acorn@2233: /* dbuck@8067: * Copyright (c) 1998, 2015, Oracle and/or its affiliates. All rights reserved. acorn@2233: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. acorn@2233: * acorn@2233: * This code is free software; you can redistribute it and/or modify it acorn@2233: * under the terms of the GNU General Public License version 2 only, as acorn@2233: * published by the Free Software Foundation. acorn@2233: * acorn@2233: * This code is distributed in the hope that it will be useful, but WITHOUT acorn@2233: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or acorn@2233: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License acorn@2233: * version 2 for more details (a copy is included in the LICENSE file that acorn@2233: * accompanied this code). acorn@2233: * acorn@2233: * You should have received a copy of the GNU General Public License version acorn@2233: * 2 along with this work; if not, write to the Free Software Foundation, acorn@2233: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. acorn@2233: * acorn@2233: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA acorn@2233: * or visit www.oracle.com if you need additional information or have any acorn@2233: * questions. acorn@2233: * acorn@2233: */ acorn@2233: stefank@2314: #include "precompiled.hpp" stefank@2314: #include "classfile/vmSymbols.hpp" stefank@2314: #include "memory/resourceArea.hpp" stefank@2314: #include "oops/markOop.hpp" stefank@2314: #include "oops/oop.inline.hpp" stefank@2314: #include "runtime/handles.inline.hpp" stefank@2314: #include "runtime/interfaceSupport.hpp" stefank@2314: #include "runtime/mutexLocker.hpp" stefank@2314: #include "runtime/objectMonitor.hpp" stefank@2314: #include "runtime/objectMonitor.inline.hpp" goetz@6911: #include "runtime/orderAccess.inline.hpp" stefank@2314: #include "runtime/osThread.hpp" stefank@2314: #include "runtime/stubRoutines.hpp" stefank@4299: #include "runtime/thread.inline.hpp" stefank@2314: #include "services/threadService.hpp" sla@5237: #include "trace/tracing.hpp" sla@5237: #include "trace/traceMacros.hpp" stefank@2314: #include "utilities/dtrace.hpp" sla@5237: #include "utilities/macros.hpp" stefank@2314: #include "utilities/preserveException.hpp" stefank@2314: #ifdef TARGET_OS_FAMILY_linux stefank@2314: # include "os_linux.inline.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_OS_FAMILY_solaris stefank@2314: # include "os_solaris.inline.hpp" stefank@2314: #endif stefank@2314: #ifdef TARGET_OS_FAMILY_windows stefank@2314: # include "os_windows.inline.hpp" stefank@2314: #endif never@3156: #ifdef TARGET_OS_FAMILY_bsd never@3156: # include "os_bsd.inline.hpp" never@3156: #endif acorn@2233: goetz@6453: #if defined(__GNUC__) && !defined(IA64) && !defined(PPC64) acorn@2233: // Need to inhibit inlining for older versions of GCC to avoid build-time failures acorn@2233: #define ATTR __attribute__((noinline)) acorn@2233: #else acorn@2233: #define ATTR acorn@2233: #endif acorn@2233: acorn@2233: acorn@2233: #ifdef DTRACE_ENABLED acorn@2233: acorn@2233: // Only bother with this argument setup if dtrace is available acorn@2233: // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. acorn@2233: dcubed@3202: coleenp@4037: #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ dcubed@3202: char* bytes = NULL; \ dcubed@3202: int len = 0; \ dcubed@3202: jlong jtid = SharedRuntime::get_java_tid(thread); \ coleenp@4037: Symbol* klassname = ((oop)obj)->klass()->name(); \ dcubed@3202: if (klassname != NULL) { \ dcubed@3202: bytes = (char*)klassname->bytes(); \ dcubed@3202: len = klassname->utf8_length(); \ dcubed@3202: } dcubed@3202: dcubed@3202: #ifndef USDT2 dcubed@3202: acorn@2233: HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, acorn@2233: jlong, uintptr_t, char*, int); acorn@2233: HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, acorn@2233: jlong, uintptr_t, char*, int); acorn@2233: HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, acorn@2233: jlong, uintptr_t, char*, int); acorn@2233: HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, acorn@2233: jlong, uintptr_t, char*, int); acorn@2233: HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, acorn@2233: jlong, uintptr_t, char*, int); acorn@2233: coleenp@4037: #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ acorn@2233: { \ acorn@2233: if (DTraceMonitorProbes) { \ coleenp@4037: DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ acorn@2233: HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ acorn@2233: (monitor), bytes, len, (millis)); \ acorn@2233: } \ acorn@2233: } acorn@2233: coleenp@4037: #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ acorn@2233: { \ acorn@2233: if (DTraceMonitorProbes) { \ coleenp@4037: DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ acorn@2233: HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ acorn@2233: (uintptr_t)(monitor), bytes, len); \ acorn@2233: } \ acorn@2233: } acorn@2233: dcubed@3202: #else /* USDT2 */ dcubed@3202: coleenp@4037: #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ dcubed@3202: { \ dcubed@3202: if (DTraceMonitorProbes) { \ coleenp@4037: DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ dcubed@3202: HOTSPOT_MONITOR_WAIT(jtid, \ dcubed@3202: (monitor), bytes, len, (millis)); \ dcubed@3202: } \ dcubed@3202: } dcubed@3202: dcubed@3202: #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER dcubed@3202: #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED dcubed@3202: #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT dcubed@3202: #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY dcubed@3202: #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL dcubed@3202: coleenp@4037: #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ dcubed@3202: { \ dcubed@3202: if (DTraceMonitorProbes) { \ coleenp@4037: DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ dcubed@3202: HOTSPOT_MONITOR_##probe(jtid, \ dcubed@3202: (uintptr_t)(monitor), bytes, len); \ dcubed@3202: } \ dcubed@3202: } dcubed@3202: dcubed@3202: #endif /* USDT2 */ acorn@2233: #else // ndef DTRACE_ENABLED acorn@2233: coleenp@4037: #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} coleenp@4037: #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} acorn@2233: acorn@2233: #endif // ndef DTRACE_ENABLED acorn@2233: acorn@2233: // Tunables ... acorn@2233: // The knob* variables are effectively final. Once set they should acorn@2233: // never be modified hence. Consider using __read_mostly with GCC. acorn@2233: acorn@2233: int ObjectMonitor::Knob_Verbose = 0 ; acorn@2233: int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - acorn@2233: static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins acorn@2233: static int Knob_HandOff = 0 ; acorn@2233: static int Knob_ReportSettings = 0 ; acorn@2233: acorn@2233: static int Knob_SpinBase = 0 ; // Floor AKA SpinMin acorn@2233: static int Knob_SpinBackOff = 0 ; // spin-loop backoff acorn@2233: static int Knob_CASPenalty = -1 ; // Penalty for failed CAS acorn@2233: static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change acorn@2233: static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field acorn@2233: static int Knob_SpinEarly = 1 ; acorn@2233: static int Knob_SuccEnabled = 1 ; // futile wake throttling acorn@2233: static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one acorn@2233: static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs acorn@2233: static int Knob_Bonus = 100 ; // spin success bonus acorn@2233: static int Knob_BonusB = 100 ; // spin success bonus acorn@2233: static int Knob_Penalty = 200 ; // spin failure penalty acorn@2233: static int Knob_Poverty = 1000 ; acorn@2233: static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() acorn@2233: static int Knob_FixedSpin = 0 ; acorn@2233: static int Knob_OState = 3 ; // Spinner checks thread state of _owner acorn@2233: static int Knob_UsePause = 1 ; acorn@2233: static int Knob_ExitPolicy = 0 ; acorn@2233: static int Knob_PreSpin = 10 ; // 20-100 likely better acorn@2233: static int Knob_ResetEvent = 0 ; acorn@2233: static int BackOffMask = 0 ; acorn@2233: acorn@2233: static int Knob_FastHSSEC = 0 ; acorn@2233: static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee acorn@2233: static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline acorn@2233: static volatile int InitDone = 0 ; acorn@2233: acorn@2233: #define TrySpin TrySpin_VaryDuration acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Theory of operations -- Monitors lists, thread residency, etc: acorn@2233: // acorn@2233: // * A thread acquires ownership of a monitor by successfully acorn@2233: // CAS()ing the _owner field from null to non-null. acorn@2233: // acorn@2233: // * Invariant: A thread appears on at most one monitor list -- acorn@2233: // cxq, EntryList or WaitSet -- at any one time. acorn@2233: // acorn@2233: // * Contending threads "push" themselves onto the cxq with CAS acorn@2233: // and then spin/park. acorn@2233: // acorn@2233: // * After a contending thread eventually acquires the lock it must acorn@2233: // dequeue itself from either the EntryList or the cxq. acorn@2233: // acorn@2233: // * The exiting thread identifies and unparks an "heir presumptive" acorn@2233: // tentative successor thread on the EntryList. Critically, the acorn@2233: // exiting thread doesn't unlink the successor thread from the EntryList. acorn@2233: // After having been unparked, the wakee will recontend for ownership of acorn@2233: // the monitor. The successor (wakee) will either acquire the lock or acorn@2233: // re-park itself. acorn@2233: // acorn@2233: // Succession is provided for by a policy of competitive handoff. acorn@2233: // The exiting thread does _not_ grant or pass ownership to the acorn@2233: // successor thread. (This is also referred to as "handoff" succession"). acorn@2233: // Instead the exiting thread releases ownership and possibly wakes acorn@2233: // a successor, so the successor can (re)compete for ownership of the lock. acorn@2233: // If the EntryList is empty but the cxq is populated the exiting acorn@2233: // thread will drain the cxq into the EntryList. It does so by acorn@2233: // by detaching the cxq (installing null with CAS) and folding acorn@2233: // the threads from the cxq into the EntryList. The EntryList is acorn@2233: // doubly linked, while the cxq is singly linked because of the acorn@2233: // CAS-based "push" used to enqueue recently arrived threads (RATs). acorn@2233: // acorn@2233: // * Concurrency invariants: acorn@2233: // acorn@2233: // -- only the monitor owner may access or mutate the EntryList. acorn@2233: // The mutex property of the monitor itself protects the EntryList acorn@2233: // from concurrent interference. acorn@2233: // -- Only the monitor owner may detach the cxq. acorn@2233: // acorn@2233: // * The monitor entry list operations avoid locks, but strictly speaking acorn@2233: // 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. acorn@2233: // acorn@2233: // * The cxq can have multiple concurrent "pushers" but only one concurrent acorn@2233: // detaching thread. This mechanism is immune from the ABA corruption. acorn@2233: // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. acorn@2233: // acorn@2233: // * Taken together, the cxq and the EntryList constitute or form a acorn@2233: // single logical queue of threads stalled trying to acquire the lock. acorn@2233: // We use two distinct lists to improve the odds of a constant-time acorn@2233: // dequeue operation after acquisition (in the ::enter() epilog) and acorn@2233: // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). acorn@2233: // A key desideratum is to minimize queue & monitor metadata manipulation acorn@2233: // that occurs while holding the monitor lock -- that is, we want to acorn@2233: // minimize monitor lock holds times. Note that even a small amount of acorn@2233: // fixed spinning will greatly reduce the # of enqueue-dequeue operations acorn@2233: // on EntryList|cxq. That is, spinning relieves contention on the "inner" acorn@2233: // locks and monitor metadata. acorn@2233: // acorn@2233: // Cxq points to the the set of Recently Arrived Threads attempting entry. acorn@2233: // Because we push threads onto _cxq with CAS, the RATs must take the form of acorn@2233: // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when acorn@2233: // the unlocking thread notices that EntryList is null but _cxq is != null. acorn@2233: // acorn@2233: // The EntryList is ordered by the prevailing queue discipline and acorn@2233: // can be organized in any convenient fashion, such as a doubly-linked list or acorn@2233: // a circular doubly-linked list. Critically, we want insert and delete operations acorn@2233: // to operate in constant-time. If we need a priority queue then something akin acorn@2233: // to Solaris' sleepq would work nicely. Viz., acorn@2233: // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. acorn@2233: // Queue discipline is enforced at ::exit() time, when the unlocking thread acorn@2233: // drains the cxq into the EntryList, and orders or reorders the threads on the acorn@2233: // EntryList accordingly. acorn@2233: // acorn@2233: // Barring "lock barging", this mechanism provides fair cyclic ordering, acorn@2233: // somewhat similar to an elevator-scan. acorn@2233: // acorn@2233: // * The monitor synchronization subsystem avoids the use of native acorn@2233: // synchronization primitives except for the narrow platform-specific acorn@2233: // park-unpark abstraction. See the comments in os_solaris.cpp regarding acorn@2233: // the semantics of park-unpark. Put another way, this monitor implementation acorn@2233: // depends only on atomic operations and park-unpark. The monitor subsystem acorn@2233: // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the acorn@2233: // underlying OS manages the READY<->RUN transitions. acorn@2233: // acorn@2233: // * Waiting threads reside on the WaitSet list -- wait() puts acorn@2233: // the caller onto the WaitSet. acorn@2233: // acorn@2233: // * notify() or notifyAll() simply transfers threads from the WaitSet to acorn@2233: // either the EntryList or cxq. Subsequent exit() operations will acorn@2233: // unpark the notifyee. Unparking a notifee in notify() is inefficient - acorn@2233: // it's likely the notifyee would simply impale itself on the lock held acorn@2233: // by the notifier. acorn@2233: // acorn@2233: // * An interesting alternative is to encode cxq as (List,LockByte) where acorn@2233: // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary acorn@2233: // variable, like _recursions, in the scheme. The threads or Events that form acorn@2233: // the list would have to be aligned in 256-byte addresses. A thread would acorn@2233: // try to acquire the lock or enqueue itself