1 /*
2 * Copyright (c) 1998, 2013, 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 "runtime/thread.inline.hpp"
40 #include "utilities/dtrace.hpp"
41 #include "utilities/events.hpp"
42 #include "utilities/preserveException.hpp"
43 #ifdef TARGET_OS_FAMILY_linux
44 # include "os_linux.inline.hpp"
45 #endif
46 #ifdef TARGET_OS_FAMILY_solaris
47 # include "os_solaris.inline.hpp"
48 #endif
49 #ifdef TARGET_OS_FAMILY_windows
50 # include "os_windows.inline.hpp"
51 #endif
52 #ifdef TARGET_OS_FAMILY_bsd
53 # include "os_bsd.inline.hpp"
54 #endif
55
56 #if defined(__GNUC__)
57 // Need to inhibit inlining for older versions of GCC to avoid build-time failures
58 #define ATTR __attribute__((noinline))
59 #else
60 #define ATTR
61 #endif
62
63 // The "core" versions of monitor enter and exit reside in this file.
64 // The interpreter and compilers contain specialized transliterated
65 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(),
66 // for instance. If you make changes here, make sure to modify the
67 // interpreter, and both C1 and C2 fast-path inline locking code emission.
68 //
69 //
70 // -----------------------------------------------------------------------------
71
72 #ifdef DTRACE_ENABLED
73
74 // Only bother with this argument setup if dtrace is available
75 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
76
77 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
78 char* bytes = NULL; \
79 int len = 0; \
80 jlong jtid = SharedRuntime::get_java_tid(thread); \
81 Symbol* klassname = ((oop)(obj))->klass()->name(); \
82 if (klassname != NULL) { \
83 bytes = (char*)klassname->bytes(); \
84 len = klassname->utf8_length(); \
85 }
86
87 #ifndef USDT2
88 HS_DTRACE_PROBE_DECL5(hotspot, monitor__wait,
89 jlong, uintptr_t, char*, int, long);
90 HS_DTRACE_PROBE_DECL4(hotspot, monitor__waited,
91 jlong, uintptr_t, char*, int);
92
93 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
94 { \
95 if (DTraceMonitorProbes) { \
96 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
97 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \
98 (monitor), bytes, len, (millis)); \
99 } \
100 }
101
102 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
103 { \
104 if (DTraceMonitorProbes) { \
105 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
106 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \
107 (uintptr_t)(monitor), bytes, len); \
108 } \
109 }
110
111 #else /* USDT2 */
112
113 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
114 { \
115 if (DTraceMonitorProbes) { \
116 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
117 HOTSPOT_MONITOR_WAIT(jtid, \
118 (uintptr_t)(monitor), bytes, len, (millis)); \
119 } \
120 }
121
122 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_PROBE_WAITED
123
124 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
125 { \
126 if (DTraceMonitorProbes) { \
127 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
128 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
129 (uintptr_t)(monitor), bytes, len); \
130 } \
131 }
132
133 #endif /* USDT2 */
134 #else // ndef DTRACE_ENABLED
135
136 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
137 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
138
139 #endif // ndef DTRACE_ENABLED
140
141 // This exists only as a workaround of dtrace bug 6254741
142 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
143 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
144 return 0;
145 }
146
147 #define NINFLATIONLOCKS 256
148 static volatile intptr_t InflationLocks [NINFLATIONLOCKS] ;
149
150 ObjectMonitor * ObjectSynchronizer::gBlockList = NULL ;
151 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL ;
152 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL ;
153 int ObjectSynchronizer::gOmInUseCount = 0;
154 static volatile intptr_t ListLock = 0 ; // protects global monitor free-list cache
155 static volatile int MonitorFreeCount = 0 ; // # on gFreeList
156 static volatile int MonitorPopulation = 0 ; // # Extant -- in circulation
157 #define CHAINMARKER ((oop)-1)
158
159 // -----------------------------------------------------------------------------
160 // Fast Monitor Enter/Exit
161 // This the fast monitor enter. The interpreter and compiler use
162 // some assembly copies of this code. Make sure update those code
163 // if the following function is changed. The implementation is
164 // extremely sensitive to race condition. Be careful.
165
166 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) {
167 if (UseBiasedLocking) {
168 if (!SafepointSynchronize::is_at_safepoint()) {
169 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD);
170 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) {
171 return;
172 }
173 } else {
174 assert(!attempt_rebias, "can not rebias toward VM thread");
175 BiasedLocking::revoke_at_safepoint(obj);
176 }
177 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
178 }
179
180 slow_enter (obj, lock, THREAD) ;
181 }
182
183 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) {
184 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here");
185 // if displaced header is null, the previous enter is recursive enter, no-op
186 markOop dhw = lock->displaced_header();
187 markOop mark ;
188 if (dhw == NULL) {
189 // Recursive stack-lock.
190 // Diagnostics -- Could be: stack-locked, inflating, inflated.
191 mark = object->mark() ;
192 assert (!mark->is_neutral(), "invariant") ;
193 if (mark->has_locker() && mark != markOopDesc::INFLATING()) {
194 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant") ;
195 }
196 if (mark->has_monitor()) {
197 ObjectMonitor * m = mark->monitor() ;
198 assert(((oop)(m->object()))->mark() == mark, "invariant") ;
199 assert(m->is_entered(THREAD), "invariant") ;
200 }
201 return ;
202 }
203
204 mark = object->mark() ;
205
206 // If the object is stack-locked by the current thread, try to
207 // swing the displaced header from the box back to the mark.
208 if (mark == (markOop) lock) {
209 assert (dhw->is_neutral(), "invariant") ;
210 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) {
211 TEVENT (fast_exit: release stacklock) ;
212 return;
213 }
214 }
215
216 ObjectSynchronizer::inflate(THREAD, object)->exit (THREAD) ;
217 }
218
219 // -----------------------------------------------------------------------------
220 // Interpreter/Compiler Slow Case
221 // This routine is used to handle interpreter/compiler slow case
222 // We don't need to use fast path here, because it must have been
223 // failed in the interpreter/compiler code.
224 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
225 markOop mark = obj->mark();
226 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
227
228 if (mark->is_neutral()) {
229 // Anticipate successful CAS -- the ST of the displaced mark must
230 // be visible <= the ST performed by the CAS.
231 lock->set_displaced_header(mark);
232 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) {
233 TEVENT (slow_enter: release stacklock) ;
234 return ;
235 }
236 // Fall through to inflate() ...
237 } else
238 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
239 assert(lock != mark->locker(), "must not re-lock the same lock");
240 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
241 lock->set_displaced_header(NULL);
242 return;
243 }
244
245 #if 0
246 // The following optimization isn't particularly useful.
247 if (mark->has_monitor() && mark->monitor()->is_entered(THREAD)) {
248 lock->set_displaced_header (NULL) ;
249 return ;
250 }
251 #endif
252
253 // The object header will never be displaced to this lock,
254 // so it does not matter what the value is, except that it
255 // must be non-zero to avoid looking like a re-entrant lock,
256 // and must not look locked either.
257 lock->set_displaced_header(markOopDesc::unused_mark());
258 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
259 }
260
261 // This routine is used to handle interpreter/compiler slow case
262 // We don't need to use fast path here, because it must have
263 // failed in the interpreter/compiler code. Simply use the heavy
264 // weight monitor should be ok, unless someone find otherwise.
