot point 645 646 // Resample LockWord/cxq to recover from possible race. 647 // For instance, while this thread T1 held OnDeck, some other thread T2 might 648 // acquire the outer lock. Another thread T3 might try to acquire the outer 649 // lock, but encounter contention and enqueue itself on cxq. T2 then drops the 650 // outer lock, but skips succession as this thread T1 still holds OnDeck. 651 // T1 is and remains responsible for ensuring succession of T3. 652 // 653 // Note that we don't need to recheck EntryList, just cxq. 654 // If threads moved onto EntryList since we dropped OnDeck 655 // that implies some other thread forced succession. 656 cxq = _LockWord.FullWord; 657 if ((cxq & ~_LBIT) != 0 && (cxq & _LBIT) == 0) { 658 goto Succession; // potential race -- re-run succession 659 } 660 return; 661 } 662 663 bool Monitor::notify() { 664 assert(_owner == Thread::current(), "invariant"); 665 assert(ILocked(), "invariant"); 666 if (_WaitSet == NULL) return true; 667 668 // Transfer one thread from the WaitSet to the EntryList or cxq. 669 // Currently we just unlink the head of the WaitSet and prepend to the cxq. 670 // And of course we could just unlink it and unpark it, too, but 671 // in that case it'd likely impale itself on the reentry. 672 Thread::muxAcquire(_WaitLock, "notify:WaitLock"); 673 ParkEvent * nfy = _WaitSet; 674 if (nfy != NULL) { // DCL idiom 675 _WaitSet = nfy->ListNext; 676 assert(nfy->Notified == 0, "invariant"); 677 // push nfy onto the cxq 678 for (;;) { 679 const intptr_t v = _LockWord.FullWord; 680 assert((v & 0xFF) == _LBIT, "invariant"); 681 nfy->ListNext = (ParkEvent *)(v & ~_LBIT); 682 if (Atomic::cmpxchg(intptr_t(nfy)|_LBIT, &_LockWord.FullWord, v) == v) break; 683 // interference - _LockWord changed -- just retry 684 } 685 // Note that setting Notified before pushing nfy onto the cxq is 686 // also legal and safe, but the safety properties are much more 687 // subtle, so for the sake of code stewardship ... 688 OrderAccess::fence(); 689 nfy->Notified = 1; 690 } 691 Thread::muxRelease(_WaitLock); 692 assert(ILocked(), "invariant"); 693 return true; 694 } 695 696 // Currently notifyAll() transfers the waiters one-at-a-time from the waitset 697 // to the cxq. This could be done more efficiently with a single bulk en-mass transfer, 698 // but in practice notifyAll() for large #s of threads is rare and not time-critical. 699 // Beware too, that we invert the order of the waiters. Lets say that the 700 // waitset is "ABCD" and the cxq is "XYZ". After a notifyAll() the waitset 701 // will be empty and the cxq will be "DCBAXYZ". This is benign, of course. 702 703 bool Monitor::notify_all() { 704 assert(_owner == Thread::current(), "invariant"); 705 assert(ILocked(), "invariant"); 706 while (_WaitSet != NULL) notify(); 707 return true; 708 } 709 710 int Monitor::IWait(Thread * Self, jlong timo) { 711 assert(ILocked(), "invariant"); 712 713 // Phases: 714 // 1. Enqueue Self on WaitSet - currently prepend 715 // 2. unlock - drop the outer lock 716 // 3. wait for either notification or timeout 717 // 4. lock - reentry - reacquire the outer lock 718 719 ParkEvent * const ESelf = Self->_MutexEvent; 720 ESelf->Notified = 0; 721 ESelf->reset(); 722 OrderAccess::fence(); 723 724 // Add Self to WaitSet 725 // Ideally only the holder of the outer lock would manipulate the WaitSet - 726 // That is, the outer lock would implicitly protect the WaitSet. 727 // But if a thread in wait() encounters a timeout it will need to dequeue itself 728 // from the WaitSet _before it becomes the owner of the lock. We need to dequeue 729 // as the ParkEvent -- which serves as a proxy for the thread -- can't reside 730 // on both the WaitSet and the EntryList|cxq at the same time.. That is, a thread 731 // on the WaitSet can't be allowed to compete for the lock until it has managed to 732 // unlink its ParkEvent from WaitSet. Thus the need for WaitLock. 