1.1 --- /dev/null Thu Jan 01 00:00:00 1970 +0000
1.2 +++ b/src/share/vm/runtime/mutex.cpp Wed Apr 27 01:25:04 2016 +0800
1.3 @@ -0,0 +1,1390 @@
1.4 +
1.5 +/*
1.6 + * Copyright (c) 1998, 2014, Oracle and/or its affiliates. All rights reserved.
1.7 + * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
1.8 + *
1.9 + * This code is free software; you can redistribute it and/or modify it
1.10 + * under the terms of the GNU General Public License version 2 only, as
1.11 + * published by the Free Software Foundation.
1.12 + *
1.13 + * This code is distributed in the hope that it will be useful, but WITHOUT
1.14 + * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
1.15 + * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
1.16 + * version 2 for more details (a copy is included in the LICENSE file that
1.17 + * accompanied this code).
1.18 + *
1.19 + * You should have received a copy of the GNU General Public License version
1.20 + * 2 along with this work; if not, write to the Free Software Foundation,
1.21 + * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
1.22 + *
1.23 + * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
1.24 + * or visit www.oracle.com if you need additional information or have any
1.25 + * questions.
1.26 + *
1.27 + */
1.28 +
1.29 +#include "precompiled.hpp"
1.30 +#include "runtime/mutex.hpp"
1.31 +#include "runtime/osThread.hpp"
1.32 +#include "runtime/thread.inline.hpp"
1.33 +#include "utilities/events.hpp"
1.34 +#ifdef TARGET_OS_FAMILY_linux
1.35 +# include "mutex_linux.inline.hpp"
1.36 +#endif
1.37 +#ifdef TARGET_OS_FAMILY_solaris
1.38 +# include "mutex_solaris.inline.hpp"
1.39 +#endif
1.40 +#ifdef TARGET_OS_FAMILY_windows
1.41 +# include "mutex_windows.inline.hpp"
1.42 +#endif
1.43 +#ifdef TARGET_OS_FAMILY_bsd
1.44 +# include "mutex_bsd.inline.hpp"
1.45 +#endif
1.46 +
1.47 +PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
1.48 +
1.49 +// 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
1.50 +//
1.51 +// Native Monitor-Mutex locking - theory of operations
1.52 +//
1.53 +// * Native Monitors are completely unrelated to Java-level monitors,
1.54 +// although the "back-end" slow-path implementations share a common lineage.
1.55 +// See objectMonitor:: in synchronizer.cpp.
1.56 +// Native Monitors do *not* support nesting or recursion but otherwise
1.57 +// they're basically Hoare-flavor monitors.
1.58 +//
1.59 +// * A thread acquires ownership of a Monitor/Mutex by CASing the LockByte
1.60 +// in the _LockWord from zero to non-zero. Note that the _Owner field
1.61 +// is advisory and is used only to verify that the thread calling unlock()
1.62 +// is indeed the last thread to have acquired the lock.
1.63 +//
1.64 +// * Contending threads "push" themselves onto the front of the contention
1.65 +// queue -- called the cxq -- with CAS and then spin/park.
1.66 +// The _LockWord contains the LockByte as well as the pointer to the head
1.67 +// of the cxq. Colocating the LockByte with the cxq precludes certain races.
1.68 +//
1.69 +// * Using a separately addressable LockByte allows for CAS:MEMBAR or CAS:0
1.70 +// idioms. We currently use MEMBAR in the uncontended unlock() path, as
1.71 +// MEMBAR often has less latency than CAS. If warranted, we could switch to
1.72 +// a CAS:0 mode, using timers to close the resultant race, as is done
1.73 +// with Java Monitors in synchronizer.cpp.
1.74 +//
1.75 +// See the following for a discussion of the relative cost of atomics (CAS)
1.76 +// MEMBAR, and ways to eliminate such instructions from the common-case paths:
1.77 +// -- http://blogs.sun.com/dave/entry/biased_locking_in_hotspot
1.78 +// -- http://blogs.sun.com/dave/resource/MustangSync.pdf
1.79 +// -- http://blogs.sun.com/dave/resource/synchronization-public2.pdf
1.80 +// -- synchronizer.cpp
1.81 +//
1.82 +// * Overall goals - desiderata
1.83 +// 1. Minimize context switching
1.84 +// 2. Minimize lock migration
1.85 +// 3. Minimize CPI -- affinity and locality
1.86 +// 4. Minimize the execution of high-latency instructions such as CAS or MEMBAR
1.87 +// 5. Minimize outer lock hold times
1.88 +// 6. Behave gracefully on a loaded system
1.89 +//
1.90 +// * Thread flow and list residency:
1.91 +//
1.92 +// Contention queue --> EntryList --> OnDeck --> Owner --> !Owner
1.93 +// [..resident on monitor list..]
1.94 +// [...........contending..................]
1.95 +//
1.96 +// -- The contention queue (cxq) contains recently-arrived threads (RATs).
1.97 +// Threads on the cxq eventually drain into the EntryList.
1.98 +// -- Invariant: a thread appears on at most one list -- cxq, EntryList
1.99 +// or WaitSet -- at any one time.
1.100 +// -- For a given monitor there can be at most one "OnDeck" thread at any
1.101 +// given time but if needbe this particular invariant could be relaxed.
1.102 +//
1.103 +// * The WaitSet and EntryList linked lists are composed of ParkEvents.
1.104 +// I use ParkEvent instead of threads as ParkEvents are immortal and
1.105 +// type-stable, meaning we can safely unpark() a possibly stale
1.106 +// list element in the unlock()-path. (That's benign).
1.107 +//
1.108 +// * Succession policy - providing for progress:
1.109 +//
1.110 +// As necessary, the unlock()ing thread identifies, unlinks, and unparks
1.111 +// an "heir presumptive" tentative successor thread from the EntryList.
1.112 +// This becomes the so-called "OnDeck" thread, of which there can be only
1.113 +// one at any given time for a given monitor. The wakee will recontend
1.114 +// for ownership of monitor.
1.115 +//
1.116 +// Succession is provided for by a policy of competitive handoff.
1.117 +// The exiting thread does _not_ grant or pass ownership to the
1.118 +// successor thread. (This is also referred to as "handoff" succession").
1.119 +// Instead the exiting thread releases ownership and possibly wakes
1.120 +// a successor, so the successor can (re)compete for ownership of the lock.
1.121 +//
1.122 +// Competitive handoff provides excellent overall throughput at the expense
1.123 +// of short-term fairness. If fairness is a concern then one remedy might
1.124 +// be to add an AcquireCounter field to the monitor. After a thread acquires
1.125 +// the lock it will decrement the AcquireCounter field. When the count
1.126 +// reaches 0 the thread would reset the AcquireCounter variable, abdicate
1.127 +// the lock directly to some thread on the EntryList, and then move itself to the
1.128 +// tail of the EntryList.
1.129 +//
1.130 +// But in practice most threads engage or otherwise participate in resource
1.131 +// bounded producer-consumer relationships, so lock domination is not usually
1.132 +// a practical concern. Recall too, that in general it's easier to construct
1.133 +// a fair lock from a fast lock, but not vice-versa.
1.134 +//
1.135 +// * The cxq can have multiple concurrent "pushers" but only one concurrent
1.136 +// detaching thread. This mechanism is immune from the ABA corruption.
1.137 +// More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
1.138 +// We use OnDeck as a pseudo-lock to enforce the at-most-one detaching
1.139 +// thread constraint.
1.140 +//
1.141 +// * Taken together, the cxq and the EntryList constitute or form a
1.142 +// single logical queue of threads stalled trying to acquire the lock.
1.143 +// We use two distinct lists to reduce heat on the list ends.
1.144 +// Threads in lock() enqueue onto cxq while threads in unlock() will
1.145 +// dequeue from the EntryList. (c.f. Michael Scott's "2Q" algorithm).
1.146 +// A key desideratum is to minimize queue & monitor metadata manipulation
1.147 +// that occurs while holding the "outer" monitor lock -- that is, we want to
1.148 +// minimize monitor lock holds times.
1.149 +//
1.150 +// The EntryList is ordered by the prevailing queue discipline and
1.151 +// can be organized in any convenient fashion, such as a doubly-linked list or
1.152 +// a circular doubly-linked list. If we need a priority queue then something akin
1.153 +// to Solaris' sleepq would work nicely. Viz.,
1.154 +// -- http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
1.155 +// -- http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/uts/common/os/sleepq.c
1.156 +// Queue discipline is enforced at ::unlock() time, when the unlocking thread
1.157 +// drains the cxq into the EntryList, and orders or reorders the threads on the
1.158 +// EntryList accordingly.
1.159 +//
1.160 +// Barring "lock barging", this mechanism provides fair cyclic ordering,
1.161 +// somewhat similar to an elevator-scan.
1.162 +//
1.163 +// * OnDeck
1.164 +// -- For a given monitor there can be at most one OnDeck thread at any given
1.165 +// instant. The OnDeck thread is contending for the lock, but has been
1.166 +// unlinked from the EntryList and cxq by some previous unlock() operations.
1.167 +// Once a thread has been designated the OnDeck thread it will remain so
1.168 +// until it manages to acquire the lock -- being OnDeck is a stable property.
1.169 +// -- Threads on the EntryList or cxq are _not allowed to attempt lock acquisition.
