Tue, 09 Oct 2012 10:09:34 -0700
7197424: update copyright year to match last edit in jdk8 hotspot repository
Summary: Update copyright year to 2012 for relevant files
Reviewed-by: dholmes, coleenp
1 /*
2 * Copyright (c) 1998, 2012, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "classfile/vmSymbols.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "oops/markOop.hpp"
29 #include "oops/oop.inline.hpp"
30 #include "runtime/biasedLocking.hpp"
31 #include "runtime/handles.inline.hpp"
32 #include "runtime/interfaceSupport.hpp"
33 #include "runtime/mutexLocker.hpp"
34 #include "runtime/objectMonitor.hpp"
35 #include "runtime/objectMonitor.inline.hpp"
36 #include "runtime/osThread.hpp"
37 #include "runtime/stubRoutines.hpp"
38 #include "runtime/synchronizer.hpp"
39 #include "utilities/dtrace.hpp"
40 #include "utilities/events.hpp"
41 #include "utilities/preserveException.hpp"
42 #ifdef TARGET_OS_FAMILY_linux
43 # include "os_linux.inline.hpp"
44 # include "thread_linux.inline.hpp"
45 #endif
46 #ifdef TARGET_OS_FAMILY_solaris
47 # include "os_solaris.inline.hpp"
48 # include "thread_solaris.inline.hpp"
49 #endif
50 #ifdef TARGET_OS_FAMILY_windows
51 # include "os_windows.inline.hpp"
52 # include "thread_windows.inline.hpp"
53 #endif
54 #ifdef TARGET_OS_FAMILY_bsd
55 # include "os_bsd.inline.hpp"
56 # include "thread_bsd.inline.hpp"
57 #endif
59 #if defined(__GNUC__) && !defined(IA64)
60 // Need to inhibit inlining for older versions of GCC to avoid build-time failures
61 #define ATTR __attribute__((noinline))
62 #else
63 #define ATTR
64 #endif
66 // The "core" versions of monitor enter and exit reside in this file.
67 // The interpreter and compilers contain specialized transliterated
68 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(),
69 // for instance. If you make changes here, make sure to modify the
70 // interpreter, and both C1 and C2 fast-path inline locking code emission.
71 //
72 //
73 // -----------------------------------------------------------------------------
75 #ifdef DTRACE_ENABLED
77 // Only bother with this argument setup if dtrace is available
78 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
80 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
81 char* bytes = NULL; \
82 int len = 0; \
83 jlong jtid = SharedRuntime::get_java_tid(thread); \
84 Symbol* klassname = ((oop)(obj))->klass()->name(); \
85 if (klassname != NULL) { \
86 bytes = (char*)klassname->bytes(); \
87 len = klassname->utf8_length(); \
88 }
90 #ifndef USDT2
91 HS_DTRACE_PROBE_DECL5(hotspot, monitor__wait,
92 jlong, uintptr_t, char*, int, long);
93 HS_DTRACE_PROBE_DECL4(hotspot, monitor__waited,
94 jlong, uintptr_t, char*, int);
96 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
97 { \
98 if (DTraceMonitorProbes) { \
99 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
100 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \
101 (monitor), bytes, len, (millis)); \
102 } \
103 }
105 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
106 { \
107 if (DTraceMonitorProbes) { \
108 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
109 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \
110 (uintptr_t)(monitor), bytes, len); \
111 } \
112 }
114 #else /* USDT2 */
116 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
117 { \
118 if (DTraceMonitorProbes) { \
119 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
120 HOTSPOT_MONITOR_WAIT(jtid, \
121 (uintptr_t)(monitor), bytes, len, (millis)); \
122 } \
123 }
125 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_PROBE_WAITED
127 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
128 { \
129 if (DTraceMonitorProbes) { \
130 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
131 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
132 (uintptr_t)(monitor), bytes, len); \
133 } \
134 }
136 #endif /* USDT2 */
137 #else // ndef DTRACE_ENABLED
139 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
140 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
142 #endif // ndef DTRACE_ENABLED
144 // This exists only as a workaround of dtrace bug 6254741
145 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
146 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
147 return 0;
148 }
150 #define NINFLATIONLOCKS 256
151 static volatile intptr_t InflationLocks [NINFLATIONLOCKS] ;
153 ObjectMonitor * ObjectSynchronizer::gBlockList = NULL ;
154 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL ;
155 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL ;
156 int ObjectSynchronizer::gOmInUseCount = 0;
157 static volatile intptr_t ListLock = 0 ; // protects global monitor free-list cache
158 static volatile int MonitorFreeCount = 0 ; // # on gFreeList
159 static volatile int MonitorPopulation = 0 ; // # Extant -- in circulation
160 #define CHAINMARKER ((oop)-1)
162 // -----------------------------------------------------------------------------
163 // Fast Monitor Enter/Exit
164 // This the fast monitor enter. The interpreter and compiler use
165 // some assembly copies of this code. Make sure update those code
166 // if the following function is changed. The implementation is
167 // extremely sensitive to race condition. Be careful.
169 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) {
170 if (UseBiasedLocking) {
171 if (!SafepointSynchronize::is_at_safepoint()) {
172 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD);
173 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) {
174 return;
175 }
176 } else {
177 assert(!attempt_rebias, "can not rebias toward VM thread");
178 BiasedLocking::revoke_at_safepoint(obj);
179 }
180 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
181 }
183 slow_enter (obj, lock, THREAD) ;
184 }
186 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) {
187 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here");
188 // if displaced header is null, the previous enter is recursive enter, no-op
189 markOop dhw = lock->displaced_header();
190 markOop mark ;
191 if (dhw == NULL) {
192 // Recursive stack-lock.
193 // Diagnostics -- Could be: stack-locked, inflating, inflated.
194 mark = object->mark() ;
195 assert (!mark->is_neutral(), "invariant") ;
196 if (mark->has_locker() && mark != markOopDesc::INFLATING()) {
197 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant") ;
198 }
199 if (mark->has_monitor()) {
200 ObjectMonitor * m = mark->monitor() ;
201 assert(((oop)(m->object()))->mark() == mark, "invariant") ;
202 assert(m->is_entered(THREAD), "invariant") ;
203 }
204 return ;
205 }
207 mark = object->mark() ;
209 // If the object is stack-locked by the current thread, try to
210 // swing the displaced header from the box back to the mark.
211 if (mark == (markOop) lock) {
212 assert (dhw->is_neutral(), "invariant") ;
213 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) {
214 TEVENT (fast_exit: release stacklock) ;
215 return;
216 }
217 }
219 ObjectSynchronizer::inflate(THREAD, object)->exit (THREAD) ;
220 }
222 // -----------------------------------------------------------------------------
223 // Interpreter/Compiler Slow Case
224 // This routine is used to handle interpreter/compiler slow case
225 // We don't need to use fast path here, because it must have been
226 // failed in the interpreter/compiler code.
227 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
228 markOop mark = obj->mark();
229 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
231 if (mark->is_neutral()) {
232 // Anticipate successful CAS -- the ST of the displaced mark must
233 // be visible <= the ST performed by the CAS.
234 lock->set_displaced_header(mark);
235 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) {
236 TEVENT (slow_enter: release stacklock) ;
237 return ;
238 }
239 // Fall through to inflate() ...
240 } else
241 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
242 assert(lock != mark->locker(), "must not re-lock the same lock");
243 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
244 lock->set_displaced_header(NULL);
245 return;
246 }
248 #if 0
249 // The following optimization isn't particularly useful.