733 // Contention on the WaitLock is minimal. 734 // 735 // Another viable approach would be add another ParkEvent, "WaitEvent" to the 736 // thread class. The WaitSet would be composed of WaitEvents. Only the 737 // owner of the outer lock would manipulate the WaitSet. A thread in wait() 738 // could then compete for the outer lock, and then, if necessary, unlink itself 739 // from the WaitSet only after having acquired the outer lock. More precisely, 740 // there would be no WaitLock. A thread in in wait() would enqueue its WaitEvent 741 // on the WaitSet; release the outer lock; wait for either notification or timeout; 742 // reacquire the inner lock; and then, if needed, unlink itself from the WaitSet. 743 // 744 // Alternatively, a 2nd set of list link fields in the ParkEvent might suffice. 745 // One set would be for the WaitSet and one for the EntryList. 746 // We could also deconstruct the ParkEvent into a "pure" event and add a 747 // new immortal/TSM "ListElement" class that referred to ParkEvents. 748 // In that case we could have one ListElement on the WaitSet and another 749 // on the EntryList, with both referring to the same pure Event. 750 751 Thread::muxAcquire(_WaitLock, "wait:WaitLock:Add"); 752 ESelf->ListNext = _WaitSet; 753 _WaitSet = ESelf; 754 Thread::muxRelease(_WaitLock); 755 756 // Release the outer lock 757 // We call IUnlock (RelaxAssert=true) as a thread T1 might 758 // enqueue itself on the WaitSet, call IUnlock(), drop the lock, 759 // and then stall before it can attempt to wake a successor. 760 // Some other thread T2 acquires the lock, and calls notify(), moving 761 // T1 from the WaitSet to the cxq. T2 then drops the lock. T1 resumes, 762 // and then finds *itself* on the cxq. During the course of a normal 763 // IUnlock() call a thread should _never find itself on the EntryList 764 // or cxq, but in the case of wait() it's possible. 765 // See synchronizer.cpp objectMonitor::wait(). 766 IUnlock(true); 767 768 // Wait for either notification or timeout 769 // Beware that in some circumstances we might propagate 770 // spurious wakeups back to the caller. 771 772 for (;;) { 773 if (ESelf->Notified) break; 774 int err = ParkCommon(ESelf, timo); 775 if (err == OS_TIMEOUT) break; 776 } 777 778 // Prepare for reentry - if necessary, remove ESelf from WaitSet 779 // ESelf can be: 780 // 1. Still on the WaitSet. This can happen if we exited the loop by timeout. 781 // 2. On the cxq or EntryList 782 // 3. Not resident on cxq, EntryList or WaitSet, but in the OnDeck position. 783 784 OrderAccess::fence(); 785 int WasOnWaitSet = 0; 786 if (ESelf->Notified == 0) { 787 Thread::muxAcquire(_WaitLock, "wait:WaitLock:remove"); 788 if (ESelf->Notified == 0) { // DCL idiom 789 assert(_OnDeck != ESelf, "invariant"); // can't be both OnDeck and on WaitSet 790 // ESelf is resident on the WaitSet -- unlink it. 791 // A doubly-linked list would be better here so we can unlink in constant-time. 792 // We have to unlink before we potentially recontend as ESelf might otherwise 793 // end up on the cxq|EntryList -- it can't be on two lists at once. 794 ParkEvent * p = _WaitSet; 795 ParkEvent * q = NULL; // classic q chases p 796 while (p != NULL && p != ESelf) { 797 q = p; 798 p = p->ListNext; 799 } 800 assert(p == ESelf, "invariant"); 801 if (p == _WaitSet) { // found at head 802 assert(q == NULL, "invariant"); 803 _WaitSet = p->ListNext; 804 } else { // found in interior 805 assert(q->ListNext == p, "invariant"); 806 q->ListNext = p->ListNext; 807 } 808 WasOnWaitSet = 1; // We were *not* notified but instead encountered timeout 809 } 810 Thread::muxRelease(_WaitLock); 811 } 812 813 // Reentry phase - reacquire the lock 814 if (WasOnWaitSet) { 815 // ESelf was previously on the WaitSet but we just unlinked it above 816 // because of a timeout. ESelf is not resident on any list and is not OnDeck 817 assert(_OnDeck != ESelf, "invariant"); 818 ILock(Self); 819 } else { 820 // A prior notify() operation moved ESelf from the WaitSet to the cxq. 821 // ESelf is now on the cxq, EntryList or at the OnDeck position. 822 // The following fragment is extracted from Monitor::ILock() 823 for (;;) { 824 if (OrderAccess::load_acquire(&_OnDeck) == ESelf && TrySpin(Self)) break; 825 ParkCommon(ESelf, 0); 826 } 827 assert(_OnDeck == ESelf, "invariant"); 828 _OnDeck = NULL; 829 } 830 831 assert(ILocked(), "invariant"); 832 return WasOnWaitSet != 0; // return true IFF timeout 833 } 834 835 836 // ON THE VMTHREAD SNEAKING PAST HELD LOCKS: 837 // In particular, there are certain types of global lock that may be held 838 // by a Java thread while it is blocked at a safepoint but before it has 839 // written the _owner field. These locks may be sneakily acquired by the 840 // VM thread during a safepoint to avoid deadlocks. Alternatively, one should 841 // identify all such locks, and ensure that Java threads never block at 842 // safepoints while holding them (_no_safepoint_check_flag). While it 843 // seems as though this could increase the time to reach a safepoint 844 // (or at least increase the mean, if not the variance), the latter 845 // approach might make for a cleaner, more maintainable JVM design. 846 // 847 // Sneaking is vile and reprehensible and should be excised at the 1st 848 // opportunity. It's possible that the need for sneaking could be obviated 849 // as follows. Currently, a thread might (a) while TBIVM, call pthread_mutex_lock 850 // or ILock() thus acquiring the "physical" lock underlying Monitor/Mutex. 851 // (b) stall at the TBIVM exit point as a safepoint is in effect. Critically, 852 // it'll stall at the TBIVM reentry state transition after having acquired the 853 // underlying lock, but before having set _owner and having entered the actual 854 // critical section. The lock-sneaking facility leverages that fact and allowed the 855 // VM thread to logically acquire locks that had already be physically locked by mutators 856 // but where mutators were known blocked by the reentry thread state transition. 857 // 858 // If we were to modify the Monitor-Mutex so that TBIVM state transitions tightly 859 // wrapped calls to park(), then we could likely do away with sneaking. We'd 860 // decouple lock acquisition and parking. The critical invariant to eliminating 861 // sneaking is to ensure that we never "physically" acquire the lock while TBIVM. 862 // An easy way to accomplish this is to wrap the park calls in a narrow TBIVM jacket. 863 // One difficulty with this approach is that the TBIVM wrapper could recurse and 864 // call lock() deep from within a lock() call, while the MutexEvent was already enqueued. 865 // Using a stack (N=2 at minimum) of ParkEvents would take care of that problem. 866 // 867 // But of course the proper ultimate approach is to avoid schemes that require explicit 868 // sneaking or dependence on any any clever invariants or subtle implementation properties 869 // of Mutex-Monitor and instead directly address the underlying design flaw. 870 871 void Monitor::lock(Thread * Self) { 872 // Ensure that the Monitor requires/allows safepoint checks. 