1.170 +// -- OnDeck also serves as an "inner lock" as follows. Threads in unlock() will, after
1.171 +// having cleared the LockByte and dropped the outer lock, attempt to "trylock"
1.172 +// OnDeck by CASing the field from null to non-null. If successful, that thread
1.173 +// is then responsible for progress and succession and can use CAS to detach and
1.174 +// drain the cxq into the EntryList. By convention, only this thread, the holder of
1.175 +// the OnDeck inner lock, can manipulate the EntryList or detach and drain the
1.176 +// RATs on the cxq into the EntryList. This avoids ABA corruption on the cxq as
1.177 +// we allow multiple concurrent "push" operations but restrict detach concurrency
1.178 +// to at most one thread. Having selected and detached a successor, the thread then
1.179 +// changes the OnDeck to refer to that successor, and then unparks the successor.
1.180 +// That successor will eventually acquire the lock and clear OnDeck. Beware
1.181 +// that the OnDeck usage as a lock is asymmetric. A thread in unlock() transiently
1.182 +// "acquires" OnDeck, performs queue manipulations, passes OnDeck to some successor,
1.183 +// and then the successor eventually "drops" OnDeck. Note that there's never
1.184 +// any sense of contention on the inner lock, however. Threads never contend
1.185 +// or wait for the inner lock.
1.186 +// -- OnDeck provides for futile wakeup throttling a described in section 3.3 of
1.187 +// See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1.188 +// In a sense, OnDeck subsumes the ObjectMonitor _Succ and ObjectWaiter
1.189 +// TState fields found in Java-level objectMonitors. (See synchronizer.cpp).
1.190 +//
1.191 +// * Waiting threads reside on the WaitSet list -- wait() puts
1.192 +// the caller onto the WaitSet. Notify() or notifyAll() simply
1.193 +// transfers threads from the WaitSet to either the EntryList or cxq.
1.194 +// Subsequent unlock() operations will eventually unpark the notifyee.
1.195 +// Unparking a notifee in notify() proper is inefficient - if we were to do so
1.196 +// it's likely the notifyee would simply impale itself on the lock held
1.197 +// by the notifier.
1.198 +//
1.199 +// * The mechanism is obstruction-free in that if the holder of the transient
1.200 +// OnDeck lock in unlock() is preempted or otherwise stalls, other threads
1.201 +// can still acquire and release the outer lock and continue to make progress.
1.202 +// At worst, waking of already blocked contending threads may be delayed,
1.203 +// but nothing worse. (We only use "trylock" operations on the inner OnDeck
1.204 +// lock).
1.205 +//
1.206 +// * Note that thread-local storage must be initialized before a thread
1.207 +// uses Native monitors or mutexes. The native monitor-mutex subsystem
1.208 +// depends on Thread::current().
1.209 +//
1.210 +// * The monitor synchronization subsystem avoids the use of native
1.211 +// synchronization primitives except for the narrow platform-specific
1.212 +// park-unpark abstraction. See the comments in os_solaris.cpp regarding
1.213 +// the semantics of park-unpark. Put another way, this monitor implementation
1.214 +// depends only on atomic operations and park-unpark. The monitor subsystem
1.215 +// manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
1.216 +// underlying OS manages the READY<->RUN transitions.
1.217 +//
1.218 +// * The memory consistency model provide by lock()-unlock() is at least as
1.219 +// strong or stronger than the Java Memory model defined by JSR-133.
1.220 +// That is, we guarantee at least entry consistency, if not stronger.
1.221 +// See http://g.oswego.edu/dl/jmm/cookbook.html.
1.222 +//
1.223 +// * Thread:: currently contains a set of purpose-specific ParkEvents:
1.224 +// _MutexEvent, _ParkEvent, etc. A better approach might be to do away with
1.225 +// the purpose-specific ParkEvents and instead implement a general per-thread
1.226 +// stack of available ParkEvents which we could provision on-demand. The
1.227 +// stack acts as a local cache to avoid excessive calls to ParkEvent::Allocate()
1.228 +// and ::Release(). A thread would simply pop an element from the local stack before it
1.229 +// enqueued or park()ed. When the contention was over the thread would
1.230 +// push the no-longer-needed ParkEvent back onto its stack.
1.231 +//
1.232 +// * A slightly reduced form of ILock() and IUnlock() have been partially
1.233 +// model-checked (Murphi) for safety and progress at T=1,2,3 and 4.
1.234 +// It'd be interesting to see if TLA/TLC could be useful as well.
1.235 +//
1.236 +// * Mutex-Monitor is a low-level "leaf" subsystem. That is, the monitor
1.237 +// code should never call other code in the JVM that might itself need to
1.238 +// acquire monitors or mutexes. That's true *except* in the case of the
1.239 +// ThreadBlockInVM state transition wrappers. The ThreadBlockInVM DTOR handles
1.240 +// mutator reentry (ingress) by checking for a pending safepoint in which case it will
1.241 +// call SafepointSynchronize::block(), which in turn may call Safepoint_lock->lock(), etc.
1.242 +// In that particular case a call to lock() for a given Monitor can end up recursively
1.243 +// calling lock() on another monitor. While distasteful, this is largely benign
1.244 +// as the calls come from jacket that wraps lock(), and not from deep within lock() itself.
1.245 +//
1.246 +// It's unfortunate that native mutexes and thread state transitions were convolved.
1.247 +// They're really separate concerns and should have remained that way. Melding
1.248 +// them together was facile -- a bit too facile. The current implementation badly
1.249 +// conflates the two concerns.
1.250 +//
1.251 +// * TODO-FIXME:
1.252 +//
1.253 +// -- Add DTRACE probes for contended acquire, contended acquired, contended unlock
1.254 +// We should also add DTRACE probes in the ParkEvent subsystem for
1.255 +// Park-entry, Park-exit, and Unpark.
1.256 +//
1.257 +// -- We have an excess of mutex-like constructs in the JVM, namely:
1.258 +// 1. objectMonitors for Java-level synchronization (synchronizer.cpp)
1.259 +// 2. low-level muxAcquire and muxRelease
1.260 +// 3. low-level spinAcquire and spinRelease
1.261 +// 4. native Mutex:: and Monitor::
1.262 +// 5. jvm_raw_lock() and _unlock()
1.263 +// 6. JVMTI raw monitors -- distinct from (5) despite having a confusingly
1.264 +// similar name.
1.265 +//
1.266 +// 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
1.267 +
1.268 +
1.269 +// CASPTR() uses the canonical argument order that dominates in the literature.
1.270 +// Our internal cmpxchg_ptr() uses a bastardized ordering to accommodate Sun .il templates.
1.271 +
1.272 +#define CASPTR(a,c,s) intptr_t(Atomic::cmpxchg_ptr ((void *)(s),(void *)(a),(void *)(c)))
1.273 +#define UNS(x) (uintptr_t(x))
1.274 +#define TRACE(m) { static volatile int ctr = 0 ; int x = ++ctr ; if ((x & (x-1))==0) { ::printf ("%d:%s\n", x, #m); ::fflush(stdout); }}
1.275 +
1.276 +// Simplistic low-quality Marsaglia SHIFT-XOR RNG.
1.277 +// Bijective except for the trailing mask operation.
1.278 +// Useful for spin loops as the compiler can't optimize it away.
1.279 +
1.280 +static inline jint MarsagliaXORV (jint x) {
1.281 + if (x == 0) x = 1|os::random() ;
1.282 + x ^= x << 6;
1.283 + x ^= ((unsigned)x) >> 21;
1.284 + x ^= x << 7 ;
1.285 + return x & 0x7FFFFFFF ;
1.286 +}
1.287 +
1.288 +static inline jint MarsagliaXOR (jint * const a) {
1.289 + jint x = *a ;
1.290 + if (x == 0) x = UNS(a)|1 ;
1.291 + x ^= x << 6;
1.292 + x ^= ((unsigned)x) >> 21;
1.293 + x ^= x << 7 ;
1.294 + *a = x ;
1.295 + return x & 0x7FFFFFFF ;
1.296 +}
1.297 +
1.298 +static int Stall (int its) {
1.299 + static volatile jint rv = 1 ;
1.300 + volatile int OnFrame = 0 ;
1.301 + jint v = rv ^ UNS(OnFrame) ;
1.302 + while (--its >= 0) {
1.303 + v = MarsagliaXORV (v) ;
1.304 + }
1.305 + // Make this impossible for the compiler to optimize away,
1.306 + // but (mostly) avoid W coherency sharing on MP systems.
1.307 + if (v == 0x12345) rv = v ;
1.308 + return v ;
1.309 +}
1.310 +
1.311 +int Monitor::TryLock () {
1.312 + intptr_t v = _LockWord.FullWord ;
1.313 + for (;;) {
1.314 + if ((v & _LBIT) != 0) return 0 ;
1.315 + const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ;
1.316 + if (v == u) return 1 ;
1.317 + v = u ;
1.318 + }
1.319 +}
1.320 +
1.321 +int Monitor::TryFast () {
1.322 + // Optimistic fast-path form ...
1.323 + // Fast-path attempt for the common uncontended case.
1.324 + // Avoid RTS->RTO $ coherence upgrade on typical SMP systems.
1.325 + intptr_t v = CASPTR (&_LockWord, 0, _LBIT) ; // agro ...
