1 /*
2 * Copyright (c) 1998, 2018, 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 "logging/log.hpp"
28 #include "memory/allocation.inline.hpp"
29 #include "memory/metaspaceShared.hpp"
30 #include "memory/padded.hpp"
31 #include "memory/resourceArea.hpp"
32 #include "oops/markOop.hpp"
33 #include "oops/oop.inline.hpp"
34 #include "runtime/atomic.hpp"
35 #include "runtime/biasedLocking.hpp"
36 #include "runtime/handles.inline.hpp"
37 #include "runtime/interfaceSupport.inline.hpp"
38 #include "runtime/mutexLocker.hpp"
39 #include "runtime/objectMonitor.hpp"
40 #include "runtime/objectMonitor.inline.hpp"
41 #include "runtime/osThread.hpp"
42 #include "runtime/stubRoutines.hpp"
43 #include "runtime/synchronizer.hpp"
44 #include "runtime/thread.inline.hpp"
45 #include "runtime/vframe.hpp"
46 #include "runtime/vmThread.hpp"
47 #include "trace/traceMacros.hpp"
48 #include "trace/tracing.hpp"
49 #include "utilities/align.hpp"
50 #include "utilities/dtrace.hpp"
51 #include "utilities/events.hpp"
52 #include "utilities/preserveException.hpp"
53
54 // The "core" versions of monitor enter and exit reside in this file.
55 // The interpreter and compilers contain specialized transliterated
56 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(),
57 // for instance. If you make changes here, make sure to modify the
58 // interpreter, and both C1 and C2 fast-path inline locking code emission.
59 //
60 // -----------------------------------------------------------------------------
61
62 #ifdef DTRACE_ENABLED
63
64 // Only bother with this argument setup if dtrace is available
65 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
66
67 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
68 char* bytes = NULL; \
69 int len = 0; \
70 jlong jtid = SharedRuntime::get_java_tid(thread); \
71 Symbol* klassname = ((oop)(obj))->klass()->name(); \
72 if (klassname != NULL) { \
73 bytes = (char*)klassname->bytes(); \
74 len = klassname->utf8_length(); \
75 }
76
77 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
78 { \
79 if (DTraceMonitorProbes) { \
80 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
81 HOTSPOT_MONITOR_WAIT(jtid, \
82 (uintptr_t)(monitor), bytes, len, (millis)); \
83 } \
84 }
85
86 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY
87 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL
88 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
89
90 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
91 { \
92 if (DTraceMonitorProbes) { \
93 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
94 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
95 (uintptr_t)(monitor), bytes, len); \
96 } \
97 }
98
99 #else // ndef DTRACE_ENABLED
100
101 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
102 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
103
104 #endif // ndef DTRACE_ENABLED
105
106 // This exists only as a workaround of dtrace bug 6254741
107 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
108 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
109 return 0;
110 }
111
112 #define NINFLATIONLOCKS 256
113 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS];
114
115 // global list of blocks of monitors
116 PaddedEnd<ObjectMonitor> * volatile ObjectSynchronizer::gBlockList = NULL;
117 // global monitor free list
118 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL;
119 // global monitor in-use list, for moribund threads,
120 // monitors they inflated need to be scanned for deflation
121 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL;
122 // count of entries in gOmInUseList
123 int ObjectSynchronizer::gOmInUseCount = 0;
124
125 static volatile intptr_t gListLock = 0; // protects global monitor lists
126 static volatile int gMonitorFreeCount = 0; // # on gFreeList
127 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation
128
129 static void post_monitor_inflate_event(EventJavaMonitorInflate&,
130 const oop,
131 const ObjectSynchronizer::InflateCause);
132
133 #define CHAINMARKER (cast_to_oop<intptr_t>(-1))
134
135
136 // =====================> Quick functions
137
138 // The quick_* forms are special fast-path variants used to improve
139 // performance. In the simplest case, a "quick_*" implementation could
140 // simply return false, in which case the caller will perform the necessary
141 // state transitions and call the slow-path form.
142 // The fast-path is designed to handle frequently arising cases in an efficient
143 // manner and is just a degenerate "optimistic" variant of the slow-path.
144 // returns true -- to indicate the call was satisfied.
145 // returns false -- to indicate the call needs the services of the slow-path.
146 // A no-loitering ordinance is in effect for code in the quick_* family
147 // operators: safepoints or indefinite blocking (blocking that might span a
148 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon
149 // entry.
150 //
151 // Consider: An interesting optimization is to have the JIT recognize the
152 // following common idiom:
153 // synchronized (someobj) { .... ; notify(); }
154 // That is, we find a notify() or notifyAll() call that immediately precedes
155 // the monitorexit operation. In that case the JIT could fuse the operations
156 // into a single notifyAndExit() runtime primitive.
157
158 bool ObjectSynchronizer::quick_notify(oopDesc * obj, Thread * self, bool all) {
159 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
160 assert(self->is_Java_thread(), "invariant");
161 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant");
162 NoSafepointVerifier nsv;
163 if (obj == NULL) return false; // slow-path for invalid obj
164 const markOop mark = obj->mark();
165
166 if (mark->has_locker() && self->is_lock_owned((address)mark->locker())) {
167 // Degenerate notify
168 // stack-locked by caller so by definition the implied waitset is empty.
169 return true;
170 }
171
172 if (mark->has_monitor()) {
173 ObjectMonitor * const mon = mark->monitor();
174 assert(mon->object() == obj, "invariant");
175 if (mon->owner() != self) return false; // slow-path for IMS exception
176
177 if (mon->first_waiter() != NULL) {
178 // We have one or more waiters. Since this is an inflated monitor
179 // that we own, we can transfer one or more threads from the waitset
180 // to the entrylist here and now, avoiding the slow-path.
181 if (all) {
182 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self);
183 } else {
184 DTRACE_MONITOR_PROBE(notify, mon, obj, self);
185 }
186 int tally = 0;
187 do {
188 mon->INotify(self);
189 ++tally;
190 } while (mon->first_waiter() != NULL && all);
191 OM_PERFDATA_OP(Notifications, inc(tally));
192 }
193 return true;
194 }
195
196 // biased locking and any other IMS exception states take the slow-path
197 return false;
198 }
199
200
201 // The LockNode emitted directly at the synchronization site would have
202 // been too big if it were to have included support for the cases of inflated
203 // recursive enter and exit, so they go here instead.
204 // Note that we can't safely call AsyncPrintJavaStack() from within
205 // quick_enter() as our thread state remains _in_Java.
206
207 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self,
208 BasicLock * lock) {
209 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
210 assert(Self->is_Java_thread(), "invariant");
211 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant");
212 NoSafepointVerifier nsv;
213 if (obj == NULL) return false; // Need to throw NPE
214 const markOop mark = obj->mark();
215
216 if (mark->has_monitor()) {
217 ObjectMonitor * const m = mark->monitor();
218 assert(m->object() == obj, "invariant");
219 Thread * const owner = (Thread *) m->_owner;
220
221 // Lock contention and Transactional Lock Elision (TLE) diagnostics
222 // and observability
223 // Case: light contention possibly amenable to TLE
224 // Case: TLE inimical operations such as nested/recursive synchronization
225
226 if (owner == Self) {
227 m->_recursions++;
228 return true;
229 }
230
231 // This Java Monitor is inflated so obj's header will never be
232 // displaced to this thread's BasicLock. Make the displaced header
233 // non-NULL so this BasicLock is not seen as recursive nor as
234 // being locked. We do this unconditionally so that this thread's
235 // BasicLock cannot be mis-interpreted by any stack walkers. For
236 // performance reasons, stack walkers generally first check for
237 // Biased Locking in the object's header, the second check is for
238 // stack-locking in the object's header, the third check is for
239 // recursive stack-locking in the displaced header in the BasicLock,
240 // and last are the inflated Java Monitor (ObjectMonitor) checks.
241 lock->set_displaced_header(markOopDesc::unused_mark());
242
243 if (owner == NULL && Atomic::replace_if_null(Self, &(m->_owner))) {
244 assert(m->_recursions == 0, "invariant");
245 assert(m->_owner == Self, "invariant");
246 return true;
247 }
248 }
249
250 // Note that we could inflate in quick_enter.
251 // This is likely a useful optimization
252 // Critically, in quick_enter() we must not:
253 // -- perform bias revocation, or
254 // -- block indefinitely, or
255 // -- reach a safepoint
256
257 return false; // revert to slow-path
258 }
259
260 // -----------------------------------------------------------------------------
261 // Fast Monitor Enter/Exit
262 // This the fast monitor enter. The interpreter and compiler use
263 // some assembly copies of this code. Make sure update those code
264 // if the following function is changed. The implementation is
265 // extremely sensitive to race condition. Be careful.