873 assert(_safepoint_check_required != Monitor::_safepoint_check_never, 874 "This lock should never have a safepoint check: %s", name()); 875 876 #ifdef CHECK_UNHANDLED_OOPS 877 // Clear unhandled oops so we get a crash right away. Only clear for non-vm 878 // or GC threads. 879 if (Self->is_Java_thread()) { 880 Self->clear_unhandled_oops(); 881 } 882 #endif // CHECK_UNHANDLED_OOPS 883 884 debug_only(check_prelock_state(Self, StrictSafepointChecks)); 885 assert(_owner != Self, "invariant"); 886 assert(_OnDeck != Self->_MutexEvent, "invariant"); 887 888 if (TryFast()) { 889 Exeunt: 890 assert(ILocked(), "invariant"); 891 assert(owner() == NULL, "invariant"); 892 set_owner(Self); 893 return; 894 } 895 896 // The lock is contended ... 897 898 bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint(); 899 if (can_sneak && _owner == NULL) { 900 // a java thread has locked the lock but has not entered the 901 // critical region -- let's just pretend we've locked the lock 902 // and go on. we note this with _snuck so we can also 903 // pretend to unlock when the time comes. 904 _snuck = true; 905 goto Exeunt; 906 } 907 908 // Try a brief spin to avoid passing thru thread state transition ... 909 if (TrySpin(Self)) goto Exeunt; 910 911 check_block_state(Self); 912 if (Self->is_Java_thread()) { 913 // Horrible dictu - we suffer through a state transition 914 assert(rank() > Mutex::special, "Potential deadlock with special or lesser rank mutex"); 915 ThreadBlockInVM tbivm((JavaThread *) Self); 916 ILock(Self); 917 } else { 918 // Mirabile dictu 919 ILock(Self); 920 } 921 goto Exeunt; 922 } 923 924 void Monitor::lock() { 925 this->lock(Thread::current()); 926 } 927 928 // Lock without safepoint check - a degenerate variant of lock(). 929 // Should ONLY be used by safepoint code and other code 930 // that is guaranteed not to block while running inside the VM. If this is called with 931 // thread state set to be in VM, the safepoint synchronization code will deadlock! 932 933 void Monitor::lock_without_safepoint_check(Thread * Self) { 934 // Ensure that the Monitor does not require or allow safepoint checks. 935 assert(_safepoint_check_required != Monitor::_safepoint_check_always, 936 "This lock should always have a safepoint check: %s", name()); 937 assert(_owner != Self, "invariant"); 938 ILock(Self); 939 assert(_owner == NULL, "invariant"); 940 set_owner(Self); 941 } 942 943 void Monitor::lock_without_safepoint_check() { 944 lock_without_safepoint_check(Thread::current()); 945 } 946 947 948 // Returns true if thread succeeds in grabbing the lock, otherwise false. 949 950 bool Monitor::try_lock() { 951 Thread * const Self = Thread::current(); 952 debug_only(check_prelock_state(Self, false)); 953 // assert(!thread->is_inside_signal_handler(), "don't lock inside signal handler"); 954 955 // Special case, where all Java threads are stopped. 956 // The lock may have been acquired but _owner is not yet set. 957 // In that case the VM thread can safely grab the lock. 958 // It strikes me this should appear _after the TryLock() fails, below. 959 bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint(); 960 if (can_sneak && _owner == NULL) { 961 set_owner(Self); // Do not need to be atomic, since we are at a safepoint 962 _snuck = true; 963 return true; 964 } 965 966 if (TryLock()) { 967 // We got the lock 968 assert(_owner == NULL, "invariant"); 969 set_owner(Self); 970 return true; 971 } 972 return false; 973 } 974 975 void Monitor::unlock() { 976 assert(_owner == Thread::current(), "invariant"); 977 assert(_OnDeck != Thread::current()->_MutexEvent, "invariant"); 978 set_owner(NULL); 979 if (_snuck) { 980 assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak"); 981 _snuck = false; 982 return; 983 } 984 IUnlock(false); 985 } 986 987 // Yet another degenerate version of Monitor::lock() or lock_without_safepoint_check() 988 // jvm_raw_lock() and _unlock() can be called by non-Java threads via JVM_RawMonitorEnter. 