250 if (mark->has_monitor() && mark->monitor()->is_entered(THREAD)) {
251 lock->set_displaced_header (NULL) ;
252 return ;
253 }
254 #endif
256 // The object header will never be displaced to this lock,
257 // so it does not matter what the value is, except that it
258 // must be non-zero to avoid looking like a re-entrant lock,
259 // and must not look locked either.
260 lock->set_displaced_header(markOopDesc::unused_mark());
261 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
262 }
264 // This routine is used to handle interpreter/compiler slow case
265 // We don't need to use fast path here, because it must have
266 // failed in the interpreter/compiler code. Simply use the heavy
267 // weight monitor should be ok, unless someone find otherwise.
268 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
269 fast_exit (object, lock, THREAD) ;
270 }
272 // -----------------------------------------------------------------------------
273 // Class Loader support to workaround deadlocks on the class loader lock objects
274 // Also used by GC
275 // complete_exit()/reenter() are used to wait on a nested lock
276 // i.e. to give up an outer lock completely and then re-enter
277 // Used when holding nested locks - lock acquisition order: lock1 then lock2
278 // 1) complete_exit lock1 - saving recursion count
279 // 2) wait on lock2
280 // 3) when notified on lock2, unlock lock2
281 // 4) reenter lock1 with original recursion count
282 // 5) lock lock2
283 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
284 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
285 TEVENT (complete_exit) ;
286 if (UseBiasedLocking) {
287 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
288 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
289 }
291 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
293 return monitor->complete_exit(THREAD);
294 }
296 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
297 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
298 TEVENT (reenter) ;
299 if (UseBiasedLocking) {
300 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
301 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
302 }
304 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
306 monitor->reenter(recursion, THREAD);
307 }
308 // -----------------------------------------------------------------------------
309 // JNI locks on java objects
310 // NOTE: must use heavy weight monitor to handle jni monitor enter
311 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { // possible entry from jni enter
312 // the current locking is from JNI instead of Java code
313 TEVENT (jni_enter) ;
314 if (UseBiasedLocking) {
315 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
316 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
317 }
318 THREAD->set_current_pending_monitor_is_from_java(false);
319 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
320 THREAD->set_current_pending_monitor_is_from_java(true);
321 }
323 // NOTE: must use heavy weight monitor to handle jni monitor enter
324 bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) {
325 if (UseBiasedLocking) {
326 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
327 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
328 }
330 ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj());
331 return monitor->try_enter(THREAD);
332 }
335 // NOTE: must use heavy weight monitor to handle jni monitor exit
336 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
337 TEVENT (jni_exit) ;
338 if (UseBiasedLocking) {
339 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
340 }
341 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
343 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj);
344 // If this thread has locked the object, exit the monitor. Note: can't use
345 // monitor->check(CHECK); must exit even if an exception is pending.
346 if (monitor->check(THREAD)) {
347 monitor->exit(THREAD);
348 }
349 }
351 // -----------------------------------------------------------------------------
352 // Internal VM locks on java objects
353 // standard constructor, allows locking failures
354 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
355 _dolock = doLock;
356 _thread = thread;
357 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
358 _obj = obj;
360 if (_dolock) {
361 TEVENT (ObjectLocker) ;
363 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
364 }
365 }
367 ObjectLocker::~ObjectLocker() {
368 if (_dolock) {
369 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
370 }
371 }
374 // -----------------------------------------------------------------------------
375 // Wait/Notify/NotifyAll
376 // NOTE: must use heavy weight monitor to handle wait()
377 void ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
378 if (UseBiasedLocking) {
379 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
380 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
381 }
382 if (millis < 0) {
383 TEVENT (wait - throw IAX) ;
384 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
385 }
386 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
387 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
388 monitor->wait(millis, true, THREAD);
390 /* This dummy call is in place to get around dtrace bug 6254741. Once
391 that's fixed we can uncomment the following line and remove the call */
392 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
393 dtrace_waited_probe(monitor, obj, THREAD);
394 }
396 void ObjectSynchronizer::waitUninterruptibly (Handle obj, jlong millis, TRAPS) {
397 if (UseBiasedLocking) {
398 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
399 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
400 }
401 if (millis < 0) {
402 TEVENT (wait - throw IAX) ;
403 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
404 }
405 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD) ;
406 }
408 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
409 if (UseBiasedLocking) {
410 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
411 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
412 }
414 markOop mark = obj->mark();
415 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
416 return;
417 }
418 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD);
419 }
421 // NOTE: see comment of notify()
422 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
423 if (UseBiasedLocking) {
424 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
425 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
426 }
428 markOop mark = obj->mark();
429 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
430 return;
431 }
432 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD);
433 }
435 // -----------------------------------------------------------------------------
436 // Hash Code handling
437 //
438 // Performance concern:
439 // OrderAccess::storestore() calls release() which STs 0 into the global volatile
440 // OrderAccess::Dummy variable. This store is unnecessary for correctness.
441 // Many threads STing into a common location causes considerable cache migration
442 // or "sloshing" on large SMP system. As such, I avoid using OrderAccess::storestore()
443 // until it's repaired. In some cases OrderAccess::fence() -- which incurs local
444 // latency on the executing processor -- is a better choice as it scales on SMP
445 // systems. See http://blogs.sun.com/dave/entry/biased_locking_in_hotspot for a
446 // discussion of coherency costs. Note that all our current reference platforms
447 // provide strong ST-ST order, so the issue is moot on IA32, x64, and SPARC.
448 //
449 // As a general policy we use "volatile" to control compiler-based reordering
450 // and explicit fences (barriers) to control for architectural reordering performed
451 // by the CPU(s) or platform.
453 static int MBFence (int x) { OrderAccess::fence(); return x; }
455 struct SharedGlobals {
456 // These are highly shared mostly-read variables.
457 // To avoid false-sharing they need to be the sole occupants of a $ line.
458 double padPrefix [8];
459 volatile int stwRandom ;
460 volatile int stwCycle ;
462 // Hot RW variables -- Sequester to avoid false-sharing
463 double padSuffix [16];
464 volatile int hcSequence ;
465 double padFinal [8] ;
466 } ;
468 static SharedGlobals GVars ;
469 static int MonitorScavengeThreshold = 1000000 ;
470 static volatile int ForceMonitorScavenge = 0 ; // Scavenge required and pending
472 static markOop ReadStableMark (oop obj) {
473 markOop mark = obj->mark() ;
474 if (!mark->is_being_inflated()) {
475 return mark ; // normal fast-path return
476 }
478 int its = 0 ;
479 for (;;) {
480 markOop mark = obj->mark() ;
481 if (!mark->is_being_inflated()) {
482 return mark ; // normal fast-path return
483 }
485 // The object is being inflated by some other thread.
486 // The caller of ReadStableMark() must wait for inflation to complete.
487 // Avoid live-lock
488 // TODO: consider calling SafepointSynchronize::do_call_back() while
489 // spinning to see if there's a safepoint pending. If so, immediately
490 // yielding or blocking would be appropriate. Avoid spinning while
491 // there is a safepoint pending.
492 // TODO: add inflation contention performance counters.
493 // TODO: restrict the aggregate number of spinners.
495 ++its ;
496 if (its > 10000 || !os::is_MP()) {
497 if (its & 1) {
498 os::NakedYield() ;
499 TEVENT (Inflate: INFLATING - yield) ;
500 } else {
501 // Note that the following code attenuates the livelock problem but is not
502 // a complete remedy. A more complete solution would require that the inflating
503 // thread hold the associated inflation lock. The following code simply restricts
504 // the number of spinners to at most one. We'll have N-2 threads blocked
505 // on the inflationlock, 1 thread holding the inflation lock and using
506 // a yield/park strategy, and 1 thread in the midst of inflation.