265 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
266 fast_exit (object, lock, THREAD) ;
267 }
268
269 // -----------------------------------------------------------------------------
270 // Class Loader support to workaround deadlocks on the class loader lock objects
271 // Also used by GC
272 // complete_exit()/reenter() are used to wait on a nested lock
273 // i.e. to give up an outer lock completely and then re-enter
274 // Used when holding nested locks - lock acquisition order: lock1 then lock2
275 // 1) complete_exit lock1 - saving recursion count
276 // 2) wait on lock2
277 // 3) when notified on lock2, unlock lock2
278 // 4) reenter lock1 with original recursion count
279 // 5) lock lock2
280 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
281 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
282 TEVENT (complete_exit) ;
283 if (UseBiasedLocking) {
284 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
285 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
286 }
287
288 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
289
290 return monitor->complete_exit(THREAD);
291 }
292
293 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
294 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
295 TEVENT (reenter) ;
296 if (UseBiasedLocking) {
297 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
298 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
299 }
300
301 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
302
303 monitor->reenter(recursion, THREAD);
304 }
305 // -----------------------------------------------------------------------------
306 // JNI locks on java objects
307 // NOTE: must use heavy weight monitor to handle jni monitor enter
308 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { // possible entry from jni enter
309 // the current locking is from JNI instead of Java code
310 TEVENT (jni_enter) ;
311 if (UseBiasedLocking) {
312 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
313 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
314 }
315 THREAD->set_current_pending_monitor_is_from_java(false);
316 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
317 THREAD->set_current_pending_monitor_is_from_java(true);
318 }
319
320 // NOTE: must use heavy weight monitor to handle jni monitor enter
321 bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) {
322 if (UseBiasedLocking) {
323 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
324 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
325 }
326
327 ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj());
328 return monitor->try_enter(THREAD);
329 }
330
331
332 // NOTE: must use heavy weight monitor to handle jni monitor exit
333 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
334 TEVENT (jni_exit) ;
335 if (UseBiasedLocking) {
336 Handle h_obj(THREAD, obj);
337 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD);
338 obj = h_obj();
339 }
340 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
341
342 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj);
343 // If this thread has locked the object, exit the monitor. Note: can't use
344 // monitor->check(CHECK); must exit even if an exception is pending.
345 if (monitor->check(THREAD)) {
346 monitor->exit(THREAD);
347 }
348 }
349
350 // -----------------------------------------------------------------------------
351 // Internal VM locks on java objects
352 // standard constructor, allows locking failures
353 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
354 _dolock = doLock;
355 _thread = thread;
356 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
357 _obj = obj;
358
359 if (_dolock) {
360 TEVENT (ObjectLocker) ;
361
362 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
363 }
364 }
365
366 ObjectLocker::~ObjectLocker() {
367 if (_dolock) {
368 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
369 }
370 }
371
372
373 // -----------------------------------------------------------------------------
374 // Wait/Notify/NotifyAll
375 // NOTE: must use heavy weight monitor to handle wait()
376 void ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
377 if (UseBiasedLocking) {
378 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
379 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
380 }
381 if (millis < 0) {
382 TEVENT (wait - throw IAX) ;
383 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
384 }
385 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
386 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
387 monitor->wait(millis, true, THREAD);
388
389 /* This dummy call is in place to get around dtrace bug 6254741. Once
390 that's fixed we can uncomment the following line and remove the call */
391 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
392 dtrace_waited_probe(monitor, obj, THREAD);
393 }
394
395 void ObjectSynchronizer::waitUninterruptibly (Handle obj, jlong millis, TRAPS) {
396 if (UseBiasedLocking) {
397 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
398 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
399 }
400 if (millis < 0) {
401 TEVENT (wait - throw IAX) ;
402 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
403 }
404 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD) ;
405 }
406
407 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
408 if (UseBiasedLocking) {
409 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
410 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
411 }
412
413 markOop mark = obj->mark();
414 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
415 return;
416 }
417 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD);
418 }
419
420 // NOTE: see comment of notify()
421 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
422 if (UseBiasedLocking) {
423 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
424 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
425 }
426
427 markOop mark = obj->mark();
428 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
429 return;
430 }
431 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD);
432 }
433
434 // -----------------------------------------------------------------------------
435 // Hash Code handling
436 //
437 // Performance concern:
438 // OrderAccess::storestore() calls release() which STs 0 into the global volatile
439 // OrderAccess::Dummy variable. This store is unnecessary for correctness.
440 // Many threads STing into a common location causes considerable cache migration
441 // or "sloshing" on large SMP system. As such, I avoid using OrderAccess::storestore()
442 // until it's repaired. In some cases OrderAccess::fence() -- which incurs local
443 // latency on the executing processor -- is a better choice as it scales on SMP
444 // systems. See http://blogs.sun.com/dave/entry/biased_locking_in_hotspot for a
445 // discussion of coherency costs. Note that all our current reference platforms
446 // provide strong ST-ST order, so the issue is moot on IA32, x64, and SPARC.
447 //
448 // As a general policy we use "volatile" to control compiler-based reordering
449 // and explicit fences (barriers) to control for architectural reordering performed
450 // by the CPU(s) or platform.
451
452 struct SharedGlobals {
453 // These are highly shared mostly-read variables.
454 // To avoid false-sharing they need to be the sole occupants of a $ line.
455 double padPrefix [8];
456 volatile int stwRandom ;
457 volatile int stwCycle ;
458
459 // Hot RW variables -- Sequester to avoid false-sharing
460 double padSuffix [16];
461 volatile int hcSequence ;
462 double padFinal [8] ;
463 } ;
464
465 static SharedGlobals GVars ;
466 static int MonitorScavengeThreshold = 1000000 ;
467 static volatile int ForceMonitorScavenge = 0 ; // Scavenge required and pending
468
469 static markOop ReadStableMark (oop obj) {
470 markOop mark = obj->mark() ;
471 if (!mark->is_being_inflated()) {
472 return mark ; // normal fast-path return
473 }
474
475 int its = 0 ;
476 for (;;) {
477 markOop mark = obj->mark() ;
478 if (!mark->is_being_inflated()) {
479 return mark ; // normal fast-path return
480 }
481
482 // The object is being inflated by some other thread.
483 // The caller of ReadStableMark() must wait for inflation to complete.
484 // Avoid live-lock
485 // TODO: consider calling SafepointSynchronize::do_call_back() while
486 // spinning to see if there's a safepoint pending. If so, immediately
487 // yielding or blocking would be appropriate. Avoid spinning while
488 // there is a safepoint pending.
489 // TODO: add inflation contention performance counters.
490 // TODO: restrict the aggregate number of spinners.
491
492 ++its ;
493 if (its > 10000 || !os::is_MP()) {
494 if (its & 1) {
495 os::NakedYield() ;
496 TEVENT (Inflate: INFLATING - yield) ;
497 } else {
498 // Note that the following code attenuates the livelock problem but is not
499 // a complete remedy. A more complete solution would require that the inflating
500 // thread hold the associated inflation lock. The following code simply restricts
501 // the number of spinners to at most one. We'll have N-2 threads blocked
502 // on the inflationlock, 1 thread holding the inflation lock and using
503 // a yield/park strategy, and 1 thread in the midst of inflation.
504 // A more refined approach would be to change the encoding of INFLATING
505 // to allow encapsulation of a native thread pointer. Threads waiting for
506 // inflation to complete would use CAS to push themselves onto a singly linked
507 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
508 // and calling park(). When inflation was complete the thread that accomplished inflation
509 // would detach the list and set the markword to inflated with a single CAS and
510 // then for each thread on the list, set the flag and unpark() the thread.
511 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease
512 // wakes at most one thread whereas we need to wake the entire list.
