aoqi@0: /* dbuck@8067: * Copyright (c) 1998, 2015, Oracle and/or its affiliates. All rights reserved. aoqi@0: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. aoqi@0: * aoqi@0: * This code is free software; you can redistribute it and/or modify it aoqi@0: * under the terms of the GNU General Public License version 2 only, as aoqi@0: * published by the Free Software Foundation. aoqi@0: * aoqi@0: * This code is distributed in the hope that it will be useful, but WITHOUT aoqi@0: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or aoqi@0: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License aoqi@0: * version 2 for more details (a copy is included in the LICENSE file that aoqi@0: * accompanied this code). aoqi@0: * aoqi@0: * You should have received a copy of the GNU General Public License version aoqi@0: * 2 along with this work; if not, write to the Free Software Foundation, aoqi@0: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. aoqi@0: * aoqi@0: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA aoqi@0: * or visit www.oracle.com if you need additional information or have any aoqi@0: * questions. aoqi@0: * aoqi@0: */ aoqi@0: aoqi@0: #include "precompiled.hpp" aoqi@0: #include "classfile/vmSymbols.hpp" aoqi@0: #include "memory/resourceArea.hpp" aoqi@0: #include "oops/markOop.hpp" aoqi@0: #include "oops/oop.inline.hpp" aoqi@0: #include "runtime/handles.inline.hpp" aoqi@0: #include "runtime/interfaceSupport.hpp" aoqi@0: #include "runtime/mutexLocker.hpp" aoqi@0: #include "runtime/objectMonitor.hpp" aoqi@0: #include "runtime/objectMonitor.inline.hpp" goetz@6911: #include "runtime/orderAccess.inline.hpp" aoqi@0: #include "runtime/osThread.hpp" aoqi@0: #include "runtime/stubRoutines.hpp" aoqi@0: #include "runtime/thread.inline.hpp" aoqi@0: #include "services/threadService.hpp" aoqi@0: #include "trace/tracing.hpp" aoqi@0: #include "trace/traceMacros.hpp" aoqi@0: #include "utilities/dtrace.hpp" aoqi@0: #include "utilities/macros.hpp" aoqi@0: #include "utilities/preserveException.hpp" aoqi@0: #ifdef TARGET_OS_FAMILY_linux aoqi@0: # include "os_linux.inline.hpp" aoqi@0: #endif aoqi@0: #ifdef TARGET_OS_FAMILY_solaris aoqi@0: # include "os_solaris.inline.hpp" aoqi@0: #endif aoqi@0: #ifdef TARGET_OS_FAMILY_windows aoqi@0: # include "os_windows.inline.hpp" aoqi@0: #endif aoqi@0: #ifdef TARGET_OS_FAMILY_bsd aoqi@0: # include "os_bsd.inline.hpp" aoqi@0: #endif aoqi@0: aoqi@0: #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64) aoqi@0: // Need to inhibit inlining for older versions of GCC to avoid build-time failures aoqi@0: #define ATTR __attribute__((noinline)) aoqi@0: #else aoqi@0: #define ATTR aoqi@0: #endif aoqi@0: aoqi@0: aoqi@0: #ifdef DTRACE_ENABLED aoqi@0: aoqi@0: // Only bother with this argument setup if dtrace is available aoqi@0: // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. aoqi@0: aoqi@0: aoqi@0: #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ aoqi@0: char* bytes = NULL; \ aoqi@0: int len = 0; \ aoqi@0: jlong jtid = SharedRuntime::get_java_tid(thread); \ aoqi@0: Symbol* klassname = ((oop)obj)->klass()->name(); \ aoqi@0: if (klassname != NULL) { \ aoqi@0: bytes = (char*)klassname->bytes(); \ aoqi@0: len = klassname->utf8_length(); \ aoqi@0: } aoqi@0: aoqi@0: #ifndef USDT2 aoqi@0: aoqi@0: HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, aoqi@0: jlong, uintptr_t, char*, int); aoqi@0: HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, aoqi@0: jlong, uintptr_t, char*, int); aoqi@0: HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, aoqi@0: jlong, uintptr_t, char*, int); aoqi@0: HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, aoqi@0: jlong, uintptr_t, char*, int); aoqi@0: HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, aoqi@0: jlong, uintptr_t, char*, int); aoqi@0: aoqi@0: #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ aoqi@0: { \ aoqi@0: if (DTraceMonitorProbes) { \ aoqi@0: DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ aoqi@0: HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ aoqi@0: (monitor), bytes, len, (millis)); \ aoqi@0: } \ aoqi@0: } aoqi@0: aoqi@0: #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ aoqi@0: { \ aoqi@0: if (DTraceMonitorProbes) { \ aoqi@0: DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ aoqi@0: HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ aoqi@0: (uintptr_t)(monitor), bytes, len); \ aoqi@0: } \ aoqi@0: } aoqi@0: aoqi@0: #else /* USDT2 */ aoqi@0: aoqi@0: #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ aoqi@0: { \ aoqi@0: if (DTraceMonitorProbes) { \ aoqi@0: DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ aoqi@0: HOTSPOT_MONITOR_WAIT(jtid, \ aoqi@0: (monitor), bytes, len, (millis)); \ aoqi@0: } \ aoqi@0: } aoqi@0: aoqi@0: #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER aoqi@0: #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED aoqi@0: #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT aoqi@0: #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY aoqi@0: #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL aoqi@0: aoqi@0: #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ aoqi@0: { \ aoqi@0: if (DTraceMonitorProbes) { \ aoqi@0: DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ aoqi@0: HOTSPOT_MONITOR_##probe(jtid, \ aoqi@0: (uintptr_t)(monitor), bytes, len); \ aoqi@0: } \ aoqi@0: } aoqi@0: aoqi@0: #endif /* USDT2 */ aoqi@0: #else // ndef DTRACE_ENABLED aoqi@0: aoqi@0: #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} aoqi@0: #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} aoqi@0: aoqi@0: #endif // ndef DTRACE_ENABLED aoqi@0: aoqi@0: // Tunables ... aoqi@0: // The knob* variables are effectively final. Once set they should aoqi@0: // never be modified hence. Consider using __read_mostly with GCC. aoqi@0: aoqi@0: int ObjectMonitor::Knob_Verbose = 0 ; aoqi@0: int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - aoqi@0: static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins aoqi@0: static int Knob_HandOff = 0 ; aoqi@0: static int Knob_ReportSettings = 0 ; aoqi@0: aoqi@0: static int Knob_SpinBase = 0 ; // Floor AKA SpinMin aoqi@0: static int Knob_SpinBackOff = 0 ; // spin-loop backoff aoqi@0: static int Knob_CASPenalty = -1 ; // Penalty for failed CAS aoqi@0: static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change aoqi@0: static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field aoqi@0: static int Knob_SpinEarly = 1 ; aoqi@0: static int Knob_SuccEnabled = 1 ; // futile wake throttling aoqi@0: static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one aoqi@0: static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs aoqi@0: static int Knob_Bonus = 100 ; // spin success bonus aoqi@0: static int Knob_BonusB = 100 ; // spin success bonus aoqi@0: static int Knob_Penalty = 200 ; // spin failure penalty aoqi@0: static int Knob_Poverty = 1000 ; aoqi@0: static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() aoqi@0: static int Knob_FixedSpin = 0 ; aoqi@0: static int Knob_OState = 3 ; // Spinner checks thread state of _owner aoqi@0: static int Knob_UsePause = 1 ; aoqi@0: static int Knob_ExitPolicy = 0 ; aoqi@0: static int Knob_PreSpin = 10 ; // 20-100 likely better aoqi@0: static int Knob_ResetEvent = 0 ; aoqi@0: static int BackOffMask = 0 ; aoqi@0: aoqi@0: static int Knob_FastHSSEC = 0 ; aoqi@0: static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee aoqi@0: static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline aoqi@0: static volatile int InitDone = 0 ; aoqi@0: aoqi@0: #define TrySpin TrySpin_VaryDuration aoqi@0: aoqi@0: // ----------------------------------------------------------------------------- aoqi@0: // Theory of operations -- Monitors lists, thread residency, etc: aoqi@0: // aoqi@0: // * A thread acquires ownership of a monitor by successfully aoqi@0: // CAS()ing the _owner field from null to non-null. aoqi@0: // aoqi@0: // * Invariant: A thread appears on at most one monitor list -- aoqi@0: // cxq, EntryList or WaitSet -- at any one time. aoqi@0: // aoqi@0: // * Contending threads "push" themselves onto the cxq with CAS aoqi@0: // and then spin/park. aoqi@0: // aoqi@0: // * After a contending thread eventually acquires the lock it must aoqi@0: // dequeue itself from either the EntryList or the cxq. aoqi@0: // aoqi@0: // * The exiting thread identifies and unparks an "heir presumptive" aoqi@0: // tentative successor thread on the EntryList. Critically, the aoqi@0: // exiting thread doesn't unlink the successor thread from the EntryList. aoqi@0: // After having been unparked, the wakee will recontend for ownership of aoqi@0: // the monitor. The successor (wakee) will either acquire the lock or aoqi@0: // re-park itself. aoqi@0: // aoqi@0: // Succession is provided for by a policy of competitive handoff. aoqi@0: // The exiting thread does _not_ grant or pass ownership to the aoqi@0: // successor thread. (This is also referred to as "handoff" succession"). aoqi@0: // Instead the exiting thread releases ownership and possibly wakes aoqi@0: // a successor, so the successor can (re)compete for ownership of the lock. aoqi@0: // If the EntryList is empty but the cxq is populated the exiting aoqi@0: // thread will drain the cxq into the EntryList. It does so by aoqi@0: // by detaching the cxq (installing null with CAS) and folding aoqi@0: // the threads from the cxq into the EntryList. The EntryList is aoqi@0: // doubly linked, while the cxq is singly linked because of the aoqi@0: // CAS-based "push" used to enqueue recently arrived threads (RATs). aoqi@0: // aoqi@0: // * Concurrency invariants: aoqi@0: // aoqi@0: // -- only the monitor owner may access or mutate the EntryList. aoqi@0: // The mutex property of the monitor itself protects the EntryList aoqi@0: // from concurrent interference. aoqi@0: // -- Only the monitor owner may detach the cxq. aoqi@0: // aoqi@0: // * The monitor entry list operations avoid locks, but strictly speaking aoqi@0: // they're not lock-free. Enter is lock-free, exit is not. dbuck@8067: // For a description of 'Methods and apparatus providing non-blocking access dbuck@8067: // to a resource,' see U.S. Pat. No. 7844973. aoqi@0: // aoqi@0: // * The cxq can have multiple concurrent "pushers" but only one concurrent aoqi@0: // detaching thread. This mechanism is immune from the ABA corruption. aoqi@0: // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. aoqi@0: // aoqi@0: // * Taken together, the cxq and the EntryList constitute or form a aoqi@0: // single logical queue of threads stalled trying to acquire the lock. aoqi@0: // We use two distinct lists to improve the odds of a constant-time aoqi@0: // dequeue operation after acquisition (in the ::enter() epilog) and aoqi@0: // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). aoqi@0: // A key desideratum is to minimize queue & monitor metadata manipulation aoqi@0: // that occurs while holding the monitor lock -- that is, we want to aoqi@0: // minimize monitor lock holds times. Note that even a small amount of aoqi@0: // fixed spinning will greatly reduce the # of enqueue-dequeue operations aoqi@0: // on EntryList|cxq. That is, spinning relieves contention on the "inner" aoqi@0: // locks and monitor metadata. aoqi@0: // aoqi@0: // Cxq points to the the set of Recently Arrived Threads attempting entry. aoqi@0: // Because we push threads onto _cxq with CAS, the RATs must take the form of aoqi@0: // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when aoqi@0: // the unlocking thread notices that EntryList is null but _cxq is != null. aoqi@0: // aoqi@0: // The EntryList is ordered by the prevailing queue discipline and aoqi@0: // can be organized in any convenient fashion, such as a doubly-linked list or aoqi@0: // a circular doubly-linked list. Critically, we want insert and delete operations aoqi@0: // to operate in constant-time. If we need a priority queue then something akin aoqi@0: // to Solaris' sleepq would work nicely. Viz., aoqi@0: // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. aoqi@0: // Queue discipline is enforced at ::exit() time, when the unlocking thread aoqi@0: // drains the cxq into the EntryList, and orders or reorders the threads on the aoqi@0: // EntryList accordingly. aoqi@0: // aoqi@0: // Barring "lock barging", this mechanism provides fair cyclic ordering, aoqi@0: // somewhat similar to an elevator-scan. aoqi@0: // aoqi@0: // * The monitor synchronization subsystem avoids the use of native aoqi@0: // synchronization primitives except for the narrow platform-specific aoqi@0: // park-unpark abstraction. See the comments in os_solaris.cpp regarding aoqi@0: // the semantics of park-unpark. Put another way, this monitor implementation aoqi@0: // depends only on atomic operations and park-unpark. The monitor subsystem aoqi@0: // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the aoqi@0: // underlying OS manages the READY<->RUN transitions. aoqi@0: // aoqi@0: // * Waiting threads reside on the WaitSet list -- wait() puts aoqi@0: // the caller onto the WaitSet. aoqi@0: // aoqi@0: // * notify() or notifyAll() simply transfers threads from the WaitSet to aoqi@0: // either the EntryList or cxq. Subsequent exit() operations will aoqi@0: // unpark the notifyee. Unparking a notifee in notify() is inefficient - aoqi@0: // it's likely the notifyee would simply impale itself on the lock held aoqi@0: // by the notifier. aoqi@0: // aoqi@0: // * An interesting alternative is to encode cxq as (List,LockByte) where aoqi@0: // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary aoqi@0: // variable, like _recursions, in the scheme. The threads or Events that form aoqi@0: // the list would have to be aligned in 256-byte addresses. A thread would aoqi@0: // try to acquire the lock or enqueue itself with CAS, but exiting threads aoqi@0: // could use a 1-0 protocol and simply STB to set the LockByte to 0. aoqi@0: // Note that is is *not* word-tearing, but it does presume that full-word aoqi@0: // CAS operations are coherent