507 // A more refined approach would be to change the encoding of INFLATING
508 // to allow encapsulation of a native thread pointer. Threads waiting for
509 // inflation to complete would use CAS to push themselves onto a singly linked
510 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
511 // and calling park(). When inflation was complete the thread that accomplished inflation
512 // would detach the list and set the markword to inflated with a single CAS and
513 // then for each thread on the list, set the flag and unpark() the thread.
514 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease
515 // wakes at most one thread whereas we need to wake the entire list.
516 int ix = (intptr_t(obj) >> 5) & (NINFLATIONLOCKS-1) ;
517 int YieldThenBlock = 0 ;
518 assert (ix >= 0 && ix < NINFLATIONLOCKS, "invariant") ;
519 assert ((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant") ;
520 Thread::muxAcquire (InflationLocks + ix, "InflationLock") ;
521 while (obj->mark() == markOopDesc::INFLATING()) {
522 // Beware: NakedYield() is advisory and has almost no effect on some platforms
523 // so we periodically call Self->_ParkEvent->park(1).
524 // We use a mixed spin/yield/block mechanism.
525 if ((YieldThenBlock++) >= 16) {
526 Thread::current()->_ParkEvent->park(1) ;
527 } else {
528 os::NakedYield() ;
529 }
530 }
531 Thread::muxRelease (InflationLocks + ix ) ;
532 TEVENT (Inflate: INFLATING - yield/park) ;
533 }
534 } else {
535 SpinPause() ; // SMP-polite spinning
536 }
537 }
538 }
540 // hashCode() generation :
541 //
542 // Possibilities:
543 // * MD5Digest of {obj,stwRandom}
544 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function.
545 // * A DES- or AES-style SBox[] mechanism
546 // * One of the Phi-based schemes, such as:
547 // 2654435761 = 2^32 * Phi (golden ratio)
548 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ;
549 // * A variation of Marsaglia's shift-xor RNG scheme.
550 // * (obj ^ stwRandom) is appealing, but can result
551 // in undesirable regularity in the hashCode values of adjacent objects
552 // (objects allocated back-to-back, in particular). This could potentially
553 // result in hashtable collisions and reduced hashtable efficiency.
554 // There are simple ways to "diffuse" the middle address bits over the
555 // generated hashCode values:
556 //
558 static inline intptr_t get_next_hash(Thread * Self, oop obj) {
559 intptr_t value = 0 ;
560 if (hashCode == 0) {
561 // This form uses an unguarded global Park-Miller RNG,
562 // so it's possible for two threads to race and generate the same RNG.
563 // On MP system we'll have lots of RW access to a global, so the
564 // mechanism induces lots of coherency traffic.
565 value = os::random() ;
566 } else
567 if (hashCode == 1) {
568 // This variation has the property of being stable (idempotent)
569 // between STW operations. This can be useful in some of the 1-0
570 // synchronization schemes.
571 intptr_t addrBits = intptr_t(obj) >> 3 ;
572 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom ;
573 } else
574 if (hashCode == 2) {
575 value = 1 ; // for sensitivity testing
576 } else
577 if (hashCode == 3) {
578 value = ++GVars.hcSequence ;
579 } else
580 if (hashCode == 4) {
581 value = intptr_t(obj) ;
582 } else {
583 // Marsaglia's xor-shift scheme with thread-specific state
584 // This is probably the best overall implementation -- we'll
585 // likely make this the default in future releases.
586 unsigned t = Self->_hashStateX ;
587 t ^= (t << 11) ;
588 Self->_hashStateX = Self->_hashStateY ;
589 Self->_hashStateY = Self->_hashStateZ ;
590 Self->_hashStateZ = Self->_hashStateW ;
591 unsigned v = Self->_hashStateW ;
592 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)) ;
593 Self->_hashStateW = v ;
594 value = v ;
595 }
597 value &= markOopDesc::hash_mask;
598 if (value == 0) value = 0xBAD ;
599 assert (value != markOopDesc::no_hash, "invariant") ;
600 TEVENT (hashCode: GENERATE) ;
601 return value;
602 }
603 //
604 intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) {
605 if (UseBiasedLocking) {
606 // NOTE: many places throughout the JVM do not expect a safepoint
607 // to be taken here, in particular most operations on perm gen
608 // objects. However, we only ever bias Java instances and all of
609 // the call sites of identity_hash that might revoke biases have
610 // been checked to make sure they can handle a safepoint. The
611 // added check of the bias pattern is to avoid useless calls to
612 // thread-local storage.
613 if (obj->mark()->has_bias_pattern()) {
614 // Box and unbox the raw reference just in case we cause a STW safepoint.
615 Handle hobj (Self, obj) ;
616 // Relaxing assertion for bug 6320749.
617 assert (Universe::verify_in_progress() ||
618 !SafepointSynchronize::is_at_safepoint(),
619 "biases should not be seen by VM thread here");
620 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
621 obj = hobj() ;
622 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
623 }
624 }
626 // hashCode() is a heap mutator ...
627 // Relaxing assertion for bug 6320749.
628 assert (Universe::verify_in_progress() ||
629 !SafepointSynchronize::is_at_safepoint(), "invariant") ;
630 assert (Universe::verify_in_progress() ||
631 Self->is_Java_thread() , "invariant") ;
632 assert (Universe::verify_in_progress() ||
633 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ;
635 ObjectMonitor* monitor = NULL;
636 markOop temp, test;
637 intptr_t hash;
638 markOop mark = ReadStableMark (obj);
640 // object should remain ineligible for biased locking
641 assert (!mark->has_bias_pattern(), "invariant") ;
643 if (mark->is_neutral()) {
644 hash = mark->hash(); // this is a normal header
645 if (hash) { // if it has hash, just return it
646 return hash;
647 }
648 hash = get_next_hash(Self, obj); // allocate a new hash code
649 temp = mark->copy_set_hash(hash); // merge the hash code into header
650 // use (machine word version) atomic operation to install the hash
651 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark);
652 if (test == mark) {
653 return hash;
654 }
655 // If atomic operation failed, we must inflate the header
656 // into heavy weight monitor. We could add more code here
657 // for fast path, but it does not worth the complexity.
658 } else if (mark->has_monitor()) {
659 monitor = mark->monitor();
660 temp = monitor->header();
661 assert (temp->is_neutral(), "invariant") ;
662 hash = temp->hash();
663 if (hash) {
664 return hash;
665 }
666 // Skip to the following code to reduce code size
667 } else if (Self->is_lock_owned((address)mark->locker())) {
668 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
669 assert (temp->is_neutral(), "invariant") ;
670 hash = temp->hash(); // by current thread, check if the displaced
671 if (hash) { // header contains hash code
672 return hash;
673 }
674 // WARNING:
675 // The displaced header is strictly immutable.
676 // It can NOT be changed in ANY cases. So we have
677 // to inflate the header into heavyweight monitor
678 // even the current thread owns the lock. The reason
679 // is the BasicLock (stack slot) will be asynchronously
680 // read by other threads during the inflate() function.
681 // Any change to stack may not propagate to other threads
682 // correctly.
