Tue, 11 Nov 2014 11:05:41 +0100
8056071: compiler/whitebox/IsMethodCompilableTest.java fails with 'method() is not compilable after 3 iterations'
Summary: Always use MDO if valid and always compile trivial methods with C1 if available.
Reviewed-by: kvn, iveresov
1 /*
2 * Copyright (c) 1998, 2014, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "classfile/vmSymbols.hpp"
27 #include "memory/resourceArea.hpp"
28 #include "oops/markOop.hpp"
29 #include "oops/oop.inline.hpp"
30 #include "runtime/biasedLocking.hpp"
31 #include "runtime/handles.inline.hpp"
32 #include "runtime/interfaceSupport.hpp"
33 #include "runtime/mutexLocker.hpp"
34 #include "runtime/objectMonitor.hpp"
35 #include "runtime/objectMonitor.inline.hpp"
36 #include "runtime/osThread.hpp"
37 #include "runtime/stubRoutines.hpp"
38 #include "runtime/synchronizer.hpp"
39 #include "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
56 #if defined(__GNUC__) && !defined(PPC64)
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
63 PRAGMA_FORMAT_MUTE_WARNINGS_FOR_GCC
65 // The "core" versions of monitor enter and exit reside in this file.
66 // The interpreter and compilers contain specialized transliterated
67 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(),
68 // for instance. If you make changes here, make sure to modify the
69 // interpreter, and both C1 and C2 fast-path inline locking code emission.
70 //
71 //
72 // -----------------------------------------------------------------------------
74 #ifdef DTRACE_ENABLED
76 // Only bother with this argument setup if dtrace is available
77 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
79 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
80 char* bytes = NULL; \
81 int len = 0; \
82 jlong jtid = SharedRuntime::get_java_tid(thread); \
83 Symbol* klassname = ((oop)(obj))->klass()->name(); \
84 if (klassname != NULL) { \
85 bytes = (char*)klassname->bytes(); \
86 len = klassname->utf8_length(); \
87 }
89 #ifndef USDT2
90 HS_DTRACE_PROBE_DECL5(hotspot, monitor__wait,
91 jlong, uintptr_t, char*, int, long);
92 HS_DTRACE_PROBE_DECL4(hotspot, monitor__waited,
93 jlong, uintptr_t, char*, int);
95 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
96 { \
97 if (DTraceMonitorProbes) { \
98 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
99 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \
100 (monitor), bytes, len, (millis)); \
101 } \
102 }
104 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
105 { \
106 if (DTraceMonitorProbes) { \
107 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
108 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \
109 (uintptr_t)(monitor), bytes, len); \
110 } \
111 }
113 #else /* USDT2 */
115 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
116 { \
117 if (DTraceMonitorProbes) { \
118 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
119 HOTSPOT_MONITOR_WAIT(jtid, \
120 (uintptr_t)(monitor), bytes, len, (millis)); \
121 } \
122 }
124 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_PROBE_WAITED
126 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
127 { \
128 if (DTraceMonitorProbes) { \
129 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
130 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
131 (uintptr_t)(monitor), bytes, len); \
132 } \
133 }
135 #endif /* USDT2 */
136 #else // ndef DTRACE_ENABLED
138 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
139 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
141 #endif // ndef DTRACE_ENABLED
143 // This exists only as a workaround of dtrace bug 6254741
144 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
145 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
146 return 0;
147 }
149 #define NINFLATIONLOCKS 256
150 static volatile intptr_t InflationLocks [NINFLATIONLOCKS] ;
152 ObjectMonitor * ObjectSynchronizer::gBlockList = NULL ;
153 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL ;
154 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL ;
155 int ObjectSynchronizer::gOmInUseCount = 0;
156 static volatile intptr_t ListLock = 0 ; // protects global monitor free-list cache
157 static volatile int MonitorFreeCount = 0 ; // # on gFreeList
158 static volatile int MonitorPopulation = 0 ; // # Extant -- in circulation
159 #define CHAINMARKER (cast_to_oop<intptr_t>(-1))
161 // -----------------------------------------------------------------------------
162 // Fast Monitor Enter/Exit
163 // This the fast monitor enter. The interpreter and compiler use
164 // some assembly copies of this code. Make sure update those code
165 // if the following function is changed. The implementation is
166 // extremely sensitive to race condition. Be careful.
168 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) {
169 if (UseBiasedLocking) {
170 if (!SafepointSynchronize::is_at_safepoint()) {
171 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD);
172 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) {
173 return;
174 }
175 } else {
176 assert(!attempt_rebias, "can not rebias toward VM thread");
177 BiasedLocking::revoke_at_safepoint(obj);
178 }
179 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
180 }
182 slow_enter (obj, lock, THREAD) ;
183 }
185 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) {
186 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here");
187 // if displaced header is null, the previous enter is recursive enter, no-op
188 markOop dhw = lock->displaced_header();
189 markOop mark ;
190 if (dhw == NULL) {
191 // Recursive stack-lock.
192 // Diagnostics -- Could be: stack-locked, inflating, inflated.
193 mark = object->mark() ;
194 assert (!mark->is_neutral(), "invariant") ;
195 if (mark->has_locker() && mark != markOopDesc::INFLATING()) {
196 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant") ;
197 }
198 if (mark->has_monitor()) {
199 ObjectMonitor * m = mark->monitor() ;
200 assert(((oop)(m->object()))->mark() == mark, "invariant") ;
201 assert(m->is_entered(THREAD), "invariant") ;
202 }
203 return ;
204 }
206 mark = object->mark() ;
208 // If the object is stack-locked by the current thread, try to
209 // swing the displaced header from the box back to the mark.
210 if (mark == (markOop) lock) {
211 assert (dhw->is_neutral(), "invariant") ;
212 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) {
213 TEVENT (fast_exit: release stacklock) ;
214 return;
215 }
216 }
218 ObjectSynchronizer::inflate(THREAD, object)->exit (true, THREAD) ;
219 }
221 // -----------------------------------------------------------------------------
222 // Interpreter/Compiler Slow Case
223 // This routine is used to handle interpreter/compiler slow case
224 // We don't need to use fast path here, because it must have been
225 // failed in the interpreter/compiler code.
226 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
227 markOop mark = obj->mark();
228 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
230 if (mark->is_neutral()) {
231 // Anticipate successful CAS -- the ST of the displaced mark must
232 // be visible <= the ST performed by the CAS.
233 lock->set_displaced_header(mark);
234 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) {
235 TEVENT (slow_enter: release stacklock) ;
236 return ;
237 }
238 // Fall through to inflate() ...
239 } else
240 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
241 assert(lock != mark->locker(), "must not re-lock the same lock");
242 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
243 lock->set_displaced_header(NULL);
244 return;
245 }
247 #if 0
248 // The following optimization isn't particularly useful.
249 if (mark->has_monitor() && mark->monitor()->is_entered(THREAD)) {
250 lock->set_displaced_header (NULL) ;
251 return ;
252 }
253 #endif
255 // The object header will never be displaced to this lock,
256 // so it does not matter what the value is, except that it
257 // must be non-zero to avoid looking like a re-entrant lock,
258 // and must not look locked either.
259 lock->set_displaced_header(markOopDesc::unused_mark());
260 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
261 }
263 // This routine is used to handle interpreter/compiler slow case
264 // We don't need to use fast path here, because it must have
265 // failed in the interpreter/compiler code. Simply use the heavy
266 // weight monitor should be ok, unless someone find otherwise.