989 // 990 // There's no expectation that JVM_RawMonitors will interoperate properly with the native 991 // Mutex-Monitor constructs. We happen to implement JVM_RawMonitors in terms of 992 // native Mutex-Monitors simply as a matter of convenience. A simple abstraction layer 993 // over a pthread_mutex_t would work equally as well, but require more platform-specific 994 // code -- a "PlatformMutex". Alternatively, a simply layer over muxAcquire-muxRelease 995 // would work too. 996 // 997 // Since the caller might be a foreign thread, we don't necessarily have a Thread.MutexEvent 998 // instance available. Instead, we transiently allocate a ParkEvent on-demand if 999 // we encounter contention. That ParkEvent remains associated with the thread 1000 // until it manages to acquire the lock, at which time we return the ParkEvent 1001 // to the global ParkEvent free list. This is correct and suffices for our purposes. 1002 // 1003 // Beware that the original jvm_raw_unlock() had a "_snuck" test but that 1004 // jvm_raw_lock() didn't have the corresponding test. I suspect that's an 1005 // oversight, but I've replicated the original suspect logic in the new code ... 1006 1007 void Monitor::jvm_raw_lock() { 1008 assert(rank() == native, "invariant"); 1009 1010 if (TryLock()) { 1011 Exeunt: 1012 assert(ILocked(), "invariant"); 1013 assert(_owner == NULL, "invariant"); 1014 // This can potentially be called by non-java Threads. Thus, the Thread::current_or_null() 1015 // might return NULL. Don't call set_owner since it will break on an NULL owner 1016 // Consider installing a non-null "ANON" distinguished value instead of just NULL. 1017 _owner = Thread::current_or_null(); 1018 return; 1019 } 1020 1021 if (TrySpin(NULL)) goto Exeunt; 1022 1023 // slow-path - apparent contention 1024 // Allocate a ParkEvent for transient use. 1025 // The ParkEvent remains associated with this thread until 1026 // the time the thread manages to acquire the lock. 1027 ParkEvent * const ESelf = ParkEvent::Allocate(NULL); 1028 ESelf->reset(); 1029 OrderAccess::storeload(); 1030 1031 // Either Enqueue Self on cxq or acquire the outer lock. 1032 if (AcquireOrPush (ESelf)) { 1033 ParkEvent::Release(ESelf); // surrender the ParkEvent 1034 goto Exeunt; 1035 } 1036 1037 // At any given time there is at most one ondeck thread. 1038 // ondeck implies not resident on cxq and not resident on EntryList 1039 // Only the OnDeck thread can try to acquire -- contend for -- the lock. 1040 // CONSIDER: use Self->OnDeck instead of m->OnDeck. 1041 for (;;) { 1042 if (OrderAccess::load_acquire(&_OnDeck) == ESelf && TrySpin(NULL)) break; 1043 ParkCommon(ESelf, 0); 1044 } 1045 1046 assert(_OnDeck == ESelf, "invariant"); 1047 _OnDeck = NULL; 1048 ParkEvent::Release(ESelf); // surrender the ParkEvent 1049 goto Exeunt; 1050 } 1051 1052 void Monitor::jvm_raw_unlock() { 1053 // Nearly the same as Monitor::unlock() ... 1054 // directly set _owner instead of using set_owner(null) 1055 _owner = NULL; 1056 if (_snuck) { // ??? 1057 assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak"); 1058 _snuck = false; 1059 return; 1060 } 1061 IUnlock(false); 1062 } 1063 1064 bool Monitor::wait(bool no_safepoint_check, long timeout, 1065 bool as_suspend_equivalent) { 1066 // Make sure safepoint checking is used properly. 1067 assert(!(_safepoint_check_required == Monitor::_safepoint_check_never && no_safepoint_check == false), 1068 "This lock should never have a safepoint check: %s", name()); 1069 assert(!