513 int ix = (intptr_t(obj) >> 5) & (NINFLATIONLOCKS-1) ;
514 int YieldThenBlock = 0 ;
515 assert (ix >= 0 && ix < NINFLATIONLOCKS, "invariant") ;
516 assert ((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant") ;
517 Thread::muxAcquire (InflationLocks + ix, "InflationLock") ;
518 while (obj->mark() == markOopDesc::INFLATING()) {
519 // Beware: NakedYield() is advisory and has almost no effect on some platforms
520 // so we periodically call Self->_ParkEvent->park(1).
521 // We use a mixed spin/yield/block mechanism.
522 if ((YieldThenBlock++) >= 16) {
523 Thread::current()->_ParkEvent->park(1) ;
524 } else {
525 os::NakedYield() ;
526 }
527 }
528 Thread::muxRelease (InflationLocks + ix ) ;
529 TEVENT (Inflate: INFLATING - yield/park) ;
530 }
531 } else {
532 SpinPause() ; // SMP-polite spinning
533 }
534 }
535 }
536
537 // hashCode() generation :
538 //
539 // Possibilities:
540 // * MD5Digest of {obj,stwRandom}
541 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function.
542 // * A DES- or AES-style SBox[] mechanism
543 // * One of the Phi-based schemes, such as:
544 // 2654435761 = 2^32 * Phi (golden ratio)
545 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ;
546 // * A variation of Marsaglia's shift-xor RNG scheme.
547 // * (obj ^ stwRandom) is appealing, but can result
548 // in undesirable regularity in the hashCode values of adjacent objects
549 // (objects allocated back-to-back, in particular). This could potentially
550 // result in hashtable collisions and reduced hashtable efficiency.
551 // There are simple ways to "diffuse" the middle address bits over the
552 // generated hashCode values:
553 //
554
555 static inline intptr_t get_next_hash(Thread * Self, oop obj) {
556 intptr_t value = 0 ;
557 if (hashCode == 0) {
558 // This form uses an unguarded global Park-Miller RNG,
559 // so it's possible for two threads to race and generate the same RNG.
560 // On MP system we'll have lots of RW access to a global, so the
561 // mechanism induces lots of coherency traffic.
562 value = os::random() ;
563 } else
564 if (hashCode == 1) {
565 // This variation has the property of being stable (idempotent)
566 // between STW operations. This can be useful in some of the 1-0
567 // synchronization schemes.
568 intptr_t addrBits = intptr_t(obj) >> 3 ;
569 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom ;
570 } else
571 if (hashCode == 2) {
572 value = 1 ; // for sensitivity testing
573 } else
574 if (hashCode == 3) {
575 value = ++GVars.hcSequence ;
576 } else
577 if (hashCode == 4) {
578 value = intptr_t(obj) ;
579 } else {
580 // Marsaglia's xor-shift scheme with thread-specific state
581 // This is probably the best overall implementation -- we'll
582 // likely make this the default in future releases.
583 unsigned t = Self->_hashStateX ;
584 t ^= (t << 11) ;
585 Self->_hashStateX = Self->_hashStateY ;
586 Self->_hashStateY = Self->_hashStateZ ;
587 Self->_hashStateZ = Self->_hashStateW ;
588 unsigned v = Self->_hashStateW ;
589 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)) ;
590 Self->_hashStateW = v ;
591 value = v ;
592 }
593
594 value &= markOopDesc::hash_mask;
595 if (value == 0) value = 0xBAD ;
596 assert (value != markOopDesc::no_hash, "invariant") ;
597 TEVENT (hashCode: GENERATE) ;
598 return value;
599 }
600 //
601 intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) {
602 if (UseBiasedLocking) {
603 // NOTE: many places throughout the JVM do not expect a safepoint
604 // to be taken here, in particular most operations on perm gen
605 // objects. However, we only ever bias Java instances and all of
606 // the call sites of identity_hash that might revoke biases have
607 // been checked to make sure they can handle a safepoint. The
608 // added check of the bias pattern is to avoid useless calls to
609 // thread-local storage.
610 if (obj->mark()->has_bias_pattern()) {
611 // Box and unbox the raw reference just in case we cause a STW safepoint.
612 Handle hobj (Self, obj) ;
613 // Relaxing assertion for bug 6320749.
614 assert (Universe::verify_in_progress() ||
615 !SafepointSynchronize::is_at_safepoint(),
616 "biases should not be seen by VM thread here");
617 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
618 obj = hobj() ;
619 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
620 }
621 }
622
623 // hashCode() is a heap mutator ...
624 // Relaxing assertion for bug 6320749.
625 assert (Universe::verify_in_progress() ||
626 !SafepointSynchronize::is_at_safepoint(), "invariant") ;
627 assert (Universe::verify_in_progress() ||
628 Self->is_Java_thread() , "invariant") ;
629 assert (Universe::verify_in_progress() ||
630 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ;
631
632 ObjectMonitor* monitor = NULL;
633 markOop temp, test;
634 intptr_t hash;
635 markOop mark = ReadStableMark (obj);
636
637 // object should remain ineligible for biased locking
638 assert (!mark->has_bias_pattern(), "invariant") ;
639
640 if (mark->is_neutral()) {
641 hash = mark->hash(); // this is a normal header
642 if (hash) { // if it has hash, just return it
643 return hash;
644 }
645 hash = get_next_hash(Self, obj); // allocate a new hash code
646 temp = mark->copy_set_hash(hash); // merge the hash code into header
647 // use (machine word version) atomic operation to install the hash
648 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark);
649 if (test == mark) {
650 return hash;
651 }
652 // If atomic operation failed, we must inflate the header
653 // into heavy weight monitor. We could add more code here
654 // for fast path, but it does not worth the complexity.
655 } else if (mark->has_monitor()) {
656 monitor = mark->monitor();
657 temp = monitor->header();
658 assert (temp->is_neutral(), "invariant") ;
659 hash = temp->hash();
660 if (hash) {
661 return hash;
662 }
663 // Skip to the following code to reduce code size
664 } else if (Self->is_lock_owned((address)mark->locker())) {
665 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
666 assert (temp->is_neutral(), "invariant") ;
667 hash = temp->hash(); // by current thread, check if the displaced
668 if (hash) { // header contains hash code
669 return hash;
670 }
671 // WARNING:
672 // The displaced header is strictly immutable.
673 // It can NOT be changed in ANY cases. So we have
674 // to inflate the header into heavyweight monitor
675 // even the current thread owns the lock. The reason
676 // is the BasicLock (stack slot) will be asynchronously
677 // read by other threads during the inflate() function.
678 // Any change to stack may not propagate to other threads
679 // correctly.
680 }
681
682 // Inflate the monitor to set hash code
683 monitor = ObjectSynchronizer::inflate(Self, obj);
684 // Load displaced header and check it has hash code
685 mark = monitor->header();
686 assert (mark->is_neutral(), "invariant") ;
687 hash = mark->hash();
688 if (hash == 0) {
689 hash = get_next_hash(Self, obj);
690 temp = mark->copy_set_hash(hash); // merge hash code into header
691 assert (temp->is_neutral(), "invariant") ;
692 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark);
693 if (test != mark) {
694 // The only update to the header in the monitor (outside GC)
695 // is install the hash code. If someone add new usage of
696 // displaced header, please update this code
697 hash = test->hash();
698 assert (test->is_neutral(), "invariant") ;
699 assert (hash != 0, "Trivial unexpected object/monitor header usage.");
700 }
701 }
702 // We finally get the hash
703 return hash;
704 }
705
706 // Deprecated -- use FastHashCode() instead.