267 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
268 fast_exit (object, lock, THREAD) ;
269 }
271 // -----------------------------------------------------------------------------
272 // Class Loader support to workaround deadlocks on the class loader lock objects
273 // Also used by GC
274 // complete_exit()/reenter() are used to wait on a nested lock
275 // i.e. to give up an outer lock completely and then re-enter
276 // Used when holding nested locks - lock acquisition order: lock1 then lock2
277 // 1) complete_exit lock1 - saving recursion count
278 // 2) wait on lock2
279 // 3) when notified on lock2, unlock lock2
280 // 4) reenter lock1 with original recursion count
281 // 5) lock lock2
282 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
283 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
284 TEVENT (complete_exit) ;
285 if (UseBiasedLocking) {
286 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
287 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
288 }
290 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
292 return monitor->complete_exit(THREAD);
293 }
295 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
296 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
297 TEVENT (reenter) ;
298 if (UseBiasedLocking) {
299 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
300 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
301 }
303 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
305 monitor->reenter(recursion, THREAD);
306 }
307 // -----------------------------------------------------------------------------
308 // JNI locks on java objects
309 // NOTE: must use heavy weight monitor to handle jni monitor enter
310 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { // possible entry from jni enter
311 // the current locking is from JNI instead of Java code
312 TEVENT (jni_enter) ;
313 if (UseBiasedLocking) {
314 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
315 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
316 }
317 THREAD->set_current_pending_monitor_is_from_java(false);
318 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD);
319 THREAD->set_current_pending_monitor_is_from_java(true);
320 }
322 // NOTE: must use heavy weight monitor to handle jni monitor enter
323 bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) {
324 if (UseBiasedLocking) {
325 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
326 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
327 }
329 ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj());
330 return monitor->try_enter(THREAD);
331 }
334 // NOTE: must use heavy weight monitor to handle jni monitor exit
335 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
336 TEVENT (jni_exit) ;
337 if (UseBiasedLocking) {
338 Handle h_obj(THREAD, obj);
339 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD);
340 obj = h_obj();
341 }
342 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
344 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj);
345 // If this thread has locked the object, exit the monitor. Note: can't use
346 // monitor->check(CHECK); must exit even if an exception is pending.
347 if (monitor->check(THREAD)) {
348 monitor->exit(true, THREAD);
349 }
350 }
352 // -----------------------------------------------------------------------------
353 // Internal VM locks on java objects
354 // standard constructor, allows locking failures
355 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
356 _dolock = doLock;
357 _thread = thread;
358 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
359 _obj = obj;
361 if (_dolock) {
362 TEVENT (ObjectLocker) ;
364 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
365 }
366 }
368 ObjectLocker::~ObjectLocker() {
369 if (_dolock) {
370 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
371 }
372 }
375 // -----------------------------------------------------------------------------
376 // Wait/Notify/NotifyAll
377 // NOTE: must use heavy weight monitor to handle wait()
378 void ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
379 if (UseBiasedLocking) {
380 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
381 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
382 }
383 if (millis < 0) {
384 TEVENT (wait - throw IAX) ;
385 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
386 }
387 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj());
388 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
389 monitor->wait(millis, true, THREAD);
391 /* This dummy call is in place to get around dtrace bug 6254741. Once
392 that's fixed we can uncomment the following line and remove the call */
393 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
394 dtrace_waited_probe(monitor, obj, THREAD);
395 }
397 void ObjectSynchronizer::waitUninterruptibly (Handle obj, jlong millis, TRAPS) {
398 if (UseBiasedLocking) {
399 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
400 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
401 }
402 if (millis < 0) {
403 TEVENT (wait - throw IAX) ;
404 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
405 }
406 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD) ;
407 }
409 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
410 if (UseBiasedLocking) {
411 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
412 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
413 }
415 markOop mark = obj->mark();
416 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
417 return;
418 }
419 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD);
420 }
422 // NOTE: see comment of notify()
423 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
424 if (UseBiasedLocking) {
425 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
426 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
427 }
429 markOop mark = obj->mark();
430 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
431 return;
432 }
433 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD);
434 }
436 // -----------------------------------------------------------------------------
437 // Hash Code handling
438 //
439 // Performance concern:
440 // OrderAccess::storestore() calls release() which STs 0 into the global volatile
441 // OrderAccess::Dummy variable. This store is unnecessary for correctness.
442 // Many threads STing into a common location causes considerable cache migration
443 // or "sloshing" on large SMP system. As such, I avoid using OrderAccess::storestore()
444 // until it's repaired. In some cases OrderAccess::fence() -- which incurs local
445 // latency on the executing processor -- is a better choice as it scales on SMP
446 // systems. See http://blogs.sun.com/dave/entry/biased_locking_in_hotspot for a
447 // discussion of coherency costs. Note that all our current reference platforms
448 // provide strong ST-ST order, so the issue is moot on IA32, x64, and SPARC.
449 //
450 // As a general policy we use "volatile" to control compiler-based reordering
451 // and explicit fences (barriers) to control for architectural reordering performed
452 // by the CPU(s) or platform.
454 struct SharedGlobals {
455 // These are highly shared mostly-read variables.
456 // To avoid false-sharing they need to be the sole occupants of a $ line.
457 double padPrefix [8];
458 volatile int stwRandom ;
459 volatile int stwCycle ;
461 // Hot RW variables -- Sequester to avoid false-sharing
462 double padSuffix [16];
463 volatile int hcSequence ;
464 double padFinal [8] ;
465 } ;
467 static SharedGlobals GVars ;
468 static int MonitorScavengeThreshold = 1000000 ;
469 static volatile int ForceMonitorScavenge = 0 ; // Scavenge required and pending
471 static markOop ReadStableMark (oop obj) {
472 markOop mark = obj->mark() ;
473 if (!mark->is_being_inflated()) {
474 return mark ; // normal fast-path return
475 }
477 int its = 0 ;
478 for (;;) {
479 markOop mark = obj->mark() ;
480 if (!mark->is_being_inflated()) {
481 return mark ; // normal fast-path return
482 }
484 // The object is being inflated by some other thread.
485 // The caller of ReadStableMark() must wait for inflation to complete.
486 // Avoid live-lock
487 // TODO: consider calling SafepointSynchronize::do_call_back() while
488 // spinning to see if there's a safepoint pending. If so, immediately
489 // yielding or blocking would be appropriate. Avoid spinning while
490 // there is a safepoint pending.
491 // TODO: add inflation contention performance counters.
492 // TODO: restrict the aggregate number of spinners.
494 ++its ;
495 if (its > 10000 || !os::is_MP()) {
496 if (its & 1) {
497 os::NakedYield() ;
498 TEVENT (Inflate: INFLATING - yield) ;
499 } else {
500 // Note that the following code attenuates the livelock problem but is not
501 // a complete remedy. A more complete solution would require that the inflating
502 // thread hold the associated inflation lock. The following code simply restricts
503 // the number of spinners to at most one. We'll have N-2 threads blocked
504 // on the inflationlock, 1 thread holding the inflation lock and using
505 // a yield/park strategy, and 1 thread in the midst of inflation.
506 // A more refined approach would be to change the encoding of INFLATING
507 // to allow encapsulation of a native thread pointer. Threads waiting for
508 // inflation to complete would use CAS to push themselves onto a singly linked
509 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
510 // and calling park(). When inflation was complete the thread that accomplished inflation
511 // would detach the list and set the markword to inflated with a single CAS and
512 // then for each thread on the list, set the flag and unpark() the thread.