with CAS, but exiting threads acorn@2233: // could use a 1-0 protocol and simply STB to set the LockByte to 0. acorn@2233: // Note that is is *not* word-tearing, but it does presume that full-word acorn@2233: // CAS operations are coherent with intermix with STB operations. That's true acorn@2233: // on most common processors. acorn@2233: // acorn@2233: // * See also http://blogs.sun.com/dave acorn@2233: acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Enter support acorn@2233: acorn@2233: bool ObjectMonitor::try_enter(Thread* THREAD) { acorn@2233: if (THREAD != _owner) { acorn@2233: if (THREAD->is_lock_owned ((address)_owner)) { acorn@2233: assert(_recursions == 0, "internal state error"); acorn@2233: _owner = THREAD ; acorn@2233: _recursions = 1 ; acorn@2233: OwnerIsThread = 1 ; acorn@2233: return true; acorn@2233: } acorn@2233: if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { acorn@2233: return false; acorn@2233: } acorn@2233: return true; acorn@2233: } else { acorn@2233: _recursions++; acorn@2233: return true; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: void ATTR ObjectMonitor::enter(TRAPS) { acorn@2233: // The following code is ordered to check the most common cases first acorn@2233: // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. acorn@2233: Thread * const Self = THREAD ; acorn@2233: void * cur ; acorn@2233: acorn@2233: cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; acorn@2233: if (cur == NULL) { acorn@2233: // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. acorn@2233: assert (_recursions == 0 , "invariant") ; acorn@2233: assert (_owner == Self, "invariant") ; acorn@2233: // CONSIDER: set or assert OwnerIsThread == 1 acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: if (cur == Self) { acorn@2233: // TODO-FIXME: check for integer overflow! BUGID 6557169. acorn@2233: _recursions ++ ; acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: if (Self->is_lock_owned ((address)cur)) { acorn@2233: assert (_recursions == 0, "internal state error"); acorn@2233: _recursions = 1 ; acorn@2233: // Commute owner from a thread-specific on-stack BasicLockObject address to acorn@2233: // a full-fledged "Thread *". acorn@2233: _owner = Self ; acorn@2233: OwnerIsThread = 1 ; acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: // We've encountered genuine contention. acorn@2233: assert (Self->_Stalled == 0, "invariant") ; acorn@2233: Self->_Stalled = intptr_t(this) ; acorn@2233: acorn@2233: // Try one round of spinning *before* enqueueing Self acorn@2233: // and before going through the awkward and expensive state acorn@2233: // transitions. The following spin is strictly optional ... acorn@2233: // Note that if we acquire the monitor from an initial spin acorn@2233: // we forgo posting JVMTI events and firing DTRACE probes. acorn@2233: if (Knob_SpinEarly && TrySpin (Self) > 0) { acorn@2233: assert (_owner == Self , "invariant") ; acorn@2233: assert (_recursions == 0 , "invariant") ; acorn@2233: assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; acorn@2233: Self->_Stalled = 0 ; acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: assert (_owner != Self , "invariant") ; acorn@2233: assert (_succ != Self , "invariant") ; acorn@2233: assert (Self->is_Java_thread() , "invariant") ; acorn@2233: JavaThread * jt = (JavaThread *) Self ; acorn@2233: assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; acorn@2233: assert (jt->thread_state() != _thread_blocked , "invariant") ; acorn@2233: assert (this->object() != NULL , "invariant") ; acorn@2233: assert (_count >= 0, "invariant") ; acorn@2233: acorn@2233: // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). acorn@2233: // Ensure the object-monitor relationship remains stable while there's contention. acorn@2233: Atomic::inc_ptr(&_count); acorn@2233: sla@5237: EventJavaMonitorEnter event; sla@5237: acorn@2233: { // Change java thread status to indicate blocked on monitor enter. acorn@2233: JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); acorn@2233: dbuck@8887: Self->set_current_pending_monitor(this); dbuck@8887: acorn@2233: DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); acorn@2233: if (JvmtiExport::should_post_monitor_contended_enter()) { acorn@2233: JvmtiExport::post_monitor_contended_enter(jt, this); dcubed@6335: dcubed@6335: // The current thread does not yet own the monitor and does not dcubed@6335: // yet appear on any queues that would get it made the successor. dcubed@6335: // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event dcubed@6335: // handler cannot accidentally consume an unpark() meant for the dcubed@6335: // ParkEvent associated with this ObjectMonitor. acorn@2233: } acorn@2233: acorn@2233: OSThreadContendState osts(Self->osthread()); acorn@2233: ThreadBlockInVM tbivm(jt); acorn@2233: acorn@2233: // TODO-FIXME: change the following for(;;) loop to straight-line code. acorn@2233: for (;;) { acorn@2233: jt->set_suspend_equivalent(); acorn@2233: // cleared by handle_special_suspend_equivalent_condition() acorn@2233: // or java_suspend_self() acorn@2233: acorn@2233: EnterI (THREAD) ; acorn@2233: acorn@2233: if (!ExitSuspendEquivalent(jt)) break ; acorn@2233: acorn@2233: // acorn@2233: // We have acquired the contended monitor, but while we were acorn@2233: // waiting another thread suspended us. We don't want to enter acorn@2233: // the monitor while suspended because that would surprise the acorn@2233: // thread that suspended us. acorn@2233: // acorn@2233: _recursions = 0 ; acorn@2233: _succ = NULL ; sla@5237: exit (false, Self) ; acorn@2233: acorn@2233: jt->java_suspend_self(); acorn@2233: } acorn@2233: Self->set_current_pending_monitor(NULL); dcubed@6708: dcubed@6708: // We cleared the pending monitor info since we've just gotten past dcubed@6708: // the enter-check-for-suspend dance and we now own the monitor free dcubed@6708: // and clear, i.e., it is no longer pending. The ThreadBlockInVM dcubed@6708: // destructor can go to a safepoint at the end of this block. If we dcubed@6708: // do a thread dump during that safepoint, then this thread will show dcubed@6708: // as having "-locked" the monitor, but the OS and java.lang.Thread dcubed@6708: // states will still report that the thread is blocked trying to dcubed@6708: // acquire it. acorn@2233: } acorn@2233: acorn@2233: Atomic::dec_ptr(&_count); acorn@2233: assert (_count >= 0, "invariant") ; acorn@2233: Self->_Stalled = 0 ; acorn@2233: acorn@2233: // Must either set _recursions = 0 or ASSERT _recursions == 0. acorn@2233: assert (_recursions == 0 , "invariant") ; acorn@2233: assert (_owner == Self , "invariant") ; acorn@2233: assert (_succ != Self , "invariant") ; acorn@2233: assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; acorn@2233: acorn@2233: // The thread -- now the owner -- is back in vm mode. acorn@2233: // Report the glorious news via TI,DTrace and jvmstat. acorn@2233: // The probe effect is non-trivial. All the reportage occurs acorn@2233: // while we hold the monitor, increasing the length of the critical acorn@2233: // section. Amdahl's parallel speedup law comes vividly into play. acorn@2233: // acorn@2233: // Another option might be to aggregate the events (thread local or acorn@2233: // per-monitor aggregation) and defer reporting until a more opportune acorn@2233: // time -- such as next time some thread encounters contention but has acorn@2233: // yet to acquire the lock. While spinning that thread could acorn@2233: // spinning we could increment JVMStat counters, etc. acorn@2233: acorn@2233: DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); acorn@2233: if (JvmtiExport::should_post_monitor_contended_entered()) { acorn@2233: JvmtiExport::post_monitor_contended_entered(jt, this); dcubed@6335: dcubed@6335: // The current thread already owns the monitor and is not going to dcubed@6335: // call park() for the remainder of the monitor enter protocol. So dcubed@6335: // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED dcubed@6335: // event handler consumed an unpark() issued by the thread that dcubed@6335: // just exited the monitor. acorn@2233: } sla@5237: sla@5237: if (event.should_commit()) { sla@5237: event.set_klass(((oop)this->object())->klass()); sla@5237: event.set_previousOwner((TYPE_JAVALANGTHREAD)_previous_owner_tid); sla@5237: event.set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); sla@5237: event.commit(); sla@5237: } sla@5237: acorn@2233: if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { acorn@2233: ObjectMonitor::_sync_ContendedLockAttempts->inc() ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: acorn@2233: // Caveat: TryLock() is not necessarily serializing if it returns failure. acorn@2233: // Callers must compensate as needed. acorn@2233: acorn@2233: int ObjectMonitor::TryLock (Thread * Self) { acorn@2233: for (;;) { acorn@2233: void * own = _owner ; acorn@2233: if (own != NULL) return 0 ; acorn@2233: if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { acorn@2233: // Either guarantee _recursions == 0 or set _recursions = 0. acorn@2233: assert (_recursions == 0, "invariant") ; acorn@2233: assert (_owner == Self, "invariant") ; acorn@2233: // CONSIDER: set or assert that OwnerIsThread == 1 acorn@2233: return 1 ; acorn@2233: } acorn@2233: // The lock had been free momentarily, but we lost the race to the lock. acorn@2233: // Interference -- the CAS failed. acorn@2233: // We can either return -1 or retry. acorn@2233: // Retry doesn't make as much sense because the lock was just acquired. acorn@2233: if (true) return -1 ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: void ATTR ObjectMonitor::EnterI (TRAPS) { acorn@2233: Thread * Self = THREAD ; acorn@2233: assert (Self->is_Java_thread(), "invariant") ; acorn@2233: assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; acorn@2233: acorn@2233: // Try the lock - TATAS acorn@2233: if (TryLock (Self) > 0) { acorn@2233: assert (_succ != Self , "invariant") ; acorn@2233: assert (_owner == Self , "invariant") ; acorn@2233: assert (_Responsible != Self , "invariant") ; acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: DeferredInitialize () ; acorn@2233: acorn@2233: // We try one round of spinning *before* enqueueing Self. acorn@2233: // acorn@2233: // If the _owner is ready but OFFPROC we could use a YieldTo() acorn@2233: // operation to donate the remainder of this thread's quantum acorn@2233: // to the owner. This has subtle but beneficial affinity acorn@2233: // effects. acorn@2233: acorn@2233: if (TrySpin (Self) > 0) { acorn@2233: assert (_owner == Self , "invariant") ; acorn@2233: assert (_succ != Self , "invariant") ; acorn@2233: assert (_Responsible != Self , "invariant") ; acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: // The Spin failed -- Enqueue and park the thread ... acorn@2233: assert (_succ != Self , "invariant") ; acorn@2233: assert (_owner != Self , "invariant") ; acorn@2233: assert (_Responsible != Self , "invariant") ; acorn@2233: acorn@2233: // Enqueue "Self" on ObjectMonitor's _cxq. acorn@2233: // acorn@2233: // Node acts as a proxy for Self. acorn@2233: // As an aside, if were to ever rewrite the synchronization code mostly acorn@2233: // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class acorn@2233: // Java objects. This would avoid awkward lifecycle and liveness issues, acorn@2233: // as well as eliminate a subset of ABA issues. acorn@2233: // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. acorn@2233: // acorn@2233: acorn@2233: ObjectWaiter node(Self) ; acorn@2233: Self->_ParkEvent->reset() ; acorn@2233: node._prev = (ObjectWaiter *) 0xBAD ; acorn@2233: node.TState = ObjectWaiter::TS_CXQ ; acorn@2233: acorn@2233: // Push "Self" onto the front of the _cxq. acorn@2233: // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. acorn@2233: // Note that spinning tends to reduce the rate at which threads acorn@2233: // enqueue and dequeue on EntryList|cxq. acorn@2233: ObjectWaiter * nxt ; acorn@2233: for (;;) { acorn@2233: node._next = nxt = _cxq ; acorn@2233: if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; acorn@2233: acorn@2233: // Interference - the CAS failed because _cxq changed. Just retry. acorn@2233: // As an optional optimization we retry the lock. acorn@2233: if (TryLock (Self) > 0) { acorn@2233: assert (_succ != Self , "invariant") ; acorn@2233: assert (_owner == Self , "invariant") ; acorn@2233: assert (_Responsible != Self , "invariant") ; acorn@2233: return ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: // Check for cxq|EntryList edge transition to non-null. This indicates acorn@2233: // the onset of contention. While contention persists exiting threads acorn@2233: // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit acorn@2233: // operations revert to the faster 1-0 mode. This enter operation may interleave acorn@2233: // (race) a concurrent 1-0 exit operation, resulting in stranding, so we acorn@2233: // arrange for one of the contending thread to use a timed park() operations acorn@2233: // to detect and recover from the race. (Stranding is form of progress failure acorn@2233: // where the monitor is unlocked but all the contending threads remain parked). acorn@2233: // That is, at least one of the contended threads will periodically poll _owner. acorn@2233: // One of the contending threads will become the designated "Responsible" thread. acorn@2233: // The Responsible thread uses a timed park instead of a normal indefinite park acorn@2233: // operation -- it periodically wakes and checks for and recovers from potential acorn@2233: // strandings admitted by 1-0 exit operations. We need at most one Responsible acorn@2233: // thread per-monitor at any given moment. Only threads on cxq|EntryList may acorn@2233: // be responsible for a monitor. acorn@2233: // acorn@2233: // Currently, one of the contended threads takes on the added role of "Responsible". acorn@2233: // A viable alternative would be to use a dedicated "stranding checker" thread acorn@2233: // that periodically iterated over all the threads (or active monitors) and unparked acorn@2233: // successors where there was risk of stranding. This would help eliminate the acorn@2233: // timer scalability issues we see on some platforms as we'd only have one thread acorn@2233: // -- the checker -- parked on a timer. acorn@2233: acorn@2233: if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { acorn@2233: // Try to assume the role of responsible thread for the monitor. acorn@2233: // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } acorn@2233: Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; acorn@2233: } acorn@2233: acorn@2233: // The lock have been released while this thread was occupied queueing acorn@2233: // itself onto _cxq. To close the race and avoid "stranding" and