(_safepoint_check_required == Monitor::_safepoint_check_always && no_safepoint_check == true), 1070 "This lock should always have a safepoint check: %s", name()); 1071 1072 Thread * const Self = Thread::current(); 1073 assert(_owner == Self, "invariant"); 1074 assert(ILocked(), "invariant"); 1075 1076 // as_suspend_equivalent logically implies !no_safepoint_check 1077 guarantee(!as_suspend_equivalent || !no_safepoint_check, "invariant"); 1078 // !no_safepoint_check logically implies java_thread 1079 guarantee(no_safepoint_check || Self->is_Java_thread(), "invariant"); 1080 1081 #ifdef ASSERT 1082 Monitor * least = get_least_ranked_lock_besides_this(Self->owned_locks()); 1083 assert(least != this, "Specification of get_least_... call above"); 1084 if (least != NULL && least->rank() <= special) { 1085 tty->print("Attempting to wait on monitor %s/%d while holding" 1086 " lock %s/%d -- possible deadlock", 1087 name(), rank(), least->name(), least->rank()); 1088 assert(false, "Shouldn't block(wait) while holding a lock of rank special"); 1089 } 1090 #endif // ASSERT 1091 1092 int wait_status; 1093 // conceptually set the owner to NULL in anticipation of 1094 // abdicating the lock in wait 1095 set_owner(NULL); 1096 if (no_safepoint_check) { 1097 wait_status = IWait(Self, timeout); 1098 } else { 1099 assert(Self->is_Java_thread(), "invariant"); 1100 JavaThread *jt = (JavaThread *)Self; 1101 1102 // Enter safepoint region - ornate and Rococo ... 1103 ThreadBlockInVM tbivm(jt); 1104 OSThreadWaitState osts(Self->osthread(), false /* not Object.wait() */); 1105 1106 if (as_suspend_equivalent) { 1107 jt->set_suspend_equivalent(); 1108 // cleared by handle_special_suspend_equivalent_condition() or 1109 // java_suspend_self() 1110 } 1111 1112 wait_status = IWait(Self, timeout); 1113 1114 // were we externally suspended while we were waiting? 1115 if (as_suspend_equivalent && jt->handle_special_suspend_equivalent_condition()) { 1116 // Our event wait has finished and we own the lock, but 1117 // while we were waiting another thread suspended us. We don't 1118 // want to hold the lock while suspended because that 1119 // would surprise the thread that suspended us. 1120 assert(ILocked(), "invariant"); 1121 IUnlock(true); 1122 jt->java_suspend_self(); 1123 ILock(Self); 1124 assert(ILocked(), "invariant"); 1125 } 1126 } 1127 1128 // Conceptually reestablish ownership of the lock. 1129 // The "real" lock -- the LockByte -- was reacquired by IWait(). 1130 assert(ILocked(), "invariant"); 1131 assert(_owner == NULL, "invariant"); 1132 set_owner(Self); 1133 return wait_status != 0; // return true IFF timeout 1134 } 1135 1136 Monitor::~Monitor() { 1137 #ifdef ASSERT 1138 uintptr_t owner = UNS(_owner); 1139 uintptr_t lockword = UNS(_LockWord.FullWord); 1140 uintptr_t entrylist = UNS(_EntryList); 1141 uintptr_t waitset = UNS(_WaitSet); 1142 uintptr_t ondeck = UNS(_OnDeck); 1143 // Print _name with precision limit, in case failure is due to memory 1144 // corruption that also trashed _name. 1145 assert((owner|lockword|entrylist|waitset|ondeck) == 0, 1146 "%.*s: _owner(" INTPTR_FORMAT ")|_LockWord(" INTPTR_FORMAT ")|_EntryList(" INTPTR_FORMAT ")|_WaitSet(" 1147 INTPTR_FORMAT ")|_OnDeck(" INTPTR_FORMAT ") != 0", 1148 MONITOR_NAME_LEN, _name, owner, lockword, entrylist, waitset, ondeck); 1149 #endif 1150 } 1151 1152 void Monitor::ClearMonitor(Monitor * m, const char *name) { 1153 m->_owner = NULL; 1154 m->_snuck = false; 1155 if (name == NULL) { 1156 strcpy(m->_name, "UNKNOWN"); 1157 } else { 1158 strncpy(m->_name, name, MONITOR_NAME_LEN - 1); 1159 m->_name[MONITOR_NAME_LEN - 1] = '\0'; 1160 } 1161 m->_LockWord.FullWord = 0; 1162 m->_EntryList = NULL; 1163 m->_OnDeck = NULL; 1164 m->_WaitSet = NULL; 1165 m->_WaitLock[0] = 0; 1166 } 1167 1168 Monitor::Monitor() { ClearMonitor(this); } 1169 1170 Monitor::Monitor(int Rank, const char * name, bool allow_vm_block, 1171 SafepointCheckRequired safepoint_check_required) { 1172 ClearMonitor(this, name); 1173 #ifdef ASSERT 1174 _allow_vm_block = allow_vm_block; 1175 _rank = Rank; 1176 NOT_PRODUCT(_safepoint_check_required = safepoint_check_required;) 1177 #endif 1178 } 1179 1180 Mutex::Mutex(int Rank, const char * name, bool allow_vm_block, 1181 SafepointCheckRequired safepoint_check_required) { 1182 ClearMonitor((Monitor *) this, name); 1183 #ifdef ASSERT 1184 _allow_vm_block = allow_vm_block; 1185 _rank = Rank; 1186 NOT_PRODUCT(_safepoint_check_required = safepoint_check_required;) 1187 #endif 1188 } 1189 1190 bool Monitor::owned_by_self() const { 1191 bool ret = _owner == Thread::current(); 1192 assert(!ret || _LockWord.Bytes[_LSBINDEX] != 0, "invariant"); 1193 return ret; 1194 } 1195 1196 void Monitor::print_on_error(outputStream* st) const { 1197 st->print("[" PTR_FORMAT, p2i(this)); 1198 st->print("] %s", _name); 1199 st->print(" - owner thread: " PTR_FORMAT, p2i(_owner)); 1200 } 1201 1202 1203 1204 1205 // ---------------------------------------------------------------------------------- 1206 // Non-product code 1207 1208 #ifndef PRODUCT 1209 void Monitor::print_on(outputStream* st) const { 1210 st->print_cr("Mutex: [" PTR_FORMAT "/" PTR_FORMAT "] %s - owner: " PTR_FORMAT, 1211 p2i(this), _LockWord.FullWord, _name, p2i(_owner)); 1212 } 1213 #endif 1214 1215 #ifndef PRODUCT 1216 #ifdef ASSERT 1217 Monitor * Monitor::get_least_ranked_lock(Monitor * locks) { 1218 Monitor *res, *tmp; 1219 for (res = tmp = locks; tmp != NULL; tmp = tmp->next()) { 1220 if (tmp->rank() < res->rank()) { 1221 res = tmp; 1222 } 1223 } 1224 if (!SafepointSynchronize::is_at_safepoint()) { 1225 // In this case, we expect the held locks to be 1226 // in increasing rank order (modulo any native ranks) 1227 for (tmp = locks; tmp != NULL; tmp = tmp->next()) { 1228 if (tmp->next() != NULL) { 1229 assert(tmp->rank() == Mutex::native || 1230 tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?"); 1231 } 1232 } 1233 } 1234 return res; 1235 } 1236 1237 Monitor* Monitor::get_least_ranked_lock_besides_this(Monitor* locks) { 1238 Monitor *res, *tmp; 1239 for (res = NULL, tmp = locks; tmp != NULL; tmp = tmp->next()) { 1240 if (tmp != this && (res == NULL || 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 1258 bool Monitor::contains(Monitor* locks, Monitor * lock) { 1259 for (; locks != NULL; locks = locks->next()) { 1260 if (locks == lock) { 1261 return true; 1262 } 1263 } 1264 return false; 1265 } 1266 #endif 1267 1268 // Called immediately after lock acquisition or release as a diagnostic 1269 // to track the lock-set of the thread and test for rank violations that 1270 // might indicate exposure to deadlock. 1271 // Rather like an EventListener for _owner (:>). 1272 1273 void Monitor::set_owner_implementation(Thread *new_owner) { 1274 // This function is solely responsible for maintaining 1275 // and checking the invariant that threads and locks 1276 // are in a 1/N relation, with some some locks unowned. 1277 // It uses the Mutex::_owner, Mutex::_next, and 1278 // Thread::_owned_locks fields, and no other function 1279 // changes those fields. 1280 // It is illegal to set the mutex from one non-NULL 1281 // owner to another--it must be owned by NULL as an 1282 // intermediate state. 