707
708 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
709 return FastHashCode (Thread::current(), obj()) ;
710 }
711
712
713 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
714 Handle h_obj) {
715 if (UseBiasedLocking) {
716 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
717 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
718 }
719
720 assert(thread == JavaThread::current(), "Can only be called on current thread");
721 oop obj = h_obj();
722
723 markOop mark = ReadStableMark (obj) ;
724
725 // Uncontended case, header points to stack
726 if (mark->has_locker()) {
727 return thread->is_lock_owned((address)mark->locker());
728 }
729 // Contended case, header points to ObjectMonitor (tagged pointer)
730 if (mark->has_monitor()) {
731 ObjectMonitor* monitor = mark->monitor();
732 return monitor->is_entered(thread) != 0 ;
733 }
734 // Unlocked case, header in place
735 assert(mark->is_neutral(), "sanity check");
736 return false;
737 }
738
739 // Be aware of this method could revoke bias of the lock object.
740 // This method querys the ownership of the lock handle specified by 'h_obj'.
741 // If the current thread owns the lock, it returns owner_self. If no
742 // thread owns the lock, it returns owner_none. Otherwise, it will return
743 // ower_other.
744 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
745 (JavaThread *self, Handle h_obj) {
746 // The caller must beware this method can revoke bias, and
747 // revocation can result in a safepoint.
748 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ;
749 assert (self->thread_state() != _thread_blocked , "invariant") ;
750
751 // Possible mark states: neutral, biased, stack-locked, inflated
752
753 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
754 // CASE: biased
755 BiasedLocking::revoke_and_rebias(h_obj, false, self);
756 assert(!h_obj->mark()->has_bias_pattern(),
757 "biases should be revoked by now");
758 }
759
760 assert(self == JavaThread::current(), "Can only be called on current thread");
761 oop obj = h_obj();
762 markOop mark = ReadStableMark (obj) ;
763
764 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
765 if (mark->has_locker()) {
766 return self->is_lock_owned((address)mark->locker()) ?
767 owner_self : owner_other;
768 }
769
770 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor.
771 // The Object:ObjectMonitor relationship is stable as long as we're
772 // not at a safepoint.
773 if (mark->has_monitor()) {
774 void * owner = mark->monitor()->_owner ;
775 if (owner == NULL) return owner_none ;
776 return (owner == self ||
777 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
778 }
779
780 // CASE: neutral
781 assert(mark->is_neutral(), "sanity check");
782 return owner_none ; // it's unlocked
783 }
784
785 // FIXME: jvmti should call this
786 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) {
787 if (UseBiasedLocking) {
788 if (SafepointSynchronize::is_at_safepoint()) {
789 BiasedLocking::revoke_at_safepoint(h_obj);
790 } else {
791 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
792 }
793 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
794 }
795
796 oop obj = h_obj();
797 address owner = NULL;
798
799 markOop mark = ReadStableMark (obj) ;
800
801 // Uncontended case, header points to stack
802 if (mark->has_locker()) {
803 owner = (address) mark->locker();
804 }
805
806 // Contended case, header points to ObjectMonitor (tagged pointer)
807 if (mark->has_monitor()) {
808 ObjectMonitor* monitor = mark->monitor();
809 assert(monitor != NULL, "monitor should be non-null");
810 owner = (address) monitor->owner();
811 }
812
813 if (owner != NULL) {
814 // owning_thread_from_monitor_owner() may also return NULL here
815 return Threads::owning_thread_from_monitor_owner(owner, doLock);
816 }
817
818 // Unlocked case, header in place
819 // Cannot have assertion since this object may have been
820 // locked by another thread when reaching here.
821 // assert(mark->is_neutral(), "sanity check");
822
823 return NULL;
824 }
825 // Visitors ...
826
827 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
828 ObjectMonitor* block = gBlockList;
829 ObjectMonitor* mid;
830 while (block) {
831 assert(block->object() == CHAINMARKER, "must be a block header");
832 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
833 mid = block + i;
834 oop object = (oop) mid->object();
835 if (object != NULL) {
836 closure->do_monitor(mid);
837 }
838 }
839 block = (ObjectMonitor*) block->FreeNext;
840 }
841 }
842
843 // Get the next block in the block list.
844 static inline ObjectMonitor* next(ObjectMonitor* block) {
845 assert(block->object() == CHAINMARKER, "must be a block header");
846 block = block->FreeNext ;
847 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
848 return block;
849 }
850
851
852 void ObjectSynchronizer::oops_do(OopClosure* f) {
853 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
854 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) {
855 assert(block->object() == CHAINMARKER, "must be a block header");
856 for (int i = 1; i < _BLOCKSIZE; i++) {
857 ObjectMonitor* mid = &block[i];
858 if (mid->object() != NULL) {
859 f->do_oop((oop*)mid->object_addr());
860 }
861 }
862 }
863 }
864
865
866 // -----------------------------------------------------------------------------
867 // ObjectMonitor Lifecycle
868 // -----------------------
869 // Inflation unlinks monitors from the global gFreeList and
870 // associates them with objects. Deflation -- which occurs at
871 // STW-time -- disassociates idle monitors from objects. Such
872 // scavenged monitors are returned to the gFreeList.
873 //
874 // The global list is protected by ListLock. All the critical sections
875 // are short and operate in constant-time.
876 //
877 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
878 //
879 // Lifecycle:
880 // -- unassigned and on the global free list
881 // -- unassigned and on a thread's private omFreeList
882 // -- assigned to an object. The object is inflated and the mark refers
883 // to the objectmonitor.
884 //
885
886
887 // Constraining monitor pool growth via MonitorBound ...
888 //
889 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
890 // the rate of scavenging is driven primarily by GC. As such, we can find
891 // an inordinate number of monitors in circulation.
892 // To avoid that scenario we can artificially induce a STW safepoint
893 // if the pool appears to be growing past some reasonable bound.
894 // Generally we favor time in space-time tradeoffs, but as there's no
895 // natural back-pressure on the # of extant monitors we need to impose some
896 // type of limit. Beware that if MonitorBound is set to too low a value
897 // we could just loop. In addition, if MonitorBound is set to a low value
898 // we'll incur more safepoints, which are harmful to performance.
899 // See also: GuaranteedSafepointInterval
900 //
901 // The current implementation uses asynchronous VM operations.
902 //
903
904 static void InduceScavenge (Thread * Self, const char * Whence) {
905 // Induce STW safepoint to trim monitors
906 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
907 // More precisely, trigger an asynchronous STW safepoint as the number
908 // of active monitors passes the specified threshold.
909 // TODO: assert thread state is reasonable
910
911 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
912 if (ObjectMonitor::Knob_Verbose) {
913 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ;
914 ::fflush(stdout) ;
915 }
916 // Induce a 'null' safepoint to scavenge monitors
917 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
918 // to the VMthread and have a lifespan longer than that of this activation record.
919 // The VMThread will delete the op when completed.
920 VMThread::execute (new VM_ForceAsyncSafepoint()) ;
921
922 if (ObjectMonitor::Knob_Verbose) {
923 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ;
924 ::fflush(stdout) ;
925 }
926 }
927 }
928 /* Too slow for general assert or debug
929 void ObjectSynchronizer::verifyInUse (Thread *Self) {
930 ObjectMonitor* mid;
931 int inusetally = 0;
932 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) {
933 inusetally ++;
934 }
935 assert(inusetally == Self->omInUseCount, "inuse count off");
936
937 int freetally = 0;
938 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) {
939 freetally ++;
940 }
941 assert(freetally == Self->omFreeCount, "free count off");
942 }
943 */
944 ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) {
945 // A large MAXPRIVATE value reduces both list lock contention
946 // and list coherency traffic, but also tends to increase the
947 // number of objectMonitors in circulation as well as the STW
948 // scavenge costs. As usual, we lean toward time in space-time
949 // tradeoffs.
