1 /*
2 * Copyright (c) 1998, 2011, 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 */
24
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
55 #if defined(__GNUC__) && !defined(IA64)
56 // Need to inhibit inlining for older versions of GCC to avoid build-time failures
57 #define ATTR __attribute__((noinline))
58 #else
59 #define ATTR
60 #endif
61
62 // The "core" versions of monitor enter and exit reside in this file.
63 // The interpreter and compilers contain specialized transliterated
64 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(),
65 // for instance. If you make changes here, make sure to modify the
66 // interpreter, and both C1 and C2 fast-path inline locking code emission.
67 //
68 //
69 // -----------------------------------------------------------------------------
70
71 #ifdef DTRACE_ENABLED
72
73 // Only bother with this argument setup if dtrace is available
74 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
75
76 HS_DTRACE_PROBE_DECL5(hotspot, monitor__wait,
77 jlong, uintptr_t, char*, int, long);
78 HS_DTRACE_PROBE_DECL4(hotspot, monitor__waited,
79 jlong, uintptr_t, char*, int);
80
81 #define DTRACE_MONITOR_PROBE_COMMON(klassOop, thread) \
82 char* bytes = NULL; \
83 int len = 0; \
84 jlong jtid = SharedRuntime::get_java_tid(thread); \
85 Symbol* klassname = ((oop)(klassOop))->klass()->klass_part()->name(); \
86 if (klassname != NULL) { \
87 bytes = (char*)klassname->bytes(); \
88 len = klassname->utf8_length(); \
89 }
90
91 #define DTRACE_MONITOR_WAIT_PROBE(monitor, klassOop, thread, millis) \
92 { \
93 if (DTraceMonitorProbes) { \
94 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \
95 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \
96 (monitor), bytes, len, (millis)); \
97 } \
98 }
99
100 #define DTRACE_MONITOR_PROBE(probe, monitor, klassOop, thread) \
101 { \
102 if (DTraceMonitorProbes) { \
103 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \
104 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \
105 (uintptr_t)(monitor), bytes, len); \
106 } \
107 }
108
109 #else // ndef DTRACE_ENABLED
110
111 #define DTRACE_MONITOR_WAIT_PROBE(klassOop, thread, millis, mon) {;}
112 #define DTRACE_MONITOR_PROBE(probe, klassOop, thread, mon) {;}
113
114 #endif // ndef DTRACE_ENABLED
115
116 // This exists only as a workaround of dtrace bug 6254741
117 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
118 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
119 return 0;
120 }
121
122 #define NINFLATIONLOCKS 256
123 static volatile intptr_t InflationLocks [NINFLATIONLOCKS] ;
124
125 ObjectMonitor * ObjectSynchronizer::gBlockList = NULL ;
126 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL ;
127 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL ;
128 int ObjectSynchronizer::gOmInUseCount = 0;
129 static volatile intptr_t ListLock = 0 ; // protects global monitor free-list cache
130 static volatile int MonitorFreeCount = 0 ; // # on gFreeList
131 static volatile int MonitorPopulation = 0 ; // # Extant -- in circulation
132 #define CHAINMARKER ((oop)-1)
133
134 // -----------------------------------------------------------------------------
135 // Fast Monitor Enter/Exit
136 // This the fast monitor enter. The interpreter and compiler use
137 // some assembly copies of this code. Make sure update those code
138 // if the following function is changed. The implementation is
139 // extremely sensitive to race condition. Be careful.
140
141 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) {
142 if (UseBiasedLocking) {
143 if (!SafepointSynchronize::is_at_safepoint()) {
144 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD);
145 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) {
146 return;
147 }
148 } else {
149 assert(!attempt_rebias, "can not rebias toward VM thread");
150 BiasedLocking::revoke_at_safepoint(obj);
151 }
152 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
153 }
154
155 slow_enter (obj, lock, THREAD) ;
156 }
157
158 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) {
159 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here");
160 // if displaced header is null, the previous enter is recursive enter, no-op
161 markOop dhw = lock->displaced_header();
162 markOop mark ;
163 if (dhw == NULL) {
164 // Recursive stack-lock.
165 // Diagnostics -- Could be: stack-locked, inflating, inflated.
166 mark = object->mark() ;
167 assert (!mark->is_neutral(), "invariant") ;
168 if (mark->has_locker() && mark != markOopDesc::INFLATING()) {
169 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant") ;
170 }
171 if (mark->has_monitor()) {
172 ObjectMonitor * m = mark->monitor() ;
173 assert(((oop)(m->object()))->mark() == mark, "invariant") ;
174 assert(m->is_entered(THREAD), "invariant") ;
175 }
176 return ;
177 }
178
179 mark = object->mark() ;
180
181 // If the object is stack-locked by the current thread, try to
182 // swing the displaced header from the box back to the mark.
183 if (mark == (markOop) lock) {
184 assert (dhw->is_neutral(), "invariant") ;
185 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) {
186 TEVENT (fast_exit: release stacklock) ;
187 return;
188 }
189 }
190
191 ObjectSynchronizer::inflate(THREAD, object)->exit (THREAD) ;
192 }
193
194 // -----------------------------------------------------------------------------
195 // Interpreter/Compiler Slow Case
196 // This routine is used to handle interpreter/compiler slow case
197 // We don't need to use fast path here, because it must have been
198 // failed in the interpreter/compiler code.
199 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
200 markOop mark = obj->mark();
201 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
202
203 if (mark->is_neutral()) {
204 // Anticipate successful CAS -- the ST of the displaced mark must
205 // be visible <= the ST performed by the CAS.
206 lock->set_displaced_header(mark);
207 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) {
208 TEVENT (slow_enter: release stacklock) ;
209 return ;
210 }
211 // Fall through to inflate() ...
212 } else
213 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
214 assert(lock != mark->locker(), "must not re-lock the same lock");
215 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
216 lock->set_displaced_header(NULL);
217 return;
218 }
219
220 #if 0
221 // The following optimization isn't particularly useful.
222 if (mark->has_monitor() && mark->monitor()->is_entered(THREAD)) {
223 lock->set_displaced_header (NULL) ;
224 return ;
225 }
226 #endif
227
228 // The object header will never be displaced to this lock,
229 // so it does not matter what the value is, except that it
230 // must be non-zero to avoid looking like a re-entrant lock,
231 // and must not look locked either.
232 lock->set_displaced_header(markOopDesc::unused_mark());
233 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
234 }
235
236 // This routine is used to handle interpreter/compiler slow case
237 // We don't need to use fast path here, because it must have
238 // failed in the interpreter/compiler code. Simply use the heavy
239 // weight monitor should be ok, unless someone find otherwise.