683 }
685 // Inflate the monitor to set hash code
686 monitor = ObjectSynchronizer::inflate(Self, obj);
687 // Load displaced header and check it has hash code
688 mark = monitor->header();
689 assert (mark->is_neutral(), "invariant") ;
690 hash = mark->hash();
691 if (hash == 0) {
692 hash = get_next_hash(Self, obj);
693 temp = mark->copy_set_hash(hash); // merge hash code into header
694 assert (temp->is_neutral(), "invariant") ;
695 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark);
696 if (test != mark) {
697 // The only update to the header in the monitor (outside GC)
698 // is install the hash code. If someone add new usage of
699 // displaced header, please update this code
700 hash = test->hash();
701 assert (test->is_neutral(), "invariant") ;
702 assert (hash != 0, "Trivial unexpected object/monitor header usage.");
703 }
704 }
705 // We finally get the hash
706 return hash;
707 }
709 // Deprecated -- use FastHashCode() instead.
711 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
712 return FastHashCode (Thread::current(), obj()) ;
713 }
716 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
717 Handle h_obj) {
718 if (UseBiasedLocking) {
719 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
720 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
721 }
723 assert(thread == JavaThread::current(), "Can only be called on current thread");
724 oop obj = h_obj();
726 markOop mark = ReadStableMark (obj) ;
728 // Uncontended case, header points to stack
729 if (mark->has_locker()) {
730 return thread->is_lock_owned((address)mark->locker());
731 }
732 // Contended case, header points to ObjectMonitor (tagged pointer)
733 if (mark->has_monitor()) {
734 ObjectMonitor* monitor = mark->monitor();
735 return monitor->is_entered(thread) != 0 ;
736 }
737 // Unlocked case, header in place
738 assert(mark->is_neutral(), "sanity check");
739 return false;
740 }
742 // Be aware of this method could revoke bias of the lock object.
743 // This method querys the ownership of the lock handle specified by 'h_obj'.
744 // If the current thread owns the lock, it returns owner_self. If no
745 // thread owns the lock, it returns owner_none. Otherwise, it will return
746 // ower_other.
747 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
748 (JavaThread *self, Handle h_obj) {
749 // The caller must beware this method can revoke bias, and
750 // revocation can result in a safepoint.
751 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ;
752 assert (self->thread_state() != _thread_blocked , "invariant") ;
754 // Possible mark states: neutral, biased, stack-locked, inflated
756 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
757 // CASE: biased
758 BiasedLocking::revoke_and_rebias(h_obj, false, self);
759 assert(!h_obj->mark()->has_bias_pattern(),
760 "biases should be revoked by now");
761 }
763 assert(self == JavaThread::current(), "Can only be called on current thread");
764 oop obj = h_obj();
765 markOop mark = ReadStableMark (obj) ;
767 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
768 if (mark->has_locker()) {
769 return self->is_lock_owned((address)mark->locker()) ?
770 owner_self : owner_other;
771 }
773 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor.
774 // The Object:ObjectMonitor relationship is stable as long as we're
775 // not at a safepoint.
776 if (mark->has_monitor()) {
777 void * owner = mark->monitor()->_owner ;
778 if (owner == NULL) return owner_none ;
779 return (owner == self ||
780 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
781 }
783 // CASE: neutral
784 assert(mark->is_neutral(), "sanity check");
785 return owner_none ; // it's unlocked
786 }
788 // FIXME: jvmti should call this
789 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) {
790 if (UseBiasedLocking) {
791 if (SafepointSynchronize::is_at_safepoint()) {
792 BiasedLocking::revoke_at_safepoint(h_obj);
793 } else {
794 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
795 }
796 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
797 }
799 oop obj = h_obj();
800 address owner = NULL;
802 markOop mark = ReadStableMark (obj) ;
804 // Uncontended case, header points to stack
805 if (mark->has_locker()) {
806 owner = (address) mark->locker();
807 }
809 // Contended case, header points to ObjectMonitor (tagged pointer)
810 if (mark->has_monitor()) {
811 ObjectMonitor* monitor = mark->monitor();
812 assert(monitor != NULL, "monitor should be non-null");
813 owner = (address) monitor->owner();
814 }
816 if (owner != NULL) {
817 return Threads::owning_thread_from_monitor_owner(owner, doLock);
818 }
820 // Unlocked case, header in place
821 // Cannot have assertion since this object may have been
822 // locked by another thread when reaching here.
823 // assert(mark->is_neutral(), "sanity check");
825 return NULL;
826 }
827 // Visitors ...
829 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
830 ObjectMonitor* block = gBlockList;
831 ObjectMonitor* mid;
832 while (block) {
833 assert(block->object() == CHAINMARKER, "must be a block header");
834 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
835 mid = block + i;
836 oop object = (oop) mid->object();
837 if (object != NULL) {
838 closure->do_monitor(mid);
839 }
840 }
841 block = (ObjectMonitor*) block->FreeNext;
842 }
843 }
845 // Get the next block in the block list.
846 static inline ObjectMonitor* next(ObjectMonitor* block) {
847 assert(block->object() == CHAINMARKER, "must be a block header");
848 block = block->FreeNext ;
849 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
850 return block;
851 }
854 void ObjectSynchronizer::oops_do(OopClosure* f) {
855 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
856 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) {
857 assert(block->object() == CHAINMARKER, "must be a block header");
858 for (int i = 1; i < _BLOCKSIZE; i++) {
859 ObjectMonitor* mid = &block[i];
860 if (mid->object() != NULL) {
861 f->do_oop((oop*)mid->object_addr());
862 }
863 }
864 }
865 }
868 // -----------------------------------------------------------------------------
869 // ObjectMonitor Lifecycle
870 // -----------------------
871 // Inflation unlinks monitors from the global gFreeList and
872 // associates them with objects. Deflation -- which occurs at
873 // STW-time -- disassociates idle monitors from objects. Such
874 // scavenged monitors are returned to the gFreeList.
875 //
876 // The global list is protected by ListLock. All the critical sections
877 // are short and operate in constant-time.
878 //
879 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
880 //
881 // Lifecycle:
882 // -- unassigned and on the global free list
883 // -- unassigned and on a thread's private omFreeList
884 // -- assigned to an object. The object is inflated and the mark refers
885 // to the objectmonitor.
886 //
889 // Constraining monitor pool growth via MonitorBound ...
890 //
891 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
892 // the rate of scavenging is driven primarily by GC. As such, we can find
893 // an inordinate number of monitors in circulation.
894 // To avoid that scenario we can artificially induce a STW safepoint
895 // if the pool appears to be growing past some reasonable bound.
896 // Generally we favor time in space-time tradeoffs, but as there's no
897 // natural back-pressure on the # of extant monitors we need to impose some
898 // type of limit. Beware that if MonitorBound is set to too low a value
899 // we could just loop. In addition, if MonitorBound is set to a low value
900 // we'll incur more safepoints, which are harmful to performance.
901 // See also: GuaranteedSafepointInterval
902 //
903 // The current implementation uses asynchronous VM operations.
904 //
906 static void InduceScavenge (Thread * Self, const char * Whence) {
907 // Induce STW safepoint to trim monitors
908 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
909 // More precisely, trigger an asynchronous STW safepoint as the number
910 // of active monitors passes the specified threshold.
911 // TODO: assert thread state is reasonable
913 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
914 if (ObjectMonitor::Knob_Verbose) {
915 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ;
916 ::fflush(stdout) ;
917 }
918 // Induce a 'null' safepoint to scavenge monitors
919 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
920 // to the VMthread and have a lifespan longer than that of this activation record.
921 // The VMThread will delete the op when completed.