1.326 + if (v == 0) return 1 ;
1.327 +
1.328 + for (;;) {
1.329 + if ((v & _LBIT) != 0) return 0 ;
1.330 + const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ;
1.331 + if (v == u) return 1 ;
1.332 + v = u ;
1.333 + }
1.334 +}
1.335 +
1.336 +int Monitor::ILocked () {
1.337 + const intptr_t w = _LockWord.FullWord & 0xFF ;
1.338 + assert (w == 0 || w == _LBIT, "invariant") ;
1.339 + return w == _LBIT ;
1.340 +}
1.341 +
1.342 +// Polite TATAS spinlock with exponential backoff - bounded spin.
1.343 +// Ideally we'd use processor cycles, time or vtime to control
1.344 +// the loop, but we currently use iterations.
1.345 +// All the constants within were derived empirically but work over
1.346 +// over the spectrum of J2SE reference platforms.
1.347 +// On Niagara-class systems the back-off is unnecessary but
1.348 +// is relatively harmless. (At worst it'll slightly retard
1.349 +// acquisition times). The back-off is critical for older SMP systems
1.350 +// where constant fetching of the LockWord would otherwise impair
1.351 +// scalability.
1.352 +//
1.353 +// Clamp spinning at approximately 1/2 of a context-switch round-trip.
1.354 +// See synchronizer.cpp for details and rationale.
1.355 +
1.356 +int Monitor::TrySpin (Thread * const Self) {
1.357 + if (TryLock()) return 1 ;
1.358 + if (!os::is_MP()) return 0 ;
1.359 +
1.360 + int Probes = 0 ;
1.361 + int Delay = 0 ;
1.362 + int Steps = 0 ;
1.363 + int SpinMax = NativeMonitorSpinLimit ;
1.364 + int flgs = NativeMonitorFlags ;
1.365 + for (;;) {
1.366 + intptr_t v = _LockWord.FullWord;
1.367 + if ((v & _LBIT) == 0) {
1.368 + if (CASPTR (&_LockWord, v, v|_LBIT) == v) {
1.369 + return 1 ;
1.370 + }
1.371 + continue ;
1.372 + }
1.373 +
1.374 + if ((flgs & 8) == 0) {
1.375 + SpinPause () ;
1.376 + }
1.377 +
1.378 + // Periodically increase Delay -- variable Delay form
1.379 + // conceptually: delay *= 1 + 1/Exponent
1.380 + ++ Probes;
1.381 + if (Probes > SpinMax) return 0 ;
1.382 +
1.383 + if ((Probes & 0x7) == 0) {
1.384 + Delay = ((Delay << 1)|1) & 0x7FF ;
1.385 + // CONSIDER: Delay += 1 + (Delay/4); Delay &= 0x7FF ;
1.386 + }
1.387 +
1.388 + if (flgs & 2) continue ;
1.389 +
1.390 + // Consider checking _owner's schedctl state, if OFFPROC abort spin.
1.391 + // If the owner is OFFPROC then it's unlike that the lock will be dropped
1.392 + // in a timely fashion, which suggests that spinning would not be fruitful
1.393 + // or profitable.
1.394 +
1.395 + // Stall for "Delay" time units - iterations in the current implementation.
1.396 + // Avoid generating coherency traffic while stalled.
1.397 + // Possible ways to delay:
1.398 + // PAUSE, SLEEP, MEMBAR #sync, MEMBAR #halt,
1.399 + // wr %g0,%asi, gethrtime, rdstick, rdtick, rdtsc, etc. ...
1.400 + // Note that on Niagara-class systems we want to minimize STs in the
1.401 + // spin loop. N1 and brethren write-around the L1$ over the xbar into the L2$.
1.402 + // Furthermore, they don't have a W$ like traditional SPARC processors.
1.403 + // We currently use a Marsaglia Shift-Xor RNG loop.
1.404 + Steps += Delay ;
1.405 + if (Self != NULL) {
1.406 + jint rv = Self->rng[0] ;
1.407 + for (int k = Delay ; --k >= 0; ) {
1.408 + rv = MarsagliaXORV (rv) ;
1.409 + if ((flgs & 4) == 0 && SafepointSynchronize::do_call_back()) return 0 ;
1.410 + }
1.411 + Self->rng[0] = rv ;
1.412 + } else {
1.413 + Stall (Delay) ;
1.414 + }
1.415 + }
1.416 +}
1.417 +
1.418 +static int ParkCommon (ParkEvent * ev, jlong timo) {
1.419 + // Diagnostic support - periodically unwedge blocked threads
1.420 + intx nmt = NativeMonitorTimeout ;
1.421 + if (nmt > 0 && (nmt < timo || timo <= 0)) {
1.422 + timo = nmt ;
1.423 + }
1.424 + int err = OS_OK ;
1.425 + if (0 == timo) {
1.426 + ev->park() ;
1.427 + } else {
1.428 + err = ev->park(timo) ;
1.429 + }
1.430 + return err ;
1.431 +}
1.432 +
1.433 +inline int Monitor::AcquireOrPush (ParkEvent * ESelf) {
1.434 + intptr_t v = _LockWord.FullWord ;
1.435 + for (;;) {
1.436 + if ((v & _LBIT) == 0) {
1.437 + const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ;
1.438 + if (u == v) return 1 ; // indicate acquired
1.439 + v = u ;
1.440 + } else {
1.441 + // Anticipate success ...
1.442 + ESelf->ListNext = (ParkEvent *) (v & ~_LBIT) ;
1.443 + const intptr_t u = CASPTR (&_LockWord, v, intptr_t(ESelf)|_LBIT) ;
1.444 + if (u == v) return 0 ; // indicate pushed onto cxq
1.445 + v = u ;
1.446 + }
1.447 + // Interference - LockWord change - just retry
1.448 + }
1.449 +}
1.450 +
1.451 +// ILock and IWait are the lowest level primitive internal blocking
1.452 +// synchronization functions. The callers of IWait and ILock must have
1.453 +// performed any needed state transitions beforehand.
1.454 +// IWait and ILock may directly call park() without any concern for thread state.
1.455 +// Note that ILock and IWait do *not* access _owner.
1.456 +// _owner is a higher-level logical concept.
1.457 +
1.458 +void Monitor::ILock (Thread * Self) {
1.459 + assert (_OnDeck != Self->_MutexEvent, "invariant") ;
1.460 +
1.461 + if (TryFast()) {
1.462 + Exeunt:
1.463 + assert (ILocked(), "invariant") ;
1.464 + return ;
1.465 + }
1.466 +
1.467 + ParkEvent * const ESelf = Self->_MutexEvent ;
1.468 + assert (_OnDeck != ESelf, "invariant") ;
1.469 +
1.470 + // As an optimization, spinners could conditionally try to set ONDECK to _LBIT
1.471 + // Synchronizer.cpp uses a similar optimization.
1.472 + if (TrySpin (Self)) goto Exeunt ;
1.473 +
1.474 + // Slow-path - the lock is contended.
1.475 + // Either Enqueue Self on cxq or acquire the outer lock.
1.476 + // LockWord encoding = (cxq,LOCKBYTE)
1.477 + ESelf->reset() ;
1.478 + OrderAccess::fence() ;
1.479 +
1.480 + // Optional optimization ... try barging on the inner lock
1.481 + if ((NativeMonitorFlags & 32) && CASPTR (&_OnDeck, NULL, UNS(Self)) == 0) {
1.482 + goto OnDeck_LOOP ;
1.483 + }
1.484 +
1.485 + if (AcquireOrPush (ESelf)) goto Exeunt ;
1.486 +
1.487 + // At any given time there is at most one ondeck thread.
1.488 + // ondeck implies not resident on cxq and not resident on EntryList
1.489 + // Only the OnDeck thread can try to acquire -- contended for -- the lock.
1.490 + // CONSIDER: use Self->OnDeck instead of m->OnDeck.
1.491 + // Deschedule Self so that others may run.
1.492 + while (_OnDeck != ESelf) {
1.493 + ParkCommon (ESelf, 0) ;
1.494 + }
1.495 +
1.496 + // Self is now in the ONDECK position and will remain so until it
1.497 + // manages to acquire the lock.
1.498 + OnDeck_LOOP:
1.499 + for (;;) {
1.500 + assert (_OnDeck == ESelf, "invariant") ;
1.501 + if (TrySpin (Self)) break ;
1.502 + // CONSIDER: if ESelf->TryPark() && TryLock() break ...
1.503 + // It's probably wise to spin only if we *actually* blocked
1.504 + // CONSIDER: check the lockbyte, if it remains set then
1.505 + // preemptively drain the cxq into the EntryList.
1.506 + // The best place and time to perform queue operations -- lock metadata --
1.507 + // is _before having acquired the outer lock, while waiting for the lock to drop.
1.508 + ParkCommon (ESelf, 0) ;
1.509 + }
1.510 +
1.511 + assert (_OnDeck == ESelf, "invariant") ;
1.512 + _OnDeck = NULL ;
1.513 +
1.514 + // Note that we current drop the inner lock (clear OnDeck) in the slow-path
1.515 + // epilog immediately after having acquired the outer lock.
1.516 + // But instead we could consider the following optimizations:
1.517 + // A. Shift or defer dropping the inner lock until the subsequent IUnlock() operation.
1.518 + // This might avoid potential reacquisition of the inner lock in IUlock().
