duke@435: duke@435: /* drchase@6680: * Copyright (c) 1998, 2014, Oracle and/or its affiliates. All rights reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * trims@1907: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA trims@1907: * or visit www.oracle.com if you need additional information or have any trims@1907: * questions. duke@435: * duke@435: */ duke@435: stefank@2314: #include "precompiled.hpp" stefank@2314: #include "runtime/mutex.hpp" goetz@6911: #include "runtime/orderAccess.inline.hpp" stefank@2314: #include "runtime/osThread.hpp" stefank@4299: #include "runtime/thread.inline.hpp" stefank@2314: #include "utilities/events.hpp" stefank@2314: #ifdef TARGET_OS_FAMILY_linux stefank@2314: # include "mutex_linux.inline.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_OS_FAMILY_solaris stefank@2314: # include "mutex_solaris.inline.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_OS_FAMILY_windows stefank@2314: # include "mutex_windows.inline.hpp" stefank@2314: #endif never@3156: #ifdef TARGET_OS_FAMILY_bsd never@3156: # include "mutex_bsd.inline.hpp" never@3156: #endif duke@435: drchase@6680: PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC drchase@6680: duke@435: // 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 duke@435: // duke@435: // Native Monitor-Mutex locking - theory of operations duke@435: // duke@435: // * Native Monitors are completely unrelated to Java-level monitors, duke@435: // although the "back-end" slow-path implementations share a common lineage. duke@435: // See objectMonitor:: in synchronizer.cpp. duke@435: // Native Monitors do *not* support nesting or recursion but otherwise duke@435: // they're basically Hoare-flavor monitors. duke@435: // duke@435: // * A thread acquires ownership of a Monitor/Mutex by CASing the LockByte duke@435: // in the _LockWord from zero to non-zero. Note that the _Owner field duke@435: // is advisory and is used only to verify that the thread calling unlock() duke@435: // is indeed the last thread to have acquired the lock. duke@435: // duke@435: // * Contending threads "push" themselves onto the front of the contention duke@435: // queue -- called the cxq -- with CAS and then spin/park. duke@435: // The _LockWord contains the LockByte as well as the pointer to the head duke@435: // of the cxq. Colocating the LockByte with the cxq precludes certain races. duke@435: // duke@435: // * Using a separately addressable LockByte allows for CAS:MEMBAR or CAS:0 duke@435: // idioms. We currently use MEMBAR in the uncontended unlock() path, as duke@435: // MEMBAR often has less latency than CAS. If warranted, we could switch to duke@435: // a CAS:0 mode, using timers to close the resultant race, as is done duke@435: // with Java Monitors in synchronizer.cpp. duke@435: // duke@435: // See the following for a discussion of the relative cost of atomics (CAS) duke@435: // MEMBAR, and ways to eliminate such instructions from the common-case paths: duke@435: // -- http://blogs.sun.com/dave/entry/biased_locking_in_hotspot duke@435: // -- http://blogs.sun.com/dave/resource/MustangSync.pdf duke@435: // -- http://blogs.sun.com/dave/resource/synchronization-public2.pdf duke@435: // -- synchronizer.cpp duke@435: // duke@435: // * Overall goals - desiderata duke@435: // 1. Minimize context switching duke@435: // 2. Minimize lock migration duke@435: // 3. Minimize CPI -- affinity and locality duke@435: // 4. Minimize the execution of high-latency instructions such as CAS or MEMBAR duke@435: // 5. Minimize outer lock hold times duke@435: // 6. Behave gracefully on a loaded system duke@435: // duke@435: // * Thread flow and list residency: duke@435: // duke@435: // Contention queue --> EntryList --> OnDeck --> Owner --> !Owner duke@435: // [..resident on monitor list..] duke@435: // [...........contending..................] duke@435: // duke@435: // -- The contention queue (cxq) contains recently-arrived threads (RATs). duke@435: // Threads on the cxq eventually drain into the EntryList. duke@435: // -- Invariant: a thread appears on at most one list -- cxq, EntryList duke@435: // or WaitSet -- at any one time. duke@435: // -- For a given monitor there can be at most one "OnDeck" thread at any duke@435: // given time but if needbe this particular invariant could be relaxed. duke@435: // duke@435: // * The WaitSet and EntryList linked lists are composed of ParkEvents. duke@435: // I use ParkEvent instead of threads as ParkEvents are immortal and duke@435: // type-stable, meaning we can safely unpark() a possibly stale duke@435: // list element in the unlock()-path. (That's benign). duke@435: // duke@435: // * Succession policy - providing for progress: duke@435: // duke@435: // As necessary, the unlock()ing thread identifies, unlinks, and unparks duke@435: // an "heir presumptive" tentative successor thread from the EntryList. duke@435: // This becomes the so-called "OnDeck" thread, of which there can be only duke@435: // one at any given time for a given monitor. The wakee will recontend duke@435: // for ownership of monitor. duke@435: // duke@435: // Succession is provided for by a policy of competitive handoff. duke@435: // The exiting thread does _not_ grant or pass ownership to the duke@435: // successor thread. (This is also referred to as "handoff" succession"). duke@435: // Instead the exiting thread releases ownership and possibly wakes duke@435: // a successor, so the successor can (re)compete for ownership of the lock. duke@435: // duke@435: // Competitive handoff provides excellent overall throughput at the expense duke@435: // of short-term fairness. If fairness is a concern then one remedy might duke@435: // be to add an AcquireCounter field to the monitor. After a thread acquires duke@435: // the lock it will decrement the AcquireCounter field. When the count duke@435: // reaches 0 the thread would reset the AcquireCounter variable, abdicate duke@435: // the lock directly to some thread on the EntryList, and then move itself to the duke@435: // tail of the EntryList. duke@435: // duke@435: // But in practice most threads engage or otherwise participate in resource duke@435: // bounded producer-consumer relationships, so lock domination is not usually duke@435: // a practical concern. Recall too, that in general it's easier to construct duke@435: // a fair lock from a fast lock, but not vice-versa. duke@435: // duke@435: // * The cxq can have multiple concurrent "pushers" but only one concurrent duke@435: // detaching thread. This mechanism is immune from the ABA corruption. duke@435: // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. duke@435: // We use OnDeck as a pseudo-lock to enforce the at-most-one detaching duke@435: // thread constraint. duke@435: // duke@435: // * Taken together, the cxq and the EntryList constitute or form a duke@435: // single logical queue of threads stalled trying to acquire the lock. duke@435: // We use two distinct lists to reduce heat on the list ends. duke@435: // Threads in lock() enqueue onto cxq while threads in unlock() will duke@435: // dequeue from the EntryList. (c.f. Michael Scott's "2Q" algorithm). duke@435: // A key desideratum is to minimize queue & monitor metadata manipulation duke@435: // that occurs while holding the "outer" monitor lock -- that is, we want to duke@435: // minimize monitor lock holds times. duke@435: // duke@435: // The EntryList is ordered by the prevailing queue discipline and duke@435: // can be organized in any convenient fashion, such as a doubly-linked list or duke@435: // a circular doubly-linked list. If we need a priority queue then something akin duke@435: // to Solaris' sleepq would work nicely. Viz., duke@435: // -- http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. duke@435: // -- http://cvs.opensolaris.org/source/xref/onnv/onnv-gate/usr/src/uts/common/os/sleepq.c duke@435: // Queue discipline is enforced at ::unlock() time, when the unlocking thread duke@435: // drains the cxq into the EntryList, and orders or reorders the threads on the duke@435: // EntryList accordingly. duke@435: // duke@435: // Barring "lock barging", this mechanism provides fair cyclic ordering, duke@435: // somewhat similar to an elevator-scan. duke@435: // duke@435: // * OnDeck duke@435: // -- For a given monitor there can be at most one OnDeck thread at any given duke@435: // instant. The OnDeck thread is contending for the lock, but has been duke@435: // unlinked from the EntryList and cxq by some previous unlock() operations. duke@435: // Once a thread has been designated the OnDeck thread it will remain so duke@435: // until it manages to acquire the lock -- being OnDeck is a stable property. duke@435: // -- Threads on the EntryList or cxq are _not allowed to attempt lock acquisition. duke@435: // -- OnDeck also serves as an "inner lock" as follows. Threads in unlock() will, after duke@435: // having cleared the LockByte and dropped the outer lock, attempt to "trylock" duke@435: // OnDeck by CASing the field from null to non-null. If successful, that thread duke@435: // is then responsible for progress and succession and can use CAS to detach and duke@435: // drain the cxq into the EntryList. By convention, only this thread, the holder of duke@435: // the OnDeck inner lock, can manipulate the EntryList or detach and drain the duke@435: // RATs on the cxq into the EntryList. This avoids ABA corruption on the cxq as duke@435: // we allow multiple concurrent "push" operations but restrict detach concurrency duke@435: // to at most one thread. Having selected and detached a successor, the thread then duke@435: // changes the OnDeck to refer to that successor, and then unparks the successor. duke@435: // That successor will eventually acquire the lock and clear OnDeck. Beware duke@435: // that the OnDeck usage as a lock is asymmetric. A thread in unlock() transiently duke@435: // "acquires" OnDeck, performs queue manipulations, passes OnDeck to some successor, duke@435: // and then the successor eventually "drops" OnDeck. Note that there's never duke@435: // any sense of contention on the inner lock, however. Threads never contend duke@435: // or wait for the inner lock. duke@435: // -- OnDeck provides for futile wakeup throttling a described in section 3.3 of duke@435: // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf duke@435: // In a sense, OnDeck subsumes the ObjectMonitor _Succ and ObjectWaiter duke@435: // TState fields found in Java-level objectMonitors. (See synchronizer.cpp). duke@435: // duke@435: // * Waiting threads reside on the WaitSet list -- wait() puts duke@435: // the caller onto the WaitSet. Notify() or notifyAll() simply duke@435: // transfers threads from the WaitSet to either the EntryList or cxq. duke@435: // Subsequent unlock() operations will eventually unpark the notifyee. duke@435: // Unparking a notifee in notify() proper is inefficient - if we were to do so duke@435: // it's likely the notifyee would simply impale itself on the lock held duke@435: // by the notifier. duke@435: // duke@435: // * The mechanism is obstruction-free in that if the holder of the transient duke@435: // OnDeck lock in unlock() is preempted or otherwise stalls, other threads duke@435: // can still acquire and release the outer lock and continue to make progress. duke@435: // At worst, waking of already blocked contending threads may be delayed, duke@435: // but nothing worse. (We only use "trylock" operations on the inner OnDeck duke@435: // lock). duke@435: // duke@435: // * Note that thread-local storage must be initialized before a thread duke@435: // uses Native monitors or mutexes. The native monitor-mutex subsystem duke@435: // depends on Thread::current(). duke@435: // duke@435: // * The monitor synchronization subsystem avoids the use of native duke@435: // synchronization primitives except for the narrow platform-specific duke@435: // park-unpark abstraction. See the comments in os_solaris.cpp regarding duke@435: // the semantics of park-unpark. Put another way, this monitor implementation duke@435: // depends only on atomic operations and park-unpark. The monitor subsystem duke@435: // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the duke@435: // underlying OS manages the READY<->RUN transitions. duke@435: // duke@435: // * The memory consistency model provide by lock()-unlock() is at least as duke@435: // strong or stronger than the Java Memory model defined by JSR-133. duke@435: // That is, we guarantee at least entry consistency, if not stronger. duke@435: // See http://g.oswego.edu/dl/jmm/cookbook.html. duke@435: // duke@435: // * Thread:: currently contains a set of purpose-specific ParkEvents: duke@435: // _MutexEvent, _ParkEvent, etc. A better approach might be to do away with duke@435: // the purpose-specific ParkEvents and instead implement a general per-thread duke@435: // stack of available ParkEvents which we could provision on-demand. The duke@435: // stack acts as a local cache to avoid excessive calls to ParkEvent::Allocate() duke@435: // and ::Release(). A thread would simply pop an element from the local stack before it duke@435: // enqueued or park()ed. When the contention was over the thread would duke@435: // push the no-longer-needed ParkEvent back onto its stack. duke@435: // duke@435: // * A slightly reduced form of ILock() and IUnlock() have been partially duke@435: // model-checked (Murphi) for safety and progress at T=1,2,3 and 4. duke@435: // It'd be interesting to see if TLA/TLC could be useful as well. duke@435: // duke@435: // * Mutex-Monitor is a low-level "leaf" subsystem. That is, the monitor duke@435: // code should never call other code in the JVM that might itself need to duke@435: // acquire monitors or mutexes. That's true *except* in the case of the duke@435: // ThreadBlockInVM state transition wrappers. The ThreadBlockInVM DTOR handles duke@435: // mutator reentry (ingress) by checking for a pending safepoint in which case it will duke@435: // call SafepointSynchronize::block(), which in turn may call Safepoint_lock->lock(), etc. duke@435: // In that particular case a call to lock() for a given Monitor can end up recursively duke@435: // calling lock() on another monitor. While distasteful, this is largely benign duke@435: // as the calls come from jacket that wraps lock(), and not from deep within lock() itself. duke@435: // duke@435: // It's unfortunate that native mutexes and thread state transitions were convolved. duke@435: // They're really separate concerns and should have remained that way. Melding duke@435: // them together was facile -- a bit too facile. The current implementation badly duke@435: // conflates the two concerns. duke@435: // duke@435: // * TODO-FIXME: duke@435: // duke@435: // -- Add DTRACE probes for contended acquire, contended acquired, contended unlock duke@435: // We should also add DTRACE probes in the ParkEvent subsystem for duke@435: // Park-entry, Park-exit, and Unpark. duke@435: // duke@435: // -- We have an excess of mutex-like constructs in the JVM, namely: duke@435: // 1. objectMonitors for Java-level synchronization (synchronizer.cpp) duke@435: // 2. low-level muxAcquire and muxRelease duke@435: // 3. low-level spinAcquire and spinRelease duke@435: // 4. native Mutex:: and Monitor:: duke@435: // 5. jvm_raw_lock() and _unlock() duke@435: // 6. JVMTI raw monitors -- distinct from (5) despite having a confusingly duke@435: // similar name. duke@435: // duke@435: // 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 duke@435: duke@435: duke@435: // CASPTR() uses the canonical argument order that dominates in the literature. duke@435: // Our internal cmpxchg_ptr() uses a bastardized ordering to accommodate Sun .il templates. duke@435: duke@435: #define CASPTR(a,c,s) intptr_t(Atomic::cmpxchg_ptr ((void *)(s),(void *)(a),(void *)(c))) duke@435: #define UNS(x) (uintptr_t(x)) duke@435: #define TRACE(m) { static volatile int ctr = 0 ; int x = ++ctr ; if ((x & (x-1))==0) { ::printf ("%d:%s\n", x, #m); ::fflush(stdout); }} duke@435: duke@435: // Simplistic low-quality Marsaglia SHIFT-XOR RNG. duke@435: // Bijective except for the trailing mask operation. duke@435: // Useful for spin loops as the compiler can't optimize it away. duke@435: duke@435: static inline jint MarsagliaXORV (jint x) { duke@435: if (x == 0) x = 1|os::random() ; duke@435: x ^= x << 6; duke@435: x ^= ((unsigned)x) >> 21; duke@435: x ^= x << 7 ; duke@435: return x & 0x7FFFFFFF ; duke@435: } duke@435: duke@435: static inline jint MarsagliaXOR (jint * const a) { duke@435: jint x = *a ; duke@435: if (x == 0) x = UNS(a)|1 ; duke@435: x ^= x << 6; duke@435: x ^= ((unsigned)x) >> 21; duke@435: x ^= x << 7 ; duke@435: *a = x ; duke@435: return x & 0x7FFFFFFF ; duke@435: } duke@435: duke@435: static int Stall (int its) { duke@435: static volatile jint rv = 1 ; duke@435: volatile