duke@435: /* dcubed@4470: * Copyright (c) 1998, 2013, Oracle and/or its affiliates. All rights reserved. duke@435: * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. duke@435: * duke@435: * This code is free software; you can redistribute it and/or modify it duke@435: * under the terms of the GNU General Public License version 2 only, as duke@435: * published by the Free Software Foundation. duke@435: * duke@435: * This code is distributed in the hope that it will be useful, but WITHOUT duke@435: * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or duke@435: * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License duke@435: * version 2 for more details (a copy is included in the LICENSE file that duke@435: * accompanied this code). duke@435: * duke@435: * You should have received a copy of the GNU General Public License version duke@435: * 2 along with this work; if not, write to the Free Software Foundation, duke@435: * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. duke@435: * trims@1907: * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA trims@1907: * or visit www.oracle.com if you need additional information or have any trims@1907: * questions. duke@435: * duke@435: */ duke@435: stefank@2314: #include "precompiled.hpp" stefank@2314: #include "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/biasedLocking.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" stefank@2314: #include "runtime/osThread.hpp" stefank@2314: #include "runtime/stubRoutines.hpp" stefank@2314: #include "runtime/synchronizer.hpp" stefank@4299: #include "runtime/thread.inline.hpp" stefank@2314: #include "utilities/dtrace.hpp" stefank@2314: #include "utilities/events.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 duke@435: goetz@6453: #if defined(__GNUC__) && !defined(PPC64) duke@435: // Need to inhibit inlining for older versions of GCC to avoid build-time failures duke@435: #define ATTR __attribute__((noinline)) duke@435: #else duke@435: #define ATTR duke@435: #endif duke@435: duke@435: // The "core" versions of monitor enter and exit reside in this file. duke@435: // The interpreter and compilers contain specialized transliterated duke@435: // variants of the enter-exit fast-path operations. See i486.ad fast_lock(), duke@435: // for instance. If you make changes here, make sure to modify the duke@435: // interpreter, and both C1 and C2 fast-path inline locking code emission. duke@435: // duke@435: // duke@435: // ----------------------------------------------------------------------------- duke@435: duke@435: #ifdef DTRACE_ENABLED duke@435: duke@435: // Only bother with this argument setup if dtrace is available duke@435: // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. duke@435: coleenp@4037: #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \ duke@435: char* bytes = NULL; \ duke@435: int len = 0; \ duke@435: jlong jtid = SharedRuntime::get_java_tid(thread); \ coleenp@4037: Symbol* klassname = ((oop)(obj))->klass()->name(); \ duke@435: if (klassname != NULL) { \ duke@435: bytes = (char*)klassname->bytes(); \ duke@435: len = klassname->utf8_length(); \ duke@435: } duke@435: dcubed@3202: #ifndef USDT2 dcubed@3202: HS_DTRACE_PROBE_DECL5(hotspot, monitor__wait, dcubed@3202: jlong, uintptr_t, char*, int, long); dcubed@3202: HS_DTRACE_PROBE_DECL4(hotspot, monitor__waited, dcubed@3202: jlong, uintptr_t, char*, int); dcubed@3202: coleenp@4037: #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \ duke@435: { \ duke@435: if (DTraceMonitorProbes) { \ coleenp@4037: DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ duke@435: HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ duke@435: (monitor), bytes, len, (millis)); \ duke@435: } \ duke@435: } duke@435: coleenp@4037: #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \ duke@435: { \ duke@435: if (DTraceMonitorProbes) { \ coleenp@4037: DTRACE_MONITOR_PROBE_COMMON(obj, thread); \ duke@435: HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ duke@435: (uintptr_t)(monitor), bytes, len); \ duke@435: } \ duke@435: } duke@435: 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: (uintptr_t)(monitor), bytes, len, (millis)); \ dcubed@3202: } \ dcubed@3202: } dcubed@3202: dcubed@3202: #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_PROBE_WAITED 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_##probe(jtid, /* probe = waited */ \ dcubed@3202: (uintptr_t)(monitor), bytes, len); \ dcubed@3202: } \ dcubed@3202: } dcubed@3202: dcubed@3202: #endif /* USDT2 */ duke@435: #else // ndef DTRACE_ENABLED duke@435: coleenp@4037: #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;} coleenp@4037: #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;} duke@435: duke@435: #endif // ndef DTRACE_ENABLED duke@435: acorn@2233: // This exists only as a workaround of dtrace bug 6254741 acorn@2233: int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { acorn@2233: DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); acorn@2233: return 0; acorn@2233: } duke@435: acorn@2233: #define NINFLATIONLOCKS 256 acorn@2233: static volatile intptr_t InflationLocks [NINFLATIONLOCKS] ; acorn@2233: acorn@2233: ObjectMonitor * ObjectSynchronizer::gBlockList = NULL ; acorn@2233: ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL ; acorn@2233: ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL ; acorn@2233: int ObjectSynchronizer::gOmInUseCount = 0; acorn@2233: static volatile intptr_t ListLock = 0 ; // protects global monitor free-list cache acorn@2233: static volatile int MonitorFreeCount = 0 ; // # on gFreeList acorn@2233: static volatile int MonitorPopulation = 0 ; // # Extant -- in circulation hseigel@5784: #define CHAINMARKER (cast_to_oop(-1)) acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Fast Monitor Enter/Exit acorn@2233: // This the fast monitor enter. The interpreter and compiler use acorn@2233: // some assembly copies of this code. Make sure update those code acorn@2233: // if the following function is changed. The implementation is acorn@2233: // extremely sensitive to race condition. Be careful. acorn@2233: acorn@2233: void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) { acorn@2233: if (UseBiasedLocking) { acorn@2233: if (!SafepointSynchronize::is_at_safepoint()) { acorn@2233: BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); acorn@2233: if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { acorn@2233: return; acorn@2233: } acorn@2233: } else { acorn@2233: assert(!attempt_rebias, "can not rebias toward VM thread"); acorn@2233: BiasedLocking::revoke_at_safepoint(obj); acorn@2233: } acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } acorn@2233: acorn@2233: slow_enter (obj, lock, THREAD) ; acorn@2233: } acorn@2233: acorn@2233: void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { acorn@2233: assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here"); acorn@2233: // if displaced header is null, the previous enter is recursive enter, no-op acorn@2233: markOop dhw = lock->displaced_header(); acorn@2233: markOop mark ; acorn@2233: if (dhw == NULL) { acorn@2233: // Recursive stack-lock. acorn@2233: // Diagnostics -- Could be: stack-locked, inflating, inflated. acorn@2233: mark = object->mark() ; acorn@2233: assert (!mark->is_neutral(), "invariant") ; acorn@2233: if (mark->has_locker() && mark != markOopDesc::INFLATING()) { acorn@2233: assert(THREAD->is_lock_owned((address)mark->locker()), "invariant") ; acorn@2233: } acorn@2233: if (mark->has_monitor()) { acorn@2233: ObjectMonitor * m = mark->monitor() ; acorn@2233: assert(((oop)(m->object()))->mark() == mark, "invariant") ; acorn@2233: assert(m->is_entered(THREAD), "invariant") ; acorn@2233: } acorn@2233: return ; duke@435: } duke@435: acorn@2233: mark = object->mark() ; acorn@2233: acorn@2233: // If the object is stack-locked by the current thread, try to acorn@2233: // swing the displaced header from the box back to the mark. acorn@2233: if (mark == (markOop) lock) { acorn@2233: assert (dhw->is_neutral(), "invariant") ; acorn@2233: if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) { acorn@2233: TEVENT (fast_exit: release stacklock) ; acorn@2233: return; acorn@2233: } duke@435: } duke@435: sla@5237: ObjectSynchronizer::inflate(THREAD, object)->exit (true, THREAD) ; acorn@2233: } acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Interpreter/Compiler Slow Case acorn@2233: // This routine is used to handle interpreter/compiler slow case acorn@2233: // We don't need to use fast path here, because it must have been acorn@2233: // failed in the interpreter/compiler code. acorn@2233: void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { acorn@2233: markOop mark = obj->mark(); acorn@2233: assert(!mark->has_bias_pattern(), "should not see bias pattern here"); acorn@2233: acorn@2233: if (mark->is_neutral()) { acorn@2233: // Anticipate successful CAS -- the ST of the displaced mark must acorn@2233: // be visible <= the ST performed by the CAS. acorn@2233: lock->set_displaced_header(mark); acorn@2233: if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) { acorn@2233: TEVENT (slow_enter: release stacklock) ; acorn@2233: return ; acorn@2233: } acorn@2233: // Fall through to inflate() ... acorn@2233: } else acorn@2233: if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { acorn@2233: assert(lock != mark->locker(), "must not re-lock the same lock"); acorn@2233: assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); acorn@2233: lock->set_displaced_header(NULL); acorn@2233: return; duke@435: } duke@435: acorn@2233: #if 0 acorn@2233: // The following optimization isn't particularly useful. acorn@2233: if (mark->has_monitor() && mark->monitor()->is_entered(THREAD)) { acorn@2233: lock->set_displaced_header (NULL) ; acorn@2233: return ; acorn@2233: } acorn@2233: #endif duke@435: acorn@2233: // The object header will never be displaced to this lock, acorn@2233: // so it does not matter what the value is, except that it acorn@2233: // must be non-zero to avoid looking like a re-entrant lock, acorn@2233: // and must not look locked either. acorn@2233: lock->set_displaced_header(markOopDesc::unused_mark()); acorn@2233: ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); acorn@2233: } duke@435: acorn@2233: // This routine is used to handle interpreter/compiler slow case acorn@2233: // We don't need to use fast path here, because it must have acorn@2233: // failed in the interpreter/compiler code. Simply use the heavy acorn@2233: // weight monitor should be ok, unless someone find otherwise. acorn@2233: void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { acorn@2233: fast_exit (object, lock, THREAD) ; acorn@2233: } duke@435: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Class Loader support to workaround deadlocks on the class loader lock objects acorn@2233: // Also used by GC acorn@2233: // complete_exit()/reenter() are used to wait on a nested lock acorn@2233: // i.e. to give up an outer lock completely and then re-enter acorn@2233: // Used when holding nested locks - lock acquisition order: lock1 then lock2 acorn@2233: // 1) complete_exit lock1 - saving recursion count acorn@2233: // 2) wait on lock2 acorn@2233: // 3) when notified on lock2, unlock lock2 acorn@2233: // 4) reenter lock1 with original recursion count acorn@2233: // 5) lock lock2 acorn@2233: // NOTE: must use heavy weight monitor to handle complete_exit/reenter() acorn@2233: intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { acorn@2233: TEVENT (complete_exit) ; acorn@2233: if (UseBiasedLocking) { acorn@2233: BiasedLocking::revoke_and_rebias(obj, false, THREAD); acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } duke@435: acorn@2233: ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); duke@435: acorn@2233: return monitor->complete_exit(THREAD); acorn@2233: } acorn@2233: acorn@2233: // NOTE: must use heavy weight monitor to handle complete_exit/reenter() acorn@2233: void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { acorn@2233: TEVENT (reenter) ; acorn@2233: if (UseBiasedLocking) { acorn@2233: BiasedLocking::revoke_and_rebias(obj, false, THREAD); acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } acorn@2233: acorn@2233: ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); acorn@2233: acorn@2233: monitor->reenter(recursion, THREAD); acorn@2233: } acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // JNI locks on java objects acorn@2233: // NOTE: must use heavy weight monitor to handle jni monitor enter acorn@2233: void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { // possible entry from jni enter acorn@2233: // the current locking is from JNI instead of Java code acorn@2233: TEVENT (jni_enter) ; acorn@2233: if (UseBiasedLocking) { acorn@2233: BiasedLocking::revoke_and_rebias(obj, false, THREAD); acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } acorn@2233: THREAD->set_current_pending_monitor_is_from_java(false); acorn@2233: ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); acorn@2233: THREAD->set_current_pending_monitor_is_from_java(true); acorn@2233: } acorn@2233: acorn@2233: // NOTE: must use heavy weight monitor to handle jni monitor enter acorn@2233: bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) { acorn@2233: if (UseBiasedLocking) { acorn@2233: BiasedLocking::revoke_and_rebias(obj, false, THREAD); acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } acorn@2233: acorn@2233: ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj()); acorn@2233: return monitor->try_enter(THREAD); acorn@2233: } acorn@2233: acorn@2233: acorn@2233: // NOTE: must use heavy weight monitor to handle jni monitor exit acorn@2233: void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { acorn@2233: TEVENT (jni_exit) ; acorn@2233: if (UseBiasedLocking) { dcubed@4470: Handle h_obj(THREAD, obj); dcubed@4470: BiasedLocking::revoke_and_rebias(h_obj, false, THREAD); dcubed@4470: obj = h_obj(); acorn@2233: } acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: acorn@2233: ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj); acorn@2233: // If this thread has locked the object, exit the monitor. Note: can't use acorn@2233: // monitor->check(CHECK); must exit even if an exception is pending. acorn@2233: if (monitor->check(THREAD)) { sla@5237: monitor->exit(true, THREAD); acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Internal VM locks on java objects acorn@2233: // standard constructor, allows locking failures acorn@2233: ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { acorn@2233: _dolock = doLock; acorn@2233: _thread = thread; acorn@2233: debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) acorn@2233: _obj = obj; acorn@2233: acorn@2233: if (_dolock) { acorn@2233: TEVENT (ObjectLocker) ; acorn@2233: acorn@2233: ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: ObjectLocker::~ObjectLocker() { acorn@2233: if (_dolock) { acorn@2233: ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); acorn@2233: } acorn@2233: } acorn@2233: acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Wait/Notify/NotifyAll acorn@2233: // NOTE: must use heavy weight monitor to handle wait() acorn@2233: void ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { acorn@2233: if (UseBiasedLocking) { acorn@2233: BiasedLocking::revoke_and_rebias(obj, false, THREAD); acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } acorn@2233: if (millis < 0) { acorn@2233: TEVENT (wait - throw IAX) ; acorn@2233: THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); acorn@2233: } acorn@2233: ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); acorn@2233: DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); acorn@2233: monitor->wait(millis, true, THREAD); acorn@2233: acorn@2233: /* This dummy call is in place to get around dtrace bug 6254741. Once acorn@2233: that's fixed we can uncomment the following line and remove the call */ acorn@2233: // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); acorn@2233: dtrace_waited_probe(monitor, obj, THREAD); acorn@2233: } acorn@2233: acorn@2233: void ObjectSynchronizer::waitUninterruptibly (Handle obj, jlong millis, TRAPS) { acorn@2233: if (UseBiasedLocking) { acorn@2233: BiasedLocking::revoke_and_rebias(obj, false, THREAD); acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } acorn@2233: if (millis < 0) { acorn@2233: TEVENT (wait - throw IAX) ; acorn@2233: THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); acorn@2233: } acorn@2233: ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD) ; acorn@2233: } acorn@2233: acorn@2233: void ObjectSynchronizer::notify(Handle obj, TRAPS) { acorn@2233: if (UseBiasedLocking) { acorn@2233: BiasedLocking::revoke_and_rebias(obj, false, THREAD); acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } acorn@2233: acorn@2233: markOop mark = obj->mark(); acorn@2233: if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { acorn@2233: return; acorn@2233: } acorn@2233: ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD); acorn@2233: } acorn@2233: acorn@2233: // NOTE: see comment of notify() acorn@2233: void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { acorn@2233: if (UseBiasedLocking) { acorn@2233: BiasedLocking::revoke_and_rebias(obj, false, THREAD); acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } acorn@2233: acorn@2233: markOop mark = obj->mark(); acorn@2233: if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { acorn@2233: return; acorn@2233: } acorn@2233: ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD); acorn@2233: } acorn@2233: acorn@2233: // ----------------------------------------------------------------------------- acorn@2233: // Hash Code handling acorn@2233: // duke@435: // Performance concern: duke@435: // OrderAccess::storestore() calls release() which STs 0 into the global volatile duke@435: // OrderAccess::Dummy variable. This store is unnecessary for correctness. duke@435: // Many threads STing into a common location causes considerable cache migration duke@435: // or "sloshing" on large SMP system. As such, I avoid using OrderAccess::storestore() duke@435: // until it's repaired. In some cases OrderAccess::fence() -- which incurs local duke@435: // latency on the executing processor -- is a better choice as it scales on SMP duke@435: // systems. See http://blogs.sun.com/dave/entry/biased_locking_in_hotspot for a duke@435: // discussion of coherency costs. Note that all our current reference platforms duke@435: // provide strong ST-ST order, so the issue is moot on IA32, x64, and SPARC. duke@435: // duke@435: // As a general policy we use "volatile" to control compiler-based reordering duke@435: // and explicit fences (barriers) to control for architectural reordering performed duke@435: // by the CPU(s) or platform. duke@435: duke@435: struct SharedGlobals { duke@435: // These are highly