1.519 + // B. While still holding the inner lock, attempt to opportunistically select
1.520 + // and unlink the next ONDECK thread from the EntryList.
1.521 + // If successful, set ONDECK to refer to that thread, otherwise clear ONDECK.
1.522 + // It's critical that the select-and-unlink operation run in constant-time as
1.523 + // it executes when holding the outer lock and may artificially increase the
1.524 + // effective length of the critical section.
1.525 + // Note that (A) and (B) are tantamount to succession by direct handoff for
1.526 + // the inner lock.
1.527 + goto Exeunt ;
1.528 +}
1.529 +
1.530 +void Monitor::IUnlock (bool RelaxAssert) {
1.531 + assert (ILocked(), "invariant") ;
1.532 + // Conceptually we need a MEMBAR #storestore|#loadstore barrier or fence immediately
1.533 + // before the store that releases the lock. Crucially, all the stores and loads in the
1.534 + // critical section must be globally visible before the store of 0 into the lock-word
1.535 + // that releases the lock becomes globally visible. That is, memory accesses in the
1.536 + // critical section should not be allowed to bypass or overtake the following ST that
1.537 + // releases the lock. As such, to prevent accesses within the critical section
1.538 + // from "leaking" out, we need a release fence between the critical section and the
1.539 + // store that releases the lock. In practice that release barrier is elided on
1.540 + // platforms with strong memory models such as TSO.
1.541 + //
1.542 + // Note that the OrderAccess::storeload() fence that appears after unlock store
1.543 + // provides for progress conditions and succession and is _not related to exclusion
1.544 + // safety or lock release consistency.
1.545 + OrderAccess::release_store(&_LockWord.Bytes[_LSBINDEX], 0); // drop outer lock
1.546 +
1.547 + OrderAccess::storeload ();
1.548 + ParkEvent * const w = _OnDeck ;
1.549 + assert (RelaxAssert || w != Thread::current()->_MutexEvent, "invariant") ;
1.550 + if (w != NULL) {
1.551 + // Either we have a valid ondeck thread or ondeck is transiently "locked"
1.552 + // by some exiting thread as it arranges for succession. The LSBit of
1.553 + // OnDeck allows us to discriminate two cases. If the latter, the
1.554 + // responsibility for progress and succession lies with that other thread.
1.555 + // For good performance, we also depend on the fact that redundant unpark()
1.556 + // operations are cheap. That is, repeated Unpark()ing of the ONDECK thread
1.557 + // is inexpensive. This approach provides implicit futile wakeup throttling.
1.558 + // Note that the referent "w" might be stale with respect to the lock.
1.559 + // In that case the following unpark() is harmless and the worst that'll happen
1.560 + // is a spurious return from a park() operation. Critically, if "w" _is stale,
1.561 + // then progress is known to have occurred as that means the thread associated
1.562 + // with "w" acquired the lock. In that case this thread need take no further
1.563 + // action to guarantee progress.
1.564 + if ((UNS(w) & _LBIT) == 0) w->unpark() ;
1.565 + return ;
1.566 + }
1.567 +
1.568 + intptr_t cxq = _LockWord.FullWord ;
1.569 + if (((cxq & ~_LBIT)|UNS(_EntryList)) == 0) {
1.570 + return ; // normal fast-path exit - cxq and EntryList both empty
1.571 + }
1.572 + if (cxq & _LBIT) {
1.573 + // Optional optimization ...
1.574 + // Some other thread acquired the lock in the window since this
1.575 + // thread released it. Succession is now that thread's responsibility.
1.576 + return ;
1.577 + }
1.578 +
1.579 + Succession:
1.580 + // Slow-path exit - this thread must ensure succession and progress.
1.581 + // OnDeck serves as lock to protect cxq and EntryList.
1.582 + // Only the holder of OnDeck can manipulate EntryList or detach the RATs from cxq.
1.583 + // Avoid ABA - allow multiple concurrent producers (enqueue via push-CAS)
1.584 + // but only one concurrent consumer (detacher of RATs).
1.585 + // Consider protecting this critical section with schedctl on Solaris.
1.586 + // Unlike a normal lock, however, the exiting thread "locks" OnDeck,
1.587 + // picks a successor and marks that thread as OnDeck. That successor
1.588 + // thread will then clear OnDeck once it eventually acquires the outer lock.
1.589 + if (CASPTR (&_OnDeck, NULL, _LBIT) != UNS(NULL)) {
1.590 + return ;
1.591 + }
1.592 +
1.593 + ParkEvent * List = _EntryList ;
1.594 + if (List != NULL) {
1.595 + // Transfer the head of the EntryList to the OnDeck position.
1.596 + // Once OnDeck, a thread stays OnDeck until it acquires the lock.
1.597 + // For a given lock there is at most OnDeck thread at any one instant.
1.598 + WakeOne:
1.599 + assert (List == _EntryList, "invariant") ;
1.600 + ParkEvent * const w = List ;
1.601 + assert (RelaxAssert || w != Thread::current()->_MutexEvent, "invariant") ;
1.602 + _EntryList = w->ListNext ;
1.603 + // as a diagnostic measure consider setting w->_ListNext = BAD
1.604 + assert (UNS(_OnDeck) == _LBIT, "invariant") ;
1.605 + _OnDeck = w ; // pass OnDeck to w.
1.606 + // w will clear OnDeck once it acquires the outer lock
1.607 +
1.608 + // Another optional optimization ...
1.609 + // For heavily contended locks it's not uncommon that some other
1.610 + // thread acquired the lock while this thread was arranging succession.
1.611 + // Try to defer the unpark() operation - Delegate the responsibility
1.612 + // for unpark()ing the OnDeck thread to the current or subsequent owners
1.613 + // That is, the new owner is responsible for unparking the OnDeck thread.
1.614 + OrderAccess::storeload() ;
1.615 + cxq = _LockWord.FullWord ;
1.616 + if (cxq & _LBIT) return ;
1.617 +
1.618 + w->unpark() ;
1.619 + return ;
1.620 + }
1.621 +
1.622 + cxq = _LockWord.FullWord ;
1.623 + if ((cxq & ~_LBIT) != 0) {
1.624 + // The EntryList is empty but the cxq is populated.
1.625 + // drain RATs from cxq into EntryList
1.626 + // Detach RATs segment with CAS and then merge into EntryList
1.627 + for (;;) {
1.628 + // optional optimization - if locked, the owner is responsible for succession
1.629 + if (cxq & _LBIT) goto Punt ;
1.630 + const intptr_t vfy = CASPTR (&_LockWord, cxq, cxq & _LBIT) ;
1.631 + if (vfy == cxq) break ;
1.632 + cxq = vfy ;
1.633 + // Interference - LockWord changed - Just retry
1.634 + // We can see concurrent interference from contending threads
1.635 + // pushing themselves onto the cxq or from lock-unlock operations.
1.636 + // From the perspective of this thread, EntryList is stable and
1.637 + // the cxq is prepend-only -- the head is volatile but the interior
1.638 + // of the cxq is stable. In theory if we encounter interference from threads
1.639 + // pushing onto cxq we could simply break off the original cxq suffix and
1.640 + // move that segment to the EntryList, avoiding a 2nd or multiple CAS attempts
1.641 + // on the high-traffic LockWord variable. For instance lets say the cxq is "ABCD"
1.642 + // when we first fetch cxq above. Between the fetch -- where we observed "A"
1.643 + // -- and CAS -- where we attempt to CAS null over A -- "PQR" arrive,
1.644 + // yielding cxq = "PQRABCD". In this case we could simply set A.ListNext
1.645 + // null, leaving cxq = "PQRA" and transfer the "BCD" segment to the EntryList.
1.646 + // Note too, that it's safe for this thread to traverse the cxq
1.647 + // without taking any special concurrency precautions.
1.648 + }
1.649 +
1.650 + // We don't currently reorder the cxq segment as we move it onto
1.651 + // the EntryList, but it might make sense to reverse the order
1.652 + // or perhaps sort by thread priority. See the comments in
1.653 + // synchronizer.cpp objectMonitor::exit().
1.654 + assert (_EntryList == NULL, "invariant") ;
1.655 + _EntryList = List = (ParkEvent *)(cxq & ~_LBIT) ;
1.656 + assert (List != NULL, "invariant") ;
1.657 + goto WakeOne ;
1.658 + }
1.659 +
1.660 + // cxq|EntryList is empty.
1.661 + // w == NULL implies that cxq|EntryList == NULL in the past.
1.662 + // Possible race - rare inopportune interleaving.
1.663 + // A thread could have added itself to cxq since this thread previously checked.
1.664 + // Detect and recover by refetching cxq.
1.665 + Punt:
1.666 + assert (UNS(_OnDeck) == _LBIT, "invariant") ;
1.667 + _OnDeck = NULL ; // Release inner lock.
1.668 + OrderAccess::storeload(); // Dekker duality - pivot point
1.669 +
1.670 + // Resample LockWord/cxq to recover from possible race.
1.671 + // For instance, while this thread T1 held OnDeck, some other thread T2 might
1.672 + // acquire the outer lock. Another thread T3 might try to acquire the outer
1.673 + // lock, but encounter contention and enqueue itself on cxq. T2 then drops the
1.674 + // outer lock, but skips succession as this thread T1 still holds OnDeck.