922 VMThread::execute (new VM_ForceAsyncSafepoint()) ;
924 if (ObjectMonitor::Knob_Verbose) {
925 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ;
926 ::fflush(stdout) ;
927 }
928 }
929 }
930 /* Too slow for general assert or debug
931 void ObjectSynchronizer::verifyInUse (Thread *Self) {
932 ObjectMonitor* mid;
933 int inusetally = 0;
934 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) {
935 inusetally ++;
936 }
937 assert(inusetally == Self->omInUseCount, "inuse count off");
939 int freetally = 0;
940 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) {
941 freetally ++;
942 }
943 assert(freetally == Self->omFreeCount, "free count off");
944 }
945 */
946 ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) {
947 // A large MAXPRIVATE value reduces both list lock contention
948 // and list coherency traffic, but also tends to increase the
949 // number of objectMonitors in circulation as well as the STW
950 // scavenge costs. As usual, we lean toward time in space-time
951 // tradeoffs.
952 const int MAXPRIVATE = 1024 ;
953 for (;;) {
954 ObjectMonitor * m ;
956 // 1: try to allocate from the thread's local omFreeList.
957 // Threads will attempt to allocate first from their local list, then
958 // from the global list, and only after those attempts fail will the thread
959 // attempt to instantiate new monitors. Thread-local free lists take
960 // heat off the ListLock and improve allocation latency, as well as reducing
961 // coherency traffic on the shared global list.
962 m = Self->omFreeList ;
963 if (m != NULL) {
964 Self->omFreeList = m->FreeNext ;
965 Self->omFreeCount -- ;
966 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene
967 guarantee (m->object() == NULL, "invariant") ;
968 if (MonitorInUseLists) {
969 m->FreeNext = Self->omInUseList;
970 Self->omInUseList = m;
971 Self->omInUseCount ++;
972 // verifyInUse(Self);
973 } else {
974 m->FreeNext = NULL;
975 }
976 return m ;
977 }
979 // 2: try to allocate from the global gFreeList
980 // CONSIDER: use muxTry() instead of muxAcquire().
981 // If the muxTry() fails then drop immediately into case 3.
982 // If we're using thread-local free lists then try
983 // to reprovision the caller's free list.
984 if (gFreeList != NULL) {
985 // Reprovision the thread's omFreeList.
986 // Use bulk transfers to reduce the allocation rate and heat
987 // on various locks.
988 Thread::muxAcquire (&ListLock, "omAlloc") ;
989 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL; ) {
990 MonitorFreeCount --;
991 ObjectMonitor * take = gFreeList ;
992 gFreeList = take->FreeNext ;
993 guarantee (take->object() == NULL, "invariant") ;
994 guarantee (!take->is_busy(), "invariant") ;
995 take->Recycle() ;
996 omRelease (Self, take, false) ;
997 }
998 Thread::muxRelease (&ListLock) ;
999 Self->omFreeProvision += 1 + (Self->omFreeProvision/2) ;
1000 if (Self->omFreeProvision > MAXPRIVATE ) Self->omFreeProvision = MAXPRIVATE ;
1001 TEVENT (omFirst - reprovision) ;
1003 const int mx = MonitorBound ;
1004 if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) {
1005 // We can't safely induce a STW safepoint from omAlloc() as our thread
1006 // state may not be appropriate for such activities and callers may hold
1007 // naked oops, so instead we defer the action.
1008 InduceScavenge (Self, "omAlloc") ;
1009 }
1010 continue;
1011 }
1013 // 3: allocate a block of new ObjectMonitors
1014 // Both the local and global free lists are empty -- resort to malloc().
1015 // In the current implementation objectMonitors are TSM - immortal.
1016 assert (_BLOCKSIZE > 1, "invariant") ;
1017 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE];
1019 // NOTE: (almost) no way to recover if allocation failed.
1020 // We might be able to induce a STW safepoint and scavenge enough
1021 // objectMonitors to permit progress.
1022 if (temp == NULL) {
1023 vm_exit_out_of_memory (sizeof (ObjectMonitor[_BLOCKSIZE]), "Allocate ObjectMonitors") ;
1024 }
1026 // Format the block.
1027 // initialize the linked list, each monitor points to its next
1028 // forming the single linked free list, the very first monitor
1029 // will points to next block, which forms the block list.
1030 // The trick of using the 1st element in the block as gBlockList
1031 // linkage should be reconsidered. A better implementation would
1032 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1034 for (int i = 1; i < _BLOCKSIZE ; i++) {
1035 temp[i].FreeNext = &temp[i+1];
1036 }
1038 // terminate the last monitor as the end of list
1039 temp[_BLOCKSIZE - 1].FreeNext = NULL ;
1041 // Element [0] is reserved for global list linkage
1042 temp[0].set_object(CHAINMARKER);
1044 // Consider carving out this thread's current request from the
1045 // block in hand. This avoids some lock traffic and redundant
1046 // list activity.
1048 // Acquire the ListLock to manipulate BlockList and FreeList.
1049 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1050 Thread::muxAcquire (&ListLock, "omAlloc [2]") ;
1051 MonitorPopulation += _BLOCKSIZE-1;
1052 MonitorFreeCount += _BLOCKSIZE-1;
1054 // Add the new block to the list of extant blocks (gBlockList).
1055 // The very first objectMonitor in a block is reserved and dedicated.
1056 // It serves as blocklist "next" linkage.
1057 temp[0].FreeNext = gBlockList;
1058 gBlockList = temp;
1060 // Add the new string of objectMonitors to the global free list
1061 temp[_BLOCKSIZE - 1].FreeNext = gFreeList ;
1062 gFreeList = temp + 1;
1063 Thread::muxRelease (&ListLock) ;
1064 TEVENT (Allocate block of monitors) ;
1065 }
1066 }
1068 // Place "m" on the caller's private per-thread omFreeList.
1069 // In practice there's no need to clamp or limit the number of
1070 // monitors on a thread's omFreeList as the only time we'll call
1071 // omRelease is to return a monitor to the free list after a CAS
1072 // attempt failed. This doesn't allow unbounded #s of monitors to
1073 // accumulate on a thread's free list.
1074 //
1076 void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m, bool fromPerThreadAlloc) {
1077 guarantee (m->object() == NULL, "invariant") ;
1079 // Remove from omInUseList
1080 if (MonitorInUseLists && fromPerThreadAlloc) {
1081 ObjectMonitor* curmidinuse = NULL;
1082 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; ) {
1083 if (m == mid) {
1084 // extract from per-thread in-use-list
1085 if (mid == Self->omInUseList) {
1086 Self->omInUseList = mid->FreeNext;
1087 } else if (curmidinuse != NULL) {
1088 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist
1089 }
1090 Self->omInUseCount --;
1091 // verifyInUse(Self);
1092 break;
1093 } else {
1094 curmidinuse = mid;
1095 mid = mid->FreeNext;
1096 }
1097 }
1098 }
1100 // FreeNext is used for both onInUseList and omFreeList, so clear old before setting new
1101 m->FreeNext = Self->omFreeList ;
1102 Self->omFreeList = m ;
1103 Self->omFreeCount ++ ;
1104 }
1106 // Return the monitors of a moribund thread's local free list to
1107 // the global free list. Typically a thread calls omFlush() when
1108 // it's dying. We could also consider having the VM thread steal
1109 // monitors from threads that have not run java code over a few
1110 // consecutive STW safepoints. Relatedly, we might decay
1111 // omFreeProvision at STW safepoints.
1112 //
1113 // Also return the monitors of a moribund thread"s omInUseList to
1114 // a global gOmInUseList under the global list lock so these
1115 // will continue to be scanned.