acorn@2233: // progress-liveness failure we must resample-retry _owner before parking. acorn@2233: // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. acorn@2233: // In this case the ST-MEMBAR is accomplished with CAS(). acorn@2233: // acorn@2233: // TODO: Defer all thread state transitions until park-time. acorn@2233: // Since state transitions are heavy and inefficient we'd like acorn@2233: // to defer the state transitions until absolutely necessary, acorn@2233: // and in doing so avoid some transitions ... acorn@2233: acorn@2233: TEVENT (Inflated enter - Contention) ; acorn@2233: int nWakeups = 0 ; acorn@2233: int RecheckInterval = 1 ; acorn@2233: acorn@2233: for (;;) { acorn@2233: acorn@2233: if (TryLock (Self) > 0) break ; acorn@2233: assert (_owner != Self, "invariant") ; acorn@2233: acorn@2233: if ((SyncFlags & 2) && _Responsible == NULL) { acorn@2233: Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; acorn@2233: } acorn@2233: acorn@2233: // park self acorn@2233: if (_Responsible == Self || (SyncFlags & 1)) { acorn@2233: TEVENT (Inflated enter - park TIMED) ; acorn@2233: Self->_ParkEvent->park ((jlong) RecheckInterval) ; acorn@2233: // Increase the RecheckInterval, but clamp the value. acorn@2233: RecheckInterval *= 8 ; acorn@2233: if (RecheckInterval > 1000) RecheckInterval = 1000 ; acorn@2233: } else { acorn@2233: TEVENT (Inflated enter - park UNTIMED) ; acorn@2233: Self->_ParkEvent->park() ; acorn@2233: } acorn@2233: acorn@2233: if (TryLock(Self) > 0) break ; acorn@2233: acorn@2233: // The lock is still contested. acorn@2233: // Keep a tally of the # of futile wakeups. acorn@2233: // Note that the counter is not protected by a lock or updated by atomics. acorn@2233: // That is by design - we trade "lossy" counters which are exposed to acorn@2233: // races during updates for a lower probe effect. acorn@2233: TEVENT (Inflated enter - Futile wakeup) ; acorn@2233: if (ObjectMonitor::_sync_FutileWakeups != NULL) { acorn@2233: ObjectMonitor::_sync_FutileWakeups->inc() ; acorn@2233: } acorn@2233: ++ nWakeups ; acorn@2233: acorn@2233: // Assuming this is not a spurious wakeup we'll normally find _succ == Self. acorn@2233: // We can defer clearing _succ until after the spin completes acorn@2233: // TrySpin() must tolerate being called with _succ == Self. acorn@2233: // Try yet another round of adaptive spinning. acorn@2233: if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; acorn@2233: acorn@2233: // We can find that we were unpark()ed and redesignated _succ while acorn@2233: // we were spinning. That's harmless. If we iterate and call park(), acorn@2233: // park() will consume the event and return immediately and we'll acorn@2233: // just spin again. This pattern can repeat, leaving _succ to simply acorn@2233: // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). acorn@2233: // Alternately, we can sample fired() here, and if set, forgo spinning acorn@2233: // in the next iteration. acorn@2233: acorn@2233: if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { acorn@2233: Self->_ParkEvent->reset() ; acorn@2233: OrderAccess::fence() ; acorn@2233: } acorn@2233: if (_succ == Self) _succ = NULL ; acorn@2233: acorn@2233: // Invariant: after clearing _succ a thread *must* retry _owner before parking. acorn@2233: OrderAccess::fence() ; acorn@2233: } acorn@2233: acorn@2233: // Egress : acorn@2233: // Self has acquired the lock -- Unlink Self from the cxq or EntryList. acorn@2233: // Normally we'll find Self on the EntryList . acorn@2233: // From the perspective of the lock owner (this thread), the acorn@2233: // EntryList is stable and cxq is prepend-only. acorn@2233: // The head of cxq is volatile but the interior is stable. acorn@2233: // In addition, Self.TState is stable. acorn@2233: acorn@2233: assert (_owner == Self , "invariant") ; acorn@2233: assert (object() != NULL , "invariant") ; acorn@2233: // I'd like to write: acorn@2233: // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; acorn@2233: // but as we're at a safepoint that's not safe. acorn@2233: acorn@2233: UnlinkAfterAcquire (Self, &node) ; acorn@2233: if (_succ == Self) _succ = NULL ; acorn@2233: acorn@2233: assert (_succ != Self, "invariant") ; acorn@2233: if (_Responsible == Self) { acorn@2233: _Responsible = NULL ; dcubed@4471: OrderAccess::fence(); // Dekker pivot-point acorn@2233: acorn@2233: // We may leave threads on cxq|EntryList without a designated acorn@2233: // "Responsible" thread. This is benign. When this thread subsequently acorn@2233: // exits the monitor it can "see" such preexisting "old" threads -- acorn@2233: // threads that arrived on the cxq|EntryList before the fence, above -- acorn@2233: // by LDing cxq|EntryList. Newly arrived threads -- that is, threads acorn@2233: // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible acorn@2233: // non-null and elect a new "Responsible" timer thread. acorn@2233: // acorn@2233: // This thread executes: acorn@2233: // ST Responsible=null; MEMBAR (in enter epilog - here) acorn@2233: // LD cxq|EntryList (in subsequent exit) acorn@2233: // acorn@2233: // Entering threads in the slow/contended path execute: acorn@2233: // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) acorn@2233: // The (ST cxq; MEMBAR) is accomplished with CAS(). acorn@2233: // acorn@2233: // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent acorn@2233: // exit operation from floating above the ST Responsible=null. acorn@2233: } acorn@2233: acorn@2233: // We've acquired ownership with CAS(). acorn@2233: // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. acorn@2233: // But since the CAS() this thread may have also stored into _succ, acorn@2233: // EntryList, cxq or Responsible. These meta-data updates must be acorn@2233: // visible __before this thread subsequently drops the lock. acorn@2233: // Consider what could occur if we didn't enforce this constraint -- acorn@2233: // STs to monitor meta-data and user-data could reorder with (become acorn@2233: // visible after) the ST in exit that drops ownership of the lock. acorn@2233: // Some other thread could then acquire the lock, but observe inconsistent acorn@2233: // or old monitor meta-data and heap data. That violates the JMM. acorn@2233: // To that end, the 1-0 exit() operation must have at least STST|LDST acorn@2233: // "release" barrier semantics. Specifically, there must be at least a acorn@2233: // STST|LDST barrier in exit() before the ST of null into _owner that drops acorn@2233: // the lock. The barrier ensures that changes to monitor meta-data and data acorn@2233: // protected by the lock will be visible before we release the lock, and acorn@2233: // therefore before some other thread (CPU) has a chance to acquire the lock. acorn@2233: // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. acorn@2233: // acorn@2233: // Critically, any prior STs to _succ or EntryList must be visible before acorn@2233: // the ST of null into _owner in the *subsequent* (following) corresponding acorn@2233: // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily acorn@2233: // execute a serializing instruction. acorn@2233: acorn@2233: if (SyncFlags & 8) { acorn@2233: OrderAccess::fence() ; acorn@2233: } acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: // ReenterI() is a specialized inline form of the latter half of the acorn@2233: // contended slow-path from EnterI(). We use ReenterI() only for acorn@2233: // monitor reentry in wait(). acorn@2233: // acorn@2233: // In the future we should reconcile EnterI() and ReenterI(), adding acorn@2233: // Knob_Reset and Knob_SpinAfterFutile support and restructuring the acorn@2233: // loop accordingly. acorn@2233: acorn@2233: void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { acorn@2233: assert (Self != NULL , "invariant") ; acorn@2233: assert (SelfNode != NULL , "invariant") ; acorn@2233: assert (SelfNode->_thread == Self , "invariant") ; acorn@2233: assert (_waiters > 0 , "invariant") ; acorn@2233: assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; acorn@2233: assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; acorn@2233: JavaThread * jt = (JavaThread *) Self ; acorn@2233: acorn@2233: int nWakeups = 0 ; acorn@2233: for (;;) { acorn@2233: ObjectWaiter::TStates v = SelfNode->TState ; acorn@2233: guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; acorn@2233: assert (_owner != Self, "invariant") ; acorn@2233: acorn@2233: if (TryLock (Self) > 0) break ; acorn@2233: if (TrySpin (Self) > 0) break ; acorn@2233: acorn@2233: TEVENT (Wait Reentry - parking) ; acorn@2233: acorn@2233: // State transition wrappers around park() ... acorn@2233: // ReenterI() wisely defers state transitions until acorn@2233: // it's clear we must park the thread. acorn@2233: { acorn@2233: OSThreadContendState osts(Self->osthread()); acorn@2233: ThreadBlockInVM tbivm(jt); acorn@2233: acorn@2233: // cleared by handle_special_suspend_equivalent_condition() acorn@2233: // or java_suspend_self() acorn@2233: jt->set_suspend_equivalent(); acorn@2233: if (SyncFlags & 1) { acorn@2233: Self->_ParkEvent->park ((jlong)1000) ; acorn@2233: } else { acorn@2233: Self->_ParkEvent->park () ; acorn@2233: } acorn@2233: acorn@2233: // were we externally suspended while we were waiting? acorn@2233: for (;;) { acorn@2233: if (!ExitSuspendEquivalent (jt)) break ; acorn@2233: if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } acorn@2233: jt->java_suspend_self(); acorn@2233: jt->set_suspend_equivalent(); acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: // Try again, but just so we distinguish between futile wakeups and acorn@2233: // successful wakeups. The following test isn't algorithmically acorn@2233: // necessary, but it helps us maintain sensible statistics. acorn@2233: if (TryLock(Self) > 0) break ; acorn@2233: acorn@2233: // The lock is still contested. acorn@2233: // Keep a tally of the # of futile wakeups. acorn@2233: // Note that the counter is not protected by a lock or updated by atomics. acorn@2233: // That is by design - we trade "lossy" counters which are exposed to acorn@2233: // races during updates for a lower probe effect. acorn@2233: TEVENT (Wait Reentry - futile wakeup) ; acorn@2233: ++ nWakeups ; acorn@2233: acorn@2233: // Assuming this is not a spurious wakeup we'll normally acorn@2233: // find that _succ == Self. acorn@2233: if (_succ == Self) _succ = NULL ; acorn@2233: acorn@2233: // Invariant: after clearing _succ a contending thread acorn@2233: // *must* retry _owner before parking. acorn@2233: OrderAccess::fence() ; acorn@2233: acorn@2233: if (ObjectMonitor::_sync_FutileWakeups != NULL) { acorn@2233: ObjectMonitor::_sync_FutileWakeups->inc() ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: // Self has acquired the lock -- Unlink Self from the cxq or EntryList . acorn@2233: // Normally we'll find Self on the EntryList. acorn@2233: // Unlinking from the EntryList is constant-time and atomic-free. acorn@2233: // From the perspective of the lock owner (this thread), the acorn@2233: // EntryList is stable and cxq is prepend-only. acorn@2233: // The head of cxq is volatile but the interior is stable. acorn@2233: // In addition, Self.TState is stable. acorn@2233: acorn@2233: assert (_owner == Self, "invariant") ; acorn@2233: assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; acorn@2233: UnlinkAfterAcquire (Self, SelfNode) ; acorn@2233: if (_succ == Self) _succ = NULL ; acorn@2233: assert (_succ != Self, "invariant") ; acorn@2233: SelfNode->TState = ObjectWaiter::TS_RUN ; acorn@2233: OrderAccess::fence() ; // see comments at the end of EnterI() acorn@2233: } acorn@2233: acorn@2233: // after the thread acquires the lock in ::enter(). Equally, we could defer acorn@2233: // unlinking the thread until ::exit()-time. acorn@2233: acorn@2233: void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) acorn@2233: { acorn@2233: assert (_owner == Self, "invariant") ; acorn@2233: assert (SelfNode->_thread == Self, "invariant") ; acorn@2233: acorn@2233: if (SelfNode->TState == ObjectWaiter::TS_ENTER) { acorn@2233: // Normal case: remove Self from the DLL EntryList . acorn@2233: // This is a constant-time operation. acorn@2233: ObjectWaiter * nxt = SelfNode->_next ; acorn@2233: ObjectWaiter * prv = SelfNode->_prev ; acorn@2233: if (nxt != NULL) nxt->_prev = prv ; acorn@2233: if (prv != NULL) prv->_next = nxt ; acorn@2233: if (SelfNode == _EntryList ) _EntryList = nxt ; acorn@2233: assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; acorn@2233: assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; acorn@2233: TEVENT (Unlink from EntryList) ; acorn@2233: } else { acorn@2233: guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; acorn@2233: // Inopportune interleaving -- Self is still on the cxq. acorn@2233: // This usually means the enqueue of self raced an exiting thread. acorn@2233: // Normally we'll find Self near the front of the cxq, so acorn@2233: // dequeueing is typically fast. If needbe we can accelerate acorn@2233: // this with some MCS/CHL-like bidirectional list hints and advisory acorn@2233: // back-links so dequeueing from the interior will normally operate acorn@2233: // in constant-time. acorn@2233: // Dequeue Self from either the head (with CAS) or from the interior acorn@2233: // with a linear-time scan and normal non-atomic memory operations. acorn@2233: // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList acorn@2233: // and then unlink Self from EntryList. We have to drain eventually, acorn@2233: // so it might as well be now. acorn@2233: acorn@2233: ObjectWaiter * v = _cxq ; acorn@2233: assert (v != NULL, "invariant") ; acorn@2233: if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { acorn@2233: // The CAS above can fail from interference IFF a "RAT" arrived. acorn@2233: // In that case Self must be in the interior and can no longer be acorn@2233: // at the head of cxq. acorn@2233: if (v == SelfNode) { acorn@2233: assert (_cxq != v, "invariant") ; acorn@2233: v = _cxq ; // CAS above failed - start scan at head of list acorn@2233: } acorn@2233: ObjectWaiter * p ; acorn@2233: ObjectWaiter * q = NULL ; acorn@2233: for (p = v ; p != NULL && p != SelfNode; p = p->_next) { acorn@2233: q = p ; acorn@2233: assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; acorn@2233: } acorn@2233: assert (v != SelfNode, "invariant") ; acorn@2233: assert (p == SelfNode, "Node not found on cxq") ; acorn@2233: assert (p != _cxq, "invariant") ; acorn@2233: assert (q != NULL, "invariant") ; acorn@2233: assert (q->_next == p, "invariant") ; acorn@2233: q->_next = p->_next ; acorn@2233: } acorn@2233: TEVENT (Unlink from cxq) ; acorn@2233: } acorn@2233: acorn@2233: // Diagnostic hygiene ... acorn@2233: SelfNode->_prev = (ObjectWaiter *) 0xBAD ; acorn@2233: SelfNode->_next = (ObjectWaiter *) 0xBAD ; acorn@2233: SelfNode->TState = ObjectWaiter::TS_RUN ; acorn@2233: } acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Exit support acorn@2233: // acorn@2233: // exit() acorn@2233: // ~~~~~~ acorn@2233: // Note that the collector can't reclaim the objectMonitor or deflate acorn@2233: // the object out from underneath the thread calling ::exit() as the acorn@2233: // thread calling ::exit() never transitions to a stable state. acorn@2233: // This inhibits GC, which in turn inhibits asynchronous (and acorn@2233: // inopportune) reclamation of "this". acorn@2233: // acorn@2233: // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; acorn@2233: // There's one exception to the claim above, however. EnterI() can call acorn@2233: // exit() to drop a lock if the acquirer has been externally suspended. acorn@2233: // In that case exit() is called with _thread_state as _thread_blocked, acorn@2233: // but the monitor's _count field is > 0, which inhibits reclamation. acorn@2233: // acorn@2233: // 1-0 exit acorn@2233: // ~~~~~~~~ acorn@2233: // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of acorn@2233: // the fast-path operators have been optimized so the common ::exit() acorn@2233: // operation is 1-0. See i486.ad fast_unlock(), for instance. acorn@2233: // The code emitted by fast_unlock() elides the usual MEMBAR. This acorn@2233: // greatly improves latency -- MEMBAR and CAS having considerable local acorn@2233: // latency on modern processors -- but at the cost of "stranding". Absent the acorn@2233: // MEMBAR, a thread in fast_unlock() can race a thread in the slow acorn@2233: // ::enter() path, resulting in the entering thread being stranding acorn@2233: // and a progress-liveness failure. Stranding is extremely rare. acorn@2233: // We use timers (timed park operations) & periodic polling to detect acorn@2233: // and recover from stranding. Potentially stranded threads periodically acorn@2233: // wake up and poll the lock. See the usage of the _Responsible variable. acorn@2233: // acorn@2233: // The CAS() in enter provides for safety and exclusion, while the CAS or acorn@2233: // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking acorn@2233: // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. acorn@2233: // We detect and recover from stranding with timers. acorn@2233: // acorn@2233: // If a thread transiently strands it'll park until (a) another acorn@2233: // thread acquires the lock and then drops the lock, at which time the acorn@2233: // exiting thread will notice and unpark the stranded thread, or, (b) acorn@2233: // the timer expires. If the lock is high traffic then the stranding latency acorn@2233: // will be low due to (a). If the lock is low traffic then the odds of acorn@2233: // stranding are lower, although the worst-case stranding latency acorn@2233: // is longer. Critically, we don't want to put excessive load in the acorn@2233: // platform's timer subsystem. We want to minimize both the timer injection acorn@2233: // rate (timers created/sec) as well as the number of timers active at acorn@2233: // any one time. (more precisely, we want to minimize timer-seconds, which is acorn@2233: // the integral of the # of active timers at any instant over time). acorn@2233: // Both impinge on OS scalability. Given that, at most one thread parked on acorn@2233: // a monitor will use a timer. acorn@2233: sla@5237: void ATTR ObjectMonitor::exit(bool not_suspended, TRAPS) { acorn@2233: Thread * Self = THREAD ; acorn@2233: if (THREAD != _owner) { acorn@2233: if (THREAD->is_lock_owned((address) _owner)) { acorn@2233: // Transmute _owner from a BasicLock pointer to a Thread address. acorn@2233: // We don't need to hold _mutex for this transition. acorn@2233: // Non-null to Non-null is safe as long as all readers can acorn@2233: // tolerate either flavor. acorn@2233: assert (_recursions == 0, "invariant") ; acorn@2233: _owner = THREAD ; acorn@2233: _recursions = 0 ; acorn@2233: OwnerIsThread = 1 ; acorn@2233: } else { acorn@2233: // NOTE: we need to handle unbalanced monitor enter/exit acorn@2233: // in native code by throwing an exception. acorn@2233: // TODO: Throw an IllegalMonitorStateException ? acorn@2233: TEVENT (Exit - Throw IMSX) ; acorn@2233: assert(false, "Non-balanced monitor enter/exit!"); acorn@2233: if (false) { acorn@2233: THROW(vmSymbols::java_lang_IllegalMonitorStateException()); acorn@2233: } acorn@2233: return; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: if (_recursions != 0) { acorn@2233: _recursions--; // this is simple recursive enter acorn@2233: TEVENT (Inflated exit - recursive) ; acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: // Invariant: after setting Responsible=null an thread must execute acorn@2233: // a MEMBAR or other serializing instruction before fetching EntryList|cxq. acorn@2233: if ((SyncFlags & 4) == 0) { acorn@2233: _Responsible = NULL ; acorn@2233: } acorn@2233: sla@5237: #if INCLUDE_TRACE sla@5237: // get the owner's thread id for the MonitorEnter event sla@5237: // if it is enabled and the thread isn't suspended sla@5237: if (not_suspended && Tracing::is_event_enabled(TraceJavaMonitorEnterEvent)) { sla@5237: _previous_owner_tid = SharedRuntime::get_java_tid(Self); sla@5237: } sla@5237: #endif sla@5237: acorn@2233: for (;;) { acorn@2233: assert (THREAD == _owner, "invariant") ; acorn@2233: acorn@2233: acorn@2233: if (Knob_ExitPolicy == 0) { acorn@2233: // release semantics: prior loads and stores from within the critical section acorn@2233: // must not float (reorder) past the following store that drops the lock. acorn@2233: // On SPARC that requires MEMBAR #loadstore|#storestore. acorn@2233: // But of course in TSO #loadstore|#storestore is not required. acorn@2233: // I'd like to write one of the following: acorn@2233: // A. OrderAccess::release() ; _owner = NULL acorn@2233: // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; acorn@2233: // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both acorn@2233: // store into a _dummy variable. That store is not needed, but can result acorn@2233: // in massive wasteful coherency traffic on classic SMP systems. acorn@2233: // Instead, I use release_store(), which is implemented as just a simple acorn@2233: // ST on x64, x86 and SPARC. acorn@2233: OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock acorn@2233: OrderAccess::storeload() ; // See if we need to wake a successor acorn@2233: if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { acorn@2233: TEVENT (Inflated exit - simple egress) ; acorn@2233: return ; acorn@2233: } acorn@2233: TEVENT (Inflated exit - complex egress) ; acorn@2233: acorn@2233: // Normally the exiting thread is responsible for ensuring succession, acorn@2233: // but if other successors are ready or other entering threads are spinning acorn@2233: // then this thread can simply store NULL into _owner and exit without acorn@2233: // waking a successor. The existence of spinners or ready successors acorn@2233: // guarantees proper succession (liveness). Responsibility passes to the acorn@2233: // ready or running successors. The exiting thread delegates the duty. acorn@2233: // More precisely, if a successor already exists this thread is absolved acorn@2233: // of the responsibility of waking (unparking) one. acorn@2233: // acorn@2233: // The _succ variable is critical to reducing futile wakeup frequency. acorn@2233: // _succ identifies the "heir presumptive" thread that has been made acorn@2233: // ready (unparked) but that has not yet run. We need only one such acorn@2233: // successor thread to guarantee progress. acorn@2233: // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf acorn@2233: // section 3.3 "Futile Wakeup Throttling" for details. acorn@2233: // acorn@2233: // Note that spinners in Enter() also set _succ non-null. acorn@2233: // In the current implementation spinners opportunistically set acorn@2233: // _succ so that exiting threads might avoid waking a successor. acorn@2233: // Another less appealing alternative would be for the exiting thread acorn@2233: // to drop the lock and then spin briefly to see if a spinner managed acorn@2233: // to acquire the lock. If so, the exiting thread could exit acorn@2233: // immediately without waking a successor, otherwise the exiting acorn@2233: // thread would need to dequeue and wake a successor. acorn@2233: // (Note that we'd need to make the post-drop spin short, but no acorn@2233: // shorter than the worst-case round-trip cache-line migration time. acorn@2233: // The dropped lock needs to become visible to the spinner, and then acorn@2233: // the acquisition of the lock by the spinner must become visible to acorn@2233: // the exiting thread). acorn@2233: // acorn@2233: acorn@2233: // It appears that an heir-presumptive (successor) must be made ready. acorn@2233: // Only the current lock owner can manipulate the EntryList or acorn@2233: // drain _cxq, so we need to reacquire the lock. If we fail acorn@2233: // to reacquire the lock the responsibility for ensuring succession acorn@2233: // falls to the new owner. acorn@2233: // acorn@2233: if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { acorn@2233: return ; acorn@2233: } acorn@2233: TEVENT (Exit - Reacquired) ; acorn@2233: } else { acorn@2233: if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { acorn@2233: OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock acorn@2233: OrderAccess::storeload() ; acorn@2233: // Ratify the previously observed values. acorn@2233: if (_cxq == NULL || _succ != NULL) { acorn@2233: TEVENT (Inflated exit - simple egress) ; acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: // inopportune interleaving -- the exiting thread (this thread) acorn@2233: // in the fast-exit path raced an entering thread in the slow-enter acorn@2233: // path. acorn@2233: // We have two choices: acorn@2233: // A. Try to reacquire the lock. acorn@2233: // If the CAS() fails return immediately, otherwise acorn@2233: // we either restart/rerun the exit operation, or simply acorn@2233: // fall-through into the code below which wakes a successor. acorn@2233: // B. If the elements forming the EntryList|cxq are TSM acorn@2233: // we could simply unpark() the lead thread and return acorn@2233: // without having set _succ. acorn@2233: if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { acorn@2233: TEVENT (Inflated exit - reacquired succeeded) ; acorn@2233: return ; acorn@2233: } acorn@2233: TEVENT (Inflated exit - reacquired failed) ; acorn@2233: } else { acorn@2233: TEVENT (Inflated exit - complex egress) ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: guarantee (_owner == THREAD, "invariant") ; acorn@2233: acorn@2233: ObjectWaiter * w = NULL ; acorn@2233: int QMode = Knob_QMode ; acorn@2233: acorn@2233: if (QMode == 2 && _cxq != NULL) { acorn@2233: // QMode == 2 : cxq has precedence over EntryList. acorn@2233: // Try to directly wake a successor from the cxq. acorn@2233: // If successful, the successor will need to unlink itself from cxq. acorn@2233: w = _cxq ; acorn@2233: assert (w != NULL, "invariant") ; acorn@2233: assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; acorn@2233: ExitEpilog (Self, w) ; acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: if (QMode == 3 && _cxq != NULL) { acorn@2233: // Aggressively drain cxq into EntryList at the first opportunity. acorn@2233: // This policy ensure that recently-run threads live at the head of EntryList. acorn@2233: // Drain _cxq into EntryList - bulk transfer. acorn@2233: // First, detach _cxq. acorn@2233: // The following loop is tantamount to: w = swap (&cxq, NULL) acorn@2233: w = _cxq ; acorn@2233: for (;;) { acorn@2233: assert (w != NULL, "Invariant") ; acorn@2233: ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; acorn@2233: if (u == w) break ; acorn@2233: w = u ; acorn@2233: } acorn@2233: assert (w != NULL , "invariant") ; acorn@2233: acorn@2233: ObjectWaiter * q = NULL ; acorn@2233: ObjectWaiter * p ; acorn@2233: for (p = w ; p != NULL ; p = p->_next) { acorn@2233: guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; acorn@2233: p->TState = ObjectWaiter::TS_ENTER ; acorn@2233: p->_prev = q ; acorn@2233: q = p ; acorn@2233: } acorn@2233: acorn@2233: // Append the RATs to the EntryList acorn@2233: // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. acorn@2233: ObjectWaiter * Tail ; acorn@2233: for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; acorn@2233: if (Tail == NULL) { acorn@2233: _EntryList = w ; acorn@2233: } else { acorn@2233: Tail->_next = w ; acorn@2233: w->_prev = Tail ; acorn@2233: } acorn@2233: acorn@2233: // Fall thru into code that tries to wake a successor from EntryList acorn@2233: } acorn@2233: acorn@2233: if (QMode == 4 && _cxq != NULL) { acorn@2233: // Aggressively drain cxq into EntryList at the first opportunity. acorn@2233: // This policy ensure that recently-run threads live at the head of EntryList. acorn@2233: acorn@2233: // Drain _cxq into EntryList - bulk transfer. acorn@2233: // First, detach _cxq. acorn@2233: // The following loop is tantamount to: w = swap (&cxq, NULL) acorn@2233: w = _cxq ; acorn@2233: for (;;) { acorn@2233: assert (w != NULL, "Invariant") ; acorn@2233: ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; acorn@2233: if (u == w) break ; acorn@2233: w = u ; acorn@2233: } acorn@2233: assert (w != NULL , "invariant") ; acorn@2233: acorn@2233: ObjectWaiter * q = NULL ; acorn@2233: ObjectWaiter * p ; acorn@2233: for (p = w ; p != NULL ; p = p->_next) { acorn@2233: guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; acorn@2233: p->TState = ObjectWaiter::TS_ENTER ; acorn@2233: p->_prev = q ; acorn@2233: q = p ; acorn@2233: } acorn@2233: acorn@2233: // Prepend the RATs to the EntryList acorn@2233: if (_EntryList != NULL) { acorn@2233: q->_next = _EntryList ; acorn@2233: _EntryList->_prev = q ; acorn@2233: } acorn@2233: _EntryList = w ; acorn@2233: acorn@2233: // Fall thru into code that tries to wake a successor from EntryList acorn@2233: } acorn@2233: acorn@2233: w = _EntryList ; acorn@2233: if (w != NULL) { acorn@2233: // I'd like to write: guarantee (w->_thread != Self). acorn@2233: // But in practice an exiting thread may find itself on the EntryList. acorn@2233: // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and acorn@2233: // then calls exit(). Exit release the lock by setting O._owner to NULL. acorn@2233: // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The acorn@2233: // notify() operation moves T1 from O's waitset to O's EntryList. T2 then acorn@2233: // release the lock "O". T2 resumes immediately after the ST of null into acorn@2233: // _owner, above. T2 notices that the EntryList is populated, so it acorn@2233: // reacquires the lock and then finds itself on the