513 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease
514 // wakes at most one thread whereas we need to wake the entire list.
515 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1) ;
516 int YieldThenBlock = 0 ;
517 assert (ix >= 0 && ix < NINFLATIONLOCKS, "invariant") ;
518 assert ((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant") ;
519 Thread::muxAcquire (InflationLocks + ix, "InflationLock") ;
520 while (obj->mark() == markOopDesc::INFLATING()) {
521 // Beware: NakedYield() is advisory and has almost no effect on some platforms
522 // so we periodically call Self->_ParkEvent->park(1).
523 // We use a mixed spin/yield/block mechanism.
524 if ((YieldThenBlock++) >= 16) {
525 Thread::current()->_ParkEvent->park(1) ;
526 } else {
527 os::NakedYield() ;
528 }
529 }
530 Thread::muxRelease (InflationLocks + ix ) ;
531 TEVENT (Inflate: INFLATING - yield/park) ;
532 }
533 } else {
534 SpinPause() ; // SMP-polite spinning
535 }
536 }
537 }
539 // hashCode() generation :
540 //
541 // Possibilities:
542 // * MD5Digest of {obj,stwRandom}
543 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function.
544 // * A DES- or AES-style SBox[] mechanism
545 // * One of the Phi-based schemes, such as:
546 // 2654435761 = 2^32 * Phi (golden ratio)
547 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ;
548 // * A variation of Marsaglia's shift-xor RNG scheme.
549 // * (obj ^ stwRandom) is appealing, but can result
550 // in undesirable regularity in the hashCode values of adjacent objects
551 // (objects allocated back-to-back, in particular). This could potentially
552 // result in hashtable collisions and reduced hashtable efficiency.
553 // There are simple ways to "diffuse" the middle address bits over the
554 // generated hashCode values:
555 //
557 static inline intptr_t get_next_hash(Thread * Self, oop obj) {
558 intptr_t value = 0 ;
559 if (hashCode == 0) {
560 // This form uses an unguarded global Park-Miller RNG,
561 // so it's possible for two threads to race and generate the same RNG.
562 // On MP system we'll have lots of RW access to a global, so the
563 // mechanism induces lots of coherency traffic.
564 value = os::random() ;
565 } else
566 if (hashCode == 1) {
567 // This variation has the property of being stable (idempotent)
568 // between STW operations. This can be useful in some of the 1-0
569 // synchronization schemes.
570 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3 ;
571 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom ;
572 } else
573 if (hashCode == 2) {
574 value = 1 ; // for sensitivity testing
575 } else
576 if (hashCode == 3) {
577 value = ++GVars.hcSequence ;
578 } else
579 if (hashCode == 4) {
580 value = cast_from_oop<intptr_t>(obj) ;
581 } else {
582 // Marsaglia's xor-shift scheme with thread-specific state
583 // This is probably the best overall implementation -- we'll
584 // likely make this the default in future releases.
585 unsigned t = Self->_hashStateX ;
586 t ^= (t << 11) ;
587 Self->_hashStateX = Self->_hashStateY ;
588 Self->_hashStateY = Self->_hashStateZ ;
589 Self->_hashStateZ = Self->_hashStateW ;
590 unsigned v = Self->_hashStateW ;
591 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)) ;
592 Self->_hashStateW = v ;
593 value = v ;
594 }
596 value &= markOopDesc::hash_mask;
597 if (value == 0) value = 0xBAD ;
598 assert (value != markOopDesc::no_hash, "invariant") ;
599 TEVENT (hashCode: GENERATE) ;
600 return value;
601 }
602 //
603 intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) {
604 if (UseBiasedLocking) {
605 // NOTE: many places throughout the JVM do not expect a safepoint
606 // to be taken here, in particular most operations on perm gen
607 // objects. However, we only ever bias Java instances and all of
608 // the call sites of identity_hash that might revoke biases have
609 // been checked to make sure they can handle a safepoint. The
610 // added check of the bias pattern is to avoid useless calls to
611 // thread-local storage.
612 if (obj->mark()->has_bias_pattern()) {
613 // Box and unbox the raw reference just in case we cause a STW safepoint.
614 Handle hobj (Self, obj) ;
615 // Relaxing assertion for bug 6320749.
616 assert (Universe::verify_in_progress() ||
617 !SafepointSynchronize::is_at_safepoint(),
618 "biases should not be seen by VM thread here");
619 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
620 obj = hobj() ;
621 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
622 }
623 }
625 // hashCode() is a heap mutator ...
626 // Relaxing assertion for bug 6320749.
627 assert (Universe::verify_in_progress() ||
628 !SafepointSynchronize::is_at_safepoint(), "invariant") ;
629 assert (Universe::verify_in_progress() ||
630 Self->is_Java_thread() , "invariant") ;
631 assert (Universe::verify_in_progress() ||
632 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ;
634 ObjectMonitor* monitor = NULL;
635 markOop temp, test;
636 intptr_t hash;
637 markOop mark = ReadStableMark (obj);
639 // object should remain ineligible for biased locking
640 assert (!mark->has_bias_pattern(), "invariant") ;
642 if (mark->is_neutral()) {
643 hash = mark->hash(); // this is a normal header
644 if (hash) { // if it has hash, just return it
645 return hash;
646 }
647 hash = get_next_hash(Self, obj); // allocate a new hash code
648 temp = mark->copy_set_hash(hash); // merge the hash code into header
649 // use (machine word version) atomic operation to install the hash
650 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark);
651 if (test == mark) {
652 return hash;
653 }
654 // If atomic operation failed, we must inflate the header
655 // into heavy weight monitor. We could add more code here
656 // for fast path, but it does not worth the complexity.
657 } else if (mark->has_monitor()) {
658 monitor = mark->monitor();
659 temp = monitor->header();
660 assert (temp->is_neutral(), "invariant") ;
661 hash = temp->hash();
662 if (hash) {
663 return hash;
664 }
665 // Skip to the following code to reduce code size
666 } else if (Self->is_lock_owned((address)mark->locker())) {
667 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
668 assert (temp->is_neutral(), "invariant") ;
669 hash = temp->hash(); // by current thread, check if the displaced
670 if (hash) { // header contains hash code
671 return hash;
672 }
673 // WARNING:
674 // The displaced header is strictly immutable.
675 // It can NOT be changed in ANY cases. So we have
676 // to inflate the header into heavyweight monitor
677 // even the current thread owns the lock. The reason
678 // is the BasicLock (stack slot) will be asynchronously
679 // read by other threads during the inflate() function.
680 // Any change to stack may not propagate to other threads
681 // correctly.
682 }
684 // Inflate the monitor to set hash code
685 monitor = ObjectSynchronizer::inflate(Self, obj);
686 // Load displaced header and check it has hash code
687 mark = monitor->header();
688 assert (mark->is_neutral(), "invariant") ;
689 hash = mark->hash();
690 if (hash == 0) {
691 hash = get_next_hash(Self, obj);
692 temp = mark->copy_set_hash(hash); // merge hash code into header
693 assert (temp->is_neutral(), "invariant") ;
694 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark);
695 if (test != mark) {
696 // The only update to the header in the monitor (outside GC)
697 // is install the hash code. If someone add new usage of
698 // displaced header, please update this code
699 hash = test->hash();
700 assert (test->is_neutral(), "invariant") ;
701 assert (hash != 0, "Trivial unexpected object/monitor header usage.");
702 }
703 }
704 // We finally get the hash
705 return hash;
706 }
708 // Deprecated -- use FastHashCode() instead.