1116 //
1117 // We currently call omFlush() from the Thread:: dtor _after the thread
1118 // has been excised from the thread list and is no longer a mutator.
1119 // That means that omFlush() can run concurrently with a safepoint and
1120 // the scavenge operator. Calling omFlush() from JavaThread::exit() might
1121 // be a better choice as we could safely reason that that the JVM is
1122 // not at a safepoint at the time of the call, and thus there could
1123 // be not inopportune interleavings between omFlush() and the scavenge
1124 // operator.
1126 void ObjectSynchronizer::omFlush (Thread * Self) {
1127 ObjectMonitor * List = Self->omFreeList ; // Null-terminated SLL
1128 Self->omFreeList = NULL ;
1129 ObjectMonitor * Tail = NULL ;
1130 int Tally = 0;
1131 if (List != NULL) {
1132 ObjectMonitor * s ;
1133 for (s = List ; s != NULL ; s = s->FreeNext) {
1134 Tally ++ ;
1135 Tail = s ;
1136 guarantee (s->object() == NULL, "invariant") ;
1137 guarantee (!s->is_busy(), "invariant") ;
1138 s->set_owner (NULL) ; // redundant but good hygiene
1139 TEVENT (omFlush - Move one) ;
1140 }
1141 guarantee (Tail != NULL && List != NULL, "invariant") ;
1142 }
1144 ObjectMonitor * InUseList = Self->omInUseList;
1145 ObjectMonitor * InUseTail = NULL ;
1146 int InUseTally = 0;
1147 if (InUseList != NULL) {
1148 Self->omInUseList = NULL;
1149 ObjectMonitor *curom;
1150 for (curom = InUseList; curom != NULL; curom = curom->FreeNext) {
1151 InUseTail = curom;
1152 InUseTally++;
1153 }
1154 // TODO debug
1155 assert(Self->omInUseCount == InUseTally, "inuse count off");
1156 Self->omInUseCount = 0;
1157 guarantee (InUseTail != NULL && InUseList != NULL, "invariant");
1158 }
1160 Thread::muxAcquire (&ListLock, "omFlush") ;
1161 if (Tail != NULL) {
1162 Tail->FreeNext = gFreeList ;
1163 gFreeList = List ;
1164 MonitorFreeCount += Tally;
1165 }
1167 if (InUseTail != NULL) {
1168 InUseTail->FreeNext = gOmInUseList;
1169 gOmInUseList = InUseList;
1170 gOmInUseCount += InUseTally;
1171 }
1173 Thread::muxRelease (&ListLock) ;
1174 TEVENT (omFlush) ;
1175 }
1177 // Fast path code shared by multiple functions
1178 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) {
1179 markOop mark = obj->mark();
1180 if (mark->has_monitor()) {
1181 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid");
1182 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header");
1183 return mark->monitor();
1184 }
1185 return ObjectSynchronizer::inflate(Thread::current(), obj);
1186 }
1189 // Note that we could encounter some performance loss through false-sharing as
1190 // multiple locks occupy the same $ line. Padding might be appropriate.
1193 ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) {
1194 // Inflate mutates the heap ...
1195 // Relaxing assertion for bug 6320749.
1196 assert (Universe::verify_in_progress() ||
1197 !SafepointSynchronize::is_at_safepoint(), "invariant") ;
1199 for (;;) {
1200 const markOop mark = object->mark() ;
1201 assert (!mark->has_bias_pattern(), "invariant") ;
1203 // The mark can be in one of the following states:
1204 // * Inflated - just return
1205 // * Stack-locked - coerce it to inflated
1206 // * INFLATING - busy wait for conversion to complete
1207 // * Neutral - aggressively inflate the object.
1208 // * BIASED - Illegal. We should never see this
1210 // CASE: inflated
1211 if (mark->has_monitor()) {
1212 ObjectMonitor * inf = mark->monitor() ;
1213 assert (inf->header()->is_neutral(), "invariant");
1214 assert (inf->object() == object, "invariant") ;
1215 assert (ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1216 return inf ;
1217 }
1219 // CASE: inflation in progress - inflating over a stack-lock.
1220 // Some other thread is converting from stack-locked to inflated.
1221 // Only that thread can complete inflation -- other threads must wait.
1222 // The INFLATING value is transient.
1223 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1224 // We could always eliminate polling by parking the thread on some auxiliary list.
1225 if (mark == markOopDesc::INFLATING()) {
1226 TEVENT (Inflate: spin while INFLATING) ;
1227 ReadStableMark(object) ;
1228 continue ;
1229 }
1231 // CASE: stack-locked
1232 // Could be stack-locked either by this thread or by some other thread.
1233 //
1234 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1235 // to install INFLATING into the mark word. We originally installed INFLATING,
1236 // allocated the objectmonitor, and then finally STed the address of the
1237 // objectmonitor into the mark. This was correct, but artificially lengthened
1238 // the interval in which INFLATED appeared in the mark, thus increasing
1239 // the odds of inflation contention.
1240 //
1241 // We now use per-thread private objectmonitor free lists.
1242 // These list are reprovisioned from the global free list outside the
1243 // critical INFLATING...ST interval. A thread can transfer
1244 // multiple objectmonitors en-mass from the global free list to its local free list.
1245 // This reduces coherency traffic and lock contention on the global free list.
1246 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1247 // before or after the CAS(INFLATING) operation.
1248 // See the comments in omAlloc().
1250 if (mark->has_locker()) {
1251 ObjectMonitor * m = omAlloc (Self) ;
1252 // Optimistically prepare the objectmonitor - anticipate successful CAS
1253 // We do this before the CAS in order to minimize the length of time
1254 // in which INFLATING appears in the mark.
1255 m->Recycle();
1256 m->_Responsible = NULL ;
1257 m->OwnerIsThread = 0 ;
1258 m->_recursions = 0 ;
1259 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // Consider: maintain by type/class
1261 markOop cmp = (markOop) Atomic::cmpxchg_ptr (markOopDesc::INFLATING(), object->mark_addr(), mark) ;
1262 if (cmp != mark) {
1263 omRelease (Self, m, true) ;
1264 continue ; // Interference -- just retry
1265 }
1267 // We've successfully installed INFLATING (0) into the mark-word.
1268 // This is the only case where 0 will appear in a mark-work.
1269 // Only the singular thread that successfully swings the mark-word
1270 // to 0 can perform (or more precisely, complete) inflation.
1271 //
1272 // Why do we CAS a 0 into the mark-word instead of just CASing the
1273 // mark-word from the stack-locked value directly to the new inflated state?
1274 // Consider what happens when a thread unlocks a stack-locked object.
1275 // It attempts to use CAS to swing the displaced header value from the
1276 // on-stack basiclock back into the object header. Recall also that the
1277 // header value (hashcode, etc) can reside in (a) the object header, or
1278 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1279 // header in an objectMonitor. The inflate() routine must copy the header
1280 // value from the basiclock on the owner's stack to the objectMonitor, all
1281 // the while preserving the hashCode stability invariants. If the owner
1282 // decides to release the lock while the value is 0, the unlock will fail
1283 // and control will eventually pass from slow_exit() to inflate. The owner
1284 // will then spin, waiting for the 0 value to disappear. Put another way,
1285 // the 0 causes the owner to stall if the owner happens to try to
1286 // drop the lock (restoring the header from the basiclock to the object)
1287 // while inflation is in-progress. This protocol avoids races that might
1288 // would otherwise permit hashCode values to change or "flicker" for an object.
1289 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1290 // 0 serves as a "BUSY" inflate-in-progress indicator.
1293 // fetch the displaced mark from the owner's stack.