950 const int MAXPRIVATE = 1024 ;
951 for (;;) {
952 ObjectMonitor * m ;
953
954 // 1: try to allocate from the thread's local omFreeList.
955 // Threads will attempt to allocate first from their local list, then
956 // from the global list, and only after those attempts fail will the thread
957 // attempt to instantiate new monitors. Thread-local free lists take
958 // heat off the ListLock and improve allocation latency, as well as reducing
959 // coherency traffic on the shared global list.
960 m = Self->omFreeList ;
961 if (m != NULL) {
962 Self->omFreeList = m->FreeNext ;
963 Self->omFreeCount -- ;
964 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene
965 guarantee (m->object() == NULL, "invariant") ;
966 if (MonitorInUseLists) {
967 m->FreeNext = Self->omInUseList;
968 Self->omInUseList = m;
969 Self->omInUseCount ++;
970 // verifyInUse(Self);
971 } else {
972 m->FreeNext = NULL;
973 }
974 return m ;
975 }
976
977 // 2: try to allocate from the global gFreeList
978 // CONSIDER: use muxTry() instead of muxAcquire().
979 // If the muxTry() fails then drop immediately into case 3.
980 // If we're using thread-local free lists then try
981 // to reprovision the caller's free list.
982 if (gFreeList != NULL) {
983 // Reprovision the thread's omFreeList.
984 // Use bulk transfers to reduce the allocation rate and heat
985 // on various locks.
986 Thread::muxAcquire (&ListLock, "omAlloc") ;
987 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL; ) {
988 MonitorFreeCount --;
989 ObjectMonitor * take = gFreeList ;
990 gFreeList = take->FreeNext ;
991 guarantee (take->object() == NULL, "invariant") ;
992 guarantee (!take->is_busy(), "invariant") ;
993 take->Recycle() ;
994 omRelease (Self, take, false) ;
995 }
996 Thread::muxRelease (&ListLock) ;
997 Self->omFreeProvision += 1 + (Self->omFreeProvision/2) ;
998 if (Self->omFreeProvision > MAXPRIVATE ) Self->omFreeProvision = MAXPRIVATE ;
999 TEVENT (omFirst - reprovision) ;
1000
1001 const int mx = MonitorBound ;
1002 if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) {
1003 // We can't safely induce a STW safepoint from omAlloc() as our thread
1004 // state may not be appropriate for such activities and callers may hold
1005 // naked oops, so instead we defer the action.
1006 InduceScavenge (Self, "omAlloc") ;
1007 }
1008 continue;
1009 }
1010
1011 // 3: allocate a block of new ObjectMonitors
1012 // Both the local and global free lists are empty -- resort to malloc().
1013 // In the current implementation objectMonitors are TSM - immortal.
1014 assert (_BLOCKSIZE > 1, "invariant") ;
1015 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE];
1016
1017 // NOTE: (almost) no way to recover if allocation failed.
1018 // We might be able to induce a STW safepoint and scavenge enough
1019 // objectMonitors to permit progress.
1020 if (temp == NULL) {
1021 vm_exit_out_of_memory (sizeof (ObjectMonitor[_BLOCKSIZE]), "Allocate ObjectMonitors") ;
1022 }
1023
1024 // Format the block.
1025 // initialize the linked list, each monitor points to its next
1026 // forming the single linked free list, the very first monitor
1027 // will points to next block, which forms the block list.
1028 // The trick of using the 1st element in the block as gBlockList
1029 // linkage should be reconsidered. A better implementation would
1030 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1031
1032 for (int i = 1; i < _BLOCKSIZE ; i++) {
1033 temp[i].FreeNext = &temp[i+1];
1034 }
1035
1036 // terminate the last monitor as the end of list
1037 temp[_BLOCKSIZE - 1].FreeNext = NULL ;
1038
1039 // Element [0] is reserved for global list linkage
1040 temp[0].set_object(CHAINMARKER);
1041
1042 // Consider carving out this thread's current request from the
1043 // block in hand. This avoids some lock traffic and redundant
1044 // list activity.
1045
1046 // Acquire the ListLock to manipulate BlockList and FreeList.
1047 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1048 Thread::muxAcquire (&ListLock, "omAlloc [2]") ;
1049 MonitorPopulation += _BLOCKSIZE-1;
1050 MonitorFreeCount += _BLOCKSIZE-1;
1051
1052 // Add the new block to the list of extant blocks (gBlockList).
1053 // The very first objectMonitor in a block is reserved and dedicated.
1054 // It serves as blocklist "next" linkage.
1055 temp[0].FreeNext = gBlockList;
1056 gBlockList = temp;
1057
1058 // Add the new string of objectMonitors to the global free list
1059 temp[_BLOCKSIZE - 1].FreeNext = gFreeList ;
1060 gFreeList = temp + 1;
1061 Thread::muxRelease (&ListLock) ;
1062 TEVENT (Allocate block of monitors) ;
1063 }
1064 }
1065
1066 // Place "m" on the caller's private per-thread omFreeList.
1067 // In practice there's no need to clamp or limit the number of
1068 // monitors on a thread's omFreeList as the only time we'll call
1069 // omRelease is to return a monitor to the free list after a CAS
1070 // attempt failed. This doesn't allow unbounded #s of monitors to
1071 // accumulate on a thread's free list.
1072 //
1073
1074 void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m, bool fromPerThreadAlloc) {
1075 guarantee (m->object() == NULL, "invariant") ;
1076
1077 // Remove from omInUseList
1078 if (MonitorInUseLists && fromPerThreadAlloc) {
1079 ObjectMonitor* curmidinuse = NULL;
1080 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; ) {
1081 if (m == mid) {
1082 // extract from per-thread in-use-list
1083 if (mid == Self->omInUseList) {
1084 Self->omInUseList = mid->FreeNext;
1085 } else if (curmidinuse != NULL) {
1086 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist
1087 }
1088 Self->omInUseCount --;
1089 // verifyInUse(Self);
1090 break;
1091 } else {
1092 curmidinuse = mid;
1093 mid = mid->FreeNext;
1094 }
1095 }
1096 }
1097
1098 // FreeNext is used for both onInUseList and omFreeList, so clear old before setting new
1099 m->FreeNext = Self->omFreeList ;
1100 Self->omFreeList = m ;
1101 Self->omFreeCount ++ ;
1102 }
1103
1104 // Return the monitors of a moribund thread's local free list to
1105 // the global free list. Typically a thread calls omFlush() when
1106 // it's dying. We could also consider having the VM thread steal
1107 // monitors from threads that have not run java code over a few
1108 // consecutive STW safepoints. Relatedly, we might decay
1109 // omFreeProvision at STW safepoints.
1110 //
1111 // Also return the monitors of a moribund thread"s omInUseList to
1112 // a global gOmInUseList under the global list lock so these
1113 // will continue to be scanned.
1114 //
1115 // We currently call omFlush() from the Thread:: dtor _after the thread
1116 // has been excised from the thread list and is no longer a mutator.
1117 // That means that omFlush() can run concurrently with a safepoint and
1118 // the scavenge operator. Calling omFlush() from JavaThread::exit() might
1119 // be a better choice as we could safely reason that that the JVM is
1120 // not at a safepoint at the time of the call, and thus there could
1121 // be not inopportune interleavings between omFlush() and the scavenge
1122 // operator.