EntryList. acorn@2233: // Given all that, we have to tolerate the circumstance where "w" is acorn@2233: // associated with Self. acorn@2233: assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; acorn@2233: ExitEpilog (Self, w) ; acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: // If we find that both _cxq and EntryList are null then just acorn@2233: // re-run the exit protocol from the top. acorn@2233: w = _cxq ; acorn@2233: if (w == NULL) continue ; acorn@2233: acorn@2233: // Drain _cxq into EntryList - bulk transfer. acorn@2233: // First, detach _cxq. acorn@2233: // The following loop is tantamount to: w = swap (&cxq, NULL) acorn@2233: for (;;) { acorn@2233: assert (w != NULL, "Invariant") ; acorn@2233: ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; acorn@2233: if (u == w) break ; acorn@2233: w = u ; acorn@2233: } acorn@2233: TEVENT (Inflated exit - drain cxq into EntryList) ; acorn@2233: acorn@2233: assert (w != NULL , "invariant") ; acorn@2233: assert (_EntryList == NULL , "invariant") ; acorn@2233: acorn@2233: // Convert the LIFO SLL anchored by _cxq into a DLL. acorn@2233: // The list reorganization step operates in O(LENGTH(w)) time. acorn@2233: // It's critical that this step operate quickly as acorn@2233: // "Self" still holds the outer-lock, restricting parallelism acorn@2233: // and effectively lengthening the critical section. acorn@2233: // Invariant: s chases t chases u. acorn@2233: // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so acorn@2233: // we have faster access to the tail. acorn@2233: acorn@2233: if (QMode == 1) { acorn@2233: // QMode == 1 : drain cxq to EntryList, reversing order acorn@2233: // We also reverse the order of the list. acorn@2233: ObjectWaiter * s = NULL ; acorn@2233: ObjectWaiter * t = w ; acorn@2233: ObjectWaiter * u = NULL ; acorn@2233: while (t != NULL) { acorn@2233: guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; acorn@2233: t->TState = ObjectWaiter::TS_ENTER ; acorn@2233: u = t->_next ; acorn@2233: t->_prev = u ; acorn@2233: t->_next = s ; acorn@2233: s = t; acorn@2233: t = u ; acorn@2233: } acorn@2233: _EntryList = s ; acorn@2233: assert (s != NULL, "invariant") ; acorn@2233: } else { acorn@2233: // QMode == 0 or QMode == 2 acorn@2233: _EntryList = w ; acorn@2233: ObjectWaiter * q = NULL ; acorn@2233: ObjectWaiter * p ; acorn@2233: for (p = w ; p != NULL ; p = p->_next) { acorn@2233: guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; acorn@2233: p->TState = ObjectWaiter::TS_ENTER ; acorn@2233: p->_prev = q ; acorn@2233: q = p ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL acorn@2233: // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). acorn@2233: acorn@2233: // See if we can abdicate to a spinner instead of waking a thread. acorn@2233: // A primary goal of the implementation is to reduce the acorn@2233: // context-switch rate. acorn@2233: if (_succ != NULL) continue; acorn@2233: acorn@2233: w = _EntryList ; acorn@2233: if (w != NULL) { acorn@2233: guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; acorn@2233: ExitEpilog (Self, w) ; acorn@2233: return ; acorn@2233: } acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: // ExitSuspendEquivalent: acorn@2233: // A faster alternate to handle_special_suspend_equivalent_condition() acorn@2233: // acorn@2233: // handle_special_suspend_equivalent_condition() unconditionally acorn@2233: // acquires the SR_lock. On some platforms uncontended MutexLocker() acorn@2233: // operations have high latency. Note that in ::enter() we call HSSEC acorn@2233: // while holding the monitor, so we effectively lengthen the critical sections. acorn@2233: // acorn@2233: // There are a number of possible solutions: acorn@2233: // acorn@2233: // A. To ameliorate the problem we might also defer state transitions acorn@2233: // to as late as possible -- just prior to parking. acorn@2233: // Given that, we'd call HSSEC after having returned from park(), acorn@2233: // but before attempting to acquire the monitor. This is only a acorn@2233: // partial solution. It avoids calling HSSEC while holding the acorn@2233: // monitor (good), but it still increases successor reacquisition latency -- acorn@2233: // the interval between unparking a successor and the time the successor acorn@2233: // resumes and retries the lock. See ReenterI(), which defers state transitions. acorn@2233: // If we use this technique we can also avoid EnterI()-exit() loop acorn@2233: // in ::enter() where we iteratively drop the lock and then attempt acorn@2233: // to reacquire it after suspending. acorn@2233: // acorn@2233: // B. In the future we might fold all the suspend bits into a acorn@2233: // composite per-thread suspend flag and then update it with CAS(). acorn@2233: // Alternately, a Dekker-like mechanism with multiple variables acorn@2233: // would suffice: acorn@2233: // ST Self->_suspend_equivalent = false acorn@2233: // MEMBAR acorn@2233: // LD Self_>_suspend_flags acorn@2233: // acorn@2233: acorn@2233: acorn@2233: bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { acorn@2233: int Mode = Knob_FastHSSEC ; acorn@2233: if (Mode && !jSelf->is_external_suspend()) { acorn@2233: assert (jSelf->is_suspend_equivalent(), "invariant") ; acorn@2233: jSelf->clear_suspend_equivalent() ; acorn@2233: if (2 == Mode) OrderAccess::storeload() ; acorn@2233: if (!jSelf->is_external_suspend()) return false ; acorn@2233: // We raced a suspension -- fall thru into the slow path acorn@2233: TEVENT (ExitSuspendEquivalent - raced) ; acorn@2233: jSelf->set_suspend_equivalent() ; acorn@2233: } acorn@2233: return jSelf->handle_special_suspend_equivalent_condition() ; acorn@2233: } acorn@2233: acorn@2233: acorn@2233: void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { acorn@2233: assert (_owner == Self, "invariant") ; acorn@2233: acorn@2233: // Exit protocol: acorn@2233: // 1. ST _succ = wakee acorn@2233: // 2. membar #loadstore|#storestore; acorn@2233: // 2. ST _owner = NULL acorn@2233: // 3. unpark(wakee) acorn@2233: acorn@2233: _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; acorn@2233: ParkEvent * Trigger = Wakee->_event ; acorn@2233: acorn@2233: // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. acorn@2233: // The thread associated with Wakee may have grabbed the lock and "Wakee" may be acorn@2233: // out-of-scope (non-extant). acorn@2233: Wakee = NULL ; acorn@2233: acorn@2233: // Drop the lock acorn@2233: OrderAccess::release_store_ptr (&_owner, NULL) ; acorn@2233: OrderAccess::fence() ; // ST _owner vs LD in unpark() acorn@2233: acorn@2233: if (SafepointSynchronize::do_call_back()) { acorn@2233: TEVENT (unpark before SAFEPOINT) ; acorn@2233: } acorn@2233: acorn@2233: DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); acorn@2233: Trigger->unpark() ; acorn@2233: acorn@2233: // Maintain stats and report events to JVMTI acorn@2233: if (ObjectMonitor::_sync_Parks != NULL) { acorn@2233: ObjectMonitor::_sync_Parks->inc() ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Class Loader deadlock handling. acorn@2233: // acorn@2233: // complete_exit exits a lock returning recursion count acorn@2233: // complete_exit/reenter operate as a wait without waiting acorn@2233: // complete_exit requires an inflated monitor acorn@2233: // The _owner field is not always the Thread addr even with an acorn@2233: // inflated monitor, e.g. the monitor can be inflated by a non-owning acorn@2233: // thread due to contention. acorn@2233: intptr_t ObjectMonitor::complete_exit(TRAPS) { acorn@2233: Thread * const Self = THREAD; acorn@2233: assert(Self->is_Java_thread(), "Must be Java thread!"); acorn@2233: JavaThread *jt = (JavaThread *)THREAD; acorn@2233: acorn@2233: DeferredInitialize(); acorn@2233: acorn@2233: if (THREAD != _owner) { acorn@2233: if (THREAD->is_lock_owned ((address)_owner)) { acorn@2233: assert(_recursions == 0, "internal state error"); acorn@2233: _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ acorn@2233: _recursions = 0 ; acorn@2233: OwnerIsThread = 1 ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: guarantee(Self == _owner, "complete_exit not owner"); acorn@2233: intptr_t save = _recursions; // record the old recursion count acorn@2233: _recursions = 0; // set the recursion level to be 0 sla@5237: exit (true, Self) ; // exit the monitor acorn@2233: guarantee (_owner != Self, "invariant"); acorn@2233: return save; acorn@2233: } acorn@2233: acorn@2233: // reenter() enters a lock and sets recursion count acorn@2233: // complete_exit/reenter operate as a wait without waiting acorn@2233: void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { acorn@2233: Thread * const Self = THREAD; acorn@2233: assert(Self->is_Java_thread(), "Must be Java thread!"); acorn@2233: JavaThread *jt = (JavaThread *)THREAD; acorn@2233: acorn@2233: guarantee(_owner != Self, "reenter already owner"); acorn@2233: enter (THREAD); // enter the monitor acorn@2233: guarantee (_recursions == 0, "reenter recursion"); acorn@2233: _recursions = recursions; acorn@2233: return; acorn@2233: } acorn@2233: acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // A macro is used below because there may already be a pending acorn@2233: // exception which should not abort the execution of the routines acorn@2233: // which use this (which is why we don't put this into check_slow and acorn@2233: // call it with a CHECK argument). acorn@2233: acorn@2233: #define CHECK_OWNER() \ acorn@2233: do { \ acorn@2233: if (THREAD != _owner) { \ acorn@2233: if (THREAD->is_lock_owned((address) _owner)) { \ acorn@2233: _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ acorn@2233: _recursions = 0; \ acorn@2233: OwnerIsThread = 1 ; \ acorn@2233: } else { \ acorn@2233: TEVENT (Throw IMSX) ; \ acorn@2233: THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ acorn@2233: } \ acorn@2233: } \ acorn@2233: } while (false) acorn@2233: acorn@2233: // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. acorn@2233: // TODO-FIXME: remove check_slow() -- it's likely dead. acorn@2233: acorn@2233: void ObjectMonitor::check_slow(TRAPS) { acorn@2233: TEVENT (check_slow - throw IMSX) ; acorn@2233: assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); acorn@2233: THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); acorn@2233: } acorn@2233: acorn@2233: static int Adjust (volatile int * adr, int dx) { acorn@2233: int v ; acorn@2233: for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; acorn@2233: return v ; acorn@2233: } sla@5237: sla@5237: // helper method for posting a monitor wait event sla@5237: void ObjectMonitor::post_monitor_wait_event(EventJavaMonitorWait* event, sla@5237: jlong notifier_tid, sla@5237: jlong timeout, sla@5237: bool timedout) { sla@5237: event->set_klass(((oop)this->object())->klass()); sla@5237: event->set_timeout((TYPE_ULONG)timeout); sla@5237: event->set_address((TYPE_ADDRESS)(uintptr_t)(this->object_addr())); sla@5237: event->set_notifier((TYPE_OSTHREAD)notifier_tid); sla@5237: event->set_timedOut((TYPE_BOOLEAN)timedout); sla@5237: event->commit(); sla@5237: } sla@5237: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Wait/Notify/NotifyAll acorn@2233: // acorn@2233: // Note: a subset of changes to ObjectMonitor::wait() acorn@2233: // will need to be replicated in complete_exit above acorn@2233: void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { acorn@2233: Thread * const Self = THREAD ; acorn@2233: assert(Self->is_Java_thread(), "Must be Java thread!"); acorn@2233: JavaThread *jt = (JavaThread *)THREAD; acorn@2233: acorn@2233: DeferredInitialize () ; acorn@2233: acorn@2233: // Throw IMSX or IEX. acorn@2233: CHECK_OWNER(); acorn@2233: sla@5237: EventJavaMonitorWait event; sla@5237: acorn@2233: // check for a pending interrupt acorn@2233: if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { acorn@2233: // post monitor waited event. Note that this is past-tense, we are done waiting. acorn@2233: if (JvmtiExport::should_post_monitor_waited()) { acorn@2233: // Note: 'false' parameter is passed here because the acorn@2233: // wait was not timed out due to thread interrupt. acorn@2233: JvmtiExport::post_monitor_waited(jt, this, false); dcubed@6335: dcubed@6335: // In this short circuit of the monitor wait protocol, the dcubed@6335: // current thread never drops ownership of the monitor and dcubed@6335: // never gets added to the wait queue so the current thread dcubed@6335: // cannot be made the successor. This means that the dcubed@6335: // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally dcubed@6335: // consume an unpark() meant for the ParkEvent associated with dcubed@6335: // this ObjectMonitor. acorn@2233: } sla@5237: if (event.should_commit()) { sla@5237: post_monitor_wait_event(&event, 0, millis, false); sla@5237: } acorn@2233: TEVENT (Wait - Throw IEX) ; acorn@2233: THROW(vmSymbols::java_lang_InterruptedException()); acorn@2233: return ; acorn@2233: } sla@5237: acorn@2233: TEVENT (Wait) ; acorn@2233: acorn@2233: assert (Self->_Stalled == 0, "invariant") ; acorn@2233: Self->_Stalled = intptr_t(this) ; acorn@2233: jt->set_current_waiting_monitor(this); acorn@2233: acorn@2233: // create a node to be put into the queue acorn@2233: // Critically, after we reset() the event but prior to park(), we must check acorn@2233: // for a pending interrupt. acorn@2233: ObjectWaiter node(Self); acorn@2233: node.TState = ObjectWaiter::TS_WAIT ; acorn@2233: Self->_ParkEvent->reset() ; acorn@2233: OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag acorn@2233: acorn@2233: // Enter the waiting queue, which is a circular doubly linked list in this case acorn@2233: // but it could be a priority queue or any data structure. acorn@2233: // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only acorn@2233: // by the the owner of the monitor *except* in the case where park() acorn@2233: // returns because of a timeout of interrupt. Contention is exceptionally rare acorn@2233: // so we use a simple spin-lock instead of a heavier-weight blocking lock. acorn@2233: acorn@2233: Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; acorn@2233: AddWaiter (&node) ; acorn@2233: Thread::SpinRelease (&_WaitSetLock) ; acorn@2233: acorn@2233: if ((SyncFlags & 4) == 0) { acorn@2233: _Responsible = NULL ; acorn@2233: } acorn@2233: intptr_t save = _recursions; // record the old recursion count acorn@2233: _waiters++; // increment the number of waiters acorn@2233: _recursions = 0; // set the recursion level to be 1 sla@5237: exit (true, Self) ; // exit the monitor acorn@2233: guarantee (_owner != Self, "invariant") ; acorn@2233: acorn@2233: // The thread is on the WaitSet list - now park() it. acorn@2233: // On MP systems it's conceivable that a brief spin before we park acorn@2233: // could be profitable. acorn@2233: // acorn@2233: // TODO-FIXME: change the following