shared mostly-read variables. duke@435: // To avoid false-sharing they need to be the sole occupants of a $ line. duke@435: double padPrefix [8]; duke@435: volatile int stwRandom ; duke@435: volatile int stwCycle ; duke@435: duke@435: // Hot RW variables -- Sequester to avoid false-sharing duke@435: double padSuffix [16]; duke@435: volatile int hcSequence ; duke@435: double padFinal [8] ; duke@435: } ; duke@435: duke@435: static SharedGlobals GVars ; acorn@1942: static int MonitorScavengeThreshold = 1000000 ; acorn@1942: static volatile int ForceMonitorScavenge = 0 ; // Scavenge required and pending duke@435: acorn@2233: static markOop ReadStableMark (oop obj) { acorn@2233: markOop mark = obj->mark() ; acorn@2233: if (!mark->is_being_inflated()) { acorn@2233: return mark ; // normal fast-path return acorn@2233: } duke@435: acorn@2233: int its = 0 ; acorn@2233: for (;;) { acorn@2233: markOop mark = obj->mark() ; acorn@2233: if (!mark->is_being_inflated()) { acorn@2233: return mark ; // normal fast-path return acorn@2233: } duke@435: acorn@2233: // The object is being inflated by some other thread. acorn@2233: // The caller of ReadStableMark() must wait for inflation to complete. acorn@2233: // Avoid live-lock acorn@2233: // TODO: consider calling SafepointSynchronize::do_call_back() while acorn@2233: // spinning to see if there's a safepoint pending. If so, immediately acorn@2233: // yielding or blocking would be appropriate. Avoid spinning while acorn@2233: // there is a safepoint pending. acorn@2233: // TODO: add inflation contention performance counters. acorn@2233: // TODO: restrict the aggregate number of spinners. duke@435: acorn@2233: ++its ; acorn@2233: if (its > 10000 || !os::is_MP()) { acorn@2233: if (its & 1) { acorn@2233: os::NakedYield() ; acorn@2233: TEVENT (Inflate: INFLATING - yield) ; acorn@2233: } else { acorn@2233: // Note that the following code attenuates the livelock problem but is not acorn@2233: // a complete remedy. A more complete solution would require that the inflating acorn@2233: // thread hold the associated inflation lock. The following code simply restricts acorn@2233: // the number of spinners to at most one. We'll have N-2 threads blocked acorn@2233: // on the inflationlock, 1 thread holding the inflation lock and using acorn@2233: // a yield/park strategy, and 1 thread in the midst of inflation. acorn@2233: // A more refined approach would be to change the encoding of INFLATING acorn@2233: // to allow encapsulation of a native thread pointer. Threads waiting for acorn@2233: // inflation to complete would use CAS to push themselves onto a singly linked acorn@2233: // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag acorn@2233: // and calling park(). When inflation was complete the thread that accomplished inflation acorn@2233: // would detach the list and set the markword to inflated with a single CAS and acorn@2233: // then for each thread on the list, set the flag and unpark() the thread. acorn@2233: // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease acorn@2233: // wakes at most one thread whereas we need to wake the entire list. hseigel@5784: int ix = (cast_from_oop(obj) >> 5) & (NINFLATIONLOCKS-1) ; acorn@2233: int YieldThenBlock = 0 ; acorn@2233: assert (ix >= 0 && ix < NINFLATIONLOCKS, "invariant") ; acorn@2233: assert ((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant") ; acorn@2233: Thread::muxAcquire (InflationLocks + ix, "InflationLock") ; acorn@2233: while (obj->mark() == markOopDesc::INFLATING()) { acorn@2233: // Beware: NakedYield() is advisory and has almost no effect on some platforms acorn@2233: // so we periodically call Self->_ParkEvent->park(1). acorn@2233: // We use a mixed spin/yield/block mechanism. acorn@2233: if ((YieldThenBlock++) >= 16) { acorn@2233: Thread::current()->_ParkEvent->park(1) ; acorn@2233: } else { acorn@2233: os::NakedYield() ; acorn@2233: } acorn@2233: } acorn@2233: Thread::muxRelease (InflationLocks + ix ) ; acorn@2233: TEVENT (Inflate: INFLATING - yield/park) ; acorn@2233: } acorn@2233: } else { acorn@2233: SpinPause() ; // SMP-polite spinning acorn@2233: } acorn@2233: } acorn@2233: } duke@435: duke@435: // hashCode() generation : duke@435: // duke@435: // Possibilities: duke@435: // * MD5Digest of {obj,stwRandom} duke@435: // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. duke@435: // * A DES- or AES-style SBox[] mechanism duke@435: // * One of the Phi-based schemes, such as: duke@435: // 2654435761 = 2^32 * Phi (golden ratio) duke@435: // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; duke@435: // * A variation of Marsaglia's shift-xor RNG scheme. duke@435: // * (obj ^ stwRandom) is appealing, but can result duke@435: // in undesirable regularity in the hashCode values of adjacent objects duke@435: // (objects allocated back-to-back, in particular). This could potentially duke@435: // result in hashtable collisions and reduced hashtable efficiency. duke@435: // There are simple ways to "diffuse" the middle address bits over the duke@435: // generated hashCode values: duke@435: // duke@435: duke@435: static inline intptr_t get_next_hash(Thread * Self, oop obj) { duke@435: intptr_t value = 0 ; duke@435: if (hashCode == 0) { duke@435: // This form uses an unguarded global Park-Miller RNG, duke@435: // so it's possible for two threads to race and generate the same RNG. duke@435: // On MP system we'll have lots of RW access to a global, so the duke@435: // mechanism induces lots of coherency traffic. duke@435: value = os::random() ; duke@435: } else duke@435: if (hashCode == 1) { duke@435: // This variation has the property of being stable (idempotent) duke@435: // between STW operations. This can be useful in some of the 1-0 duke@435: // synchronization schemes. hseigel@5784: intptr_t addrBits = cast_from_oop(obj) >> 3 ; duke@435: value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom ; duke@435: } else duke@435: if (hashCode == 2) { duke@435: value = 1 ; // for sensitivity testing duke@435: } else duke@435: if (hashCode == 3) { duke@435: value = ++GVars.hcSequence ; duke@435: } else duke@435: if (hashCode == 4) { hseigel@5784: value = cast_from_oop(obj) ; duke@435: } else { duke@435: // Marsaglia's xor-shift scheme with thread-specific state duke@435: // This is probably the best overall implementation -- we'll duke@435: // likely make this the default in future releases. duke@435: unsigned t = Self->_hashStateX ; duke@435: t ^= (t << 11) ; duke@435: Self->_hashStateX = Self->_hashStateY ; duke@435: Self->_hashStateY = Self->_hashStateZ ; duke@435: Self->_hashStateZ = Self->_hashStateW ; duke@435: unsigned v = Self->_hashStateW ; duke@435: v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)) ; duke@435: Self->_hashStateW = v ; duke@435: value = v ; duke@435: } duke@435: duke@435: value &= markOopDesc::hash_mask; duke@435: if (value == 0) value = 0xBAD ; duke@435: assert (value != markOopDesc::no_hash, "invariant") ; duke@435: TEVENT (hashCode: GENERATE) ; duke@435: return value; duke@435: } acorn@2233: // acorn@2233: intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) { acorn@2233: if (UseBiasedLocking) { acorn@2233: // NOTE: many places throughout the JVM do not expect a safepoint acorn@2233: // to be taken here, in particular most operations on perm gen acorn@2233: // objects. However, we only ever bias Java instances and all of acorn@2233: // the call sites of identity_hash that might revoke biases have acorn@2233: // been checked to make sure they can handle a safepoint. The acorn@2233: // added check of the bias pattern is to avoid useless calls to acorn@2233: // thread-local storage. acorn@2233: if (obj->mark()->has_bias_pattern()) { acorn@2233: // Box and unbox the raw reference just in case we cause a STW safepoint. acorn@2233: Handle hobj (Self, obj) ; acorn@2233: // Relaxing assertion for bug 6320749. acorn@2233: assert (Universe::verify_in_progress() || acorn@2233: !SafepointSynchronize::is_at_safepoint(), acorn@2233: "biases should not be seen by VM thread here"); acorn@2233: BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); acorn@2233: obj = hobj() ; acorn@2233: assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } acorn@2233: } duke@435: acorn@2233: // hashCode() is a heap mutator ... acorn@2233: // Relaxing assertion for bug 6320749. acorn@2233: assert (Universe::verify_in_progress() || acorn@2233: !SafepointSynchronize::is_at_safepoint(), "invariant") ; acorn@2233: assert (Universe::verify_in_progress() || acorn@2233: Self->is_Java_thread() , "invariant") ; acorn@2233: assert (Universe::verify_in_progress() || acorn@2233: ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; acorn@2233: acorn@2233: ObjectMonitor* monitor = NULL; acorn@2233: markOop temp, test; acorn@2233: intptr_t hash; acorn@2233: markOop mark = ReadStableMark (obj); acorn@2233: acorn@2233: // object should remain ineligible for biased locking acorn@2233: assert (!mark->has_bias_pattern(), "invariant") ; acorn@2233: acorn@2233: if (mark->is_neutral()) { acorn@2233: hash = mark->hash(); // this is a normal header acorn@2233: if (hash) { // if it has hash, just return it acorn@2233: return hash; acorn@2233: } acorn@2233: hash = get_next_hash(Self, obj); // allocate a new hash code acorn@2233: temp = mark->copy_set_hash(hash); // merge the hash code into header acorn@2233: // use (machine word version) atomic operation to install the