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 "jfr/jfrEvents.hpp"
29 #include "memory/allocation.inline.hpp"
30 #include "memory/metaspaceShared.hpp"
31 #include "memory/padded.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "oops/markOop.hpp"
34 #include "oops/oop.inline.hpp"
35 #include "runtime/atomic.hpp"
36 #include "runtime/biasedLocking.hpp"
37 #include "runtime/handles.inline.hpp"
38 #include "runtime/interfaceSupport.inline.hpp"
39 #include "runtime/mutexLocker.hpp"
40 #include "runtime/objectMonitor.hpp"
41 #include "runtime/objectMonitor.inline.hpp"
42 #include "runtime/osThread.hpp"
43 #include "runtime/safepointVerifiers.hpp"
44 #include "runtime/sharedRuntime.hpp"
45 #include "runtime/stubRoutines.hpp"
46 #include "runtime/synchronizer.hpp"
47 #include "runtime/thread.inline.hpp"
48 #include "runtime/vframe.hpp"
49 #include "runtime/vmThread.hpp"
50 #include "utilities/align.hpp"
51 #include "utilities/dtrace.hpp"
52 #include "utilities/events.hpp"
53 #include "utilities/preserveException.hpp"
54
55 // The "core" versions of monitor enter and exit reside in this file.
56 // The interpreter and compilers contain specialized transliterated
57 // variants of the enter-exit fast-path operations. See i486.ad fast_lock(),
58 // for instance. If you make changes here, make sure to modify the
59 // interpreter, and both C1 and C2 fast-path inline locking code emission.
60 //
61 // -----------------------------------------------------------------------------
62
63 #ifdef DTRACE_ENABLED
64
65 // Only bother with this argument setup if dtrace is available
66 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
67
68 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
69 char* bytes = NULL; \
70 int len = 0; \
71 jlong jtid = SharedRuntime::get_java_tid(thread); \
72 Symbol* klassname = ((oop)(obj))->klass()->name(); \
73 if (klassname != NULL) { \
74 bytes = (char*)klassname->bytes(); \
75 len = klassname->utf8_length(); \
76 }
77
78 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
79 { \
80 if (DTraceMonitorProbes) { \
81 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
82 HOTSPOT_MONITOR_WAIT(jtid, \
83 (uintptr_t)(monitor), bytes, len, (millis)); \
84 } \
85 }
86
87 #define HOTSPOT_MONITOR_PROBE_notify HOTSPOT_MONITOR_NOTIFY
88 #define HOTSPOT_MONITOR_PROBE_notifyAll HOTSPOT_MONITOR_NOTIFYALL
89 #define HOTSPOT_MONITOR_PROBE_waited HOTSPOT_MONITOR_WAITED
90
91 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
92 { \
93 if (DTraceMonitorProbes) { \
94 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
95 HOTSPOT_MONITOR_PROBE_##probe(jtid, /* probe = waited */ \
96 (uintptr_t)(monitor), bytes, len); \
97 } \
98 }
99
100 #else // ndef DTRACE_ENABLED
101
102 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
103 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
104
105 #endif // ndef DTRACE_ENABLED
106
107 // This exists only as a workaround of dtrace bug 6254741
dtrace_waited_probe(ObjectMonitor * monitor,Handle obj,Thread * thr)108 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) {
109 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr);
110 return 0;
111 }
112
113 #define NINFLATIONLOCKS 256
114 static volatile intptr_t gInflationLocks[NINFLATIONLOCKS];
115
116 // global list of blocks of monitors
117 PaddedEnd<ObjectMonitor> * volatile ObjectSynchronizer::gBlockList = NULL;
118 // global monitor free list
119 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL;
120 // global monitor in-use list, for moribund threads,
121 // monitors they inflated need to be scanned for deflation
122 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL;
123 // count of entries in gOmInUseList
124 int ObjectSynchronizer::gOmInUseCount = 0;
125
126 static volatile intptr_t gListLock = 0; // protects global monitor lists
127 static volatile int gMonitorFreeCount = 0; // # on gFreeList
128 static volatile int gMonitorPopulation = 0; // # Extant -- in circulation
129
130 #define CHAINMARKER (cast_to_oop<intptr_t>(-1))
131
132
133 // =====================> Quick functions
134
135 // The quick_* forms are special fast-path variants used to improve
136 // performance. In the simplest case, a "quick_*" implementation could
137 // simply return false, in which case the caller will perform the necessary
138 // state transitions and call the slow-path form.
139 // The fast-path is designed to handle frequently arising cases in an efficient
140 // manner and is just a degenerate "optimistic" variant of the slow-path.
141 // returns true -- to indicate the call was satisfied.
142 // returns false -- to indicate the call needs the services of the slow-path.
143 // A no-loitering ordinance is in effect for code in the quick_* family
144 // operators: safepoints or indefinite blocking (blocking that might span a
145 // safepoint) are forbidden. Generally the thread_state() is _in_Java upon
146 // entry.
147 //
148 // Consider: An interesting optimization is to have the JIT recognize the
149 // following common idiom:
150 // synchronized (someobj) { .... ; notify(); }
151 // That is, we find a notify() or notifyAll() call that immediately precedes
152 // the monitorexit operation. In that case the JIT could fuse the operations
153 // into a single notifyAndExit() runtime primitive.
154
quick_notify(oopDesc * obj,Thread * self,bool all)155 bool ObjectSynchronizer::quick_notify(oopDesc * obj, Thread * self, bool all) {
156 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
157 assert(self->is_Java_thread(), "invariant");
158 assert(((JavaThread *) self)->thread_state() == _thread_in_Java, "invariant");
159 NoSafepointVerifier nsv;
160 if (obj == NULL) return false; // slow-path for invalid obj
161 const markOop mark = obj->mark();
162
163 if (mark->has_locker() && self->is_lock_owned((address)mark->locker())) {
164 // Degenerate notify
165 // stack-locked by caller so by definition the implied waitset is empty.
166 return true;
167 }
168
169 if (mark->has_monitor()) {
170 ObjectMonitor * const mon = mark->monitor();
171 assert(mon->object() == obj, "invariant");
172 if (mon->owner() != self) return false; // slow-path for IMS exception
173
174 if (mon->first_waiter() != NULL) {
175 // We have one or more waiters. Since this is an inflated monitor
176 // that we own, we can transfer one or more threads from the waitset
177 // to the entrylist here and now, avoiding the slow-path.
178 if (all) {
179 DTRACE_MONITOR_PROBE(notifyAll, mon, obj, self);
180 } else {
181 DTRACE_MONITOR_PROBE(notify, mon, obj, self);
182 }
183 int tally = 0;
184 do {
185 mon->INotify(self);
186 ++tally;
187 } while (mon->first_waiter() != NULL && all);
188 OM_PERFDATA_OP(Notifications, inc(tally));
189 }
190 return true;
191 }
192
193 // biased locking and any other IMS exception states take the slow-path
194 return false;
195 }
196
197
198 // The LockNode emitted directly at the synchronization site would have
199 // been too big if it were to have included support for the cases of inflated
200 // recursive enter and exit, so they go here instead.
201 // Note that we can't safely call AsyncPrintJavaStack() from within
202 // quick_enter() as our thread state remains _in_Java.
203
quick_enter(oop obj,Thread * Self,BasicLock * lock)204 bool ObjectSynchronizer::quick_enter(oop obj, Thread * Self,
205 BasicLock * lock) {
206 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
207 assert(Self->is_Java_thread(), "invariant");
208 assert(((JavaThread *) Self)->thread_state() == _thread_in_Java, "invariant");
209 NoSafepointVerifier nsv;
210 if (obj == NULL) return false; // Need to throw NPE
211 const markOop mark = obj->mark();
212
213 if (mark->has_monitor()) {
214 ObjectMonitor * const m = mark->monitor();
215 assert(m->object() == obj, "invariant");
216 Thread * const owner = (Thread *) m->_owner;
217
218 // Lock contention and Transactional Lock Elision (TLE) diagnostics
219 // and observability
220 // Case: light contention possibly amenable to TLE
221 // Case: TLE inimical operations such as nested/recursive synchronization
222
223 if (owner == Self) {
224 m->_recursions++;
225 return true;
226 }
227
228 // This Java Monitor is inflated so obj's header will never be
229 // displaced to this thread's BasicLock. Make the displaced header
230 // non-NULL so this BasicLock is not seen as recursive nor as
231 // being locked. We do this unconditionally so that this thread's
232 // BasicLock cannot be mis-interpreted by any stack walkers. For
233 // performance reasons, stack walkers generally first check for
234 // Biased Locking in the object's header, the second check is for
235 // stack-locking in the object's header, the third check is for
236 // recursive stack-locking in the displaced header in the BasicLock,
237 // and last are the inflated Java Monitor (ObjectMonitor) checks.
238 lock->set_displaced_header(markOopDesc::unused_mark());
239
240 if (owner == NULL && Atomic::replace_if_null(Self, &(m->_owner))) {
241 assert(m->_recursions == 0, "invariant");
242 assert(m->_owner == Self, "invariant");
243 return true;
244 }
245 }
246
247 // Note that we could inflate in quick_enter.
248 // This is likely a useful optimization
249 // Critically, in quick_enter() we must not:
250 // -- perform bias revocation, or
251 // -- block indefinitely, or
252 // -- reach a safepoint
253
254 return false; // revert to slow-path
255 }
256
257 // -----------------------------------------------------------------------------
258 // Fast Monitor Enter/Exit
259 // This the fast monitor enter. The interpreter and compiler use
260 // some assembly copies of this code. Make sure update those code
261 // if the following function is changed. The implementation is
262 // extremely sensitive to race condition. Be careful.