logic to a loop of the form acorn@2233: // while (!timeout && !interrupted && _notified == 0) park() acorn@2233: acorn@2233: int ret = OS_OK ; acorn@2233: int WasNotified = 0 ; acorn@2233: { // State transition wrappers acorn@2233: OSThread* osthread = Self->osthread(); acorn@2233: OSThreadWaitState osts(osthread, true); acorn@2233: { acorn@2233: ThreadBlockInVM tbivm(jt); acorn@2233: // Thread is in thread_blocked state and oop access is unsafe. acorn@2233: jt->set_suspend_equivalent(); acorn@2233: acorn@2233: if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { acorn@2233: // Intentionally empty acorn@2233: } else acorn@2233: if (node._notified == 0) { acorn@2233: if (millis <= 0) { acorn@2233: Self->_ParkEvent->park () ; acorn@2233: } else { acorn@2233: ret = Self->_ParkEvent->park (millis) ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: // were we externally suspended while we were waiting? acorn@2233: if (ExitSuspendEquivalent (jt)) { acorn@2233: // TODO-FIXME: add -- if succ == Self then succ = null. acorn@2233: jt->java_suspend_self(); acorn@2233: } acorn@2233: acorn@2233: } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm acorn@2233: acorn@2233: acorn@2233: // Node may be on the WaitSet, the EntryList (or cxq), or in transition acorn@2233: // from the WaitSet to the EntryList. acorn@2233: // See if we need to remove Node from the WaitSet. acorn@2233: // We use double-checked locking to avoid grabbing _WaitSetLock acorn@2233: // if the thread is not on the wait queue. acorn@2233: // acorn@2233: // Note that we don't need a fence before the fetch of TState. acorn@2233: // In the worst case we'll fetch a old-stale value of TS_WAIT previously acorn@2233: // written by the is thread. (perhaps the fetch might even be satisfied acorn@2233: // by a look-aside into the processor's own store buffer, although given acorn@2233: // the length of the code path between the prior ST and this load that's acorn@2233: // highly unlikely). If the following LD fetches a stale TS_WAIT value acorn@2233: // then we'll acquire the lock and then re-fetch a fresh TState value. acorn@2233: // That is, we fail toward safety. acorn@2233: acorn@2233: if (node.TState == ObjectWaiter::TS_WAIT) { acorn@2233: Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; acorn@2233: if (node.TState == ObjectWaiter::TS_WAIT) { acorn@2233: DequeueSpecificWaiter (&node) ; // unlink from WaitSet acorn@2233: assert(node._notified == 0, "invariant"); acorn@2233: node.TState = ObjectWaiter::TS_RUN ; acorn@2233: } acorn@2233: Thread::SpinRelease (&_WaitSetLock) ; acorn@2233: } acorn@2233: acorn@2233: // The thread is now either on off-list (TS_RUN), acorn@2233: // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). acorn@2233: // The Node's TState variable is stable from the perspective of this thread. acorn@2233: // No other threads will asynchronously modify TState. acorn@2233: guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; acorn@2233: OrderAccess::loadload() ; acorn@2233: if (_succ == Self) _succ = NULL ; acorn@2233: WasNotified = node._notified ; acorn@2233: acorn@2233: // Reentry phase -- reacquire the monitor. acorn@2233: // re-enter contended monitor after object.wait(). acorn@2233: // retain OBJECT_WAIT state until re-enter successfully completes acorn@2233: // Thread state is thread_in_vm and oop access is again safe, acorn@2233: // although the raw address of the object may have changed. acorn@2233: // (Don't cache naked oops over safepoints, of course). acorn@2233: acorn@2233: // post monitor waited event. Note that this is past-tense, we are done waiting. acorn@2233: if (JvmtiExport::should_post_monitor_waited()) { acorn@2233: JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); sla@5237: dcubed@6436: if (node._notified != 0 && _succ == Self) { dcubed@6436: // In this part of the monitor wait-notify-reenter protocol it dcubed@6436: // is possible (and normal) for another thread to do a fastpath dcubed@6436: // monitor enter-exit while this thread is still trying to get dcubed@6436: // to the reenter portion of the protocol. dcubed@6436: // dcubed@6436: // The ObjectMonitor was notified and the current thread is dcubed@6436: // the successor which also means that an unpark() has already dcubed@6436: // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can dcubed@6436: // consume the unpark() that was done when the successor was dcubed@6436: // set because the same ParkEvent is shared between Java dcubed@6436: // monitors and JVM/TI RawMonitors (for now). dcubed@6436: // dcubed@6436: // We redo the unpark() to ensure forward progress, i.e., we dcubed@6436: // don't want all pending threads hanging (parked) with none dcubed@6436: // entering the unlocked monitor. dcubed@6436: node._event->unpark(); dcubed@6436: } dcubed@6335: } dcubed@6335: sla@5237: if (event.should_commit()) { sla@5237: post_monitor_wait_event(&event, node._notifier_tid, millis, ret == OS_TIMEOUT); sla@5237: } sla@5237: acorn@2233: OrderAccess::fence() ; acorn@2233: acorn@2233: assert (Self->_Stalled != 0, "invariant") ; acorn@2233: Self->_Stalled = 0 ; acorn@2233: acorn@2233: assert (_owner != Self, "invariant") ; acorn@2233: ObjectWaiter::TStates v = node.TState ; acorn@2233: if (v == ObjectWaiter::TS_RUN) { acorn@2233: enter (Self) ; acorn@2233: } else { acorn@2233: guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; acorn@2233: ReenterI (Self, &node) ; acorn@2233: node.wait_reenter_end(this); acorn@2233: } acorn@2233: acorn@2233: // Self has reacquired the lock. acorn@2233: // Lifecycle - the node representing Self must not appear on any queues. acorn@2233: // Node is about to go out-of-scope, but even if it were immortal we wouldn't acorn@2233: // want residual elements associated with this thread left on any lists. acorn@2233: guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; acorn@2233: assert (_owner == Self, "invariant") ; acorn@2233: assert (_succ != Self , "invariant") ; acorn@2233: } // OSThreadWaitState() acorn@2233: acorn@2233: jt->set_current_waiting_monitor(NULL); acorn@2233: acorn@2233: guarantee (_recursions == 0, "invariant") ; acorn@2233: _recursions = save; // restore the old recursion count acorn@2233: _waiters--; // decrement the number of waiters acorn@2233: acorn@2233: // Verify a few postconditions acorn@2233: assert (_owner == Self , "invariant") ; acorn@2233: assert (_succ != Self , "invariant") ; acorn@2233: assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; acorn@2233: acorn@2233: if (SyncFlags & 32) { acorn@2233: OrderAccess::fence() ; acorn@2233: } acorn@2233: acorn@2233: // check if the notification happened acorn@2233: if (!WasNotified) { acorn@2233: // no, it could be timeout or Thread.interrupt() or both acorn@2233: // check for interrupt event, otherwise it is timeout acorn@2233: if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { acorn@2233: TEVENT (Wait - throw IEX from epilog) ; acorn@2233: THROW(vmSymbols::java_lang_InterruptedException()); acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: // NOTE: Spurious wake up will be consider as timeout. acorn@2233: // Monitor notify has precedence over thread interrupt. acorn@2233: } acorn@2233: acorn@2233: acorn@2233: // Consider: acorn@2233: // If the lock is cool (cxq == null && succ == null) and we're on an MP system acorn@2233: // then instead of transferring a thread from the WaitSet to the EntryList acorn@2233: // we might just dequeue a thread from the WaitSet and directly unpark() it. acorn@2233: acorn@2233: void ObjectMonitor::notify(TRAPS) { acorn@2233: CHECK_OWNER(); acorn@2233: if (_WaitSet == NULL) { acorn@2233: TEVENT (Empty-Notify) ; acorn@2233: return ; acorn@2233: } acorn@2233: DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); acorn@2233: acorn@2233: int Policy = Knob_MoveNotifyee ; acorn@2233: acorn@2233: Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; acorn@2233: ObjectWaiter * iterator = DequeueWaiter() ; acorn@2233: if (iterator != NULL) { acorn@2233: TEVENT (Notify1 - Transfer) ; acorn@2233: guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; acorn@2233: guarantee (iterator->_notified == 0, "invariant") ; acorn@2233: if (Policy != 4) { acorn@2233: iterator->TState = ObjectWaiter::TS_ENTER ; acorn@2233: } acorn@2233: iterator->_notified = 1 ; sla@5237: Thread * Self = THREAD; sla@5237: iterator->_notifier_tid = Self->osthread()->thread_id(); acorn@2233: acorn@2233: ObjectWaiter * List = _EntryList ; acorn@2233: if (List != NULL) { acorn@2233: assert (List->_prev == NULL, "invariant") ; acorn@2233: assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; acorn@2233: assert (List != iterator, "invariant") ; acorn@2233: } acorn@2233: acorn@2233: if (Policy == 0) { // prepend to EntryList acorn@2233: if (List == NULL) { acorn@2233: iterator->_next = iterator->_prev = NULL ; acorn@2233: _EntryList = iterator ; acorn@2233: } else { acorn@2233: List->_prev = iterator ; acorn@2233: iterator->_next = List ; acorn@2233: iterator->_prev = NULL ; acorn@2233: _EntryList = iterator ; acorn@2233: } acorn@2233: } else acorn@2233: if (Policy == 1) { // append to EntryList acorn@2233: if (List == NULL) { acorn@2233: iterator->_next = iterator->_prev = NULL ; acorn@2233: _EntryList = iterator ; acorn@2233: } else { acorn@2233: // CONSIDER: finding the tail currently requires a linear-time walk of acorn@2233: // the EntryList. We can make tail access constant-time by converting to acorn@2233: // a CDLL instead of using our current DLL. acorn@2233: ObjectWaiter * Tail ; acorn@2233: for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; acorn@2233: assert (Tail != NULL && Tail->_next == NULL, "invariant") ; acorn@2233: Tail->_next = iterator ; acorn@2233: iterator->_prev = Tail ; acorn@2233: iterator->_next = NULL ; acorn@2233: } acorn@2233: } else acorn@2233: if (Policy == 2) { // prepend to cxq acorn@2233: // prepend to cxq acorn@2233: if (List == NULL) { acorn@2233: iterator->_next = iterator->_prev = NULL ; acorn@2233: _EntryList = iterator ; acorn@2233: } else { acorn@2233: iterator->TState = ObjectWaiter::TS_CXQ ; acorn@2233: for (;;) { acorn@2233: ObjectWaiter * Front = _cxq ; acorn@2233: iterator->_next = Front ; acorn@2233: if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { acorn@2233: break ; acorn@2233: } acorn@2233: } acorn@2233: } acorn@2233: } else acorn@2233: if (Policy == 3) { // append to cxq acorn@2233: iterator->TState = ObjectWaiter::TS_CXQ ; acorn@2233: for (;;) { acorn@2233: ObjectWaiter * Tail ; acorn@2233: Tail = _cxq ; acorn@2233: if (Tail == NULL) { acorn@2233: iterator->_next = NULL ; acorn@2233: if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { acorn@2233: break ; acorn@2233: } acorn@2233: } else { acorn@2233: while (Tail->_next != NULL) Tail = Tail->_next ; acorn@2233: Tail->_next = iterator ; acorn@2233: iterator->_prev = Tail ; acorn@2233: iterator->_next = NULL ; acorn@2233: break ; acorn@2233: } acorn@2233: } acorn@2233: } else { acorn@2233: ParkEvent * ev = iterator->_event ; acorn@2233: iterator->TState = ObjectWaiter::TS_RUN ; acorn@2233: OrderAccess::fence() ; acorn@2233: ev->unpark() ; acorn@2233: } acorn@2233: acorn@2233: if (Policy < 4) { acorn@2233: iterator->wait_reenter_begin(this); acorn@2233: } acorn@2233: acorn@2233: // _WaitSetLock protects the wait queue, not the EntryList. We could acorn@2233: // move the add-to-EntryList operation, above, outside the critical section acorn@2233: // protected by _WaitSetLock. In practice that's not useful. With the acorn@2233: // exception of wait() timeouts and interrupts the monitor owner acorn@2233: // is the only thread that grabs _WaitSetLock. There's almost no contention acorn@2233: // on _WaitSetLock so it's not profitable to reduce the length of the acorn@2233: // critical section. acorn@2233: } acorn@2233: acorn@2233: Thread::SpinRelease (&_WaitSetLock) ; acorn@2233: acorn@2233: if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { acorn@2233: ObjectMonitor::_sync_Notifications->inc() ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: acorn@2233: void ObjectMonitor::notifyAll(TRAPS) { acorn@2233: CHECK_OWNER(); acorn@2233: ObjectWaiter* iterator; acorn@2233: if (_WaitSet == NULL) { acorn@2233: TEVENT (Empty-NotifyAll) ; acorn@2233: return ; acorn@2233: } acorn@2233: DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); acorn@2233: acorn@2233: int Policy = Knob_MoveNotifyee ; acorn@2233: int Tally = 0 ; acorn@2233: Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; acorn@2233: acorn@2233: for (;;) { acorn@2233: iterator = DequeueWaiter () ; acorn@2233: if (iterator == NULL) break ; acorn@2233: TEVENT (NotifyAll - Transfer1) ; acorn@2233: ++Tally ; acorn@2233: acorn@2233: // Disposition - what might we do with iterator ? acorn@2233: // a. add it directly to the EntryList - either tail or head. acorn@2233: // b. push it onto the front of the _cxq. acorn@2233: // For now we use (a). acorn@2233: acorn@2233: guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; acorn@2233: guarantee (iterator->_notified == 0, "invariant") ; acorn@2233: iterator->_notified = 1 ; sla@5237: Thread * Self = THREAD; sla@5237: iterator->_notifier_tid = Self->osthread()->thread_id(); acorn@2233: if (Policy != 4) { acorn@2233: iterator->TState = ObjectWaiter::TS_ENTER ; acorn@2233: } acorn@2233: acorn@2233: ObjectWaiter * List = _EntryList ; acorn@2233: if (List != NULL) { acorn@2233: assert (List->_prev == NULL, "invariant") ; acorn@2233: assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; acorn@2233: assert (List != iterator, "invariant") ; acorn@2233: } acorn@2233: acorn@2233: if (Policy == 0) { // prepend to EntryList acorn@2233: if (List == NULL) { acorn@2233: iterator->_next = iterator->_prev = NULL ; acorn@2233: _EntryList = iterator ; acorn@2233: } else { acorn@2233: List->_prev = iterator ; acorn@2233: iterator->_next = List ; acorn@2233: iterator->_prev = NULL ; acorn@2233: _EntryList = iterator ; acorn@2233: } acorn@2233: } else acorn@2233: if (Policy == 1) { // append to EntryList acorn@2233: if (List == NULL) { acorn@2233: iterator->_next = iterator->_prev = NULL ; acorn@2233: _EntryList = iterator ; acorn@2233: } else { acorn@2233: // CONSIDER: finding the tail currently requires a linear-time walk of acorn@2233: // the EntryList. We can make tail access constant-time by converting to acorn@2233: // a CDLL instead of using our current DLL. acorn@2233: ObjectWaiter * Tail ; acorn@2233: for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; acorn@2233: assert (Tail != NULL && Tail->_next == NULL, "invariant") ; acorn@2233: Tail->_next = iterator ; acorn@2233: iterator->_prev = Tail ; acorn@2233: iterator->_next = NULL ; acorn@2233: } acorn@2233: } else acorn@2233: if (Policy == 2) { // prepend to cxq acorn@2233: // prepend to