1123
1124 void ObjectSynchronizer::omFlush (Thread * Self) {
1125 ObjectMonitor * List = Self->omFreeList ; // Null-terminated SLL
1126 Self->omFreeList = NULL ;
1127 ObjectMonitor * Tail = NULL ;
1128 int Tally = 0;
1129 if (List != NULL) {
1130 ObjectMonitor * s ;
1131 for (s = List ; s != NULL ; s = s->FreeNext) {
1132 Tally ++ ;
1133 Tail = s ;
1134 guarantee (s->object() == NULL, "invariant") ;
1135 guarantee (!s->is_busy(), "invariant") ;
1136 s->set_owner (NULL) ; // redundant but good hygiene
1137 TEVENT (omFlush - Move one) ;
1138 }
1139 guarantee (Tail != NULL && List != NULL, "invariant") ;
1140 }
1141
1142 ObjectMonitor * InUseList = Self->omInUseList;
1143 ObjectMonitor * InUseTail = NULL ;
1144 int InUseTally = 0;
1145 if (InUseList != NULL) {
1146 Self->omInUseList = NULL;
1147 ObjectMonitor *curom;
1148 for (curom = InUseList; curom != NULL; curom = curom->FreeNext) {
1149 InUseTail = curom;
1150 InUseTally++;
1151 }
1152 // TODO debug
1153 assert(Self->omInUseCount == InUseTally, "inuse count off");
1154 Self->omInUseCount = 0;
1155 guarantee (InUseTail != NULL && InUseList != NULL, "invariant");
1156 }
1157
1158 Thread::muxAcquire (&ListLock, "omFlush") ;
1159 if (Tail != NULL) {
1160 Tail->FreeNext = gFreeList ;
1161 gFreeList = List ;
1162 MonitorFreeCount += Tally;
1163 }
1164
1165 if (InUseTail != NULL) {
1166 InUseTail->FreeNext = gOmInUseList;
1167 gOmInUseList = InUseList;
1168 gOmInUseCount += InUseTally;
1169 }
1170
1171 Thread::muxRelease (&ListLock) ;
1172 TEVENT (omFlush) ;
1173 }
1174
1175 // Fast path code shared by multiple functions
1176 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) {
1177 markOop mark = obj->mark();
1178 if (mark->has_monitor()) {
1179 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid");
1180 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header");
1181 return mark->monitor();
1182 }
1183 return ObjectSynchronizer::inflate(Thread::current(), obj);
1184 }
1185
1186
1187 // Note that we could encounter some performance loss through false-sharing as
1188 // multiple locks occupy the same $ line. Padding might be appropriate.
1189
1190
1191 ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) {
1192 // Inflate mutates the heap ...
1193 // Relaxing assertion for bug 6320749.
1194 assert (Universe::verify_in_progress() ||
1195 !SafepointSynchronize::is_at_safepoint(), "invariant") ;
1196
1197 for (;;) {
1198 const markOop mark = object->mark() ;
1199 assert (!mark->has_bias_pattern(), "invariant") ;
1200
1201 // The mark can be in one of the following states:
1202 // * Inflated - just return
1203 // * Stack-locked - coerce it to inflated
1204 // * INFLATING - busy wait for conversion to complete
1205 // * Neutral - aggressively inflate the object.
1206 // * BIASED - Illegal. We should never see this
1207
1208 // CASE: inflated
1209 if (mark->has_monitor()) {
1210 ObjectMonitor * inf = mark->monitor() ;
1211 assert (inf->header()->is_neutral(), "invariant");
1212 assert (inf->object() == object, "invariant") ;
1213 assert (ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1214 return inf ;
1215 }
1216
1217 // CASE: inflation in progress - inflating over a stack-lock.
1218 // Some other thread is converting from stack-locked to inflated.
1219 // Only that thread can complete inflation -- other threads must wait.
1220 // The INFLATING value is transient.
1221 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1222 // We could always eliminate polling by parking the thread on some auxiliary list.
1223 if (mark == markOopDesc::INFLATING()) {
1224 TEVENT (Inflate: spin while INFLATING) ;
1225 ReadStableMark(object) ;
1226 continue ;
1227 }
1228
1229 // CASE: stack-locked
1230 // Could be stack-locked either by this thread or by some other thread.
1231 //
1232 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1233 // to install INFLATING into the mark word. We originally installed INFLATING,
1234 // allocated the objectmonitor, and then finally STed the address of the
1235 // objectmonitor into the mark. This was correct, but artificially lengthened
1236 // the interval in which INFLATED appeared in the mark, thus increasing
1237 // the odds of inflation contention.
1238 //
1239 // We now use per-thread private objectmonitor free lists.
1240 // These list are reprovisioned from the global free list outside the
1241 // critical INFLATING...ST interval. A thread can transfer
1242 // multiple objectmonitors en-mass from the global free list to its local free list.
1243 // This reduces coherency traffic and lock contention on the global free list.
1244 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1245 // before or after the CAS(INFLATING) operation.
1246 // See the comments in omAlloc().
1247
1248 if (mark->has_locker()) {
1249 ObjectMonitor * m = omAlloc (Self) ;
1250 // Optimistically prepare the objectmonitor - anticipate successful CAS
1251 // We do this before the CAS in order to minimize the length of time
1252 // in which INFLATING appears in the mark.
1253 m->Recycle();
1254 m->_Responsible = NULL ;
1255 m->OwnerIsThread = 0 ;
1256 m->_recursions = 0 ;
1257 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // Consider: maintain by type/class
1258
1259 markOop cmp = (markOop) Atomic::cmpxchg_ptr (markOopDesc::INFLATING(), object->mark_addr(), mark) ;
1260 if (cmp != mark) {
1261 omRelease (Self, m, true) ;
1262 continue ; // Interference -- just retry
1263 }
1264
1265 // We've successfully installed INFLATING (0) into the mark-word.
1266 // This is the only case where 0 will appear in a mark-work.
1267 // Only the singular thread that successfully swings the mark-word
1268 // to 0 can perform (or more precisely, complete) inflation.
1269 //
1270 // Why do we CAS a 0 into the mark-word instead of just CASing the
1271 // mark-word from the stack-locked value directly to the new inflated state?
1272 // Consider what happens when a thread unlocks a stack-locked object.
1273 // It attempts to use CAS to swing the displaced header value from the
1274 // on-stack basiclock back into the object header. Recall also that the
1275 // header value (hashcode, etc) can reside in (a) the object header, or
1276 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1277 // header in an objectMonitor. The inflate() routine must copy the header
1278 // value from the basiclock on the owner's stack to the objectMonitor, all
1279 // the while preserving the hashCode stability invariants. If the owner
1280 // decides to release the lock while the value is 0, the unlock will fail
1281 // and control will eventually pass from slow_exit() to inflate. The owner
1282 // will then spin, waiting for the 0 value to disappear. Put another way,
1283 // the 0 causes the owner to stall if the owner happens to try to
1284 // drop the lock (restoring the header from the basiclock to the object)
1285 // while inflation is in-progress. This protocol avoids races that might
1286 // would otherwise permit hashCode values to change or "flicker" for an object.
1287 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1288 // 0 serves as a "BUSY" inflate-in-progress indicator.
1289
1290
1291 // fetch the displaced mark from the owner's stack.
1292 // The owner can't die or unwind past the lock while our INFLATING
1293 // object is in the mark. Furthermore the owner can't complete
1294 // an unlock on the object, either.