266
267 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock,
268 bool attempt_rebias, TRAPS) {
269 if (UseBiasedLocking) {
270 if (!SafepointSynchronize::is_at_safepoint()) {
271 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD);
272 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) {
273 return;
274 }
275 } else {
276 assert(!attempt_rebias, "can not rebias toward VM thread");
277 BiasedLocking::revoke_at_safepoint(obj);
278 }
279 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
280 }
281
282 slow_enter(obj, lock, THREAD);
283 }
284
285 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) {
286 markOop mark = object->mark();
287 // We cannot check for Biased Locking if we are racing an inflation.
288 assert(mark == markOopDesc::INFLATING() ||
289 !mark->has_bias_pattern(), "should not see bias pattern here");
290
291 markOop dhw = lock->displaced_header();
292 if (dhw == NULL) {
293 // If the displaced header is NULL, then this exit matches up with
294 // a recursive enter. No real work to do here except for diagnostics.
295 #ifndef PRODUCT
296 if (mark != markOopDesc::INFLATING()) {
297 // Only do diagnostics if we are not racing an inflation. Simply
298 // exiting a recursive enter of a Java Monitor that is being
299 // inflated is safe; see the has_monitor() comment below.
300 assert(!mark->is_neutral(), "invariant");
301 assert(!mark->has_locker() ||
302 THREAD->is_lock_owned((address)mark->locker()), "invariant");
303 if (mark->has_monitor()) {
304 // The BasicLock's displaced_header is marked as a recursive
305 // enter and we have an inflated Java Monitor (ObjectMonitor).
306 // This is a special case where the Java Monitor was inflated
307 // after this thread entered the stack-lock recursively. When a
308 // Java Monitor is inflated, we cannot safely walk the Java
309 // Monitor owner's stack and update the BasicLocks because a
310 // Java Monitor can be asynchronously inflated by a thread that
311 // does not own the Java Monitor.
312 ObjectMonitor * m = mark->monitor();
313 assert(((oop)(m->object()))->mark() == mark, "invariant");
314 assert(m->is_entered(THREAD), "invariant");
315 }
316 }
317 #endif
318 return;
319 }
320
321 if (mark == (markOop) lock) {
322 // If the object is stack-locked by the current thread, try to
323 // swing the displaced header from the BasicLock back to the mark.
324 assert(dhw->is_neutral(), "invariant");
325 if (object->cas_set_mark(dhw, mark) == mark) {
326 TEVENT(fast_exit: release stack-lock);
327 return;
328 }
329 }
330
331 // We have to take the slow-path of possible inflation and then exit.
332 ObjectSynchronizer::inflate(THREAD,
333 object,
334 inflate_cause_vm_internal)->exit(true, THREAD);
335 }
336
337 // -----------------------------------------------------------------------------
338 // Interpreter/Compiler Slow Case
339 // This routine is used to handle interpreter/compiler slow case
340 // We don't need to use fast path here, because it must have been
341 // failed in the interpreter/compiler code.
342 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
343 markOop mark = obj->mark();
344 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
345
346 if (mark->is_neutral()) {
347 // Anticipate successful CAS -- the ST of the displaced mark must
348 // be visible <= the ST performed by the CAS.
349 lock->set_displaced_header(mark);
350 if (mark == obj()->cas_set_mark((markOop) lock, mark)) {
351 TEVENT(slow_enter: release stacklock);
352 return;
353 }
354 // Fall through to inflate() ...
355 } else if (mark->has_locker() &&
356 THREAD->is_lock_owned((address)mark->locker())) {
357 assert(lock != mark->locker(), "must not re-lock the same lock");
358 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
359 lock->set_displaced_header(NULL);
360 return;
361 }
362
363 // The object header will never be displaced to this lock,
364 // so it does not matter what the value is, except that it
365 // must be non-zero to avoid looking like a re-entrant lock,
366 // and must not look locked either.
367 lock->set_displaced_header(markOopDesc::unused_mark());
368 ObjectSynchronizer::inflate(THREAD,
369 obj(),
370 inflate_cause_monitor_enter)->enter(THREAD);
371 }
372
373 // This routine is used to handle interpreter/compiler slow case
374 // We don't need to use fast path here, because it must have
375 // failed in the interpreter/compiler code. Simply use the heavy
376 // weight monitor should be ok, unless someone find otherwise.
377 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
378 fast_exit(object, lock, THREAD);
379 }
380
381 // -----------------------------------------------------------------------------
382 // Class Loader support to workaround deadlocks on the class loader lock objects
383 // Also used by GC
384 // complete_exit()/reenter() are used to wait on a nested lock
385 // i.e. to give up an outer lock completely and then re-enter
386 // Used when holding nested locks - lock acquisition order: lock1 then lock2
387 // 1) complete_exit lock1 - saving recursion count
388 // 2) wait on lock2
389 // 3) when notified on lock2, unlock lock2
390 // 4) reenter lock1 with original recursion count
391 // 5) lock lock2
392 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
393 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
394 TEVENT(complete_exit);
395 if (UseBiasedLocking) {
396 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
397 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
398 }
399
400 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
401 obj(),
402 inflate_cause_vm_internal);
403
404 return monitor->complete_exit(THREAD);
405 }
406
407 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
408 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
409 TEVENT(reenter);
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 }
414
415 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
416 obj(),
417 inflate_cause_vm_internal);
418
419 monitor->reenter(recursion, THREAD);
420 }
421 // -----------------------------------------------------------------------------
422 // JNI locks on java objects
423 // NOTE: must use heavy weight monitor to handle jni monitor enter
424 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) {
425 // the current locking is from JNI instead of Java code
426 TEVENT(jni_enter);
427 if (UseBiasedLocking) {
428 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
429 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
430 }
431 THREAD->set_current_pending_monitor_is_from_java(false);
432 ObjectSynchronizer::inflate(THREAD, obj(), inflate_cause_jni_enter)->enter(THREAD);
433 THREAD->set_current_pending_monitor_is_from_java(true);
434 }
435
436 // NOTE: must use heavy weight monitor to handle jni monitor exit
437 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
438 TEVENT(jni_exit);
439 if (UseBiasedLocking) {
440 Handle h_obj(THREAD, obj);
441 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD);
442 obj = h_obj();
443 }
444 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
445
446 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
447 obj,
448 inflate_cause_jni_exit);
449 // If this thread has locked the object, exit the monitor. Note: can't use
450 // monitor->check(CHECK); must exit even if an exception is pending.
451 if (monitor->check(THREAD)) {
452 monitor->exit(true, THREAD);
453 }
454 }
455
456 // -----------------------------------------------------------------------------
457 // Internal VM locks on java objects
458 // standard constructor, allows locking failures
459 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
460 _dolock = doLock;
461 _thread = thread;
462 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
463 _obj = obj;
464
465 if (_dolock) {
466 TEVENT(ObjectLocker);
467
468 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
469 }
470 }
471
472 ObjectLocker::~ObjectLocker() {
473 if (_dolock) {
474 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
475 }
476 }
477
478
479 // -----------------------------------------------------------------------------
480 // Wait/Notify/NotifyAll
481 // NOTE: must use heavy weight monitor to handle wait()
482 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
483 if (UseBiasedLocking) {
484 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
485 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
486 }
487 if (millis < 0) {
488 TEVENT(wait - throw IAX);
489 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
490 }
491 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
492 obj(),
493 inflate_cause_wait);
494
495 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
496 monitor->wait(millis, true, THREAD);
497
498 // This dummy call is in place to get around dtrace bug 6254741. Once
499 // that's fixed we can uncomment the following line, remove the call
500 // and change this function back into a "void" func.
501 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
502 return dtrace_waited_probe(monitor, obj, THREAD);
503 }
504
505 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
506 if (UseBiasedLocking) {
507 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
508 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
509 }
510 if (millis < 0) {
511 TEVENT(wait - throw IAX);
512 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
513 }
514 ObjectSynchronizer::inflate(THREAD,
515 obj(),
516 inflate_cause_wait)->wait(millis, false, THREAD);
517 }
518
519 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
520 if (UseBiasedLocking) {
521 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
522 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
523 }
524
525 markOop mark = obj->mark();
526 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
527 return;
528 }
529 ObjectSynchronizer::inflate(THREAD,
530 obj(),
531 inflate_cause_notify)->notify(THREAD);
532 }
533
534 // NOTE: see comment of notify()
535 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
536 if (UseBiasedLocking) {
537 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
538 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
539 }
540
541 markOop mark = obj->mark();
542 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
543 return;
544 }
545 ObjectSynchronizer::inflate(THREAD,
546 obj(),
547 inflate_cause_notify)->notifyAll(THREAD);
548 }
549
550 // -----------------------------------------------------------------------------
551 // Hash Code handling
552 //
553 // Performance concern:
554 // OrderAccess::storestore() calls release() which at one time stored 0
555 // into the global volatile OrderAccess::dummy variable. This store was
556 // unnecessary for correctness. Many threads storing into a common location
557 // causes considerable cache migration or "sloshing" on large SMP systems.