1.675 + // T1 is and remains responsible for ensuring succession of T3.
1.676 + //
1.677 + // Note that we don't need to recheck EntryList, just cxq.
1.678 + // If threads moved onto EntryList since we dropped OnDeck
1.679 + // that implies some other thread forced succession.
1.680 + cxq = _LockWord.FullWord ;
1.681 + if ((cxq & ~_LBIT) != 0 && (cxq & _LBIT) == 0) {
1.682 + goto Succession ; // potential race -- re-run succession
1.683 + }
1.684 + return ;
1.685 +}
1.686 +
1.687 +bool Monitor::notify() {
1.688 + assert (_owner == Thread::current(), "invariant") ;
1.689 + assert (ILocked(), "invariant") ;
1.690 + if (_WaitSet == NULL) return true ;
1.691 + NotifyCount ++ ;
1.692 +
1.693 + // Transfer one thread from the WaitSet to the EntryList or cxq.
1.694 + // Currently we just unlink the head of the WaitSet and prepend to the cxq.
1.695 + // And of course we could just unlink it and unpark it, too, but
1.696 + // in that case it'd likely impale itself on the reentry.
1.697 + Thread::muxAcquire (_WaitLock, "notify:WaitLock") ;
1.698 + ParkEvent * nfy = _WaitSet ;
1.699 + if (nfy != NULL) { // DCL idiom
1.700 + _WaitSet = nfy->ListNext ;
1.701 + assert (nfy->Notified == 0, "invariant") ;
1.702 + // push nfy onto the cxq
1.703 + for (;;) {
1.704 + const intptr_t v = _LockWord.FullWord ;
1.705 + assert ((v & 0xFF) == _LBIT, "invariant") ;
1.706 + nfy->ListNext = (ParkEvent *)(v & ~_LBIT);
1.707 + if (CASPTR (&_LockWord, v, UNS(nfy)|_LBIT) == v) break;
1.708 + // interference - _LockWord changed -- just retry
1.709 + }
1.710 + // Note that setting Notified before pushing nfy onto the cxq is
1.711 + // also legal and safe, but the safety properties are much more
1.712 + // subtle, so for the sake of code stewardship ...
1.713 + OrderAccess::fence() ;
1.714 + nfy->Notified = 1;
1.715 + }
1.716 + Thread::muxRelease (_WaitLock) ;
1.717 + if (nfy != NULL && (NativeMonitorFlags & 16)) {
1.718 + // Experimental code ... light up the wakee in the hope that this thread (the owner)
1.719 + // will drop the lock just about the time the wakee comes ONPROC.
1.720 + nfy->unpark() ;
1.721 + }
1.722 + assert (ILocked(), "invariant") ;
1.723 + return true ;
1.724 +}
1.725 +
1.726 +// Currently notifyAll() transfers the waiters one-at-a-time from the waitset
1.727 +// to the cxq. This could be done more efficiently with a single bulk en-mass transfer,
1.728 +// but in practice notifyAll() for large #s of threads is rare and not time-critical.
1.729 +// Beware too, that we invert the order of the waiters. Lets say that the
1.730 +// waitset is "ABCD" and the cxq is "XYZ". After a notifyAll() the waitset
1.731 +// will be empty and the cxq will be "DCBAXYZ". This is benign, of course.
1.732 +
1.733 +bool Monitor::notify_all() {
1.734 + assert (_owner == Thread::current(), "invariant") ;
1.735 + assert (ILocked(), "invariant") ;
1.736 + while (_WaitSet != NULL) notify() ;
1.737 + return true ;
1.738 +}
1.739 +
1.740 +int Monitor::IWait (Thread * Self, jlong timo) {
1.741 + assert (ILocked(), "invariant") ;
1.742 +
1.743 + // Phases:
1.744 + // 1. Enqueue Self on WaitSet - currently prepend
1.745 + // 2. unlock - drop the outer lock
1.746 + // 3. wait for either notification or timeout
1.747 + // 4. lock - reentry - reacquire the outer lock
1.748 +
1.749 + ParkEvent * const ESelf = Self->_MutexEvent ;
1.750 + ESelf->Notified = 0 ;
1.751 + ESelf->reset() ;
1.752 + OrderAccess::fence() ;
1.753 +
1.754 + // Add Self to WaitSet
1.755 + // Ideally only the holder of the outer lock would manipulate the WaitSet -
1.756 + // That is, the outer lock would implicitly protect the WaitSet.
1.757 + // But if a thread in wait() encounters a timeout it will need to dequeue itself
1.758 + // from the WaitSet _before it becomes the owner of the lock. We need to dequeue
1.759 + // as the ParkEvent -- which serves as a proxy for the thread -- can't reside
1.760 + // on both the WaitSet and the EntryList|cxq at the same time.. That is, a thread
1.761 + // on the WaitSet can't be allowed to compete for the lock until it has managed to
1.762 + // unlink its ParkEvent from WaitSet. Thus the need for WaitLock.
1.763 + // Contention on the WaitLock is minimal.
1.764 + //
1.765 + // Another viable approach would be add another ParkEvent, "WaitEvent" to the
1.766 + // thread class. The WaitSet would be composed of WaitEvents. Only the
1.767 + // owner of the outer lock would manipulate the WaitSet. A thread in wait()
1.768 + // could then compete for the outer lock, and then, if necessary, unlink itself
1.769 + // from the WaitSet only after having acquired the outer lock. More precisely,
1.770 + // there would be no WaitLock. A thread in in wait() would enqueue its WaitEvent
1.771 + // on the WaitSet; release the outer lock; wait for either notification or timeout;
1.772 + // reacquire the inner lock; and then, if needed, unlink itself from the WaitSet.
1.773 + //
1.774 + // Alternatively, a 2nd set of list link fields in the ParkEvent might suffice.
1.775 + // One set would be for the WaitSet and one for the EntryList.
1.776 + // We could also deconstruct the ParkEvent into a "pure" event and add a
1.777 + // new immortal/TSM "ListElement" class that referred to ParkEvents.
1.778 + // In that case we could have one ListElement on the WaitSet and another
1.779 + // on the EntryList, with both referring to the same pure Event.
1.780 +
1.781 + Thread::muxAcquire (_WaitLock, "wait:WaitLock:Add") ;
1.782 + ESelf->ListNext = _WaitSet ;
1.783 + _WaitSet = ESelf ;
1.784 + Thread::muxRelease (_WaitLock) ;
1.785 +
1.786 + // Release the outer lock
1.787 + // We call IUnlock (RelaxAssert=true) as a thread T1 might
1.788 + // enqueue itself on the WaitSet, call IUnlock(), drop the lock,
1.789 + // and then stall before it can attempt to wake a successor.
1.790 + // Some other thread T2 acquires the lock, and calls notify(), moving
1.791 + // T1 from the WaitSet to the cxq. T2 then drops the lock. T1 resumes,
1.792 + // and then finds *itself* on the cxq. During the course of a normal
1.793 + // IUnlock() call a thread should _never find itself on the EntryList
1.794 + // or cxq, but in the case of wait() it's possible.
1.795 + // See synchronizer.cpp objectMonitor::wait().
1.796 + IUnlock (true) ;
1.797 +
1.798 + // Wait for either notification or timeout
1.799 + // Beware that in some circumstances we might propagate
1.800 + // spurious wakeups back to the caller.
1.801 +
1.802 + for (;;) {
1.803 + if (ESelf->Notified) break ;
1.804 + int err = ParkCommon (ESelf, timo) ;
1.805 + if (err == OS_TIMEOUT || (NativeMonitorFlags & 1)) break ;
1.806 + }
1.807 +
1.808 + // Prepare for reentry - if necessary, remove ESelf from WaitSet
1.809 + // ESelf can be:
1.810 + // 1. Still on the WaitSet. This can happen if we exited the loop by timeout.
1.811 + // 2. On the cxq or EntryList
1.812 + // 3. Not resident on cxq, EntryList or WaitSet, but in the OnDeck position.
1.813 +
1.814 + OrderAccess::fence() ;
1.815 + int WasOnWaitSet = 0 ;
1.816 + if (ESelf->Notified == 0) {
1.817 + Thread::muxAcquire (_WaitLock, "wait:WaitLock:remove") ;
1.818 + if (ESelf->Notified == 0) { // DCL idiom
1.819 + assert (_OnDeck != ESelf, "invariant") ; // can't be both OnDeck and on WaitSet
1.820 + // ESelf is resident on the WaitSet -- unlink it.
1.821 + // A doubly-linked list would be better here so we can unlink in constant-time.
1.822 + // We have to unlink before we potentially recontend as ESelf might otherwise
1.823 + // end up on the cxq|EntryList -- it can't be on two lists at once.