710 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
711 return FastHashCode (Thread::current(), obj()) ;
712 }
715 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
716 Handle h_obj) {
717 if (UseBiasedLocking) {
718 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
719 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
720 }
722 assert(thread == JavaThread::current(), "Can only be called on current thread");
723 oop obj = h_obj();
725 markOop mark = ReadStableMark (obj) ;
727 // Uncontended case, header points to stack
728 if (mark->has_locker()) {
729 return thread->is_lock_owned((address)mark->locker());
730 }
731 // Contended case, header points to ObjectMonitor (tagged pointer)
732 if (mark->has_monitor()) {
733 ObjectMonitor* monitor = mark->monitor();
734 return monitor->is_entered(thread) != 0 ;
735 }
736 // Unlocked case, header in place
737 assert(mark->is_neutral(), "sanity check");
738 return false;
739 }
741 // Be aware of this method could revoke bias of the lock object.
742 // This method querys the ownership of the lock handle specified by 'h_obj'.
743 // If the current thread owns the lock, it returns owner_self. If no
744 // thread owns the lock, it returns owner_none. Otherwise, it will return
745 // ower_other.
746 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
747 (JavaThread *self, Handle h_obj) {
748 // The caller must beware this method can revoke bias, and
749 // revocation can result in a safepoint.
750 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ;
751 assert (self->thread_state() != _thread_blocked , "invariant") ;
753 // Possible mark states: neutral, biased, stack-locked, inflated
755 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
756 // CASE: biased
757 BiasedLocking::revoke_and_rebias(h_obj, false, self);
758 assert(!h_obj->mark()->has_bias_pattern(),
759 "biases should be revoked by now");
760 }
762 assert(self == JavaThread::current(), "Can only be called on current thread");
763 oop obj = h_obj();
764 markOop mark = ReadStableMark (obj) ;
766 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
767 if (mark->has_locker()) {
768 return self->is_lock_owned((address)mark->locker()) ?
769 owner_self : owner_other;
770 }
772 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor.
773 // The Object:ObjectMonitor relationship is stable as long as we're
774 // not at a safepoint.
775 if (mark->has_monitor()) {
776 void * owner = mark->monitor()->_owner ;
777 if (owner == NULL) return owner_none ;
778 return (owner == self ||
779 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
780 }
782 // CASE: neutral
783 assert(mark->is_neutral(), "sanity check");
784 return owner_none ; // it's unlocked
785 }
787 // FIXME: jvmti should call this
788 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) {
789 if (UseBiasedLocking) {
790 if (SafepointSynchronize::is_at_safepoint()) {
791 BiasedLocking::revoke_at_safepoint(h_obj);
792 } else {
793 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
794 }
795 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
796 }
798 oop obj = h_obj();
799 address owner = NULL;
801 markOop mark = ReadStableMark (obj) ;
803 // Uncontended case, header points to stack
804 if (mark->has_locker()) {
805 owner = (address) mark->locker();
806 }
808 // Contended case, header points to ObjectMonitor (tagged pointer)
809 if (mark->has_monitor()) {
810 ObjectMonitor* monitor = mark->monitor();
811 assert(monitor != NULL, "monitor should be non-null");
812 owner = (address) monitor->owner();
813 }
815 if (owner != NULL) {
816 // owning_thread_from_monitor_owner() may also return NULL here
817 return Threads::owning_thread_from_monitor_owner(owner, doLock);
818 }
820 // Unlocked case, header in place
821 // Cannot have assertion since this object may have been
822 // locked by another thread when reaching here.
823 // assert(mark->is_neutral(), "sanity check");
825 return NULL;
826 }
827 // Visitors ...
829 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
830 ObjectMonitor* block = gBlockList;
831 ObjectMonitor* mid;
832 while (block) {
833 assert(block->object() == CHAINMARKER, "must be a block header");
834 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
835 mid = block + i;
836 oop object = (oop) mid->object();
837 if (object != NULL) {
838 closure->do_monitor(mid);
839 }
840 }
841 block = (ObjectMonitor*) block->FreeNext;
842 }
843 }
845 // Get the next block in the block list.
846 static inline ObjectMonitor* next(ObjectMonitor* block) {
847 assert(block->object() == CHAINMARKER, "must be a block header");
848 block = block->FreeNext ;
849 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
850 return block;
851 }
854 void ObjectSynchronizer::oops_do(OopClosure* f) {
855 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
856 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) {
857 assert(block->object() == CHAINMARKER, "must be a block header");
858 for (int i = 1; i < _BLOCKSIZE; i++) {
859 ObjectMonitor* mid = &block[i];
860 if (mid->object() != NULL) {
861 f->do_oop((oop*)mid->object_addr());
862 }
863 }
864 }
865 }
868 // -----------------------------------------------------------------------------
869 // ObjectMonitor Lifecycle
870 // -----------------------
871 // Inflation unlinks monitors from the global gFreeList and
872 // associates them with objects. Deflation -- which occurs at
873 // STW-time -- disassociates idle monitors from objects. Such
874 // scavenged monitors are returned to the gFreeList.
875 //
876 // The global list is protected by ListLock. All the critical sections
877 // are short and operate in constant-time.
878 //
879 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
880 //
881 // Lifecycle:
882 // -- unassigned and on the global free list
883 // -- unassigned and on a thread's private omFreeList
884 // -- assigned to an object. The object is inflated and the mark refers
885 // to the objectmonitor.
886 //
889 // Constraining monitor pool growth via MonitorBound ...
890 //
891 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
892 // the rate of scavenging is driven primarily by GC. As such, we can find
893 // an inordinate number of monitors in circulation.
894 // To avoid that scenario we can artificially induce a STW safepoint
895 // if the pool appears to be growing past some reasonable bound.
896 // Generally we favor time in space-time tradeoffs, but as there's no
897 // natural back-pressure on the # of extant monitors we need to impose some
898 // type of limit. Beware that if MonitorBound is set to too low a value
899 // we could just loop. In addition, if MonitorBound is set to a low value
900 // we'll incur more safepoints, which are harmful to performance.
901 // See also: GuaranteedSafepointInterval
902 //
903 // The current implementation uses asynchronous VM operations.
904 //
906 static void InduceScavenge (Thread * Self, const char * Whence) {
907 // Induce STW safepoint to trim monitors
908 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
909 // More precisely, trigger an asynchronous STW safepoint as the number
910 // of active monitors passes the specified threshold.
911 // TODO: assert thread state is reasonable
913 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
914 if (ObjectMonitor::Knob_Verbose) {
915 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ;
916 ::fflush(stdout) ;
917 }
918 // Induce a 'null' safepoint to scavenge monitors
919 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
920 // to the VMthread and have a lifespan longer than that of this activation record.
921 // The VMThread will delete the op when completed.
922 VMThread::execute (new VM_ForceAsyncSafepoint()) ;
924 if (ObjectMonitor::Knob_Verbose) {
925 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ;
926 ::fflush(stdout) ;
927 }
928 }
929 }
930 /* Too slow for general assert or debug
931 void ObjectSynchronizer::verifyInUse (Thread *Self) {
932 ObjectMonitor* mid;
933 int inusetally = 0;
934 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) {
935 inusetally ++;
936 }
937 assert(inusetally == Self->omInUseCount, "inuse count off");
939 int freetally = 0;
940 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) {
941 freetally ++;
942 }
943 assert(freetally == Self->omFreeCount, "free count off");
944 }
945 */
946 ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) {
947 // A large MAXPRIVATE value reduces both list lock contention
948 // and list coherency traffic, but also tends to increase the
949 // number of objectMonitors in circulation as well as the STW
950 // scavenge costs. As usual, we lean toward time in space-time
951 // tradeoffs.