263
fast_enter(Handle obj,BasicLock * lock,bool attempt_rebias,TRAPS)264 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock,
265 bool attempt_rebias, TRAPS) {
266 if (UseBiasedLocking) {
267 if (!SafepointSynchronize::is_at_safepoint()) {
268 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD);
269 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) {
270 return;
271 }
272 } else {
273 assert(!attempt_rebias, "can not rebias toward VM thread");
274 BiasedLocking::revoke_at_safepoint(obj);
275 }
276 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
277 }
278
279 slow_enter(obj, lock, THREAD);
280 }
281
fast_exit(oop object,BasicLock * lock,TRAPS)282 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) {
283 markOop mark = object->mark();
284 // We cannot check for Biased Locking if we are racing an inflation.
285 assert(mark == markOopDesc::INFLATING() ||
286 !mark->has_bias_pattern(), "should not see bias pattern here");
287
288 markOop dhw = lock->displaced_header();
289 if (dhw == NULL) {
290 // If the displaced header is NULL, then this exit matches up with
291 // a recursive enter. No real work to do here except for diagnostics.
292 #ifndef PRODUCT
293 if (mark != markOopDesc::INFLATING()) {
294 // Only do diagnostics if we are not racing an inflation. Simply
295 // exiting a recursive enter of a Java Monitor that is being
296 // inflated is safe; see the has_monitor() comment below.
297 assert(!mark->is_neutral(), "invariant");
298 assert(!mark->has_locker() ||
299 THREAD->is_lock_owned((address)mark->locker()), "invariant");
300 if (mark->has_monitor()) {
301 // The BasicLock's displaced_header is marked as a recursive
302 // enter and we have an inflated Java Monitor (ObjectMonitor).
303 // This is a special case where the Java Monitor was inflated
304 // after this thread entered the stack-lock recursively. When a
305 // Java Monitor is inflated, we cannot safely walk the Java
306 // Monitor owner's stack and update the BasicLocks because a
307 // Java Monitor can be asynchronously inflated by a thread that
308 // does not own the Java Monitor.
309 ObjectMonitor * m = mark->monitor();
310 assert(((oop)(m->object()))->mark() == mark, "invariant");
311 assert(m->is_entered(THREAD), "invariant");
312 }
313 }
314 #endif
315 return;
316 }
317
318 if (mark == (markOop) lock) {
319 // If the object is stack-locked by the current thread, try to
320 // swing the displaced header from the BasicLock back to the mark.
321 assert(dhw->is_neutral(), "invariant");
322 if (object->cas_set_mark(dhw, mark) == mark) {
323 TEVENT(fast_exit: release stack-lock);
324 return;
325 }
326 }
327
328 // We have to take the slow-path of possible inflation and then exit.
329 ObjectSynchronizer::inflate(THREAD,
330 object,
331 inflate_cause_vm_internal)->exit(true, THREAD);
332 }
333
334 // -----------------------------------------------------------------------------
335 // Interpreter/Compiler Slow Case
336 // This routine is used to handle interpreter/compiler slow case
337 // We don't need to use fast path here, because it must have been
338 // failed in the interpreter/compiler code.
slow_enter(Handle obj,BasicLock * lock,TRAPS)339 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) {
340 markOop mark = obj->mark();
341 assert(!mark->has_bias_pattern(), "should not see bias pattern here");
342
343 if (mark->is_neutral()) {
344 // Anticipate successful CAS -- the ST of the displaced mark must
345 // be visible <= the ST performed by the CAS.
346 lock->set_displaced_header(mark);
347 if (mark == obj()->cas_set_mark((markOop) lock, mark)) {
348 TEVENT(slow_enter: release stacklock);
349 return;
350 }
351 // Fall through to inflate() ...
352 } else if (mark->has_locker() &&
353 THREAD->is_lock_owned((address)mark->locker())) {
354 assert(lock != mark->locker(), "must not re-lock the same lock");
355 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock");
356 lock->set_displaced_header(NULL);
357 return;
358 }
359
360 // The object header will never be displaced to this lock,
361 // so it does not matter what the value is, except that it
362 // must be non-zero to avoid looking like a re-entrant lock,
363 // and must not look locked either.
364 lock->set_displaced_header(markOopDesc::unused_mark());
365 ObjectSynchronizer::inflate(THREAD,
366 obj(),
367 inflate_cause_monitor_enter)->enter(THREAD);
368 }
369
370 // This routine is used to handle interpreter/compiler slow case
371 // We don't need to use fast path here, because it must have
372 // failed in the interpreter/compiler code. Simply use the heavy
373 // weight monitor should be ok, unless someone find otherwise.
slow_exit(oop object,BasicLock * lock,TRAPS)374 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) {
375 fast_exit(object, lock, THREAD);
376 }
377
378 // -----------------------------------------------------------------------------
379 // Class Loader support to workaround deadlocks on the class loader lock objects
380 // Also used by GC
381 // complete_exit()/reenter() are used to wait on a nested lock
382 // i.e. to give up an outer lock completely and then re-enter
383 // Used when holding nested locks - lock acquisition order: lock1 then lock2
384 // 1) complete_exit lock1 - saving recursion count
385 // 2) wait on lock2
386 // 3) when notified on lock2, unlock lock2
387 // 4) reenter lock1 with original recursion count
388 // 5) lock lock2
389 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
complete_exit(Handle obj,TRAPS)390 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) {
391 TEVENT(complete_exit);
392 if (UseBiasedLocking) {
393 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
394 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
395 }
396
397 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
398 obj(),
399 inflate_cause_vm_internal);
400
401 return monitor->complete_exit(THREAD);
402 }
403
404 // NOTE: must use heavy weight monitor to handle complete_exit/reenter()
reenter(Handle obj,intptr_t recursion,TRAPS)405 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) {
406 TEVENT(reenter);
407 if (UseBiasedLocking) {
408 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
409 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
410 }
411
412 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
413 obj(),
414 inflate_cause_vm_internal);
415
416 monitor->reenter(recursion, THREAD);
417 }
418 // -----------------------------------------------------------------------------
419 // JNI locks on java objects
420 // NOTE: must use heavy weight monitor to handle jni monitor enter
jni_enter(Handle obj,TRAPS)421 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) {
422 // the current locking is from JNI instead of Java code
423 TEVENT(jni_enter);
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 }
428 THREAD->set_current_pending_monitor_is_from_java(false);
429 ObjectSynchronizer::inflate(THREAD, obj(), inflate_cause_jni_enter)->enter(THREAD);
430 THREAD->set_current_pending_monitor_is_from_java(true);
431 }
432
433 // NOTE: must use heavy weight monitor to handle jni monitor exit
jni_exit(oop obj,Thread * THREAD)434 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) {
435 TEVENT(jni_exit);
436 if (UseBiasedLocking) {
437 Handle h_obj(THREAD, obj);
438 BiasedLocking::revoke_and_rebias(h_obj, false, THREAD);
439 obj = h_obj();
440 }
441 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
442
443 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
444 obj,
445 inflate_cause_jni_exit);
446 // If this thread has locked the object, exit the monitor. Note: can't use
447 // monitor->check(CHECK); must exit even if an exception is pending.
448 if (monitor->check(THREAD)) {
449 monitor->exit(true, THREAD);
450 }
451 }
452
453 // -----------------------------------------------------------------------------
454 // Internal VM locks on java objects
455 // standard constructor, allows locking failures
ObjectLocker(Handle obj,Thread * thread,bool doLock)456 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) {
457 _dolock = doLock;
458 _thread = thread;
459 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);)
460 _obj = obj;
461
462 if (_dolock) {
463 TEVENT(ObjectLocker);
464
465 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread);
466 }
467 }
468
~ObjectLocker()469 ObjectLocker::~ObjectLocker() {
470 if (_dolock) {
471 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread);
472 }
473 }
474
475
476 // -----------------------------------------------------------------------------
477 // Wait/Notify/NotifyAll
478 // NOTE: must use heavy weight monitor to handle wait()
wait(Handle obj,jlong millis,TRAPS)479 int ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) {
480 if (UseBiasedLocking) {
481 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
482 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
483 }
484 if (millis < 0) {
485 TEVENT(wait - throw IAX);
486 THROW_MSG_0(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
487 }
488 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD,
489 obj(),
490 inflate_cause_wait);
491
492 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis);
493 monitor->wait(millis, true, THREAD);
494
495 // This dummy call is in place to get around dtrace bug 6254741. Once
496 // that's fixed we can uncomment the following line, remove the call
497 // and change this function back into a "void" func.
498 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD);
499 return dtrace_waited_probe(monitor, obj, THREAD);
500 }
501
waitUninterruptibly(Handle obj,jlong millis,TRAPS)502 void ObjectSynchronizer::waitUninterruptibly(Handle obj, jlong millis, TRAPS) {
503 if (UseBiasedLocking) {
504 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
505 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
506 }
507 if (millis < 0) {
508 TEVENT(wait - throw IAX);
509 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative");
510 }
511 ObjectSynchronizer::inflate(THREAD,
512 obj(),
513 inflate_cause_wait)->wait(millis, false, THREAD);
514 }
515
notify(Handle obj,TRAPS)516 void ObjectSynchronizer::notify(Handle obj, TRAPS) {
517 if (UseBiasedLocking) {
518 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
519 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
520 }
521
522 markOop mark = obj->mark();
523 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
524 return;
525 }
526 ObjectSynchronizer::inflate(THREAD,
527 obj(),
528 inflate_cause_notify)->notify(THREAD);
529 }
530
531 // NOTE: see comment of notify()
notifyall(Handle obj,TRAPS)532 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) {
533 if (UseBiasedLocking) {
534 BiasedLocking::revoke_and_rebias(obj, false, THREAD);
535 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
536 }
537
538 markOop mark = obj->mark();
539 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) {
540 return;
541 }
542 ObjectSynchronizer::inflate(THREAD,
543 obj(),
544 inflate_cause_notify)->notifyAll(THREAD);
545 }
546
547 // -----------------------------------------------------------------------------
548 // Hash Code handling
549 //
550 // Performance concern:
551 // OrderAccess::storestore() calls release() which at one time stored 0
552 // into the global volatile OrderAccess::dummy variable. This store was
553 // unnecessary for correctness. Many threads storing into a common location
554 // causes considerable cache migration or "sloshing" on large SMP systems.