240 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
241 fast_exit (object, lock, THREAD) ;
242 }
243
244 // -----------------------------------------------------------------------------
245 // Class Loader support to workaround deadlocks on the class loader lock objects
246 // Also used by GC
247 // complete_exit()/reenter() are used to wait on a nested lock
248 // i.e. to give up an outer lock completely and then re-enter
249 // Used when holding nested locks - lock acquisition order: lock1 then lock2
250 // 1) complete_exit lock1 - saving recursion count
251 // 2) wait on lock2
252 // 3) when notified on lock2, unlock lock2
253 // 4) reenter lock1 with original recursion count
254 // 5) lock lock2
255 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
256 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
257 TEVENT (complete_exit) ;
258 if (UseBiasedLocking) {
259 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
260 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
261 }
262
263 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
264
265 return monitor->complete_exit(THREAD);
266 }
267
268 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
269 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
270 TEVENT (reenter) ;
271 if (UseBiasedLocking) {
272 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
273 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
274 }
275
276 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
277
278 monitor->reenter(recursion, THREAD);
279 }
280 // -----------------------------------------------------------------------------
281 // JNI locks on java objects
282 // NOTE: must use heavy weight monitor to handle jni monitor enter
283 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { // possible entry from jni enter
284 // the current locking is from JNI instead of Java code
285 TEVENT (jni_enter) ;
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 }
290 THREAD->set_current_pending_monitor_is_from_java(false);
291 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
292 THREAD->set_current_pending_monitor_is_from_java(true);
293 }
294
295 // NOTE: must use heavy weight monitor to handle jni monitor enter
296 bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) {
297 if (UseBiasedLocking) {
298 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
299 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
300 }
301
302 ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj());
303 return monitor->try_enter(THREAD);
304 }
305
306
307 // NOTE: must use heavy weight monitor to handle jni monitor exit
308 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
309 TEVENT (jni_exit) ;
310 if (UseBiasedLocking) {
311 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
312 }
313 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
314
315 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj);
316 // If this thread has locked the object, exit the monitor. Note: can't use
317 // monitor->check(CHECK); must exit even if an exception is pending.
318 if (monitor->check(THREAD)) {
319 monitor->exit(THREAD);
320 }
321 }
322
323 // -----------------------------------------------------------------------------
324 // Internal VM locks on java objects
325 // standard constructor, allows locking failures
326 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
327 _dolock = doLock;
328 _thread = thread;
329 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
330 _obj = obj;
331
332 if (_dolock) {
333 TEVENT (ObjectLocker) ;
334
335 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
336 }
337 }
338
339 ObjectLocker::~ObjectLocker() {
340 if (_dolock) {
341 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
342 }
343 }
344
345
346 // -----------------------------------------------------------------------------
347 // Wait/Notify/NotifyAll
348 // NOTE: must use heavy weight monitor to handle wait()
349 void ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
350 if (UseBiasedLocking) {
351 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
352 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
353 }
354 if (millis < 0) {
355 TEVENT (wait - throw IAX) ;
356 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
357 }
358 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
359 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
360 monitor->wait(millis, true, THREAD);
361
362 /* This dummy call is in place to get around dtrace bug 6254741. Once
363 that's fixed we can uncomment the following line and remove the call */
364 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
365 dtrace_waited_probe(monitor, obj, THREAD);
366 }
367
368 void ObjectSynchronizer::waitUninterruptibly (Handle obj, jlong millis, TRAPS) {
369 if (UseBiasedLocking) {
370 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
371 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
372 }
373 if (millis < 0) {
374 TEVENT (wait - throw IAX) ;
375 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
376 }
377 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD) ;
378 }
379
380 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
381 if (UseBiasedLocking) {
382 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
383 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
384 }
385
386 markOop mark = obj->mark();
387 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
388 return;
389 }
390 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD);
391 }
392
393 // NOTE: see comment of notify()
394 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
395 if (UseBiasedLocking) {
396 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
397 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
398 }
399
400 markOop mark = obj->mark();
401 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
402 return;
403 }
404 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD);
405 }
406
407 // -----------------------------------------------------------------------------
408 // Hash Code handling
409 //
410 // Performance concern:
411 // OrderAccess::storestore() calls release() which STs 0 into the global volatile
412 // OrderAccess::Dummy variable. This store is unnecessary for correctness.
413 // Many threads STing into a common location causes considerable cache migration
414 // or "sloshing" on large SMP system. As such, I avoid using OrderAccess::storestore()
415 // until it's repaired. In some cases OrderAccess::fence() -- which incurs local
416 // latency on the executing processor -- is a better choice as it scales on SMP
417 // systems. See http://blogs.sun.com/dave/entry/biased_locking_in_hotspot for a
418 // discussion of coherency costs. Note that all our current reference platforms
419 // provide strong ST-ST order, so the issue is moot on IA32, x64, and SPARC.
420 //
421 // As a general policy we use "volatile" to control compiler-based reordering
422 // and explicit fences (barriers) to control for architectural reordering performed
423 // by the CPU(s) or platform.
424
425 static int MBFence (int x) { OrderAccess::fence(); return x; }
426
427 struct SharedGlobals {
428 // These are highly shared mostly-read variables.
429 // To avoid false-sharing they need to be the sole occupants of a $ line.
430 double padPrefix [8];
431 volatile int stwRandom ;
432 volatile int stwCycle ;
433
434 // Hot RW variables -- Sequester to avoid false-sharing
435 double padSuffix [16];
436 volatile int hcSequence ;
437 double padFinal [8] ;
438 } ;
439
440 static SharedGlobals GVars ;
441 static int MonitorScavengeThreshold = 1000000 ;
442 static volatile int ForceMonitorScavenge = 0 ; // Scavenge required and pending
443
444 static markOop ReadStableMark (oop obj) {
445 markOop mark = obj->mark() ;
446 if (!mark->is_being_inflated()) {
447 return mark ; // normal fast-path return
448 }
449
450 int its = 0 ;
451 for (;;) {
452 markOop mark = obj->mark() ;
453 if (!mark->is_being_inflated()) {
454 return mark ; // normal fast-path return
455 }
456
457 // The object is being inflated by some other thread.
458 // The caller of ReadStableMark() must wait for inflation to complete.
459 // Avoid live-lock
460 // TODO: consider calling SafepointSynchronize::do_call_back() while
461 // spinning to see if there's a safepoint pending. If so, immediately
462 // yielding or blocking would be appropriate. Avoid spinning while
463 // there is a safepoint pending.
464 // TODO: add inflation contention performance counters.
465 // TODO: restrict the aggregate number of spinners.
466
467 ++its ;
468 if (its > 10000 || !os::is_MP()) {
469 if (its & 1) {
470 os::NakedYield() ;
471 TEVENT (Inflate: INFLATING - yield) ;
472 } else {
473 // Note that the following code attenuates the livelock problem but is not
474 // a complete remedy. A more complete solution would require that the inflating
475 // thread hold the associated inflation lock. The following code simply restricts
476 // the number of spinners to at most one. We'll have N-2 threads blocked
477 // on the inflationlock, 1 thread holding the inflation lock and using
478 // a yield/park strategy, and 1 thread in the midst of inflation.
479 // A more refined approach would be to change the encoding of INFLATING
480 // to allow encapsulation of a native thread pointer. Threads waiting for
481 // inflation to complete would use CAS to push themselves onto a singly linked
482 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
483 // and calling park(). When inflation was complete the thread that accomplished inflation
484 // would detach the list and set the markword to inflated with a single CAS and
485 // then for each thread on the list, set the flag and unpark() the thread.
486 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease
487 // wakes at most one thread whereas we need to wake the entire list.
488 int ix = (intptr_t(obj) >> 5) & (NINFLATIONLOCKS-1) ;
489 int YieldThenBlock = 0 ;
490 assert (ix >= 0 && ix < NINFLATIONLOCKS, "invariant") ;
491 assert ((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant") ;
492 Thread::muxAcquire (InflationLocks + ix, "InflationLock") ;
493 while (obj->mark() == markOopDesc::INFLATING()) {
494 // Beware: NakedYield() is advisory and has almost no effect on some platforms
495 // so we periodically call Self->_ParkEvent->park(1).
496 // We use a mixed spin/yield/block mechanism.