hash acorn@2233: test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); acorn@2233: if (test == mark) { acorn@2233: return hash; acorn@2233: } acorn@2233: // If atomic operation failed, we must inflate the header acorn@2233: // into heavy weight monitor. We could add more code here acorn@2233: // for fast path, but it does not worth the complexity. acorn@2233: } else if (mark->has_monitor()) { acorn@2233: monitor = mark->monitor(); acorn@2233: temp = monitor->header(); acorn@2233: assert (temp->is_neutral(), "invariant") ; acorn@2233: hash = temp->hash(); acorn@2233: if (hash) { acorn@2233: return hash; acorn@2233: } acorn@2233: // Skip to the following code to reduce code size acorn@2233: } else if (Self->is_lock_owned((address)mark->locker())) { acorn@2233: temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned acorn@2233: assert (temp->is_neutral(), "invariant") ; acorn@2233: hash = temp->hash(); // by current thread, check if the displaced acorn@2233: if (hash) { // header contains hash code acorn@2233: return hash; acorn@2233: } acorn@2233: // WARNING: acorn@2233: // The displaced header is strictly immutable. acorn@2233: // It can NOT be changed in ANY cases. So we have acorn@2233: // to inflate the header into heavyweight monitor acorn@2233: // even the current thread owns the lock. The reason acorn@2233: // is the BasicLock (stack slot) will be asynchronously acorn@2233: // read by other threads during the inflate() function. acorn@2233: // Any change to stack may not propagate to other threads acorn@2233: // correctly. acorn@2233: } acorn@2233: acorn@2233: // Inflate the monitor to set hash code acorn@2233: monitor = ObjectSynchronizer::inflate(Self, obj); acorn@2233: // Load displaced header and check it has hash code acorn@2233: mark = monitor->header(); acorn@2233: assert (mark->is_neutral(), "invariant") ; acorn@2233: hash = mark->hash(); acorn@2233: if (hash == 0) { acorn@2233: hash = get_next_hash(Self, obj); acorn@2233: temp = mark->copy_set_hash(hash); // merge hash code into header acorn@2233: assert (temp->is_neutral(), "invariant") ; acorn@2233: test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); acorn@2233: if (test != mark) { acorn@2233: // The only update to the header in the monitor (outside GC) acorn@2233: // is install the hash code. If someone add new usage of acorn@2233: // displaced header, please update this code acorn@2233: hash = test->hash(); acorn@2233: assert (test->is_neutral(), "invariant") ; acorn@2233: assert (hash != 0, "Trivial unexpected object/monitor header usage."); acorn@2233: } acorn@2233: } acorn@2233: // We finally get the hash acorn@2233: return hash; duke@435: } duke@435: acorn@2233: // Deprecated -- use FastHashCode() instead. duke@435: acorn@2233: intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { acorn@2233: return FastHashCode (Thread::current(), obj()) ; duke@435: } duke@435: duke@435: acorn@2233: bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, acorn@2233: Handle h_obj) { acorn@2233: if (UseBiasedLocking) { acorn@2233: BiasedLocking::revoke_and_rebias(h_obj, false, thread); acorn@2233: assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } duke@435: acorn@2233: assert(thread == JavaThread::current(), "Can only be called on current thread"); acorn@2233: oop obj = h_obj(); duke@435: acorn@2233: markOop mark = ReadStableMark (obj) ; acorn@2233: acorn@2233: // Uncontended case, header points to stack acorn@2233: if (mark->has_locker()) { acorn@2233: return thread->is_lock_owned((address)mark->locker()); acorn@2233: } acorn@2233: // Contended case, header points to ObjectMonitor (tagged pointer) acorn@2233: if (mark->has_monitor()) { acorn@2233: ObjectMonitor* monitor = mark->monitor(); acorn@2233: return monitor->is_entered(thread) != 0 ; acorn@2233: } acorn@2233: // Unlocked case, header in place acorn@2233: assert(mark->is_neutral(), "sanity check"); acorn@2233: return false; acorn@2233: } acorn@2233: acorn@2233: // Be aware of this method could revoke bias of the lock object. acorn@2233: // This method querys the ownership of the lock handle specified by 'h_obj'. acorn@2233: // If the current thread owns the lock, it returns owner_self. If no acorn@2233: // thread owns the lock, it returns owner_none. Otherwise, it will return acorn@2233: // ower_other. acorn@2233: ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership acorn@2233: (JavaThread *self, Handle h_obj) { acorn@2233: // The caller must beware this method can revoke bias, and acorn@2233: // revocation can result in a safepoint. acorn@2233: assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; acorn@2233: assert (self->thread_state() != _thread_blocked , "invariant") ; acorn@2233: acorn@2233: // Possible mark states: neutral, biased, stack-locked, inflated acorn@2233: acorn@2233: if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { acorn@2233: // CASE: biased acorn@2233: BiasedLocking::revoke_and_rebias(h_obj, false, self); acorn@2233: assert(!h_obj->mark()->has_bias_pattern(), acorn@2233: "biases should be revoked by now"); acorn@2233: } acorn@2233: acorn@2233: assert(self == JavaThread::current(), "Can only be called on current thread"); acorn@2233: oop obj = h_obj(); acorn@2233: markOop mark = ReadStableMark (obj) ; acorn@2233: acorn@2233: // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. acorn@2233: if (mark->has_locker()) { acorn@2233: return self->is_lock_owned((address)mark->locker()) ? acorn@2233: owner_self : owner_other; acorn@2233: } acorn@2233: acorn@2233: // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. acorn@2233: // The Object:ObjectMonitor relationship is stable as long as we're acorn@2233: // not at a safepoint. acorn@2233: if (mark->has_monitor()) { acorn@2233: void * owner = mark->monitor()->_owner ; acorn@2233: if (owner == NULL) return owner_none ; acorn@2233: return (owner == self || acorn@2233: self->is_lock_owned((address)owner)) ? owner_self : owner_other; acorn@2233: } acorn@2233: acorn@2233: // CASE: neutral acorn@2233: assert(mark->is_neutral(), "sanity check"); acorn@2233: return owner_none ; // it's unlocked acorn@2233: } acorn@2233: acorn@2233: // FIXME: jvmti should call this acorn@2233: JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { acorn@2233: if (UseBiasedLocking) { acorn@2233: if (SafepointSynchronize::is_at_safepoint()) { acorn@2233: BiasedLocking::revoke_at_safepoint(h_obj); acorn@2233: } else { acorn@2233: BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); acorn@2233: } acorn@2233: assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); acorn@2233: } acorn@2233: acorn@2233: oop obj = h_obj(); acorn@2233: address owner = NULL; acorn@2233: acorn@2233: markOop mark = ReadStableMark (obj) ; acorn@2233: acorn@2233: // Uncontended case, header points to stack acorn@2233: if (mark->has_locker()) { acorn@2233: owner = (address) mark->locker(); acorn@2233: } acorn@2233: acorn@2233: // Contended case, header points to ObjectMonitor (tagged pointer) acorn@2233: if (mark->has_monitor()) { acorn@2233: ObjectMonitor* monitor = mark->monitor(); acorn@2233: assert(monitor != NULL, "monitor should be non-null"); acorn@2233: owner = (address) monitor->owner(); acorn@2233: } acorn@2233: acorn@2233: if (owner != NULL) { dcubed@4673: // owning_thread_from_monitor_owner() may also return NULL here acorn@2233: return Threads::owning_thread_from_monitor_owner(owner, doLock); acorn@2233: } acorn@2233: acorn@2233: // Unlocked case, header in place acorn@2233: // Cannot have assertion since this object may have been acorn@2233: // locked by another thread when reaching here. acorn@2233: // assert(mark->is_neutral(), "sanity check"); acorn@2233: acorn@2233: return NULL; acorn@2233: } acorn@2233: // Visitors ... acorn@2233: acorn@2233: void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { acorn@2233: ObjectMonitor* block = gBlockList; acorn@2233: ObjectMonitor* mid; acorn@2233: while (block) { acorn@2233: assert(block->object() == CHAINMARKER, "must be a block header"); acorn@2233: for (int i = _BLOCKSIZE - 1; i > 0; i--) { acorn@2233: mid = block + i; acorn@2233: oop object = (oop) mid->object(); acorn@2233: if (object != NULL) { acorn@2233: closure->do_monitor(mid); acorn@2233: } acorn@2233: } acorn@2233: block = (ObjectMonitor*) block->FreeNext; duke@435: } duke@435: } duke@435: acorn@2233: // Get the next block in the block list. acorn@2233: static inline ObjectMonitor* next(ObjectMonitor* block) { acorn@2233: assert(block->object() == CHAINMARKER, "must be a block header"); acorn@2233: block = block->FreeNext ; acorn@2233: assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); acorn@2233: return block; duke@435: } duke@435: duke@435: acorn@2233: void ObjectSynchronizer::oops_do(OopClosure* f) { acorn@2233: assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); acorn@2233: for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { acorn@2233: assert(block->object() == CHAINMARKER, "must be a block header"); acorn@2233: for (int i = 1; i < _BLOCKSIZE; i++) { acorn@2233: ObjectMonitor* mid = &block[i]; acorn@2233: if (mid->object() != NULL) { acorn@2233: f->do_oop((oop*)mid->object_addr()); duke@435: } duke@435: } duke@435: } duke@435: } duke@435: duke@435: acorn@2233: // ----------------------------------------------------------------------------- duke@435: // ObjectMonitor Lifecycle duke@435: // ----------------------- duke@435: // Inflation unlinks monitors from the global gFreeList and duke@435: // associates them with objects. Deflation -- which occurs at duke@435: // STW-time -- disassociates idle monitors from objects. Such duke@435: // scavenged monitors are returned to the gFreeList. duke@435: // duke@435: // The global list is protected by ListLock. All the critical sections duke@435: // are short and operate in constant-time. duke@435: // duke@435: // ObjectMonitors reside in type-stable memory (TSM) and are immortal. duke@435: // duke@435: // Lifecycle: duke@435: // -- unassigned and on the global free list duke@435: // -- unassigned and on a thread's private omFreeList duke@435: // -- assigned to an object. The object is inflated and the mark refers duke@435: // to the objectmonitor. duke@435: // duke@435: duke@435: acorn@1942: // Constraining monitor pool growth via MonitorBound ... acorn@1942: // acorn@1942: // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the acorn@1942: // the rate of scavenging is driven primarily by GC. As such, we can find acorn@1942: // an inordinate number of monitors in circulation. acorn@1942: // To avoid that scenario we can artificially induce a STW safepoint acorn@1942: // if the pool appears to be growing past some reasonable bound. acorn@1942: // Generally we favor time in space-time tradeoffs, but as there's no acorn@1942: // natural back-pressure on the # of extant monitors we need to impose some acorn@1942: // type of limit. Beware that if MonitorBound is set to too low a value acorn@1942: // we could just loop. In addition, if MonitorBound is set to a low value acorn@1942: // we'll incur more safepoints, which are harmful to performance. acorn@1942: // See also: GuaranteedSafepointInterval acorn@1942: // acorn@1942: // The current implementation uses asynchronous VM operations. acorn@1942: // acorn@1942: acorn@1942: static void InduceScavenge (Thread * Self, const char * Whence) { acorn@1942: // Induce STW safepoint to trim monitors acorn@1942: // Ultimately, this results in a call to deflate_idle_monitors() in the near future. acorn@1942: // More precisely, trigger an asynchronous STW safepoint as the number acorn@1942: // of active monitors passes the specified threshold. acorn@1942: // TODO: assert thread state is reasonable acorn@1942: acorn@1942: if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { acorn@2233: if (ObjectMonitor::Knob_Verbose) { acorn@1942: ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ; acorn@1942: ::fflush(stdout) ; acorn@1942: } acorn@1942: // Induce a 'null' safepoint to scavenge monitors acorn@1942: // Must VM_Operation instance be heap allocated as the op will be enqueue and posted acorn@1942: // to the VMthread and have a lifespan longer than that of this activation record. acorn@1942: // The VMThread will delete the op when completed. acorn@1942: VMThread::execute (new VM_ForceAsyncSafepoint()) ; acorn@1942: acorn@2233: if (ObjectMonitor::Knob_Verbose) { acorn@1942: ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ; acorn@1942: ::fflush(stdout) ; acorn@1942: } acorn@1942: } acorn@1942: } acorn@1995: /* Too slow for general assert or debug acorn@1995: void ObjectSynchronizer::verifyInUse (Thread *Self) { acorn@1995: ObjectMonitor* mid; acorn@1995: int inusetally = 0; acorn@1995: for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { acorn@1995: inusetally ++; acorn@1995: } acorn@1995: assert(inusetally == Self->omInUseCount, "inuse count off"); acorn@1995: acorn@1995: int freetally = 0; acorn@1995: for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { acorn@1995: freetally ++; acorn@1995: } acorn@1995: assert(freetally == Self->omFreeCount, "free count off"); acorn@1995: } acorn@1995: */ duke@435: ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) { duke@435: // A large MAXPRIVATE value reduces both list lock contention duke@435: // and list coherency traffic, but also tends to increase the duke@435: // number of objectMonitors in circulation as well as the STW duke@435: // scavenge costs. As usual, we lean toward time in space-time duke@435: // tradeoffs. duke@435: const int MAXPRIVATE = 1024 ; duke@435: for (;;) { duke@435: ObjectMonitor * m ; duke@435: duke@435: // 1: try to allocate from the thread's local omFreeList. duke@435: // Threads will attempt to allocate first from their local list, then duke@435: // from the global list, and only after those attempts fail will the thread duke@435: // attempt to instantiate new monitors. Thread-local free lists take duke@435: // heat off the ListLock and improve allocation latency, as well as reducing duke@435: // coherency traffic on the shared global list. duke@435: m = Self->omFreeList ; duke@435: if (m != NULL) { duke@435: Self->omFreeList = m->FreeNext ; duke@435: Self->omFreeCount -- ; duke@435: // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene duke@435: guarantee (m->object() == NULL, "invariant") ; acorn@1942: if (MonitorInUseLists) { acorn@1942: m->FreeNext = Self->omInUseList; acorn@1942: Self->omInUseList = m; acorn@1942: Self->omInUseCount ++; acorn@1995: // verifyInUse(Self); acorn@1995: } else { acorn@1995: m->FreeNext = NULL; acorn@1942: } duke@435: return m ; duke@435: } duke@435: duke@435: // 2: try to allocate from the global gFreeList duke@435: // CONSIDER: use muxTry() instead of muxAcquire(). duke@435: // If the muxTry() fails then drop immediately into case 3. duke@435: // If we're using thread-local free lists then try duke@435: // to reprovision the caller's free list. duke@435: if (gFreeList != NULL) { duke@435: // Reprovision the thread's omFreeList. duke@435: // Use bulk transfers to reduce the allocation rate and heat duke@435: // on various locks. duke@435: Thread::muxAcquire (&ListLock, "omAlloc") ; duke@435: for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL; ) { acorn@1942: MonitorFreeCount --; duke@435: ObjectMonitor * take = gFreeList ; duke@435: gFreeList = take->FreeNext ; duke@435: guarantee (take->object() == NULL, "invariant") ; duke@435: guarantee (!take->is_busy(), "invariant") ; duke@435: take->Recycle() ; acorn@1995: omRelease (Self, take, false) ; duke@435: } duke@435: Thread::muxRelease (&ListLock) ; duke@435: Self->omFreeProvision += 1 + (Self->omFreeProvision/2) ; duke@435: if (Self->omFreeProvision > MAXPRIVATE ) Self->omFreeProvision = MAXPRIVATE ; duke@435: TEVENT (omFirst - reprovision) ; acorn@1942: acorn@1942: const int mx = MonitorBound ; acorn@1942: if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) { acorn@1942: // We can't safely induce a STW safepoint from omAlloc() as our thread acorn@1942: // state may not be appropriate for such activities and callers may hold acorn@1942: // naked oops, so instead we defer the action. acorn@1942: InduceScavenge (Self, "omAlloc") ; acorn@1942: } acorn@1942: continue; duke@435: } duke@435: duke@435: // 3: allocate a block of new ObjectMonitors duke@435: // Both the local and global free lists are empty -- resort to malloc(). duke@435: // In the current implementation objectMonitors are TSM - immortal. duke@435: assert (_BLOCKSIZE > 1, "invariant") ; dcubed@4967: ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE]; duke@435: duke@435: // NOTE: (almost) no way to recover if allocation failed. duke@435: // We might be able to induce a STW safepoint and scavenge enough duke@435: // objectMonitors to permit progress. duke@435: if (temp == NULL) { ccheung@4993: vm_exit_out_of_memory (sizeof (ObjectMonitor[_BLOCKSIZE]), OOM_MALLOC_ERROR, ccheung@4993: "Allocate ObjectMonitors"); duke@435: } duke@435: duke@435: // Format the block. duke@435: // initialize the linked list, each monitor points to its next duke@435: // forming the single linked free list, the very first monitor duke@435: // will points to next block, which forms the block list. duke@435: // The trick of using the 1st element in the block as gBlockList duke@435: // linkage should be reconsidered. A better implementation would duke@435: // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } duke@435: duke@435: for (int i = 1; i < _BLOCKSIZE ; i++) { duke@435: temp[i].FreeNext = &temp[i+1]; duke@435: } duke@435: duke@435: // terminate the last monitor as the end of list duke@435: temp[_BLOCKSIZE - 1].FreeNext = NULL ; duke@435: duke@435: // Element [0] is reserved for global list linkage duke@435: temp[0].set_object(CHAINMARKER); duke@435: duke@435: // Consider carving out this thread's current request from the duke@435: // block in hand. This avoids some lock traffic and redundant duke@435: // list activity. duke@435: duke@435: // Acquire the ListLock to manipulate BlockList and FreeList. duke@435: // An Oyama-Taura-Yonezawa scheme might be more efficient. duke@435: Thread::muxAcquire (&ListLock, "omAlloc [2]") ; acorn@1942: MonitorPopulation += _BLOCKSIZE-1; acorn@1942: MonitorFreeCount += _BLOCKSIZE-1; duke@435: duke@435: // Add the new block to the list of extant blocks (gBlockList). duke@435: // The very first objectMonitor in a block is reserved and dedicated. duke@435: // It serves as blocklist "next" linkage. duke@435: temp[0].FreeNext = gBlockList; duke@435: gBlockList = temp; duke@435: duke@435: // Add the new string of objectMonitors to the global free list duke@435: temp[_BLOCKSIZE - 1].FreeNext = gFreeList ; duke@435: gFreeList = temp + 1; duke@435: Thread::muxRelease (&ListLock) ; duke@435: TEVENT (Allocate block of monitors) ; duke@435: } duke@435: } duke@435: duke@435: // Place "m" on the caller's private per-thread omFreeList. duke@435: // In practice there's no need to clamp or limit the number of duke@435: // monitors on a thread's omFreeList as the only time we'll call duke@435: // omRelease is to return a monitor to the free list after a CAS duke@435: // attempt failed. This doesn't allow unbounded #s of monitors to duke@435: // accumulate on a thread's free list. duke@435: // duke@435: acorn@1995: void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m, bool fromPerThreadAlloc) { duke@435: guarantee (m->object() == NULL, "invariant") ; acorn@1995: acorn@1995: // Remove from omInUseList acorn@1995: if (MonitorInUseLists && fromPerThreadAlloc) { acorn@1995: ObjectMonitor* curmidinuse = NULL; acorn@1995: for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; ) { acorn@1995: if (m == mid) { acorn@1995: // extract from per-thread in-use-list acorn@1995: if (mid == Self->omInUseList) { acorn@1995: Self->omInUseList = mid->FreeNext; acorn@1995: } else if (curmidinuse != NULL) { acorn@1995: curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist acorn@1995: } acorn@1995: Self->omInUseCount --; acorn@1995: // verifyInUse(Self); acorn@1995: break; acorn@1995: } else { acorn@1995: curmidinuse = mid; acorn@1995: mid = mid->FreeNext; acorn@1995: } acorn@1995: } acorn@1995: } acorn@1995: acorn@1995: // FreeNext is used for both onInUseList and omFreeList, so clear old before setting new acorn@1995: m->FreeNext = Self->omFreeList ; acorn@1995: Self->omFreeList = m ; acorn@1995: Self->omFreeCount ++ ; duke@435: } duke@435: duke@435: // Return the monitors of a moribund thread's local free list to duke@435: // the global free list. Typically a thread calls omFlush() when duke@435: // it's dying. We could also consider having the VM thread steal duke@435: // monitors from threads that have not run java code over a few duke@435: // consecutive STW safepoints. Relatedly, we might decay duke@435: // omFreeProvision at STW safepoints. duke@435: // acorn@1995: // Also return the monitors of a moribund thread"s omInUseList to acorn@1995: // a global gOmInUseList under the global list lock so these acorn@1995: // will continue to be scanned. acorn@1995: // duke@435: // We currently call omFlush() from the Thread:: dtor _after the thread duke@435: // has been excised from the thread list and is no longer a mutator. duke@435: // That means that omFlush() can run concurrently with a safepoint and duke@435: // the scavenge operator. Calling omFlush() from JavaThread::exit() might duke@435: // be a better choice as we could safely reason that that the JVM is duke@435: // not at a safepoint at the time of the call, and thus there could duke@435: // be not inopportune interleavings between omFlush() and the scavenge duke@435: // operator. duke@435: duke@435: void ObjectSynchronizer::omFlush (Thread * Self) { duke@435: ObjectMonitor * List = Self->omFreeList ; // Null-terminated SLL duke@435: Self->omFreeList = NULL ; duke@435: ObjectMonitor * Tail = NULL ; acorn@1942: int Tally = 0; acorn@1995: if (List != NULL) { acorn@1995: ObjectMonitor * s ; acorn@1995: for (s = List ; s != NULL ; s = s->FreeNext) { acorn@1995: Tally ++ ; acorn@1995: Tail = s ; acorn@1995: guarantee (s->object() == NULL, "invariant") ; acorn@1995: guarantee (!s->is_busy(), "invariant") ; acorn@1995: s->set_owner (NULL) ; // redundant but good hygiene acorn@1995: TEVENT (omFlush - Move one) ; acorn@1995: } acorn@1995: guarantee (Tail != NULL && List != NULL, "invariant") ; duke@435: } duke@435: acorn@1995: ObjectMonitor * InUseList = Self->omInUseList; acorn@1995: ObjectMonitor * InUseTail = NULL ; acorn@1995: int InUseTally = 0; acorn@1995: if (InUseList != NULL) { acorn@1995: Self->omInUseList = NULL; acorn@1995: ObjectMonitor *curom; acorn@1995: for (curom = InUseList; curom != NULL; curom = curom->FreeNext) { acorn@1995: InUseTail = curom; acorn@1995: InUseTally++; acorn@1995: } acorn@1995: // TODO debug acorn@1995: assert(Self->omInUseCount == InUseTally, "inuse count off"); acorn@1995: Self->omInUseCount = 0; acorn@1995: guarantee (InUseTail != NULL && InUseList != NULL, "invariant"); acorn@1995: } acorn@1995: duke@435: Thread::muxAcquire (&ListLock, "omFlush") ; acorn@1995: if (Tail != NULL) { acorn@1995: Tail->FreeNext = gFreeList ; acorn@1995: gFreeList = List ; acorn@1995: MonitorFreeCount += Tally; acorn@1995: } acorn@1995: acorn@1995: if (InUseTail != NULL) { acorn@1995: InUseTail->FreeNext = gOmInUseList; acorn@1995: gOmInUseList = InUseList; acorn@1995: gOmInUseCount += InUseTally; acorn@1995: } acorn@1995: duke@435: Thread::muxRelease (&ListLock) ; duke@435: TEVENT (omFlush) ; duke@435: } duke@435: duke@435: // Fast path code shared by multiple functions duke@435: ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { duke@435: markOop mark = obj->mark(); duke@435: if (mark->has_monitor()) { duke@435: assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); duke@435: assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); duke@435: return mark->monitor(); duke@435: } duke@435: return ObjectSynchronizer::inflate(Thread::current(), obj); duke@435: } duke@435: acorn@2233: duke@435: // Note that we could encounter some performance loss through false-sharing as duke@435: // multiple locks occupy the same $ line. Padding might be appropriate. duke@435: duke@435: duke@435: ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) { duke@435: // Inflate mutates the heap ... duke@435: // Relaxing assertion for bug 6320749. duke@435: assert (Universe::verify_in_progress() || duke@435: !SafepointSynchronize::is_at_safepoint(), "invariant") ; duke@435: duke@435: for (;;) { duke@435: const markOop mark = object->mark() ; duke@435: assert (!mark->has_bias_pattern(), "invariant") ; duke@435: duke@435: // The mark can be in one of the following states: duke@435: // * Inflated - just return duke@435: // * Stack-locked - coerce it to inflated duke@435: // * INFLATING - busy wait for conversion to complete duke@435: // * Neutral - aggressively inflate the object. duke@435: // * BIASED - Illegal. We should never see this duke@435: duke@435: // CASE: inflated duke@435: if (mark->has_monitor()) { duke@435: ObjectMonitor * inf = mark->monitor() ; duke@435: assert (inf->header()->is_neutral(), "invariant"); duke@435: assert (inf->object() == object, "invariant") ; duke@435: assert (ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid"); duke@435: return inf ; duke@435: } duke@435: duke@435: // CASE: inflation in progress - inflating over a stack-lock. duke@435: // Some other thread is converting from stack-locked to inflated. duke@435: // Only that thread can complete inflation -- other threads must wait. duke@435: // The INFLATING value is transient. duke@435: // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish. duke@435: // We could always eliminate polling by parking the thread on some auxiliary list. duke@435: if (mark == markOopDesc::INFLATING()) { duke@435: TEVENT (Inflate: spin while INFLATING) ; duke@435: ReadStableMark(object) ; duke@435: continue ; duke@435: } duke@435: duke@435: // CASE: stack-locked duke@435: // Could be stack-locked either by this thread or by some other thread. duke@435: // duke@435: // Note that we allocate the objectmonitor speculatively, _before_ attempting duke@435: // to install INFLATING into the mark word. We originally installed INFLATING, duke@435: // allocated the objectmonitor, and then finally STed the address of the duke@435: // objectmonitor into the mark. This was correct, but artificially lengthened duke@435: // the interval in which INFLATED appeared in the mark, thus increasing duke@435: // the odds of inflation contention. duke@435: // duke@435: // We now use per-thread private objectmonitor free lists. duke@435: // These list are reprovisioned from the global free list outside the duke@435: // critical INFLATING...ST interval. A thread can transfer duke@435: // multiple objectmonitors en-mass from the global free list to its local free list. duke@435: // This reduces coherency traffic and lock contention on the global free list. duke@435: // Using such local free lists, it doesn't matter if the omAlloc() call appears duke@435: // before or after the CAS(INFLATING) operation. duke@435: // See the comments in omAlloc(). duke@435: duke@435: if (mark->has_locker()) { duke@435: ObjectMonitor * m = omAlloc (Self) ; duke@435: // Optimistically prepare the objectmonitor - anticipate successful CAS duke@435: // We do this before the CAS in order to minimize the length of time duke@435: // in which INFLATING appears in the mark. duke@435: m->Recycle(); duke@435: m->_Responsible = NULL ; duke@435: m->OwnerIsThread = 0 ; duke@435: m->_recursions = 0 ; acorn@2233: m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // Consider: maintain by type/class duke@435: duke@435: markOop cmp = (markOop) Atomic::cmpxchg_ptr (markOopDesc::INFLATING(), object->mark_addr(), mark) ; duke@435: if (cmp != mark) { acorn@1995: omRelease (Self, m, true) ; duke@435: continue ; // Interference -- just retry duke@435: } duke@435: duke@435: // We've successfully installed INFLATING (0) into the mark-word. duke@435: // This is the only case where 0 will appear in a mark-work. duke@435: // Only the singular thread that successfully swings the mark-word duke@435: // to 0 can perform (or more precisely, complete) inflation. duke@435: // duke@435: // Why do we CAS a 0 into the mark-word instead of just CASing the duke@435: // mark-word from the stack-locked value directly to the new inflated state? duke@435: // Consider what happens when a thread unlocks a stack-locked object. duke@435: // It attempts to use CAS to swing the displaced header value from the duke@435: // on-stack basiclock back into the object header. Recall also that the duke@435: // header value (hashcode, etc) can reside in (a) the object header, or duke@435: // (b) a displaced header associated with the stack-lock, or (c) a displaced duke@435: // header in an objectMonitor. The inflate() routine must copy the header duke@435: // value from the basiclock on the owner's stack to the objectMonitor, all duke@435: // the while preserving the hashCode stability invariants. If the owner duke@435: // decides to release the lock while the value is 0, the unlock will fail duke@435: // and control will eventually pass from slow_exit() to inflate. The owner duke@435: // will then spin, waiting for the 0 value to disappear. Put another way, duke@435: // the 0 causes the owner to stall if the owner happens to try to duke@435: // drop the lock (restoring the header from the basiclock to the object) duke@435: // while inflation is in-progress. This protocol avoids races that might duke@435: // would otherwise permit hashCode values to change or "flicker" for an object. duke@435: // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable. duke@435: // 0 serves as a "BUSY" inflate-in-progress indicator. duke@435: duke@435: duke@435: // fetch the displaced mark from the owner's stack. duke@435: // The owner can't die or unwind past the lock while our INFLATING duke@435: // object is in the mark. Furthermore the owner can't complete duke@435: // an unlock on the object, either. duke@435: markOop dmw = mark->displaced_mark_helper() ; duke@435: assert (dmw->is_neutral(), "invariant") ; duke@435: duke@435: // Setup monitor fields to proper values -- prepare the monitor duke@435: m->set_header(dmw) ; duke@435: duke@435: // Optimization: if the mark->locker stack address is associated duke@435: // with this thread we could simply set m->_owner = Self and xlu@1137: // m->OwnerIsThread = 1. Note that a thread can inflate an object duke@435: // that it has stack-locked -- as might happen in wait() -- directly duke@435: // with CAS. That is, we can avoid the xchg-NULL .... ST idiom. xlu@1137: m->set_owner(mark->locker()); duke@435: m->set_object(object); duke@435: // TODO-FIXME: assert BasicLock->dhw != 0. duke@435: duke@435: // Must preserve store ordering. The monitor state must duke@435: // be stable at the time of publishing the monitor address. duke@435: guarantee (object->mark() == markOopDesc::INFLATING(), "invariant") ; duke@435: object->release_set_mark(markOopDesc::encode(m)); duke@435: duke@435: // Hopefully the performance counters are allocated on distinct cache lines duke@435: // to avoid false sharing on MP systems ... acorn@2233: if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ; duke@435: TEVENT(Inflate: overwrite stacklock) ; duke@435: if (TraceMonitorInflation) { duke@435: if (object->is_instance()) { duke@435: ResourceMark rm; duke@435: tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", hseigel@5784: (void *) object, (intptr_t) object->mark(), hseigel@4278: object->klass()->external_name()); duke@435: } duke@435: } duke@435: return m ; duke@435: } duke@435: duke@435: // CASE: neutral duke@435: // TODO-FIXME: for entry we currently inflate and then try to CAS _owner. duke@435: // If we know we're inflating for entry it's better to inflate by swinging a duke@435: // pre-locked objectMonitor pointer into the object header. A successful duke@435: // CAS inflates the object *and* confers ownership to the inflating thread. duke@435: // In the current implementation we use a 2-step mechanism where we CAS() duke@435: // to inflate and then CAS() again to try to swing _owner from NULL to Self. duke@435: // An inflateTry() method that we could call from fast_enter() and slow_enter() duke@435: // would be useful. duke@435: duke@435: assert (mark->is_neutral(), "invariant"); duke@435: ObjectMonitor * m = omAlloc (Self) ; duke@435: // prepare m for installation - set monitor to initial state duke@435: m->Recycle(); duke@435: m->set_header(mark); duke@435: m->set_owner(NULL); duke@435: m->set_object(object); duke@435: m->OwnerIsThread = 1 ; duke@435: m->_recursions = 0 ; duke@435: m->_Responsible = NULL ; acorn@2233: m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // consider: keep metastats by type/class duke@435: duke@435: if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) { duke@435: m->set_object (NULL) ; duke@435: m->set_owner (NULL) ; duke@435: m->OwnerIsThread = 0 ; duke@435: m->Recycle() ; acorn@1995: omRelease (Self, m, true) ; duke@435: m = NULL ; duke@435: continue ; duke@435: // interference - the markword changed - just retry. duke@435: // The state-transitions are one-way, so there's no chance of duke@435: // live-lock -- "Inflated" is an absorbing state. duke@435: } duke@435: duke@435: // Hopefully the performance counters are allocated on distinct duke@435: // cache lines to avoid false sharing on MP systems ... acorn@2233: if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ; duke@435: TEVENT(Inflate: overwrite neutral) ; duke@435: if (TraceMonitorInflation) { duke@435: if (object->is_instance()) { duke@435: ResourceMark rm; duke@435: tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", hseigel@5784: (void *) object, (intptr_t) object->mark(), hseigel@4278: object->klass()->external_name()); duke@435: } duke@435: } duke@435: return m ; duke@435: } duke@435: } duke@435: acorn@2233: // Note that we could encounter some performance loss through false-sharing as acorn@2233: // multiple locks occupy the same $ line. Padding might be appropriate. duke@435: duke@435: duke@435: // Deflate_idle_monitors() is called at all safepoints, immediately duke@435: // after all mutators are stopped, but before any objects have moved. duke@435: // It traverses the list of known monitors, deflating where possible. duke@435: // The scavenged monitor are returned to the monitor free list. duke@435: // duke@435: // Beware that we scavenge at *every* stop-the-world point. duke@435: // Having a large number of monitors in-circulation negatively duke@435: // impacts the performance of some applications (e.g., PointBase). duke@435: // Broadly, we want to minimize the # of monitors in circulation. acorn@1942: // acorn@1942: // We have added a flag, MonitorInUseLists, which creates a list acorn@1942: // of active monitors for each thread. deflate_idle_monitors() acorn@1942: // only scans the per-thread inuse lists. omAlloc() puts all acorn@1942: // assigned monitors on the per-thread list. deflate_idle_monitors() acorn@1942: // returns the non-busy monitors to the global free list. acorn@1995: // When a thread dies, omFlush() adds the list of active monitors for acorn@1995: // that thread to a global gOmInUseList acquiring the acorn@1995: // global list lock. deflate_idle_monitors() acquires the global acorn@1995: // list lock to scan for non-busy monitors to the global free list. acorn@1942: // An alternative could have used a single global inuse list. The acorn@1942: // downside would have been the additional cost of acquiring the global list lock acorn@1942: // for every omAlloc(). duke@435: // duke@435: // Perversely, the heap size -- and thus the STW safepoint rate -- duke@435: // typically drives the scavenge rate. Large heaps can mean infrequent GC, duke@435: // which in turn can mean large(r) numbers of objectmonitors in circulation. duke@435: // This is an unfortunate aspect of this design. duke@435: // duke@435: acorn@2233: enum ManifestConstants { acorn@2233: ClearResponsibleAtSTW = 0, acorn@2233: MaximumRecheckInterval = 1000 acorn@2233: } ; acorn@1942: acorn@1942: // Deflate a single monitor if not in use acorn@1942: // Return true if deflated, false if in use acorn@1942: bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj, acorn@1942: ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) { acorn@1942: bool deflated; acorn@1942: // Normal case ... The monitor is associated with obj. acorn@1942: guarantee (obj->mark() == markOopDesc::encode(mid), "invariant") ; acorn@1942: guarantee (mid == obj->mark()->monitor(), "invariant"); acorn@1942: guarantee (mid->header()->is_neutral(), "invariant"); acorn@1942: acorn@1942: if (mid->is_busy()) { acorn@1942: if (ClearResponsibleAtSTW) mid->_Responsible = NULL ; acorn@1942: deflated = false; acorn@1942: } else { acorn@1942: // Deflate the monitor if it is no longer being used acorn@1942: // It's idle - scavenge and return to the global free list acorn@1942: // plain old deflation ... acorn@1942: TEVENT (deflate_idle_monitors - scavenge1) ; acorn@1942: if (TraceMonitorInflation) { acorn@1942: if (obj->is_instance()) { acorn@1942: ResourceMark rm; acorn@1942: tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s", hseigel@5784: (void *) obj, (intptr_t) obj->mark(), obj->klass()->external_name()); acorn@1942: } acorn@1942: } acorn@1942: acorn@1942: // Restore the header back to obj acorn@1942: obj->release_set_mark(mid->header()); acorn@1942: mid->clear(); acorn@1942: acorn@1942: assert (mid->object() == NULL, "invariant") ; acorn@1942: acorn@1942: // Move