555 // As such, I avoided using OrderAccess::storestore(). In some cases
556 // OrderAccess::fence() -- which incurs local latency on the executing
557 // processor -- is a better choice as it scales on SMP systems.
558 //
559 // See http://blogs.oracle.com/dave/entry/biased_locking_in_hotspot for
560 // a discussion of coherency costs. Note that all our current reference
561 // platforms provide strong ST-ST order, so the issue is moot on IA32,
562 // x64, and SPARC.
563 //
564 // As a general policy we use "volatile" to control compiler-based reordering
565 // and explicit fences (barriers) to control for architectural reordering
566 // performed by the CPU(s) or platform.
567
568 struct SharedGlobals {
569 char _pad_prefix[DEFAULT_CACHE_LINE_SIZE];
570 // These are highly shared mostly-read variables.
571 // To avoid false-sharing they need to be the sole occupants of a cache line.
572 volatile int stwRandom;
573 volatile int stwCycle;
574 DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int) * 2);
575 // Hot RW variable -- Sequester to avoid false-sharing
576 volatile int hcSequence;
577 DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(volatile int));
578 };
579
580 static SharedGlobals GVars;
581 static int MonitorScavengeThreshold = 1000000;
582 static volatile int ForceMonitorScavenge = 0; // Scavenge required and pending
583
ReadStableMark(oop obj)584 static markOop ReadStableMark(oop obj) {
585 markOop mark = obj->mark();
586 if (!mark->is_being_inflated()) {
587 return mark; // normal fast-path return
588 }
589
590 int its = 0;
591 for (;;) {
592 markOop mark = obj->mark();
593 if (!mark->is_being_inflated()) {
594 return mark; // normal fast-path return
595 }
596
597 // The object is being inflated by some other thread.
598 // The caller of ReadStableMark() must wait for inflation to complete.
599 // Avoid live-lock
600 // TODO: consider calling SafepointSynchronize::do_call_back() while
601 // spinning to see if there's a safepoint pending. If so, immediately
602 // yielding or blocking would be appropriate. Avoid spinning while
603 // there is a safepoint pending.
604 // TODO: add inflation contention performance counters.
605 // TODO: restrict the aggregate number of spinners.
606
607 ++its;
608 if (its > 10000 || !os::is_MP()) {
609 if (its & 1) {
610 os::naked_yield();
611 TEVENT(Inflate: INFLATING - yield);
612 } else {
613 // Note that the following code attenuates the livelock problem but is not
614 // a complete remedy. A more complete solution would require that the inflating
615 // thread hold the associated inflation lock. The following code simply restricts
616 // the number of spinners to at most one. We'll have N-2 threads blocked
617 // on the inflationlock, 1 thread holding the inflation lock and using
618 // a yield/park strategy, and 1 thread in the midst of inflation.
619 // A more refined approach would be to change the encoding of INFLATING
620 // to allow encapsulation of a native thread pointer. Threads waiting for
621 // inflation to complete would use CAS to push themselves onto a singly linked
622 // list rooted at the markword. Once enqueued, they'd loop, checking a per-thread flag
623 // and calling park(). When inflation was complete the thread that accomplished inflation
624 // would detach the list and set the markword to inflated with a single CAS and
625 // then for each thread on the list, set the flag and unpark() the thread.
626 // This is conceptually similar to muxAcquire-muxRelease, except that muxRelease
627 // wakes at most one thread whereas we need to wake the entire list.
628 int ix = (cast_from_oop<intptr_t>(obj) >> 5) & (NINFLATIONLOCKS-1);
629 int YieldThenBlock = 0;
630 assert(ix >= 0 && ix < NINFLATIONLOCKS, "invariant");
631 assert((NINFLATIONLOCKS & (NINFLATIONLOCKS-1)) == 0, "invariant");
632 Thread::muxAcquire(gInflationLocks + ix, "gInflationLock");
633 while (obj->mark() == markOopDesc::INFLATING()) {
634 // Beware: NakedYield() is advisory and has almost no effect on some platforms
635 // so we periodically call Self->_ParkEvent->park(1).
636 // We use a mixed spin/yield/block mechanism.
637 if ((YieldThenBlock++) >= 16) {
638 Thread::current()->_ParkEvent->park(1);
639 } else {
640 os::naked_yield();
641 }
642 }
643 Thread::muxRelease(gInflationLocks + ix);
644 TEVENT(Inflate: INFLATING - yield/park);
645 }
646 } else {
647 SpinPause(); // SMP-polite spinning
648 }
649 }
650 }
651
652 // hashCode() generation :
653 //
654 // Possibilities:
655 // * MD5Digest of {obj,stwRandom}
656 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function.
657 // * A DES- or AES-style SBox[] mechanism
658 // * One of the Phi-based schemes, such as:
659 // 2654435761 = 2^32 * Phi (golden ratio)
660 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ;
661 // * A variation of Marsaglia's shift-xor RNG scheme.
662 // * (obj ^ stwRandom) is appealing, but can result
663 // in undesirable regularity in the hashCode values of adjacent objects
664 // (objects allocated back-to-back, in particular). This could potentially
665 // result in hashtable collisions and reduced hashtable efficiency.
666 // There are simple ways to "diffuse" the middle address bits over the
667 // generated hashCode values:
668
get_next_hash(Thread * Self,oop obj)669 static inline intptr_t get_next_hash(Thread * Self, oop obj) {
670 intptr_t value = 0;
671 if (hashCode == 0) {
672 // This form uses global Park-Miller RNG.
673 // On MP system we'll have lots of RW access to a global, so the
674 // mechanism induces lots of coherency traffic.
675 value = os::random();
676 } else if (hashCode == 1) {
677 // This variation has the property of being stable (idempotent)
678 // between STW operations. This can be useful in some of the 1-0
679 // synchronization schemes.
680 intptr_t addrBits = cast_from_oop<intptr_t>(obj) >> 3;
681 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom;
682 } else if (hashCode == 2) {
683 value = 1; // for sensitivity testing
684 } else if (hashCode == 3) {
685 value = ++GVars.hcSequence;
686 } else if (hashCode == 4) {
687 value = cast_from_oop<intptr_t>(obj);
688 } else {
689 // Marsaglia's xor-shift scheme with thread-specific state
690 // This is probably the best overall implementation -- we'll
691 // likely make this the default in future releases.
692 unsigned t = Self->_hashStateX;
693 t ^= (t << 11);
694 Self->_hashStateX = Self->_hashStateY;
695 Self->_hashStateY = Self->_hashStateZ;
696 Self->_hashStateZ = Self->_hashStateW;
697 unsigned v = Self->_hashStateW;
698 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8));
699 Self->_hashStateW = v;
700 value = v;
701 }
702
703 value &= markOopDesc::hash_mask;
704 if (value == 0) value = 0xBAD;
705 assert(value != markOopDesc::no_hash, "invariant");
706 TEVENT(hashCode: GENERATE);
707 return value;
708 }
709
FastHashCode(Thread * Self,oop obj)710 intptr_t ObjectSynchronizer::FastHashCode(Thread * Self, oop obj) {
711 if (UseBiasedLocking) {
712 // NOTE: many places throughout the JVM do not expect a safepoint
713 // to be taken here, in particular most operations on perm gen
714 // objects. However, we only ever bias Java instances and all of
715 // the call sites of identity_hash that might revoke biases have
716 // been checked to make sure they can handle a safepoint. The
717 // added check of the bias pattern is to avoid useless calls to
718 // thread-local storage.
719 if (obj->mark()->has_bias_pattern()) {
720 // Handle for oop obj in case of STW safepoint
721 Handle hobj(Self, obj);
722 // Relaxing assertion for bug 6320749.
723 assert(Universe::verify_in_progress() ||
724 !SafepointSynchronize::is_at_safepoint(),
725 "biases should not be seen by VM thread here");
726 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current());
727 obj = hobj();
728 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now");
729 }
730 }
731
732 // hashCode() is a heap mutator ...
733 // Relaxing assertion for bug 6320749.
734 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
735 !SafepointSynchronize::is_at_safepoint(), "invariant");
736 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
737 Self->is_Java_thread() , "invariant");
738 assert(Universe::verify_in_progress() || DumpSharedSpaces ||
739 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
740
741 ObjectMonitor* monitor = NULL;
742 markOop temp, test;
743 intptr_t hash;
744 markOop mark = ReadStableMark(obj);
745
746 // object should remain ineligible for biased locking
747 assert(!mark->has_bias_pattern(), "invariant");
748
749 if (mark->is_neutral()) {
750 hash = mark->hash(); // this is a normal header
751 if (hash) { // if it has hash, just return it
752 return hash;
753 }
754 hash = get_next_hash(Self, obj); // allocate a new hash code
755 temp = mark->copy_set_hash(hash); // merge the hash code into header
756 // use (machine word version) atomic operation to install the hash
757 test = obj->cas_set_mark(temp, mark);
758 if (test == mark) {
759 return hash;
760 }
761 // If atomic operation failed, we must inflate the header
762 // into heavy weight monitor. We could add more code here
763 // for fast path, but it does not worth the complexity.
764 } else if (mark->has_monitor()) {
765 monitor = mark->monitor();
766 temp = monitor->header();
767 assert(temp->is_neutral(), "invariant");
768 hash = temp->hash();
769 if (hash) {
770 return hash;
771 }
772 // Skip to the following code to reduce code size
773 } else if (Self->is_lock_owned((address)mark->locker())) {
774 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned
775 assert(temp->is_neutral(), "invariant");
776 hash = temp->hash(); // by current thread, check if the displaced
777 if (hash) { // header contains hash code
778 return hash;
779 }
780 // WARNING:
781 // The displaced header is strictly immutable.