int OnFrame = 0 ; duke@435: jint v = rv ^ UNS(OnFrame) ; duke@435: while (--its >= 0) { duke@435: v = MarsagliaXORV (v) ; duke@435: } duke@435: // Make this impossible for the compiler to optimize away, duke@435: // but (mostly) avoid W coherency sharing on MP systems. duke@435: if (v == 0x12345) rv = v ; duke@435: return v ; duke@435: } duke@435: duke@435: int Monitor::TryLock () { duke@435: intptr_t v = _LockWord.FullWord ; duke@435: for (;;) { duke@435: if ((v & _LBIT) != 0) return 0 ; duke@435: const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ; duke@435: if (v == u) return 1 ; duke@435: v = u ; duke@435: } duke@435: } duke@435: duke@435: int Monitor::TryFast () { duke@435: // Optimistic fast-path form ... duke@435: // Fast-path attempt for the common uncontended case. duke@435: // Avoid RTS->RTO $ coherence upgrade on typical SMP systems. duke@435: intptr_t v = CASPTR (&_LockWord, 0, _LBIT) ; // agro ... duke@435: if (v == 0) return 1 ; duke@435: duke@435: for (;;) { duke@435: if ((v & _LBIT) != 0) return 0 ; duke@435: const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ; duke@435: if (v == u) return 1 ; duke@435: v = u ; duke@435: } duke@435: } duke@435: duke@435: int Monitor::ILocked () { duke@435: const intptr_t w = _LockWord.FullWord & 0xFF ; duke@435: assert (w == 0 || w == _LBIT, "invariant") ; duke@435: return w == _LBIT ; duke@435: } duke@435: duke@435: // Polite TATAS spinlock with exponential backoff - bounded spin. duke@435: // Ideally we'd use processor cycles, time or vtime to control duke@435: // the loop, but we currently use iterations. duke@435: // All the constants within were derived empirically but work over duke@435: // over the spectrum of J2SE reference platforms. duke@435: // On Niagara-class systems the back-off is unnecessary but duke@435: // is relatively harmless. (At worst it'll slightly retard duke@435: // acquisition times). The back-off is critical for older SMP systems duke@435: // where constant fetching of the LockWord would otherwise impair duke@435: // scalability. duke@435: // duke@435: // Clamp spinning at approximately 1/2 of a context-switch round-trip. duke@435: // See synchronizer.cpp for details and rationale. duke@435: duke@435: int Monitor::TrySpin (Thread * const Self) { duke@435: if (TryLock()) return 1 ; duke@435: if (!os::is_MP()) return 0 ; duke@435: duke@435: int Probes = 0 ; duke@435: int Delay = 0 ; duke@435: int Steps = 0 ; duke@435: int SpinMax = NativeMonitorSpinLimit ; duke@435: int flgs = NativeMonitorFlags ; duke@435: for (;;) { duke@435: intptr_t v = _LockWord.FullWord; duke@435: if ((v & _LBIT) == 0) { duke@435: if (CASPTR (&_LockWord, v, v|_LBIT) == v) { duke@435: return 1 ; duke@435: } duke@435: continue ; duke@435: } duke@435: duke@435: if ((flgs & 8) == 0) { duke@435: SpinPause () ; duke@435: } duke@435: duke@435: // Periodically increase Delay -- variable Delay form duke@435: // conceptually: delay *= 1 + 1/Exponent duke@435: ++ Probes; duke@435: if (Probes > SpinMax) return 0 ; duke@435: duke@435: if ((Probes & 0x7) == 0) { duke@435: Delay = ((Delay << 1)|1) & 0x7FF ; duke@435: // CONSIDER: Delay += 1 + (Delay/4); Delay &= 0x7FF ; duke@435: } duke@435: duke@435: if (flgs & 2) continue ; duke@435: duke@435: // Consider checking _owner's schedctl state, if OFFPROC abort spin. duke@435: // If the owner is OFFPROC then it's unlike that the lock will be dropped duke@435: // in a timely fashion, which suggests that spinning would not be fruitful duke@435: // or profitable. duke@435: duke@435: // Stall for "Delay" time units - iterations in the current implementation. duke@435: // Avoid generating coherency traffic while stalled. duke@435: // Possible ways to delay: duke@435: // PAUSE, SLEEP, MEMBAR #sync, MEMBAR #halt, duke@435: // wr %g0,%asi, gethrtime, rdstick, rdtick, rdtsc, etc. ... duke@435: // Note that on Niagara-class systems we want to minimize STs in the duke@435: // spin loop. N1 and brethren write-around the L1$ over the xbar into the L2$. duke@435: // Furthermore, they don't have a W$ like traditional SPARC processors. duke@435: // We currently use a Marsaglia Shift-Xor RNG loop. duke@435: Steps += Delay ; duke@435: if (Self != NULL) { duke@435: jint rv = Self->rng[0] ; duke@435: for (int k = Delay ; --k >= 0; ) { duke@435: rv = MarsagliaXORV (rv) ; duke@435: if ((flgs & 4) == 0 && SafepointSynchronize::do_call_back()) return 0 ; duke@435: } duke@435: Self->rng[0] = rv ; duke@435: } else { duke@435: Stall (Delay) ; duke@435: } duke@435: } duke@435: } duke@435: duke@435: static int ParkCommon (ParkEvent * ev, jlong timo) { duke@435: // Diagnostic support - periodically unwedge blocked threads duke@435: intx nmt = NativeMonitorTimeout ; duke@435: if (nmt > 0 && (nmt < timo || timo <= 0)) { duke@435: timo = nmt ; duke@435: } duke@435: int err = OS_OK ; duke@435: if (0 == timo) { duke@435: ev->park() ; duke@435: } else { duke@435: err = ev->park(timo) ; duke@435: } duke@435: return err ; duke@435: } duke@435: duke@435: inline int Monitor::AcquireOrPush (ParkEvent * ESelf) { duke@435: intptr_t v = _LockWord.FullWord ; duke@435: for (;;) { duke@435: if ((v & _LBIT) == 0) { duke@435: const intptr_t u = CASPTR (&_LockWord, v, v|_LBIT) ; duke@435: if (u == v) return 1 ; // indicate acquired duke@435: v = u ; duke@435: } else { duke@435: // Anticipate success ... duke@435: ESelf->ListNext = (ParkEvent *) (v & ~_LBIT) ; duke@435: const intptr_t u = CASPTR (&_LockWord, v, intptr_t(ESelf)|_LBIT) ; duke@435: if (u == v) return 0 ; // indicate pushed onto cxq duke@435: v = u ; duke@435: } duke@435: // Interference - LockWord change - just retry duke@435: } duke@435: } duke@435: duke@435: // ILock and IWait are the lowest level primitive internal blocking duke@435: // synchronization functions. The callers of IWait and ILock must have duke@435: // performed any needed state transitions beforehand. duke@435: // IWait and ILock may directly call park() without any concern for thread state. duke@435: // Note that ILock and IWait do *not* access _owner. duke@435: // _owner is a higher-level logical concept. duke@435: duke@435: void Monitor::ILock (Thread * Self) { duke@435: assert (_OnDeck != Self->_MutexEvent, "invariant") ; duke@435: duke@435: if (TryFast()) { duke@435: Exeunt: duke@435: assert (ILocked(), "invariant") ; duke@435: return ; duke@435: } duke@435: duke@435: ParkEvent * const ESelf = Self->_MutexEvent ; duke@435: assert (_OnDeck != ESelf, "invariant") ; duke@435: duke@435: // As an optimization, spinners could conditionally try to set ONDECK to _LBIT duke@435: // Synchronizer.cpp uses a similar optimization. duke@435: if (TrySpin (Self)) goto Exeunt ; duke@435: duke@435: // Slow-path - the lock is contended. duke@435: // Either Enqueue Self on cxq or acquire the outer lock. duke@435: // LockWord encoding = (cxq,LOCKBYTE) duke@435: ESelf->reset() ; duke@435: OrderAccess::fence() ; duke@435: duke@435: // Optional optimization ... try barging on the inner lock duke@435: if ((NativeMonitorFlags & 32) && CASPTR (&_OnDeck, NULL, UNS(Self)) == 0) { duke@435: goto OnDeck_LOOP ; duke@435: } duke@435: duke@435: if (AcquireOrPush (ESelf)) goto Exeunt ; duke@435: duke@435: // At any given time there is at most one ondeck thread. duke@435: // ondeck implies not resident on cxq and not resident on EntryList duke@435: // Only the OnDeck thread can try to acquire -- contended for -- the lock. duke@435: // CONSIDER: use Self->OnDeck instead of m->OnDeck. duke@435: // Deschedule Self so that others may run. duke@435: while (_OnDeck != ESelf) { duke@435: ParkCommon (ESelf, 0) ; duke@435: } duke@435: duke@435: // Self is now in the ONDECK position and will remain so until it duke@435: // manages to acquire the lock. duke@435: OnDeck_LOOP: duke@435: for (;;) { duke@435: assert (_OnDeck == ESelf, "invariant") ; duke@435: if (TrySpin (Self)) break ; duke@435: // CONSIDER: if ESelf->TryPark() && TryLock() break ... duke@435: // It's probably wise to spin only if we *actually* blocked duke@435: // CONSIDER: check the lockbyte, if it remains set then duke@435: // preemptively drain the cxq into the EntryList. duke@435: // The best place and time to perform queue operations -- lock metadata -- duke@435: // is _before having acquired the outer lock, while waiting for the lock to drop. duke@435: ParkCommon (ESelf, 0) ; duke@435: } duke@435: duke@435: assert (_OnDeck == ESelf, "invariant") ; duke@435: _OnDeck = NULL ; duke@435: duke@435: // Note that we current drop the inner lock (clear OnDeck) in the slow-path duke@435: // epilog immediately after having acquired the outer lock. duke@435: // But instead we could consider the following optimizations: duke@435: // A. Shift or defer dropping the inner lock until the subsequent IUnlock() operation. duke@435: // This might avoid potential reacquisition of the inner lock in IUlock(). duke@435: // B. While still holding the inner lock, attempt to opportunistically select duke@435: // and unlink the next ONDECK thread from the EntryList. duke@435: // If successful, set ONDECK to refer to that thread, otherwise clear ONDECK. duke@435: // It's critical that the select-and-unlink operation run in constant-time as duke@435: // it executes when holding the outer lock and may artificially increase the duke@435: // effective length of the critical section. duke@435: // Note that (A) and (B) are tantamount to succession by direct handoff for duke@435: // the inner lock. duke@435: goto Exeunt ; duke@435: } duke@435: duke@435: void Monitor::IUnlock (bool RelaxAssert) { duke@435: assert (ILocked(), "invariant") ; vladidan@3369: // Conceptually we need a MEMBAR #storestore|#loadstore barrier or fence immediately vladidan@3369: // before the store that releases the lock. Crucially, all the stores and loads in the vladidan@3369: // critical section must be globally visible before the store of 0 into the lock-word vladidan@3369: // that releases the lock becomes globally visible. That is, memory accesses in the vladidan@3369: // critical section should not be allowed to bypass or overtake the following ST that vladidan@3369: // releases the lock. As such, to prevent accesses within the critical section vladidan@3369: // from "leaking" out, we need a release fence between the critical section and the vladidan@3369: // store that releases the lock. In practice that release barrier is elided on vladidan@3369: // platforms with strong memory models such as TSO. vladidan@3369: // vladidan@3369: // Note that the OrderAccess::storeload() fence that appears after unlock store vladidan@3369: // provides for progress conditions and succession and is _not related to exclusion vladidan@3369: // safety or lock release consistency. vladidan@3369: OrderAccess::release_store(&_LockWord.Bytes[_LSBINDEX], 0); // drop outer lock vladidan@3369: duke@435: OrderAccess::storeload (); duke@435: ParkEvent * const w = _OnDeck ; duke@435: assert (RelaxAssert || w != Thread::current()->_MutexEvent, "invariant") ; duke@435: if (w != NULL) { duke@435: // Either we have a valid ondeck thread or ondeck is transiently "locked" duke@435: // by some exiting thread as it arranges for succession. The LSBit of duke@435: // OnDeck allows us to discriminate two cases. If the latter, the duke@435: // responsibility for progress and succession lies with that other thread. duke@435: // For good performance, we also depend on the fact that redundant unpark() duke@435: // operations are cheap. That is, repeated Unpark()ing of the ONDECK thread duke@435: // is inexpensive. This approach provides implicit futile wakeup throttling. duke@435: // Note that the referent "w" might be stale with respect to the lock. duke@435: // In that case the following unpark() is harmless and the worst that'll happen duke@435: // is a spurious return from a park() operation. Critically, if "w" _is stale, duke@435: // then progress is known to have occurred as that means the thread associated duke@435: // with "w" acquired the lock. In that case this thread need take no further duke@435: // action to guarantee progress. duke@435: if ((UNS(w) & _LBIT) == 0) w->unpark() ; duke@435: return ; duke@435: } duke@435: duke@435: intptr_t cxq = _LockWord.FullWord ; duke@435: if (((cxq & ~_LBIT)|UNS(_EntryList)) == 0) { duke@435: return ; // normal fast-path exit - cxq and EntryList both empty duke@435: } duke@435: if (cxq & _LBIT) { duke@435: // Optional optimization ... duke@435: // Some other thread acquired the lock in the window since this duke@435: // thread released it. Succession is now that thread's responsibility. duke@435: return ; duke@435: } duke@435: duke@435: Succession: duke@435: // Slow-path exit - this thread must ensure succession and progress. duke@435: // OnDeck serves as lock to protect cxq and EntryList. duke@435: // Only the holder of OnDeck can manipulate EntryList or detach the RATs from cxq. duke@435: // Avoid ABA - allow multiple concurrent producers (enqueue via push-CAS) duke@435: // but only one concurrent consumer (detacher of RATs). duke@435: // Consider protecting this critical section with schedctl on Solaris. duke@435: // Unlike a normal lock, however, the exiting thread "locks" OnDeck, duke@435: // picks a successor and marks that thread as OnDeck. That successor duke@435: // thread will then clear OnDeck once it eventually acquires the outer lock. duke@435: if (CASPTR (&_OnDeck, NULL, _LBIT) != UNS(NULL)) { duke@435: return ; duke@435: } duke@435: duke@435: ParkEvent * List = _EntryList ; duke@435: if (List != NULL) { duke@435: // Transfer the head of the EntryList to the OnDeck position. duke@435: // Once OnDeck, a thread stays OnDeck until it acquires the lock. duke@435: // For a given lock there is at most OnDeck thread at any one instant. duke@435: WakeOne: duke@435: assert (List == _EntryList, "invariant") ; duke@435: ParkEvent * const w = List ; duke@435: assert (RelaxAssert || w != Thread::current()->_MutexEvent, "invariant") ; duke@435: _EntryList = w->ListNext ; duke@435: // as a diagnostic measure consider setting w->_ListNext = BAD duke@435: assert (UNS(_OnDeck) == _LBIT, "invariant") ; duke@435: _OnDeck = w ; // pass OnDeck to w. duke@435: // w will clear OnDeck once it acquires the outer lock duke@435: duke@435: // Another optional optimization ... duke@435: // For heavily contended locks it's not uncommon that some other duke@435: // thread acquired the lock while this thread was arranging succession. duke@435: // Try to defer the unpark() operation - Delegate the responsibility duke@435: // for unpark()ing the OnDeck thread to the current or subsequent owners duke@435: // That is, the new owner is responsible for unparking the OnDeck thread. duke@435: OrderAccess::storeload() ; duke@435: cxq = _LockWord.FullWord ; duke@435: if (cxq & _LBIT) return ; duke@435: duke@435: w->unpark() ; duke@435: return ; duke@435: } duke@435: duke@435: cxq = _LockWord.FullWord ; duke@435: if ((cxq & ~_LBIT) != 0) { duke@435: // The EntryList is empty but the cxq is populated. duke@435: // drain RATs from cxq into EntryList duke@435: // Detach RATs segment with CAS and then merge into EntryList duke@435: for (;;) { duke@435: // optional optimization - if locked, the owner is responsible for succession duke@435: if (cxq & _LBIT) goto Punt ; duke@435: const intptr_t vfy = CASPTR (&_LockWord, cxq, cxq & _LBIT) ; duke@435: if (vfy == cxq) break ; duke@435: cxq = vfy ; duke@435: // Interference - LockWord changed - Just retry duke@435: // We can see concurrent interference from contending threads duke@435: // pushing themselves onto the cxq or from lock-unlock operations. duke@435: // From the perspective of this thread, EntryList is stable and duke@435: // the cxq is prepend-only -- the head is volatile but the interior duke@435: // of the cxq is stable. In theory if we encounter interference from threads duke@435: // pushing onto cxq we could simply break off the original cxq suffix and duke@435: // move that segment to the EntryList, avoiding a 2nd or multiple CAS attempts duke@435: // on the high-traffic LockWord variable. For instance lets say the cxq is "ABCD" duke@435: // when we first fetch cxq above. Between the fetch -- where we observed "A" duke@435: // -- and CAS -- where we attempt to CAS null over A -- "PQR" arrive, duke@435: // yielding cxq = "PQRABCD". In this case we could simply set A.ListNext duke@435: // null, leaving cxq = "PQRA" and transfer the "BCD" segment to the EntryList. duke@435: // Note too, that it's safe for this thread to traverse the cxq duke@435: // without taking any special concurrency precautions. duke@435: } duke@435: duke@435: // We don't currently reorder the cxq segment as we move it onto duke@435: // the EntryList, but it might make sense to reverse the order duke@435: // or perhaps sort by thread priority. See the comments in duke@435: // synchronizer.cpp objectMonitor::exit(). duke@435: assert (_EntryList == NULL, "invariant") ; duke@435: _EntryList = List = (ParkEvent *)(cxq & ~_LBIT) ; duke@435: assert (List != NULL, "invariant") ; duke@435: goto WakeOne ; duke@435: } duke@435: duke@435: // cxq|EntryList is empty. duke@435: // w == NULL implies that cxq|EntryList == NULL in the past. duke@435: // Possible race - rare inopportune interleaving. duke@435: // A thread could have added itself to cxq since this thread previously checked. duke@435: // Detect and recover by refetching cxq. duke@435: Punt: duke@435: assert (UNS(_OnDeck) == _LBIT, "invariant") ; duke@435: _OnDeck = NULL ; // Release inner lock. duke@435: OrderAccess::storeload(); // Dekker duality - pivot point duke@435: duke@435: // Resample LockWord/cxq to recover from possible race. duke@435: // For instance, while this thread T1 held OnDeck, some other thread T2 might duke@435: // acquire the outer lock. Another thread T3 might try to acquire the outer duke@435: // lock, but encounter contention and enqueue itself on cxq. T2 then drops the duke@435: // outer lock, but skips succession as this thread T1 still holds OnDeck. duke@435: // T1 is and remains responsible for ensuring succession of T3. duke@435: // duke@435: // Note that we don't need to recheck EntryList, just cxq. duke@435: // If threads moved onto EntryList since we dropped OnDeck duke@435: // that implies some other thread forced succession. duke@435: cxq = _LockWord.FullWord ; duke@435: if ((cxq & ~_LBIT) != 0 && (cxq & _LBIT) == 0) { duke@435: goto Succession ; // potential race -- re-run succession duke@435: } duke@435: return ; duke@435: } duke@435: duke@435: bool Monitor::notify() { duke@435: assert (_owner == Thread::current(), "invariant") ; duke@435: assert (ILocked(), "invariant") ; duke@435: if (_WaitSet == NULL) return true ; duke@435: NotifyCount ++ ; duke@435: duke@435: // Transfer one thread from the WaitSet to the EntryList or cxq. duke@435: // Currently we just unlink the head of the WaitSet and prepend to the cxq. duke@435: // And of course we could just unlink it and unpark it, too, but duke@435: // in that case it'd likely impale itself on the reentry. duke@435: Thread::muxAcquire (_WaitLock, "notify:WaitLock") ; duke@435: ParkEvent * nfy = _WaitSet ; duke@435: if (nfy != NULL) { // DCL idiom duke@435: _WaitSet = nfy->ListNext ; duke@435: assert (nfy->Notified == 0, "invariant") ; duke@435: // push nfy onto the cxq duke@435: for (;;) { duke@435: const intptr_t v = _LockWord.FullWord ; duke@435: assert ((v & 0xFF) == _LBIT, "invariant") ; duke@435: nfy->ListNext = (ParkEvent *)(v & ~_LBIT); duke@435: if (CASPTR (&_LockWord, v, UNS(nfy)|_LBIT) == v) break; duke@435: // interference - _LockWord changed -- just retry duke@435: } duke@435: // Note that setting Notified before pushing nfy onto the cxq is duke@435: // also legal and safe, but the safety properties are much more duke@435: // subtle, so for the sake of code stewardship ... duke@435: OrderAccess::fence() ; duke@435: nfy->Notified = 1; duke@435: } duke@435: Thread::muxRelease (_WaitLock) ; duke@435: if (nfy != NULL && (NativeMonitorFlags & 16)) { duke@435: // Experimental code ... light up the wakee in the hope that this thread (the owner) duke@435: // will drop the lock just about the time the wakee comes ONPROC. duke@435: nfy->unpark() ; duke@435: } duke@435: assert (ILocked(), "invariant") ; duke@435: return true ; duke@435: } duke@435: duke@435: // Currently notifyAll() transfers the waiters one-at-a-time from the waitset duke@435: // to the cxq. This could be done more efficiently with a single bulk en-mass transfer, duke@435: // but in practice notifyAll() for large #s of threads is rare and not time-critical. duke@435: // Beware too, that we invert the order of the waiters. Lets say that the duke@435: // waitset is "ABCD" and the cxq is "XYZ". After a notifyAll() the waitset duke@435: // will be empty and the cxq will be "DCBAXYZ". This is benign, of course. duke@435: duke@435: bool Monitor::notify_all() { duke@435: assert (_owner == Thread::current(), "invariant") ; duke@435: assert (ILocked(), "invariant") ; duke@435: while (_WaitSet != NULL) notify() ; duke@435: return true ; duke@435: } duke@435: duke@435: int Monitor::IWait (Thread * Self, jlong timo) { duke@435: assert (ILocked(), "invariant") ; duke@435: duke@435: // Phases: duke@435: // 1. Enqueue Self on WaitSet - currently prepend duke@435: // 2. unlock - drop the outer lock duke@435: // 3. wait for either notification or timeout duke@435: // 4. lock - reentry - reacquire the outer lock duke@435: duke@435: ParkEvent * const ESelf = Self->_MutexEvent ; duke@435: ESelf->Notified = 0 ; duke@435: ESelf->reset() ; duke@435: OrderAccess::fence() ; duke@435: duke@435: // Add Self to WaitSet duke@435: // Ideally only the holder of the outer lock would manipulate the WaitSet - duke@435: // That is, the outer lock would implicitly protect the WaitSet. duke@435: // But if a thread in wait() encounters a timeout it will need to dequeue itself duke@435: // from the WaitSet _before it becomes the owner of the lock. We need to dequeue duke@435: // as the ParkEvent -- which serves as a proxy for the thread -- can't reside duke@435: // on both the WaitSet and the EntryList|cxq at the same time.. That is, a thread duke@435: // on the WaitSet can't be allowed to compete for the lock until it has managed to duke@435: // unlink its ParkEvent from WaitSet. Thus the need for WaitLock. duke@435: // Contention on the WaitLock is minimal. duke@435: // duke@435: // Another viable approach would be add another ParkEvent, "WaitEvent" to the duke@435: // thread class. The WaitSet would be composed of WaitEvents. Only the duke@435: // owner of the outer lock would manipulate the WaitSet. A thread in wait() duke@435: // could then compete for the outer lock, and then, if necessary, unlink itself duke@435: // from the WaitSet only after having acquired the outer lock. More precisely, duke@435: // there would be no WaitLock. A thread in in wait() would enqueue its WaitEvent duke@435: // on the WaitSet; release the outer lock; wait for either notification or timeout; duke@435: // reacquire the inner lock; and then, if needed, unlink itself from the WaitSet. duke@435: // duke@435: // Alternatively, a 2nd set of list link fields in the ParkEvent might suffice. duke@435: // One set would be for the WaitSet and one for the EntryList. duke@435: // We could also deconstruct the ParkEvent into a "pure" event and add a duke@435: // new immortal/TSM "ListElement" class that referred to ParkEvents. duke@435: // In that case we could have one ListElement on the WaitSet and another duke@435: // on the EntryList, with both referring to the same pure Event. duke@435: duke@435: Thread::muxAcquire (_WaitLock, "wait:WaitLock:Add") ; duke@435: ESelf->ListNext = _WaitSet ; duke@435: _WaitSet = ESelf ; duke@435: Thread::muxRelease (_WaitLock) ; duke@435: duke@435: // Release the outer lock duke@435: // We call IUnlock (RelaxAssert=true) as a thread T1 might duke@435: // enqueue itself on the WaitSet, call IUnlock(), drop the lock, duke@435: // and then stall before it can attempt to wake a successor. duke@435: // Some other thread T2 acquires the lock, and calls notify(), moving duke@435: // T1 from the WaitSet to the cxq. T2 then drops the lock. T1 resumes, duke@435: // and then finds *itself* on the cxq. During the course of a normal duke@435: // IUnlock() call a thread should _never find itself on the EntryList duke@435: // or cxq, but in the case of wait() it's possible. duke@435: // See synchronizer.cpp objectMonitor::wait(). duke@435: IUnlock (true) ; duke@435: duke@435: // Wait for either notification or timeout duke@435: // Beware that in some circumstances we might propagate duke@435: // spurious wakeups back to the caller. duke@435: duke@435: for (;;) { duke@435: if (ESelf->Notified) break ; duke@435: int err = ParkCommon (ESelf, timo) ; duke@435: if (err == OS_TIMEOUT || (NativeMonitorFlags & 1)) break ; duke@435: } duke@435: duke@435: // Prepare for reentry - if necessary, remove ESelf from WaitSet duke@435: // ESelf can be: duke@435: // 1. Still on the WaitSet. This can happen if we exited the loop by timeout. duke@435: // 2. On the cxq or EntryList duke@435: // 3. Not resident on cxq, EntryList or WaitSet, but in the OnDeck position. duke@435: duke@435: OrderAccess::fence() ; duke@435: int WasOnWaitSet = 0 ; duke@435: if (ESelf->Notified == 