497 if ((YieldThenBlock++) >= 16) {
498 Thread::current()->_ParkEvent->park(1) ;
499 } else {
500 os::NakedYield() ;
501 }
502 }
503 Thread::muxRelease (InflationLocks + ix ) ;
504 TEVENT (Inflate: INFLATING - yield/park) ;
505 }
506 } else {
507 SpinPause() ; // SMP-polite spinning
508 }
509 }
510 }
511
512 // hashCode() generation :
513 //
514 // Possibilities:
515 // * MD5Digest of {obj,stwRandom}
516 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function.
517 // * A DES- or AES-style SBox[] mechanism
518 // * One of the Phi-based schemes, such as:
519 // 2654435761 = 2^32 * Phi (golden ratio)
520 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ;
521 // * A variation of Marsaglia's shift-xor RNG scheme.
522 // * (obj ^ stwRandom) is appealing, but can result
523 // in undesirable regularity in the hashCode values of adjacent objects
524 // (objects allocated back-to-back, in particular). This could potentially
525 // result in hashtable collisions and reduced hashtable efficiency.
526 // There are simple ways to "diffuse" the middle address bits over the
527 // generated hashCode values:
528 //
529
530 static inline intptr_t get_next_hash(Thread * Self, oop obj) {
531 intptr_t value = 0 ;
532 if (hashCode == 0) {
533 // This form uses an unguarded global Park-Miller RNG,
534 // so it's possible for two threads to race and generate the same RNG.
535 // On MP system we'll have lots of RW access to a global, so the
536 // mechanism induces lots of coherency traffic.
537 value = os::random() ;
538 } else
539 if (hashCode == 1) {
540 // This variation has the property of being stable (idempotent)
541 // between STW operations. This can be useful in some of the 1-0
542 // synchronization schemes.
543 intptr_t addrBits = intptr_t(obj) >> 3 ;
544 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom ;
545 } else
546 if (hashCode == 2) {
547 value = 1 ; // for sensitivity testing
548 } else
549 if (hashCode == 3) {
550 value = ++GVars.hcSequence ;
551 } else
552 if (hashCode == 4) {
553 value = intptr_t(obj) ;
554 } else {
555 // Marsaglia's xor-shift scheme with thread-specific state
556 // This is probably the best overall implementation -- we'll
557 // likely make this the default in future releases.
558 unsigned t = Self->_hashStateX ;
559 t ^= (t << 11) ;
560 Self->_hashStateX = Self->_hashStateY ;
561 Self->_hashStateY = Self->_hashStateZ ;
562 Self->_hashStateZ = Self->_hashStateW ;
563 unsigned v = Self->_hashStateW ;
564 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)) ;
565 Self->_hashStateW = v ;
566 value = v ;
567 }
568
569 value &= markOopDesc::hash_mask;
570 if (value == 0) value = 0xBAD ;
571 assert (value != markOopDesc::no_hash, "invariant") ;
572 TEVENT (hashCode: GENERATE) ;
573 return value;
574 }
575 //
576 intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) {
577 if (UseBiasedLocking) {
578 // NOTE: many places throughout the JVM do not expect a safepoint
579 // to be taken here, in particular most operations on perm gen
580 // objects. However, we only ever bias Java instances and all of
581 // the call sites of identity_hash that might revoke biases have
582 // been checked to make sure they can handle a safepoint. The
583 // added check of the bias pattern is to avoid useless calls to
584 // thread-local storage.
585 if (obj->mark()->has_bias_pattern()) {
586 // Box and unbox the raw reference just in case we cause a STW safepoint.
587 Handle hobj (Self, obj) ;
588 // Relaxing assertion for bug 6320749.
589 assert (Universe::verify_in_progress() ||
590 !SafepointSynchronize::is_at_safepoint(),
591 "biases should not be seen by VM thread here");
592 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
593 obj = hobj() ;
594 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
595 }
596 }
597
598 // hashCode() is a heap mutator ...
599 // Relaxing assertion for bug 6320749.
600 assert (Universe::verify_in_progress() ||
601 !SafepointSynchronize::is_at_safepoint(), "invariant") ;
602 assert (Universe::verify_in_progress() ||
603 Self->is_Java_thread() , "invariant") ;
604 assert (Universe::verify_in_progress() ||
605 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ;
606
607 ObjectMonitor* monitor = NULL;
608 markOop temp, test;
609 intptr_t hash;
610 markOop mark = ReadStableMark (obj);
611
612 // object should remain ineligible for biased locking
613 assert (!mark->has_bias_pattern(), "invariant") ;
614
615 if (mark->is_neutral()) {
616 hash = mark->hash(); // this is a normal header
617 if (hash) { // if it has hash, just return it
618 return hash;
619 }
620 hash = get_next_hash(Self, obj); // allocate a new hash code
621 temp = mark->copy_set_hash(hash); // merge the hash code into header
622 // use (machine word version) atomic operation to install the hash
623 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark);
624 if (test == mark) {
625 return hash;
626 }
627 // If atomic operation failed, we must inflate the header
628 // into heavy weight monitor. We could add more code here
629 // for fast path, but it does not worth the complexity.
630 } else if (mark->has_monitor()) {
631 monitor = mark->monitor();
632 temp = monitor->header();
633 assert (temp->is_neutral(), "invariant") ;
634 hash = temp->hash();
635 if (hash) {
636 return hash;
637 }
638 // Skip to the following code to reduce code size
639 } else if (Self->is_lock_owned((address)mark->locker())) {
640 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
641 assert (temp->is_neutral(), "invariant") ;
642 hash = temp->hash(); // by current thread, check if the displaced
643 if (hash) { // header contains hash code
644 return hash;
645 }
646 // WARNING:
647 // The displaced header is strictly immutable.
648 // It can NOT be changed in ANY cases. So we have
649 // to inflate the header into heavyweight monitor
650 // even the current thread owns the lock. The reason
651 // is the BasicLock (stack slot) will be asynchronously
652 // read by other threads during the inflate() function.
653 // Any change to stack may not propagate to other threads
654 // correctly.
655 }
656
657 // Inflate the monitor to set hash code
658 monitor = ObjectSynchronizer::inflate(Self, obj);
659 // Load displaced header and check it has hash code
660 mark = monitor->header();
661 assert (mark->is_neutral(), "invariant") ;
662 hash = mark->hash();
663 if (hash == 0) {
664 hash = get_next_hash(Self, obj);
665 temp = mark->copy_set_hash(hash); // merge hash code into header
666 assert (temp->is_neutral(), "invariant") ;
667 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark);
668 if (test != mark) {
669 // The only update to the header in the monitor (outside GC)
670 // is install the hash code. If someone add new usage of
671 // displaced header, please update this code
672 hash = test->hash();
673 assert (test->is_neutral(), "invariant") ;
674 assert (hash != 0, "Trivial unexpected object/monitor header usage.");
675 }
676 }
677 // We finally get the hash
678 return hash;
679 }
680
681 // Deprecated -- use FastHashCode() instead.
682
683 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
684 return FastHashCode (Thread::current(), obj()) ;
685 }
686
687
688 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
689 Handle h_obj) {
690 if (UseBiasedLocking) {
691 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
692 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
693 }
694
695 assert(thread == JavaThread::current(), "Can only be called on current thread");
696 oop obj = h_obj();
697
698 markOop mark = ReadStableMark (obj) ;
699
700 // Uncontended case, header points to stack
701 if (mark->has_locker()) {
702 return thread->is_lock_owned((address)mark->locker());
703 }
704 // Contended case, header points to ObjectMonitor (tagged pointer)
705 if (mark->has_monitor()) {
706 ObjectMonitor* monitor = mark->monitor();
707 return monitor->is_entered(thread) != 0 ;
708 }
709 // Unlocked case, header in place
710 assert(mark->is_neutral(), "sanity check");
711 return false;
712 }
713
714 // Be aware of this method could revoke bias of the lock object.