1.824 + ParkEvent * p = _WaitSet ;
1.825 + ParkEvent * q = NULL ; // classic q chases p
1.826 + while (p != NULL && p != ESelf) {
1.827 + q = p ;
1.828 + p = p->ListNext ;
1.829 + }
1.830 + assert (p == ESelf, "invariant") ;
1.831 + if (p == _WaitSet) { // found at head
1.832 + assert (q == NULL, "invariant") ;
1.833 + _WaitSet = p->ListNext ;
1.834 + } else { // found in interior
1.835 + assert (q->ListNext == p, "invariant") ;
1.836 + q->ListNext = p->ListNext ;
1.837 + }
1.838 + WasOnWaitSet = 1 ; // We were *not* notified but instead encountered timeout
1.839 + }
1.840 + Thread::muxRelease (_WaitLock) ;
1.841 + }
1.842 +
1.843 + // Reentry phase - reacquire the lock
1.844 + if (WasOnWaitSet) {
1.845 + // ESelf was previously on the WaitSet but we just unlinked it above
1.846 + // because of a timeout. ESelf is not resident on any list and is not OnDeck
1.847 + assert (_OnDeck != ESelf, "invariant") ;
1.848 + ILock (Self) ;
1.849 + } else {
1.850 + // A prior notify() operation moved ESelf from the WaitSet to the cxq.
1.851 + // ESelf is now on the cxq, EntryList or at the OnDeck position.
1.852 + // The following fragment is extracted from Monitor::ILock()
1.853 + for (;;) {
1.854 + if (_OnDeck == ESelf && TrySpin(Self)) break ;
1.855 + ParkCommon (ESelf, 0) ;
1.856 + }
1.857 + assert (_OnDeck == ESelf, "invariant") ;
1.858 + _OnDeck = NULL ;
1.859 + }
1.860 +
1.861 + assert (ILocked(), "invariant") ;
1.862 + return WasOnWaitSet != 0 ; // return true IFF timeout
1.863 +}
1.864 +
1.865 +
1.866 +// ON THE VMTHREAD SNEAKING PAST HELD LOCKS:
1.867 +// In particular, there are certain types of global lock that may be held
1.868 +// by a Java thread while it is blocked at a safepoint but before it has
1.869 +// written the _owner field. These locks may be sneakily acquired by the
1.870 +// VM thread during a safepoint to avoid deadlocks. Alternatively, one should
1.871 +// identify all such locks, and ensure that Java threads never block at
1.872 +// safepoints while holding them (_no_safepoint_check_flag). While it
1.873 +// seems as though this could increase the time to reach a safepoint
1.874 +// (or at least increase the mean, if not the variance), the latter
1.875 +// approach might make for a cleaner, more maintainable JVM design.
1.876 +//
1.877 +// Sneaking is vile and reprehensible and should be excised at the 1st
1.878 +// opportunity. It's possible that the need for sneaking could be obviated
1.879 +// as follows. Currently, a thread might (a) while TBIVM, call pthread_mutex_lock
1.880 +// or ILock() thus acquiring the "physical" lock underlying Monitor/Mutex.
1.881 +// (b) stall at the TBIVM exit point as a safepoint is in effect. Critically,
1.882 +// it'll stall at the TBIVM reentry state transition after having acquired the
1.883 +// underlying lock, but before having set _owner and having entered the actual
1.884 +// critical section. The lock-sneaking facility leverages that fact and allowed the
1.885 +// VM thread to logically acquire locks that had already be physically locked by mutators
1.886 +// but where mutators were known blocked by the reentry thread state transition.
1.887 +//
1.888 +// If we were to modify the Monitor-Mutex so that TBIVM state transitions tightly
1.889 +// wrapped calls to park(), then we could likely do away with sneaking. We'd
1.890 +// decouple lock acquisition and parking. The critical invariant to eliminating
1.891 +// sneaking is to ensure that we never "physically" acquire the lock while TBIVM.
1.892 +// An easy way to accomplish this is to wrap the park calls in a narrow TBIVM jacket.
1.893 +// One difficulty with this approach is that the TBIVM wrapper could recurse and
1.894 +// call lock() deep from within a lock() call, while the MutexEvent was already enqueued.
1.895 +// Using a stack (N=2 at minimum) of ParkEvents would take care of that problem.
1.896 +//
1.897 +// But of course the proper ultimate approach is to avoid schemes that require explicit
1.898 +// sneaking or dependence on any any clever invariants or subtle implementation properties
1.899 +// of Mutex-Monitor and instead directly address the underlying design flaw.
1.900 +
1.901 +void Monitor::lock (Thread * Self) {
1.902 +#ifdef CHECK_UNHANDLED_OOPS
1.903 + // Clear unhandled oops so we get a crash right away. Only clear for non-vm
1.904 + // or GC threads.
1.905 + if (Self->is_Java_thread()) {
1.906 + Self->clear_unhandled_oops();
1.907 + }
1.908 +#endif // CHECK_UNHANDLED_OOPS
1.909 +
1.910 + debug_only(check_prelock_state(Self));
1.911 + assert (_owner != Self , "invariant") ;
1.912 + assert (_OnDeck != Self->_MutexEvent, "invariant") ;
1.913 +
1.914 + if (TryFast()) {
1.915 + Exeunt:
1.916 + assert (ILocked(), "invariant") ;
1.917 + assert (owner() == NULL, "invariant");
1.918 + set_owner (Self);
1.919 + return ;
1.920 + }
1.921 +
1.922 + // The lock is contended ...
1.923 +
1.924 + bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint();
1.925 + if (can_sneak && _owner == NULL) {
1.926 + // a java thread has locked the lock but has not entered the
1.927 + // critical region -- let's just pretend we've locked the lock
1.928 + // and go on. we note this with _snuck so we can also
1.929 + // pretend to unlock when the time comes.
1.930 + _snuck = true;
1.931 + goto Exeunt ;
1.932 + }
1.933 +
1.934 + // Try a brief spin to avoid passing thru thread state transition ...
1.935 + if (TrySpin (Self)) goto Exeunt ;
1.936 +
1.937 + check_block_state(Self);
1.938 + if (Self->is_Java_thread()) {
1.939 + // Horribile dictu - we suffer through a state transition
1.940 + assert(rank() > Mutex::special, "Potential deadlock with special or lesser rank mutex");
1.941 + ThreadBlockInVM tbivm ((JavaThread *) Self) ;
1.942 + ILock (Self) ;
1.943 + } else {
1.944 + // Mirabile dictu
1.945 + ILock (Self) ;
1.946 + }
1.947 + goto Exeunt ;
1.948 +}
1.949 +
1.950 +void Monitor::lock() {
1.951 + this->lock(Thread::current());
1.952 +}
1.953 +
1.954 +// Lock without safepoint check - a degenerate variant of lock().
1.955 +// Should ONLY be used by safepoint code and other code
1.956 +// that is guaranteed not to block while running inside the VM. If this is called with
1.957 +// thread state set to be in VM, the safepoint synchronization code will deadlock!
1.958 +
1.959 +void Monitor::lock_without_safepoint_check (Thread * Self) {
1.960 + assert (_owner != Self, "invariant") ;
1.961 + ILock (Self) ;
1.962 + assert (_owner == NULL, "invariant");
1.963 + set_owner (Self);
1.964 +}
1.965 +
1.966 +void Monitor::lock_without_safepoint_check () {
1.967 + lock_without_safepoint_check (Thread::current()) ;
1.968 +}
1.969 +
1.970 +
1.971 +// Returns true if thread succeceed [sic] in grabbing the lock, otherwise false.
1.972 +
1.973 +bool Monitor::try_lock() {
1.974 + Thread * const Self = Thread::current();
1.975 + debug_only(check_prelock_state(Self));
1.976 + // assert(!thread->is_inside_signal_handler(), "don't lock inside signal handler");
1.977 +
1.978 + // Special case, where all Java threads are stopped.
1.979 + // The lock may have been acquired but _owner is not yet set.
1.980 + // In that case the VM thread can safely grab the lock.
1.981 + // It strikes me this should appear _after the TryLock() fails, below.
1.982 + bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint();
1.983 + if (can_sneak && _owner == NULL) {
1.984 + set_owner(Self); // Do not need to be atomic, since we are at a safepoint
1.985 + _snuck = true;
1.986 + return true;
1.987 + }
1.988 +
1.989 + if (TryLock()) {
1.990 + // We got the lock
1.991 + assert (_owner == NULL, "invariant");
1.992 + set_owner (Self);
1.993 + return true;
1.994 + }
1.995 + return false;
1.996 +}
1.997 +
1.998 +void Monitor::unlock() {
1.999 + assert (_owner == Thread::current(), "invariant") ;
1.1000 + assert (_OnDeck != Thread::current()->_MutexEvent , "invariant") ;
1.1001 + set_owner (NULL) ;
1.1002 + if (_snuck) {
1.1003 + assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak");
1.1004 + _snuck = false;
1.1005 + return ;
1.1006 + }
1.1007 + IUnlock (false) ;
1.1008 +}
1.1009 +
1.1010 +// Yet another degenerate version of Monitor::lock() or lock_without_safepoint_check()
1.1011 +// jvm_raw_lock() and _unlock() can be called by non-Java threads via JVM_RawMonitorEnter.
1.1012 +//
1.1013 +// There's no expectation that JVM_RawMonitors will interoperate properly with the native
1.1014 +// Mutex-Monitor constructs. We happen to implement JVM_RawMonitors in terms of
1.1015 +// native Mutex-Monitors simply as a matter of convenience. A simple abstraction layer
1.1016 +// over a pthread_mutex_t would work equally as well, but require more platform-specific
1.1017 +// code -- a "PlatformMutex". Alternatively, a simply layer over muxAcquire-muxRelease
1.1018 +// would work too.