782 // It can NOT be changed in ANY cases. So we have
783 // to inflate the header into heavyweight monitor
784 // even the current thread owns the lock. The reason
785 // is the BasicLock (stack slot) will be asynchronously
786 // read by other threads during the inflate() function.
787 // Any change to stack may not propagate to other threads
788 // correctly.
789 }
790
791 // Inflate the monitor to set hash code
792 monitor = ObjectSynchronizer::inflate(Self, obj, inflate_cause_hash_code);
793 // Load displaced header and check it has hash code
794 mark = monitor->header();
795 assert(mark->is_neutral(), "invariant");
796 hash = mark->hash();
797 if (hash == 0) {
798 hash = get_next_hash(Self, obj);
799 temp = mark->copy_set_hash(hash); // merge hash code into header
800 assert(temp->is_neutral(), "invariant");
801 test = Atomic::cmpxchg(temp, monitor->header_addr(), mark);
802 if (test != mark) {
803 // The only update to the header in the monitor (outside GC)
804 // is install the hash code. If someone add new usage of
805 // displaced header, please update this code
806 hash = test->hash();
807 assert(test->is_neutral(), "invariant");
808 assert(hash != 0, "Trivial unexpected object/monitor header usage.");
809 }
810 }
811 // We finally get the hash
812 return hash;
813 }
814
815 // Deprecated -- use FastHashCode() instead.
816
identity_hash_value_for(Handle obj)817 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) {
818 return FastHashCode(Thread::current(), obj());
819 }
820
821
current_thread_holds_lock(JavaThread * thread,Handle h_obj)822 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread,
823 Handle h_obj) {
824 if (UseBiasedLocking) {
825 BiasedLocking::revoke_and_rebias(h_obj, false, thread);
826 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
827 }
828
829 assert(thread == JavaThread::current(), "Can only be called on current thread");
830 oop obj = h_obj();
831
832 markOop mark = ReadStableMark(obj);
833
834 // Uncontended case, header points to stack
835 if (mark->has_locker()) {
836 return thread->is_lock_owned((address)mark->locker());
837 }
838 // Contended case, header points to ObjectMonitor (tagged pointer)
839 if (mark->has_monitor()) {
840 ObjectMonitor* monitor = mark->monitor();
841 return monitor->is_entered(thread) != 0;
842 }
843 // Unlocked case, header in place
844 assert(mark->is_neutral(), "sanity check");
845 return false;
846 }
847
848 // Be aware of this method could revoke bias of the lock object.
849 // This method queries the ownership of the lock handle specified by 'h_obj'.
850 // If the current thread owns the lock, it returns owner_self. If no
851 // thread owns the lock, it returns owner_none. Otherwise, it will return
852 // owner_other.
query_lock_ownership(JavaThread * self,Handle h_obj)853 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership
854 (JavaThread *self, Handle h_obj) {
855 // The caller must beware this method can revoke bias, and
856 // revocation can result in a safepoint.
857 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
858 assert(self->thread_state() != _thread_blocked, "invariant");
859
860 // Possible mark states: neutral, biased, stack-locked, inflated
861
862 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) {
863 // CASE: biased
864 BiasedLocking::revoke_and_rebias(h_obj, false, self);
865 assert(!h_obj->mark()->has_bias_pattern(),
866 "biases should be revoked by now");
867 }
868
869 assert(self == JavaThread::current(), "Can only be called on current thread");
870 oop obj = h_obj();
871 markOop mark = ReadStableMark(obj);
872
873 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack.
874 if (mark->has_locker()) {
875 return self->is_lock_owned((address)mark->locker()) ?
876 owner_self : owner_other;
877 }
878
879 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor.
880 // The Object:ObjectMonitor relationship is stable as long as we're
881 // not at a safepoint.
882 if (mark->has_monitor()) {
883 void * owner = mark->monitor()->_owner;
884 if (owner == NULL) return owner_none;
885 return (owner == self ||
886 self->is_lock_owned((address)owner)) ? owner_self : owner_other;
887 }
888
889 // CASE: neutral
890 assert(mark->is_neutral(), "sanity check");
891 return owner_none; // it's unlocked
892 }
893
894 // FIXME: jvmti should call this
get_lock_owner(ThreadsList * t_list,Handle h_obj)895 JavaThread* ObjectSynchronizer::get_lock_owner(ThreadsList * t_list, Handle h_obj) {
896 if (UseBiasedLocking) {
897 if (SafepointSynchronize::is_at_safepoint()) {
898 BiasedLocking::revoke_at_safepoint(h_obj);
899 } else {
900 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current());
901 }
902 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now");
903 }
904
905 oop obj = h_obj();
906 address owner = NULL;
907
908 markOop mark = ReadStableMark(obj);
909
910 // Uncontended case, header points to stack
911 if (mark->has_locker()) {
912 owner = (address) mark->locker();
913 }
914
915 // Contended case, header points to ObjectMonitor (tagged pointer)
916 if (mark->has_monitor()) {
917 ObjectMonitor* monitor = mark->monitor();
918 assert(monitor != NULL, "monitor should be non-null");
919 owner = (address) monitor->owner();
920 }
921
922 if (owner != NULL) {
923 // owning_thread_from_monitor_owner() may also return NULL here
924 return Threads::owning_thread_from_monitor_owner(t_list, owner);
925 }
926
927 // Unlocked case, header in place
928 // Cannot have assertion since this object may have been
929 // locked by another thread when reaching here.
930 // assert(mark->is_neutral(), "sanity check");
931
932 return NULL;
933 }
934
935 // Visitors ...
936
monitors_iterate(MonitorClosure * closure)937 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) {
938 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
939 while (block != NULL) {
940 assert(block->object() == CHAINMARKER, "must be a block header");
941 for (int i = _BLOCKSIZE - 1; i > 0; i--) {
942 ObjectMonitor* mid = (ObjectMonitor *)(block + i);
943 oop object = (oop)mid->object();
944 if (object != NULL) {
945 closure->do_monitor(mid);
946 }
947 }
948 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
949 }
950 }
951
952 // Get the next block in the block list.
next(PaddedEnd<ObjectMonitor> * block)953 static inline PaddedEnd<ObjectMonitor>* next(PaddedEnd<ObjectMonitor>* block) {
954 assert(block->object() == CHAINMARKER, "must be a block header");
955 block = (PaddedEnd<ObjectMonitor>*) block->FreeNext;
956 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header");
957 return block;
958 }
959
monitors_used_above_threshold()960 static bool monitors_used_above_threshold() {
961 if (gMonitorPopulation == 0) {
962 return false;
963 }
964 int monitors_used = gMonitorPopulation - gMonitorFreeCount;
965 int monitor_usage = (monitors_used * 100LL) / gMonitorPopulation;
966 return monitor_usage > MonitorUsedDeflationThreshold;
967 }
968
is_cleanup_needed()969 bool ObjectSynchronizer::is_cleanup_needed() {
970 if (MonitorUsedDeflationThreshold > 0) {
971 return monitors_used_above_threshold();
972 }
973 return false;
974 }
975
oops_do(OopClosure * f)976 void ObjectSynchronizer::oops_do(OopClosure* f) {
977 if (MonitorInUseLists) {
978 // When using thread local monitor lists, we only scan the
979 // global used list here (for moribund threads), and
980 // the thread-local monitors in Thread::oops_do().
981 global_used_oops_do(f);
982 } else {
983 global_oops_do(f);
984 }
985 }
986
global_oops_do(OopClosure * f)987 void ObjectSynchronizer::global_oops_do(OopClosure* f) {
988 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
989 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
990 for (; block != NULL; block = next(block)) {
991 assert(block->object() == CHAINMARKER, "must be a block header");
992 for (int i = 1; i < _BLOCKSIZE; i++) {
993 ObjectMonitor* mid = (ObjectMonitor *)&block[i];
994 if (mid->object() != NULL) {
995 f->do_oop((oop*)mid->object_addr());
996 }
997 }
998 }
999 }
1000
global_used_oops_do(OopClosure * f)1001 void ObjectSynchronizer::global_used_oops_do(OopClosure* f) {
1002 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1003 list_oops_do(gOmInUseList, f);
1004 }
1005
thread_local_used_oops_do(Thread * thread,OopClosure * f)1006 void ObjectSynchronizer::thread_local_used_oops_do(Thread* thread, OopClosure* f) {
1007 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1008 list_oops_do(thread->omInUseList, f);
1009 }
1010
list_oops_do(ObjectMonitor * list,OopClosure * f)1011 void ObjectSynchronizer::list_oops_do(ObjectMonitor* list, OopClosure* f) {
1012 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1013 ObjectMonitor* mid;
1014 for (mid = list; mid != NULL; mid = mid->FreeNext) {
1015 if (mid->object() != NULL) {
1016 f->do_oop((oop*)mid->object_addr());
1017 }
1018 }
1019 }
1020
1021
1022 // -----------------------------------------------------------------------------
1023 // ObjectMonitor Lifecycle
1024 // -----------------------
1025 // Inflation unlinks monitors from the global gFreeList and
1026 // associates them with objects. Deflation -- which occurs at
1027 // STW-time -- disassociates idle monitors from objects. Such
1028 // scavenged monitors are returned to the gFreeList.
1029 //
1030 // The global list is protected by gListLock. All the critical sections
1031 // are short and operate in constant-time.
1032 //
1033 // ObjectMonitors reside in type-stable memory (TSM) and are immortal.
1034 //
1035 // Lifecycle:
1036 // -- unassigned and on the global free list
1037 // -- unassigned and on a thread's private omFreeList
1038 // -- assigned to an object. The object is inflated and the mark refers
1039 // to the objectmonitor.