715 // This method querys the ownership of the lock handle specified by 'h_obj'.
716 // If the current thread owns the lock, it returns owner_self. If no
717 // thread owns the lock, it returns owner_none. Otherwise, it will return
718 // ower_other.
719 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
720 (JavaThread *self, Handle h_obj) {
721 // The caller must beware this method can revoke bias, and
722 // revocation can result in a safepoint.
723 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ;
724 assert (self->thread_state() != _thread_blocked , "invariant") ;
725
726 // Possible mark states: neutral, biased, stack-locked, inflated
727
728 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
729 // CASE: biased
730 BiasedLocking::revoke_and_rebias(h_obj, false, self);
731 assert(!h_obj->mark()->has_bias_pattern(),
732 "biases should be revoked by now");
733 }
734
735 assert(self == JavaThread::current(), "Can only be called on current thread");
736 oop obj = h_obj();
737 markOop mark = ReadStableMark (obj) ;
738
739 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
740 if (mark->has_locker()) {
741 return self->is_lock_owned((address)mark->locker()) ?
742 owner_self : owner_other;
743 }
744
745 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor.
746 // The Object:ObjectMonitor relationship is stable as long as we're
747 // not at a safepoint.
748 if (mark->has_monitor()) {
749 void * owner = mark->monitor()->_owner ;
750 if (owner == NULL) return owner_none ;
751 return (owner == self ||
752 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
753 }
754
755 // CASE: neutral
756 assert(mark->is_neutral(), "sanity check");
757 return owner_none ; // it's unlocked
758 }
759
760 // FIXME: jvmti should call this
761 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) {
762 if (UseBiasedLocking) {
763 if (SafepointSynchronize::is_at_safepoint()) {
764 BiasedLocking::revoke_at_safepoint(h_obj);
765 } else {
766 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
767 }
768 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
769 }
770
771 oop obj = h_obj();
772 address owner = NULL;
773
774 markOop mark = ReadStableMark (obj) ;
775
776 // Uncontended case, header points to stack
777 if (mark->has_locker()) {
778 owner = (address) mark->locker();
779 }
780
781 // Contended case, header points to ObjectMonitor (tagged pointer)
782 if (mark->has_monitor()) {
783 ObjectMonitor* monitor = mark->monitor();
784 assert(monitor != NULL, "monitor should be non-null");
785 owner = (address) monitor->owner();
786 }
787
788 if (owner != NULL) {
789 return Threads::owning_thread_from_monitor_owner(owner, doLock);
790 }
791
792 // Unlocked case, header in place
793 // Cannot have assertion since this object may have been
794 // locked by another thread when reaching here.
795 // assert(mark->is_neutral(), "sanity check");
796
797 return NULL;
798 }
799 // Visitors ...
800
801 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
802 ObjectMonitor* block = gBlockList;
803 ObjectMonitor* mid;
804 while (block) {
805 assert(block->object() == CHAINMARKER, "must be a block header");
806 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
807 mid = block + i;
808 oop object = (oop) mid->object();
809 if (object != NULL) {
810 closure->do_monitor(mid);
811 }
812 }
813 block = (ObjectMonitor*) block->FreeNext;
814 }
815 }
816
817 // Get the next block in the block list.
818 static inline ObjectMonitor* next(ObjectMonitor* block) {
819 assert(block->object() == CHAINMARKER, "must be a block header");
820 block = block->FreeNext ;
821 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
822 return block;
823 }
824
825
826 void ObjectSynchronizer::oops_do(OopClosure* f) {
827 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
828 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) {
829 assert(block->object() == CHAINMARKER, "must be a block header");
830 for (int i = 1; i < _BLOCKSIZE; i++) {
831 ObjectMonitor* mid = &block[i];
832 if (mid->object() != NULL) {
833 f->do_oop((oop*)mid->object_addr());
834 }
835 }
836 }
837 }
838
839
840 // -----------------------------------------------------------------------------
841 // ObjectMonitor Lifecycle
842 // -----------------------
843 // Inflation unlinks monitors from the global gFreeList and
844 // associates them with objects. Deflation -- which occurs at
845 // STW-time -- disassociates idle monitors from objects. Such
846 // scavenged monitors are returned to the gFreeList.
847 //
848 // The global list is protected by ListLock. All the critical sections
849 // are short and operate in constant-time.
850 //
851 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
852 //
853 // Lifecycle:
854 // -- unassigned and on the global free list
855 // -- unassigned and on a thread's private omFreeList
856 // -- assigned to an object. The object is inflated and the mark refers
857 // to the objectmonitor.
858 //
859
860
861 // Constraining monitor pool growth via MonitorBound ...
862 //
863 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
864 // the rate of scavenging is driven primarily by GC. As such, we can find
865 // an inordinate number of monitors in circulation.
866 // To avoid that scenario we can artificially induce a STW safepoint
867 // if the pool appears to be growing past some reasonable bound.
868 // Generally we favor time in space-time tradeoffs, but as there's no
869 // natural back-pressure on the # of extant monitors we need to impose some
870 // type of limit. Beware that if MonitorBound is set to too low a value
871 // we could just loop. In addition, if MonitorBound is set to a low value
872 // we'll incur more safepoints, which are harmful to performance.
873 // See also: GuaranteedSafepointInterval
874 //
875 // The current implementation uses asynchronous VM operations.
876 //
877
878 static void InduceScavenge (Thread * Self, const char * Whence) {
879 // Induce STW safepoint to trim monitors
880 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
881 // More precisely, trigger an asynchronous STW safepoint as the number
882 // of active monitors passes the specified threshold.
883 // TODO: assert thread state is reasonable
884
885 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
886 if (ObjectMonitor::Knob_Verbose) {
887 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ;
888 ::fflush(stdout) ;
889 }
890 // Induce a 'null' safepoint to scavenge monitors
891 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
892 // to the VMthread and have a lifespan longer than that of this activation record.
893 // The VMThread will delete the op when completed.
894 VMThread::execute (new VM_ForceAsyncSafepoint()) ;
895
896 if (ObjectMonitor::Knob_Verbose) {
897 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ;
898 ::fflush(stdout) ;
899 }
900 }
901 }
902 /* Too slow for general assert or debug
903 void ObjectSynchronizer::verifyInUse (Thread *Self) {
904 ObjectMonitor* mid;
905 int inusetally = 0;
906 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) {
907 inusetally ++;
908 }
909 assert(inusetally == Self->omInUseCount, "inuse count off");
910
911 int freetally = 0;
912 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) {
913 freetally ++;
914 }
915 assert(freetally == Self->omFreeCount, "free count off");
916 }
917 */
918 ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) {
919 // A large MAXPRIVATE value reduces both list lock contention
920 // and list coherency traffic, but also tends to increase the
921 // number of objectMonitors in circulation as well as the STW
922 // scavenge costs. As usual, we lean toward time in space-time
923 // tradeoffs.
924 const int MAXPRIVATE = 1024 ;
925 for (;;) {
926 ObjectMonitor * m ;
927
928 // 1: try to allocate from the thread's local omFreeList.