the object to the working free list defined by FreeHead,FreeTail. acorn@1942: if (*FreeHeadp == NULL) *FreeHeadp = mid; acorn@1942: if (*FreeTailp != NULL) { acorn@1942: ObjectMonitor * prevtail = *FreeTailp; acorn@1995: assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); // TODO KK acorn@1942: prevtail->FreeNext = mid; acorn@1942: } acorn@1942: *FreeTailp = mid; acorn@1942: deflated = true; acorn@1942: } acorn@1942: return deflated; acorn@1942: } acorn@1942: acorn@1995: // Caller acquires ListLock acorn@1995: int ObjectSynchronizer::walk_monitor_list(ObjectMonitor** listheadp, acorn@1995: ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) { acorn@1995: ObjectMonitor* mid; acorn@1995: ObjectMonitor* next; acorn@1995: ObjectMonitor* curmidinuse = NULL; acorn@1995: int deflatedcount = 0; acorn@1995: acorn@1995: for (mid = *listheadp; mid != NULL; ) { acorn@1995: oop obj = (oop) mid->object(); acorn@1995: bool deflated = false; acorn@1995: if (obj != NULL) { acorn@1995: deflated = deflate_monitor(mid, obj, FreeHeadp, FreeTailp); acorn@1995: } acorn@1995: if (deflated) { acorn@1995: // extract from per-thread in-use-list acorn@1995: if (mid == *listheadp) { acorn@1995: *listheadp = mid->FreeNext; acorn@1995: } else if (curmidinuse != NULL) { acorn@1995: curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist acorn@1995: } acorn@1995: next = mid->FreeNext; acorn@1995: mid->FreeNext = NULL; // This mid is current tail in the FreeHead list acorn@1995: mid = next; acorn@1995: deflatedcount++; acorn@1995: } else { acorn@1995: curmidinuse = mid; acorn@1995: mid = mid->FreeNext; acorn@1995: } acorn@1995: } acorn@1995: return deflatedcount; acorn@1995: } acorn@1995: duke@435: void ObjectSynchronizer::deflate_idle_monitors() { duke@435: assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); duke@435: int nInuse = 0 ; // currently associated with objects duke@435: int nInCirculation = 0 ; // extant duke@435: int nScavenged = 0 ; // reclaimed acorn@1942: bool deflated = false; duke@435: duke@435: ObjectMonitor * FreeHead = NULL ; // Local SLL of scavenged monitors duke@435: ObjectMonitor * FreeTail = NULL ; duke@435: acorn@1942: TEVENT (deflate_idle_monitors) ; acorn@1942: // Prevent omFlush from changing mids in Thread dtor's during deflation acorn@1942: // And in case the vm thread is acquiring a lock during a safepoint acorn@1942: // See e.g. 6320749 acorn@1942: Thread::muxAcquire (&ListLock, "scavenge - return") ; acorn@1942: acorn@1942: if (MonitorInUseLists) { acorn@1995: int inUse = 0; acorn@1942: for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) { acorn@1995: nInCirculation+= cur->omInUseCount; acorn@1995: int deflatedcount = walk_monitor_list(cur->omInUseList_addr(), &FreeHead, &FreeTail); acorn@1995: cur->omInUseCount-= deflatedcount; acorn@1995: // verifyInUse(cur); acorn@1995: nScavenged += deflatedcount; acorn@1995: nInuse += cur->omInUseCount; acorn@1942: } acorn@1995: acorn@1995: // For moribund threads, scan gOmInUseList acorn@1995: if (gOmInUseList) { acorn@1995: nInCirculation += gOmInUseCount; acorn@1995: int deflatedcount = walk_monitor_list((ObjectMonitor **)&gOmInUseList, &FreeHead, &FreeTail); acorn@1995: gOmInUseCount-= deflatedcount; acorn@1995: nScavenged += deflatedcount; acorn@1995: nInuse += gOmInUseCount; acorn@1995: } acorn@1995: acorn@1942: } else for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { duke@435: // Iterate over all extant monitors - Scavenge all idle monitors. duke@435: assert(block->object() == CHAINMARKER, "must be a block header"); duke@435: nInCirculation += _BLOCKSIZE ; duke@435: for (int i = 1 ; i < _BLOCKSIZE; i++) { duke@435: ObjectMonitor* mid = &block[i]; duke@435: oop obj = (oop) mid->object(); duke@435: duke@435: if (obj == NULL) { duke@435: // The monitor is not associated with an object. duke@435: // The monitor should either be a thread-specific private duke@435: // free list or the global free list. duke@435: // obj == NULL IMPLIES mid->is_busy() == 0 duke@435: guarantee (!mid->is_busy(), "invariant") ; duke@435: continue ; duke@435: } acorn@1942: deflated = deflate_monitor(mid, obj, &FreeHead, &FreeTail); acorn@1942: acorn@1942: if (deflated) { acorn@1942: mid->FreeNext = NULL ; acorn@1942: nScavenged ++ ; duke@435: } else { acorn@1942: nInuse ++; duke@435: } duke@435: } duke@435: } duke@435: acorn@1942: MonitorFreeCount += nScavenged; acorn@1942: acorn@1942: // Consider: audit gFreeList to ensure that MonitorFreeCount and list agree. acorn@1942: acorn@2233: if (ObjectMonitor::Knob_Verbose) { acorn@1942: ::printf ("Deflate: InCirc=%d InUse=%d Scavenged=%d ForceMonitorScavenge=%d : pop=%d free=%d\n", acorn@1942: nInCirculation, nInuse, nScavenged, ForceMonitorScavenge, acorn@1942: MonitorPopulation, MonitorFreeCount) ; acorn@1942: ::fflush(stdout) ; acorn@1942: } acorn@1942: acorn@1942: ForceMonitorScavenge = 0; // Reset acorn@1942: duke@435: // Move the scavenged monitors back to the global free list. duke@435: if (FreeHead != NULL) { duke@435: guarantee (FreeTail != NULL && nScavenged > 0, "invariant") ; duke@435: assert (FreeTail->FreeNext == NULL, "invariant") ; duke@435: // constant-time list splice - prepend scavenged segment to gFreeList duke@435: FreeTail->FreeNext = gFreeList ; duke@435: gFreeList = FreeHead ; duke@435: } acorn@1942: Thread::muxRelease (&ListLock) ; duke@435: acorn@2233: if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged) ; acorn@2233: if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation); duke@435: duke@435: // TODO: Add objectMonitor leak detection. duke@435: // Audit/inventory the objectMonitors -- make sure they're all accounted for. duke@435: GVars.stwRandom = os::random() ; duke@435: GVars.stwCycle ++ ; duke@435: } duke@435: acorn@2233: // Monitor cleanup on JavaThread::exit duke@435: acorn@2233: // Iterate through monitor cache and attempt to release thread's monitors acorn@2233: // Gives up on a particular monitor if an exception occurs, but continues acorn@2233: // the overall iteration, swallowing the exception. acorn@2233: class ReleaseJavaMonitorsClosure: public MonitorClosure { acorn@2233: private: acorn@2233: TRAPS; duke@435: acorn@2233: public: acorn@2233: ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} acorn@2233: void do_monitor(ObjectMonitor* mid) { acorn@2233: if (mid->owner() == THREAD) { acorn@2233: (void)mid->complete_exit(CHECK); duke@435: } duke@435: } acorn@2233: }; acorn@2233: acorn@2233: // Release all inflated monitors owned by THREAD. Lightweight monitors are acorn@2233: // ignored. This is meant to be called during JNI thread detach which assumes acorn@2233: // all remaining monitors are heavyweight. All exceptions are swallowed. acorn@2233: // Scanning the extant monitor list can be time consuming. acorn@2233: // A simple optimization is to add a per-thread flag that indicates a thread acorn@2233: // called jni_monitorenter() during its lifetime. acorn@2233: // acorn@2233: // Instead of No_Savepoint_Verifier it might be cheaper to acorn@2233: // use an idiom of the form: acorn@2233: // auto int tmp = SafepointSynchronize::_safepoint_counter ; acorn@2233: // acorn@2233: // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; acorn@2233: // Since the tests are extremely cheap we could leave them enabled acorn@2233: // for normal product builds. acorn@2233: acorn@2233: void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { acorn@2233: assert(THREAD == JavaThread::current(), "must be current Java thread"); acorn@2233: No_Safepoint_Verifier nsv ; acorn@2233: ReleaseJavaMonitorsClosure rjmc(THREAD); acorn@2233: Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread"); acorn@2233: ObjectSynchronizer::monitors_iterate(&rjmc); acorn@2233: Thread::muxRelease(&ListLock); acorn@2233: THREAD->clear_pending_exception(); duke@435: } duke@435: duke@435: //------------------------------------------------------------------------------ duke@435: // Non-product code duke@435: duke@435: #ifndef PRODUCT duke@435: duke@435: // Verify all monitors in the monitor cache, the verification is weak. duke@435: void ObjectSynchronizer::verify() { duke@435: ObjectMonitor* block = gBlockList; duke@435: ObjectMonitor* mid; duke@435: while (block) { duke@435: assert(block->object() == CHAINMARKER, "must be a block header"); duke@435: for (int i = 1; i < _BLOCKSIZE; i++) { duke@435: mid = block + i; duke@435: oop object = (oop) mid->object(); duke@435: if (object != NULL) { duke@435: mid->verify(); duke@435: } duke@435: } duke@435: block = (ObjectMonitor*) block->FreeNext; duke@435: } duke@435: } duke@435: duke@435: // Check if monitor belongs to the monitor cache duke@435: // The list is grow-only so it's *relatively* safe to traverse duke@435: // the list of extant blocks without taking a lock. duke@435: duke@435: int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) { duke@435: ObjectMonitor* block = gBlockList; duke@435: duke@435: while (block) { duke@435: assert(block->object() == CHAINMARKER, "must be a block header"); duke@435: if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) { duke@435: address mon = (address) monitor; duke@435: address blk = (address) block; duke@435: size_t diff = mon - blk; duke@435: assert((diff % sizeof(ObjectMonitor)) == 0, "check"); duke@435: return 1; duke@435: } duke@435: block = (ObjectMonitor*) block->FreeNext; duke@435: } duke@435: return 0; duke@435: } duke@435: duke@435: #endif