558 // As such, I avoided using OrderAccess::storestore(). In some cases
559 // OrderAccess::fence() -- which incurs local latency on the executing
560 // processor -- is a better choice as it scales on SMP systems.
561 //
562 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for
563 // a discussion of coherency costs. Note that all our current reference
564 // platforms provide strong ST-ST order, so the issue is moot on IA32,
565 // x64, and SPARC.
566 //
567 // As a general policy we use "volatile" to control compiler-based reordering
568 // and explicit fences (barriers) to control for architectural reordering
569 // performed by the CPU(s) or platform.
570
571 struct SharedGlobals {
572 char _pad_prefix[DEFAULT_CACHE_LINE_SIZE];
573 // These are highly shared mostly-read variables.
574 // To avoid false-sharing they need to be the sole occupants of a cache line.
575 volatile int stwRandom;
576 volatile int stwCycle;
577 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int) * 2);
578 // Hot RW variable -- Sequester to avoid false-sharing
579 volatile int hcSequence;
580 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int));
581 };
582
583 static SharedGlobals GVars;
584 static int MonitorScavengeThreshold = 1000000;
585 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending
586
587 static markOop ReadStableMark(oop obj) {
588 markOop mark = obj->mark();
589 if (!mark->is_being_inflated()) {
590 return mark; // normal fast-path return
591 }
592
593 int its = 0;
594 for (;;) {
595 markOop mark = obj->mark();
596 if (!mark->is_being_inflated()) {
597 return mark; // normal fast-path return
598 }
599
600 // The object is being inflated by some other thread.
601 // The caller of ReadStableMark() must wait for inflation to complete.
602 // Avoid live-lock
603 // TODO: consider calling SafepointSynchronize::do_call_back() while
604 // spinning to see if there's a safepoint pending. If so, immediately
605 // yielding or blocking would be appropriate. Avoid spinning while
606 // there is a safepoint pending.
607 // TODO: add inflation contention performance counters.
608 // TODO: restrict the aggregate number of spinners.
609
610 ++its;
611 if (its > 10000 || !os::is_MP()) {
612 if (its & 1) {
613 os::naked_yield();
614 TEVENT(Inflate: INFLATING - yield);
615 } else {
616 // Note that the following code attenuates the livelock problem but is not
617 // a complete remedy. A more complete solution would require that the inflating
618 // thread hold the associated inflation lock. The following code simply restricts
619 // the number of spinners to at most one. We'll have N-2 threads blocked
620 // on the inflationlock, 1 thread holding the inflation lock and using
621 // a yield/park strategy, and 1 thread in the midst of inflation.
622 // A more refined approach would be to change the encoding of INFLATING
623 // to allow encapsulation of a native thread pointer. Threads waiting for
624 // inflation to complete would use CAS to push themselves onto a singly linked
625 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
626 // and calling park(). When inflation was complete the thread that accomplished inflation
627 // would detach the list and set the markword to inflated with a single CAS and
628 // then for each thread on the list, set the flag and unpark() the thread.
629 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease
630 // wakes at most one thread whereas we need to wake the entire list.
631 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1);
632 int YieldThenBlock = 0;
633 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant");
634 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant");
635 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock");
636 while (obj->mark() == markOopDesc::INFLATING()) {
637 // Beware: NakedYield() is advisory and has almost no effect on some platforms
638 // so we periodically call Self->_ParkEvent->park(1).
639 // We use a mixed spin/yield/block mechanism.
640 if ((YieldThenBlock++) >= 16) {
641 Thread::current()->_ParkEvent->park(1);
642 } else {
643 os::naked_yield();
644 }
645 }
646 Thread::muxRelease(gInflationLocks + ix);
647 TEVENT(Inflate: INFLATING - yield/park);
648 }
649 } else {
650 SpinPause(); // SMP-polite spinning
651 }
652 }
653 }
654
655 // hashCode() generation :
656 //
657 // Possibilities:
658 // * MD5Digest of {obj,stwRandom}
659 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function.
660 // * A DES- or AES-style SBox[] mechanism
661 // * One of the Phi-based schemes, such as:
662 // 2654435761 = 2^32 * Phi (golden ratio)
663 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ;
664 // * A variation of Marsaglia's shift-xor RNG scheme.
665 // * (obj ^ stwRandom) is appealing, but can result
666 // in undesirable regularity in the hashCode values of adjacent objects
667 // (objects allocated back-to-back, in particular). This could potentially
668 // result in hashtable collisions and reduced hashtable efficiency.
669 // There are simple ways to "diffuse" the middle address bits over the
670 // generated hashCode values:
671
672 static inline intptr_t get_next_hash(Thread * Self, oop obj) {
673 intptr_t value = 0;
674 if (hashCode == 0) {
675 // This form uses global Park-Miller RNG.
676 // On MP system we'll have lots of RW access to a global, so the
677 // mechanism induces lots of coherency traffic.
678 value = os::random();
679 } else if (hashCode == 1) {
680 // This variation has the property of being stable (idempotent)
681 // between STW operations. This can be useful in some of the 1-0
682 // synchronization schemes.
683 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3;
684 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom;
685 } else if (hashCode == 2) {
686 value = 1; // for sensitivity testing
687 } else if (hashCode == 3) {
688 value = ++GVars.hcSequence;
689 } else if (hashCode == 4) {
690 value = cast_from_oop<intptr_t>(obj);
691 } else {
692 // Marsaglia's xor-shift scheme with thread-specific state
693 // This is probably the best overall implementation -- we'll
694 // likely make this the default in future releases.
695 unsigned t = Self->_hashStateX;
696 t ^= (t << 11);
697 Self->_hashStateX = Self->_hashStateY;
698 Self->_hashStateY = Self->_hashStateZ;
699 Self->_hashStateZ = Self->_hashStateW;
700 unsigned v = Self->_hashStateW;
701 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
702 Self->_hashStateW = v;
703 value = v;
704 }
705
706 value &= markOopDesc::hash_mask;
707 if (value == 0) value = 0xBAD;
708 assert(value != markOopDesc::no_hash, "invariant");
709 TEVENT(hashCode: GENERATE);
710 return value;
711 }
712
713 intptr_t ObjectSynchronizer::FastHashCode(Thread * Self, oop obj) {
714 if (UseBiasedLocking) {
715 // NOTE: many places throughout the JVM do not expect a safepoint
716 // to be taken here, in particular most operations on perm gen
717 // objects. However, we only ever bias Java instances and all of
718 // the call sites of identity_hash that might revoke biases have
719 // been checked to make sure they can handle a safepoint. The
720 // added check of the bias pattern is to avoid useless calls to
721 // thread-local storage.
722 if (obj->mark()->has_bias_pattern()) {
723 // Handle for oop obj in case of STW safepoint
724 Handle hobj(Self, obj);
725 // Relaxing assertion for bug 6320749.
726 assert(Universe::verify_in_progress() ||
727 !SafepointSynchronize::is_at_safepoint(),
728 "biases should not be seen by VM thread here");
729 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
730 obj = hobj();
731 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
732 }
733 }
734
735 // hashCode() is a heap mutator ...
736 // Relaxing assertion for bug 6320749.
737 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
738 !SafepointSynchronize::is_at_safepoint(), "invariant");
739 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
740 Self->is_Java_thread() , "invariant");
741 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
742 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
743
744 ObjectMonitor* monitor = NULL;
745 markOop temp, test;
746 intptr_t hash;
747 markOop mark = ReadStableMark(obj);
748
749 // object should remain ineligible for biased locking
750 assert(!mark->has_bias_pattern(), "invariant");
751
752 if (mark->is_neutral()) {
753 hash = mark->hash(); // this is a normal header
754 if (hash) { // if it has hash, just return it
755 return hash;
756 }
757 hash = get_next_hash(Self, obj); // allocate a new hash code
758 temp = mark->copy_set_hash(hash); // merge the hash code into header
759 // use (machine word version) atomic operation to install the hash
760 test = obj->cas_set_mark(temp, mark);
761 if (test == mark) {
762 return hash;
763 }
764 // If atomic operation failed, we must inflate the header
765 // into heavy weight monitor. We could add more code here
766 // for fast path, but it does not worth the complexity.
767 } else if (mark->has_monitor()) {
768 monitor = mark->monitor();
769 temp = monitor->header();
770 assert(temp->is_neutral(), "invariant");
771 hash = temp->hash();
772 if (hash) {
773 return hash;
774 }
775 // Skip to the following code to reduce code size
776 } else if (Self->is_lock_owned((address)mark->locker())) {
777 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
778 assert(temp->is_neutral(), "invariant");
779 hash = temp->hash(); // by current thread, check if the displaced
780 if (hash) { // header contains hash code
781 return hash;
782 }
783 // WARNING:
784 // The displaced header is strictly immutable.
785 // It can NOT be changed in ANY cases. So we have
786 // to inflate the header into heavyweight monitor
787 // even the current thread owns the lock. The reason
788 // is the BasicLock (stack slot) will be asynchronously
789 // read by other threads during the inflate() function.