1294 // The owner can't die or unwind past the lock while our INFLATING
1295 // object is in the mark. Furthermore the owner can't complete
1296 // an unlock on the object, either.
1297 markOop dmw = mark->displaced_mark_helper() ;
1298 assert (dmw->is_neutral(), "invariant") ;
1300 // Setup monitor fields to proper values -- prepare the monitor
1301 m->set_header(dmw) ;
1303 // Optimization: if the mark->locker stack address is associated
1304 // with this thread we could simply set m->_owner = Self and
1305 // m->OwnerIsThread = 1. Note that a thread can inflate an object
1306 // that it has stack-locked -- as might happen in wait() -- directly
1307 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1308 m->set_owner(mark->locker());
1309 m->set_object(object);
1310 // TODO-FIXME: assert BasicLock->dhw != 0.
1312 // Must preserve store ordering. The monitor state must
1313 // be stable at the time of publishing the monitor address.
1314 guarantee (object->mark() == markOopDesc::INFLATING(), "invariant") ;
1315 object->release_set_mark(markOopDesc::encode(m));
1317 // Hopefully the performance counters are allocated on distinct cache lines
1318 // to avoid false sharing on MP systems ...
1319 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ;
1320 TEVENT(Inflate: overwrite stacklock) ;
1321 if (TraceMonitorInflation) {
1322 if (object->is_instance()) {
1323 ResourceMark rm;
1324 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1325 (intptr_t) object, (intptr_t) object->mark(),
1326 Klass::cast(object->klass())->external_name());
1327 }
1328 }
1329 return m ;
1330 }
1332 // CASE: neutral
1333 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1334 // If we know we're inflating for entry it's better to inflate by swinging a
1335 // pre-locked objectMonitor pointer into the object header. A successful
1336 // CAS inflates the object *and* confers ownership to the inflating thread.
1337 // In the current implementation we use a 2-step mechanism where we CAS()
1338 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1339 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1340 // would be useful.
1342 assert (mark->is_neutral(), "invariant");
1343 ObjectMonitor * m = omAlloc (Self) ;
1344 // prepare m for installation - set monitor to initial state
1345 m->Recycle();
1346 m->set_header(mark);
1347 m->set_owner(NULL);
1348 m->set_object(object);
1349 m->OwnerIsThread = 1 ;
1350 m->_recursions = 0 ;
1351 m->_Responsible = NULL ;
1352 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // consider: keep metastats by type/class
1354 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) {
1355 m->set_object (NULL) ;
1356 m->set_owner (NULL) ;
1357 m->OwnerIsThread = 0 ;
1358 m->Recycle() ;
1359 omRelease (Self, m, true) ;
1360 m = NULL ;
1361 continue ;
1362 // interference - the markword changed - just retry.
1363 // The state-transitions are one-way, so there's no chance of
1364 // live-lock -- "Inflated" is an absorbing state.
1365 }
1367 // Hopefully the performance counters are allocated on distinct
1368 // cache lines to avoid false sharing on MP systems ...
1369 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ;
1370 TEVENT(Inflate: overwrite neutral) ;
1371 if (TraceMonitorInflation) {
1372 if (object->is_instance()) {
1373 ResourceMark rm;
1374 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1375 (intptr_t) object, (intptr_t) object->mark(),
1376 Klass::cast(object->klass())->external_name());
1377 }
1378 }
1379 return m ;
1380 }
1381 }
1383 // Note that we could encounter some performance loss through false-sharing as
1384 // multiple locks occupy the same $ line. Padding might be appropriate.
1387 // Deflate_idle_monitors() is called at all safepoints, immediately
1388 // after all mutators are stopped, but before any objects have moved.
1389 // It traverses the list of known monitors, deflating where possible.
1390 // The scavenged monitor are returned to the monitor free list.
1391 //
1392 // Beware that we scavenge at *every* stop-the-world point.
1393 // Having a large number of monitors in-circulation negatively
1394 // impacts the performance of some applications (e.g., PointBase).
1395 // Broadly, we want to minimize the # of monitors in circulation.
1396 //
1397 // We have added a flag, MonitorInUseLists, which creates a list
1398 // of active monitors for each thread. deflate_idle_monitors()
1399 // only scans the per-thread inuse lists. omAlloc() puts all
1400 // assigned monitors on the per-thread list. deflate_idle_monitors()
1401 // returns the non-busy monitors to the global free list.
1402 // When a thread dies, omFlush() adds the list of active monitors for
1403 // that thread to a global gOmInUseList acquiring the
1404 // global list lock. deflate_idle_monitors() acquires the global
1405 // list lock to scan for non-busy monitors to the global free list.
1406 // An alternative could have used a single global inuse list. The
1407 // downside would have been the additional cost of acquiring the global list lock
1408 // for every omAlloc().
1409 //
1410 // Perversely, the heap size -- and thus the STW safepoint rate --
1411 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1412 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1413 // This is an unfortunate aspect of this design.
1414 //
1416 enum ManifestConstants {
1417 ClearResponsibleAtSTW = 0,
1418 MaximumRecheckInterval = 1000
1419 } ;
1421 // Deflate a single monitor if not in use
1422 // Return true if deflated, false if in use
1423 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1424 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) {
1425 bool deflated;
1426 // Normal case ... The monitor is associated with obj.
1427 guarantee (obj->mark() == markOopDesc::encode(mid), "invariant") ;
1428 guarantee (mid == obj->mark()->monitor(), "invariant");
1429 guarantee (mid->header()->is_neutral(), "invariant");
1431 if (mid->is_busy()) {
1432 if (ClearResponsibleAtSTW) mid->_Responsible = NULL ;
1433 deflated = false;
1434 } else {
1435 // Deflate the monitor if it is no longer being used
1436 // It's idle - scavenge and return to the global free list
1437 // plain old deflation ...
1438 TEVENT (deflate_idle_monitors - scavenge1) ;
1439 if (TraceMonitorInflation) {
1440 if (obj->is_instance()) {
1441 ResourceMark rm;
1442 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1443 (intptr_t) obj, (intptr_t) obj->mark(), Klass::cast(obj->klass())->external_name());
1444 }
1445 }
1447 // Restore the header back to obj
1448 obj->release_set_mark(mid->header());
1449 mid->clear();
1451 assert (mid->object() == NULL, "invariant") ;
1453 // Move the object to the working free list defined by FreeHead,FreeTail.