952 const int MAXPRIVATE = 1024 ;
953 for (;;) {
954 ObjectMonitor * m ;
956 // 1: try to allocate from the thread's local omFreeList.
957 // Threads will attempt to allocate first from their local list, then
958 // from the global list, and only after those attempts fail will the thread
959 // attempt to instantiate new monitors. Thread-local free lists take
960 // heat off the ListLock and improve allocation latency, as well as reducing
961 // coherency traffic on the shared global list.
962 m = Self->omFreeList ;
963 if (m != NULL) {
964 Self->omFreeList = m->FreeNext ;
965 Self->omFreeCount -- ;
966 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene
967 guarantee (m->object() == NULL, "invariant") ;
968 if (MonitorInUseLists) {
969 m->FreeNext = Self->omInUseList;
970 Self->omInUseList = m;
971 Self->omInUseCount ++;
972 // verifyInUse(Self);
973 } else {
974 m->FreeNext = NULL;
975 }
976 return m ;
977 }
979 // 2: try to allocate from the global gFreeList
980 // CONSIDER: use muxTry() instead of muxAcquire().
981 // If the muxTry() fails then drop immediately into case 3.
982 // If we're using thread-local free lists then try
983 // to reprovision the caller's free list.
984 if (gFreeList != NULL) {
985 // Reprovision the thread's omFreeList.
986 // Use bulk transfers to reduce the allocation rate and heat
987 // on various locks.
988 Thread::muxAcquire (&ListLock, "omAlloc") ;
989 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL; ) {
990 MonitorFreeCount --;
991 ObjectMonitor * take = gFreeList ;
992 gFreeList = take->FreeNext ;
993 guarantee (take->object() == NULL, "invariant") ;
994 guarantee (!take->is_busy(), "invariant") ;
995 take->Recycle() ;
996 omRelease (Self, take, false) ;
997 }
998 Thread::muxRelease (&ListLock) ;
999 Self->omFreeProvision += 1 + (Self->omFreeProvision/2) ;
1000 if (Self->omFreeProvision > MAXPRIVATE ) Self->omFreeProvision = MAXPRIVATE ;
1001 TEVENT (omFirst - reprovision) ;
1003 const int mx = MonitorBound ;
1004 if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) {
1005 // We can't safely induce a STW safepoint from omAlloc() as our thread
1006 // state may not be appropriate for such activities and callers may hold
1007 // naked oops, so instead we defer the action.
1008 InduceScavenge (Self, "omAlloc") ;
1009 }
1010 continue;
1011 }
1013 // 3: allocate a block of new ObjectMonitors
1014 // Both the local and global free lists are empty -- resort to malloc().
1015 // In the current implementation objectMonitors are TSM - immortal.
1016 assert (_BLOCKSIZE > 1, "invariant") ;
1017 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE];
1019 // NOTE: (almost) no way to recover if allocation failed.
1020 // We might be able to induce a STW safepoint and scavenge enough
1021 // objectMonitors to permit progress.
1022 if (temp == NULL) {
1023 vm_exit_out_of_memory (sizeof (ObjectMonitor[_BLOCKSIZE]), OOM_MALLOC_ERROR,
1024 "Allocate ObjectMonitors");
1025 }
1027 // Format the block.
1028 // initialize the linked list, each monitor points to its next
1029 // forming the single linked free list, the very first monitor
1030 // will points to next block, which forms the block list.
1031 // The trick of using the 1st element in the block as gBlockList
1032 // linkage should be reconsidered. A better implementation would
1033 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1035 for (int i = 1; i < _BLOCKSIZE ; i++) {
1036 temp[i].FreeNext = &temp[i+1];
1037 }
1039 // terminate the last monitor as the end of list
1040 temp[_BLOCKSIZE - 1].FreeNext = NULL ;
1042 // Element [0] is reserved for global list linkage
1043 temp[0].set_object(CHAINMARKER);
1045 // Consider carving out this thread's current request from the
1046 // block in hand. This avoids some lock traffic and redundant
1047 // list activity.
1049 // Acquire the ListLock to manipulate BlockList and FreeList.
1050 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1051 Thread::muxAcquire (&ListLock, "omAlloc [2]") ;
1052 MonitorPopulation += _BLOCKSIZE-1;
1053 MonitorFreeCount += _BLOCKSIZE-1;
1055 // Add the new block to the list of extant blocks (gBlockList).
1056 // The very first objectMonitor in a block is reserved and dedicated.
1057 // It serves as blocklist "next" linkage.
1058 temp[0].FreeNext = gBlockList;
1059 gBlockList = temp;
1061 // Add the new string of objectMonitors to the global free list
1062 temp[_BLOCKSIZE - 1].FreeNext = gFreeList ;
1063 gFreeList = temp + 1;
1064 Thread::muxRelease (&ListLock) ;
1065 TEVENT (Allocate block of monitors) ;
1066 }
1067 }
1069 // Place "m" on the caller's private per-thread omFreeList.
1070 // In practice there's no need to clamp or limit the number of
1071 // monitors on a thread's omFreeList as the only time we'll call
1072 // omRelease is to return a monitor to the free list after a CAS
1073 // attempt failed. This doesn't allow unbounded #s of monitors to
1074 // accumulate on a thread's free list.
1075 //
1077 void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m, bool fromPerThreadAlloc) {
1078 guarantee (m->object() == NULL, "invariant") ;
1080 // Remove from omInUseList
1081 if (MonitorInUseLists && fromPerThreadAlloc) {
1082 ObjectMonitor* curmidinuse = NULL;
1083 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; ) {
1084 if (m == mid) {
1085 // extract from per-thread in-use-list
1086 if (mid == Self->omInUseList) {
1087 Self->omInUseList = mid->FreeNext;
1088 } else if (curmidinuse != NULL) {
1089 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist
1090 }
1091 Self->omInUseCount --;
1092 // verifyInUse(Self);
1093 break;
1094 } else {
1095 curmidinuse = mid;
1096 mid = mid->FreeNext;
1097 }
1098 }
1099 }
1101 // FreeNext is used for both onInUseList and omFreeList, so clear old before setting new
1102 m->FreeNext = Self->omFreeList ;
1103 Self->omFreeList = m ;
1104 Self->omFreeCount ++ ;
1105 }
1107 // Return the monitors of a moribund thread's local free list to
1108 // the global free list. Typically a thread calls omFlush() when
1109 // it's dying. We could also consider having the VM thread steal
1110 // monitors from threads that have not run java code over a few
1111 // consecutive STW safepoints. Relatedly, we might decay
1112 // omFreeProvision at STW safepoints.
1113 //
1114 // Also return the monitors of a moribund thread"s omInUseList to
1115 // a global gOmInUseList under the global list lock so these
1116 // will continue to be scanned.
1117 //
1118 // We currently call omFlush() from the Thread:: dtor _after the thread
1119 // has been excised from the thread list and is no longer a mutator.
1120 // That means that omFlush() can run concurrently with a safepoint and
1121 // the scavenge operator. Calling omFlush() from JavaThread::exit() might
1122 // be a better choice as we could safely reason that that the JVM is
1123 // not at a safepoint at the time of the call, and thus there could
1124 // be not inopportune interleavings between omFlush() and the scavenge
1125 // operator.