790 // Any change to stack may not propagate to other threads
791 // correctly.
792 }
793
794 // Inflate the monitor to set hash code
795 monitor = ObjectSynchronizer::inflate(Self, obj, inflate_cause_hash_code);
796 // Load displaced header and check it has hash code
797 mark = monitor->header();
798 assert(mark->is_neutral(), "invariant");
799 hash = mark->hash();
800 if (hash == 0) {
801 hash = get_next_hash(Self, obj);
802 temp = mark->copy_set_hash(hash); // merge hash code into header
803 assert(temp->is_neutral(), "invariant");
804 test = Atomic::cmpxchg(temp, monitor->header_addr(), mark);
805 if (test != mark) {
806 // The only update to the header in the monitor (outside GC)
807 // is install the hash code. If someone add new usage of
808 // displaced header, please update this code
809 hash = test->hash();
810 assert(test->is_neutral(), "invariant");
811 assert(hash != 0, "Trivial unexpected object/monitor header usage.");
812 }
813 }
814 // We finally get the hash
815 return hash;
816 }
817
818 // Deprecated -- use FastHashCode() instead.
819
820 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
821 return FastHashCode(Thread::current(), obj());
822 }
823
824
825 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
826 Handle h_obj) {
827 if (UseBiasedLocking) {
828 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
829 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
830 }
831
832 assert(thread == JavaThread::current(), "Can only be called on current thread");
833 oop obj = h_obj();
834
835 markOop mark = ReadStableMark(obj);
836
837 // Uncontended case, header points to stack
838 if (mark->has_locker()) {
839 return thread->is_lock_owned((address)mark->locker());
840 }
841 // Contended case, header points to ObjectMonitor (tagged pointer)
842 if (mark->has_monitor()) {
843 ObjectMonitor* monitor = mark->monitor();
844 return monitor->is_entered(thread) != 0;
845 }
846 // Unlocked case, header in place
847 assert(mark->is_neutral(), "sanity check");
848 return false;
849 }
850
851 // Be aware of this method could revoke bias of the lock object.
852 // This method queries the ownership of the lock handle specified by 'h_obj'.
853 // If the current thread owns the lock, it returns owner_self. If no
854 // thread owns the lock, it returns owner_none. Otherwise, it will return
855 // owner_other.
856 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
857 (JavaThread *self, Handle h_obj) {
858 // The caller must beware this method can revoke bias, and
859 // revocation can result in a safepoint.
860 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
861 assert(self->thread_state() != _thread_blocked, "invariant");
862
863 // Possible mark states: neutral, biased, stack-locked, inflated
864
865 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
866 // CASE: biased
867 BiasedLocking::revoke_and_rebias(h_obj, false, self);
868 assert(!h_obj->mark()->has_bias_pattern(),
869 "biases should be revoked by now");
870 }
871
872 assert(self == JavaThread::current(), "Can only be called on current thread");
873 oop obj = h_obj();
874 markOop mark = ReadStableMark(obj);
875
876 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
877 if (mark->has_locker()) {
878 return self->is_lock_owned((address)mark->locker()) ?
879 owner_self : owner_other;
880 }
881
882 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor.
883 // The Object:ObjectMonitor relationship is stable as long as we're
884 // not at a safepoint.
885 if (mark->has_monitor()) {
886 void * owner = mark->monitor()->_owner;
887 if (owner == NULL) return owner_none;
888 return (owner == self ||
889 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
890 }
891
892 // CASE: neutral
893 assert(mark->is_neutral(), "sanity check");
894 return owner_none; // it's unlocked
895 }
896
897 // FIXME: jvmti should call this
898 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
899 if (UseBiasedLocking) {
900 if (SafepointSynchronize::is_at_safepoint()) {
901 BiasedLocking::revoke_at_safepoint(h_obj);
902 } else {
903 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
904 }
905 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
906 }
907
908 oop obj = h_obj();
909 address owner = NULL;
910
911 markOop mark = ReadStableMark(obj);
912
913 // Uncontended case, header points to stack
914 if (mark->has_locker()) {
915 owner = (address) mark->locker();
916 }
917
918 // Contended case, header points to ObjectMonitor (tagged pointer)
919 if (mark->has_monitor()) {
920 ObjectMonitor* monitor = mark->monitor();
921 assert(monitor != NULL, "monitor should be non-null");
922 owner = (address) monitor->owner();
923 }
924
925 if (owner != NULL) {
926 // owning_thread_from_monitor_owner() may also return NULL here
927 return Threads::owning_thread_from_monitor_owner(t_list, owner);
928 }
929
930 // Unlocked case, header in place
931 // Cannot have assertion since this object may have been
932 // locked by another thread when reaching here.
933 // assert(mark->is_neutral(), "sanity check");
934
935 return NULL;
936 }
937
938 // Visitors ...
939
940 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
941 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
942 while (block != NULL) {
943 assert(block->object() == CHAINMARKER, "must be a block header");
944 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
945 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
946 oop object = (oop)mid->object();
947 if (object != NULL) {
948 closure->do_monitor(mid);
949 }
950 }
951 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
952 }
953 }
954
955 // Get the next block in the block list.
956 static inline PaddedEnd<ObjectMonitor>* next(PaddedEnd<ObjectMonitor>* block) {
957 assert(block->object() == CHAINMARKER, "must be a block header");
958 block = (PaddedEnd<ObjectMonitor>*) block->FreeNext;
959 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
960 return block;
961 }
962
963 static bool monitors_used_above_threshold() {
964 if (gMonitorPopulation == 0) {
965 return false;
966 }
967 int monitors_used = gMonitorPopulation - gMonitorFreeCount;
968 int monitor_usage = (monitors_used * 100LL) / gMonitorPopulation;
969 return monitor_usage > MonitorUsedDeflationThreshold;
970 }
971
972 bool ObjectSynchronizer::is_cleanup_needed() {
973 if (MonitorUsedDeflationThreshold > 0) {
974 return monitors_used_above_threshold();
975 }
976 return false;
977 }
978
979 void ObjectSynchronizer::oops_do(OopClosure* f) {
980 if (MonitorInUseLists) {
981 // When using thread local monitor lists, we only scan the
982 // global used list here (for moribund threads), and
983 // the thread-local monitors in Thread::oops_do().
984 global_used_oops_do(f);
985 } else {
986 global_oops_do(f);
987 }
988 }
989
990 void ObjectSynchronizer::global_oops_do(OopClosure* f) {
991 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
992 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
993 for (; block != NULL; block = next(block)) {
994 assert(block->object() == CHAINMARKER, "must be a block header");
995 for (int i = 1; i < _BLOCKSIZE; i++) {
996 ObjectMonitor* mid = (ObjectMonitor *)&block[i];
997 if (mid->object() != NULL) {
998 f->do_oop((oop*)mid->object_addr());
999 }
1000 }
1001 }
1002 }
1003
1004 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) {
1005 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1006 list_oops_do(gOmInUseList, f);
1007 }
1008
1009 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) {
1010 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1011 list_oops_do(thread->omInUseList, f);
1012 }
1013
1014 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) {
1015 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1016 ObjectMonitor* mid;
1017 for (mid = list; mid != NULL; mid = mid->FreeNext) {
1018 if (mid->object() != NULL) {
1019 f->do_oop((oop*)mid->object_addr());
1020 }
1021 }
1022 }
1023
1024
1025 // -----------------------------------------------------------------------------
1026 // ObjectMonitor Lifecycle
1027 // -----------------------
1028 // Inflation unlinks monitors from the global gFreeList and
1029 // associates them with objects. Deflation -- which occurs at
1030 // STW-time -- disassociates idle monitors from objects. Such
1031 // scavenged monitors are returned to the gFreeList.
1032 //
1033 // The global list is protected by gListLock. All the critical sections
1034 // are short and operate in constant-time.
1035 //
1036 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
1037 //
1038 // Lifecycle:
1039 // -- unassigned and on the global free list
1040 // -- unassigned and on a thread's private omFreeList
1041 // -- assigned to an object. The object is inflated and the mark refers
1042 // to the objectmonitor.
1043
1044
1045 // Constraining monitor pool growth via MonitorBound ...
1046 //
1047 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
1048 // the rate of scavenging is driven primarily by GC. As such, we can find
1049 // an inordinate number of monitors in circulation.
1050 // To avoid that scenario we can artificially induce a STW safepoint
1051 // if the pool appears to be growing past some reasonable bound.
1052 // Generally we favor time in space-time tradeoffs, but as there's no
1053 // natural back-pressure on the # of extant monitors we need to impose some
1054 // type of limit. Beware that if MonitorBound is set to too low a value
1055 // we could just loop. In addition, if MonitorBound is set to a low value
1056 // we'll incur more safepoints, which are harmful to performance.
1057 // See also: GuaranteedSafepointInterval
1058 //
1059 // The current implementation uses asynchronous VM operations.