929 // Threads will attempt to allocate first from their local list, then
930 // from the global list, and only after those attempts fail will the thread
931 // attempt to instantiate new monitors. Thread-local free lists take
932 // heat off the ListLock and improve allocation latency, as well as reducing
933 // coherency traffic on the shared global list.
934 m = Self->omFreeList ;
935 if (m != NULL) {
936 Self->omFreeList = m->FreeNext ;
937 Self->omFreeCount -- ;
938 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene
939 guarantee (m->object() == NULL, "invariant") ;
940 if (MonitorInUseLists) {
941 m->FreeNext = Self->omInUseList;
942 Self->omInUseList = m;
943 Self->omInUseCount ++;
944 // verifyInUse(Self);
945 } else {
946 m->FreeNext = NULL;
947 }
948 return m ;
949 }
950
951 // 2: try to allocate from the global gFreeList
952 // CONSIDER: use muxTry() instead of muxAcquire().
953 // If the muxTry() fails then drop immediately into case 3.
954 // If we're using thread-local free lists then try
955 // to reprovision the caller's free list.
956 if (gFreeList != NULL) {
957 // Reprovision the thread's omFreeList.
958 // Use bulk transfers to reduce the allocation rate and heat
959 // on various locks.
960 Thread::muxAcquire (&ListLock, "omAlloc") ;
961 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL; ) {
962 MonitorFreeCount --;
963 ObjectMonitor * take = gFreeList ;
964 gFreeList = take->FreeNext ;
965 guarantee (take->object() == NULL, "invariant") ;
966 guarantee (!take->is_busy(), "invariant") ;
967 take->Recycle() ;
968 omRelease (Self, take, false) ;
969 }
970 Thread::muxRelease (&ListLock) ;
971 Self->omFreeProvision += 1 + (Self->omFreeProvision/2) ;
972 if (Self->omFreeProvision > MAXPRIVATE ) Self->omFreeProvision = MAXPRIVATE ;
973 TEVENT (omFirst - reprovision) ;
974
975 const int mx = MonitorBound ;
976 if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) {
977 // We can't safely induce a STW safepoint from omAlloc() as our thread
978 // state may not be appropriate for such activities and callers may hold
979 // naked oops, so instead we defer the action.
980 InduceScavenge (Self, "omAlloc") ;
981 }
982 continue;
983 }
984
985 // 3: allocate a block of new ObjectMonitors
986 // Both the local and global free lists are empty -- resort to malloc().
987 // In the current implementation objectMonitors are TSM - immortal.
988 assert (_BLOCKSIZE > 1, "invariant") ;
989 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE];
990
991 // NOTE: (almost) no way to recover if allocation failed.
992 // We might be able to induce a STW safepoint and scavenge enough
993 // objectMonitors to permit progress.
994 if (temp == NULL) {
995 vm_exit_out_of_memory (sizeof (ObjectMonitor[_BLOCKSIZE]), "Allocate ObjectMonitors") ;
996 }
997
998 // Format the block.
999 // initialize the linked list, each monitor points to its next
1000 // forming the single linked free list, the very first monitor
1001 // will points to next block, which forms the block list.
1002 // The trick of using the 1st element in the block as gBlockList
1003 // linkage should be reconsidered. A better implementation would
1004 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1005
1006 for (int i = 1; i < _BLOCKSIZE ; i++) {
1007 temp[i].FreeNext = &temp[i+1];
1008 }
1009
1010 // terminate the last monitor as the end of list
1011 temp[_BLOCKSIZE - 1].FreeNext = NULL ;
1012
1013 // Element [0] is reserved for global list linkage
1014 temp[0].set_object(CHAINMARKER);
1015
1016 // Consider carving out this thread's current request from the
1017 // block in hand. This avoids some lock traffic and redundant
1018 // list activity.
1019
1020 // Acquire the ListLock to manipulate BlockList and FreeList.
1021 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1022 Thread::muxAcquire (&ListLock, "omAlloc [2]") ;
1023 MonitorPopulation += _BLOCKSIZE-1;
1024 MonitorFreeCount += _BLOCKSIZE-1;
1025
1026 // Add the new block to the list of extant blocks (gBlockList).
1027 // The very first objectMonitor in a block is reserved and dedicated.
1028 // It serves as blocklist "next" linkage.
1029 temp[0].FreeNext = gBlockList;
1030 gBlockList = temp;
1031
1032 // Add the new string of objectMonitors to the global free list
1033 temp[_BLOCKSIZE - 1].FreeNext = gFreeList ;
1034 gFreeList = temp + 1;
1035 Thread::muxRelease (&ListLock) ;
1036 TEVENT (Allocate block of monitors) ;
1037 }
1038 }
1039
1040 // Place "m" on the caller's private per-thread omFreeList.
1041 // In practice there's no need to clamp or limit the number of
1042 // monitors on a thread's omFreeList as the only time we'll call
1043 // omRelease is to return a monitor to the free list after a CAS
1044 // attempt failed. This doesn't allow unbounded #s of monitors to
1045 // accumulate on a thread's free list.
1046 //
1047
1048 void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m, bool fromPerThreadAlloc) {
1049 guarantee (m->object() == NULL, "invariant") ;
1050
1051 // Remove from omInUseList
1052 if (MonitorInUseLists && fromPerThreadAlloc) {
1053 ObjectMonitor* curmidinuse = NULL;
1054 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; ) {
1055 if (m == mid) {
1056 // extract from per-thread in-use-list
1057 if (mid == Self->omInUseList) {
1058 Self->omInUseList = mid->FreeNext;
1059 } else if (curmidinuse != NULL) {
1060 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist
1061 }
1062 Self->omInUseCount --;
1063 // verifyInUse(Self);
1064 break;
1065 } else {
1066 curmidinuse = mid;
1067 mid = mid->FreeNext;
1068 }
1069 }
1070 }
1071
1072 // FreeNext is used for both onInUseList and omFreeList, so clear old before setting new
1073 m->FreeNext = Self->omFreeList ;
1074 Self->omFreeList = m ;
1075 Self->omFreeCount ++ ;
1076 }
1077
1078 // Return the monitors of a moribund thread's local free list to
1079 // the global free list. Typically a thread calls omFlush() when
1080 // it's dying. We could also consider having the VM thread steal
1081 // monitors from threads that have not run java code over a few
1082 // consecutive STW safepoints. Relatedly, we might decay
1083 // omFreeProvision at STW safepoints.
1084 //
1085 // Also return the monitors of a moribund thread"s omInUseList to
1086 // a global gOmInUseList under the global list lock so these
1087 // will continue to be scanned.
1088 //
1089 // We currently call omFlush() from the Thread:: dtor _after the thread
1090 // has been excised from the thread list and is no longer a mutator.
1091 // That means that omFlush() can run concurrently with a safepoint and
1092 // the scavenge operator. Calling omFlush() from JavaThread::exit() might
1093 // be a better choice as we could safely reason that that the JVM is
1094 // not at a safepoint at the time of the call, and thus there could
1095 // be not inopportune interleavings between omFlush() and the scavenge
1096 // operator.