1127 void ObjectSynchronizer::omFlush (Thread * Self) {
1128 ObjectMonitor * List = Self->omFreeList ; // Null-terminated SLL
1129 Self->omFreeList = NULL ;
1130 ObjectMonitor * Tail = NULL ;
1131 int Tally = 0;
1132 if (List != NULL) {
1133 ObjectMonitor * s ;
1134 for (s = List ; s != NULL ; s = s->FreeNext) {
1135 Tally ++ ;
1136 Tail = s ;
1137 guarantee (s->object() == NULL, "invariant") ;
1138 guarantee (!s->is_busy(), "invariant") ;
1139 s->set_owner (NULL) ; // redundant but good hygiene
1140 TEVENT (omFlush - Move one) ;
1141 }
1142 guarantee (Tail != NULL && List != NULL, "invariant") ;
1143 }
1145 ObjectMonitor * InUseList = Self->omInUseList;
1146 ObjectMonitor * InUseTail = NULL ;
1147 int InUseTally = 0;
1148 if (InUseList != NULL) {
1149 Self->omInUseList = NULL;
1150 ObjectMonitor *curom;
1151 for (curom = InUseList; curom != NULL; curom = curom->FreeNext) {
1152 InUseTail = curom;
1153 InUseTally++;
1154 }
1155 // TODO debug
1156 assert(Self->omInUseCount == InUseTally, "inuse count off");
1157 Self->omInUseCount = 0;
1158 guarantee (InUseTail != NULL && InUseList != NULL, "invariant");
1159 }
1161 Thread::muxAcquire (&ListLock, "omFlush") ;
1162 if (Tail != NULL) {
1163 Tail->FreeNext = gFreeList ;
1164 gFreeList = List ;
1165 MonitorFreeCount += Tally;
1166 }
1168 if (InUseTail != NULL) {
1169 InUseTail->FreeNext = gOmInUseList;
1170 gOmInUseList = InUseList;
1171 gOmInUseCount += InUseTally;
1172 }
1174 Thread::muxRelease (&ListLock) ;
1175 TEVENT (omFlush) ;
1176 }
1178 // Fast path code shared by multiple functions
1179 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) {
1180 markOop mark = obj->mark();
1181 if (mark->has_monitor()) {
1182 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid");
1183 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header");
1184 return mark->monitor();
1185 }
1186 return ObjectSynchronizer::inflate(Thread::current(), obj);
1187 }
1190 // Note that we could encounter some performance loss through false-sharing as
1191 // multiple locks occupy the same $ line. Padding might be appropriate.
1194 ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) {
1195 // Inflate mutates the heap ...
1196 // Relaxing assertion for bug 6320749.
1197 assert (Universe::verify_in_progress() ||
1198 !SafepointSynchronize::is_at_safepoint(), "invariant") ;
1200 for (;;) {
1201 const markOop mark = object->mark() ;
1202 assert (!mark->has_bias_pattern(), "invariant") ;
1204 // The mark can be in one of the following states:
1205 // * Inflated - just return
1206 // * Stack-locked - coerce it to inflated
1207 // * INFLATING - busy wait for conversion to complete
1208 // * Neutral - aggressively inflate the object.
1209 // * BIASED - Illegal. We should never see this
1211 // CASE: inflated
1212 if (mark->has_monitor()) {
1213 ObjectMonitor * inf = mark->monitor() ;
1214 assert (inf->header()->is_neutral(), "invariant");
1215 assert (inf->object() == object, "invariant") ;
1216 assert (ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1217 return inf ;
1218 }
1220 // CASE: inflation in progress - inflating over a stack-lock.
1221 // Some other thread is converting from stack-locked to inflated.
1222 // Only that thread can complete inflation -- other threads must wait.
1223 // The INFLATING value is transient.
1224 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1225 // We could always eliminate polling by parking the thread on some auxiliary list.
1226 if (mark == markOopDesc::INFLATING()) {
1227 TEVENT (Inflate: spin while INFLATING) ;
1228 ReadStableMark(object) ;
1229 continue ;
1230 }
1232 // CASE: stack-locked
1233 // Could be stack-locked either by this thread or by some other thread.
1234 //
1235 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1236 // to install INFLATING into the mark word. We originally installed INFLATING,
1237 // allocated the objectmonitor, and then finally STed the address of the
1238 // objectmonitor into the mark. This was correct, but artificially lengthened
1239 // the interval in which INFLATED appeared in the mark, thus increasing
1240 // the odds of inflation contention.
1241 //
1242 // We now use per-thread private objectmonitor free lists.
1243 // These list are reprovisioned from the global free list outside the
1244 // critical INFLATING...ST interval. A thread can transfer
1245 // multiple objectmonitors en-mass from the global free list to its local free list.
1246 // This reduces coherency traffic and lock contention on the global free list.
1247 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1248 // before or after the CAS(INFLATING) operation.
1249 // See the comments in omAlloc().
1251 if (mark->has_locker()) {
1252 ObjectMonitor * m = omAlloc (Self) ;
1253 // Optimistically prepare the objectmonitor - anticipate successful CAS
1254 // We do this before the CAS in order to minimize the length of time
1255 // in which INFLATING appears in the mark.
1256 m->Recycle();
1257 m->_Responsible = NULL ;
1258 m->OwnerIsThread = 0 ;
1259 m->_recursions = 0 ;
1260 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // Consider: maintain by type/class
1262 markOop cmp = (markOop) Atomic::cmpxchg_ptr (markOopDesc::INFLATING(), object->mark_addr(), mark) ;
1263 if (cmp != mark) {
1264 omRelease (Self, m, true) ;
1265 continue ; // Interference -- just retry
1266 }
1268 // We've successfully installed INFLATING (0) into the mark-word.
1269 // This is the only case where 0 will appear in a mark-work.
1270 // Only the singular thread that successfully swings the mark-word
1271 // to 0 can perform (or more precisely, complete) inflation.
1272 //
1273 // Why do we CAS a 0 into the mark-word instead of just CASing the
1274 // mark-word from the stack-locked value directly to the new inflated state?
1275 // Consider what happens when a thread unlocks a stack-locked object.
1276 // It attempts to use CAS to swing the displaced header value from the
1277 // on-stack basiclock back into the object header. Recall also that the
1278 // header value (hashcode, etc) can reside in (a) the object header, or
1279 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1280 // header in an objectMonitor. The inflate() routine must copy the header
1281 // value from the basiclock on the owner's stack to the objectMonitor, all
1282 // the while preserving the hashCode stability invariants. If the owner
1283 // decides to release the lock while the value is 0, the unlock will fail
1284 // and control will eventually pass from slow_exit() to inflate. The owner
1285 // will then spin, waiting for the 0 value to disappear. Put another way,
1286 // the 0 causes the owner to stall if the owner happens to try to
1287 // drop the lock (restoring the header from the basiclock to the object)
1288 // while inflation is in-progress. This protocol avoids races that might
1289 // would otherwise permit hashCode values to change or "flicker" for an object.
1290 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1291 // 0 serves as a "BUSY" inflate-in-progress indicator.
1294 // fetch the displaced mark from the owner's stack.
1295 // The owner can't die or unwind past the lock while our INFLATING
1296 // object is in the mark. Furthermore the owner can't complete
1297 // an unlock on the object, either.