1454 if (*FreeHeadp == NULL) *FreeHeadp = mid;
1455 if (*FreeTailp != NULL) {
1456 ObjectMonitor * prevtail = *FreeTailp;
1457 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); // TODO KK
1458 prevtail->FreeNext = mid;
1459 }
1460 *FreeTailp = mid;
1461 deflated = true;
1462 }
1463 return deflated;
1464 }
1466 // Caller acquires ListLock
1467 int ObjectSynchronizer::walk_monitor_list(ObjectMonitor** listheadp,
1468 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) {
1469 ObjectMonitor* mid;
1470 ObjectMonitor* next;
1471 ObjectMonitor* curmidinuse = NULL;
1472 int deflatedcount = 0;
1474 for (mid = *listheadp; mid != NULL; ) {
1475 oop obj = (oop) mid->object();
1476 bool deflated = false;
1477 if (obj != NULL) {
1478 deflated = deflate_monitor(mid, obj, FreeHeadp, FreeTailp);
1479 }
1480 if (deflated) {
1481 // extract from per-thread in-use-list
1482 if (mid == *listheadp) {
1483 *listheadp = mid->FreeNext;
1484 } else if (curmidinuse != NULL) {
1485 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist
1486 }
1487 next = mid->FreeNext;
1488 mid->FreeNext = NULL; // This mid is current tail in the FreeHead list
1489 mid = next;
1490 deflatedcount++;
1491 } else {
1492 curmidinuse = mid;
1493 mid = mid->FreeNext;
1494 }
1495 }
1496 return deflatedcount;
1497 }
1499 void ObjectSynchronizer::deflate_idle_monitors() {
1500 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1501 int nInuse = 0 ; // currently associated with objects
1502 int nInCirculation = 0 ; // extant
1503 int nScavenged = 0 ; // reclaimed
1504 bool deflated = false;
1506 ObjectMonitor * FreeHead = NULL ; // Local SLL of scavenged monitors
1507 ObjectMonitor * FreeTail = NULL ;
1509 TEVENT (deflate_idle_monitors) ;
1510 // Prevent omFlush from changing mids in Thread dtor's during deflation
1511 // And in case the vm thread is acquiring a lock during a safepoint
1512 // See e.g. 6320749
1513 Thread::muxAcquire (&ListLock, "scavenge - return") ;
1515 if (MonitorInUseLists) {
1516 int inUse = 0;
1517 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) {
1518 nInCirculation+= cur->omInUseCount;
1519 int deflatedcount = walk_monitor_list(cur->omInUseList_addr(), &FreeHead, &FreeTail);
1520 cur->omInUseCount-= deflatedcount;
1521 // verifyInUse(cur);
1522 nScavenged += deflatedcount;
1523 nInuse += cur->omInUseCount;
1524 }
1526 // For moribund threads, scan gOmInUseList
1527 if (gOmInUseList) {
1528 nInCirculation += gOmInUseCount;
1529 int deflatedcount = walk_monitor_list((ObjectMonitor **)&gOmInUseList, &FreeHead, &FreeTail);
1530 gOmInUseCount-= deflatedcount;
1531 nScavenged += deflatedcount;
1532 nInuse += gOmInUseCount;
1533 }
1535 } else for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) {
1536 // Iterate over all extant monitors - Scavenge all idle monitors.
1537 assert(block->object() == CHAINMARKER, "must be a block header");
1538 nInCirculation += _BLOCKSIZE ;
1539 for (int i = 1 ; i < _BLOCKSIZE; i++) {
1540 ObjectMonitor* mid = &block[i];
1541 oop obj = (oop) mid->object();
1543 if (obj == NULL) {
1544 // The monitor is not associated with an object.
1545 // The monitor should either be a thread-specific private
1546 // free list or the global free list.
1547 // obj == NULL IMPLIES mid->is_busy() == 0
1548 guarantee (!mid->is_busy(), "invariant") ;
1549 continue ;
1550 }
1551 deflated = deflate_monitor(mid, obj, &FreeHead, &FreeTail);
1553 if (deflated) {
1554 mid->FreeNext = NULL ;
1555 nScavenged ++ ;
1556 } else {
1557 nInuse ++;
1558 }
1559 }
1560 }
1562 MonitorFreeCount += nScavenged;
1564 // Consider: audit gFreeList to ensure that MonitorFreeCount and list agree.
1566 if (ObjectMonitor::Knob_Verbose) {
1567 ::printf ("Deflate: InCirc=%d InUse=%d Scavenged=%d ForceMonitorScavenge=%d : pop=%d free=%d\n",
1568 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge,
1569 MonitorPopulation, MonitorFreeCount) ;
1570 ::fflush(stdout) ;
1571 }
1573 ForceMonitorScavenge = 0; // Reset
1575 // Move the scavenged monitors back to the global free list.
1576 if (FreeHead != NULL) {
1577 guarantee (FreeTail != NULL && nScavenged > 0, "invariant") ;
1578 assert (FreeTail->FreeNext == NULL, "invariant") ;
1579 // constant-time list splice - prepend scavenged segment to gFreeList
1580 FreeTail->FreeNext = gFreeList ;
1581 gFreeList = FreeHead ;
1582 }
1583 Thread::muxRelease (&ListLock) ;
1585 if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged) ;
1586 if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation);
1588 // TODO: Add objectMonitor leak detection.
1589 // Audit/inventory the objectMonitors -- make sure they're all accounted for.
1590 GVars.stwRandom = os::random() ;
1591 GVars.stwCycle ++ ;
1592 }
1594 // Monitor cleanup on JavaThread::exit
1596 // Iterate through monitor cache and attempt to release thread's monitors
1597 // Gives up on a particular monitor if an exception occurs, but continues
1598 // the overall iteration, swallowing the exception.
1599 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1600 private:
1601 TRAPS;
1603 public:
1604 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {}
1605 void do_monitor(ObjectMonitor* mid) {
1606 if (mid->owner() == THREAD) {
1607 (void)mid->complete_exit(CHECK);
1608 }
1609 }
1610 };
1612 // Release all inflated monitors owned by THREAD. Lightweight monitors are
1613 // ignored. This is meant to be called during JNI thread detach which assumes
1614 // all remaining monitors are heavyweight. All exceptions are swallowed.
1615 // Scanning the extant monitor list can be time consuming.
1616 // A simple optimization is to add a per-thread flag that indicates a thread
1617 // called jni_monitorenter() during its lifetime.
1618 //
1619 // Instead of No_Savepoint_Verifier it might be cheaper to
1620 // use an idiom of the form:
1621 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1622 // <code that must not run at safepoint>
1623 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1624 // Since the tests are extremely cheap we could leave them enabled
1625 // for normal product builds.
1627 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) {
1628 assert(THREAD == JavaThread::current(), "must be current Java thread");
1629 No_Safepoint_Verifier nsv ;
1630 ReleaseJavaMonitorsClosure rjmc(THREAD);
1631 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread");
1632 ObjectSynchronizer::monitors_iterate(&rjmc);
1633 Thread::muxRelease(&ListLock);
1634 THREAD->clear_pending_exception();
1635 }
1637 //------------------------------------------------------------------------------
1638 // Non-product code
1640 #ifndef PRODUCT
1642 void ObjectSynchronizer::trace_locking(Handle locking_obj, bool is_compiled,
1643 bool is_method, bool is_locking) {
1644 // Don't know what to do here
1645 }
1647 // Verify all monitors in the monitor cache, the verification is weak.
1648 void ObjectSynchronizer::verify() {
1649 ObjectMonitor* block = gBlockList;
1650 ObjectMonitor* mid;
1651 while (block) {
1652 assert(block->object() == CHAINMARKER, "must be a block header");
1653 for (int i = 1; i < _BLOCKSIZE; i++) {
1654 mid = block + i;
1655 oop object = (oop) mid->object();
1656 if (object != NULL) {
1657 mid->verify();
1658 }
1659 }
1660 block = (ObjectMonitor*) block->FreeNext;
1661 }
1662 }
1664 // Check if monitor belongs to the monitor cache
1665 // The list is grow-only so it's *relatively* safe to traverse
1666 // the list of extant blocks without taking a lock.
1668 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) {
1669 ObjectMonitor* block = gBlockList;
1671 while (block) {
1672 assert(block->object() == CHAINMARKER, "must be a block header");
1673 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) {
1674 address mon = (address) monitor;
1675 address blk = (address) block;
1676 size_t diff = mon - blk;
1677 assert((diff % sizeof(ObjectMonitor)) == 0, "check");
1678 return 1;
1679 }
1680 block = (ObjectMonitor*) block->FreeNext;
1681 }
1682 return 0;
1683 }
1685 #endif