1097
1098 void ObjectSynchronizer::omFlush (Thread * Self) {
1099 ObjectMonitor * List = Self->omFreeList ; // Null-terminated SLL
1100 Self->omFreeList = NULL ;
1101 ObjectMonitor * Tail = NULL ;
1102 int Tally = 0;
1103 if (List != NULL) {
1104 ObjectMonitor * s ;
1105 for (s = List ; s != NULL ; s = s->FreeNext) {
1106 Tally ++ ;
1107 Tail = s ;
1108 guarantee (s->object() == NULL, "invariant") ;
1109 guarantee (!s->is_busy(), "invariant") ;
1110 s->set_owner (NULL) ; // redundant but good hygiene
1111 TEVENT (omFlush - Move one) ;
1112 }
1113 guarantee (Tail != NULL && List != NULL, "invariant") ;
1114 }
1115
1116 ObjectMonitor * InUseList = Self->omInUseList;
1117 ObjectMonitor * InUseTail = NULL ;
1118 int InUseTally = 0;
1119 if (InUseList != NULL) {
1120 Self->omInUseList = NULL;
1121 ObjectMonitor *curom;
1122 for (curom = InUseList; curom != NULL; curom = curom->FreeNext) {
1123 InUseTail = curom;
1124 InUseTally++;
1125 }
1126 // TODO debug
1127 assert(Self->omInUseCount == InUseTally, "inuse count off");
1128 Self->omInUseCount = 0;
1129 guarantee (InUseTail != NULL && InUseList != NULL, "invariant");
1130 }
1131
1132 Thread::muxAcquire (&ListLock, "omFlush") ;
1133 if (Tail != NULL) {
1134 Tail->FreeNext = gFreeList ;
1135 gFreeList = List ;
1136 MonitorFreeCount += Tally;
1137 }
1138
1139 if (InUseTail != NULL) {
1140 InUseTail->FreeNext = gOmInUseList;
1141 gOmInUseList = InUseList;
1142 gOmInUseCount += InUseTally;
1143 }
1144
1145 Thread::muxRelease (&ListLock) ;
1146 TEVENT (omFlush) ;
1147 }
1148
1149 // Fast path code shared by multiple functions
1150 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) {
1151 markOop mark = obj->mark();
1152 if (mark->has_monitor()) {
1153 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid");
1154 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header");
1155 return mark->monitor();
1156 }
1157 return ObjectSynchronizer::inflate(Thread::current(), obj);
1158 }
1159
1160
1161 // Note that we could encounter some performance loss through false-sharing as
1162 // multiple locks occupy the same $ line. Padding might be appropriate.
1163
1164
1165 ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) {
1166 // Inflate mutates the heap ...
1167 // Relaxing assertion for bug 6320749.
1168 assert (Universe::verify_in_progress() ||
1169 !SafepointSynchronize::is_at_safepoint(), "invariant") ;
1170
1171 for (;;) {
1172 const markOop mark = object->mark() ;
1173 assert (!mark->has_bias_pattern(), "invariant") ;
1174
1175 // The mark can be in one of the following states:
1176 // * Inflated - just return
1177 // * Stack-locked - coerce it to inflated
1178 // * INFLATING - busy wait for conversion to complete
1179 // * Neutral - aggressively inflate the object.
1180 // * BIASED - Illegal. We should never see this
1181
1182 // CASE: inflated
1183 if (mark->has_monitor()) {
1184 ObjectMonitor * inf = mark->monitor() ;
1185 assert (inf->header()->is_neutral(), "invariant");
1186 assert (inf->object() == object, "invariant") ;
1187 assert (ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1188 return inf ;
1189 }
1190
1191 // CASE: inflation in progress - inflating over a stack-lock.
1192 // Some other thread is converting from stack-locked to inflated.
1193 // Only that thread can complete inflation -- other threads must wait.
1194 // The INFLATING value is transient.
1195 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1196 // We could always eliminate polling by parking the thread on some auxiliary list.
1197 if (mark == markOopDesc::INFLATING()) {
1198 TEVENT (Inflate: spin while INFLATING) ;
1199 ReadStableMark(object) ;
1200 continue ;
1201 }
1202
1203 // CASE: stack-locked
1204 // Could be stack-locked either by this thread or by some other thread.
1205 //
1206 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1207 // to install INFLATING into the mark word. We originally installed INFLATING,
1208 // allocated the objectmonitor, and then finally STed the address of the
1209 // objectmonitor into the mark. This was correct, but artificially lengthened
1210 // the interval in which INFLATED appeared in the mark, thus increasing
1211 // the odds of inflation contention.
1212 //
1213 // We now use per-thread private objectmonitor free lists.
1214 // These list are reprovisioned from the global free list outside the
1215 // critical INFLATING...ST interval. A thread can transfer
1216 // multiple objectmonitors en-mass from the global free list to its local free list.
1217 // This reduces coherency traffic and lock contention on the global free list.
1218 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1219 // before or after the CAS(INFLATING) operation.
1220 // See the comments in omAlloc().
1221
1222 if (mark->has_locker()) {
1223 ObjectMonitor * m = omAlloc (Self) ;
1224 // Optimistically prepare the objectmonitor - anticipate successful CAS
1225 // We do this before the CAS in order to minimize the length of time
1226 // in which INFLATING appears in the mark.
1227 m->Recycle();
1228 m->_Responsible = NULL ;
1229 m->OwnerIsThread = 0 ;
1230 m->_recursions = 0 ;
1231 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // Consider: maintain by type/class
1232
1233 markOop cmp = (markOop) Atomic::cmpxchg_ptr (markOopDesc::INFLATING(), object->mark_addr(), mark) ;
1234 if (cmp != mark) {
1235 omRelease (Self, m, true) ;
1236 continue ; // Interference -- just retry
1237 }
1238
1239 // We've successfully installed INFLATING (0) into the mark-word.
1240 // This is the only case where 0 will appear in a mark-work.
1241 // Only the singular thread that successfully swings the mark-word
1242 // to 0 can perform (or more precisely, complete) inflation.
1243 //
1244 // Why do we CAS a 0 into the mark-word instead of just CASing the
1245 // mark-word from the stack-locked value directly to the new inflated state?
1246 // Consider what happens when a thread unlocks a stack-locked object.
1247 // It attempts to use CAS to swing the displaced header value from the
1248 // on-stack basiclock back into the object header. Recall also that the
1249 // header value (hashcode, etc) can reside in (a) the object header, or
1250 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1251 // header in an objectMonitor. The inflate() routine must copy the header
1252 // value from the basiclock on the owner's stack to the objectMonitor, all
1253 // the while preserving the hashCode stability invariants. If the owner
1254 // decides to release the lock while the value is 0, the unlock will fail
1255 // and control will eventually pass from slow_exit() to inflate. The owner
1256 // will then spin, waiting for the 0 value to disappear. Put another way,
1257 // the 0 causes the owner to stall if the owner happens to try to
1258 // drop the lock (restoring the header from the basiclock to the object)
1259 // while inflation is in-progress. This protocol avoids races that might
1260 // would otherwise permit hashCode values to change or "flicker" for an object.
1261 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1262 // 0 serves as a "BUSY" inflate-in-progress indicator.
1263
1264
1265 // fetch the displaced mark from the owner's stack.
1266 // The owner can't die or unwind past the lock while our INFLATING
1267 // object is in the mark. Furthermore the owner can't complete
1268 // an unlock on the object, either.
1269 markOop dmw = mark->displaced_mark_helper() ;
1270 assert (dmw->is_neutral(), "invariant") ;
1271
1272 // Setup monitor fields to proper values -- prepare the monitor
1273 m->set_header(dmw) ;
1274
1275 // Optimization: if the mark->locker stack address is associated
1276 // with this thread we could simply set m->_owner = Self and
1277 // m->OwnerIsThread = 1. Note that a thread can inflate an object
1278 // that it has stack-locked -- as might happen in wait() -- directly
1279 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1280 m->set_owner(mark->locker());
1281 m->set_object(object);
1282 // TODO-FIXME: assert BasicLock->dhw != 0.