1295 markOop dmw = mark->displaced_mark_helper() ;
1296 assert (dmw->is_neutral(), "invariant") ;
1297
1298 // Setup monitor fields to proper values -- prepare the monitor
1299 m->set_header(dmw) ;
1300
1301 // Optimization: if the mark->locker stack address is associated
1302 // with this thread we could simply set m->_owner = Self and
1303 // m->OwnerIsThread = 1. Note that a thread can inflate an object
1304 // that it has stack-locked -- as might happen in wait() -- directly
1305 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1306 m->set_owner(mark->locker());
1307 m->set_object(object);
1308 // TODO-FIXME: assert BasicLock->dhw != 0.
1309
1310 // Must preserve store ordering. The monitor state must
1311 // be stable at the time of publishing the monitor address.
1312 guarantee (object->mark() == markOopDesc::INFLATING(), "invariant") ;
1313 object->release_set_mark(markOopDesc::encode(m));
1314
1315 // Hopefully the performance counters are allocated on distinct cache lines
1316 // to avoid false sharing on MP systems ...
1317 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ;
1318 TEVENT(Inflate: overwrite stacklock) ;
1319 if (TraceMonitorInflation) {
1320 if (object->is_instance()) {
1321 ResourceMark rm;
1322 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1323 (intptr_t) object, (intptr_t) object->mark(),
1324 object->klass()->external_name());
1325 }
1326 }
1327 return m ;
1328 }
1329
1330 // CASE: neutral
1331 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1332 // If we know we're inflating for entry it's better to inflate by swinging a
1333 // pre-locked objectMonitor pointer into the object header. A successful
1334 // CAS inflates the object *and* confers ownership to the inflating thread.
1335 // In the current implementation we use a 2-step mechanism where we CAS()
1336 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1337 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1338 // would be useful.
1339
1340 assert (mark->is_neutral(), "invariant");
1341 ObjectMonitor * m = omAlloc (Self) ;
1342 // prepare m for installation - set monitor to initial state
1343 m->Recycle();
1344 m->set_header(mark);
1345 m->set_owner(NULL);
1346 m->set_object(object);
1347 m->OwnerIsThread = 1 ;
1348 m->_recursions = 0 ;
1349 m->_Responsible = NULL ;
1350 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // consider: keep metastats by type/class
1351
1352 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) {
1353 m->set_object (NULL) ;
1354 m->set_owner (NULL) ;
1355 m->OwnerIsThread = 0 ;
1356 m->Recycle() ;
1357 omRelease (Self, m, true) ;
1358 m = NULL ;
1359 continue ;
1360 // interference - the markword changed - just retry.
1361 // The state-transitions are one-way, so there's no chance of
1362 // live-lock -- "Inflated" is an absorbing state.
1363 }
1364
1365 // Hopefully the performance counters are allocated on distinct
1366 // cache lines to avoid false sharing on MP systems ...
1367 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ;
1368 TEVENT(Inflate: overwrite neutral) ;
1369 if (TraceMonitorInflation) {
1370 if (object->is_instance()) {
1371 ResourceMark rm;
1372 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1373 (intptr_t) object, (intptr_t) object->mark(),
1374 object->klass()->external_name());
1375 }
1376 }
1377 return m ;
1378 }
1379 }
1380
1381 // Note that we could encounter some performance loss through false-sharing as
1382 // multiple locks occupy the same $ line. Padding might be appropriate.
1383
1384
1385 // Deflate_idle_monitors() is called at all safepoints, immediately
1386 // after all mutators are stopped, but before any objects have moved.
1387 // It traverses the list of known monitors, deflating where possible.
1388 // The scavenged monitor are returned to the monitor free list.
1389 //
1390 // Beware that we scavenge at *every* stop-the-world point.
1391 // Having a large number of monitors in-circulation negatively
1392 // impacts the performance of some applications (e.g., PointBase).
1393 // Broadly, we want to minimize the # of monitors in circulation.
1394 //
1395 // We have added a flag, MonitorInUseLists, which creates a list
1396 // of active monitors for each thread. deflate_idle_monitors()
1397 // only scans the per-thread inuse lists. omAlloc() puts all
1398 // assigned monitors on the per-thread list. deflate_idle_monitors()
1399 // returns the non-busy monitors to the global free list.
1400 // When a thread dies, omFlush() adds the list of active monitors for
1401 // that thread to a global gOmInUseList acquiring the
1402 // global list lock. deflate_idle_monitors() acquires the global
1403 // list lock to scan for non-busy monitors to the global free list.
1404 // An alternative could have used a single global inuse list. The
1405 // downside would have been the additional cost of acquiring the global list lock
1406 // for every omAlloc().
1407 //
1408 // Perversely, the heap size -- and thus the STW safepoint rate --
1409 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1410 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1411 // This is an unfortunate aspect of this design.
1412 //
1413
1414 enum ManifestConstants {
1415 ClearResponsibleAtSTW = 0,
1416 MaximumRecheckInterval = 1000
1417 } ;
1418
1419 // Deflate a single monitor if not in use
1420 // Return true if deflated, false if in use
1421 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1422 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) {
1423 bool deflated;
1424 // Normal case ... The monitor is associated with obj.
1425 guarantee (obj->mark() == markOopDesc::encode(mid), "invariant") ;
1426 guarantee (mid == obj->mark()->monitor(), "invariant");
1427 guarantee (mid->header()->is_neutral(), "invariant");
1428
1429 if (mid->is_busy()) {
1430 if (ClearResponsibleAtSTW) mid->_Responsible = NULL ;
1431 deflated = false;
1432 } else {
1433 // Deflate the monitor if it is no longer being used
1434 // It's idle - scavenge and return to the global free list
1435 // plain old deflation ...
1436 TEVENT (deflate_idle_monitors - scavenge1) ;
1437 if (TraceMonitorInflation) {
1438 if (obj->is_instance()) {
1439 ResourceMark rm;
1440 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1441 (intptr_t) obj, (intptr_t) obj->mark(), obj->klass()->external_name());
1442 }
1443 }
1444
1445 // Restore the header back to obj
1446 obj->release_set_mark(mid->header());
1447 mid->clear();
1448
1449 assert (mid->object() == NULL, "invariant") ;
1450
1451 // Move the object to the working free list defined by FreeHead,FreeTail.