1060
1061 static void InduceScavenge(Thread * Self, const char * Whence) {
1062 // Induce STW safepoint to trim monitors
1063 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
1064 // More precisely, trigger an asynchronous STW safepoint as the number
1065 // of active monitors passes the specified threshold.
1066 // TODO: assert thread state is reasonable
1067
1068 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
1069 if (ObjectMonitor::Knob_Verbose) {
1070 tty->print_cr("INFO: Monitor scavenge - Induced STW @%s (%d)",
1071 Whence, ForceMonitorScavenge) ;
1072 tty->flush();
1073 }
1074 // Induce a 'null' safepoint to scavenge monitors
1075 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
1076 // to the VMthread and have a lifespan longer than that of this activation record.
1077 // The VMThread will delete the op when completed.
1078 VMThread::execute(new VM_ScavengeMonitors());
1079
1080 if (ObjectMonitor::Knob_Verbose) {
1081 tty->print_cr("INFO: Monitor scavenge - STW posted @%s (%d)",
1082 Whence, ForceMonitorScavenge) ;
1083 tty->flush();
1084 }
1085 }
1086 }
1087
1088 void ObjectSynchronizer::verifyInUse(Thread *Self) {
1089 ObjectMonitor* mid;
1090 int in_use_tally = 0;
1091 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) {
1092 in_use_tally++;
1093 }
1094 assert(in_use_tally == Self->omInUseCount, "in-use count off");
1095
1096 int free_tally = 0;
1097 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) {
1098 free_tally++;
1099 }
1100 assert(free_tally == Self->omFreeCount, "free count off");
1101 }
1102
1103 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) {
1104 // A large MAXPRIVATE value reduces both list lock contention
1105 // and list coherency traffic, but also tends to increase the
1106 // number of objectMonitors in circulation as well as the STW
1107 // scavenge costs. As usual, we lean toward time in space-time
1108 // tradeoffs.
1109 const int MAXPRIVATE = 1024;
1110 for (;;) {
1111 ObjectMonitor * m;
1112
1113 // 1: try to allocate from the thread's local omFreeList.
1114 // Threads will attempt to allocate first from their local list, then
1115 // from the global list, and only after those attempts fail will the thread
1116 // attempt to instantiate new monitors. Thread-local free lists take
1117 // heat off the gListLock and improve allocation latency, as well as reducing
1118 // coherency traffic on the shared global list.
1119 m = Self->omFreeList;
1120 if (m != NULL) {
1121 Self->omFreeList = m->FreeNext;
1122 Self->omFreeCount--;
1123 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene
1124 guarantee(m->object() == NULL, "invariant");
1125 if (MonitorInUseLists) {
1126 m->FreeNext = Self->omInUseList;
1127 Self->omInUseList = m;
1128 Self->omInUseCount++;
1129 if (ObjectMonitor::Knob_VerifyInUse) {
1130 verifyInUse(Self);
1131 }
1132 } else {
1133 m->FreeNext = NULL;
1134 }
1135 return m;
1136 }
1137
1138 // 2: try to allocate from the global gFreeList
1139 // CONSIDER: use muxTry() instead of muxAcquire().
1140 // If the muxTry() fails then drop immediately into case 3.
1141 // If we're using thread-local free lists then try
1142 // to reprovision the caller's free list.
1143 if (gFreeList != NULL) {
1144 // Reprovision the thread's omFreeList.
1145 // Use bulk transfers to reduce the allocation rate and heat
1146 // on various locks.
1147 Thread::muxAcquire(&gListLock, "omAlloc");
1148 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) {
1149 gMonitorFreeCount--;
1150 ObjectMonitor * take = gFreeList;
1151 gFreeList = take->FreeNext;
1152 guarantee(take->object() == NULL, "invariant");
1153 guarantee(!take->is_busy(), "invariant");
1154 take->Recycle();
1155 omRelease(Self, take, false);
1156 }
1157 Thread::muxRelease(&gListLock);
1158 Self->omFreeProvision += 1 + (Self->omFreeProvision/2);
1159 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE;
1160 TEVENT(omFirst - reprovision);
1161
1162 const int mx = MonitorBound;
1163 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) {
1164 // We can't safely induce a STW safepoint from omAlloc() as our thread
1165 // state may not be appropriate for such activities and callers may hold
1166 // naked oops, so instead we defer the action.
1167 InduceScavenge(Self, "omAlloc");
1168 }
1169 continue;
1170 }
1171
1172 // 3: allocate a block of new ObjectMonitors
1173 // Both the local and global free lists are empty -- resort to malloc().
1174 // In the current implementation objectMonitors are TSM - immortal.
1175 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want
1176 // each ObjectMonitor to start at the beginning of a cache line,
1177 // so we use align_up().
1178 // A better solution would be to use C++ placement-new.
1179 // BEWARE: As it stands currently, we don't run the ctors!
1180 assert(_BLOCKSIZE > 1, "invariant");
1181 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE;
1182 PaddedEnd<ObjectMonitor> * temp;
1183 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1);
1184 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size,
1185 mtInternal);
1186 temp = (PaddedEnd<ObjectMonitor> *)
1187 align_up(real_malloc_addr, DEFAULT_CACHE_LINE_SIZE);
1188
1189 // NOTE: (almost) no way to recover if allocation failed.
1190 // We might be able to induce a STW safepoint and scavenge enough
1191 // objectMonitors to permit progress.
1192 if (temp == NULL) {
1193 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR,
1194 "Allocate ObjectMonitors");
1195 }
1196 (void)memset((void *) temp, 0, neededsize);
1197
1198 // Format the block.
1199 // initialize the linked list, each monitor points to its next
1200 // forming the single linked free list, the very first monitor
1201 // will points to next block, which forms the block list.
1202 // The trick of using the 1st element in the block as gBlockList
1203 // linkage should be reconsidered. A better implementation would
1204 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1205
1206 for (int i = 1; i < _BLOCKSIZE; i++) {
1207 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1];
1208 }
1209
1210 // terminate the last monitor as the end of list
1211 temp[_BLOCKSIZE - 1].FreeNext = NULL;
1212
1213 // Element [0] is reserved for global list linkage
1214 temp[0].set_object(CHAINMARKER);
1215
1216 // Consider carving out this thread's current request from the
1217 // block in hand. This avoids some lock traffic and redundant
1218 // list activity.
1219
1220 // Acquire the gListLock to manipulate gBlockList and gFreeList.
1221 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1222 Thread::muxAcquire(&gListLock, "omAlloc [2]");
1223 gMonitorPopulation += _BLOCKSIZE-1;
1224 gMonitorFreeCount += _BLOCKSIZE-1;
1225
1226 // Add the new block to the list of extant blocks (gBlockList).
1227 // The very first objectMonitor in a block is reserved and dedicated.
1228 // It serves as blocklist "next" linkage.
1229 temp[0].FreeNext = gBlockList;
1230 // There are lock-free uses of gBlockList so make sure that
1231 // the previous stores happen before we update gBlockList.
1232 OrderAccess::release_store(&gBlockList, temp);
1233
1234 // Add the new string of objectMonitors to the global free list
1235 temp[_BLOCKSIZE - 1].FreeNext = gFreeList;
1236 gFreeList = temp + 1;
1237 Thread::muxRelease(&gListLock);
1238 TEVENT(Allocate block of monitors);
1239 }
1240 }
1241
1242 // Place "m" on the caller's private per-thread omFreeList.
1243 // In practice there's no need to clamp or limit the number of
1244 // monitors on a thread's omFreeList as the only time we'll call
1245 // omRelease is to return a monitor to the free list after a CAS
1246 // attempt failed. This doesn't allow unbounded #s of monitors to
1247 // accumulate on a thread's free list.
1248 //
1249 // Key constraint: all ObjectMonitors on a thread's free list and the global
1250 // free list must have their object field set to null. This prevents the
1251 // scavenger -- deflate_idle_monitors -- from reclaiming them.
1252
1253 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m,
1254 bool fromPerThreadAlloc) {
1255 guarantee(m->object() == NULL, "invariant");
1256 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor");
1257 // Remove from omInUseList
1258 if (MonitorInUseLists && fromPerThreadAlloc) {
1259 ObjectMonitor* cur_mid_in_use = NULL;
1260 bool extracted = false;
1261 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) {
1262 if (m == mid) {
1263 // extract from per-thread in-use list
1264 if (mid == Self->omInUseList) {
1265 Self->omInUseList = mid->FreeNext;
1266 } else if (cur_mid_in_use != NULL) {
1267 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1268 }
1269 extracted = true;
1270 Self->omInUseCount--;
1271 if (ObjectMonitor::Knob_VerifyInUse) {
1272 verifyInUse(Self);
1273 }
1274 break;
1275 }
1276 }
1277 assert(extracted, "Should have extracted from in-use list");
1278 }
1279
1280 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new
1281 m->FreeNext = Self->omFreeList;
1282 Self->omFreeList = m;
1283 Self->omFreeCount++;
1284 }
1285
1286 // Return the monitors of a moribund thread's local free list to
1287 // the global free list. Typically a thread calls omFlush() when
1288 // it's dying. We could also consider having the VM thread steal
1289 // monitors from threads that have not run java code over a few
1290 // consecutive STW safepoints. Relatedly, we might decay
1291 // omFreeProvision at STW safepoints.