1283 1284 if (new_owner != NULL) { 1285 // the thread is acquiring this lock 1286 1287 assert(new_owner == Thread::current(), "Should I be doing this?"); 1288 assert(_owner == NULL, "setting the owner thread of an already owned mutex"); 1289 _owner = new_owner; // set the owner 1290 1291 // link "this" into the owned locks list 1292 1293 #ifdef ASSERT // Thread::_owned_locks is under the same ifdef 1294 Monitor* locks = get_least_ranked_lock(new_owner->owned_locks()); 1295 // Mutex::set_owner_implementation is a friend of Thread 1296 1297 assert(this->rank() >= 0, "bad lock rank"); 1298 1299 // Deadlock avoidance rules require us to acquire Mutexes only in 1300 // a global total order. For example m1 is the lowest ranked mutex 1301 // that the thread holds and m2 is the mutex the thread is trying 1302 // to acquire, then deadlock avoidance rules require that the rank 1303 // of m2 be less than the rank of m1. 1304 // The rank Mutex::native is an exception in that it is not subject 1305 // to the verification rules. 1306 // Here are some further notes relating to mutex acquisition anomalies: 1307 // . it is also ok to acquire Safepoint_lock at the very end while we 1308 // already hold Terminator_lock - may happen because of periodic safepoints 1309 if (this->rank() != Mutex::native && 1310 this->rank() != Mutex::suspend_resume && 1311 locks != NULL && locks->rank() <= this->rank() && 1312 !SafepointSynchronize::is_at_safepoint() && 1313 !(this == Safepoint_lock && contains(locks, Terminator_lock) && 1314 SafepointSynchronize::is_synchronizing())) { 1315 new_owner->print_owned_locks(); 1316 fatal("acquiring lock %s/%d out of order with lock %s/%d -- " 1317 "possible deadlock", this->name(), this->rank(), 1318 locks->name(), locks->rank()); 1319 } 1320 1321 this->_next = new_owner->_owned_locks; 1322 new_owner->_owned_locks = this; 1323 #endif 1324 1325 } else { 1326 // the thread is releasing this lock 1327 1328 Thread* old_owner = _owner; 1329 debug_only(_last_owner = old_owner); 1330 1331 assert(old_owner != NULL, "removing the owner thread of an unowned mutex"); 1332 assert(old_owner == Thread::current(), "removing the owner thread of an unowned mutex"); 1333 1334 _owner = NULL; // set the owner 1335 1336 #ifdef ASSERT 1337 Monitor *locks = old_owner->owned_locks(); 1338 1339 // remove "this" from the owned locks list 1340 1341 Monitor *prev = NULL; 1342 bool found = false; 1343 for (; locks != NULL; prev = locks, locks = locks->next()) { 1344 if (locks == this) { 1345 found = true; 1346 break; 1347 } 1348 } 1349 assert(found, "Removing a lock not owned"); 1350 if (prev == NULL) { 1351 old_owner->_owned_locks = _next; 1352 } else { 1353 prev->_next = _next; 1354 } 1355 _next = NULL; 1356 #endif 1357 } 1358 } 1359 1360 1361 // Factored out common sanity checks for locking mutex'es. Used by lock() and try_lock() 1362 void Monitor::check_prelock_state(Thread *thread, bool safepoint_check) { 1363 if (safepoint_check) { 1364 assert((!thread->is_Java_thread() || ((JavaThread *)thread)->thread_state() == _thread_in_vm) 1365 || rank() == Mutex::special, "wrong thread state for using locks"); 1366 if (thread->is_VM_thread() && !allow_vm_block()) { 1367 fatal("VM thread using lock %s (not allowed to block on)", name()); 1368 } 1369 debug_only(if (rank() != Mutex::special) \ 1370 thread->check_for_valid_safepoint_state(false);) 1371 } 1372 assert(!os::ThreadCrashProtection::is_crash_protected(thread), 1373 "locking not allowed when crash protection is set"); 1374 } 1375 1376 void Monitor::check_block_state(Thread *thread) { 1377 if (!_allow_vm_block && thread->is_VM_thread()) { 1378 warning("VM thread blocked on lock"); 1379 print(); 1380 BREAKPOINT; 1381 } 1382 assert(_owner != thread, "deadlock: blocking on monitor owned by current thread"); 1383 } 1384 1385 #endif // PRODUCT