1298 markOop dmw = mark->displaced_mark_helper() ;
1299 assert (dmw->is_neutral(), "invariant") ;
1301 // Setup monitor fields to proper values -- prepare the monitor
1302 m->set_header(dmw) ;
1304 // Optimization: if the mark->locker stack address is associated
1305 // with this thread we could simply set m->_owner = Self and
1306 // m->OwnerIsThread = 1. Note that a thread can inflate an object
1307 // that it has stack-locked -- as might happen in wait() -- directly
1308 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1309 m->set_owner(mark->locker());
1310 m->set_object(object);
1311 // TODO-FIXME: assert BasicLock->dhw != 0.
1313 // Must preserve store ordering. The monitor state must
1314 // be stable at the time of publishing the monitor address.
1315 guarantee (object->mark() == markOopDesc::INFLATING(), "invariant") ;
1316 object->release_set_mark(markOopDesc::encode(m));
1318 // Hopefully the performance counters are allocated on distinct cache lines
1319 // to avoid false sharing on MP systems ...
1320 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ;
1321 TEVENT(Inflate: overwrite stacklock) ;
1322 if (TraceMonitorInflation) {
1323 if (object->is_instance()) {
1324 ResourceMark rm;
1325 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1326 (void *) object, (intptr_t) object->mark(),
1327 object->klass()->external_name());
1328 }
1329 }
1330 return m ;
1331 }
1333 // CASE: neutral
1334 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1335 // If we know we're inflating for entry it's better to inflate by swinging a
1336 // pre-locked objectMonitor pointer into the object header. A successful
1337 // CAS inflates the object *and* confers ownership to the inflating thread.
1338 // In the current implementation we use a 2-step mechanism where we CAS()
1339 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1340 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1341 // would be useful.
1343 assert (mark->is_neutral(), "invariant");
1344 ObjectMonitor * m = omAlloc (Self) ;
1345 // prepare m for installation - set monitor to initial state
1346 m->Recycle();
1347 m->set_header(mark);
1348 m->set_owner(NULL);
1349 m->set_object(object);
1350 m->OwnerIsThread = 1 ;
1351 m->_recursions = 0 ;
1352 m->_Responsible = NULL ;
1353 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit ; // consider: keep metastats by type/class
1355 if (Atomic::cmpxchg_ptr (markOopDesc::encode(m), object->mark_addr(), mark) != mark) {
1356 m->set_object (NULL) ;
1357 m->set_owner (NULL) ;
1358 m->OwnerIsThread = 0 ;
1359 m->Recycle() ;
1360 omRelease (Self, m, true) ;
1361 m = NULL ;
1362 continue ;
1363 // interference - the markword changed - just retry.
1364 // The state-transitions are one-way, so there's no chance of
1365 // live-lock -- "Inflated" is an absorbing state.
1366 }
1368 // Hopefully the performance counters are allocated on distinct
1369 // cache lines to avoid false sharing on MP systems ...
1370 if (ObjectMonitor::_sync_Inflations != NULL) ObjectMonitor::_sync_Inflations->inc() ;
1371 TEVENT(Inflate: overwrite neutral) ;
1372 if (TraceMonitorInflation) {
1373 if (object->is_instance()) {
1374 ResourceMark rm;
1375 tty->print_cr("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1376 (void *) object, (intptr_t) object->mark(),
1377 object->klass()->external_name());
1378 }
1379 }
1380 return m ;
1381 }
1382 }
1384 // Note that we could encounter some performance loss through false-sharing as
1385 // multiple locks occupy the same $ line. Padding might be appropriate.
1388 // Deflate_idle_monitors() is called at all safepoints, immediately
1389 // after all mutators are stopped, but before any objects have moved.
1390 // It traverses the list of known monitors, deflating where possible.
1391 // The scavenged monitor are returned to the monitor free list.
1392 //
1393 // Beware that we scavenge at *every* stop-the-world point.
1394 // Having a large number of monitors in-circulation negatively
1395 // impacts the performance of some applications (e.g., PointBase).
1396 // Broadly, we want to minimize the # of monitors in circulation.
1397 //
1398 // We have added a flag, MonitorInUseLists, which creates a list
1399 // of active monitors for each thread. deflate_idle_monitors()
1400 // only scans the per-thread inuse lists. omAlloc() puts all
1401 // assigned monitors on the per-thread list. deflate_idle_monitors()
1402 // returns the non-busy monitors to the global free list.
1403 // When a thread dies, omFlush() adds the list of active monitors for
1404 // that thread to a global gOmInUseList acquiring the
1405 // global list lock. deflate_idle_monitors() acquires the global
1406 // list lock to scan for non-busy monitors to the global free list.
1407 // An alternative could have used a single global inuse list. The
1408 // downside would have been the additional cost of acquiring the global list lock
1409 // for every omAlloc().
1410 //
1411 // Perversely, the heap size -- and thus the STW safepoint rate --
1412 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1413 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1414 // This is an unfortunate aspect of this design.
1415 //
1417 enum ManifestConstants {
1418 ClearResponsibleAtSTW = 0,
1419 MaximumRecheckInterval = 1000
1420 } ;
1422 // Deflate a single monitor if not in use
1423 // Return true if deflated, false if in use
1424 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1425 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) {
1426 bool deflated;
1427 // Normal case ... The monitor is associated with obj.
1428 guarantee (obj->mark() == markOopDesc::encode(mid), "invariant") ;
1429 guarantee (mid == obj->mark()->monitor(), "invariant");
1430 guarantee (mid->header()->is_neutral(), "invariant");
1432 if (mid->is_busy()) {
1433 if (ClearResponsibleAtSTW) mid->_Responsible = NULL ;
1434 deflated = false;
1435 } else {
1436 // Deflate the monitor if it is no longer being used
1437 // It's idle - scavenge and return to the global free list
1438 // plain old deflation ...
1439 TEVENT (deflate_idle_monitors - scavenge1) ;
1440 if (TraceMonitorInflation) {
1441 if (obj->is_instance()) {
1442 ResourceMark rm;
1443 tty->print_cr("Deflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1444 (void *) obj, (intptr_t) obj->mark(), obj->klass()->external_name());
1445 }
1446 }
1448 // Restore the header back to obj
1449 obj->release_set_mark(mid->header());
1450 mid->clear();
1452 assert (mid->object() == NULL, "invariant") ;
1454 // Move the object to the working free list defined by FreeHead,FreeTail.