1292 //
1293 // Also return the monitors of a moribund thread's omInUseList to
1294 // a global gOmInUseList under the global list lock so these
1295 // will continue to be scanned.
1296 //
1297 // We currently call omFlush() from Threads::remove() _before the thread
1298 // has been excised from the thread list and is no longer a mutator.
1299 // This means that omFlush() can not run concurrently with a safepoint and
1300 // interleave with the scavenge operator. In particular, this ensures that
1301 // the thread's monitors are scanned by a GC safepoint, either via
1302 // Thread::oops_do() (if safepoint happens before omFlush()) or via
1303 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's
1304 // monitors have been transferred to the global in-use list).
1305
1306 void ObjectSynchronizer::omFlush(Thread * Self) {
1307 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL
1308 Self->omFreeList = NULL;
1309 ObjectMonitor * tail = NULL;
1310 int tally = 0;
1311 if (list != NULL) {
1312 ObjectMonitor * s;
1313 // The thread is going away, the per-thread free monitors
1314 // are freed via set_owner(NULL)
1315 // Link them to tail, which will be linked into the global free list
1316 // gFreeList below, under the gListLock
1317 for (s = list; s != NULL; s = s->FreeNext) {
1318 tally++;
1319 tail = s;
1320 guarantee(s->object() == NULL, "invariant");
1321 guarantee(!s->is_busy(), "invariant");
1322 s->set_owner(NULL); // redundant but good hygiene
1323 TEVENT(omFlush - Move one);
1324 }
1325 guarantee(tail != NULL && list != NULL, "invariant");
1326 }
1327
1328 ObjectMonitor * inUseList = Self->omInUseList;
1329 ObjectMonitor * inUseTail = NULL;
1330 int inUseTally = 0;
1331 if (inUseList != NULL) {
1332 Self->omInUseList = NULL;
1333 ObjectMonitor *cur_om;
1334 // The thread is going away, however the omInUseList inflated
1335 // monitors may still be in-use by other threads.
1336 // Link them to inUseTail, which will be linked into the global in-use list
1337 // gOmInUseList below, under the gListLock
1338 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) {
1339 inUseTail = cur_om;
1340 inUseTally++;
1341 }
1342 assert(Self->omInUseCount == inUseTally, "in-use count off");
1343 Self->omInUseCount = 0;
1344 guarantee(inUseTail != NULL && inUseList != NULL, "invariant");
1345 }
1346
1347 Thread::muxAcquire(&gListLock, "omFlush");
1348 if (tail != NULL) {
1349 tail->FreeNext = gFreeList;
1350 gFreeList = list;
1351 gMonitorFreeCount += tally;
1352 assert(Self->omFreeCount == tally, "free-count off");
1353 Self->omFreeCount = 0;
1354 }
1355
1356 if (inUseTail != NULL) {
1357 inUseTail->FreeNext = gOmInUseList;
1358 gOmInUseList = inUseList;
1359 gOmInUseCount += inUseTally;
1360 }
1361
1362 Thread::muxRelease(&gListLock);
1363 TEVENT(omFlush);
1364 }
1365
1366 // Fast path code shared by multiple functions
1367 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) {
1368 markOop mark = obj->mark();
1369 if (mark->has_monitor()) {
1370 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid");
1371 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header");
1372 return mark->monitor();
1373 }
1374 return ObjectSynchronizer::inflate(Thread::current(),
1375 obj,
1376 inflate_cause_vm_internal);
1377 }
1378
1379 ObjectMonitor* ObjectSynchronizer::inflate(Thread * Self,
1380 oop object,
1381 const InflateCause cause) {
1382
1383 // Inflate mutates the heap ...
1384 // Relaxing assertion for bug 6320749.
1385 assert(Universe::verify_in_progress() ||
1386 !SafepointSynchronize::is_at_safepoint(), "invariant");
1387
1388 EventJavaMonitorInflate event;
1389
1390 for (;;) {
1391 const markOop mark = object->mark();
1392 assert(!mark->has_bias_pattern(), "invariant");
1393
1394 // The mark can be in one of the following states:
1395 // * Inflated - just return
1396 // * Stack-locked - coerce it to inflated
1397 // * INFLATING - busy wait for conversion to complete
1398 // * Neutral - aggressively inflate the object.
1399 // * BIASED - Illegal. We should never see this
1400
1401 // CASE: inflated
1402 if (mark->has_monitor()) {
1403 ObjectMonitor * inf = mark->monitor();
1404 assert(inf->header()->is_neutral(), "invariant");
1405 assert(inf->object() == object, "invariant");
1406 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1407 return inf;
1408 }
1409
1410 // CASE: inflation in progress - inflating over a stack-lock.
1411 // Some other thread is converting from stack-locked to inflated.
1412 // Only that thread can complete inflation -- other threads must wait.
1413 // The INFLATING value is transient.
1414 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1415 // We could always eliminate polling by parking the thread on some auxiliary list.
1416 if (mark == markOopDesc::INFLATING()) {
1417 TEVENT(Inflate: spin while INFLATING);
1418 ReadStableMark(object);
1419 continue;
1420 }
1421
1422 // CASE: stack-locked
1423 // Could be stack-locked either by this thread or by some other thread.
1424 //
1425 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1426 // to install INFLATING into the mark word. We originally installed INFLATING,
1427 // allocated the objectmonitor, and then finally STed the address of the
1428 // objectmonitor into the mark. This was correct, but artificially lengthened
1429 // the interval in which INFLATED appeared in the mark, thus increasing
1430 // the odds of inflation contention.
1431 //
1432 // We now use per-thread private objectmonitor free lists.
1433 // These list are reprovisioned from the global free list outside the
1434 // critical INFLATING...ST interval. A thread can transfer
1435 // multiple objectmonitors en-mass from the global free list to its local free list.
1436 // This reduces coherency traffic and lock contention on the global free list.
1437 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1438 // before or after the CAS(INFLATING) operation.
1439 // See the comments in omAlloc().
1440
1441 if (mark->has_locker()) {
1442 ObjectMonitor * m = omAlloc(Self);
1443 // Optimistically prepare the objectmonitor - anticipate successful CAS
1444 // We do this before the CAS in order to minimize the length of time
1445 // in which INFLATING appears in the mark.
1446 m->Recycle();
1447 m->_Responsible = NULL;
1448 m->_recursions = 0;
1449 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1450
1451 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark);
1452 if (cmp != mark) {
1453 omRelease(Self, m, true);
1454 continue; // Interference -- just retry
1455 }
1456
1457 // We've successfully installed INFLATING (0) into the mark-word.
1458 // This is the only case where 0 will appear in a mark-word.
1459 // Only the singular thread that successfully swings the mark-word
1460 // to 0 can perform (or more precisely, complete) inflation.
1461 //
1462 // Why do we CAS a 0 into the mark-word instead of just CASing the
1463 // mark-word from the stack-locked value directly to the new inflated state?
1464 // Consider what happens when a thread unlocks a stack-locked object.
1465 // It attempts to use CAS to swing the displaced header value from the
1466 // on-stack basiclock back into the object header. Recall also that the
1467 // header value (hashcode, etc) can reside in (a) the object header, or
1468 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1469 // header in an objectMonitor. The inflate() routine must copy the header
1470 // value from the basiclock on the owner's stack to the objectMonitor, all
1471 // the while preserving the hashCode stability invariants. If the owner
1472 // decides to release the lock while the value is 0, the unlock will fail
1473 // and control will eventually pass from slow_exit() to inflate. The owner
1474 // will then spin, waiting for the 0 value to disappear. Put another way,
1475 // the 0 causes the owner to stall if the owner happens to try to
1476 // drop the lock (restoring the header from the basiclock to the object)
1477 // while inflation is in-progress. This protocol avoids races that might
1478 // would otherwise permit hashCode values to change or "flicker" for an object.
1479 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1480 // 0 serves as a "BUSY" inflate-in-progress indicator.
1481
1482
1483 // fetch the displaced mark from the owner's stack.
1484 // The owner can't die or unwind past the lock while our INFLATING
1485 // object is in the mark. Furthermore the owner can't complete
1486 // an unlock on the object, either.
1487 markOop dmw = mark->displaced_mark_helper();
1488 assert(dmw->is_neutral(), "invariant");
1489
1490 // Setup monitor fields to proper values -- prepare the monitor
1491 m->set_header(dmw);
1492
1493 // Optimization: if the mark->locker stack address is associated
1494 // with this thread we could simply set m->_owner = Self.