with intermix with STB operations. That's true aoqi@0: // on most common processors. aoqi@0: // aoqi@0: // * See also http://blogs.sun.com/dave aoqi@0: aoqi@0: aoqi@0: // ----------------------------------------------------------------------------- aoqi@0: // Enter support aoqi@0: aoqi@0: bool ObjectMonitor::try_enter(Thread* THREAD) { aoqi@0: if (THREAD != _owner) { aoqi@0: if (THREAD->is_lock_owned ((address)_owner)) { aoqi@0: assert(_recursions == 0, "internal state error"); aoqi@0: _owner = THREAD ; aoqi@0: _recursions = 1 ; aoqi@0: OwnerIsThread = 1 ; aoqi@0: return true; aoqi@0: } aoqi@0: if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { aoqi@0: return false; aoqi@0: } aoqi@0: return true; aoqi@0: } else { aoqi@0: _recursions++; aoqi@0: return true; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: void ATTR ObjectMonitor::enter(TRAPS) { aoqi@0: // The following code is ordered to check the most common cases first aoqi@0: // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. aoqi@0: Thread * const Self = THREAD ; aoqi@0: void * cur ; aoqi@0: aoqi@0: cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; aoqi@0: if (cur == NULL) { aoqi@0: // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. aoqi@0: assert (_recursions == 0 , "invariant") ; aoqi@0: assert (_owner == Self, "invariant") ; aoqi@0: // CONSIDER: set or assert OwnerIsThread == 1 aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: if (cur == Self) { aoqi@0: // TODO-FIXME: check for integer overflow! BUGID 6557169. aoqi@0: _recursions ++ ; aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: if (Self->is_lock_owned ((address)cur)) { aoqi@0: assert (_recursions == 0, "internal state error"); aoqi@0: _recursions = 1 ; aoqi@0: // Commute owner from a thread-specific on-stack BasicLockObject address to aoqi@0: // a full-fledged "Thread *". aoqi@0: _owner = Self ; aoqi@0: OwnerIsThread = 1 ; aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: // We've encountered genuine contention. aoqi@0: assert (Self->_Stalled == 0, "invariant") ; aoqi@0: Self->_Stalled = intptr_t(this) ; aoqi@0: aoqi@0: // Try one round of spinning *before* enqueueing Self aoqi@0: // and before going through the awkward and expensive state aoqi@0: // transitions. The following spin is strictly optional ... aoqi@0: // Note that if we acquire the monitor from an initial spin aoqi@0: // we forgo posting JVMTI events and firing DTRACE probes. aoqi@0: if (Knob_SpinEarly && TrySpin (Self) > 0) { aoqi@0: assert (_owner == Self , "invariant") ; aoqi@0: assert (_recursions == 0 , "invariant") ; aoqi@0: assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; aoqi@0: Self->_Stalled = 0 ; aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: assert (_owner != Self , "invariant") ; aoqi@0: assert (_succ != Self , "invariant") ; aoqi@0: assert (Self->is_Java_thread() , "invariant") ; aoqi@0: JavaThread * jt = (JavaThread *) Self ; aoqi@0: assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; aoqi@0: assert (jt->thread_state() != _thread_blocked , "invariant") ; aoqi@0: assert (this->object() != NULL , "invariant") ; aoqi@0: assert (_count >= 0, "invariant") ; aoqi@0: aoqi@0: // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). aoqi@0: // Ensure the object-monitor relationship remains stable while there's contention. aoqi@0: Atomic::inc_ptr(&_count); aoqi@0: aoqi@0: EventJavaMonitorEnter event; aoqi@0: aoqi@0: { // Change java thread status to indicate blocked on monitor enter. aoqi@0: JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); aoqi@0: dbuck@8887: Self->set_current_pending_monitor(this); dbuck@8887: aoqi@0: DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); aoqi@0: if (JvmtiExport::should_post_monitor_contended_enter()) { aoqi@0: JvmtiExport::post_monitor_contended_enter(jt, this); aoqi@0: aoqi@0: // The current thread does not yet own the monitor and does not aoqi@0: // yet appear on any queues that would get it made the successor. aoqi@0: // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event aoqi@0: // handler cannot accidentally consume an unpark() meant for the aoqi@0: // ParkEvent associated with this ObjectMonitor. aoqi@0: } aoqi@0: aoqi@0: OSThreadContendState osts(Self->osthread()); aoqi@0: ThreadBlockInVM tbivm(jt); aoqi@0: aoqi@0: // TODO-FIXME: change the following for(;;) loop to straight-line code. aoqi@0: for (;;) { aoqi@0: jt->set_suspend_equivalent(); aoqi@0: // cleared by handle_special_suspend_equivalent_condition() aoqi@0: // or java_suspend_self() aoqi@0: aoqi@0: EnterI (THREAD) ; aoqi@0: aoqi@0: if (!ExitSuspendEquivalent(jt)) break ; aoqi@0: aoqi@0: // aoqi@0: // We have acquired the contended monitor, but while we were aoqi@0: // waiting another thread suspended us. We don't want to enter aoqi@0: // the monitor while suspended because that would surprise the aoqi@0: // thread that suspended us. aoqi@0: // aoqi@0: _recursions = 0 ; aoqi@0: _succ = NULL ; aoqi@0: exit (false, Self) ; aoqi@0: aoqi@0: jt->java_suspend_self(); aoqi@0: } aoqi@0: Self->set_current_pending_monitor(NULL); aoqi@0: aoqi@0: // We cleared the pending monitor info since we've just gotten past aoqi@0: // the enter-check-for-suspend dance and we now own the monitor free aoqi@0: // and clear, i.e., it is no longer pending. The ThreadBlockInVM aoqi@0: // destructor can go to a safepoint at the end of this block. If we aoqi@0: // do a thread dump during that safepoint, then this thread will show aoqi@0: // as having "-locked" the monitor, but the OS and java.lang.Thread aoqi@0: // states will still report that the thread is blocked trying to aoqi@0: // acquire it. aoqi@0: } aoqi@0: aoqi@0: Atomic::dec_ptr(&_count); aoqi@0: assert (_count >= 0, "invariant") ; aoqi@0: Self->_Stalled = 0 ; aoqi@0: aoqi@0: // Must either set _recursions = 0 or ASSERT _recursions == 0. aoqi@0: assert (_recursions == 0 , "invariant") ; aoqi@0: assert (_owner == Self , "invariant") ; aoqi@0: assert (_succ != Self , "invariant") ; aoqi@0: assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; aoqi@0: aoqi@0: // The thread -- now the owner -- is back in vm mode. aoqi@0: // Report the glorious news via TI,DTrace and jvmstat. aoqi@0: // The probe effect is non-trivial. All the reportage occurs aoqi@0: // while we hold the monitor, increasing the length of the critical aoqi@0: // section. Amdahl's parallel speedup law comes vividly into play. aoqi@0: // aoqi@0: // Another option might be to aggregate the events (thread local or aoqi@0: // per-monitor aggregation) and defer reporting until a more opportune aoqi@0: // time -- such as next time some thread encounters contention but has aoqi@0: // yet to acquire the lock. While spinning that thread could aoqi@0: // spinning we could increment JVMStat counters, etc. aoqi@0: aoqi@0: DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); aoqi@0: if (JvmtiExport::should_post_monitor_contended_entered()) { aoqi@0: JvmtiExport::post_monitor_contended_entered(jt, this); aoqi@0: aoqi@0: // The current thread already owns the monitor and is not going to aoqi@0: // call park() for the remainder of the monitor enter protocol. So aoqi@0: // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED aoqi@0: // event handler consumed an unpark() issued by the thread that aoqi@0: // just exited the monitor. aoqi@0: } aoqi@0: aoqi@0: if (event.should_commit()) { aoqi@0: event.set_klass(((oop)this->object())->klass()); aoqi@0: event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); aoqi@0: event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); aoqi@0: event.commit(); aoqi@0: } aoqi@0: aoqi@0: if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { aoqi@0: ObjectMonitor::_sync_ContendedLockAttempts->inc() ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: aoqi@0: // Caveat: TryLock() is not necessarily serializing if it returns failure. aoqi@0: // Callers must compensate as needed. aoqi@0: aoqi@0: int ObjectMonitor::TryLock (Thread * Self) { aoqi@0: for (;;) { aoqi@0: void * own = _owner ; aoqi@0: if (own != NULL) return 0 ; aoqi@0: if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { aoqi@0: // Either guarantee _recursions == 0 or set _recursions = 0. aoqi@0: assert (_recursions == 0, "invariant") ; aoqi@0: assert (_owner == Self, "invariant") ; aoqi@0: // CONSIDER: set or assert that OwnerIsThread == 1 aoqi@0: return 1 ; aoqi@0: } aoqi@0: // The lock had been free momentarily, but we lost the race to the lock. aoqi@0: // Interference -- the CAS failed. aoqi@0: // We can either return -1 or retry. aoqi@0: // Retry doesn't make as much sense because the lock was just acquired. aoqi@0: if (true) return -1 ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: void ATTR ObjectMonitor::EnterI (TRAPS) { aoqi@0: Thread * Self = THREAD ; aoqi@0: assert (Self->is_Java_thread(), "invariant") ; aoqi@0: assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; aoqi@0: aoqi@0: // Try the lock - TATAS aoqi@0: if (TryLock (Self) > 0) { aoqi@0: assert (_succ != Self , "invariant") ; aoqi@0: assert (_owner == Self , "invariant") ; aoqi@0: assert (_Responsible != Self , "invariant") ; aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: DeferredInitialize () ; aoqi@0: aoqi@0: // We try one round of spinning *before* enqueueing Self. aoqi@0: // aoqi@0: // If the _owner is ready but OFFPROC we could use a YieldTo() aoqi@0: // operation to donate the remainder of this thread's quantum aoqi@0: // to the owner. This has subtle but beneficial affinity aoqi@0: // effects. aoqi@0: aoqi@0: if (TrySpin (Self) > 0) { aoqi@0: assert (_owner == Self , "invariant") ; aoqi@0: assert (_succ != Self , "invariant") ; aoqi@0: assert (_Responsible != Self , "invariant") ; aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: // The Spin failed -- Enqueue and park the thread ... aoqi@0: assert (_succ != Self , "invariant") ; aoqi@0: assert (_owner != Self , "invariant") ; aoqi@0: assert (_Responsible != Self , "invariant") ; aoqi@0: aoqi@0: // Enqueue "Self" on ObjectMonitor's _cxq. aoqi@0: // aoqi@0: // Node acts as a proxy for Self. aoqi@0: // As an aside, if were to ever rewrite the synchronization code mostly aoqi@0: // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class aoqi@0: // Java objects. This would avoid awkward lifecycle and liveness issues, aoqi@0: // as well as eliminate a subset of ABA issues. aoqi@0: // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. aoqi@0: // aoqi@0: aoqi@0: ObjectWaiter node(Self) ; aoqi@0: Self->_ParkEvent->reset() ; aoqi@0: node._prev = (ObjectWaiter *) 0xBAD ; aoqi@0: node.TState = ObjectWaiter::TS_CXQ ; aoqi@0: aoqi@0: // Push "Self" onto the front of the _cxq. aoqi@0: // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. aoqi@0: // Note that spinning tends to reduce the rate at which threads aoqi@0: // enqueue and dequeue on EntryList|cxq. aoqi@0: ObjectWaiter * nxt ; aoqi@0: for (;;) { aoqi@0: node._next = nxt = _cxq ; aoqi@0: if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; aoqi@0: aoqi@0: // Interference - the CAS failed because _cxq changed. Just retry. aoqi@0: // As an optional optimization we retry the lock. aoqi@0: if (TryLock (Self) > 0) { aoqi@0: assert (_succ != Self , "invariant") ; aoqi@0: assert (_owner == Self , "invariant") ; aoqi@0: assert (_Responsible != Self , "invariant") ; aoqi@0: return ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // Check for cxq|EntryList edge transition to non-null. This indicates aoqi@0: // the onset of contention. While contention persists exiting threads aoqi@0: // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit aoqi@0: // operations revert to the faster 1-0 mode. This enter operation may interleave aoqi@0: // (race) a concurrent 1-0 exit operation, resulting in stranding, so we aoqi@0: // arrange for one of the contending thread to use a timed park() operations aoqi@0: // to detect and recover from the race. (Stranding is form of progress failure aoqi@0: // where the monitor is unlocked but all the contending threads remain parked). aoqi@0: // That is, at least one of the contended threads will periodically poll _owner. aoqi@0: // One of the contending threads will become the designated "Responsible" thread. aoqi@0: // The Responsible thread uses a timed park instead of a normal indefinite park aoqi@0: // operation -- it periodically wakes and checks for and recovers from potential aoqi@0: // strandings admitted by 1-0 exit operations. We need at most one Responsible aoqi@0: // thread per-monitor at any given moment. Only threads on cxq|EntryList may aoqi@0: // be responsible for a monitor. aoqi@0: // aoqi@0: // Currently, one of the contended threads takes on the added role of "Responsible". aoqi@0: // A viable alternative would be to use a dedicated "stranding checker" thread aoqi@0: // that periodically iterated over all the threads (or active monitors) and unparked aoqi@0: // successors where there was risk of stranding. This would help eliminate the aoqi@0: // timer scalability issues we see on some platforms as we'd only have one thread aoqi@0: // -- the checker -- parked on a timer. aoqi@0: aoqi@0: if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { aoqi@0: // Try to assume the role of responsible thread for the monitor. aoqi@0: // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } aoqi@0: Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; aoqi@0: } aoqi@0: aoqi@0: // The lock have been released while this thread was occupied queueing aoqi@0: // itself onto _cxq. To close the race and avoid "stranding" and aoqi@0: // progress-liveness failure we must resample-retry _owner before parking. aoqi@0: // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. aoqi@0: // In this case the ST-MEMBAR is accomplished with CAS(). aoqi@0: // aoqi@0: // TODO: Defer all thread state transitions until park-time. aoqi@0: // Since state transitions are heavy and inefficient we'd like aoqi@0: // to defer the state transitions until absolutely necessary, aoqi@0: // and in doing