1.1019 +//
1.1020 +// Since the caller might be a foreign thread, we don't necessarily have a Thread.MutexEvent
1.1021 +// instance available. Instead, we transiently allocate a ParkEvent on-demand if
1.1022 +// we encounter contention. That ParkEvent remains associated with the thread
1.1023 +// until it manages to acquire the lock, at which time we return the ParkEvent
1.1024 +// to the global ParkEvent free list. This is correct and suffices for our purposes.
1.1025 +//
1.1026 +// Beware that the original jvm_raw_unlock() had a "_snuck" test but that
1.1027 +// jvm_raw_lock() didn't have the corresponding test. I suspect that's an
1.1028 +// oversight, but I've replicated the original suspect logic in the new code ...
1.1029 +
1.1030 +void Monitor::jvm_raw_lock() {
1.1031 + assert(rank() == native, "invariant");
1.1032 +
1.1033 + if (TryLock()) {
1.1034 + Exeunt:
1.1035 + assert (ILocked(), "invariant") ;
1.1036 + assert (_owner == NULL, "invariant");
1.1037 + // This can potentially be called by non-java Threads. Thus, the ThreadLocalStorage
1.1038 + // might return NULL. Don't call set_owner since it will break on an NULL owner
1.1039 + // Consider installing a non-null "ANON" distinguished value instead of just NULL.
1.1040 + _owner = ThreadLocalStorage::thread();
1.1041 + return ;
1.1042 + }
1.1043 +
1.1044 + if (TrySpin(NULL)) goto Exeunt ;
1.1045 +
1.1046 + // slow-path - apparent contention
1.1047 + // Allocate a ParkEvent for transient use.
1.1048 + // The ParkEvent remains associated with this thread until
1.1049 + // the time the thread manages to acquire the lock.
1.1050 + ParkEvent * const ESelf = ParkEvent::Allocate(NULL) ;
1.1051 + ESelf->reset() ;
1.1052 + OrderAccess::storeload() ;
1.1053 +
1.1054 + // Either Enqueue Self on cxq or acquire the outer lock.
1.1055 + if (AcquireOrPush (ESelf)) {
1.1056 + ParkEvent::Release (ESelf) ; // surrender the ParkEvent
1.1057 + goto Exeunt ;
1.1058 + }
1.1059 +
1.1060 + // At any given time there is at most one ondeck thread.
1.1061 + // ondeck implies not resident on cxq and not resident on EntryList
1.1062 + // Only the OnDeck thread can try to acquire -- contended for -- the lock.
1.1063 + // CONSIDER: use Self->OnDeck instead of m->OnDeck.
1.1064 + for (;;) {
1.1065 + if (_OnDeck == ESelf && TrySpin(NULL)) break ;
1.1066 + ParkCommon (ESelf, 0) ;
1.1067 + }
1.1068 +
1.1069 + assert (_OnDeck == ESelf, "invariant") ;
1.1070 + _OnDeck = NULL ;
1.1071 + ParkEvent::Release (ESelf) ; // surrender the ParkEvent
1.1072 + goto Exeunt ;
1.1073 +}
1.1074 +
1.1075 +void Monitor::jvm_raw_unlock() {
1.1076 + // Nearly the same as Monitor::unlock() ...
1.1077 + // directly set _owner instead of using set_owner(null)
1.1078 + _owner = NULL ;
1.1079 + if (_snuck) { // ???
1.1080 + assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak");
1.1081 + _snuck = false;
1.1082 + return ;
1.1083 + }
1.1084 + IUnlock(false) ;
1.1085 +}
1.1086 +
1.1087 +bool Monitor::wait(bool no_safepoint_check, long timeout, bool as_suspend_equivalent) {
1.1088 + Thread * const Self = Thread::current() ;
1.1089 + assert (_owner == Self, "invariant") ;
1.1090 + assert (ILocked(), "invariant") ;
1.1091 +
1.1092 + // as_suspend_equivalent logically implies !no_safepoint_check
1.1093 + guarantee (!as_suspend_equivalent || !no_safepoint_check, "invariant") ;
1.1094 + // !no_safepoint_check logically implies java_thread
1.1095 + guarantee (no_safepoint_check || Self->is_Java_thread(), "invariant") ;
1.1096 +
1.1097 + #ifdef ASSERT
1.1098 + Monitor * least = get_least_ranked_lock_besides_this(Self->owned_locks());
1.1099 + assert(least != this, "Specification of get_least_... call above");
1.1100 + if (least != NULL && least->rank() <= special) {
1.1101 + tty->print("Attempting to wait on monitor %s/%d while holding"
1.1102 + " lock %s/%d -- possible deadlock",
1.1103 + name(), rank(), least->name(), least->rank());
1.1104 + assert(false, "Shouldn't block(wait) while holding a lock of rank special");
1.1105 + }
1.1106 + #endif // ASSERT
1.1107 +
1.1108 + int wait_status ;
1.1109 + // conceptually set the owner to NULL in anticipation of
1.1110 + // abdicating the lock in wait
1.1111 + set_owner(NULL);
1.1112 + if (no_safepoint_check) {
1.1113 + wait_status = IWait (Self, timeout) ;
1.1114 + } else {
1.1115 + assert (Self->is_Java_thread(), "invariant") ;
1.1116 + JavaThread *jt = (JavaThread *)Self;
1.1117 +
1.1118 + // Enter safepoint region - ornate and Rococo ...
1.1119 + ThreadBlockInVM tbivm(jt);
1.1120 + OSThreadWaitState osts(Self->osthread(), false /* not Object.wait() */);
1.1121 +
1.1122 + if (as_suspend_equivalent) {
1.1123 + jt->set_suspend_equivalent();
1.1124 + // cleared by handle_special_suspend_equivalent_condition() or
1.1125 + // java_suspend_self()
1.1126 + }
1.1127 +
1.1128 + wait_status = IWait (Self, timeout) ;
1.1129 +
1.1130 + // were we externally suspended while we were waiting?
1.1131 + if (as_suspend_equivalent && jt->handle_special_suspend_equivalent_condition()) {
1.1132 + // Our event wait has finished and we own the lock, but
1.1133 + // while we were waiting another thread suspended us. We don't
1.1134 + // want to hold the lock while suspended because that
1.1135 + // would surprise the thread that suspended us.
1.1136 + assert (ILocked(), "invariant") ;
1.1137 + IUnlock (true) ;
1.1138 + jt->java_suspend_self();
1.1139 + ILock (Self) ;
1.1140 + assert (ILocked(), "invariant") ;
1.1141 + }
1.1142 + }
1.1143 +
1.1144 + // Conceptually reestablish ownership of the lock.
1.1145 + // The "real" lock -- the LockByte -- was reacquired by IWait().