1040
1041
1042 // Constraining monitor pool growth via MonitorBound ...
1043 //
1044 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the
1045 // the rate of scavenging is driven primarily by GC. As such, we can find
1046 // an inordinate number of monitors in circulation.
1047 // To avoid that scenario we can artificially induce a STW safepoint
1048 // if the pool appears to be growing past some reasonable bound.
1049 // Generally we favor time in space-time tradeoffs, but as there's no
1050 // natural back-pressure on the # of extant monitors we need to impose some
1051 // type of limit. Beware that if MonitorBound is set to too low a value
1052 // we could just loop. In addition, if MonitorBound is set to a low value
1053 // we'll incur more safepoints, which are harmful to performance.
1054 // See also: GuaranteedSafepointInterval
1055 //
1056 // The current implementation uses asynchronous VM operations.
1057
InduceScavenge(Thread * Self,const char * Whence)1058 static void InduceScavenge(Thread * Self, const char * Whence) {
1059 // Induce STW safepoint to trim monitors
1060 // Ultimately, this results in a call to deflate_idle_monitors() in the near future.
1061 // More precisely, trigger an asynchronous STW safepoint as the number
1062 // of active monitors passes the specified threshold.
1063 // TODO: assert thread state is reasonable
1064
1065 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) {
1066 if (ObjectMonitor::Knob_Verbose) {
1067 tty->print_cr("INFO: Monitor scavenge - Induced STW @%s (%d)",
1068 Whence, ForceMonitorScavenge) ;
1069 tty->flush();
1070 }
1071 // Induce a 'null' safepoint to scavenge monitors
1072 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted
1073 // to the VMthread and have a lifespan longer than that of this activation record.
1074 // The VMThread will delete the op when completed.
1075 VMThread::execute(new VM_ScavengeMonitors());
1076
1077 if (ObjectMonitor::Knob_Verbose) {
1078 tty->print_cr("INFO: Monitor scavenge - STW posted @%s (%d)",
1079 Whence, ForceMonitorScavenge) ;
1080 tty->flush();
1081 }
1082 }
1083 }
1084
verifyInUse(Thread * Self)1085 void ObjectSynchronizer::verifyInUse(Thread *Self) {
1086 ObjectMonitor* mid;
1087 int in_use_tally = 0;
1088 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) {
1089 in_use_tally++;
1090 }
1091 assert(in_use_tally == Self->omInUseCount, "in-use count off");
1092
1093 int free_tally = 0;
1094 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) {
1095 free_tally++;
1096 }
1097 assert(free_tally == Self->omFreeCount, "free count off");
1098 }
1099
omAlloc(Thread * Self)1100 ObjectMonitor* ObjectSynchronizer::omAlloc(Thread * Self) {
1101 // A large MAXPRIVATE value reduces both list lock contention
1102 // and list coherency traffic, but also tends to increase the
1103 // number of objectMonitors in circulation as well as the STW
1104 // scavenge costs. As usual, we lean toward time in space-time
1105 // tradeoffs.
1106 const int MAXPRIVATE = 1024;
1107 for (;;) {
1108 ObjectMonitor * m;
1109
1110 // 1: try to allocate from the thread's local omFreeList.
1111 // Threads will attempt to allocate first from their local list, then
1112 // from the global list, and only after those attempts fail will the thread
1113 // attempt to instantiate new monitors. Thread-local free lists take
1114 // heat off the gListLock and improve allocation latency, as well as reducing
1115 // coherency traffic on the shared global list.
1116 m = Self->omFreeList;
1117 if (m != NULL) {
1118 Self->omFreeList = m->FreeNext;
1119 Self->omFreeCount--;
1120 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene
1121 guarantee(m->object() == NULL, "invariant");
1122 if (MonitorInUseLists) {
1123 m->FreeNext = Self->omInUseList;
1124 Self->omInUseList = m;
1125 Self->omInUseCount++;
1126 if (ObjectMonitor::Knob_VerifyInUse) {
1127 verifyInUse(Self);
1128 }
1129 } else {
1130 m->FreeNext = NULL;
1131 }
1132 return m;
1133 }
1134
1135 // 2: try to allocate from the global gFreeList
1136 // CONSIDER: use muxTry() instead of muxAcquire().
1137 // If the muxTry() fails then drop immediately into case 3.
1138 // If we're using thread-local free lists then try
1139 // to reprovision the caller's free list.
1140 if (gFreeList != NULL) {
1141 // Reprovision the thread's omFreeList.
1142 // Use bulk transfers to reduce the allocation rate and heat
1143 // on various locks.
1144 Thread::muxAcquire(&gListLock, "omAlloc");
1145 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL;) {
1146 gMonitorFreeCount--;
1147 ObjectMonitor * take = gFreeList;
1148 gFreeList = take->FreeNext;
1149 guarantee(take->object() == NULL, "invariant");
1150 guarantee(!take->is_busy(), "invariant");
1151 take->Recycle();
1152 omRelease(Self, take, false);
1153 }
1154 Thread::muxRelease(&gListLock);
1155 Self->omFreeProvision += 1 + (Self->omFreeProvision/2);
1156 if (Self->omFreeProvision > MAXPRIVATE) Self->omFreeProvision = MAXPRIVATE;
1157 TEVENT(omFirst - reprovision);
1158
1159 const int mx = MonitorBound;
1160 if (mx > 0 && (gMonitorPopulation-gMonitorFreeCount) > mx) {
1161 // We can't safely induce a STW safepoint from omAlloc() as our thread
1162 // state may not be appropriate for such activities and callers may hold
1163 // naked oops, so instead we defer the action.
1164 InduceScavenge(Self, "omAlloc");
1165 }
1166 continue;
1167 }
1168
1169 // 3: allocate a block of new ObjectMonitors
1170 // Both the local and global free lists are empty -- resort to malloc().
1171 // In the current implementation objectMonitors are TSM - immortal.
1172 // Ideally, we'd write "new ObjectMonitor[_BLOCKSIZE], but we want
1173 // each ObjectMonitor to start at the beginning of a cache line,
1174 // so we use align_up().
1175 // A better solution would be to use C++ placement-new.
1176 // BEWARE: As it stands currently, we don't run the ctors!
1177 assert(_BLOCKSIZE > 1, "invariant");
1178 size_t neededsize = sizeof(PaddedEnd<ObjectMonitor>) * _BLOCKSIZE;
1179 PaddedEnd<ObjectMonitor> * temp;
1180 size_t aligned_size = neededsize + (DEFAULT_CACHE_LINE_SIZE - 1);
1181 void* real_malloc_addr = (void *)NEW_C_HEAP_ARRAY(char, aligned_size,
1182 mtInternal);
1183 temp = (PaddedEnd<ObjectMonitor> *)
1184 align_up(real_malloc_addr, DEFAULT_CACHE_LINE_SIZE);
1185
1186 // NOTE: (almost) no way to recover if allocation failed.
1187 // We might be able to induce a STW safepoint and scavenge enough
1188 // objectMonitors to permit progress.
1189 if (temp == NULL) {
1190 vm_exit_out_of_memory(neededsize, OOM_MALLOC_ERROR,
1191 "Allocate ObjectMonitors");
1192 }
1193 (void)memset((void *) temp, 0, neededsize);
1194
1195 // Format the block.
1196 // initialize the linked list, each monitor points to its next
1197 // forming the single linked free list, the very first monitor
1198 // will points to next block, which forms the block list.
1199 // The trick of using the 1st element in the block as gBlockList
1200 // linkage should be reconsidered. A better implementation would
1201 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; }
1202
1203 for (int i = 1; i < _BLOCKSIZE; i++) {
1204 temp[i].FreeNext = (ObjectMonitor *)&temp[i+1];
1205 }
1206
1207 // terminate the last monitor as the end of list
1208 temp[_BLOCKSIZE - 1].FreeNext = NULL;
1209
1210 // Element [0] is reserved for global list linkage
1211 temp[0].set_object(CHAINMARKER);
1212
1213 // Consider carving out this thread's current request from the
1214 // block in hand. This avoids some lock traffic and redundant
1215 // list activity.
1216
1217 // Acquire the gListLock to manipulate gBlockList and gFreeList.
1218 // An Oyama-Taura-Yonezawa scheme might be more efficient.
1219 Thread::muxAcquire(&gListLock, "omAlloc [2]");
1220 gMonitorPopulation += _BLOCKSIZE-1;
1221 gMonitorFreeCount += _BLOCKSIZE-1;
1222
1223 // Add the new block to the list of extant blocks (gBlockList).
1224 // The very first objectMonitor in a block is reserved and dedicated.
1225 // It serves as blocklist "next" linkage.
1226 temp[0].FreeNext = gBlockList;
1227 // There are lock-free uses of gBlockList so make sure that
1228 // the previous stores happen before we update gBlockList.
1229 OrderAccess::release_store(&gBlockList, temp);
1230
1231 // Add the new string of objectMonitors to the global free list
1232 temp[_BLOCKSIZE - 1].FreeNext = gFreeList;
1233 gFreeList = temp + 1;
1234 Thread::muxRelease(&gListLock);
1235 TEVENT(Allocate block of monitors);
1236 }
1237 }
1238
1239 // Place "m" on the caller's private per-thread omFreeList.
1240 // In practice there's no need to clamp or limit the number of
1241 // monitors on a thread's omFreeList as the only time we'll call
1242 // omRelease is to return a monitor to the free list after a CAS
1243 // attempt failed. This doesn't allow unbounded #s of monitors to
1244 // accumulate on a thread's free list.
1245 //
1246 // Key constraint: all ObjectMonitors on a thread's free list and the global
1247 // free list must have their object field set to null. This prevents the
1248 // scavenger -- deflate_idle_monitors -- from reclaiming them.