1455 if (*FreeHeadp == NULL) *FreeHeadp = mid;
1456 if (*FreeTailp != NULL) {
1457 ObjectMonitor * prevtail = *FreeTailp;
1458 assert(prevtail->FreeNext == NULL, "cleaned up deflated?"); // TODO KK
1459 prevtail->FreeNext = mid;
1460 }
1461 *FreeTailp = mid;
1462 deflated = true;
1463 }
1464 return deflated;
1465 }
1467 // Caller acquires ListLock
1468 int ObjectSynchronizer::walk_monitor_list(ObjectMonitor** listheadp,
1469 ObjectMonitor** FreeHeadp, ObjectMonitor** FreeTailp) {
1470 ObjectMonitor* mid;
1471 ObjectMonitor* next;
1472 ObjectMonitor* curmidinuse = NULL;
1473 int deflatedcount = 0;
1475 for (mid = *listheadp; mid != NULL; ) {
1476 oop obj = (oop) mid->object();
1477 bool deflated = false;
1478 if (obj != NULL) {
1479 deflated = deflate_monitor(mid, obj, FreeHeadp, FreeTailp);
1480 }
1481 if (deflated) {
1482 // extract from per-thread in-use-list
1483 if (mid == *listheadp) {
1484 *listheadp = mid->FreeNext;
1485 } else if (curmidinuse != NULL) {
1486 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist
1487 }
1488 next = mid->FreeNext;
1489 mid->FreeNext = NULL; // This mid is current tail in the FreeHead list
1490 mid = next;
1491 deflatedcount++;
1492 } else {
1493 curmidinuse = mid;
1494 mid = mid->FreeNext;
1495 }
1496 }
1497 return deflatedcount;
1498 }
1500 void ObjectSynchronizer::deflate_idle_monitors() {
1501 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1502 int nInuse = 0 ; // currently associated with objects
1503 int nInCirculation = 0 ; // extant
1504 int nScavenged = 0 ; // reclaimed
1505 bool deflated = false;
1507 ObjectMonitor * FreeHead = NULL ; // Local SLL of scavenged monitors
1508 ObjectMonitor * FreeTail = NULL ;
1510 TEVENT (deflate_idle_monitors) ;
1511 // Prevent omFlush from changing mids in Thread dtor's during deflation
1512 // And in case the vm thread is acquiring a lock during a safepoint
1513 // See e.g. 6320749
1514 Thread::muxAcquire (&ListLock, "scavenge - return") ;
1516 if (MonitorInUseLists) {
1517 int inUse = 0;
1518 for (JavaThread* cur = Threads::first(); cur != NULL; cur = cur->next()) {
1519 nInCirculation+= cur->omInUseCount;
1520 int deflatedcount = walk_monitor_list(cur->omInUseList_addr(), &FreeHead, &FreeTail);
1521 cur->omInUseCount-= deflatedcount;
1522 // verifyInUse(cur);
1523 nScavenged += deflatedcount;
1524 nInuse += cur->omInUseCount;
1525 }
1527 // For moribund threads, scan gOmInUseList
1528 if (gOmInUseList) {
1529 nInCirculation += gOmInUseCount;
1530 int deflatedcount = walk_monitor_list((ObjectMonitor **)&gOmInUseList, &FreeHead, &FreeTail);
1531 gOmInUseCount-= deflatedcount;
1532 nScavenged += deflatedcount;
1533 nInuse += gOmInUseCount;
1534 }
1536 } else for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) {
1537 // Iterate over all extant monitors - Scavenge all idle monitors.
1538 assert(block->object() == CHAINMARKER, "must be a block header");
1539 nInCirculation += _BLOCKSIZE ;
1540 for (int i = 1 ; i < _BLOCKSIZE; i++) {
1541 ObjectMonitor* mid = &block[i];
1542 oop obj = (oop) mid->object();
1544 if (obj == NULL) {
1545 // The monitor is not associated with an object.
1546 // The monitor should either be a thread-specific private
1547 // free list or the global free list.
1548 // obj == NULL IMPLIES mid->is_busy() == 0
1549 guarantee (!mid->is_busy(), "invariant") ;
1550 continue ;
1551 }
1552 deflated = deflate_monitor(mid, obj, &FreeHead, &FreeTail);
1554 if (deflated) {
1555 mid->FreeNext = NULL ;
1556 nScavenged ++ ;
1557 } else {
1558 nInuse ++;
1559 }
1560 }
1561 }
1563 MonitorFreeCount += nScavenged;
1565 // Consider: audit gFreeList to ensure that MonitorFreeCount and list agree.
1567 if (ObjectMonitor::Knob_Verbose) {
1568 ::printf ("Deflate: InCirc=%d InUse=%d Scavenged=%d ForceMonitorScavenge=%d : pop=%d free=%d\n",
1569 nInCirculation, nInuse, nScavenged, ForceMonitorScavenge,
1570 MonitorPopulation, MonitorFreeCount) ;
1571 ::fflush(stdout) ;
1572 }
1574 ForceMonitorScavenge = 0; // Reset
1576 // Move the scavenged monitors back to the global free list.
1577 if (FreeHead != NULL) {
1578 guarantee (FreeTail != NULL && nScavenged > 0, "invariant") ;
1579 assert (FreeTail->FreeNext == NULL, "invariant") ;
1580 // constant-time list splice - prepend scavenged segment to gFreeList
1581 FreeTail->FreeNext = gFreeList ;
1582 gFreeList = FreeHead ;
1583 }
1584 Thread::muxRelease (&ListLock) ;
1586 if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged) ;
1587 if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation);
1589 // TODO: Add objectMonitor leak detection.
1590 // Audit/inventory the objectMonitors -- make sure they're all accounted for.
1591 GVars.stwRandom = os::random() ;
1592 GVars.stwCycle ++ ;
1593 }
1595 // Monitor cleanup on JavaThread::exit
1597 // Iterate through monitor cache and attempt to release thread's monitors
1598 // Gives up on a particular monitor if an exception occurs, but continues
1599 // the overall iteration, swallowing the exception.
1600 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1601 private:
1602 TRAPS;
1604 public:
1605 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {}
1606 void do_monitor(ObjectMonitor* mid) {
1607 if (mid->owner() == THREAD) {
1608 (void)mid->complete_exit(CHECK);
1609 }
1610 }
1611 };
1613 // Release all inflated monitors owned by THREAD. Lightweight monitors are
1614 // ignored. This is meant to be called during JNI thread detach which assumes
1615 // all remaining monitors are heavyweight. All exceptions are swallowed.
1616 // Scanning the extant monitor list can be time consuming.
1617 // A simple optimization is to add a per-thread flag that indicates a thread
1618 // called jni_monitorenter() during its lifetime.
1619 //
1620 // Instead of No_Savepoint_Verifier it might be cheaper to
1621 // use an idiom of the form:
1622 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1623 // <code that must not run at safepoint>
1624 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1625 // Since the tests are extremely cheap we could leave them enabled
1626 // for normal product builds.
1628 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) {
1629 assert(THREAD == JavaThread::current(), "must be current Java thread");
1630 No_Safepoint_Verifier nsv ;
1631 ReleaseJavaMonitorsClosure rjmc(THREAD);
1632 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread");
1633 ObjectSynchronizer::monitors_iterate(&rjmc);
1634 Thread::muxRelease(&ListLock);
1635 THREAD->clear_pending_exception();
1636 }
1638 //------------------------------------------------------------------------------
1639 // Non-product code
1641 #ifndef PRODUCT
1643 // Verify all monitors in the monitor cache, the verification is weak.
1644 void ObjectSynchronizer::verify() {
1645 ObjectMonitor* block = gBlockList;
1646 ObjectMonitor* mid;
1647 while (block) {
1648 assert(block->object() == CHAINMARKER, "must be a block header");
1649 for (int i = 1; i < _BLOCKSIZE; i++) {
1650 mid = block + i;
1651 oop object = (oop) mid->object();
1652 if (object != NULL) {
1653 mid->verify();
1654 }
1655 }
1656 block = (ObjectMonitor*) block->FreeNext;
1657 }
1658 }
1660 // Check if monitor belongs to the monitor cache
1661 // The list is grow-only so it's *relatively* safe to traverse
1662 // the list of extant blocks without taking a lock.
1664 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) {
1665 ObjectMonitor* block = gBlockList;
1667 while (block) {
1668 assert(block->object() == CHAINMARKER, "must be a block header");
1669 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) {
1670 address mon = (address) monitor;
1671 address blk = (address) block;
1672 size_t diff = mon - blk;
1673 assert((diff % sizeof(ObjectMonitor)) == 0, "check");
1674 return 1;
1675 }
1676 block = (ObjectMonitor*) block->FreeNext;
1677 }
1678 return 0;
1679 }
1681 #endif