1.1146 + assert (ILocked(), "invariant") ;
1.1147 + assert (_owner == NULL, "invariant") ;
1.1148 + set_owner (Self) ;
1.1149 + return wait_status != 0 ; // return true IFF timeout
1.1150 +}
1.1151 +
1.1152 +Monitor::~Monitor() {
1.1153 + assert ((UNS(_owner)|UNS(_LockWord.FullWord)|UNS(_EntryList)|UNS(_WaitSet)|UNS(_OnDeck)) == 0, "") ;
1.1154 +}
1.1155 +
1.1156 +void Monitor::ClearMonitor (Monitor * m, const char *name) {
1.1157 + m->_owner = NULL ;
1.1158 + m->_snuck = false ;
1.1159 + if (name == NULL) {
1.1160 + strcpy(m->_name, "UNKNOWN") ;
1.1161 + } else {
1.1162 + strncpy(m->_name, name, MONITOR_NAME_LEN - 1);
1.1163 + m->_name[MONITOR_NAME_LEN - 1] = '\0';
1.1164 + }
1.1165 + m->_LockWord.FullWord = 0 ;
1.1166 + m->_EntryList = NULL ;
1.1167 + m->_OnDeck = NULL ;
1.1168 + m->_WaitSet = NULL ;
1.1169 + m->_WaitLock[0] = 0 ;
1.1170 +}
1.1171 +
1.1172 +Monitor::Monitor() { ClearMonitor(this); }
1.1173 +
1.1174 +Monitor::Monitor (int Rank, const char * name, bool allow_vm_block) {
1.1175 + ClearMonitor (this, name) ;
1.1176 +#ifdef ASSERT
1.1177 + _allow_vm_block = allow_vm_block;
1.1178 + _rank = Rank ;
1.1179 +#endif
1.1180 +}
1.1181 +
1.1182 +Mutex::~Mutex() {
1.1183 + assert ((UNS(_owner)|UNS(_LockWord.FullWord)|UNS(_EntryList)|UNS(_WaitSet)|UNS(_OnDeck)) == 0, "") ;
1.1184 +}
1.1185 +
1.1186 +Mutex::Mutex (int Rank, const char * name, bool allow_vm_block) {
1.1187 + ClearMonitor ((Monitor *) this, name) ;
1.1188 +#ifdef ASSERT
1.1189 + _allow_vm_block = allow_vm_block;
1.1190 + _rank = Rank ;
1.1191 +#endif
1.1192 +}
1.1193 +
1.1194 +bool Monitor::owned_by_self() const {
1.1195 + bool ret = _owner == Thread::current();
1.1196 + assert (!ret || _LockWord.Bytes[_LSBINDEX] != 0, "invariant") ;
1.1197 + return ret;
1.1198 +}
1.1199 +
1.1200 +void Monitor::print_on_error(outputStream* st) const {
1.1201 + st->print("[" PTR_FORMAT, this);
1.1202 + st->print("] %s", _name);
1.1203 + st->print(" - owner thread: " PTR_FORMAT, _owner);
1.1204 +}
1.1205 +
1.1206 +
1.1207 +
1.1208 +
1.1209 +// ----------------------------------------------------------------------------------
1.1210 +// Non-product code
1.1211 +
1.1212 +#ifndef PRODUCT
1.1213 +void Monitor::print_on(outputStream* st) const {
1.1214 + st->print_cr("Mutex: [0x%lx/0x%lx] %s - owner: 0x%lx", this, _LockWord.FullWord, _name, _owner);
1.1215 +}
1.1216 +#endif
1.1217 +
1.1218 +#ifndef PRODUCT
1.1219 +#ifdef ASSERT
1.1220 +Monitor * Monitor::get_least_ranked_lock(Monitor * locks) {
1.1221 + Monitor *res, *tmp;
1.1222 + for (res = tmp = locks; tmp != NULL; tmp = tmp->next()) {
1.1223 + if (tmp->rank() < res->rank()) {
1.1224 + res = tmp;
1.1225 + }
1.1226 + }
1.1227 + if (!SafepointSynchronize::is_at_safepoint()) {
1.1228 + // In this case, we expect the held locks to be
1.1229 + // in increasing rank order (modulo any native ranks)
1.1230 + for (tmp = locks; tmp != NULL; tmp = tmp->next()) {
1.1231 + if (tmp->next() != NULL) {
1.1232 + assert(tmp->rank() == Mutex::native ||
1.1233 + tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?");
1.1234 + }
1.1235 + }
1.1236 + }
1.1237 + return res;
1.1238 +}
1.1239 +
1.1240 +Monitor* Monitor::get_least_ranked_lock_besides_this(Monitor* locks) {
1.1241 + Monitor *res, *tmp;
1.1242 + for (res = NULL, tmp = locks; tmp != NULL; tmp = tmp->next()) {
1.1243 + if (tmp != this && (res == NULL || tmp->rank() < res->rank())) {
1.1244 + res = tmp;
1.1245 + }
1.1246 + }
1.1247 + if (!SafepointSynchronize::is_at_safepoint()) {
1.1248 + // In this case, we expect the held locks to be
1.1249 + // in increasing rank order (modulo any native ranks)
1.1250 + for (tmp = locks; tmp != NULL; tmp = tmp->next()) {
1.1251 + if (tmp->next() != NULL) {
1.1252 + assert(tmp->rank() == Mutex::native ||
1.1253 + tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?");
1.1254 + }
1.1255 + }
1.1256 + }
1.1257 + return res;
1.1258 +}
1.1259 +
1.1260 +
1.1261 +bool Monitor::contains(Monitor* locks, Monitor * lock) {
1.1262 + for (; locks != NULL; locks = locks->next()) {
1.1263 + if (locks == lock)
1.1264 + return true;
1.1265 + }
1.1266 + return false;
1.1267 +}
1.1268 +#endif
1.1269 +
1.1270 +// Called immediately after lock acquisition or release as a diagnostic
1.1271 +// to track the lock-set of the thread and test for rank violations that
1.1272 +// might indicate exposure to deadlock.
1.1273 +// Rather like an EventListener for _owner (:>).
1.1274 +
1.1275 +void Monitor::set_owner_implementation(Thread *new_owner) {
1.1276 + // This function is solely responsible for maintaining
1.1277 + // and checking the invariant that threads and locks
1.1278 + // are in a 1/N relation, with some some locks unowned.
1.1279 + // It uses the Mutex::_owner, Mutex::_next, and
1.1280 + // Thread::_owned_locks fields, and no other function
1.1281 + // changes those fields.
1.1282 + // It is illegal to set the mutex from one non-NULL
1.1283 + // owner to another--it must be owned by NULL as an
1.1284 + // intermediate state.
1.1285 +
1.1286 + if (new_owner != NULL) {
1.1287 + // the thread is acquiring this lock
1.1288 +
1.1289 + assert(new_owner == Thread::current(), "Should I be doing this?");
1.1290 + assert(_owner == NULL, "setting the owner thread of an already owned mutex");
1.1291 + _owner = new_owner; // set the owner
1.1292 +
1.1293 + // link "this" into the owned locks list
1.1294 +
1.1295 + #ifdef ASSERT // Thread::_owned_locks is under the same ifdef
1.1296 + Monitor* locks = get_least_ranked_lock(new_owner->owned_locks());
1.1297 + // Mutex::set_owner_implementation is a friend of Thread
1.1298 +
1.1299 + assert(this->rank() >= 0, "bad lock rank");
1.1300 +
1.1301 + // Deadlock avoidance rules require us to acquire Mutexes only in
1.1302 + // a global total order. For example m1 is the lowest ranked mutex
1.1303 + // that the thread holds and m2 is the mutex the thread is trying
1.1304 + // to acquire, then deadlock avoidance rules require that the rank
1.1305 + // of m2 be less than the rank of m1.
1.1306 + // The rank Mutex::native is an exception in that it is not subject
1.1307 + // to the verification rules.
1.1308 + // Here are some further notes relating to mutex acquisition anomalies:
1.1309 + // . under Solaris, the interrupt lock gets acquired when doing
1.1310 + // profiling, so any lock could be held.
1.1311 + // . it is also ok to acquire Safepoint_lock at the very end while we
1.1312 + // already hold Terminator_lock - may happen because of periodic safepoints
1.1313 + if (this->rank() != Mutex::native &&
1.1314 + this->rank() != Mutex::suspend_resume &&
1.1315 + locks != NULL && locks->rank() <= this->rank() &&
1.1316 + !SafepointSynchronize::is_at_safepoint() &&
1.1317 + this != Interrupt_lock && this != ProfileVM_lock &&
1.1318 + !(this == Safepoint_lock && contains(locks, Terminator_lock) &&
1.1319 + SafepointSynchronize::is_synchronizing())) {
1.1320 + new_owner->print_owned_locks();
1.1321 + fatal(err_msg("acquiring lock %s/%d out of order with lock %s/%d -- "
1.1322 + "possible deadlock", this->name(), this->rank(),
1.1323 + locks->name(), locks->rank()));
1.1324 + }
1.1325 +
1.1326 + this->_next = new_owner->_owned_locks;
1.1327 + new_owner->_owned_locks = this;
1.1328 + #endif
1.1329 +
1.1330 + } else {
1.1331 + // the thread is releasing this lock
1.1332 +
1.1333 + Thread* old_owner = _owner;
1.1334 + debug_only(_last_owner = old_owner);
1.1335 +
1.1336 + assert(old_owner != NULL, "removing the owner thread of an unowned mutex");
1.1337 + assert(old_owner == Thread::current(), "removing the owner thread of an unowned mutex");
1.1338 +
1.1339 + _owner = NULL; // set the owner
1.1340 +
1.1341 + #ifdef ASSERT
1.1342 + Monitor *locks = old_owner->owned_locks();
1.1343 +
1.1344 + // remove "this" from the owned locks list
1.1345 +
1.1346 + Monitor *prev = NULL;
1.1347 + bool found = false;
1.1348 + for (; locks != NULL; prev = locks, locks = locks->next()) {
1.1349 + if (locks == this) {
1.1350 + found = true;
1.1351 + break;
1.1352 + }
1.1353 + }
1.1354 + assert(found, "Removing a lock not owned");
1.1355 + if (prev == NULL) {
1.1356 + old_owner->_owned_locks = _next;
1.1357 + } else {
1.1358 + prev->_next = _next;
1.1359 + }
1.1360 + _next = NULL;
1.1361 + #endif
1.1362 + }
1.1363 +}
1.1364 +
1.1365 +
1.1366 +// Factored out common sanity checks for locking mutex'es. Used by lock() and try_lock()
1.1367 +void Monitor::check_prelock_state(Thread *thread) {
1.1368 + assert((!thread->is_Java_thread() || ((JavaThread *)thread)->thread_state() == _thread_in_vm)
1.1369 + || rank() == Mutex::special, "wrong thread state for using locks");
1.1370 + if (StrictSafepointChecks) {
1.1371 + if (thread->is_VM_thread() && !allow_vm_block()) {
1.1372 + fatal(err_msg("VM thread using lock %s (not allowed to block on)",
1.1373 + name()));
1.1374 + }
1.1375 + debug_only(if (rank() != Mutex::special) \
1.1376 + thread->check_for_valid_safepoint_state(false);)
1.1377 + }
1.1378 + if (thread->is_Watcher_thread()) {
1.1379 + assert(!WatcherThread::watcher_thread()->has_crash_protection(),
1.1380 + "locking not allowed when crash protection is set");
1.1381 + }
1.1382 +}
1.1383 +
1.1384 +void Monitor::check_block_state(Thread *thread) {
1.1385 + if (!_allow_vm_block && thread->is_VM_thread()) {
1.1386 + warning("VM thread blocked on lock");
1.1387 + print();
1.1388 + BREAKPOINT;
1.1389 + }
1.1390 + assert(_owner != thread, "deadlock: blocking on monitor owned by current thread");
1.1391 +}
1.1392 +
1.1393 +#endif // PRODUCT