1249
omRelease(Thread * Self,ObjectMonitor * m,bool fromPerThreadAlloc)1250 void ObjectSynchronizer::omRelease(Thread * Self, ObjectMonitor * m,
1251 bool fromPerThreadAlloc) {
1252 guarantee(m->object() == NULL, "invariant");
1253 guarantee(((m->is_busy()|m->_recursions) == 0), "freeing in-use monitor");
1254 // Remove from omInUseList
1255 if (MonitorInUseLists && fromPerThreadAlloc) {
1256 ObjectMonitor* cur_mid_in_use = NULL;
1257 bool extracted = false;
1258 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; cur_mid_in_use = mid, mid = mid->FreeNext) {
1259 if (m == mid) {
1260 // extract from per-thread in-use list
1261 if (mid == Self->omInUseList) {
1262 Self->omInUseList = mid->FreeNext;
1263 } else if (cur_mid_in_use != NULL) {
1264 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1265 }
1266 extracted = true;
1267 Self->omInUseCount--;
1268 if (ObjectMonitor::Knob_VerifyInUse) {
1269 verifyInUse(Self);
1270 }
1271 break;
1272 }
1273 }
1274 assert(extracted, "Should have extracted from in-use list");
1275 }
1276
1277 // FreeNext is used for both omInUseList and omFreeList, so clear old before setting new
1278 m->FreeNext = Self->omFreeList;
1279 Self->omFreeList = m;
1280 Self->omFreeCount++;
1281 }
1282
1283 // Return the monitors of a moribund thread's local free list to
1284 // the global free list. Typically a thread calls omFlush() when
1285 // it's dying. We could also consider having the VM thread steal
1286 // monitors from threads that have not run java code over a few
1287 // consecutive STW safepoints. Relatedly, we might decay
1288 // omFreeProvision at STW safepoints.
1289 //
1290 // Also return the monitors of a moribund thread's omInUseList to
1291 // a global gOmInUseList under the global list lock so these
1292 // will continue to be scanned.
1293 //
1294 // We currently call omFlush() from Threads::remove() _before the thread
1295 // has been excised from the thread list and is no longer a mutator.
1296 // This means that omFlush() can not run concurrently with a safepoint and
1297 // interleave with the scavenge operator. In particular, this ensures that
1298 // the thread's monitors are scanned by a GC safepoint, either via
1299 // Thread::oops_do() (if safepoint happens before omFlush()) or via
1300 // ObjectSynchronizer::oops_do() (if it happens after omFlush() and the thread's
1301 // monitors have been transferred to the global in-use list).
1302
omFlush(Thread * Self)1303 void ObjectSynchronizer::omFlush(Thread * Self) {
1304 ObjectMonitor * list = Self->omFreeList; // Null-terminated SLL
1305 Self->omFreeList = NULL;
1306 ObjectMonitor * tail = NULL;
1307 int tally = 0;
1308 if (list != NULL) {
1309 ObjectMonitor * s;
1310 // The thread is going away, the per-thread free monitors
1311 // are freed via set_owner(NULL)
1312 // Link them to tail, which will be linked into the global free list
1313 // gFreeList below, under the gListLock
1314 for (s = list; s != NULL; s = s->FreeNext) {
1315 tally++;
1316 tail = s;
1317 guarantee(s->object() == NULL, "invariant");
1318 guarantee(!s->is_busy(), "invariant");
1319 s->set_owner(NULL); // redundant but good hygiene
1320 TEVENT(omFlush - Move one);
1321 }
1322 guarantee(tail != NULL && list != NULL, "invariant");
1323 }
1324
1325 ObjectMonitor * inUseList = Self->omInUseList;
1326 ObjectMonitor * inUseTail = NULL;
1327 int inUseTally = 0;
1328 if (inUseList != NULL) {
1329 Self->omInUseList = NULL;
1330 ObjectMonitor *cur_om;
1331 // The thread is going away, however the omInUseList inflated
1332 // monitors may still be in-use by other threads.
1333 // Link them to inUseTail, which will be linked into the global in-use list
1334 // gOmInUseList below, under the gListLock
1335 for (cur_om = inUseList; cur_om != NULL; cur_om = cur_om->FreeNext) {
1336 inUseTail = cur_om;
1337 inUseTally++;
1338 }
1339 assert(Self->omInUseCount == inUseTally, "in-use count off");
1340 Self->omInUseCount = 0;
1341 guarantee(inUseTail != NULL && inUseList != NULL, "invariant");
1342 }
1343
1344 Thread::muxAcquire(&gListLock, "omFlush");
1345 if (tail != NULL) {
1346 tail->FreeNext = gFreeList;
1347 gFreeList = list;
1348 gMonitorFreeCount += tally;
1349 assert(Self->omFreeCount == tally, "free-count off");
1350 Self->omFreeCount = 0;
1351 }
1352
1353 if (inUseTail != NULL) {
1354 inUseTail->FreeNext = gOmInUseList;
1355 gOmInUseList = inUseList;
1356 gOmInUseCount += inUseTally;
1357 }
1358
1359 Thread::muxRelease(&gListLock);
1360 TEVENT(omFlush);
1361 }
1362
post_monitor_inflate_event(EventJavaMonitorInflate * event,const oop obj,ObjectSynchronizer::InflateCause cause)1363 static void post_monitor_inflate_event(EventJavaMonitorInflate* event,
1364 const oop obj,
1365 ObjectSynchronizer::InflateCause cause) {
1366 assert(event != NULL, "invariant");
1367 assert(event->should_commit(), "invariant");
1368 event->set_monitorClass(obj->klass());
1369 event->set_address((uintptr_t)(void*)obj);
1370 event->set_cause((u1)cause);
1371 event->commit();
1372 }
1373
1374 // Fast path code shared by multiple functions
inflate_helper(oop obj)1375 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) {
1376 markOop mark = obj->mark();
1377 if (mark->has_monitor()) {
1378 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid");
1379 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header");
1380 return mark->monitor();
1381 }
1382 return ObjectSynchronizer::inflate(Thread::current(),
1383 obj,
1384 inflate_cause_vm_internal);
1385 }
1386
inflate(Thread * Self,oop object,const InflateCause cause)1387 ObjectMonitor* ObjectSynchronizer::inflate(Thread * Self,
1388 oop object,
1389 const InflateCause cause) {
1390
1391 // Inflate mutates the heap ...
1392 // Relaxing assertion for bug 6320749.
1393 assert(Universe::verify_in_progress() ||
1394 !SafepointSynchronize::is_at_safepoint(), "invariant");
1395
1396 EventJavaMonitorInflate event;
1397
1398 for (;;) {
1399 const markOop mark = object->mark();
1400 assert(!mark->has_bias_pattern(), "invariant");
1401
1402 // The mark can be in one of the following states:
1403 // * Inflated - just return
1404 // * Stack-locked - coerce it to inflated
1405 // * INFLATING - busy wait for conversion to complete
1406 // * Neutral - aggressively inflate the object.
1407 // * BIASED - Illegal. We should never see this
1408
1409 // CASE: inflated
1410 if (mark->has_monitor()) {
1411 ObjectMonitor * inf = mark->monitor();
1412 assert(inf->header()->is_neutral(), "invariant");
1413 assert(inf->object() == object, "invariant");
1414 assert(ObjectSynchronizer::verify_objmon_isinpool(inf), "monitor is invalid");
1415 return inf;
1416 }
1417
1418 // CASE: inflation in progress - inflating over a stack-lock.
1419 // Some other thread is converting from stack-locked to inflated.
1420 // Only that thread can complete inflation -- other threads must wait.
1421 // The INFLATING value is transient.
1422 // Currently, we spin/yield/park and poll the markword, waiting for inflation to finish.
1423 // We could always eliminate polling by parking the thread on some auxiliary list.
1424 if (mark == markOopDesc::INFLATING()) {
1425 TEVENT(Inflate: spin while INFLATING);
1426 ReadStableMark(object);
1427 continue;
1428 }
1429
1430 // CASE: stack-locked
1431 // Could be stack-locked either by this thread or by some other thread.
1432 //
1433 // Note that we allocate the objectmonitor speculatively, _before_ attempting
1434 // to install INFLATING into the mark word. We originally installed INFLATING,
1435 // allocated the objectmonitor, and then finally STed the address of the
1436 // objectmonitor into the mark. This was correct, but artificially lengthened
1437 // the interval in which INFLATED appeared in the mark, thus increasing
1438 // the odds of inflation contention.
1439 //
1440 // We now use per-thread private objectmonitor free lists.
1441 // These list are reprovisioned from the global free list outside the
1442 // critical INFLATING...ST interval. A thread can transfer
1443 // multiple objectmonitors en-mass from the global free list to its local free list.
1444 // This reduces coherency traffic and lock contention on the global free list.
1445 // Using such local free lists, it doesn't matter if the omAlloc() call appears
1446 // before or after the CAS(INFLATING) operation.
1447 // See the comments in omAlloc().
1448
1449 if (mark->has_locker()) {
1450 ObjectMonitor * m = omAlloc(Self);
1451 // Optimistically prepare the objectmonitor - anticipate successful CAS
1452 // We do this before the CAS in order to minimize the length of time
1453 // in which INFLATING appears in the mark.
1454 m->Recycle();
1455 m->_Responsible = NULL;
1456 m->_recursions = 0;
1457 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // Consider: maintain by type/class
1458
1459 markOop cmp = object->cas_set_mark(markOopDesc::INFLATING(), mark);
1460 if (cmp != mark) {
1461 omRelease(Self, m, true);
1462 continue; // Interference -- just retry
1463 }
1464
1465 // We've successfully installed INFLATING (0) into the mark-word.
1466 // This is the only case where 0 will appear in a mark-word.
1467 // Only the singular thread that successfully swings the mark-word
1468 // to 0 can perform (or more precisely, complete) inflation.
1469 //
1470 // Why do we CAS a 0 into the mark-word instead of just CASing the
1471 // mark-word from the stack-locked value directly to the new inflated state?