cxq acorn@2233: iterator->TState = ObjectWaiter::TS_CXQ ; acorn@2233: for (;;) { acorn@2233: ObjectWaiter * Front = _cxq ; acorn@2233: iterator->_next = Front ; acorn@2233: if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { acorn@2233: break ; acorn@2233: } acorn@2233: } acorn@2233: } else acorn@2233: if (Policy == 3) { // append to cxq acorn@2233: iterator->TState = ObjectWaiter::TS_CXQ ; acorn@2233: for (;;) { acorn@2233: ObjectWaiter * Tail ; acorn@2233: Tail = _cxq ; acorn@2233: if (Tail == NULL) { acorn@2233: iterator->_next = NULL ; acorn@2233: if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { acorn@2233: break ; acorn@2233: } acorn@2233: } else { acorn@2233: while (Tail->_next != NULL) Tail = Tail->_next ; acorn@2233: Tail->_next = iterator ; acorn@2233: iterator->_prev = Tail ; acorn@2233: iterator->_next = NULL ; acorn@2233: break ; acorn@2233: } acorn@2233: } acorn@2233: } else { acorn@2233: ParkEvent * ev = iterator->_event ; acorn@2233: iterator->TState = ObjectWaiter::TS_RUN ; acorn@2233: OrderAccess::fence() ; acorn@2233: ev->unpark() ; acorn@2233: } acorn@2233: acorn@2233: if (Policy < 4) { acorn@2233: iterator->wait_reenter_begin(this); acorn@2233: } acorn@2233: acorn@2233: // _WaitSetLock protects the wait queue, not the EntryList. We could acorn@2233: // move the add-to-EntryList operation, above, outside the critical section acorn@2233: // protected by _WaitSetLock. In practice that's not useful. With the acorn@2233: // exception of wait() timeouts and interrupts the monitor owner acorn@2233: // is the only thread that grabs _WaitSetLock. There's almost no contention acorn@2233: // on _WaitSetLock so it's not profitable to reduce the length of the acorn@2233: // critical section. acorn@2233: } acorn@2233: acorn@2233: Thread::SpinRelease (&_WaitSetLock) ; acorn@2233: acorn@2233: if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { acorn@2233: ObjectMonitor::_sync_Notifications->inc(Tally) ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Adaptive Spinning Support acorn@2233: // acorn@2233: // Adaptive spin-then-block - rational spinning acorn@2233: // acorn@2233: // Note that we spin "globally" on _owner with a classic SMP-polite TATAS acorn@2233: // algorithm. On high order SMP systems it would be better to start with acorn@2233: // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, acorn@2233: // a contending thread could enqueue itself on the cxq and then spin locally acorn@2233: // on a thread-specific variable such as its ParkEvent._Event flag. acorn@2233: // That's left as an exercise for the reader. Note that global spinning is acorn@2233: // not problematic on Niagara, as the L2$ serves the interconnect and has both acorn@2233: // low latency and massive bandwidth. acorn@2233: // acorn@2233: // Broadly, we can fix the spin frequency -- that is, the % of contended lock acorn@2233: // acquisition attempts where we opt to spin -- at 100% and vary the spin count acorn@2233: // (duration) or we can fix the count at approximately the duration of acorn@2233: // a context switch and vary the frequency. Of course we could also acorn@2233: // 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. acorn@2233: // acorn@2233: // This implementation varies the duration "D", where D varies with acorn@2233: // the success rate of recent spin attempts. (D is capped at approximately acorn@2233: // length of a round-trip context switch). The success rate for recent acorn@2233: // spin attempts is a good predictor of the success rate of future spin acorn@2233: // attempts. The mechanism adapts automatically to varying critical acorn@2233: // section length (lock modality), system load and degree of parallelism. acorn@2233: // D is maintained per-monitor in _SpinDuration and is initialized acorn@2233: // optimistically. Spin frequency is fixed at 100%. acorn@2233: // acorn@2233: // Note that _SpinDuration is volatile, but we update it without locks acorn@2233: // or atomics. The code is designed so that _SpinDuration stays within acorn@2233: // a reasonable range even in the presence of races. The arithmetic acorn@2233: // operations on _SpinDuration are closed over the domain of legal values, acorn@2233: // so at worst a race will install and older but still legal value. acorn@2233: // At the very worst this introduces some apparent non-determinism. acorn@2233: // We might spin when we shouldn't or vice-versa, but since the spin acorn@2233: // count are relatively short, even in the worst case, the effect is harmless. acorn@2233: // acorn@2233: // Care must be taken that a low "D" value does not become an acorn@2233: // an absorbing state. Transient spinning failures -- when spinning acorn@2233: // is overall profitable -- should not cause the system to converge acorn@2233: // on low "D" values. We want spinning to be stable and predictable acorn@2233: // and fairly responsive to change and at the same time we don't want acorn@2233: // it to oscillate, become metastable, be "too" non-deterministic, acorn@2233: // or converge on or enter undesirable stable absorbing states. acorn@2233: // acorn@2233: // We implement a feedback-based control system -- using past behavior acorn@2233: // to predict future behavior. We face two issues: (a) if the acorn@2233: // input signal is random then the spin predictor won't provide optimal acorn@2233: // results, and (b) if the signal frequency is too high then the control acorn@2233: // system, which has some natural response lag, will "chase" the signal. acorn@2233: // (b) can arise from multimodal lock hold times. Transient preemption acorn@2233: // can also result in apparent bimodal lock hold times. acorn@2233: // Although sub-optimal, neither condition is particularly harmful, as acorn@2233: // in the worst-case we'll spin when we shouldn't or vice-versa. acorn@2233: // The maximum spin duration is rather short so the failure modes aren't bad. acorn@2233: // To be conservative, I've tuned the gain in system to bias toward acorn@2233: // _not spinning. Relatedly, the system can sometimes enter a mode where it acorn@2233: // "rings" or oscillates between spinning and not spinning. This happens acorn@2233: // when spinning is just on the cusp of profitability, however, so the acorn@2233: // situation is not dire. The state is benign -- there's no need to add acorn@2233: // hysteresis control to damp the transition rate between spinning and acorn@2233: // not spinning. acorn@2233: // acorn@2233: acorn@2233: intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; acorn@2233: int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; acorn@2233: acorn@2233: // Spinning: Fixed frequency (100%), vary duration acorn@2233: acorn@2233: acorn@2233: int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { acorn@2233: acorn@2233: // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. acorn@2233: int ctr = Knob_FixedSpin ; acorn@2233: if (ctr != 0) { acorn@2233: while (--ctr >= 0) { acorn@2233: if (TryLock (Self) > 0) return 1 ; acorn@2233: SpinPause () ; acorn@2233: } acorn@2233: return 0 ; acorn@2233: } acorn@2233: acorn@2233: for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { acorn@2233: if (TryLock(Self) > 0) { acorn@2233: // Increase _SpinDuration ... acorn@2233: // Note that we don't clamp SpinDuration precisely at SpinLimit. acorn@2233: // Raising _SpurDuration to the poverty line is key. acorn@2233: int x = _SpinDuration ; acorn@2233: if (x < Knob_SpinLimit) { acorn@2233: if (x < Knob_Poverty) x = Knob_Poverty ; acorn@2233: _SpinDuration = x + Knob_BonusB ; acorn@2233: } acorn@2233: return 1 ; acorn@2233: } acorn@2233: SpinPause () ; acorn@2233: } acorn@2233: acorn@2233: // Admission control - verify preconditions for spinning acorn@2233: // acorn@2233: // We always spin a little bit, just to prevent _SpinDuration == 0 from acorn@2233: // becoming an absorbing state. Put another way, we spin briefly to acorn@2233: // sample, just in case the system load, parallelism, contention, or lock acorn@2233: // modality changed. acorn@2233: // acorn@2233: // Consider the following alternative: acorn@2233: // Periodically set _SpinDuration = _SpinLimit and try a long/full acorn@2233: // spin attempt. "Periodically" might mean after a tally of acorn@2233: // the # of failed spin attempts (or iterations) reaches some threshold. acorn@2233: // This takes us into the realm of 1-out-of-N spinning, where we acorn@2233: // hold the duration constant but vary the frequency. acorn@2233: acorn@2233: ctr = _SpinDuration ; acorn@2233: if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; acorn@2233: if (ctr <= 0) return 0 ; acorn@2233: acorn@2233: if (Knob_SuccRestrict && _succ != NULL) return 0 ; acorn@2233: if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { acorn@2233: TEVENT (Spin abort - notrunnable [TOP]); acorn@2233: return 0 ; acorn@2233: } acorn@2233: acorn@2233: int MaxSpin = Knob_MaxSpinners ; acorn@2233: if (MaxSpin >= 0) { acorn@2233: if (_Spinner > MaxSpin) { acorn@2233: TEVENT (Spin abort -- too many spinners) ; acorn@2233: return 0 ; acorn@2233: } acorn@2233: // Slighty racy, but benign ... acorn@2233: Adjust (&_Spinner, 1) ; acorn@2233: } acorn@2233: acorn@2233: // We're good to spin ... spin ingress. acorn@2233: // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades acorn@2233: // when preparing to LD...CAS _owner, etc and the CAS is likely acorn@2233: // to succeed. acorn@2233: int hits = 0 ; acorn@2233: int msk = 0 ; acorn@2233: int caspty = Knob_CASPenalty ; acorn@2233: int oxpty = Knob_OXPenalty ; acorn@2233: int sss = Knob_SpinSetSucc ; acorn@2233: if (sss && _succ == NULL ) _succ = Self ; acorn@2233: Thread * prv = NULL ; acorn@2233: acorn@2233: // There are three ways to exit the following loop: acorn@2233: // 1. A successful spin where this thread has acquired the lock. acorn@2233: // 2. Spin failure with prejudice acorn@2233: // 3. Spin failure without prejudice acorn@2233: acorn@2233: while (--ctr >= 0) { acorn@2233: acorn@2233: // Periodic polling -- Check for pending GC acorn@2233: // Threads may spin while they're unsafe. acorn@2233: // We don't want spinning threads to delay the JVM from reaching acorn@2233: // a stop-the-world safepoint or to steal cycles from GC. acorn@2233: // If we detect a pending safepoint we abort in order that acorn@2233: // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) acorn@2233: // this thread, if safe, doesn't steal cycles from GC. acorn@2233: // This is in keeping with the "no loitering in runtime" rule. acorn@2233: // We periodically check to see if there's a safepoint pending. acorn@2233: if ((ctr & 0xFF) == 0) { acorn@2233: if (SafepointSynchronize::do_call_back()) { acorn@2233: TEVENT (Spin: safepoint) ; acorn@2233: goto Abort ; // abrupt spin egress acorn@2233: } acorn@2233: if (Knob_UsePause & 1) SpinPause () ; acorn@2233: acorn@2233: int (*scb)(intptr_t,int) = SpinCallbackFunction ; acorn@2233: if (hits > 50 && scb != NULL) { acorn@2233: int abend = (*scb)(SpinCallbackArgument, 0) ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: if (Knob_UsePause & 2) SpinPause() ; acorn@2233: acorn@2233: // Exponential back-off ... Stay off the bus to reduce coherency traffic. acorn@2233: // This is useful on classic SMP systems, but is of less utility on acorn@2233: // N1-style CMT platforms. acorn@2233: // acorn@2233: // Trade-off: lock acquisition latency vs coherency bandwidth. acorn@2233: // Lock hold times are typically short. A histogram acorn@2233: // of successful spin attempts shows that we usually acquire acorn@2233: // the lock early in the spin. That suggests we want to acorn@2233: // sample _owner frequently in the early phase of the spin, acorn@2233: // but then back-off and sample less frequently as the spin acorn@2233: // progresses. The back-off makes a good citizen on SMP big acorn@2233: // SMP systems. Oversampling _owner can consume excessive acorn@2233: // coherency bandwidth. Relatedly, if we _oversample _owner we acorn@2233: // can inadvertently interfere with the the ST m->owner=null. acorn@2233: // executed by the lock owner. acorn@2233: if (ctr & msk) continue ; acorn@2233: ++hits ; acorn@2233: if ((hits & 0xF) == 0) { acorn@2233: // The 0xF, above, corresponds to the exponent. acorn@2233: // Consider: (msk+1)|msk acorn@2233: msk = ((msk << 2)|3) & BackOffMask ; acorn@2233: } acorn@2233: acorn@2233: // Probe _owner with TATAS acorn@2233: // If this thread observes the monitor transition or flicker acorn@2233: // from locked to unlocked to locked, then the odds that this acorn@2233: // thread will acquire the lock in this spin attempt go down acorn@2233: // considerably. The same argument applies if the CAS fails acorn@2233: // or if we observe _owner change from one non-null value to acorn@2233: // another non-null value. In such cases we might abort acorn@2233: // the spin without prejudice or apply a "penalty" to the acorn@2233: // spin count-down variable "ctr", reducing it by 100, say. acorn@2233: acorn@2233: Thread * ox = (Thread *) _owner ; acorn@2233: if (ox == NULL) { acorn@2233: ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; acorn@2233: if (ox == NULL) { acorn@2233: // The CAS succeeded -- this thread acquired ownership acorn@2233: // Take care of some bookkeeping to exit spin state. acorn@2233: if (sss && _succ == Self) { acorn@2233: _succ = NULL ; acorn@2233: } acorn@2233: if (MaxSpin > 0) Adjust (&_Spinner, -1) ; acorn@2233: acorn@2233: // Increase _SpinDuration : acorn@2233: // The spin was successful (profitable) so we tend toward acorn@2233: // longer spin attempts in the future. acorn@2233: // CONSIDER: factor "ctr" into the _SpinDuration adjustment. acorn@2233: // If we acquired the lock early in the spin cycle it acorn@2233: // makes sense to increase _SpinDuration proportionally. acorn@2233: // Note that we don't clamp SpinDuration precisely at SpinLimit. acorn@2233: int x = _SpinDuration ; acorn@2233: if (x < Knob_SpinLimit) { acorn@2233: if (x < Knob_Poverty) x = Knob_Poverty ; acorn@2233: _SpinDuration = x + Knob_Bonus ; acorn@2233: } acorn@2233: return 1 ; acorn@2233: } acorn@2233: acorn@2233: // The CAS failed ... we can take any of the following actions: acorn@2233: // * penalize: ctr -= Knob_CASPenalty acorn@2233: // * exit spin with prejudice -- goto Abort; acorn@2233: // * exit spin without prejudice. acorn@2233: // * Since CAS is high-latency, retry again immediately. acorn@2233: prv = ox ; acorn@2233: TEVENT (Spin: cas failed) ; acorn@2233: if (caspty == -2) break ; acorn@2233: if (caspty == -1) goto Abort ; acorn@2233: ctr -= caspty ; acorn@2233: continue ; acorn@2233: } acorn@2233: acorn@2233: // Did lock ownership change hands ? acorn@2233: if (ox != prv && prv != NULL ) { acorn@2233: TEVENT (spin: Owner changed) acorn@2233: if (oxpty == -2) break ; acorn@2233: if (oxpty == -1) goto Abort ; acorn@2233: ctr -= oxpty ; acorn@2233: } acorn@2233: prv = ox ; acorn@2233: acorn@2233: // Abort the spin if the owner is not executing. acorn@2233: // The owner must be executing in order to drop the lock. acorn@2233: // Spinning while the owner is OFFPROC is idiocy. acorn@2233: // Consider: ctr -= RunnablePenalty ; acorn@2233: if (Knob_OState && NotRunnable (Self, ox)) { acorn@2233: TEVENT (Spin abort - notrunnable); acorn@2233: goto Abort ; acorn@2233: } acorn@2233: if (sss && _succ == NULL ) _succ = Self ; acorn@2233: } acorn@2233: acorn@2233: // Spin failed with prejudice -- reduce _SpinDuration. acorn@2233: // TODO: Use an AIMD-like policy to adjust _SpinDuration. acorn@2233: // AIMD is globally stable. acorn@2233: TEVENT (Spin failure) ; acorn@2233: { acorn@2233: int x = _SpinDuration ; acorn@2233: if (x > 0) { acorn@2233: // Consider an AIMD scheme like: x -= (x >> 3) + 100 acorn@2233: // This is globally sample and tends to damp the response. acorn@2233: x -= Knob_Penalty ; acorn@2233: if (x < 0) x = 0 ; acorn@2233: _SpinDuration = x ; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: Abort: acorn@2233: if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; acorn@2233: if (sss && _succ == Self) { acorn@2233: _succ = NULL ; acorn@2233: // Invariant: after setting succ=null a contending thread acorn@2233: // must recheck-retry _owner before parking. This usually happens acorn@2233: // in the normal usage of TrySpin(), but it's safest acorn@2233: // to make TrySpin() as foolproof as possible. acorn@2233: OrderAccess::fence() ; acorn@2233: if (TryLock(Self) > 0) return 1 ; acorn@2233: } acorn@2233: return 0 ; acorn@2233: } acorn@2233: acorn@2233: // NotRunnable() -- informed spinning acorn@2233: // acorn@2233: // Don't bother spinning if the owner is not eligible to drop the lock. acorn@2233: // Peek at the owner's schedctl.sc_state and Thread._thread_values and acorn@2233: // spin only if the owner thread is _thread_in_Java or _thread_in_vm. acorn@2233: // The thread must be runnable in order to drop the lock in timely fashion. acorn@2233: // If the _owner is not runnable then spinning will not likely be acorn@2233: // successful (profitable). acorn@2233: // acorn@2233: // Beware -- the thread referenced by _owner could have died acorn@2233: // so a simply fetch from _owner->_thread_state might trap. acorn@2233: // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. acorn@2233: // Because of the lifecycle issues the schedctl and _thread_state values acorn@2233: // observed by NotRunnable() might be garbage. NotRunnable must acorn@2233: // tolerate this and consider the observed _thread_state value acorn@2233: // as advisory. acorn@2233: // acorn@2233: // Beware too, that _owner is sometimes a BasicLock address and sometimes acorn@2233: // a thread pointer. We differentiate the two cases with OwnerIsThread. acorn@2233: // Alternately, we might tag the type (thread pointer vs basiclock pointer) acorn@2233: // with the LSB of _owner. Another option would be to probablistically probe acorn@2233: // the putative _owner->TypeTag value. acorn@2233: // acorn@2233: // Checking _thread_state isn't perfect. Even if the thread is acorn@2233: // in_java it might be blocked on a page-fault or have been preempted acorn@2233: // and sitting on a ready/dispatch queue. _thread state in conjunction acorn@2233: // with schedctl.sc_state gives us a good picture of what the acorn@2233: // thread is doing, however. acorn@2233: // acorn@2233: // TODO: check schedctl.sc_state. acorn@2233: // We'll need to use SafeFetch32() to read from the schedctl block. acorn@2233: // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ acorn@2233: // acorn@2233: // The return value from NotRunnable() is *advisory* -- the acorn@2233: // result is based on sampling and is not necessarily coherent. acorn@2233: // The caller must tolerate false-negative and false-positive errors. acorn@2233: // Spinning, in general, is probabilistic anyway. acorn@2233: acorn@2233: acorn@2233: int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { acorn@2233: // Check either OwnerIsThread or ox->TypeTag == 2BAD. acorn@2233: if (!OwnerIsThread) return 0 ; acorn@2233: acorn@2233: if (ox == NULL) return 0 ; acorn@2233: acorn@2233: // Avoid transitive spinning ... acorn@2233: // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. acorn@2233: // Immediately after T1 acquires L it's possible that T2, also acorn@2233: // spinning on L, will see L.Owner=T1 and T1._Stalled=L. acorn@2233: // This occurs transiently after T1 acquired L but before acorn@2233: // T1 managed to clear T1.Stalled. T2 does not need to abort acorn@2233: // its spin in this circumstance. acorn@2233: intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; acorn@2233: acorn@2233: if (BlockedOn == 1) return 1 ; acorn@2233: if (BlockedOn != 0) { acorn@2233: return BlockedOn != intptr_t(this) && _owner == ox ; acorn@2233: } acorn@2233: acorn@2233: assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; acorn@2233: int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; acorn@2233: // consider also: jst != _thread_in_Java -- but that's overspecific. acorn@2233: return jst == _thread_blocked || jst == _thread_in_native ; acorn@2233: } acorn@2233: acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // WaitSet management ... acorn@2233: acorn@2233: ObjectWaiter::ObjectWaiter(Thread* thread) { acorn@2233: _next = NULL; acorn@2233: _prev = NULL; acorn@2233: _notified = 0; acorn@2233: TState = TS_RUN ; acorn@2233: _thread = thread; acorn@2233: _event = thread->_ParkEvent ; acorn@2233: _active = false; acorn@2233: assert (_event != NULL, "invariant") ; acorn@2233: } acorn@2233: acorn@2233: void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { acorn@2233: JavaThread *jt = (JavaThread *)this->_thread; acorn@2233: _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); acorn@2233: } acorn@2233: acorn@2233: void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { acorn@2233: JavaThread *jt = (JavaThread *)this->_thread; acorn@2233: JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); acorn@2233: } acorn@2233: acorn@2233: inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { acorn@2233: assert(node != NULL, "should not dequeue NULL node"); acorn@2233: assert(node->_prev == NULL, "node already in list"); acorn@2233: assert(node->_next == NULL, "node already in list"); acorn@2233: // put node at end of queue (circular doubly linked list) acorn@2233: if (_WaitSet == NULL) { acorn@2233: _WaitSet = node; acorn@2233: node->_prev = node; acorn@2233: node->_next = node; acorn@2233: } else { acorn@2233: ObjectWaiter* head = _WaitSet ; acorn@2233: ObjectWaiter* tail = head->_prev; acorn@2233: assert(tail->_next == head, "invariant check"); acorn@2233: tail->_next = node; acorn@2233: head->_prev = node; acorn@2233: node->_next = head; acorn@2233: node->_prev = tail; acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { acorn@2233: // dequeue the very first waiter acorn@2233: ObjectWaiter* waiter = _WaitSet; acorn@2233: if (waiter) { acorn@2233: DequeueSpecificWaiter(waiter); acorn@2233: } acorn@2233: return waiter; acorn@2233: } acorn@2233: acorn@2233: inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { acorn@2233: assert(node != NULL, "should not dequeue NULL node"); acorn@2233: assert(node->_prev != NULL, "node already removed from list"); acorn@2233: assert(node->_next != NULL, "node already removed from list"); acorn@2233: // when the waiter has woken up because of interrupt, acorn@2233: // timeout or other spurious wake-up, dequeue the acorn@2233: // waiter from waiting list acorn@2233: ObjectWaiter* next = node->_next; acorn@2233: if (next == node) { acorn@2233: assert(node->_prev == node, "invariant check"); acorn@2233: _WaitSet = NULL; acorn@2233: } else { acorn@2233: ObjectWaiter* prev = node->_prev; acorn@2233: assert(prev->_next == node, "invariant check"); acorn@2233: assert(next->_prev == node, "invariant check"); acorn@2233: next->_prev = prev; acorn@2233: prev->_next = next; acorn@2233: if (_WaitSet == node) { acorn@2233: _WaitSet = next; acorn@2233: } acorn@2233: } acorn@2233: node->_next = NULL; acorn@2233: node->_prev = NULL; acorn@2233: } acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // PerfData support acorn@2233: PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_Parks = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; acorn@2233: PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; acorn@2233: PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; acorn@2233: acorn@2233: // One-shot global initialization for the sync subsystem. acorn@2233: // We could also defer initialization and initialize on-demand acorn@2233: // the first time we call inflate(). Initialization would acorn@2233: // be protected - like so many things - by the MonitorCache_lock. acorn@2233: acorn@2233: void ObjectMonitor::Initialize () { acorn@2233: static int InitializationCompleted = 0 ; acorn@2233: assert (InitializationCompleted == 0, "invariant") ; acorn@2233: InitializationCompleted = 1 ; acorn@2233: if (UsePerfData) { acorn@2233: EXCEPTION_MARK ; acorn@2233: #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } acorn@2233: #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } acorn@2233: NEWPERFCOUNTER(_sync_Inflations) ; acorn@2233: NEWPERFCOUNTER(_sync_Deflations) ; acorn@2233: NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; acorn@2233: NEWPERFCOUNTER(_sync_FutileWakeups) ; acorn@2233: NEWPERFCOUNTER(_sync_Parks) ; acorn@2233: NEWPERFCOUNTER(_sync_EmptyNotifications) ; acorn@2233: NEWPERFCOUNTER(_sync_Notifications) ; acorn@2233: NEWPERFCOUNTER(_sync_SlowEnter) ; acorn@2233: NEWPERFCOUNTER(_sync_SlowExit) ; acorn@2233: NEWPERFCOUNTER(_sync_SlowNotify) ; acorn@2233: NEWPERFCOUNTER(_sync_SlowNotifyAll) ; acorn@2233: NEWPERFCOUNTER(_sync_FailedSpins) ; acorn@2233: NEWPERFCOUNTER(_sync_SuccessfulSpins) ; acorn@2233: NEWPERFCOUNTER(_sync_PrivateA) ; acorn@2233: NEWPERFCOUNTER(_sync_PrivateB) ; acorn@2233: NEWPERFCOUNTER(_sync_MonInCirculation) ; acorn@2233: NEWPERFCOUNTER(_sync_MonScavenged) ; acorn@2233: NEWPERFVARIABLE(_sync_MonExtant) ; acorn@2233: #undef NEWPERFCOUNTER acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: acorn@2233: // Compile-time asserts acorn@2233: // When possible, it's better to catch errors deterministically at acorn@2233: // compile-time than at runtime. The down-side to using compile-time acorn@2233: // asserts is that error message -- often something about negative array acorn@2233: // indices -- is opaque. acorn@2233: acorn@2233: #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } acorn@2233: acorn@2233: void ObjectMonitor::ctAsserts() { acorn@2233: CTASSERT(offset_of (ObjectMonitor, _header) == 0); acorn@2233: } acorn@2233: acorn@2233: acorn@2233: static char * kvGet (char * kvList, const char * Key) { acorn@2233: if (kvList == NULL) return NULL ; acorn@2233: size_t n = strlen (Key) ; acorn@2233: char * Search ; acorn@2233: for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { acorn@2233: if (strncmp (Search, Key, n) == 0) { acorn@2233: if (Search[n] == '=') return Search + n + 1 ; acorn@2233: if (Search[n] == 0) return (char *) "1" ; acorn@2233: } acorn@2233: } acorn@2233: return NULL ; acorn@2233: } acorn@2233: acorn@2233: static int kvGetInt (char * kvList, const char * Key, int Default) { acorn@2233: char * v = kvGet (kvList, Key) ; acorn@2233: int rslt = v ? ::strtol (v, NULL, 0) : Default ; acorn@2233: if (Knob_ReportSettings && v != NULL) { acorn@2233: ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; acorn@2233: ::fflush (stdout) ; acorn@2233: } acorn@2233: return rslt ; acorn@2233: } acorn@2233: acorn@2233: void ObjectMonitor::DeferredInitialize () { acorn@2233: if (InitDone > 0) return ; acorn@2233: if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { acorn@2233: while (InitDone != 1) ; acorn@2233: return ; acorn@2233: } acorn@2233: acorn@2233: // One-shot global initialization ... acorn@2233: // The initialization is idempotent, so we don't need locks. acorn@2233: // In the future consider doing this via os::init_2(). acorn@2233: // SyncKnobs consist of = pairs in the style acorn@2233: // of environment variables. Start by converting ':' to NUL. acorn@2233: acorn@2233: if (SyncKnobs == NULL) SyncKnobs = "" ; acorn@2233: acorn@2233: size_t sz = strlen (SyncKnobs) ; acorn@2233: char * knobs = (char *) malloc (sz + 2) ; acorn@2233: if (knobs == NULL) { ccheung@4993: vm_exit_out_of_memory (sz + 2, OOM_MALLOC_ERROR, "Parse SyncKnobs") ; acorn@2233: guarantee (0, "invariant") ; acorn@2233: } acorn@2233: strcpy (knobs, SyncKnobs) ; acorn@2233: knobs[sz+1] = 0 ; acorn@2233: for (char * p = knobs ; *p ; p++) { acorn@2233: if (*p == ':') *p = 0 ; acorn@2233: } acorn@2233: acorn@2233: #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } acorn@2233: SETKNOB(ReportSettings) ; acorn@2233: SETKNOB(Verbose) ; acorn@2233: SETKNOB(FixedSpin) ; acorn@2233: SETKNOB(SpinLimit) ; acorn@2233: SETKNOB(SpinBase) ; acorn@2233: SETKNOB(SpinBackOff); acorn@2233: SETKNOB(CASPenalty) ; acorn@2233: SETKNOB(OXPenalty) ; acorn@2233: SETKNOB(LogSpins) ; acorn@2233: SETKNOB(SpinSetSucc) ; acorn@2233: SETKNOB(SuccEnabled) ; acorn@2233: SETKNOB(SuccRestrict) ; acorn@2233: SETKNOB(Penalty) ; acorn@2233: SETKNOB(Bonus) ; acorn@2233: SETKNOB(BonusB) ; acorn@2233: SETKNOB(Poverty) ; acorn@2233: SETKNOB(SpinAfterFutile) ; acorn@2233: SETKNOB(UsePause) ; acorn@2233: SETKNOB(SpinEarly) ; acorn@2233: SETKNOB(OState) ; acorn@2233: SETKNOB(MaxSpinners) ; acorn@2233: SETKNOB(PreSpin) ; acorn@2233: SETKNOB(ExitPolicy) ; acorn@2233: SETKNOB(QMode); acorn@2233: SETKNOB(ResetEvent) ; acorn@2233: SETKNOB(MoveNotifyee) ; acorn@2233: SETKNOB(FastHSSEC) ; acorn@2233: #undef SETKNOB acorn@2233: kevinw@8729: if (Knob_Verbose) { kevinw@8729: sanity_checks(); kevinw@8729: } kevinw@8729: acorn@2233: if (os::is_MP()) { acorn@2233: BackOffMask = (1 << Knob_SpinBackOff) - 1 ; acorn@2233: if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; acorn@2233: // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) acorn@2233: } else { acorn@2233: Knob_SpinLimit = 0 ; acorn@2233: Knob_SpinBase = 0 ; acorn@2233: Knob_PreSpin = 0 ; acorn@2233: Knob_FixedSpin = -1 ; acorn@2233: } acorn@2233: acorn@2233: if (Knob_LogSpins == 0) { acorn@2233: ObjectMonitor::_sync_FailedSpins = NULL ; acorn@2233: } acorn@2233: acorn@2233: free (knobs) ; acorn@2233: OrderAccess::fence() ; acorn@2233: InitDone = 1 ; acorn@2233: } acorn@2233: 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: acorn@2233: #ifndef PRODUCT acorn@2233: void ObjectMonitor::verify() { acorn@2233: } acorn@2233: acorn@2233: void ObjectMonitor::print() { acorn@2233: } acorn@2233: #endif