1283
1284 // Must preserve store ordering. The monitor state must
1285 // be stable at the time of publishing the monitor address.
1286 guarantee (object->mark() == markOopDesc::INFLATING(), "invariant") ;
1287 object->release_set_mark(markOopDesc::encode(m));
1288
1289 // Hopefully the performance counters are allocated on distinct cache lines
1290 // to avoid false sharing on MP systems ...
1291 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ;
1292 TEVENT(Inflate: overwrite stacklock) ;
1293 if (TraceMonitorInflation) {
1294 if (object->is_instance()) {
1295 ResourceMark rm;
1296 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1297 (intptr_t) object, (intptr_t) object->mark(),
1298 Klass::cast(object->klass())->external_name());
1299 }
1300 }
1301 return m ;
1302 }
1303
1304 // CASE: neutral
1305 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1306 // If we know we're inflating for entry it's better to inflate by swinging a
1307 // pre-locked objectMonitor pointer into the object header. A successful
1308 // CAS inflates the object *and* confers ownership to the inflating thread.
1309 // In the current implementation we use a 2-step mechanism where we CAS()
1310 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1311 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1312 // would be useful.
1313
1314 assert (mark->is_neutral(), "invariant");
1315 ObjectMonitor * m = omAlloc (Self) ;
1316 // prepare m for installation - set monitor to initial state
1317 m->Recycle();
1318 m->set_header(mark);
1319 m->set_owner(NULL);
1320 m->set_object(object);
1321 m->OwnerIsThread = 1 ;
1322 m->_recursions = 0 ;
1323 m->_Responsible = NULL ;
1324 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // consider: keep metastats by type/class
1325
1326 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) {
1327 m->set_object (NULL) ;
1328 m->set_owner (NULL) ;
1329 m->OwnerIsThread = 0 ;
1330 m->Recycle() ;
1331 omRelease (Self, m, true) ;
1332 m = NULL ;
1333 continue ;
1334 // interference - the markword changed - just retry.
1335 // The state-transitions are one-way, so there's no chance of
1336 // live-lock -- "Inflated" is an absorbing state.
1337 }
1338
1339 // Hopefully the performance counters are allocated on distinct
1340 // cache lines to avoid false sharing on MP systems ...
1341 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ;
1342 TEVENT(Inflate: overwrite neutral) ;
1343 if (TraceMonitorInflation) {
1344 if (object->is_instance()) {
1345 ResourceMark rm;
1346 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1347 (intptr_t) object, (intptr_t) object->mark(),
1348 Klass::cast(object->klass())->external_name());
1349 }
1350 }
1351 return m ;
1352 }
1353 }
1354
1355 // Note that we could encounter some performance loss through false-sharing as
1356 // multiple locks occupy the same $ line. Padding might be appropriate.
1357
1358
1359 // Deflate_idle_monitors() is called at all safepoints, immediately
1360 // after all mutators are stopped, but before any objects have moved.
1361 // It traverses the list of known monitors, deflating where possible.
1362 // The scavenged monitor are returned to the monitor free list.
1363 //
1364 // Beware that we scavenge at *every* stop-the-world point.
1365 // Having a large number of monitors in-circulation negatively
1366 // impacts the performance of some applications (e.g., PointBase).
1367 // Broadly, we want to minimize the # of monitors in circulation.
1368 //
1369 // We have added a flag, MonitorInUseLists, which creates a list
1370 // of active monitors for each thread. deflate_idle_monitors()
1371 // only scans the per-thread inuse lists. omAlloc() puts all
1372 // assigned monitors on the per-thread list. deflate_idle_monitors()
1373 // returns the non-busy monitors to the global free list.
1374 // When a thread dies, omFlush() adds the list of active monitors for
1375 // that thread to a global gOmInUseList acquiring the
1376 // global list lock. deflate_idle_monitors() acquires the global
1377 // list lock to scan for non-busy monitors to the global free list.
1378 // An alternative could have used a single global inuse list. The
1379 // downside would have been the additional cost of acquiring the global list lock
1380 // for every omAlloc().
1381 //
1382 // Perversely, the heap size -- and thus the STW safepoint rate --
1383 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1384 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1385 // This is an unfortunate aspect of this design.
1386 //
1387
1388 enum ManifestConstants {
1389 ClearResponsibleAtSTW = 0,
1390 MaximumRecheckInterval = 1000
1391 } ;
1392
1393 // Deflate a single monitor if not in use
1394 // Return true if deflated, false if in use
1395 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1396 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) {
1397 bool deflated;
1398 // Normal case ... The monitor is associated with obj.
1399 guarantee (obj->mark() == markOopDesc::encode(mid), "invariant") ;
1400 guarantee (mid == obj->mark()->monitor(), "invariant");
1401 guarantee (mid->header()->is_neutral(), "invariant");
1402
1403 if (mid->is_busy()) {
1404 if (ClearResponsibleAtSTW) mid->_Responsible = NULL ;
1405 deflated = false;
1406 } else {
1407 // Deflate the monitor if it is no longer being used
1408 // It's idle - scavenge and return to the global free list
1409 // plain old deflation ...
1410 TEVENT (deflate_idle_monitors - scavenge1) ;
1411 if (TraceMonitorInflation) {
1412 if (obj->is_instance()) {
1413 ResourceMark rm;
1414 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1415 (intptr_t) obj, (intptr_t) obj->mark(), Klass::cast(obj->klass())->external_name());
1416 }
1417 }
1418
1419 // Restore the header back to obj
1420 obj->release_set_mark(mid->header());
1421 mid->clear();
1422
1423 assert (mid->object() == NULL, "invariant") ;
1424
1425 // Move the object to the working free list defined by FreeHead,FreeTail.