1472 // Consider what happens when a thread unlocks a stack-locked object.
1473 // It attempts to use CAS to swing the displaced header value from the
1474 // on-stack basiclock back into the object header. Recall also that the
1475 // header value (hashcode, etc) can reside in (a) the object header, or
1476 // (b) a displaced header associated with the stack-lock, or (c) a displaced
1477 // header in an objectMonitor. The inflate() routine must copy the header
1478 // value from the basiclock on the owner's stack to the objectMonitor, all
1479 // the while preserving the hashCode stability invariants. If the owner
1480 // decides to release the lock while the value is 0, the unlock will fail
1481 // and control will eventually pass from slow_exit() to inflate. The owner
1482 // will then spin, waiting for the 0 value to disappear. Put another way,
1483 // the 0 causes the owner to stall if the owner happens to try to
1484 // drop the lock (restoring the header from the basiclock to the object)
1485 // while inflation is in-progress. This protocol avoids races that might
1486 // would otherwise permit hashCode values to change or "flicker" for an object.
1487 // Critically, while object->mark is 0 mark->displaced_mark_helper() is stable.
1488 // 0 serves as a "BUSY" inflate-in-progress indicator.
1489
1490
1491 // fetch the displaced mark from the owner's stack.
1492 // The owner can't die or unwind past the lock while our INFLATING
1493 // object is in the mark. Furthermore the owner can't complete
1494 // an unlock on the object, either.
1495 markOop dmw = mark->displaced_mark_helper();
1496 assert(dmw->is_neutral(), "invariant");
1497
1498 // Setup monitor fields to proper values -- prepare the monitor
1499 m->set_header(dmw);
1500
1501 // Optimization: if the mark->locker stack address is associated
1502 // with this thread we could simply set m->_owner = Self.
1503 // Note that a thread can inflate an object
1504 // that it has stack-locked -- as might happen in wait() -- directly
1505 // with CAS. That is, we can avoid the xchg-NULL .... ST idiom.
1506 m->set_owner(mark->locker());
1507 m->set_object(object);
1508 // TODO-FIXME: assert BasicLock->dhw != 0.
1509
1510 // Must preserve store ordering. The monitor state must
1511 // be stable at the time of publishing the monitor address.
1512 guarantee(object->mark() == markOopDesc::INFLATING(), "invariant");
1513 object->release_set_mark(markOopDesc::encode(m));
1514
1515 // Hopefully the performance counters are allocated on distinct cache lines
1516 // to avoid false sharing on MP systems ...
1517 OM_PERFDATA_OP(Inflations, inc());
1518 TEVENT(Inflate: overwrite stacklock);
1519 if (log_is_enabled(Debug, monitorinflation)) {
1520 if (object->is_instance()) {
1521 ResourceMark rm;
1522 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1523 p2i(object), p2i(object->mark()),
1524 object->klass()->external_name());
1525 }
1526 }
1527 if (event.should_commit()) {
1528 post_monitor_inflate_event(&event, object, cause);
1529 }
1530 return m;
1531 }
1532
1533 // CASE: neutral
1534 // TODO-FIXME: for entry we currently inflate and then try to CAS _owner.
1535 // If we know we're inflating for entry it's better to inflate by swinging a
1536 // pre-locked objectMonitor pointer into the object header. A successful
1537 // CAS inflates the object *and* confers ownership to the inflating thread.
1538 // In the current implementation we use a 2-step mechanism where we CAS()
1539 // to inflate and then CAS() again to try to swing _owner from NULL to Self.
1540 // An inflateTry() method that we could call from fast_enter() and slow_enter()
1541 // would be useful.
1542
1543 assert(mark->is_neutral(), "invariant");
1544 ObjectMonitor * m = omAlloc(Self);
1545 // prepare m for installation - set monitor to initial state
1546 m->Recycle();
1547 m->set_header(mark);
1548 m->set_owner(NULL);
1549 m->set_object(object);
1550 m->_recursions = 0;
1551 m->_Responsible = NULL;
1552 m->_SpinDuration = ObjectMonitor::Knob_SpinLimit; // consider: keep metastats by type/class
1553
1554 if (object->cas_set_mark(markOopDesc::encode(m), mark) != mark) {
1555 m->set_object(NULL);
1556 m->set_owner(NULL);
1557 m->Recycle();
1558 omRelease(Self, m, true);
1559 m = NULL;
1560 continue;
1561 // interference - the markword changed - just retry.
1562 // The state-transitions are one-way, so there's no chance of
1563 // live-lock -- "Inflated" is an absorbing state.
1564 }
1565
1566 // Hopefully the performance counters are allocated on distinct
1567 // cache lines to avoid false sharing on MP systems ...
1568 OM_PERFDATA_OP(Inflations, inc());
1569 TEVENT(Inflate: overwrite neutral);
1570 if (log_is_enabled(Debug, monitorinflation)) {
1571 if (object->is_instance()) {
1572 ResourceMark rm;
1573 log_debug(monitorinflation)("Inflating object " INTPTR_FORMAT " , mark " INTPTR_FORMAT " , type %s",
1574 p2i(object), p2i(object->mark()),
1575 object->klass()->external_name());
1576 }
1577 }
1578 if (event.should_commit()) {
1579 post_monitor_inflate_event(&event, object, cause);
1580 }
1581 return m;
1582 }
1583 }
1584
1585
1586 // Deflate_idle_monitors() is called at all safepoints, immediately
1587 // after all mutators are stopped, but before any objects have moved.
1588 // It traverses the list of known monitors, deflating where possible.
1589 // The scavenged monitor are returned to the monitor free list.
1590 //
1591 // Beware that we scavenge at *every* stop-the-world point.
1592 // Having a large number of monitors in-circulation negatively
1593 // impacts the performance of some applications (e.g., PointBase).
1594 // Broadly, we want to minimize the # of monitors in circulation.
1595 //
1596 // We have added a flag, MonitorInUseLists, which creates a list
1597 // of active monitors for each thread. deflate_idle_monitors()
1598 // only scans the per-thread in-use lists. omAlloc() puts all
1599 // assigned monitors on the per-thread list. deflate_idle_monitors()
1600 // returns the non-busy monitors to the global free list.
1601 // When a thread dies, omFlush() adds the list of active monitors for
1602 // that thread to a global gOmInUseList acquiring the
1603 // global list lock. deflate_idle_monitors() acquires the global
1604 // list lock to scan for non-busy monitors to the global free list.
1605 // An alternative could have used a single global in-use list. The
1606 // downside would have been the additional cost of acquiring the global list lock
1607 // for every omAlloc().
1608 //
1609 // Perversely, the heap size -- and thus the STW safepoint rate --
1610 // typically drives the scavenge rate. Large heaps can mean infrequent GC,
1611 // which in turn can mean large(r) numbers of objectmonitors in circulation.
1612 // This is an unfortunate aspect of this design.
1613
1614 enum ManifestConstants {
1615 ClearResponsibleAtSTW = 0
1616 };
1617
1618 // Deflate a single monitor if not in-use
1619 // Return true if deflated, false if in-use
deflate_monitor(ObjectMonitor * mid,oop obj,ObjectMonitor ** freeHeadp,ObjectMonitor ** freeTailp)1620 bool ObjectSynchronizer::deflate_monitor(ObjectMonitor* mid, oop obj,
1621 ObjectMonitor** freeHeadp,
1622 ObjectMonitor** freeTailp) {
1623 bool deflated;
1624 // Normal case ... The monitor is associated with obj.
1625 guarantee(obj->mark() == markOopDesc::encode(mid), "invariant");
1626 guarantee(mid == obj->mark()->monitor(), "invariant");
1627 guarantee(mid->header()->is_neutral(), "invariant");
1628
1629 if (mid->is_busy()) {
1630 if (ClearResponsibleAtSTW) mid->_Responsible = NULL;
1631 deflated = false;
1632 } else {
1633 // Deflate the monitor if it is no longer being used
1634 // It's idle - scavenge and return to the global free list
1635 // plain old deflation ...
1636 TEVENT(deflate_idle_monitors - scavenge1);
1637 if (log_is_enabled(Debug, monitorinflation)) {
1638 if (obj->is_instance()) {
1639 ResourceMark rm;
1640 log_debug(monitorinflation)("Deflating object " INTPTR_FORMAT " , "
1641 "mark " INTPTR_FORMAT " , type %s",
1642 p2i(obj), p2i(obj->mark()),
1643 obj->klass()->external_name());
1644 }
1645 }
1646
1647 // Restore the header back to obj
1648 obj->release_set_mark(mid->header());
1649 mid->clear();
1650
1651 assert(mid->object() == NULL, "invariant");
1652
1653 // Move the object to the working free list defined by freeHeadp, freeTailp
1654 if (*freeHeadp == NULL) *freeHeadp = mid;
1655 if (*freeTailp != NULL) {
1656 ObjectMonitor * prevtail = *freeTailp;
1657 assert(prevtail->FreeNext == NULL, "cleaned up deflated?");
1658 prevtail->FreeNext = mid;
1659 }
1660 *freeTailp = mid;
1661 deflated = true;
1662 }
1663 return deflated;
1664 }
1665
1666 // Walk a given monitor list, and deflate idle monitors
1667 // The given list could be a per-thread list or a global list
1668 // Caller acquires gListLock.
1669 //
1670 // In the case of parallel processing of thread local monitor lists,
1671 // work is done by Threads::parallel_threads_do() which ensures that
1672 // each Java thread is processed by exactly one worker thread, and
1673 // thus avoid conflicts that would arise when worker threads would
1674 // process the same monitor lists concurrently.