1452 if (*FreeHeadp == NULL) *FreeHeadp = mid;
1453 if (*FreeTailp != NULL) {
1454 ObjectMonitor * prevtail = *FreeTailp;
1455 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); // TODO KK
1456 prevtail->FreeNext = mid;
1457 }
1458 *FreeTailp = mid;
1459 deflated = true;
1460 }
1461 return deflated;
1462 }
1463
1464 // Caller acquires ListLock
1465 int ObjectSynchronizer::walk_monitor_list(ObjectMonitor** listheadp,
1466 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) {
1467 ObjectMonitor* mid;
1468 ObjectMonitor* next;
1469 ObjectMonitor* curmidinuse = NULL;
1470 int deflatedcount = 0;
1471
1472 for (mid = *listheadp; mid != NULL; ) {
1473 oop obj = (oop) mid->object();
1474 bool deflated = false;
1475 if (obj != NULL) {
1476 deflated = deflate_monitor(mid, obj, FreeHeadp, FreeTailp);
1477 }
1478 if (deflated) {
1479 // extract from per-thread in-use-list
1480 if (mid == *listheadp) {
1481 *listheadp = mid->FreeNext;
1482 } else if (curmidinuse != NULL) {
1483 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist
1484 }
1485 next = mid->FreeNext;
1486 mid->FreeNext = NULL; // This mid is current tail in the FreeHead list
1487 mid = next;
1488 deflatedcount++;
1489 } else {
1490 curmidinuse = mid;
1491 mid = mid->FreeNext;
1492 }
1493 }
1494 return deflatedcount;
1495 }
1496
1497 void ObjectSynchronizer::deflate_idle_monitors() {
1498 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1499 int nInuse = 0 ; // currently associated with objects
1500 int nInCirculation = 0 ; // extant
1501 int nScavenged = 0 ; // reclaimed
1502 bool deflated = false;
1503
1504 ObjectMonitor * FreeHead = NULL ; // Local SLL of scavenged monitors
1505 ObjectMonitor * FreeTail = NULL ;
1506
1507 TEVENT (deflate_idle_monitors) ;
1508 // Prevent omFlush from changing mids in Thread dtor's during deflation
1509 // And in case the vm thread is acquiring a lock during a safepoint
1510 // See e.g. 6320749
1511 Thread::muxAcquire (&ListLock, "scavenge - return") ;
1512
1513 if (MonitorInUseLists) {
1514 int inUse = 0;
1515 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) {
1516 nInCirculation+= cur->omInUseCount;
1517 int deflatedcount = walk_monitor_list(cur->omInUseList_addr(), &FreeHead, &FreeTail);
1518 cur->omInUseCount-= deflatedcount;
1519 // verifyInUse(cur);
1520 nScavenged += deflatedcount;
1521 nInuse += cur->omInUseCount;
1522 }
1523
1524 // For moribund threads, scan gOmInUseList
1525 if (gOmInUseList) {
1526 nInCirculation += gOmInUseCount;
1527 int deflatedcount = walk_monitor_list((ObjectMonitor **)&gOmInUseList, &FreeHead, &FreeTail);
1528 gOmInUseCount-= deflatedcount;
1529 nScavenged += deflatedcount;
1530 nInuse += gOmInUseCount;
1531 }
1532
1533 } else for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) {
1534 // Iterate over all extant monitors - Scavenge all idle monitors.
1535 assert(block->object() == CHAINMARKER, "must be a block header");
1536 nInCirculation += _BLOCKSIZE ;
1537 for (int i = 1 ; i < _BLOCKSIZE; i++) {
1538 ObjectMonitor* mid = &block[i];
1539 oop obj = (oop) mid->object();
1540
1541 if (obj == NULL) {
1542 // The monitor is not associated with an object.
1543 // The monitor should either be a thread-specific private
1544 // free list or the global free list.
1545 // obj == NULL IMPLIES mid->is_busy() == 0
1546 guarantee (!mid->is_busy(), "invariant") ;
1547 continue ;
1548 }
1549 deflated = deflate_monitor(mid, obj, &FreeHead, &FreeTail);
1550
1551 if (deflated) {
1552 mid->FreeNext = NULL ;
1553 nScavenged ++ ;
1554 } else {
1555 nInuse ++;
1556 }
1557 }
1558 }
1559
1560 MonitorFreeCount += nScavenged;
1561
1562 // Consider: audit gFreeList to ensure that MonitorFreeCount and list agree.
1563
1564 if (ObjectMonitor::Knob_Verbose) {
1565 ::printf ("Deflate: InCirc=%d InUse=%d Scavenged=%d ForceMonitorScavenge=%d : pop=%d free=%d\n",
1566 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge,
1567 MonitorPopulation, MonitorFreeCount) ;
1568 ::fflush(stdout) ;
1569 }
1570
1571 ForceMonitorScavenge = 0; // Reset
1572
1573 // Move the scavenged monitors back to the global free list.
1574 if (FreeHead != NULL) {
1575 guarantee (FreeTail != NULL && nScavenged > 0, "invariant") ;
1576 assert (FreeTail->FreeNext == NULL, "invariant") ;
1577 // constant-time list splice - prepend scavenged segment to gFreeList
1578 FreeTail->FreeNext = gFreeList ;
1579 gFreeList = FreeHead ;
1580 }
1581 Thread::muxRelease (&ListLock) ;
1582
1583 if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged) ;
1584 if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation);
1585
1586 // TODO: Add objectMonitor leak detection.
1587 // Audit/inventory the objectMonitors -- make sure they're all accounted for.
1588 GVars.stwRandom = os::random() ;
1589 GVars.stwCycle ++ ;
1590 }
1591
1592 // Monitor cleanup on JavaThread::exit
1593
1594 // Iterate through monitor cache and attempt to release thread's monitors
1595 // Gives up on a particular monitor if an exception occurs, but continues
1596 // the overall iteration, swallowing the exception.
1597 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1598 private:
1599 TRAPS;
1600
1601 public:
1602 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {}
1603 void do_monitor(ObjectMonitor* mid) {
1604 if (mid->owner() == THREAD) {
1605 (void)mid->complete_exit(CHECK);
1606 }
1607 }
1608 };
1609
1610 // Release all inflated monitors owned by THREAD. Lightweight monitors are
1611 // ignored. This is meant to be called during JNI thread detach which assumes
1612 // all remaining monitors are heavyweight. All exceptions are swallowed.
1613 // Scanning the extant monitor list can be time consuming.
1614 // A simple optimization is to add a per-thread flag that indicates a thread
1615 // called jni_monitorenter() during its lifetime.
1616 //
1617 // Instead of No_Savepoint_Verifier it might be cheaper to
1618 // use an idiom of the form:
1619 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1620 // <code that must not run at safepoint>
1621 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1622 // Since the tests are extremely cheap we could leave them enabled
1623 // for normal product builds.
1624
1625 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) {
1626 assert(THREAD == JavaThread::current(), "must be current Java thread");
1627 No_Safepoint_Verifier nsv ;
1628 ReleaseJavaMonitorsClosure rjmc(THREAD);
1629 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread");
1630 ObjectSynchronizer::monitors_iterate(&rjmc);
1631 Thread::muxRelease(&ListLock);
1632 THREAD->clear_pending_exception();
1633 }
1634
1635 //------------------------------------------------------------------------------
1636 // Non-product code
1637
1638 #ifndef PRODUCT
1639
1640 // Verify all monitors in the monitor cache, the verification is weak.
1641 void ObjectSynchronizer::verify() {
1642 ObjectMonitor* block = gBlockList;
1643 ObjectMonitor* mid;
1644 while (block) {
1645 assert(block->object() == CHAINMARKER, "must be a block header");
1646 for (int i = 1; i < _BLOCKSIZE; i++) {
1647 mid = block + i;
1648 oop object = (oop) mid->object();
1649 if (object != NULL) {
1650 mid->verify();
1651 }
1652 }
1653 block = (ObjectMonitor*) block->FreeNext;
1654 }
1655 }
1656
1657 // Check if monitor belongs to the monitor cache
1658 // The list is grow-only so it's *relatively* safe to traverse
1659 // the list of extant blocks without taking a lock.
1660
1661 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) {
1662 ObjectMonitor* block = gBlockList;
1663
1664 while (block) {
1665 assert(block->object() == CHAINMARKER, "must be a block header");
1666 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) {
1667 address mon = (address) monitor;
1668 address blk = (address) block;
1669 size_t diff = mon - blk;
1670 assert((diff % sizeof(ObjectMonitor)) == 0, "check");
1671 return 1;
1672 }
1673 block = (ObjectMonitor*) block->FreeNext;
1674 }
1675 return 0;
1676 }
1677
1678 #endif