1426 if (*FreeHeadp == NULL) *FreeHeadp = mid;
1427 if (*FreeTailp != NULL) {
1428 ObjectMonitor * prevtail = *FreeTailp;
1429 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); // TODO KK
1430 prevtail->FreeNext = mid;
1431 }
1432 *FreeTailp = mid;
1433 deflated = true;
1434 }
1435 return deflated;
1436 }
1437
1438 // Caller acquires ListLock
1439 int ObjectSynchronizer::walk_monitor_list(ObjectMonitor** listheadp,
1440 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) {
1441 ObjectMonitor* mid;
1442 ObjectMonitor* next;
1443 ObjectMonitor* curmidinuse = NULL;
1444 int deflatedcount = 0;
1445
1446 for (mid = *listheadp; mid != NULL; ) {
1447 oop obj = (oop) mid->object();
1448 bool deflated = false;
1449 if (obj != NULL) {
1450 deflated = deflate_monitor(mid, obj, FreeHeadp, FreeTailp);
1451 }
1452 if (deflated) {
1453 // extract from per-thread in-use-list
1454 if (mid == *listheadp) {
1455 *listheadp = mid->FreeNext;
1456 } else if (curmidinuse != NULL) {
1457 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist
1458 }
1459 next = mid->FreeNext;
1460 mid->FreeNext = NULL; // This mid is current tail in the FreeHead list
1461 mid = next;
1462 deflatedcount++;
1463 } else {
1464 curmidinuse = mid;
1465 mid = mid->FreeNext;
1466 }
1467 }
1468 return deflatedcount;
1469 }
1470
1471 void ObjectSynchronizer::deflate_idle_monitors() {
1472 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1473 int nInuse = 0 ; // currently associated with objects
1474 int nInCirculation = 0 ; // extant
1475 int nScavenged = 0 ; // reclaimed
1476 bool deflated = false;
1477
1478 ObjectMonitor * FreeHead = NULL ; // Local SLL of scavenged monitors
1479 ObjectMonitor * FreeTail = NULL ;
1480
1481 TEVENT (deflate_idle_monitors) ;
1482 // Prevent omFlush from changing mids in Thread dtor's during deflation
1483 // And in case the vm thread is acquiring a lock during a safepoint
1484 // See e.g. 6320749
1485 Thread::muxAcquire (&ListLock, "scavenge - return") ;
1486
1487 if (MonitorInUseLists) {
1488 int inUse = 0;
1489 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) {
1490 nInCirculation+= cur->omInUseCount;
1491 int deflatedcount = walk_monitor_list(cur->omInUseList_addr(), &FreeHead, &FreeTail);
1492 cur->omInUseCount-= deflatedcount;
1493 // verifyInUse(cur);
1494 nScavenged += deflatedcount;
1495 nInuse += cur->omInUseCount;
1496 }
1497
1498 // For moribund threads, scan gOmInUseList
1499 if (gOmInUseList) {
1500 nInCirculation += gOmInUseCount;
1501 int deflatedcount = walk_monitor_list((ObjectMonitor **)&gOmInUseList, &FreeHead, &FreeTail);
1502 gOmInUseCount-= deflatedcount;
1503 nScavenged += deflatedcount;
1504 nInuse += gOmInUseCount;
1505 }
1506
1507 } else for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) {
1508 // Iterate over all extant monitors - Scavenge all idle monitors.
1509 assert(block->object() == CHAINMARKER, "must be a block header");
1510 nInCirculation += _BLOCKSIZE ;
1511 for (int i = 1 ; i < _BLOCKSIZE; i++) {
1512 ObjectMonitor* mid = &block[i];
1513 oop obj = (oop) mid->object();
1514
1515 if (obj == NULL) {
1516 // The monitor is not associated with an object.
1517 // The monitor should either be a thread-specific private
1518 // free list or the global free list.
1519 // obj == NULL IMPLIES mid->is_busy() == 0
1520 guarantee (!mid->is_busy(), "invariant") ;
1521 continue ;
1522 }
1523 deflated = deflate_monitor(mid, obj, &FreeHead, &FreeTail);
1524
1525 if (deflated) {
1526 mid->FreeNext = NULL ;
1527 nScavenged ++ ;
1528 } else {
1529 nInuse ++;
1530 }
1531 }
1532 }
1533
1534 MonitorFreeCount += nScavenged;
1535
1536 // Consider: audit gFreeList to ensure that MonitorFreeCount and list agree.
1537
1538 if (ObjectMonitor::Knob_Verbose) {
1539 ::printf ("Deflate: InCirc=%d InUse=%d Scavenged=%d ForceMonitorScavenge=%d : pop=%d free=%d\n",
1540 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge,
1541 MonitorPopulation, MonitorFreeCount) ;
1542 ::fflush(stdout) ;
1543 }
1544
1545 ForceMonitorScavenge = 0; // Reset
1546
1547 // Move the scavenged monitors back to the global free list.
1548 if (FreeHead != NULL) {
1549 guarantee (FreeTail != NULL && nScavenged > 0, "invariant") ;
1550 assert (FreeTail->FreeNext == NULL, "invariant") ;
1551 // constant-time list splice - prepend scavenged segment to gFreeList
1552 FreeTail->FreeNext = gFreeList ;
1553 gFreeList = FreeHead ;
1554 }
1555 Thread::muxRelease (&ListLock) ;
1556
1557 if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged) ;
1558 if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation);
1559
1560 // TODO: Add objectMonitor leak detection.
1561 // Audit/inventory the objectMonitors -- make sure they're all accounted for.
1562 GVars.stwRandom = os::random() ;
1563 GVars.stwCycle ++ ;
1564 }
1565
1566 // Monitor cleanup on JavaThread::exit
1567
1568 // Iterate through monitor cache and attempt to release thread's monitors
1569 // Gives up on a particular monitor if an exception occurs, but continues
1570 // the overall iteration, swallowing the exception.
1571 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1572 private:
1573 TRAPS;
1574
1575 public:
1576 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {}
1577 void do_monitor(ObjectMonitor* mid) {
1578 if (mid->owner() == THREAD) {
1579 (void)mid->complete_exit(CHECK);
1580 }
1581 }
1582 };
1583
1584 // Release all inflated monitors owned by THREAD. Lightweight monitors are
1585 // ignored. This is meant to be called during JNI thread detach which assumes
1586 // all remaining monitors are heavyweight. All exceptions are swallowed.
1587 // Scanning the extant monitor list can be time consuming.
1588 // A simple optimization is to add a per-thread flag that indicates a thread
1589 // called jni_monitorenter() during its lifetime.
1590 //
1591 // Instead of No_Savepoint_Verifier it might be cheaper to
1592 // use an idiom of the form:
1593 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1594 // <code that must not run at safepoint>
1595 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1596 // Since the tests are extremely cheap we could leave them enabled
1597 // for normal product builds.
1598
1599 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) {
1600 assert(THREAD == JavaThread::current(), "must be current Java thread");
1601 No_Safepoint_Verifier nsv ;
1602 ReleaseJavaMonitorsClosure rjmc(THREAD);
1603 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread");
1604 ObjectSynchronizer::monitors_iterate(&rjmc);
1605 Thread::muxRelease(&ListLock);
1606 THREAD->clear_pending_exception();
1607 }
1608
1609 //------------------------------------------------------------------------------
1610 // Non-product code
1611
1612 #ifndef PRODUCT
1613
1614 void ObjectSynchronizer::trace_locking(Handle locking_obj, bool is_compiled,
1615 bool is_method, bool is_locking) {
1616 // Don't know what to do here
1617 }
1618
1619 // Verify all monitors in the monitor cache, the verification is weak.
1620 void ObjectSynchronizer::verify() {
1621 ObjectMonitor* block = gBlockList;
1622 ObjectMonitor* mid;
1623 while (block) {
1624 assert(block->object() == CHAINMARKER, "must be a block header");
1625 for (int i = 1; i < _BLOCKSIZE; i++) {
1626 mid = block + i;
1627 oop object = (oop) mid->object();
1628 if (object != NULL) {
1629 mid->verify();
1630 }
1631 }
1632 block = (ObjectMonitor*) block->FreeNext;
1633 }
1634 }
1635
1636 // Check if monitor belongs to the monitor cache
1637 // The list is grow-only so it's *relatively* safe to traverse
1638 // the list of extant blocks without taking a lock.
1639
1640 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) {
1641 ObjectMonitor* block = gBlockList;
1642
1643 while (block) {
1644 assert(block->object() == CHAINMARKER, "must be a block header");
1645 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) {
1646 address mon = (address) monitor;
1647 address blk = (address) block;
1648 size_t diff = mon - blk;
1649 assert((diff % sizeof(ObjectMonitor)) == 0, "check");
1650 return 1;
1651 }
1652 block = (ObjectMonitor*) block->FreeNext;
1653 }
1654 return 0;
1655 }
1656
1657 #endif