so avoid some transitions ... aoqi@0: aoqi@0: TEVENT (Inflated enter - Contention) ; aoqi@0: int nWakeups = 0 ; aoqi@0: int RecheckInterval = 1 ; aoqi@0: aoqi@0: for (;;) { aoqi@0: aoqi@0: if (TryLock (Self) > 0) break ; aoqi@0: assert (_owner != Self, "invariant") ; aoqi@0: aoqi@0: if ((SyncFlags & 2) && _Responsible == NULL) { aoqi@0: Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; aoqi@0: } aoqi@0: aoqi@0: // park self aoqi@0: if (_Responsible == Self || (SyncFlags & 1)) { aoqi@0: TEVENT (Inflated enter - park TIMED) ; aoqi@0: Self->_ParkEvent->park ((jlong) RecheckInterval) ; aoqi@0: // Increase the RecheckInterval, but clamp the value. aoqi@0: RecheckInterval *= 8 ; aoqi@0: if (RecheckInterval > 1000) RecheckInterval = 1000 ; aoqi@0: } else { aoqi@0: TEVENT (Inflated enter - park UNTIMED) ; aoqi@0: Self->_ParkEvent->park() ; aoqi@0: } aoqi@0: aoqi@0: if (TryLock(Self) > 0) break ; aoqi@0: aoqi@0: // The lock is still contested. aoqi@0: // Keep a tally of the # of futile wakeups. aoqi@0: // Note that the counter is not protected by a lock or updated by atomics. aoqi@0: // That is by design - we trade "lossy" counters which are exposed to aoqi@0: // races during updates for a lower probe effect. aoqi@0: TEVENT (Inflated enter - Futile wakeup) ; aoqi@0: if (ObjectMonitor::_sync_FutileWakeups != NULL) { aoqi@0: ObjectMonitor::_sync_FutileWakeups->inc() ; aoqi@0: } aoqi@0: ++ nWakeups ; aoqi@0: aoqi@0: // Assuming this is not a spurious wakeup we'll normally find _succ == Self. aoqi@0: // We can defer clearing _succ until after the spin completes aoqi@0: // TrySpin() must tolerate being called with _succ == Self. aoqi@0: // Try yet another round of adaptive spinning. aoqi@0: if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; aoqi@0: aoqi@0: // We can find that we were unpark()ed and redesignated _succ while aoqi@0: // we were spinning. That's harmless. If we iterate and call park(), aoqi@0: // park() will consume the event and return immediately and we'll aoqi@0: // just spin again. This pattern can repeat, leaving _succ to simply aoqi@0: // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). aoqi@0: // Alternately, we can sample fired() here, and if set, forgo spinning aoqi@0: // in the next iteration. aoqi@0: aoqi@0: if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { aoqi@0: Self->_ParkEvent->reset() ; aoqi@0: OrderAccess::fence() ; aoqi@0: } aoqi@0: if (_succ == Self) _succ = NULL ; aoqi@0: aoqi@0: // Invariant: after clearing _succ a thread *must* retry _owner before parking. aoqi@0: OrderAccess::fence() ; aoqi@0: } aoqi@0: aoqi@0: // Egress : aoqi@0: // Self has acquired the lock -- Unlink Self from the cxq or EntryList. aoqi@0: // Normally we'll find Self on the EntryList . aoqi@0: // From the perspective of the lock owner (this thread), the aoqi@0: // EntryList is stable and cxq is prepend-only. aoqi@0: // The head of cxq is volatile but the interior is stable. aoqi@0: // In addition, Self.TState is stable. aoqi@0: aoqi@0: assert (_owner == Self , "invariant") ; aoqi@0: assert (object() != NULL , "invariant") ; aoqi@0: // I'd like to write: aoqi@0: // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; aoqi@0: // but as we're at a safepoint that's not safe. aoqi@0: aoqi@0: UnlinkAfterAcquire (Self, &node) ; aoqi@0: if (_succ == Self) _succ = NULL ; aoqi@0: aoqi@0: assert (_succ != Self, "invariant") ; aoqi@0: if (_Responsible == Self) { aoqi@0: _Responsible = NULL ; aoqi@0: OrderAccess::fence(); // Dekker pivot-point aoqi@0: aoqi@0: // We may leave threads on cxq|EntryList without a designated aoqi@0: // "Responsible" thread. This is benign. When this thread subsequently aoqi@0: // exits the monitor it can "see" such preexisting "old" threads -- aoqi@0: // threads that arrived on the cxq|EntryList before the fence, above -- aoqi@0: // by LDing cxq|EntryList. Newly arrived threads -- that is, threads aoqi@0: // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible aoqi@0: // non-null and elect a new "Responsible" timer thread. aoqi@0: // aoqi@0: // This thread executes: aoqi@0: // ST Responsible=null; MEMBAR (in enter epilog - here) aoqi@0: // LD cxq|EntryList (in subsequent exit) aoqi@0: // aoqi@0: // Entering threads in the slow/contended path execute: aoqi@0: // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) aoqi@0: // The (ST cxq; MEMBAR) is accomplished with CAS(). aoqi@0: // aoqi@0: // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent aoqi@0: // exit operation from floating above the ST Responsible=null. aoqi@0: } aoqi@0: aoqi@0: // We've acquired ownership with CAS(). aoqi@0: // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. aoqi@0: // But since the CAS() this thread may have also stored into _succ, aoqi@0: // EntryList, cxq or Responsible. These meta-data updates must be aoqi@0: // visible __before this thread subsequently drops the lock. aoqi@0: // Consider what could occur if we didn't enforce this constraint -- aoqi@0: // STs to monitor meta-data and user-data could reorder with (become aoqi@0: // visible after) the ST in exit that drops ownership of the lock. aoqi@0: // Some other thread could then acquire the lock, but observe inconsistent aoqi@0: // or old monitor meta-data and heap data. That violates the JMM. aoqi@0: // To that end, the 1-0 exit() operation must have at least STST|LDST aoqi@0: // "release" barrier semantics. Specifically, there must be at least a aoqi@0: // STST|LDST barrier in exit() before the ST of null into _owner that drops aoqi@0: // the lock. The barrier ensures that changes to monitor meta-data and data aoqi@0: // protected by the lock will be visible before we release the lock, and aoqi@0: // therefore before some other thread (CPU) has a chance to acquire the lock. aoqi@0: // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. aoqi@0: // aoqi@0: // Critically, any prior STs to _succ or EntryList must be visible before aoqi@0: // the ST of null into _owner in the *subsequent* (following) corresponding aoqi@0: // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily aoqi@0: // execute a serializing instruction. aoqi@0: aoqi@0: if (SyncFlags & 8) { aoqi@0: OrderAccess::fence() ; aoqi@0: } aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: // ReenterI() is a specialized inline form of the latter half of the aoqi@0: // contended slow-path from EnterI(). We use ReenterI() only for aoqi@0: // monitor reentry in wait(). aoqi@0: // aoqi@0: // In the future we should reconcile EnterI() and ReenterI(), adding aoqi@0: // Knob_Reset and Knob_SpinAfterFutile support and restructuring the aoqi@0: // loop accordingly. aoqi@0: aoqi@0: void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { aoqi@0: assert (Self != NULL , "invariant") ; aoqi@0: assert (SelfNode != NULL , "invariant") ; aoqi@0: assert (SelfNode->_thread == Self , "invariant") ; aoqi@0: assert (_waiters > 0 , "invariant") ; aoqi@0: assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; aoqi@0: assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; aoqi@0: JavaThread * jt = (JavaThread *) Self ; aoqi@0: aoqi@0: int nWakeups = 0 ; aoqi@0: for (;;) { aoqi@0: ObjectWaiter::TStates v = SelfNode->TState ; aoqi@0: guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; aoqi@0: assert (_owner != Self, "invariant") ; aoqi@0: aoqi@0: if (TryLock (Self) > 0) break ; aoqi@0: if (TrySpin (Self) > 0) break ; aoqi@0: aoqi@0: TEVENT (Wait Reentry - parking) ; aoqi@0: aoqi@0: // State transition wrappers around park() ... aoqi@0: // ReenterI() wisely defers state transitions until aoqi@0: // it's clear we must park the thread. aoqi@0: { aoqi@0: OSThreadContendState osts(Self->osthread()); aoqi@0: ThreadBlockInVM tbivm(jt); aoqi@0: aoqi@0: // cleared by handle_special_suspend_equivalent_condition() aoqi@0: // or java_suspend_self() aoqi@0: jt->set_suspend_equivalent(); aoqi@0: if (SyncFlags & 1) { aoqi@0: Self->_ParkEvent->park ((jlong)1000) ; aoqi@0: } else { aoqi@0: Self->_ParkEvent->park () ; aoqi@0: } aoqi@0: aoqi@0: // were we externally suspended while we were waiting? aoqi@0: for (;;) { aoqi@0: if (!ExitSuspendEquivalent (jt)) break ; aoqi@0: if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } aoqi@0: jt->java_suspend_self(); aoqi@0: jt->set_suspend_equivalent(); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // Try again, but just so we distinguish between futile wakeups and aoqi@0: // successful wakeups. The following test isn't algorithmically aoqi@0: // necessary, but it helps us maintain sensible statistics. aoqi@0: if (TryLock(Self) > 0) break ; aoqi@0: aoqi@0: // The lock is still contested. aoqi@0: // Keep a tally of the # of futile wakeups. aoqi@0: // Note that the counter is not protected by a lock or updated by atomics. aoqi@0: // That is by design - we trade "lossy" counters which are exposed to aoqi@0: // races during updates for a lower probe effect. aoqi@0: TEVENT (Wait Reentry - futile wakeup) ; aoqi@0: ++ nWakeups ; aoqi@0: aoqi@0: // Assuming this is not a spurious wakeup we'll normally aoqi@0: // find that _succ == Self. aoqi@0: if (_succ == Self) _succ = NULL ; aoqi@0: aoqi@0: // Invariant: after clearing _succ a contending thread aoqi@0: // *must* retry _owner before parking. aoqi@0: OrderAccess::fence() ; aoqi@0: aoqi@0: if (ObjectMonitor::_sync_FutileWakeups != NULL) { aoqi@0: ObjectMonitor::_sync_FutileWakeups->inc() ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // Self has acquired the lock -- Unlink Self from the cxq or EntryList . aoqi@0: // Normally we'll find Self on the EntryList. aoqi@0: // Unlinking from the EntryList is constant-time and atomic-free. aoqi@0: // From the perspective of the lock owner (this thread), the aoqi@0: // EntryList is stable and cxq is prepend-only. aoqi@0: // The head of cxq is volatile but the interior is stable. aoqi@0: // In addition, Self.TState is stable. aoqi@0: aoqi@0: assert (_owner == Self, "invariant") ; aoqi@0: assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; aoqi@0: UnlinkAfterAcquire (Self, SelfNode) ; aoqi@0: if (_succ == Self) _succ = NULL ; aoqi@0: assert (_succ != Self, "invariant") ; aoqi@0: SelfNode->TState = ObjectWaiter::TS_RUN ; aoqi@0: OrderAccess::fence() ; // see comments at the end of EnterI() aoqi@0: } aoqi@0: aoqi@0: // after the thread acquires the lock in ::enter(). Equally, we could defer aoqi@0: // unlinking the thread until ::exit()-time. aoqi@0: aoqi@0: void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) aoqi@0: { aoqi@0: assert (_owner == Self, "invariant") ; aoqi@0: assert (SelfNode->_thread == Self, "invariant") ; aoqi@0: aoqi@0: if (SelfNode->TState == ObjectWaiter::TS_ENTER) { aoqi@0: // Normal case: remove Self from the DLL EntryList . aoqi@0: // This is a constant-time operation. aoqi@0: ObjectWaiter * nxt = SelfNode->_next ; aoqi@0: ObjectWaiter * prv = SelfNode->_prev ; aoqi@0: if (nxt != NULL) nxt->_prev = prv ; aoqi@0: if (prv != NULL) prv->_next = nxt ; aoqi@0: if (SelfNode == _EntryList ) _EntryList = nxt ; aoqi@0: assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; aoqi@0: assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; aoqi@0: TEVENT (Unlink from EntryList) ; aoqi@0: } else { aoqi@0: guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; aoqi@0: // Inopportune interleaving -- Self is still on the cxq. aoqi@0: // This usually means the enqueue of self raced an exiting thread. aoqi@0: // Normally we'll find Self near the front of the cxq, so aoqi@0: // dequeueing is typically fast. If needbe we can accelerate aoqi@0: // this with some MCS/CHL-like bidirectional list hints and advisory aoqi@0: // back-links so dequeueing from the interior will normally operate aoqi@0: // in constant-time. aoqi@0: // Dequeue Self from either the head (with CAS) or from the interior aoqi@0: // with a linear-time scan and normal non-atomic memory operations. aoqi@0: // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList aoqi@0: // and then unlink Self from EntryList. We have to drain eventually, aoqi@0: // so it might as well be now. aoqi@0: aoqi@0: ObjectWaiter * v = _cxq ; aoqi@0: assert (v != NULL, "invariant") ; aoqi@0: if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { aoqi@0: // The CAS above can fail from interference IFF a "RAT" arrived. aoqi@0: // In that case Self must be in the interior and can no longer be aoqi@0: // at the head of cxq. aoqi@0: if (v == SelfNode) { aoqi@0: assert (_cxq != v, "invariant") ; aoqi@0: v = _cxq ; // CAS above failed - start scan at head of list aoqi@0: } aoqi@0: ObjectWaiter * p ; aoqi@0: ObjectWaiter * q = NULL ; aoqi@0: for (p = v ; p != NULL && p != SelfNode; p = p->_next) { aoqi@0: q = p ; aoqi@0: assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; aoqi@0: } aoqi@0: assert (v != SelfNode, "invariant") ; aoqi@0: assert (p == SelfNode, "Node not found on cxq") ; aoqi@0: assert (p != _cxq, "invariant") ; aoqi@0: assert (q != NULL, "invariant") ; aoqi@0: assert (q->_next == p, "invariant") ; aoqi@0: q->_next = p->_next ; aoqi@0: } aoqi@0: TEVENT (Unlink from cxq) ; aoqi@0: } aoqi@0: aoqi@0: // Diagnostic hygiene ... aoqi@0: SelfNode->_prev = (ObjectWaiter *) 0xBAD ; aoqi@0: SelfNode->_next = (ObjectWaiter *) 0xBAD ; aoqi@0: SelfNode->TState = ObjectWaiter::TS_RUN ; aoqi@0: } aoqi@0: aoqi@0: // ----------------------------------------------------------------------------- aoqi@0: // Exit support aoqi@0: // aoqi@0: // exit() aoqi@0: // ~~~~~~ aoqi@0: // Note that the collector can't reclaim the objectMonitor or deflate aoqi@0: // the object out from underneath the thread calling ::exit() as the aoqi@0: // thread calling ::exit() never transitions to a stable state. aoqi@0: // This inhibits GC, which in turn inhibits asynchronous (and aoqi@0: // inopportune) reclamation of "this". aoqi@0: // aoqi@0: // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; aoqi@0: // There's one exception to the claim above, however. EnterI() can call aoqi@0: // exit() to drop a lock if the acquirer has been externally suspended. aoqi@0: // In