0) { duke@435: Thread::muxAcquire (_WaitLock, "wait:WaitLock:remove") ; duke@435: if (ESelf->Notified == 0) { // DCL idiom duke@435: assert (_OnDeck != ESelf, "invariant") ; // can't be both OnDeck and on WaitSet duke@435: // ESelf is resident on the WaitSet -- unlink it. duke@435: // A doubly-linked list would be better here so we can unlink in constant-time. duke@435: // We have to unlink before we potentially recontend as ESelf might otherwise duke@435: // end up on the cxq|EntryList -- it can't be on two lists at once. duke@435: ParkEvent * p = _WaitSet ; duke@435: ParkEvent * q = NULL ; // classic q chases p duke@435: while (p != NULL && p != ESelf) { duke@435: q = p ; duke@435: p = p->ListNext ; duke@435: } duke@435: assert (p == ESelf, "invariant") ; duke@435: if (p == _WaitSet) { // found at head duke@435: assert (q == NULL, "invariant") ; duke@435: _WaitSet = p->ListNext ; duke@435: } else { // found in interior duke@435: assert (q->ListNext == p, "invariant") ; duke@435: q->ListNext = p->ListNext ; duke@435: } duke@435: WasOnWaitSet = 1 ; // We were *not* notified but instead encountered timeout duke@435: } duke@435: Thread::muxRelease (_WaitLock) ; duke@435: } duke@435: duke@435: // Reentry phase - reacquire the lock duke@435: if (WasOnWaitSet) { duke@435: // ESelf was previously on the WaitSet but we just unlinked it above duke@435: // because of a timeout. ESelf is not resident on any list and is not OnDeck duke@435: assert (_OnDeck != ESelf, "invariant") ; duke@435: ILock (Self) ; duke@435: } else { duke@435: // A prior notify() operation moved ESelf from the WaitSet to the cxq. duke@435: // ESelf is now on the cxq, EntryList or at the OnDeck position. duke@435: // The following fragment is extracted from Monitor::ILock() duke@435: for (;;) { duke@435: if (_OnDeck == ESelf && TrySpin(Self)) break ; duke@435: ParkCommon (ESelf, 0) ; duke@435: } duke@435: assert (_OnDeck == ESelf, "invariant") ; duke@435: _OnDeck = NULL ; duke@435: } duke@435: duke@435: assert (ILocked(), "invariant") ; duke@435: return WasOnWaitSet != 0 ; // return true IFF timeout duke@435: } duke@435: duke@435: duke@435: // ON THE VMTHREAD SNEAKING PAST HELD LOCKS: duke@435: // In particular, there are certain types of global lock that may be held duke@435: // by a Java thread while it is blocked at a safepoint but before it has duke@435: // written the _owner field. These locks may be sneakily acquired by the duke@435: // VM thread during a safepoint to avoid deadlocks. Alternatively, one should duke@435: // identify all such locks, and ensure that Java threads never block at duke@435: // safepoints while holding them (_no_safepoint_check_flag). While it duke@435: // seems as though this could increase the time to reach a safepoint duke@435: // (or at least increase the mean, if not the variance), the latter duke@435: // approach might make for a cleaner, more maintainable JVM design. duke@435: // duke@435: // Sneaking is vile and reprehensible and should be excised at the 1st duke@435: // opportunity. It's possible that the need for sneaking could be obviated duke@435: // as follows. Currently, a thread might (a) while TBIVM, call pthread_mutex_lock duke@435: // or ILock() thus acquiring the "physical" lock underlying Monitor/Mutex. duke@435: // (b) stall at the TBIVM exit point as a safepoint is in effect. Critically, duke@435: // it'll stall at the TBIVM reentry state transition after having acquired the duke@435: // underlying lock, but before having set _owner and having entered the actual duke@435: // critical section. The lock-sneaking facility leverages that fact and allowed the duke@435: // VM thread to logically acquire locks that had already be physically locked by mutators duke@435: // but where mutators were known blocked by the reentry thread state transition. duke@435: // duke@435: // If we were to modify the Monitor-Mutex so that TBIVM state transitions tightly duke@435: // wrapped calls to park(), then we could likely do away with sneaking. We'd duke@435: // decouple lock acquisition and parking. The critical invariant to eliminating duke@435: // sneaking is to ensure that we never "physically" acquire the lock while TBIVM. duke@435: // An easy way to accomplish this is to wrap the park calls in a narrow TBIVM jacket. duke@435: // One difficulty with this approach is that the TBIVM wrapper could recurse and duke@435: // call lock() deep from within a lock() call, while the MutexEvent was already enqueued. duke@435: // Using a stack (N=2 at minimum) of ParkEvents would take care of that problem. duke@435: // duke@435: // But of course the proper ultimate approach is to avoid schemes that require explicit duke@435: // sneaking or dependence on any any clever invariants or subtle implementation properties duke@435: // of Mutex-Monitor and instead directly address the underlying design flaw. duke@435: duke@435: void Monitor::lock (Thread * Self) { duke@435: #ifdef CHECK_UNHANDLED_OOPS duke@435: // Clear unhandled oops so we get a crash right away. Only clear for non-vm duke@435: // or GC threads. duke@435: if (Self->is_Java_thread()) { duke@435: Self->clear_unhandled_oops(); duke@435: } duke@435: #endif // CHECK_UNHANDLED_OOPS duke@435: duke@435: debug_only(check_prelock_state(Self)); duke@435: assert (_owner != Self , "invariant") ; duke@435: assert (_OnDeck != Self->_MutexEvent, "invariant") ; duke@435: duke@435: if (TryFast()) { duke@435: Exeunt: duke@435: assert (ILocked(), "invariant") ; duke@435: assert (owner() == NULL, "invariant"); duke@435: set_owner (Self); duke@435: return ; duke@435: } duke@435: duke@435: // The lock is contended ... duke@435: duke@435: bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint(); duke@435: if (can_sneak && _owner == NULL) { duke@435: // a java thread has locked the lock but has not entered the duke@435: // critical region -- let's just pretend we've locked the lock duke@435: // and go on. we note this with _snuck so we can also duke@435: // pretend to unlock when the time comes. duke@435: _snuck = true; duke@435: goto Exeunt ; duke@435: } duke@435: duke@435: // Try a brief spin to avoid passing thru thread state transition ... duke@435: if (TrySpin (Self)) goto Exeunt ; duke@435: duke@435: check_block_state(Self); duke@435: if (Self->is_Java_thread()) { duke@435: // Horribile dictu - we suffer through a state transition duke@435: assert(rank() > Mutex::special, "Potential deadlock with special or lesser rank mutex"); duke@435: ThreadBlockInVM tbivm ((JavaThread *) Self) ; duke@435: ILock (Self) ; duke@435: } else { duke@435: // Mirabile dictu duke@435: ILock (Self) ; duke@435: } duke@435: goto Exeunt ; duke@435: } duke@435: duke@435: void Monitor::lock() { duke@435: this->lock(Thread::current()); duke@435: } duke@435: duke@435: // Lock without safepoint check - a degenerate variant of lock(). duke@435: // Should ONLY be used by safepoint code and other code duke@435: // that is guaranteed not to block while running inside the VM. If this is called with duke@435: // thread state set to be in VM, the safepoint synchronization code will deadlock! duke@435: duke@435: void Monitor::lock_without_safepoint_check (Thread * Self) { duke@435: assert (_owner != Self, "invariant") ; duke@435: ILock (Self) ; duke@435: assert (_owner == NULL, "invariant"); duke@435: set_owner (Self); duke@435: } duke@435: duke@435: void Monitor::lock_without_safepoint_check () { duke@435: lock_without_safepoint_check (Thread::current()) ; duke@435: } duke@435: duke@435: duke@435: // Returns true if thread succeceed [sic] in grabbing the lock, otherwise false. duke@435: duke@435: bool Monitor::try_lock() { duke@435: Thread * const Self = Thread::current(); duke@435: debug_only(check_prelock_state(Self)); duke@435: // assert(!thread->is_inside_signal_handler(), "don't lock inside signal handler"); duke@435: duke@435: // Special case, where all Java threads are stopped. duke@435: // The lock may have been acquired but _owner is not yet set. duke@435: // In that case the VM thread can safely grab the lock. duke@435: // It strikes me this should appear _after the TryLock() fails, below. duke@435: bool can_sneak = Self->is_VM_thread() && SafepointSynchronize::is_at_safepoint(); duke@435: if (can_sneak && _owner == NULL) { duke@435: set_owner(Self); // Do not need to be atomic, since we are at a safepoint duke@435: _snuck = true; duke@435: return true; duke@435: } duke@435: duke@435: if (TryLock()) { duke@435: // We got the lock duke@435: assert (_owner == NULL, "invariant"); duke@435: set_owner (Self); duke@435: return true; duke@435: } duke@435: return false; duke@435: } duke@435: duke@435: void Monitor::unlock() { duke@435: assert (_owner == Thread::current(), "invariant") ; duke@435: assert (_OnDeck != Thread::current()->_MutexEvent , "invariant") ; duke@435: set_owner (NULL) ; duke@435: if (_snuck) { duke@435: assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak"); duke@435: _snuck = false; duke@435: return ; duke@435: } duke@435: IUnlock (false) ; duke@435: } duke@435: duke@435: // Yet another degenerate version of Monitor::lock() or lock_without_safepoint_check() duke@435: // jvm_raw_lock() and _unlock() can be called by non-Java threads via JVM_RawMonitorEnter. duke@435: // duke@435: // There's no expectation that JVM_RawMonitors will interoperate properly with the native duke@435: // Mutex-Monitor constructs. We happen to implement JVM_RawMonitors in terms of duke@435: // native Mutex-Monitors simply as a matter of convenience. A simple abstraction layer duke@435: // over a pthread_mutex_t would work equally as well, but require more platform-specific duke@435: // code -- a "PlatformMutex". Alternatively, a simply layer over muxAcquire-muxRelease duke@435: // would work too. duke@435: // duke@435: // Since the caller might be a foreign thread, we don't necessarily have a Thread.MutexEvent duke@435: // instance available. Instead, we transiently allocate a ParkEvent on-demand if duke@435: // we encounter contention. That ParkEvent remains associated with the thread duke@435: // until it manages to acquire the lock, at which time we return the ParkEvent duke@435: // to the global ParkEvent free list. This is correct and suffices for our purposes. duke@435: // duke@435: // Beware that the original jvm_raw_unlock() had a "_snuck" test but that duke@435: // jvm_raw_lock() didn't have the corresponding test. I suspect that's an duke@435: // oversight, but I've replicated the original suspect logic in the new code ... duke@435: duke@435: void Monitor::jvm_raw_lock() { duke@435: assert(rank() == native, "invariant"); duke@435: duke@435: if (TryLock()) { duke@435: Exeunt: duke@435: assert (ILocked(), "invariant") ; duke@435: assert (_owner == NULL, "invariant"); duke@435: // This can potentially be called by non-java Threads. Thus, the ThreadLocalStorage duke@435: // might return NULL. Don't call set_owner since it will break on an NULL owner duke@435: // Consider installing a non-null "ANON" distinguished value instead of just NULL. duke@435: _owner = ThreadLocalStorage::thread(); duke@435: return ; duke@435: } duke@435: duke@435: if (TrySpin(NULL)) goto Exeunt ; duke@435: duke@435: // slow-path - apparent contention duke@435: // Allocate a ParkEvent for transient use. duke@435: // The ParkEvent remains associated with this thread until duke@435: // the time the thread manages to acquire the lock. duke@435: ParkEvent * const ESelf = ParkEvent::Allocate(NULL) ; duke@435: ESelf->reset() ; duke@435: OrderAccess::storeload() ; duke@435: duke@435: // Either Enqueue Self on cxq or acquire the outer lock. duke@435: if (AcquireOrPush (ESelf)) { duke@435: ParkEvent::Release (ESelf) ; // surrender the ParkEvent duke@435: goto Exeunt ; duke@435: } duke@435: duke@435: // At any given time there is at most one ondeck thread. duke@435: // ondeck implies not resident on cxq and not resident on EntryList duke@435: // Only the OnDeck thread can try to acquire -- contended for -- the lock. duke@435: // CONSIDER: use Self->OnDeck instead of m->OnDeck. duke@435: for (;;) { duke@435: if (_OnDeck == ESelf && TrySpin(NULL)) break ; duke@435: ParkCommon (ESelf, 0) ; duke@435: } duke@435: duke@435: assert (_OnDeck == ESelf, "invariant") ; duke@435: _OnDeck = NULL ; duke@435: ParkEvent::Release (ESelf) ; // surrender the ParkEvent duke@435: goto Exeunt ; duke@435: } duke@435: duke@435: void Monitor::jvm_raw_unlock() { duke@435: // Nearly the same as Monitor::unlock() ... duke@435: // directly set _owner instead of using set_owner(null) duke@435: _owner = NULL ; duke@435: if (_snuck) { // ??? duke@435: assert(SafepointSynchronize::is_at_safepoint() && Thread::current()->is_VM_thread(), "sneak"); duke@435: _snuck = false; duke@435: return ; duke@435: } duke@435: IUnlock(false) ; duke@435: } duke@435: duke@435: bool Monitor::wait(bool no_safepoint_check, long timeout, bool as_suspend_equivalent) { duke@435: Thread * const Self = Thread::current() ; duke@435: assert (_owner == Self, "invariant") ; duke@435: assert (ILocked(), "invariant") ; duke@435: duke@435: // as_suspend_equivalent logically implies !no_safepoint_check duke@435: guarantee (!as_suspend_equivalent || !no_safepoint_check, "invariant") ; duke@435: // !no_safepoint_check logically implies java_thread duke@435: guarantee (no_safepoint_check || Self->is_Java_thread(), "invariant") ; duke@435: duke@435: #ifdef ASSERT duke@435: Monitor * least = get_least_ranked_lock_besides_this(Self->owned_locks()); duke@435: assert(least != this, "Specification of get_least_... call above"); duke@435: if (least != NULL && least->rank() <= special) { duke@435: tty->print("Attempting to wait on monitor %s/%d while holding" duke@435: " lock %s/%d -- possible deadlock", duke@435: name(), rank(), least->name(), least->rank()); duke@435: assert(false, "Shouldn't block(wait) while holding a lock of rank special"); duke@435: } duke@435: #endif // ASSERT duke@435: duke@435: int wait_status ; duke@435: // conceptually set the owner to NULL in anticipation of duke@435: // abdicating the lock in wait duke@435: set_owner(NULL); duke@435: if (no_safepoint_check) { duke@435: wait_status = IWait (Self, timeout) ; duke@435: } else { duke@435: assert (Self->is_Java_thread(), "invariant") ; duke@435: JavaThread *jt = (JavaThread *)Self; duke@435: duke@435: // Enter safepoint region - ornate and Rococo ... duke@435: ThreadBlockInVM tbivm(jt); duke@435: OSThreadWaitState osts(Self->osthread(), false /* not Object.wait() */); duke@435: duke@435: if (as_suspend_equivalent) { duke@435: jt->set_suspend_equivalent(); duke@435: // cleared by handle_special_suspend_equivalent_condition() or duke@435: // java_suspend_self() duke@435: } duke@435: duke@435: wait_status = IWait (Self, timeout) ; duke@435: duke@435: // were we externally suspended while we were waiting? duke@435: if (as_suspend_equivalent && jt->handle_special_suspend_equivalent_condition()) { duke@435: // Our event wait has finished and we own the lock, but duke@435: // while we were waiting another thread suspended us. We don't duke@435: // want to hold the lock while suspended because that duke@435: // would surprise the thread that suspended us. duke@435: assert (ILocked(), "invariant") ; duke@435: IUnlock (true) ; duke@435: jt->java_suspend_self(); duke@435: ILock (Self) ; duke@435: assert (ILocked(), "invariant") ; duke@435: } duke@435: } duke@435: duke@435: // Conceptually reestablish ownership of the lock. duke@435: // The "real" lock -- the LockByte -- was reacquired by IWait(). duke@435: assert (ILocked(), "invariant") ; duke@435: assert (_owner == NULL, "invariant") ; duke@435: set_owner (Self) ; duke@435: return wait_status != 0 ; // return true IFF timeout duke@435: } duke@435: duke@435: Monitor::~Monitor() { duke@435: assert ((UNS(_owner)|UNS(_LockWord.FullWord)|UNS(_EntryList)|UNS(_WaitSet)|UNS(_OnDeck)) == 0, "") ; duke@435: } duke@435: xlu@490: void Monitor::ClearMonitor (Monitor * m, const char *name) { duke@435: m->_owner = NULL ; duke@435: m->_snuck = false ; xlu@490: if (name == NULL) { xlu@490: strcpy(m->_name, "UNKNOWN") ; xlu@490: } else { xlu@490: strncpy(m->_name, name, MONITOR_NAME_LEN - 1); xlu@490: m->_name[MONITOR_NAME_LEN - 1] = '\0'; xlu@490: } duke@435: m->_LockWord.FullWord = 0 ; duke@435: m->_EntryList = NULL ; duke@435: m->_OnDeck = NULL ; duke@435: m->_WaitSet = NULL ; duke@435: m->_WaitLock[0] = 0 ; duke@435: } duke@435: duke@435: Monitor::Monitor() { ClearMonitor(this); } duke@435: duke@435: Monitor::Monitor (int Rank, const char * name, bool allow_vm_block) { xlu@490: ClearMonitor (this, name) ; duke@435: #ifdef ASSERT duke@435: _allow_vm_block = allow_vm_block; duke@435: _rank = Rank ; duke@435: #endif duke@435: } duke@435: duke@435: Mutex::~Mutex() { duke@435: assert ((UNS(_owner)|UNS(_LockWord.FullWord)|UNS(_EntryList)|UNS(_WaitSet)|UNS(_OnDeck)) == 0, "") ; duke@435: } duke@435: duke@435: Mutex::Mutex (int Rank, const char * name, bool allow_vm_block) { xlu@490: ClearMonitor ((Monitor *) this, name) ; duke@435: #ifdef ASSERT duke@435: _allow_vm_block = allow_vm_block; duke@435: _rank = Rank ; duke@435: #endif duke@435: } duke@435: duke@435: bool Monitor::owned_by_self() const { duke@435: bool ret = _owner == Thread::current(); duke@435: assert (!ret || _LockWord.Bytes[_LSBINDEX] != 0, "invariant") ; duke@435: return ret; duke@435: } duke@435: duke@435: void Monitor::print_on_error(outputStream* st) const { duke@435: st->print("[" PTR_FORMAT, this); duke@435: st->print("] %s", _name); duke@435: st->print(" - owner thread: " PTR_FORMAT, _owner); duke@435: } duke@435: duke@435: duke@435: duke@435: duke@435: // ---------------------------------------------------------------------------------- duke@435: // Non-product code duke@435: duke@435: #ifndef PRODUCT duke@435: void Monitor::print_on(outputStream* st) const { duke@435: st->print_cr("Mutex: [0x%lx/0x%lx] %s - owner: 0x%lx", this, _LockWord.FullWord, _name, _owner); duke@435: } duke@435: #endif duke@435: duke@435: #ifndef PRODUCT duke@435: #ifdef ASSERT duke@435: Monitor * Monitor::get_least_ranked_lock(Monitor * locks) { duke@435: Monitor *res, *tmp; duke@435: for (res = tmp = locks; tmp != NULL; tmp = tmp->next()) { duke@435: if (tmp->rank() < res->rank()) { duke@435: res = tmp; duke@435: } duke@435: } duke@435: if (!SafepointSynchronize::is_at_safepoint()) { duke@435: // In this case, we expect the held locks to be duke@435: // in increasing rank order (modulo any native ranks) duke@435: for (tmp = locks; tmp != NULL; tmp = tmp->next()) { duke@435: if (tmp->next() != NULL) { duke@435: assert(tmp->rank() == Mutex::native || duke@435: tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?"); duke@435: } duke@435: } duke@435: } duke@435: return res; duke@435: } duke@435: duke@435: Monitor* Monitor::get_least_ranked_lock_besides_this(Monitor* locks) { duke@435: Monitor *res, *tmp; duke@435: for (res = NULL, tmp = locks; tmp != NULL; tmp = tmp->next()) { duke@435: if (tmp != this && (res == NULL || tmp->rank() < res->rank())) { duke@435: res = tmp; duke@435: } duke@435: } duke@435: if (!SafepointSynchronize::is_at_safepoint()) { duke@435: // In this case, we expect the held locks to be duke@435: // in increasing rank order (modulo any native ranks) duke@435: for (tmp = locks; tmp != NULL; tmp = tmp->next()) { duke@435: if (tmp->next() != NULL) { duke@435: assert(tmp->rank() == Mutex::native || duke@435: tmp->rank() <= tmp->next()->rank(), "mutex rank anomaly?"); duke@435: } duke@435: } duke@435: } duke@435: return res; duke@435: } duke@435: duke@435: duke@435: bool Monitor::contains(Monitor* locks, Monitor * lock) { duke@435: for (; locks != NULL; locks = locks->next()) { duke@435: if (locks == lock) duke@435: return true; duke@435: } duke@435: return false; duke@435: } duke@435: #endif duke@435: duke@435: // Called immediately after lock acquisition or release as a diagnostic duke@435: // to track the lock-set of the thread and test for rank violations that duke@435: // might indicate exposure to deadlock. duke@435: // Rather like an EventListener for _owner (:>). duke@435: duke@435: void Monitor::set_owner_implementation(Thread *new_owner) { duke@435: // This function is solely responsible for maintaining duke@435: // and checking the invariant that threads and locks duke@435: // are in a 1/N relation, with some some locks unowned. duke@435: // It uses the Mutex::_owner, Mutex::_next, and duke@435: // Thread::_owned_locks fields, and no other function duke@435: // changes those fields. duke@435: // It is illegal to set the mutex from one non-NULL duke@435: // owner to another--it must be owned by NULL as an duke@435: // intermediate state. duke@435: duke@435: if (new_owner != NULL) { duke@435: // the thread is acquiring this lock duke@435: duke@435: assert(new_owner == Thread::current(), "Should I be doing this?"); duke@435: assert(_owner == NULL, "setting the owner thread of an already owned mutex"); duke@435: _owner = new_owner; // set the owner duke@435: duke@435: // link "this" into the owned locks list duke@435: duke@435: #ifdef ASSERT // Thread::_owned_locks is under the same ifdef duke@435: Monitor* locks = get_least_ranked_lock(new_owner->owned_locks()); duke@435: // Mutex::set_owner_implementation is a friend of Thread duke@435: duke@435: assert(this->rank() >= 0, "bad lock rank"); duke@435: duke@435: // Deadlock avoidance rules require us to acquire Mutexes only in duke@435: // a global total order. For example m1 is the lowest ranked mutex duke@435: // that the thread holds and m2 is the mutex the thread is trying duke@435: // to acquire, then deadlock avoidance rules require that the rank duke@435: // of m2 be less than the rank of m1. duke@435: // The rank Mutex::native is an exception in that it is not subject duke@435: // to the verification rules. duke@435: // Here are some further notes relating to mutex acquisition anomalies: duke@435: // . under Solaris, the interrupt lock gets acquired when doing duke@435: // profiling, so any lock could be held. duke@435: // . it is also ok to acquire Safepoint_lock at the very end while we duke@435: // already hold Terminator_lock - may happen because of periodic safepoints duke@435: if (this->rank() != Mutex::native && duke@435: this->rank() != Mutex::suspend_resume && duke@435: locks != NULL && locks->rank() <= this->rank() && duke@435: !SafepointSynchronize::is_at_safepoint() && duke@435: this != Interrupt_lock && this != ProfileVM_lock && duke@435: !(this == Safepoint_lock && contains(locks, Terminator_lock) && duke@435: SafepointSynchronize::is_synchronizing())) { duke@435: new_owner->print_owned_locks(); jcoomes@1845: fatal(err_msg("acquiring lock %s/%d out of order with lock %s/%d -- " jcoomes@1845: "possible deadlock", this->name(), this->rank(), jcoomes@1845: locks->name(), locks->rank())); duke@435: } duke@435: duke@435: this->_next = new_owner->_owned_locks; duke@435: new_owner->_owned_locks = this; duke@435: #endif duke@435: duke@435: } else { duke@435: // the thread is releasing this lock duke@435: duke@435: Thread* old_owner = _owner; duke@435: debug_only(_last_owner = old_owner); duke@435: duke@435: assert(old_owner != NULL, "removing the owner thread of an unowned mutex"); duke@435: assert(old_owner == Thread::current(), "removing the owner thread of an unowned mutex"); duke@435: duke@435: _owner = NULL; // set the owner duke@435: duke@435: #ifdef ASSERT duke@435: Monitor *locks = old_owner->owned_locks(); duke@435: duke@435: // remove "this" from the owned locks list duke@435: duke@435: Monitor *prev = NULL; duke@435: bool found = false; duke@435: for (; locks != NULL; prev = locks, locks = locks->next()) { duke@435: if (locks == this) { duke@435: found = true; duke@435: break; duke@435: } duke@435: } duke@435: assert(found, "Removing a lock not owned"); duke@435: if (prev == NULL) { duke@435: old_owner->_owned_locks = _next; duke@435: } else { duke@435: prev->_next = _next; duke@435: } duke@435: _next = NULL; duke@435: #endif duke@435: } duke@435: } duke@435: duke@435: duke@435: // Factored out common sanity checks for locking mutex'es. Used by lock() and try_lock() duke@435: void Monitor::check_prelock_state(Thread *thread) { duke@435: assert((!thread->is_Java_thread() || ((JavaThread *)thread)->thread_state() == _thread_in_vm) duke@435: || rank() == Mutex::special, "wrong thread state for using locks"); duke@435: if (StrictSafepointChecks) { duke@435: if (thread->is_VM_thread() && !allow_vm_block()) { jcoomes@1845: fatal(err_msg("VM thread using lock %s (not allowed to block on)", jcoomes@1845: name())); duke@435: } duke@435: debug_only(if (rank() != Mutex::special) \ duke@435: thread->check_for_valid_safepoint_state(false);) duke@435: } rbackman@5424: if (thread->is_Watcher_thread()) { rbackman@5424: assert(!WatcherThread::watcher_thread()->has_crash_protection(), rbackman@5424: "locking not allowed when crash protection is set"); rbackman@5424: } duke@435: } duke@435: duke@435: void Monitor::check_block_state(Thread *thread) { duke@435: if (!_allow_vm_block && thread->is_VM_thread()) { duke@435: warning("VM thread blocked on lock"); duke@435: print(); duke@435: BREAKPOINT; duke@435: } duke@435: assert(_owner != thread, "deadlock: blocking on monitor owned by current thread"); duke@435: } duke@435: duke@435: #endif // PRODUCT