1495 // Note that a thread can inflate an object
1496 // that it has stack-locked -- as might happen in wait() -- directly
1497 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1498 m->set_owner(mark->locker());
1499 m->set_object(object);
1500 // TODO-FIXME: assert BasicLock->dhw != 0.
1501
1502 // Must preserve store ordering. The monitor state must
1503 // be stable at the time of publishing the monitor address.
1504 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant");
1505 object->release_set_mark(markOopDesc::encode(m));
1506
1507 // Hopefully the performance counters are allocated on distinct cache lines
1508 // to avoid false sharing on MP systems ...
1509 OM_PERFDATA_OP(Inflations, inc());
1510 TEVENT(Inflate: overwrite stacklock);
1511 if (log_is_enabled(Debug, monitorinflation)) {
1512 if (object->is_instance()) {
1513 ResourceMark rm;
1514 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1515 p2i(object), p2i(object->mark()),
1516 object->klass()->external_name());
1517 }
1518 }
1519 if (event.should_commit()) {
1520 post_monitor_inflate_event(event, object, cause);
1521 }
1522 return m;
1523 }
1524
1525 // CASE: neutral
1526 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1527 // If we know we're inflating for entry it's better to inflate by swinging a
1528 // pre-locked objectMonitor pointer into the object header. A successful
1529 // CAS inflates the object *and* confers ownership to the inflating thread.
1530 // In the current implementation we use a 2-step mechanism where we CAS()
1531 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1532 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1533 // would be useful.
1534
1535 assert(mark->is_neutral(), "invariant");
1536 ObjectMonitor * m = omAlloc(Self);
1537 // prepare m for installation - set monitor to initial state
1538 m->Recycle();
1539 m->set_header(mark);
1540 m->set_owner(NULL);
1541 m->set_object(object);
1542 m->_recursions = 0;
1543 m->_Responsible = NULL;
1544 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1545
1546 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) {
1547 m->set_object(NULL);
1548 m->set_owner(NULL);
1549 m->Recycle();
1550 omRelease(Self, m, true);
1551 m = NULL;
1552 continue;
1553 // interference - the markword changed - just retry.
1554 // The state-transitions are one-way, so there's no chance of
1555 // live-lock -- "Inflated" is an absorbing state.
1556 }
1557
1558 // Hopefully the performance counters are allocated on distinct
1559 // cache lines to avoid false sharing on MP systems ...
1560 OM_PERFDATA_OP(Inflations, inc());
1561 TEVENT(Inflate: overwrite neutral);
1562 if (log_is_enabled(Debug, monitorinflation)) {
1563 if (object->is_instance()) {
1564 ResourceMark rm;
1565 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1566 p2i(object), p2i(object->mark()),
1567 object->klass()->external_name());
1568 }
1569 }
1570 if (event.should_commit()) {
1571 post_monitor_inflate_event(event, object, cause);
1572 }
1573 return m;
1574 }
1575 }
1576
1577
1578 // Deflate_idle_monitors() is called at all safepoints, immediately
1579 // after all mutators are stopped, but before any objects have moved.
1580 // It traverses the list of known monitors, deflating where possible.
1581 // The scavenged monitor are returned to the monitor free list.
1582 //
1583 // Beware that we scavenge at *every* stop-the-world point.
1584 // Having a large number of monitors in-circulation negatively
1585 // impacts the performance of some applications (e.g., PointBase).
1586 // Broadly, we want to minimize the # of monitors in circulation.
1587 //
1588 // We have added a flag, MonitorInUseLists, which creates a list
1589 // of active monitors for each thread. deflate_idle_monitors()
1590 // only scans the per-thread in-use lists. omAlloc() puts all
1591 // assigned monitors on the per-thread list. deflate_idle_monitors()
1592 // returns the non-busy monitors to the global free list.
1593 // When a thread dies, omFlush() adds the list of active monitors for
1594 // that thread to a global gOmInUseList acquiring the
1595 // global list lock. deflate_idle_monitors() acquires the global
1596 // list lock to scan for non-busy monitors to the global free list.
1597 // An alternative could have used a single global in-use list. The
1598 // downside would have been the additional cost of acquiring the global list lock
1599 // for every omAlloc().
1600 //
1601 // Perversely, the heap size -- and thus the STW safepoint rate --
1602 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1603 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1604 // This is an unfortunate aspect of this design.
1605
1606 enum ManifestConstants {
1607 ClearResponsibleAtSTW = 0
1608 };
1609
1610 // Deflate a single monitor if not in-use
1611 // Return true if deflated, false if in-use
1612 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1613 ObjectMonitor** freeHeadp,
1614 ObjectMonitor** freeTailp) {
1615 bool deflated;
1616 // Normal case ... The monitor is associated with obj.
1617 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant");
1618 guarantee(mid == obj->mark()->monitor(), "invariant");
1619 guarantee(mid->header()->is_neutral(), "invariant");
1620
1621 if (mid->is_busy()) {
1622 if (ClearResponsibleAtSTW) mid->_Responsible = NULL;
1623 deflated = false;
1624 } else {
1625 // Deflate the monitor if it is no longer being used
1626 // It's idle - scavenge and return to the global free list
1627 // plain old deflation ...
1628 TEVENT(deflate_idle_monitors - scavenge1);
1629 if (log_is_enabled(Debug, monitorinflation)) {
1630 if (obj->is_instance()) {
1631 ResourceMark rm;
1632 log_debug(monitorinflation)("Deflating object " INTPTR_FORMAT " , "
1633 "mark " INTPTR_FORMAT " , type %s",
1634 p2i(obj), p2i(obj->mark()),
1635 obj->klass()->external_name());
1636 }
1637 }
1638
1639 // Restore the header back to obj
1640 obj->release_set_mark(mid->header());
1641 mid->clear();
1642
1643 assert(mid->object() == NULL, "invariant");
1644
1645 // Move the object to the working free list defined by freeHeadp, freeTailp
1646 if (*freeHeadp == NULL) *freeHeadp = mid;
1647 if (*freeTailp != NULL) {
1648 ObjectMonitor * prevtail = *freeTailp;
1649 assert(prevtail->FreeNext == NULL, "cleaned up deflated?");
1650 prevtail->FreeNext = mid;
1651 }
1652 *freeTailp = mid;
1653 deflated = true;
1654 }
1655 return deflated;
1656 }
1657
1658 // Walk a given monitor list, and deflate idle monitors
1659 // The given list could be a per-thread list or a global list
1660 // Caller acquires gListLock.
1661 //
1662 // In the case of parallel processing of thread local monitor lists,
1663 // work is done by Threads::parallel_threads_do() which ensures that
1664 // each Java thread is processed by exactly one worker thread, and
1665 // thus avoid conflicts that would arise when worker threads would
1666 // process the same monitor lists concurrently.
1667 //
1668 // See also ParallelSPCleanupTask and
1669 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and
1670 // Threads::parallel_java_threads_do() in thread.cpp.
1671 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp,
1672 ObjectMonitor** freeHeadp,
1673 ObjectMonitor** freeTailp) {
1674 ObjectMonitor* mid;
1675 ObjectMonitor* next;
1676 ObjectMonitor* cur_mid_in_use = NULL;
1677 int deflated_count = 0;
1678
1679 for (mid = *listHeadp; mid != NULL;) {
1680 oop obj = (oop) mid->object();
1681 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) {
1682 // if deflate_monitor succeeded,
1683 // extract from per-thread in-use list
1684 if (mid == *listHeadp) {
1685 *listHeadp = mid->FreeNext;
1686 } else if (cur_mid_in_use != NULL) {
1687 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1688 }
1689 next = mid->FreeNext;
1690 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list
1691 mid = next;
1692 deflated_count++;
1693 } else {
1694 cur_mid_in_use = mid;
1695 mid = mid->FreeNext;
1696 }
1697 }
1698 return deflated_count;
1699 }
1700
1701 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1702 counters->nInuse = 0; // currently associated with objects
1703 counters->nInCirculation = 0; // extant
1704 counters->nScavenged = 0; // reclaimed
1705 }
1706
1707 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) {
1708 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1709 bool deflated = false;
1710
1711 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1712 ObjectMonitor * freeTailp = NULL;
1713
1714 TEVENT(deflate_idle_monitors);
1715 // Prevent omFlush from changing mids in Thread dtor's during deflation
1716 // And in case the vm thread is acquiring a lock during a safepoint
1717 // See e.g. 6320749
1718 Thread::muxAcquire(&gListLock, "scavenge - return");
1719
1720 if (MonitorInUseLists) {
1721 // Note: the thread-local monitors lists get deflated in
1722 // a separate pass. See deflate_thread_local_monitors().