that case exit() is called with _thread_state as _thread_blocked, aoqi@0: // but the monitor's _count field is > 0, which inhibits reclamation. aoqi@0: // aoqi@0: // 1-0 exit aoqi@0: // ~~~~~~~~ aoqi@0: // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of aoqi@0: // the fast-path operators have been optimized so the common ::exit() aoqi@0: // operation is 1-0. See i486.ad fast_unlock(), for instance. aoqi@0: // The code emitted by fast_unlock() elides the usual MEMBAR. This aoqi@0: // greatly improves latency -- MEMBAR and CAS having considerable local aoqi@0: // latency on modern processors -- but at the cost of "stranding". Absent the aoqi@0: // MEMBAR, a thread in fast_unlock() can race a thread in the slow aoqi@0: // ::enter() path, resulting in the entering thread being stranding aoqi@0: // and a progress-liveness failure. Stranding is extremely rare. aoqi@0: // We use timers (timed park operations) & periodic polling to detect aoqi@0: // and recover from stranding. Potentially stranded threads periodically aoqi@0: // wake up and poll the lock. See the usage of the _Responsible variable. aoqi@0: // aoqi@0: // The CAS() in enter provides for safety and exclusion, while the CAS or aoqi@0: // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking aoqi@0: // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. aoqi@0: // We detect and recover from stranding with timers. aoqi@0: // aoqi@0: // If a thread transiently strands it'll park until (a) another aoqi@0: // thread acquires the lock and then drops the lock, at which time the aoqi@0: // exiting thread will notice and unpark the stranded thread, or, (b) aoqi@0: // the timer expires. If the lock is high traffic then the stranding latency aoqi@0: // will be low due to (a). If the lock is low traffic then the odds of aoqi@0: // stranding are lower, although the worst-case stranding latency aoqi@0: // is longer. Critically, we don't want to put excessive load in the aoqi@0: // platform's timer subsystem. We want to minimize both the timer injection aoqi@0: // rate (timers created/sec) as well as the number of timers active at aoqi@0: // any one time. (more precisely, we want to minimize timer-seconds, which is aoqi@0: // the integral of the # of active timers at any instant over time). aoqi@0: // Both impinge on OS scalability. Given that, at most one thread parked on aoqi@0: // a monitor will use a timer. aoqi@0: aoqi@0: void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { aoqi@0: Thread * Self = THREAD ; aoqi@0: if (THREAD != _owner) { aoqi@0: if (THREAD->is_lock_owned((address) _owner)) { aoqi@0: // Transmute _owner from a BasicLock pointer to a Thread address. aoqi@0: // We don't need to hold _mutex for this transition. aoqi@0: // Non-null to Non-null is safe as long as all readers can aoqi@0: // tolerate either flavor. aoqi@0: assert (_recursions == 0, "invariant") ; aoqi@0: _owner = THREAD ; aoqi@0: _recursions = 0 ; aoqi@0: OwnerIsThread = 1 ; aoqi@0: } else { aoqi@0: // NOTE: we need to handle unbalanced monitor enter/exit aoqi@0: // in native code by throwing an exception. aoqi@0: // TODO: Throw an IllegalMonitorStateException ? aoqi@0: TEVENT (Exit - Throw IMSX) ; aoqi@0: assert(false, "Non-balanced monitor enter/exit!"); aoqi@0: if (false) { aoqi@0: THROW(vmSymbols::java_lang_IllegalMonitorStateException()); aoqi@0: } aoqi@0: return; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: if (_recursions != 0) { aoqi@0: _recursions--; // this is simple recursive enter aoqi@0: TEVENT (Inflated exit - recursive) ; aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: // Invariant: after setting Responsible=null an thread must execute aoqi@0: // a MEMBAR or other serializing instruction before fetching EntryList|cxq. aoqi@0: if ((SyncFlags & 4) == 0) { aoqi@0: _Responsible = NULL ; aoqi@0: } aoqi@0: aoqi@0: #if INCLUDE_TRACE aoqi@0: // get the owner's thread id for the MonitorEnter event aoqi@0: // if it is enabled and the thread isn't suspended aoqi@0: if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { aoqi@0: _previous_owner_tid = SharedRuntime::get_java_tid(Self); aoqi@0: } aoqi@0: #endif aoqi@0: aoqi@0: for (;;) { aoqi@0: assert (THREAD == _owner, "invariant") ; aoqi@0: aoqi@0: aoqi@0: if (Knob_ExitPolicy == 0) { aoqi@0: // release semantics: prior loads and stores from within the critical section aoqi@0: // must not float (reorder) past the following store that drops the lock. aoqi@0: // On SPARC that requires MEMBAR #loadstore|#storestore. aoqi@0: // But of course in TSO #loadstore|#storestore is not required. aoqi@0: // I'd like to write one of the following: aoqi@0: // A. OrderAccess::release() ; _owner = NULL aoqi@0: // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; aoqi@0: // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both aoqi@0: // store into a _dummy variable. That store is not needed, but can result aoqi@0: // in massive wasteful coherency traffic on classic SMP systems. aoqi@0: // Instead, I use release_store(), which is implemented as just a simple aoqi@0: // ST on x64, x86 and SPARC. aoqi@0: OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock aoqi@0: OrderAccess::storeload() ; // See if we need to wake a successor aoqi@0: if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { aoqi@0: TEVENT (Inflated exit - simple egress) ; aoqi@0: return ; aoqi@0: } aoqi@0: TEVENT (Inflated exit - complex egress) ; aoqi@0: aoqi@0: // Normally the exiting thread is responsible for ensuring succession, aoqi@0: // but if other successors are ready or other entering threads are spinning aoqi@0: // then this thread can simply store NULL into _owner and exit without aoqi@0: // waking a successor. The existence of spinners or ready successors aoqi@0: // guarantees proper succession (liveness). Responsibility passes to the aoqi@0: // ready or running successors. The exiting thread delegates the duty. aoqi@0: // More precisely, if a successor already exists this thread is absolved aoqi@0: // of the responsibility of waking (unparking) one. aoqi@0: // aoqi@0: // The _succ variable is critical to reducing futile wakeup frequency. aoqi@0: // _succ identifies the "heir presumptive" thread that has been made aoqi@0: // ready (unparked) but that has not yet run. We need only one such aoqi@0: // successor thread to guarantee progress. aoqi@0: // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf aoqi@0: // section 3.3 "Futile Wakeup Throttling" for details. aoqi@0: // aoqi@0: // Note that spinners in Enter() also set _succ non-null. aoqi@0: // In the current implementation spinners opportunistically set aoqi@0: // _succ so that exiting threads might avoid waking a successor. aoqi@0: // Another less appealing alternative would be for the exiting thread aoqi@0: // to drop the lock and then spin briefly to see if a spinner managed aoqi@0: // to acquire the lock. If so, the exiting thread could exit aoqi@0: // immediately without waking a successor, otherwise the exiting aoqi@0: // thread would need to dequeue and wake a successor. aoqi@0: // (Note that we'd need to make the post-drop spin short, but no aoqi@0: // shorter than the worst-case round-trip cache-line migration time. aoqi@0: // The dropped lock needs to become visible to the spinner, and then aoqi@0: // the acquisition of the lock by the spinner must become visible to aoqi@0: // the exiting thread). aoqi@0: // aoqi@0: aoqi@0: // It appears that an heir-presumptive (successor) must be made ready. aoqi@0: // Only the current lock owner can manipulate the EntryList or aoqi@0: // drain _cxq, so we need to reacquire the lock. If we fail aoqi@0: // to reacquire the lock the responsibility for ensuring succession aoqi@0: // falls to the new owner. aoqi@0: // aoqi@0: if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { aoqi@0: return ; aoqi@0: } aoqi@0: TEVENT (Exit - Reacquired) ; aoqi@0: } else { aoqi@0: if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { aoqi@0: OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock aoqi@0: OrderAccess::storeload() ; aoqi@0: // Ratify the previously observed values. aoqi@0: if (_cxq == NULL || _succ != NULL) { aoqi@0: TEVENT (Inflated exit - simple egress) ; aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: // inopportune interleaving -- the exiting thread (this thread) aoqi@0: // in the fast-exit path raced an entering thread in the slow-enter aoqi@0: // path. aoqi@0: // We have two choices: aoqi@0: // A. Try to reacquire the lock. aoqi@0: // If the CAS() fails return immediately, otherwise aoqi@0: // we either restart/rerun the exit operation, or simply aoqi@0: // fall-through into the code below which wakes a successor. aoqi@0: // B. If the elements forming the EntryList|cxq are TSM aoqi@0: // we could simply unpark() the lead thread and return aoqi@0: // without having set _succ. aoqi@0: if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { aoqi@0: TEVENT (Inflated exit - reacquired succeeded) ; aoqi@0: return ; aoqi@0: } aoqi@0: TEVENT (Inflated exit - reacquired failed) ; aoqi@0: } else { aoqi@0: TEVENT (Inflated exit - complex egress) ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: guarantee (_owner == THREAD, "invariant") ; aoqi@0: aoqi@0: ObjectWaiter * w = NULL ; aoqi@0: int QMode = Knob_QMode ; aoqi@0: aoqi@0: if (QMode == 2 && _cxq != NULL) { aoqi@0: // QMode == 2 : cxq has precedence over EntryList. aoqi@0: // Try to directly wake a successor from the cxq. aoqi@0: // If successful, the successor will need to unlink itself from cxq. aoqi@0: w = _cxq ; aoqi@0: assert (w != NULL, "invariant") ; aoqi@0: assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; aoqi@0: ExitEpilog (Self, w) ; aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: if (QMode == 3 && _cxq != NULL) { aoqi@0: // Aggressively drain cxq into EntryList at the first opportunity. aoqi@0: // This policy ensure that recently-run threads live at the head of EntryList. aoqi@0: // Drain _cxq into EntryList - bulk transfer. aoqi@0: // First, detach _cxq. aoqi@0: // The following loop is tantamount to: w = swap (&cxq, NULL) aoqi@0: w = _cxq ; aoqi@0: for (;;) { aoqi@0: assert (w != NULL, "Invariant") ; aoqi@0: ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; aoqi@0: if (u == w) break ; aoqi@0: w = u ; aoqi@0: } aoqi@0: assert (w != NULL , "invariant") ; aoqi@0: aoqi@0: ObjectWaiter * q = NULL ; aoqi@0: ObjectWaiter * p ; aoqi@0: for (p = w ; p != NULL ; p = p->_next) { aoqi@0: guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; aoqi@0: p->TState = ObjectWaiter::TS_ENTER ; aoqi@0: p->_prev = q ; aoqi@0: q = p ; aoqi@0: } aoqi@0: aoqi@0: // Append the RATs to the EntryList aoqi@0: // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. aoqi@0: ObjectWaiter * Tail ; aoqi@0: for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; aoqi@0: if (Tail == NULL) { aoqi@0: _EntryList = w ; aoqi@0: } else { aoqi@0: Tail->_next = w ; aoqi@0: w->_prev = Tail ; aoqi@0: } aoqi@0: aoqi@0: // Fall thru into code that tries to wake a successor from EntryList aoqi@0: } aoqi@0: aoqi@0: if (QMode == 4 && _cxq != NULL) { aoqi@0: // Aggressively drain cxq into EntryList at the first opportunity. aoqi@0: // This policy ensure that recently-run threads live at the head of EntryList. aoqi@0: aoqi@0: // Drain _cxq into EntryList - bulk transfer. aoqi@0: // First, detach _cxq. aoqi@0: // The following loop is tantamount to: w = swap (&cxq, NULL) aoqi@0: w = _cxq ; aoqi@0: for (;;) { aoqi@0: assert (w != NULL, "Invariant") ; aoqi@0: ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; aoqi@0: if (u == w) break ; aoqi@0: w = u ; aoqi@0: } aoqi@0: assert (w != NULL , "invariant") ; aoqi@0: aoqi@0: ObjectWaiter * q = NULL ; aoqi@0: ObjectWaiter * p ; aoqi@0: for (p = w ; p != NULL ; p = p->_next) { aoqi@0: guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; aoqi@0: p->TState = ObjectWaiter::TS_ENTER ; aoqi@0: p->_prev = q ; aoqi@0: q = p ; aoqi@0: } aoqi@0: aoqi@0: // Prepend the RATs to the EntryList aoqi@0: if (_EntryList != NULL) { aoqi@0: q->_next = _EntryList ; aoqi@0: _EntryList->_prev = q ; aoqi@0: } aoqi@0: _EntryList = w ; aoqi@0: aoqi@0: // Fall thru into code that tries to wake a successor from EntryList aoqi@0: } aoqi@0: aoqi@0: w = _EntryList ; aoqi@0: if (w != NULL) { aoqi@0: // I'd like to write: guarantee (w->_thread != Self). aoqi@0: // But in practice an exiting thread may find itself on the EntryList. aoqi@0: // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and aoqi@0: // then calls exit(). Exit release the lock by setting O._owner to NULL. aoqi@0: // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The aoqi@0: // notify() operation moves T1 from O's waitset to O's EntryList. T2 then aoqi@0: // release the lock "O". T2 resumes immediately after the ST of null into aoqi@0: // _owner, above. T2 notices that the EntryList is populated, so it aoqi@0: // reacquires the lock and then finds itself on the EntryList. aoqi@0: // Given all that, we have to tolerate the circumstance where "w" is aoqi@0: // associated with Self. aoqi@0: assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; aoqi@0: ExitEpilog (Self, w) ; aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: // If we find that both _cxq and EntryList are null then just aoqi@0: // re-run the exit protocol from the top. aoqi@0: w = _cxq ; aoqi@0: if (w == NULL) continue ; aoqi@0: aoqi@0: // Drain _cxq into EntryList - bulk transfer. aoqi@0: // First, detach _cxq. aoqi@0: // The following loop is tantamount to: w = swap (&cxq, NULL) aoqi@0: for (;;) { aoqi@0: assert (w != NULL, "Invariant") ; aoqi@0: ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; aoqi@0: if (u == w) break ; aoqi@0: w = u ; aoqi@0: } aoqi@0: TEVENT (Inflated exit - drain cxq into EntryList) ; aoqi@0: aoqi@0: assert (w != NULL , "invariant") ; aoqi@0: assert (_EntryList == NULL , "invariant") ; aoqi@0: aoqi@0: // Convert the LIFO SLL anchored by _cxq into a DLL. aoqi@0: // The list reorganization step operates in O(LENGTH(w)) time. aoqi@0: // It's critical that this step operate quickly as aoqi@0: // "Self" still holds the outer-lock, restricting parallelism aoqi@0: // and effectively lengthening the critical section. aoqi@0: // Invariant: s chases t chases u. aoqi@0: // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so aoqi@0: // we have faster access to the tail. aoqi@0: aoqi@0: if (QMode == 1) { aoqi@0: // QMode == 1 : drain cxq to EntryList, reversing order aoqi@0: // We also reverse the order of the list. aoqi@0: ObjectWaiter * s = NULL ; aoqi@0: ObjectWaiter * t = w ; aoqi@0: ObjectWaiter * u = NULL ; aoqi@0: while (t != NULL) { aoqi@0: guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; aoqi@0: t->TState = ObjectWaiter::TS_ENTER ; aoqi@0: u = t->_next ; aoqi@0: t->_prev = u ; aoqi@0: t->_next = s ; aoqi@0: s = t; aoqi@0: t = u ; aoqi@0: } aoqi@0: _EntryList = s ; aoqi@0: assert (s != NULL, "invariant") ; aoqi@0: } else { aoqi@0: // QMode == 0 or QMode == 2 aoqi@0: _EntryList = w ; aoqi@0: ObjectWaiter * q = NULL ; aoqi@0: ObjectWaiter * p ; aoqi@0: for (p = w ; p != NULL ; p = p->_next) { aoqi@0: guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; aoqi@0: p->TState = ObjectWaiter::TS_ENTER ; aoqi@0: p->_prev = q ; aoqi@0: q = p ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL aoqi@0: // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). aoqi@0: aoqi@0: // See if we can abdicate to a spinner instead of waking a thread. aoqi@0: // A primary goal of the implementation is to reduce the aoqi@0: // context-switch rate. aoqi@0: if (_succ != NULL) continue; aoqi@0: aoqi@0: w = _EntryList ; aoqi@0: if (w != NULL) { aoqi@0: guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; aoqi@0: ExitEpilog (Self, w) ; aoqi@0: return ; aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // ExitSuspendEquivalent: aoqi@0: // A faster alternate to handle_special_suspend_equivalent_condition() aoqi@0: // aoqi@0: // handle_special_suspend_equivalent_condition() unconditionally aoqi@0: // acquires the SR_lock. On some platforms uncontended MutexLocker() aoqi@0: // operations have high latency. Note that in ::enter() we call HSSEC aoqi@0: // while holding the monitor, so we effectively lengthen the critical sections. aoqi@0: // aoqi@0: // There are a number of possible solutions: aoqi@0: // aoqi@0: // A. To ameliorate the problem we might also defer state transitions aoqi@0: // to as late as possible -- just prior to parking. aoqi@0: // Given that, we'd call HSSEC after having returned from park(), aoqi@0: // but before attempting to acquire the monitor. This is only a aoqi@0: // partial solution. It avoids calling HSSEC while holding the aoqi@0: // monitor (good), but it still increases successor reacquisition latency -- aoqi@0: // the interval between unparking a successor and the time the successor aoqi@0: // resumes and retries the lock. See ReenterI(), which defers state transitions. aoqi@0: // If we use this technique we can also avoid EnterI()-exit() loop aoqi@0: // in ::enter() where we iteratively drop the lock and then attempt aoqi@0: // to reacquire it after suspending. aoqi@0: // aoqi@0: // B. In the future we might fold all the suspend bits into a aoqi@0: // composite per-thread suspend flag and then update it with CAS(). aoqi@0: // Alternately, a Dekker-like mechanism with multiple variables aoqi@0: // would suffice: aoqi@0: // ST Self->_suspend_equivalent = false aoqi@0: // MEMBAR aoqi@0: // LD Self_>_suspend_flags aoqi@0: // aoqi@0: aoqi@0: aoqi@0: bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { aoqi@0: int Mode = Knob_FastHSSEC ; aoqi@0: if (Mode && !jSelf->is_external_suspend()) { aoqi@0: assert (jSelf->is_suspend_equivalent(), "invariant") ; aoqi@0: jSelf->clear_suspend_equivalent() ; aoqi@0: if (2 == Mode) OrderAccess::storeload() ; aoqi@0: if (!jSelf->is_external_suspend()) return false ; aoqi@0: // We raced a suspension -- fall thru into the slow path aoqi@0: TEVENT (ExitSuspendEquivalent - raced) ; aoqi@0: jSelf->set_suspend_equivalent() ; aoqi@0: } aoqi@0: return jSelf->handle_special_suspend_equivalent_condition() ; aoqi@0: } aoqi@0: aoqi@0: aoqi@0: void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { aoqi@0: assert (_owner == Self, "invariant") ; aoqi@0: aoqi@0: // Exit protocol: aoqi@0: // 1. ST _succ = wakee aoqi@0: // 2. membar #loadstore|#storestore; aoqi@0: // 2. ST _owner = NULL aoqi@0: // 3. unpark(wakee) aoqi@0: aoqi@0: _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; aoqi@0: ParkEvent * Trigger = Wakee->_event ; aoqi@0: aoqi@0: // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. aoqi@0: // The thread associated with Wakee may have grabbed the lock and "Wakee" may be aoqi@0: // out-of-scope (non-extant). aoqi@0: Wakee = NULL ; aoqi@0: aoqi@0: // Drop the lock aoqi@0: OrderAccess::release_store_ptr (&_owner, NULL) ; aoqi@0: OrderAccess::fence() ; // ST _owner vs LD in unpark() aoqi@0: aoqi@0: if (SafepointSynchronize::do_call_back()) { aoqi@0: TEVENT (unpark before SAFEPOINT) ; aoqi@0: } aoqi@0: aoqi@0: DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); aoqi@0: Trigger->unpark() ; aoqi@0: aoqi@0: // Maintain stats and report events to JVMTI aoqi@0: if (ObjectMonitor::_sync_Parks != NULL) { aoqi@0: ObjectMonitor::_sync_Parks->inc() ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: aoqi@0: // ----------------------------------------------------------------------------- aoqi@0: // Class Loader deadlock handling. aoqi@0: // aoqi@0: // complete_exit exits a lock returning recursion count aoqi@0: // complete_exit/reenter operate as a wait without waiting aoqi@0: // complete_exit requires an inflated monitor aoqi@0: // The _owner field is not always the Thread addr even with an aoqi@0: // inflated monitor, e.g. the monitor can be inflated by a non-owning aoqi@0: // thread due to contention. aoqi@0: intptr_t ObjectMonitor::complete_exit(TRAPS) { aoqi@0: Thread * const Self = THREAD; aoqi@0: assert(Self->is_Java_thread(), "Must be Java thread!"); aoqi@0: JavaThread *jt = (JavaThread *)THREAD; aoqi@0: aoqi@0: DeferredInitialize(); aoqi@0: aoqi@0: if (THREAD != _owner) { aoqi@0: if (THREAD->is_lock_owned ((address)_owner)) { aoqi@0: assert(_recursions == 0, "internal state error"); aoqi@0: _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ aoqi@0: _recursions = 0 ; aoqi@0: OwnerIsThread = 1 ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: guarantee(Self == _owner, "complete_exit not owner"); aoqi@0: intptr_t save = _recursions; // record the old recursion count aoqi@0: _recursions = 0; // set the recursion level to be 0 aoqi@0: exit (true, Self) ; // exit the monitor aoqi@0: guarantee (_owner != Self, "invariant"); aoqi@0: return save; aoqi@0: } aoqi@0: aoqi@0: // reenter() enters a lock and sets recursion count aoqi@0: // complete_exit/reenter operate as a wait without waiting aoqi@0: void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { aoqi@0: Thread * const Self = THREAD; aoqi@0: assert(Self->is_Java_thread(), "Must be Java thread!"); aoqi@0: JavaThread *jt = (JavaThread *)THREAD; aoqi@0: aoqi@0: guarantee(_owner != Self, "reenter already owner"); aoqi@0: enter (THREAD); // enter the monitor aoqi@0: guarantee (_recursions == 0, "reenter recursion"); aoqi@0: _recursions = recursions; aoqi@0: return; aoqi@0: } aoqi@0: aoqi@0: aoqi@0: // ----------------------------------------------------------------------------- aoqi@0: // A macro is used below because there may already be a pending aoqi@0: // exception which should not abort the execution of the routines aoqi@0: // which use this (which is why we don't put this into check_slow and aoqi@0: // call it with a CHECK argument). aoqi@0: aoqi@0: #define CHECK_OWNER() \ aoqi@0: do { \ aoqi@0: if (THREAD != _owner) { \ aoqi@0: if (THREAD->is_lock_owned((address) _owner)) { \ aoqi@0: _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ aoqi@0: _recursions = 0; \ aoqi@0: OwnerIsThread = 1 ; \ aoqi@0: } else { \ aoqi@0: TEVENT (Throw IMSX) ; \ aoqi@0: THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ aoqi@0: } \ aoqi@0: } \ aoqi@0: } while (false) aoqi@0: aoqi@0: // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. aoqi@0: // TODO-FIXME: remove check_slow() -- it's likely dead. aoqi@0: aoqi@0: void ObjectMonitor::check_slow(TRAPS) { aoqi@0: TEVENT (check_slow - throw IMSX) ; aoqi@0: assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); aoqi@0: THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); aoqi@0: } aoqi@0: aoqi@0: static int Adjust (volatile int * adr, int dx) { aoqi@0: int v ; aoqi@0: for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; aoqi@0: return v ; aoqi@0: } aoqi@0: aoqi@0: // helper method for posting a monitor wait event aoqi@0: void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, aoqi@0: jlong notifier_tid, aoqi@0: jlong timeout, aoqi@0: bool timedout) { aoqi@0: event->set_klass(((oop)this->object())->klass()); aoqi@0: event->set_timeout((TYPE_ULONG)timeout); aoqi@0: event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); aoqi@0: event->set_notifier((TYPE_OSTHREAD)notifier_tid); aoqi@0: event->set_timedOut((TYPE_BOOLEAN)timedout); aoqi@0: event->commit(); aoqi@0: } aoqi@0: aoqi@0: // ----------------------------------------------------------------------------- aoqi@0: // Wait/Notify/NotifyAll aoqi@0: // aoqi@0: // Note: a subset of changes to ObjectMonitor::wait() aoqi@0: // will need to be replicated in complete_exit above aoqi@0: void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { aoqi@0: Thread * const Self = THREAD ; aoqi@0: assert(Self->is_Java_thread(), "Must be Java thread!"); aoqi@0: JavaThread *jt = (JavaThread *)THREAD; aoqi@0: aoqi@0: DeferredInitialize () ; aoqi@0: aoqi@0: // Throw IMSX or IEX. aoqi@0: CHECK_OWNER(); aoqi@0: aoqi@0: EventJavaMonitorWait event; aoqi@0: aoqi@0: // check for a pending interrupt aoqi@0: if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { aoqi@0: // post monitor waited event. Note that this is past-tense, we are done waiting. aoqi@0: if (JvmtiExport::should_post_monitor_waited()) { aoqi@0: // Note: 'false' parameter is passed here because the aoqi@0: // wait was not timed out due to thread interrupt. aoqi@0: JvmtiExport::post_monitor_waited(jt, this, false); aoqi@0: aoqi@0: // In this short circuit of the monitor wait protocol, the aoqi@0: // current thread never drops ownership of the monitor and aoqi@0: // never gets added to the wait queue so the current thread aoqi@0: // cannot be made the successor. This means that the aoqi@0: // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally aoqi@0: // consume an unpark() meant for the ParkEvent associated with aoqi@0: // this ObjectMonitor. aoqi@0: } aoqi@0: if (event.should_commit()) { aoqi@0: post_monitor_wait_event(&event, 0, millis, false); aoqi@0: } aoqi@0: TEVENT (Wait - Throw IEX) ; aoqi@0: THROW(vmSymbols::java_lang_InterruptedException()); aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: TEVENT (Wait) ; aoqi@0: aoqi@0: assert (Self->_Stalled == 0, "invariant") ; aoqi@0: Self->_Stalled = intptr_t(this) ; aoqi@0: jt->set_current_waiting_monitor(this); aoqi@0: aoqi@0: // create a node to be put into the queue aoqi@0: // Critically, after we reset() the event but prior to park(), we must check aoqi@0: // for a pending interrupt. aoqi@0: ObjectWaiter node(Self); aoqi@0: node.TState = ObjectWaiter::TS_WAIT ; aoqi@0: Self->_ParkEvent->reset() ; aoqi@0: OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag aoqi@0: aoqi@0: // Enter the waiting queue, which is a circular doubly linked list in this case aoqi@0: // but it could be a priority queue or any data structure. aoqi@0: // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only aoqi@0: // by the the owner of the monitor *except* in the case where park() aoqi@0: // returns because of a timeout of interrupt. Contention is exceptionally rare aoqi@0: // so we use a simple spin-lock instead of a heavier-weight blocking lock. aoqi@0: aoqi@0: Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; aoqi@0: AddWaiter (&node) ; aoqi@0: Thread::SpinRelease (&_WaitSetLock) ; aoqi@0: aoqi@0: if ((SyncFlags & 4) == 0) { aoqi@0: _Responsible = NULL ; aoqi@0: } aoqi@0: intptr_t save = _recursions; // record the old recursion count aoqi@0: _waiters++; // increment the number of waiters aoqi@0: _recursions = 0; // set the recursion level to be 1 aoqi@0: exit (true, Self) ; // exit the monitor aoqi@0: guarantee (_owner != Self, "invariant") ; aoqi@0: aoqi@0: // The thread is on the WaitSet list - now park() it. aoqi@0: // On MP systems it's conceivable that a brief spin before we park aoqi@0: // could be profitable. aoqi@0: // aoqi@0: // TODO-FIXME: change the following logic to a loop of the form aoqi@0: // while (!timeout && !interrupted && _notified == 0) park() aoqi@0: aoqi@0: int ret = OS_OK ; aoqi@0: int WasNotified = 0 ; aoqi@0: { // State transition wrappers aoqi@0: OSThread* osthread = Self->osthread(); aoqi@0: OSThreadWaitState osts(osthread, true); aoqi@0: { aoqi@0: ThreadBlockInVM tbivm(jt); aoqi@0: // Thread is in thread_blocked state and oop access is unsafe. aoqi@0: jt->set_suspend_equivalent(); aoqi@0: aoqi@0: if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { aoqi@0: // Intentionally empty aoqi@0: } else aoqi@0: if (node._notified == 0) { aoqi@0: if (millis <= 0) { aoqi@0: Self->_ParkEvent->park () ; aoqi@0: } else { aoqi@0: ret = Self->_ParkEvent->park (millis) ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // were we externally suspended while we were waiting? aoqi@0: if (ExitSuspendEquivalent (jt)) { aoqi@0: // TODO-FIXME: add -- if succ == Self then succ = null. aoqi@0: jt->java_suspend_self(); aoqi@0: } aoqi@0: aoqi@0: } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm aoqi@0: aoqi@0: aoqi@0: // Node may be on the WaitSet, the EntryList (or cxq), or in transition aoqi@0: // from the WaitSet to the EntryList. aoqi@0: // See if we need to remove Node from the WaitSet. aoqi@0: // We use double-checked locking to avoid grabbing _WaitSetLock aoqi@0: // if the thread is not on the wait queue. aoqi@0: // aoqi@0: // Note that we don't need a fence before the fetch of TState. aoqi@0: // In the worst case we'll fetch a old-stale value of TS_WAIT previously aoqi@0: // written by the is thread. (perhaps the fetch might even be satisfied aoqi@0: // by a look-aside into the processor's own store buffer, although given aoqi@0: // the length of the code path between the prior ST and this load that's aoqi@0: // highly unlikely). If the following LD fetches a stale TS_WAIT value aoqi@0: // then we'll acquire the lock and then re-fetch a fresh TState value. aoqi@0: // That is, we fail toward safety. aoqi@0: aoqi@0: if (node.TState == ObjectWaiter::TS_WAIT) { aoqi@0: Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; aoqi@0: if (node.TState == ObjectWaiter::TS_WAIT) { aoqi@0: DequeueSpecificWaiter (&node) ; // unlink from WaitSet aoqi@0: assert(node._notified == 0, "invariant"); aoqi@0: node.TState = ObjectWaiter::TS_RUN ; aoqi@0: } aoqi@0: Thread::SpinRelease (&_WaitSetLock) ; aoqi@0: } aoqi@0: aoqi@0: // The thread is now either on off-list (TS_RUN), aoqi@0: // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). aoqi@0: // The Node's TState variable is stable from the perspective of this thread. aoqi@0: // No other threads will asynchronously modify TState. aoqi@0: guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; aoqi@0: OrderAccess::loadload() ; aoqi@0: if (_succ == Self) _succ = NULL ; aoqi@0: WasNotified = node._notified ; aoqi@0: aoqi@0: // Reentry phase -- reacquire the monitor. aoqi@0: // re-enter contended monitor after object.wait(). aoqi@0: // retain OBJECT_WAIT state until re-enter successfully completes aoqi@0: // Thread state is thread_in_vm and oop access is again safe, aoqi@0: // although the raw address of the object may have changed. aoqi@0: // (Don't cache naked oops over safepoints, of course). aoqi@0: aoqi@0: // post monitor waited event. Note that this is past-tense, we are done waiting. aoqi@0: if (JvmtiExport::should_post_monitor_waited()) { aoqi@0: JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); aoqi@0: aoqi@0: if (node._notified != 0 && _succ == Self) { aoqi@0: // In this part of the monitor wait-notify-reenter protocol it aoqi@0: // is possible (and normal) for another thread to do a fastpath aoqi@0: // monitor enter-exit while this thread is still trying to get aoqi@0: // to the reenter portion of the protocol. aoqi@0: // aoqi@0: // The ObjectMonitor was notified and the current thread is aoqi@0: // the successor which also means that an unpark() has already aoqi@0: // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can aoqi@0: // consume the unpark() that was done when the successor was aoqi@0: // set because the same ParkEvent is shared between Java aoqi@0: // monitors and JVM/TI RawMonitors (for now). aoqi@0: // aoqi@0: // We redo the unpark() to ensure forward progress, i.e., we aoqi@0: // don't want all pending threads hanging (parked) with none aoqi@0: // entering the unlocked monitor. aoqi@0: node._event->unpark(); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: if (event.should_commit()) { aoqi@0: post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); aoqi@0: } aoqi@0: aoqi@0: OrderAccess::fence() ; aoqi@0: aoqi@0: assert (Self->_Stalled != 0, "invariant") ; aoqi@0: Self->_Stalled = 0 ; aoqi@0: aoqi@0: assert (_owner != Self, "invariant") ; aoqi@0: ObjectWaiter::TStates v = node.TState ; aoqi@0: if (v == ObjectWaiter::TS_RUN) { aoqi@0: enter (Self) ; aoqi@0: } else { aoqi@0: guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; aoqi@0: ReenterI (Self, &node) ; aoqi@0: node.wait_reenter_end(this); aoqi@0: } aoqi@0: aoqi@0: // Self has reacquired the lock. aoqi@0: // Lifecycle - the node representing Self must not appear on any queues. aoqi@0: // Node is about to go out-of-scope, but even if it were immortal we wouldn't aoqi@0: // want residual elements associated with this thread left on any lists. aoqi@0: guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; aoqi@0: assert (_owner == Self, "invariant") ; aoqi@0: assert (_succ != Self , "invariant") ; aoqi@0: } // OSThreadWaitState() aoqi@0: aoqi@0: jt->set_current_waiting_monitor(NULL); aoqi@0: aoqi@0: guarantee (_recursions == 0, "invariant") ; aoqi@0: _recursions = save; // restore the old recursion count aoqi@0: _waiters--; // decrement the number of waiters aoqi@0: aoqi@0: // Verify a few postconditions aoqi@0: assert (_owner == Self , "invariant") ; aoqi@0: assert (_succ != Self , "invariant") ; aoqi@0: assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; aoqi@0: aoqi@0: if (SyncFlags & 32) { aoqi@0: OrderAccess::fence() ; aoqi@0: } aoqi@0: aoqi@0: // check if the notification happened aoqi@0: if (!WasNotified) { aoqi@0: // no, it could be timeout or Thread.interrupt() or both aoqi@0: // check for interrupt event, otherwise it is timeout aoqi@0: if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { aoqi@0: TEVENT (Wait - throw IEX from epilog) ; aoqi@0: THROW(vmSymbols::java_lang_InterruptedException()); aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // NOTE: Spurious wake up will be consider as timeout. aoqi@0: // Monitor notify has precedence over thread interrupt. aoqi@0: } aoqi@0: aoqi@0: aoqi@0: // Consider: aoqi@0: // If the lock is cool (cxq == null && succ == null) and we're on an MP system aoqi@0: // then instead of transferring a thread from the WaitSet to the EntryList aoqi@0: // we might just dequeue a thread from the WaitSet and directly unpark() it. aoqi@0: aoqi@0: void ObjectMonitor::notify(TRAPS) { aoqi@0: CHECK_OWNER(); aoqi@0: if (_WaitSet == NULL) { aoqi@0: TEVENT (Empty-Notify) ; aoqi@0: return ; aoqi@0: } aoqi@0: DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); aoqi@0: aoqi@0: int Policy = Knob_MoveNotifyee ; aoqi@0: aoqi@0: Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; aoqi@0: ObjectWaiter * iterator = DequeueWaiter() ; aoqi@0: if (iterator != NULL) { aoqi@0: TEVENT (Notify1 - Transfer) ; aoqi@0: guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; aoqi@0: guarantee (iterator->_notified == 0, "invariant") ; aoqi@0: if (Policy != 4) { aoqi@0: iterator->TState = ObjectWaiter::TS_ENTER ; aoqi@0: } aoqi@0: iterator->_notified = 1 ; aoqi@0: Thread * Self = THREAD; aoqi@0: iterator->_notifier_tid = Self->osthread()->thread_id(); aoqi@0: aoqi@0: ObjectWaiter * List = _EntryList ; aoqi@0: if (List != NULL) { aoqi@0: assert (List->_prev == NULL, "invariant") ; aoqi@0: assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; aoqi@0: assert (List != iterator, "invariant") ; aoqi@0: } aoqi@0: aoqi@0: if (Policy == 0) { // prepend to EntryList aoqi@0: if (List == NULL) { aoqi@0: iterator->_next = iterator->_prev = NULL ; aoqi@0: _EntryList = iterator ; aoqi@0: } else { aoqi@0: List->_prev = iterator ; aoqi@0: iterator->_next = List ; aoqi@0: iterator->_prev = NULL ; aoqi@0: _EntryList = iterator ; aoqi@0: } aoqi@0: } else aoqi@0: if (Policy == 1) { // append to EntryList aoqi@0: if (List == NULL) { aoqi@0: iterator->_next = iterator->_prev = NULL ; aoqi@0: _EntryList = iterator ; aoqi@0: } else { aoqi@0: // CONSIDER: finding the tail currently requires a linear-time walk of aoqi@0: // the EntryList. We can make tail access constant-time by converting to aoqi@0: // a CDLL instead of using our current DLL. aoqi@0: ObjectWaiter * Tail ; aoqi@0: for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; aoqi@0: assert (Tail != NULL && Tail->_next == NULL, "invariant") ; aoqi@0: Tail->_next = iterator ; aoqi@0: iterator->_prev = Tail ; aoqi@0: iterator->_next = NULL ; aoqi@0: } aoqi@0: } else aoqi@0: if (Policy == 2) { // prepend to cxq aoqi@0: // prepend to cxq aoqi@0: if (List == NULL) { aoqi@0: iterator->_next = iterator->_prev = NULL ; aoqi@0: _EntryList = iterator ; aoqi@0: } else { aoqi@0: iterator->TState = ObjectWaiter::TS_CXQ ; aoqi@0: for (;;) { aoqi@0: ObjectWaiter * Front = _cxq ; aoqi@0: iterator->_next = Front ; aoqi@0: if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { aoqi@0: break ; aoqi@0: } aoqi@0: } aoqi@0: } aoqi@0: } else aoqi@0: if (Policy == 3) { // append to cxq aoqi@0: iterator->TState = ObjectWaiter::TS_CXQ ; aoqi@0: for (;;) { aoqi@0: ObjectWaiter * Tail ; aoqi@0: Tail = _cxq ; aoqi@0: if (Tail == NULL) { aoqi@0: iterator->_next = NULL ; aoqi@0: if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { aoqi@0: break ; aoqi@0: } aoqi@0: } else { aoqi@0: while (Tail->_next != NULL) Tail = Tail->_next ; aoqi@0: Tail->_next = iterator ; aoqi@0: iterator->_prev = Tail ; aoqi@0: iterator->_next = NULL ; aoqi@0: break ; aoqi@0: } aoqi@0: } aoqi@0: } else { aoqi@0: ParkEvent * ev = iterator->_event ; aoqi@0: iterator->TState = ObjectWaiter::TS_RUN ; aoqi@0: OrderAccess::fence() ; aoqi@0: ev->unpark() ; aoqi@0: } aoqi@0: aoqi@0: if (Policy < 4) { aoqi@0: iterator->wait_reenter_begin(this); aoqi@0: } aoqi@0: aoqi@0: // _WaitSetLock protects the wait queue, not the EntryList. We could aoqi@0: // move the add-to-EntryList operation, above, outside the critical section aoqi@0: // protected by _WaitSetLock. In practice that's not useful. With the aoqi@0: // exception of wait() timeouts and interrupts the monitor owner aoqi@0: // is the only thread that grabs _WaitSetLock. There's almost no contention aoqi@0: // on _WaitSetLock so it's not profitable to reduce the length of the aoqi@0: // critical section. aoqi@0: } aoqi@0: aoqi@0: Thread::SpinRelease (&_WaitSetLock) ; aoqi@0: aoqi@0: if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { aoqi@0: ObjectMonitor::_sync_Notifications->inc() ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: aoqi@0: void ObjectMonitor::notifyAll(TRAPS) { aoqi@0: CHECK_OWNER(); aoqi@0: ObjectWaiter* iterator; aoqi@0: if (_WaitSet == NULL) { aoqi@0: TEVENT (Empty-NotifyAll) ; aoqi@0: return ; aoqi@0: } aoqi@0: DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); aoqi@0: aoqi@0: int Policy = Knob_MoveNotifyee ; aoqi@0: int Tally = 0 ; aoqi@0: Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; aoqi@0: aoqi@0: for (;;) { aoqi@0: iterator = DequeueWaiter () ; aoqi@0: if (iterator == NULL) break ; aoqi@0: TEVENT (NotifyAll - Transfer1) ; aoqi@0: ++Tally ; aoqi@0: aoqi@0: // Disposition - what might we do with iterator ? aoqi@0: // a. add it directly to the EntryList - either tail or head. aoqi@0: // b. push it onto the front of the _cxq. aoqi@0: // For now we use (a). aoqi@0: aoqi@0: guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; aoqi@0: guarantee (iterator->_notified == 0, "invariant") ; aoqi@0: iterator->_notified = 1 ; aoqi@0: Thread * Self = THREAD; aoqi@0: iterator->_notifier_tid = Self->osthread()->thread_id(); aoqi@0: if (Policy != 4) { aoqi@0: iterator->TState = ObjectWaiter::TS_ENTER ; aoqi@0: } aoqi@0: aoqi@0: ObjectWaiter * List = _EntryList ; aoqi@0: if (List != NULL) { aoqi@0: assert (List->_prev == NULL, "invariant") ; aoqi@0: assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; aoqi@0: assert (List != iterator, "invariant") ; aoqi@0: } aoqi@0: aoqi@0: if (Policy == 0) { // prepend to EntryList aoqi@0: if (List == NULL) { aoqi@0: iterator->_next = iterator->_prev = NULL ; aoqi@0: _EntryList = iterator ; aoqi@0: } else { aoqi@0: List->_prev = iterator ; aoqi@0: iterator->_next = List ; aoqi@0: iterator->_prev = NULL ; aoqi@0: _EntryList = iterator ; aoqi@0: } aoqi@0: } else aoqi@0: if (Policy == 1) { // append to EntryList aoqi@0: if (List == NULL) { aoqi@0: iterator->_next = iterator->_prev = NULL ; aoqi@0: _EntryList = iterator ; aoqi@0: } else { aoqi@0: // CONSIDER: finding the tail currently requires a linear-time walk of aoqi@0: // the EntryList. We can make tail access constant-time by converting to aoqi@0: // a CDLL instead of using our current DLL. aoqi@0: ObjectWaiter * Tail ; aoqi@0: for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; aoqi@0: assert (Tail != NULL && Tail->_next == NULL, "invariant") ; aoqi@0: Tail->_next = iterator ; aoqi@0: iterator->_prev = Tail ; aoqi@0: iterator->_next = NULL ; aoqi@0: } aoqi@0: } else aoqi@0: if (Policy == 2) { // prepend to cxq aoqi@0: // prepend to cxq aoqi@0: iterator->TState = ObjectWaiter::TS_CXQ ; aoqi@0: for (;;) { aoqi@0: ObjectWaiter * Front = _cxq ; aoqi@0: iterator->_next = Front ; aoqi@0: if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { aoqi@0: break ; aoqi@0: } aoqi@0: } aoqi@0: } else aoqi@0: if (Policy == 3) { // append to cxq aoqi@0: iterator->TState = ObjectWaiter::TS_CXQ ; aoqi@0: for (;;) { aoqi@0: ObjectWaiter * Tail ; aoqi@0: Tail = _cxq ; aoqi@0: if (Tail == NULL) { aoqi@0: iterator->_next = NULL ; aoqi@0: if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { aoqi@0: break ; aoqi@0: } aoqi@0: } else { aoqi@0: while (Tail->_next != NULL) Tail = Tail->_next ; aoqi@0: Tail->_next = iterator ; aoqi@0: iterator->_prev = Tail ; aoqi@0: iterator->_next = NULL ; aoqi@0: break ; aoqi@0: } aoqi@0: } aoqi@0: } else { aoqi@0: ParkEvent * ev = iterator->_event ; aoqi@0: iterator->TState = ObjectWaiter::TS_RUN ; aoqi@0: OrderAccess::fence() ; aoqi@0: ev->unpark() ; aoqi@0: } aoqi@0: aoqi@0: if (Policy < 4) { aoqi@0: iterator->wait_reenter_begin(this); aoqi@0: } aoqi@0: aoqi@0: // _WaitSetLock protects the wait queue, not the EntryList. We could aoqi@0: // move the add-to-EntryList operation, above, outside the critical section aoqi@0: // protected by _WaitSetLock. In practice that's not useful. With the aoqi@0: // exception of wait() timeouts and interrupts the monitor owner aoqi@0: // is the only thread that grabs _WaitSetLock. There's almost no contention aoqi@0: // on _WaitSetLock so it's not profitable to reduce the length of the aoqi@0: // critical section. aoqi@0: } aoqi@0: aoqi@0: Thread::SpinRelease (&_WaitSetLock) ; aoqi@0: aoqi@0: if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { aoqi@0: ObjectMonitor::_sync_Notifications->inc(Tally) ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: // ----------------------------------------------------------------------------- aoqi@0: // Adaptive Spinning Support aoqi@0: // aoqi@0: // Adaptive spin-then-block - rational spinning aoqi@0: // aoqi@0: // Note that we spin "globally" on _owner with a classic SMP-polite TATAS aoqi@0: // algorithm. On high order SMP systems it would be better to start with aoqi@0: // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, aoqi@0: // a contending thread could enqueue itself on the cxq and then spin locally aoqi@0: // on a thread-specific variable such as its ParkEvent._Event flag. aoqi@0: // That's left as an exercise for the reader. Note that global spinning is aoqi@0: // not problematic on Niagara, as the L2$ serves the interconnect and has both aoqi@0: // low latency and massive bandwidth. aoqi@0: // aoqi@0: // Broadly, we can fix the spin frequency -- that is, the % of contended lock aoqi@0: // acquisition attempts where we opt to spin -- at 100% and vary the spin count aoqi@0: // (duration) or we can fix the count at approximately the duration of aoqi@0: // a context switch and vary the frequency. Of course we could also aoqi@0: // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. dbuck@8067: // For a description of 'Adaptive spin-then-block mutual exclusion in dbuck@8067: // multi-threaded processing,' see U.S. Pat. No. 8046758. aoqi@0: // aoqi@0: // This implementation varies the duration "D", where D varies with aoqi@0: // the success rate of recent spin attempts. (D is capped at approximately aoqi@0: // length of a round-trip context switch). The success rate for recent aoqi@0: // spin attempts is a good predictor of the success rate of future spin aoqi@0: // attempts. The mechanism adapts automatically to varying critical aoqi@0: // section length (lock modality), system load and degree of parallelism. aoqi@0: // D is maintained per-monitor in _SpinDuration and is initialized aoqi@0: // optimistically. Spin frequency is fixed at 100%. aoqi@0: // aoqi@0: // Note that _SpinDuration is volatile, but we update it without locks aoqi@0: // or atomics. The code is designed so that _SpinDuration stays within aoqi@0: // a reasonable range even in the presence of races. The arithmetic aoqi@0: // operations on _SpinDuration are closed over the domain of legal values, aoqi@0: // so at worst a race will install and older but still legal value. aoqi@0: // At the very worst this introduces some apparent non-determinism. aoqi@0: // We might spin when we shouldn't or vice-versa, but since the spin aoqi@0: // count are relatively short, even in the worst case, the effect is harmless. aoqi@0: // aoqi@0: // Care must be taken that a low "D" value does not become an aoqi@0: // an absorbing state. Transient spinning failures -- when spinning aoqi@0: // is overall profitable -- should not cause the system to converge aoqi@0: // on low "D" values. We want spinning to be stable and predictable aoqi@0: // and fairly responsive to change and at the same time we don't want aoqi@0: // it to oscillate, become metastable, be "too" non-deterministic, aoqi@0: // or converge on or enter undesirable stable absorbing states. aoqi@0: // aoqi@0: // We implement a feedback-based control system -- using past behavior aoqi@0: // to predict future behavior. We face two issues: (a) if the aoqi@0: // input signal is random then the spin predictor won't provide optimal aoqi@0: // results, and (b) if the signal frequency is too high then the control aoqi@0: // system, which has some natural response lag, will "chase" the signal. aoqi@0: // (b) can arise from multimodal lock hold times. Transient preemption aoqi@0: // can also result in apparent bimodal lock hold times. aoqi@0: // Although sub-optimal, neither condition is particularly harmful, as aoqi@0: // in the worst-case we'll spin when we shouldn't or vice-versa. aoqi@0: // The maximum spin duration is rather short so the failure modes aren't bad. aoqi@0: // To be conservative, I've tuned the gain in system to bias toward aoqi@0: // _not spinning. Relatedly, the system can sometimes enter a mode where it aoqi@0: // "rings" or oscillates between spinning and not spinning. This happens aoqi@0: // when spinning is just on the cusp of profitability, however, so the aoqi@0: // situation is not dire. The state is benign -- there's no need to add aoqi@0: // hysteresis control to damp the transition rate between spinning and aoqi@0: // not spinning. aoqi@0: // aoqi@0: aoqi@0: intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; aoqi@0: int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; aoqi@0: aoqi@0: // Spinning: Fixed frequency (100%), vary duration aoqi@0: aoqi@0: aoqi@0: int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { aoqi@0: aoqi@0: // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. aoqi@0: int ctr = Knob_FixedSpin ; aoqi@0: if (ctr != 0) { aoqi@0: while (--ctr >= 0) { aoqi@0: if (TryLock (Self) > 0) return 1 ; aoqi@0: SpinPause () ; aoqi@0: } aoqi@0: return 0 ; aoqi@0: } aoqi@0: aoqi@0: for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { aoqi@0: if (TryLock(Self) > 0) { aoqi@0: // Increase _SpinDuration ... aoqi@0: // Note that we don't clamp SpinDuration precisely at SpinLimit. aoqi@0: // Raising _SpurDuration to the poverty line is key. aoqi@0: int x = _SpinDuration ; aoqi@0: if (x < Knob_SpinLimit) { aoqi@0: if (x < Knob_Poverty) x = Knob_Poverty ; aoqi@0: _SpinDuration = x + Knob_BonusB ; aoqi@0: } aoqi@0: return 1 ; aoqi@0: } aoqi@0: SpinPause () ; aoqi@0: } aoqi@0: aoqi@0: // Admission control - verify preconditions for spinning aoqi@0: // aoqi@0: // We always spin a little bit, just to prevent _SpinDuration == 0 from aoqi@0: // becoming an absorbing state. Put another way, we spin briefly to aoqi@0: // sample, just in case the system load, parallelism, contention, or lock aoqi@0: // modality changed. aoqi@0: // aoqi@0: // Consider the following alternative: aoqi@0: // Periodically set _SpinDuration = _SpinLimit and try a long/full aoqi@0: // spin attempt. "Periodically" might mean after a tally of aoqi@0: // the # of failed spin attempts (or iterations) reaches some threshold. aoqi@0: // This takes us into the realm of 1-out-of-N spinning, where we aoqi@0: // hold the duration constant but vary the frequency. aoqi@0: aoqi@0: ctr = _SpinDuration ; aoqi@0: if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; aoqi@0: if (ctr <= 0) return 0 ; aoqi@0: aoqi@0: if (Knob_SuccRestrict && _succ != NULL) return 0 ; aoqi@0: if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { aoqi@0: TEVENT (Spin abort - notrunnable [TOP]); aoqi@0: return 0 ; aoqi@0: } aoqi@0: aoqi@0: int MaxSpin = Knob_MaxSpinners ; aoqi@0: if (MaxSpin >= 0) { aoqi@0: if (_Spinner > MaxSpin) { aoqi@0: TEVENT (Spin abort -- too many spinners) ; aoqi@0: return 0 ; aoqi@0: } aoqi@0: // Slighty racy, but benign ... aoqi@0: Adjust (&_Spinner, 1) ; aoqi@0: } aoqi@0: aoqi@0: // We're good to spin ... spin ingress. aoqi@0: // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades aoqi@0: // when preparing to LD...CAS _owner, etc and the CAS is likely aoqi@0: // to succeed. aoqi@0: int hits = 0 ; aoqi@0: int msk = 0 ; aoqi@0: int caspty = Knob_CASPenalty ; aoqi@0: int oxpty = Knob_OXPenalty ; aoqi@0: int sss = Knob_SpinSetSucc ; aoqi@0: if (sss && _succ == NULL ) _succ = Self ; aoqi@0: Thread * prv = NULL ; aoqi@0: aoqi@0: // There are three ways to exit the following loop: aoqi@0: // 1. A successful spin where this thread has acquired the lock. aoqi@0: // 2. Spin failure with prejudice aoqi@0: // 3. Spin failure without prejudice aoqi@0: aoqi@0: while (--ctr >= 0) { aoqi@0: aoqi@0: // Periodic polling -- Check for pending GC aoqi@0: // Threads may spin while they're unsafe. aoqi@0: // We don't want spinning threads to delay the JVM from reaching aoqi@0: // a stop-the-world safepoint or to steal cycles from GC. aoqi@0: // If we detect a pending safepoint we abort in order that aoqi@0: // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) aoqi@0: // this thread, if safe, doesn't steal cycles from GC. aoqi@0: // This is in keeping with the "no loitering in runtime" rule. aoqi@0: // We periodically check to see if there's a safepoint pending. aoqi@0: if ((ctr & 0xFF) == 0) { aoqi@0: if (SafepointSynchronize::do_call_back()) { aoqi@0: TEVENT (Spin: safepoint) ; aoqi@0: goto Abort ; // abrupt spin egress aoqi@0: } aoqi@0: if (Knob_UsePause & 1) SpinPause () ; aoqi@0: aoqi@0: int (*scb)(intptr_t,int) = SpinCallbackFunction ; aoqi@0: if (hits > 50 && scb != NULL) { aoqi@0: int abend = (*scb)(SpinCallbackArgument, 0) ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: if (Knob_UsePause & 2) SpinPause() ; aoqi@0: aoqi@0: // Exponential back-off ... Stay off the bus to reduce coherency traffic. aoqi@0: // This is useful on classic SMP systems, but is of less utility on aoqi@0: // N1-style CMT platforms. aoqi@0: // aoqi@0: // Trade-off: lock acquisition latency vs coherency bandwidth. aoqi@0: // Lock hold times are typically short. A histogram aoqi@0: // of successful spin attempts shows that we usually acquire aoqi@0: // the lock early in the spin. That suggests we want to aoqi@0: // sample _owner frequently in the early phase of the spin, aoqi@0: // but then back-off and sample less frequently as the spin aoqi@0: // progresses. The back-off makes a good citizen on SMP big aoqi@0: // SMP systems. Oversampling _owner can consume excessive aoqi@0: // coherency bandwidth. Relatedly, if we _oversample _owner we aoqi@0: // can inadvertently interfere with the the ST m->owner=null. aoqi@0: // executed by the lock owner. aoqi@0: if (ctr & msk) continue ; aoqi@0: ++hits ; aoqi@0: if ((hits & 0xF) == 0) { aoqi@0: // The 0xF, above, corresponds to the exponent. aoqi@0: // Consider: (msk+1)|msk aoqi@0: msk = ((msk << 2)|3) & BackOffMask ; aoqi@0: } aoqi@0: aoqi@0: // Probe _owner with TATAS aoqi@0: // If this thread observes the monitor transition or flicker aoqi@0: // from locked to unlocked to locked, then the odds that this aoqi@0: // thread will acquire the lock in this spin attempt go down aoqi@0: // considerably. The same argument applies if the CAS fails aoqi@0: // or if we observe _owner change from one non-null value to aoqi@0: // another non-null value. In such cases we might abort aoqi@0: // the spin without prejudice or apply a "penalty" to the aoqi@0: // spin count-down variable "ctr", reducing it by 100, say. aoqi@0: aoqi@0: Thread * ox = (Thread *) _owner ; aoqi@0: if (ox == NULL) { aoqi@0: ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; aoqi@0: if (ox == NULL) { aoqi@0: // The CAS succeeded -- this thread acquired ownership aoqi@0: // Take care of some bookkeeping to exit spin state. aoqi@0: if (sss && _succ == Self) { aoqi@0: _succ = NULL ; aoqi@0: } aoqi@0: if (MaxSpin > 0) Adjust (&_Spinner, -1) ; aoqi@0: aoqi@0: // Increase _SpinDuration : aoqi@0: // The spin was successful (profitable) so we tend toward aoqi@0: // longer spin attempts in the future. aoqi@0: // CONSIDER: factor "ctr" into the _SpinDuration adjustment. aoqi@0: // If we acquired the lock early in the spin cycle it aoqi@0: // makes sense to increase _SpinDuration proportionally. aoqi@0: // Note that we don't clamp SpinDuration precisely at SpinLimit. aoqi@0: int x = _SpinDuration ; aoqi@0: if (x < Knob_SpinLimit) { aoqi@0: if (x < Knob_Poverty) x = Knob_Poverty ; aoqi@0: _SpinDuration = x + Knob_Bonus ; aoqi@0: } aoqi@0: return 1 ; aoqi@0: } aoqi@0: aoqi@0: // The CAS failed ... we can take any of the following actions: aoqi@0: // * penalize: ctr -= Knob_CASPenalty aoqi@0: // * exit spin with prejudice -- goto Abort; aoqi@0: // * exit spin without prejudice. aoqi@0: // * Since CAS is high-latency, retry again immediately. aoqi@0: prv = ox ; aoqi@0: TEVENT (Spin: cas failed) ; aoqi@0: if (caspty == -2) break ; aoqi@0: if (caspty == -1) goto Abort ; aoqi@0: ctr -= caspty ; aoqi@0: continue ; aoqi@0: } aoqi@0: aoqi@0: // Did lock ownership change hands ? aoqi@0: if (ox != prv && prv != NULL ) { aoqi@0: TEVENT (spin: Owner changed) aoqi@0: if (oxpty == -2) break ; aoqi@0: if (oxpty == -1) goto Abort ; aoqi@0: ctr -= oxpty ; aoqi@0: } aoqi@0: prv = ox ; aoqi@0: aoqi@0: // Abort the spin if the owner is not executing. aoqi@0: // The owner must be executing in order to drop the lock. aoqi@0: // Spinning while the owner is OFFPROC is idiocy. aoqi@0: // Consider: ctr -= RunnablePenalty ; aoqi@0: if (Knob_OState && NotRunnable (Self, ox)) { aoqi@0: TEVENT (Spin abort - notrunnable); aoqi@0: goto Abort ; aoqi@0: } aoqi@0: if (sss && _succ == NULL ) _succ = Self ; aoqi@0: } aoqi@0: aoqi@0: // Spin failed with prejudice -- reduce _SpinDuration. aoqi@0: // TODO: Use an AIMD-like policy to adjust _SpinDuration. aoqi@0: // AIMD is globally stable. aoqi@0: TEVENT (Spin failure) ; aoqi@0: { aoqi@0: int x = _SpinDuration ; aoqi@0: if (x > 0) { aoqi@0: // Consider an AIMD scheme like: x -= (x >> 3) + 100 aoqi@0: // This is globally sample and tends to damp the response. aoqi@0: x -= Knob_Penalty ; aoqi@0: if (x < 0) x = 0 ; aoqi@0: _SpinDuration = x ; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: Abort: aoqi@0: if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; aoqi@0: if (sss && _succ == Self) { aoqi@0: _succ = NULL ; aoqi@0: // Invariant: after setting succ=null a contending thread aoqi@0: // must recheck-retry _owner before parking. This usually happens aoqi@0: // in the normal usage of TrySpin(), but it's safest aoqi@0: // to make TrySpin() as foolproof as possible. aoqi@0: OrderAccess::fence() ; aoqi@0: if (TryLock(Self) > 0) return 1 ; aoqi@0: } aoqi@0: return 0 ; aoqi@0: } aoqi@0: aoqi@0: // NotRunnable() -- informed spinning aoqi@0: // aoqi@0: // Don't bother spinning if the owner is not eligible to drop the lock. aoqi@0: // Peek at the owner's schedctl.sc_state and Thread._thread_values and aoqi@0: // spin only if the owner thread is _thread_in_Java or _thread_in_vm. aoqi@0: // The thread must be runnable in order to drop the lock in timely fashion. aoqi@0: // If the _owner is not runnable then spinning will not likely be aoqi@0: // successful (profitable). aoqi@0: // aoqi@0: // Beware -- the thread referenced by _owner could have died aoqi@0: // so a simply fetch from _owner->_thread_state might trap. aoqi@0: // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. aoqi@0: // Because of the lifecycle issues the schedctl and _thread_state values aoqi@0: // observed by NotRunnable() might be garbage. NotRunnable must aoqi@0: // tolerate this and consider the observed _thread_state value aoqi@0: // as advisory. aoqi@0: // aoqi@0: // Beware too, that _owner is sometimes a BasicLock address and sometimes aoqi@0: // a thread pointer. We differentiate the two cases with OwnerIsThread. aoqi@0: // Alternately, we might tag the type (thread pointer vs basiclock pointer) aoqi@0: // with the LSB of _owner. Another option would be to probablistically probe aoqi@0: // the putative _owner->TypeTag value. aoqi@0: // aoqi@0: // Checking _thread_state isn't perfect. Even if the thread is aoqi@0: // in_java it might be blocked on a page-fault or have been preempted aoqi@0: // and sitting on a ready/dispatch queue. _thread state in conjunction aoqi@0: // with schedctl.sc_state gives us a good picture of what the aoqi@0: // thread is doing, however. aoqi@0: // aoqi@0: // TODO: check schedctl.sc_state. aoqi@0: // We'll need to use SafeFetch32() to read from the schedctl block. aoqi@0: // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ aoqi@0: // aoqi@0: // The return value from NotRunnable() is *advisory* -- the aoqi@0: // result is based on sampling and is not necessarily coherent. aoqi@0: // The caller must tolerate false-negative and false-positive errors. aoqi@0: // Spinning, in general, is probabilistic anyway. aoqi@0: aoqi@0: aoqi@0: int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { aoqi@0: // Check either OwnerIsThread or ox->TypeTag == 2BAD. aoqi@0: if (!OwnerIsThread) return 0 ; aoqi@0: aoqi@0: if (ox == NULL) return 0 ; aoqi@0: aoqi@0: // Avoid transitive spinning ... aoqi@0: // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. aoqi@0: // Immediately after T1 acquires L it's possible that T2, also aoqi@0: // spinning on L, will see L.Owner=T1 and T1._Stalled=L. aoqi@0: // This occurs transiently after T1 acquired L but before aoqi@0: // T1 managed to clear T1.Stalled. T2 does not need to abort aoqi@0: // its spin in this circumstance. aoqi@0: intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; aoqi@0: aoqi@0: if (BlockedOn == 1) return 1 ; aoqi@0: if (BlockedOn != 0) { aoqi@0: return BlockedOn != intptr_t(this) && _owner == ox ; aoqi@0: } aoqi@0: aoqi@0: assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; aoqi@0: int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; aoqi@0: // consider also: jst != _thread_in_Java -- but that's overspecific. aoqi@0: return jst == _thread_blocked || jst == _thread_in_native ; aoqi@0: } aoqi@0: aoqi@0: aoqi@0: // ----------------------------------------------------------------------------- aoqi@0: // WaitSet management ... aoqi@0: aoqi@0: ObjectWaiter::ObjectWaiter(Thread* thread) { aoqi@0: _next = NULL; aoqi@0: _prev = NULL; aoqi@0: _notified = 0; aoqi@0: TState = TS_RUN ; aoqi@0: _thread = thread; aoqi@0: _event = thread->_ParkEvent ; aoqi@0: _active = false; aoqi@0: assert (_event != NULL, "invariant") ; aoqi@0: } aoqi@0: aoqi@0: void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { aoqi@0: JavaThread *jt = (JavaThread *)this->_thread; aoqi@0: _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); aoqi@0: } aoqi@0: aoqi@0: void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { aoqi@0: JavaThread *jt = (JavaThread *)this->_thread; aoqi@0: JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); aoqi@0: } aoqi@0: aoqi@0: inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { aoqi@0: assert(node != NULL, "should not dequeue NULL node"); aoqi@0: assert(node->_prev == NULL, "node already in list"); aoqi@0: assert(node->_next == NULL, "node already in list"); aoqi@0: // put node at end of queue (circular doubly linked list) aoqi@0: if (_WaitSet == NULL) { aoqi@0: _WaitSet = node; aoqi@0: node->_prev = node; aoqi@0: node->_next = node; aoqi@0: } else { aoqi@0: ObjectWaiter* head = _WaitSet ; aoqi@0: ObjectWaiter* tail = head->_prev; aoqi@0: assert(tail->_next == head, "invariant check"); aoqi@0: tail->_next = node; aoqi@0: head->_prev = node; aoqi@0: node->_next = head; aoqi@0: node->_prev = tail; aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { aoqi@0: // dequeue the very first waiter aoqi@0: ObjectWaiter* waiter = _WaitSet; aoqi@0: if (waiter) { aoqi@0: DequeueSpecificWaiter(waiter); aoqi@0: } aoqi@0: return waiter; aoqi@0: } aoqi@0: aoqi@0: inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { aoqi@0: assert(node != NULL, "should not dequeue NULL node"); aoqi@0: assert(node->_prev != NULL, "node already removed from list"); aoqi@0: assert(node->_next != NULL, "node already removed from list"); aoqi@0: // when the waiter has woken up because of interrupt, aoqi@0: // timeout or other spurious wake-up, dequeue the aoqi@0: // waiter from waiting list aoqi@0: ObjectWaiter* next = node->_next; aoqi@0: if (next == node) { aoqi@0: assert(node->_prev == node, "invariant check"); aoqi@0: _WaitSet = NULL; aoqi@0: } else { aoqi@0: ObjectWaiter* prev = node->_prev; aoqi@0: assert(prev->_next == node, "invariant check"); aoqi@0: assert(next->_prev == node, "invariant check"); aoqi@0: next->_prev = prev; aoqi@0: prev->_next = next; aoqi@0: if (_WaitSet == node) { aoqi@0: _WaitSet = next; aoqi@0: } aoqi@0: } aoqi@0: node->_next = NULL; aoqi@0: node->_prev = NULL; aoqi@0: } aoqi@0: aoqi@0: // ----------------------------------------------------------------------------- aoqi@0: // PerfData support aoqi@0: PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_Parks = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; aoqi@0: PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; aoqi@0: PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; aoqi@0: aoqi@0: // One-shot global initialization for the sync subsystem. aoqi@0: // We could also defer initialization and initialize on-demand aoqi@0: // the first time we call inflate(). Initialization would aoqi@0: // be protected - like so many things - by the MonitorCache_lock. aoqi@0: aoqi@0: void ObjectMonitor::Initialize () { aoqi@0: static int InitializationCompleted = 0 ; aoqi@0: assert (InitializationCompleted == 0, "invariant") ; aoqi@0: InitializationCompleted = 1 ; aoqi@0: if (UsePerfData) { aoqi@0: EXCEPTION_MARK ; aoqi@0: #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } aoqi@0: #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } aoqi@0: NEWPERFCOUNTER(_sync_Inflations) ; aoqi@0: NEWPERFCOUNTER(_sync_Deflations) ; aoqi@0: NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; aoqi@0: NEWPERFCOUNTER(_sync_FutileWakeups) ; aoqi@0: NEWPERFCOUNTER(_sync_Parks) ; aoqi@0: NEWPERFCOUNTER(_sync_EmptyNotifications) ; aoqi@0: NEWPERFCOUNTER(_sync_Notifications) ; aoqi@0: NEWPERFCOUNTER(_sync_SlowEnter) ; aoqi@0: NEWPERFCOUNTER(_sync_SlowExit) ; aoqi@0: NEWPERFCOUNTER(_sync_SlowNotify) ; aoqi@0: NEWPERFCOUNTER(_sync_SlowNotifyAll) ; aoqi@0: NEWPERFCOUNTER(_sync_FailedSpins) ; aoqi@0: NEWPERFCOUNTER(_sync_SuccessfulSpins) ; aoqi@0: NEWPERFCOUNTER(_sync_PrivateA) ; aoqi@0: NEWPERFCOUNTER(_sync_PrivateB) ; aoqi@0: NEWPERFCOUNTER(_sync_MonInCirculation) ; aoqi@0: NEWPERFCOUNTER(_sync_MonScavenged) ; aoqi@0: NEWPERFVARIABLE(_sync_MonExtant) ; aoqi@0: #undef NEWPERFCOUNTER aoqi@0: } aoqi@0: } aoqi@0: aoqi@0: aoqi@0: // Compile-time asserts aoqi@0: // When possible, it's better to catch errors deterministically at aoqi@0: // compile-time than at runtime. The down-side to using compile-time aoqi@0: // asserts is that error message -- often something about negative array aoqi@0: // indices -- is opaque. aoqi@0: aoqi@0: #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } aoqi@0: aoqi@0: void ObjectMonitor::ctAsserts() { aoqi@0: CTASSERT(offset_of (ObjectMonitor, _header) == 0); aoqi@0: } aoqi@0: aoqi@0: aoqi@0: static char * kvGet (char * kvList, const char * Key) { aoqi@0: if (kvList == NULL) return NULL ; aoqi@0: size_t n = strlen (Key) ; aoqi@0: char * Search ; aoqi@0: for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { aoqi@0: if (strncmp (Search, Key, n) == 0) { aoqi@0: if (Search[n] == '=') return Search + n + 1 ; aoqi@0: if (Search[n] == 0) return (char *) "1" ; aoqi@0: } aoqi@0: } aoqi@0: return NULL ; aoqi@0: } aoqi@0: aoqi@0: static int kvGetInt (char * kvList, const char * Key, int Default) { aoqi@0: char * v = kvGet (kvList, Key) ; aoqi@0: int rslt = v ? ::strtol (v, NULL, 0) : Default ; aoqi@0: if (Knob_ReportSettings && v != NULL) { aoqi@0: ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; aoqi@0: ::fflush (stdout) ; aoqi@0: } aoqi@0: return rslt ; aoqi@0: } aoqi@0: aoqi@0: void ObjectMonitor::DeferredInitialize () { aoqi@0: if (InitDone > 0) return ; aoqi@0: if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { aoqi@0: while (InitDone != 1) ; aoqi@0: return ; aoqi@0: } aoqi@0: aoqi@0: // One-shot global initialization ... aoqi@0: // The initialization is idempotent, so we don't need locks. aoqi@0: // In the future consider doing this via os::init_2(). aoqi@0: // SyncKnobs consist of = pairs in the style aoqi@0: // of environment variables. Start by converting ':' to NUL. aoqi@0: aoqi@0: if (SyncKnobs == NULL) SyncKnobs = "" ; aoqi@0: aoqi@0: size_t sz = strlen (SyncKnobs) ; aoqi@0: char * knobs = (char *) malloc (sz + 2) ; aoqi@0: if (knobs == NULL) { aoqi@0: vm_exit_out_of_memory (sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs") ; aoqi@0: guarantee (0, "invariant") ; aoqi@0: } aoqi@0: strcpy (knobs, SyncKnobs) ; aoqi@0: knobs[sz+1] = 0 ; aoqi@0: for (char * p = knobs ; *p ; p++) { aoqi@0: if (*p == ':') *p = 0 ; aoqi@0: } aoqi@0: aoqi@0: #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } aoqi@0: SETKNOB(ReportSettings) ; aoqi@0: SETKNOB(Verbose) ; aoqi@0: SETKNOB(FixedSpin) ; aoqi@0: SETKNOB(SpinLimit) ; aoqi@0: SETKNOB(SpinBase) ; aoqi@0: SETKNOB(SpinBackOff); aoqi@0: SETKNOB(CASPenalty) ; aoqi@0: SETKNOB(OXPenalty) ; aoqi@0: SETKNOB(LogSpins) ; aoqi@0: SETKNOB(SpinSetSucc) ; aoqi@0: SETKNOB(SuccEnabled) ; aoqi@0: SETKNOB(SuccRestrict) ; aoqi@0: SETKNOB(Penalty) ; aoqi@0: SETKNOB(Bonus) ; aoqi@0: SETKNOB(BonusB) ; aoqi@0: SETKNOB(Poverty) ; aoqi@0: SETKNOB(SpinAfterFutile) ; aoqi@0: SETKNOB(UsePause) ; aoqi@0: SETKNOB(SpinEarly) ; aoqi@0: SETKNOB(OState) ; aoqi@0: SETKNOB(MaxSpinners) ; aoqi@0: SETKNOB(PreSpin) ; aoqi@0: SETKNOB(ExitPolicy) ; aoqi@0: SETKNOB(QMode); aoqi@0: SETKNOB(ResetEvent) ; aoqi@0: SETKNOB(MoveNotifyee) ; aoqi@0: SETKNOB(FastHSSEC) ; aoqi@0: #undef SETKNOB aoqi@0: kevinw@8729: if (Knob_Verbose) { kevinw@8729: sanity_checks(); kevinw@8729: } kevinw@8729: aoqi@0: if (os::is_MP()) { aoqi@0: BackOffMask = (1 << Knob_SpinBackOff) - 1 ; aoqi@0: if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; aoqi@0: // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) aoqi@0: } else { aoqi@0: Knob_SpinLimit = 0 ; aoqi@0: Knob_SpinBase = 0 ; aoqi@0: Knob_PreSpin = 0 ; aoqi@0: Knob_FixedSpin = -1 ; aoqi@0: } aoqi@0: aoqi@0: if (Knob_LogSpins == 0) { aoqi@0: ObjectMonitor::_sync_FailedSpins = NULL ; aoqi@0: } aoqi@0: aoqi@0: free (knobs) ; aoqi@0: OrderAccess::fence() ; aoqi@0: InitDone = 1 ; aoqi@0: } aoqi@0: kevinw@8729: void ObjectMonitor::sanity_checks() { kevinw@8729: int error_cnt = 0; kevinw@8729: int warning_cnt = 0; kevinw@8729: bool verbose = Knob_Verbose != 0 NOT_PRODUCT(|| VerboseInternalVMTests); kevinw@8729: kevinw@8729: if (verbose) { kevinw@8729: tty->print_cr("INFO: sizeof(ObjectMonitor)=" SIZE_FORMAT, kevinw@8729: sizeof(ObjectMonitor)); kevinw@8729: } kevinw@8729: kevinw@8729: uint cache_line_size = VM_Version::L1_data_cache_line_size(); kevinw@8729: if (verbose) { kevinw@8729: tty->print_cr("INFO: L1_data_cache_line_size=%u", cache_line_size); kevinw@8729: } kevinw@8729: kevinw@8729: ObjectMonitor dummy; kevinw@8729: u_char *addr_begin = (u_char*)&dummy; kevinw@8729: u_char *addr_header = (u_char*)&dummy._header; kevinw@8729: u_char *addr_owner = (u_char*)&dummy._owner; kevinw@8729: kevinw@8729: uint offset_header = (uint)(addr_header - addr_begin); kevinw@8729: if (verbose) tty->print_cr("INFO: offset(_header)=%u", offset_header); kevinw@8729: kevinw@8729: uint offset_owner = (uint)(addr_owner - addr_begin); kevinw@8729: if (verbose) tty->print_cr("INFO: offset(_owner)=%u", offset_owner); kevinw@8729: kevinw@8729: if ((uint)(addr_header - addr_begin) != 0) { kevinw@8729: tty->print_cr("ERROR: offset(_header) must be zero (0)."); kevinw@8729: error_cnt++; kevinw@8729: } kevinw@8729: kevinw@8729: if (cache_line_size != 0) { kevinw@8729: // We were able to determine the L1 data cache line size so kevinw@8729: // do some cache line specific sanity checks kevinw@8729: kevinw@8729: if ((offset_owner - offset_header) < cache_line_size) { kevinw@8729: tty->print_cr("WARNING: the _header and _owner fields are closer " kevinw@8729: "than a cache line which permits false sharing."); kevinw@8729: warning_cnt++; kevinw@8729: } kevinw@8729: kevinw@8729: if ((sizeof(ObjectMonitor) % cache_line_size) != 0) { kevinw@8729: tty->print_cr("WARNING: ObjectMonitor size is not a multiple of " kevinw@8729: "a cache line which permits false sharing."); kevinw@8729: warning_cnt++; kevinw@8729: } kevinw@8729: } kevinw@8729: kevinw@8729: ObjectSynchronizer::sanity_checks(verbose, cache_line_size, &error_cnt, kevinw@8729: &warning_cnt); kevinw@8729: kevinw@8729: if (verbose || error_cnt != 0 || warning_cnt != 0) { kevinw@8729: tty->print_cr("INFO: error_cnt=%d", error_cnt); kevinw@8729: tty->print_cr("INFO: warning_cnt=%d", warning_cnt); kevinw@8729: } kevinw@8729: kevinw@8729: guarantee(error_cnt == 0, kevinw@8729: "Fatal error(s) found in ObjectMonitor::sanity_checks()"); kevinw@8729: } kevinw@8729: aoqi@0: #ifndef PRODUCT aoqi@0: void ObjectMonitor::verify() { aoqi@0: } aoqi@0: aoqi@0: void ObjectMonitor::print() { aoqi@0: } aoqi@0: #endif