1723
1724 // For moribund threads, scan gOmInUseList
1725 if (gOmInUseList) {
1726 counters->nInCirculation += gOmInUseCount;
1727 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp);
1728 gOmInUseCount -= deflated_count;
1729 counters->nScavenged += deflated_count;
1730 counters->nInuse += gOmInUseCount;
1731 }
1732
1733 } else {
1734 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
1735 for (; block != NULL; block = next(block)) {
1736 // Iterate over all extant monitors - Scavenge all idle monitors.
1737 assert(block->object() == CHAINMARKER, "must be a block header");
1738 counters->nInCirculation += _BLOCKSIZE;
1739 for (int i = 1; i < _BLOCKSIZE; i++) {
1740 ObjectMonitor* mid = (ObjectMonitor*)&block[i];
1741 oop obj = (oop)mid->object();
1742
1743 if (obj == NULL) {
1744 // The monitor is not associated with an object.
1745 // The monitor should either be a thread-specific private
1746 // free list or the global free list.
1747 // obj == NULL IMPLIES mid->is_busy() == 0
1748 guarantee(!mid->is_busy(), "invariant");
1749 continue;
1750 }
1751 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp);
1752
1753 if (deflated) {
1754 mid->FreeNext = NULL;
1755 counters->nScavenged++;
1756 } else {
1757 counters->nInuse++;
1758 }
1759 }
1760 }
1761 }
1762
1763 // Move the scavenged monitors back to the global free list.
1764 if (freeHeadp != NULL) {
1765 guarantee(freeTailp != NULL && counters->nScavenged > 0, "invariant");
1766 assert(freeTailp->FreeNext == NULL, "invariant");
1767 // constant-time list splice - prepend scavenged segment to gFreeList
1768 freeTailp->FreeNext = gFreeList;
1769 gFreeList = freeHeadp;
1770 }
1771 Thread::muxRelease(&gListLock);
1772
1773 }
1774
1775 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1776 gMonitorFreeCount += counters->nScavenged;
1777
1778 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree.
1779
1780 if (ObjectMonitor::Knob_Verbose) {
1781 tty->print_cr("INFO: Deflate: InCirc=%d InUse=%d Scavenged=%d "
1782 "ForceMonitorScavenge=%d : pop=%d free=%d",
1783 counters->nInCirculation, counters->nInuse, counters->nScavenged, ForceMonitorScavenge,
1784 gMonitorPopulation, gMonitorFreeCount);
1785 tty->flush();
1786 }
1787
1788 ForceMonitorScavenge = 0; // Reset
1789
1790 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged));
1791 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation));
1792
1793 // TODO: Add objectMonitor leak detection.
1794 // Audit/inventory the objectMonitors -- make sure they're all accounted for.
1795 GVars.stwRandom = os::random();
1796 GVars.stwCycle++;
1797 }
1798
1799 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) {
1800 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1801 if (!MonitorInUseLists) return;
1802
1803 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1804 ObjectMonitor * freeTailp = NULL;
1805
1806 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp);
1807
1808 Thread::muxAcquire(&gListLock, "scavenge - return");
1809
1810 // Adjust counters
1811 counters->nInCirculation += thread->omInUseCount;
1812 thread->omInUseCount -= deflated_count;
1813 if (ObjectMonitor::Knob_VerifyInUse) {
1814 verifyInUse(thread);
1815 }
1816 counters->nScavenged += deflated_count;
1817 counters->nInuse += thread->omInUseCount;
1818
1819 // Move the scavenged monitors back to the global free list.
1820 if (freeHeadp != NULL) {
1821 guarantee(freeTailp != NULL && deflated_count > 0, "invariant");
1822 assert(freeTailp->FreeNext == NULL, "invariant");
1823
1824 // constant-time list splice - prepend scavenged segment to gFreeList
1825 freeTailp->FreeNext = gFreeList;
1826 gFreeList = freeHeadp;
1827 }
1828 Thread::muxRelease(&gListLock);
1829 }
1830
1831 // Monitor cleanup on JavaThread::exit
1832
1833 // Iterate through monitor cache and attempt to release thread's monitors
1834 // Gives up on a particular monitor if an exception occurs, but continues
1835 // the overall iteration, swallowing the exception.
1836 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1837 private:
1838 TRAPS;
1839
1840 public:
1841 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {}
1842 void do_monitor(ObjectMonitor* mid) {
1843 if (mid->owner() == THREAD) {
1844 if (ObjectMonitor::Knob_VerifyMatch != 0) {
1845 ResourceMark rm;
1846 Handle obj(THREAD, (oop) mid->object());
1847 tty->print("INFO: unexpected locked object:");
1848 javaVFrame::print_locked_object_class_name(tty, obj, "locked");
1849 fatal("exiting JavaThread=" INTPTR_FORMAT
1850 " unexpectedly owns ObjectMonitor=" INTPTR_FORMAT,
1851 p2i(THREAD), p2i(mid));
1852 }
1853 (void)mid->complete_exit(CHECK);
1854 }
1855 }
1856 };
1857
1858 // Release all inflated monitors owned by THREAD. Lightweight monitors are
1859 // ignored. This is meant to be called during JNI thread detach which assumes
1860 // all remaining monitors are heavyweight. All exceptions are swallowed.
1861 // Scanning the extant monitor list can be time consuming.
1862 // A simple optimization is to add a per-thread flag that indicates a thread
1863 // called jni_monitorenter() during its lifetime.
1864 //
1865 // Instead of No_Savepoint_Verifier it might be cheaper to
1866 // use an idiom of the form:
1867 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1868 // <code that must not run at safepoint>
1869 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1870 // Since the tests are extremely cheap we could leave them enabled
1871 // for normal product builds.
1872
1873 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) {
1874 assert(THREAD == JavaThread::current(), "must be current Java thread");
1875 NoSafepointVerifier nsv;
1876 ReleaseJavaMonitorsClosure rjmc(THREAD);
1877 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread");
1878 ObjectSynchronizer::monitors_iterate(&rjmc);
1879 Thread::muxRelease(&gListLock);
1880 THREAD->clear_pending_exception();
1881 }
1882
1883 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1884 switch (cause) {
1885 case inflate_cause_vm_internal: return "VM Internal";
1886 case inflate_cause_monitor_enter: return "Monitor Enter";
1887 case inflate_cause_wait: return "Monitor Wait";
1888 case inflate_cause_notify: return "Monitor Notify";
1889 case inflate_cause_hash_code: return "Monitor Hash Code";
1890 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1891 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1892 default:
1893 ShouldNotReachHere();
1894 }
1895 return "Unknown";
1896 }
1897
1898 static void post_monitor_inflate_event(EventJavaMonitorInflate& event,
1899 const oop obj,
1900 const ObjectSynchronizer::InflateCause cause) {
1901 #if INCLUDE_TRACE
1902 assert(event.should_commit(), "check outside");
1903 event.set_monitorClass(obj->klass());
1904 event.set_address((TYPE_ADDRESS)(uintptr_t)(void*)obj);
1905 event.set_cause((u1)cause);
1906 event.commit();
1907 #endif
1908 }
1909
1910 //------------------------------------------------------------------------------
1911 // Debugging code
1912
1913 void ObjectSynchronizer::sanity_checks(const bool verbose,
1914 const uint cache_line_size,
1915 int *error_cnt_ptr,
1916 int *warning_cnt_ptr) {
1917 u_char *addr_begin = (u_char*)&GVars;
1918 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom;
1919 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence;
1920
1921 if (verbose) {
1922 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT,
1923 sizeof(SharedGlobals));
1924 }
1925
1926 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin);
1927 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom);
1928
1929 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin);
1930 if (verbose) {
1931 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence);
1932 }
1933
1934 if (cache_line_size != 0) {
1935 // We were able to determine the L1 data cache line size so
1936 // do some cache line specific sanity checks
1937
1938 if (offset_stwRandom < cache_line_size) {
1939 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer "
1940 "to the struct beginning than a cache line which permits "
1941 "false sharing.");
1942 (*warning_cnt_ptr)++;
1943 }
1944
1945 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) {
1946 tty->print_cr("WARNING: the SharedGlobals.stwRandom and "
1947 "SharedGlobals.hcSequence fields are closer than a cache "
1948 "line which permits false sharing.");
1949 (*warning_cnt_ptr)++;
1950 }
1951
1952 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) {
1953 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer "
1954 "to the struct end than a cache line which permits false "
1955 "sharing.");
1956 (*warning_cnt_ptr)++;
1957 }
1958 }
1959 }
1960
1961 #ifndef PRODUCT
1962
1963 // Check if monitor belongs to the monitor cache
1964 // The list is grow-only so it's *relatively* safe to traverse
1965 // the list of extant blocks without taking a lock.
1966
1967 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) {
1968 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
1969 while (block != NULL) {
1970 assert(block->object() == CHAINMARKER, "must be a block header");
1971 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) {
1972 address mon = (address)monitor;
1973 address blk = (address)block;
1974 size_t diff = mon - blk;
1975 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "must be aligned");
1976 return 1;
1977 }
1978 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
1979 }
1980 return 0;
1981 }
1982
1983 #endif