1675 //
1676 // See also ParallelSPCleanupTask and
1677 // SafepointSynchronize::do_cleanup_tasks() in safepoint.cpp and
1678 // Threads::parallel_java_threads_do() in thread.cpp.
deflate_monitor_list(ObjectMonitor ** listHeadp,ObjectMonitor ** freeHeadp,ObjectMonitor ** freeTailp)1679 int ObjectSynchronizer::deflate_monitor_list(ObjectMonitor** listHeadp,
1680 ObjectMonitor** freeHeadp,
1681 ObjectMonitor** freeTailp) {
1682 ObjectMonitor* mid;
1683 ObjectMonitor* next;
1684 ObjectMonitor* cur_mid_in_use = NULL;
1685 int deflated_count = 0;
1686
1687 for (mid = *listHeadp; mid != NULL;) {
1688 oop obj = (oop) mid->object();
1689 if (obj != NULL && deflate_monitor(mid, obj, freeHeadp, freeTailp)) {
1690 // if deflate_monitor succeeded,
1691 // extract from per-thread in-use list
1692 if (mid == *listHeadp) {
1693 *listHeadp = mid->FreeNext;
1694 } else if (cur_mid_in_use != NULL) {
1695 cur_mid_in_use->FreeNext = mid->FreeNext; // maintain the current thread in-use list
1696 }
1697 next = mid->FreeNext;
1698 mid->FreeNext = NULL; // This mid is current tail in the freeHeadp list
1699 mid = next;
1700 deflated_count++;
1701 } else {
1702 cur_mid_in_use = mid;
1703 mid = mid->FreeNext;
1704 }
1705 }
1706 return deflated_count;
1707 }
1708
prepare_deflate_idle_monitors(DeflateMonitorCounters * counters)1709 void ObjectSynchronizer::prepare_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1710 counters->nInuse = 0; // currently associated with objects
1711 counters->nInCirculation = 0; // extant
1712 counters->nScavenged = 0; // reclaimed
1713 }
1714
deflate_idle_monitors(DeflateMonitorCounters * counters)1715 void ObjectSynchronizer::deflate_idle_monitors(DeflateMonitorCounters* counters) {
1716 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1717 bool deflated = false;
1718
1719 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1720 ObjectMonitor * freeTailp = NULL;
1721
1722 TEVENT(deflate_idle_monitors);
1723 // Prevent omFlush from changing mids in Thread dtor's during deflation
1724 // And in case the vm thread is acquiring a lock during a safepoint
1725 // See e.g. 6320749
1726 Thread::muxAcquire(&gListLock, "scavenge - return");
1727
1728 if (MonitorInUseLists) {
1729 // Note: the thread-local monitors lists get deflated in
1730 // a separate pass. See deflate_thread_local_monitors().
1731
1732 // For moribund threads, scan gOmInUseList
1733 if (gOmInUseList) {
1734 counters->nInCirculation += gOmInUseCount;
1735 int deflated_count = deflate_monitor_list((ObjectMonitor **)&gOmInUseList, &freeHeadp, &freeTailp);
1736 gOmInUseCount -= deflated_count;
1737 counters->nScavenged += deflated_count;
1738 counters->nInuse += gOmInUseCount;
1739 }
1740
1741 } else {
1742 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
1743 for (; block != NULL; block = next(block)) {
1744 // Iterate over all extant monitors - Scavenge all idle monitors.
1745 assert(block->object() == CHAINMARKER, "must be a block header");
1746 counters->nInCirculation += _BLOCKSIZE;
1747 for (int i = 1; i < _BLOCKSIZE; i++) {
1748 ObjectMonitor* mid = (ObjectMonitor*)&block[i];
1749 oop obj = (oop)mid->object();
1750
1751 if (obj == NULL) {
1752 // The monitor is not associated with an object.
1753 // The monitor should either be a thread-specific private
1754 // free list or the global free list.
1755 // obj == NULL IMPLIES mid->is_busy() == 0
1756 guarantee(!mid->is_busy(), "invariant");
1757 continue;
1758 }
1759 deflated = deflate_monitor(mid, obj, &freeHeadp, &freeTailp);
1760
1761 if (deflated) {
1762 mid->FreeNext = NULL;
1763 counters->nScavenged++;
1764 } else {
1765 counters->nInuse++;
1766 }
1767 }
1768 }
1769 }
1770
1771 // Move the scavenged monitors back to the global free list.
1772 if (freeHeadp != NULL) {
1773 guarantee(freeTailp != NULL && counters->nScavenged > 0, "invariant");
1774 assert(freeTailp->FreeNext == NULL, "invariant");
1775 // constant-time list splice - prepend scavenged segment to gFreeList
1776 freeTailp->FreeNext = gFreeList;
1777 gFreeList = freeHeadp;
1778 }
1779 Thread::muxRelease(&gListLock);
1780
1781 }
1782
finish_deflate_idle_monitors(DeflateMonitorCounters * counters)1783 void ObjectSynchronizer::finish_deflate_idle_monitors(DeflateMonitorCounters* counters) {
1784 gMonitorFreeCount += counters->nScavenged;
1785
1786 // Consider: audit gFreeList to ensure that gMonitorFreeCount and list agree.
1787
1788 if (ObjectMonitor::Knob_Verbose) {
1789 tty->print_cr("INFO: Deflate: InCirc=%d InUse=%d Scavenged=%d "
1790 "ForceMonitorScavenge=%d : pop=%d free=%d",
1791 counters->nInCirculation, counters->nInuse, counters->nScavenged, ForceMonitorScavenge,
1792 gMonitorPopulation, gMonitorFreeCount);
1793 tty->flush();
1794 }
1795
1796 ForceMonitorScavenge = 0; // Reset
1797
1798 OM_PERFDATA_OP(Deflations, inc(counters->nScavenged));
1799 OM_PERFDATA_OP(MonExtant, set_value(counters->nInCirculation));
1800
1801 // TODO: Add objectMonitor leak detection.
1802 // Audit/inventory the objectMonitors -- make sure they're all accounted for.
1803 GVars.stwRandom = os::random();
1804 GVars.stwCycle++;
1805 }
1806
deflate_thread_local_monitors(Thread * thread,DeflateMonitorCounters * counters)1807 void ObjectSynchronizer::deflate_thread_local_monitors(Thread* thread, DeflateMonitorCounters* counters) {
1808 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint");
1809 if (!MonitorInUseLists) return;
1810
1811 ObjectMonitor * freeHeadp = NULL; // Local SLL of scavenged monitors
1812 ObjectMonitor * freeTailp = NULL;
1813
1814 int deflated_count = deflate_monitor_list(thread->omInUseList_addr(), &freeHeadp, &freeTailp);
1815
1816 Thread::muxAcquire(&gListLock, "scavenge - return");
1817
1818 // Adjust counters
1819 counters->nInCirculation += thread->omInUseCount;
1820 thread->omInUseCount -= deflated_count;
1821 if (ObjectMonitor::Knob_VerifyInUse) {
1822 verifyInUse(thread);
1823 }
1824 counters->nScavenged += deflated_count;
1825 counters->nInuse += thread->omInUseCount;
1826
1827 // Move the scavenged monitors back to the global free list.
1828 if (freeHeadp != NULL) {
1829 guarantee(freeTailp != NULL && deflated_count > 0, "invariant");
1830 assert(freeTailp->FreeNext == NULL, "invariant");
1831
1832 // constant-time list splice - prepend scavenged segment to gFreeList
1833 freeTailp->FreeNext = gFreeList;
1834 gFreeList = freeHeadp;
1835 }
1836 Thread::muxRelease(&gListLock);
1837 }
1838
1839 // Monitor cleanup on JavaThread::exit
1840
1841 // Iterate through monitor cache and attempt to release thread's monitors
1842 // Gives up on a particular monitor if an exception occurs, but continues
1843 // the overall iteration, swallowing the exception.
1844 class ReleaseJavaMonitorsClosure: public MonitorClosure {
1845 private:
1846 TRAPS;
1847
1848 public:
ReleaseJavaMonitorsClosure(Thread * thread)1849 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {}
do_monitor(ObjectMonitor * mid)1850 void do_monitor(ObjectMonitor* mid) {
1851 if (mid->owner() == THREAD) {
1852 if (ObjectMonitor::Knob_VerifyMatch != 0) {
1853 ResourceMark rm;
1854 Handle obj(THREAD, (oop) mid->object());
1855 tty->print("INFO: unexpected locked object:");
1856 javaVFrame::print_locked_object_class_name(tty, obj, "locked");
1857 fatal("exiting JavaThread=" INTPTR_FORMAT
1858 " unexpectedly owns ObjectMonitor=" INTPTR_FORMAT,
1859 p2i(THREAD), p2i(mid));
1860 }
1861 (void)mid->complete_exit(CHECK);
1862 }
1863 }
1864 };
1865
1866 // Release all inflated monitors owned by THREAD. Lightweight monitors are
1867 // ignored. This is meant to be called during JNI thread detach which assumes
1868 // all remaining monitors are heavyweight. All exceptions are swallowed.
1869 // Scanning the extant monitor list can be time consuming.
1870 // A simple optimization is to add a per-thread flag that indicates a thread
1871 // called jni_monitorenter() during its lifetime.
1872 //
1873 // Instead of No_Savepoint_Verifier it might be cheaper to
1874 // use an idiom of the form:
1875 // auto int tmp = SafepointSynchronize::_safepoint_counter ;
1876 // <code that must not run at safepoint>
1877 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ;
1878 // Since the tests are extremely cheap we could leave them enabled
1879 // for normal product builds.
1880
release_monitors_owned_by_thread(TRAPS)1881 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) {
1882 assert(THREAD == JavaThread::current(), "must be current Java thread");
1883 NoSafepointVerifier nsv;
1884 ReleaseJavaMonitorsClosure rjmc(THREAD);
1885 Thread::muxAcquire(&gListLock, "release_monitors_owned_by_thread");
1886 ObjectSynchronizer::monitors_iterate(&rjmc);
1887 Thread::muxRelease(&gListLock);
1888 THREAD->clear_pending_exception();
1889 }
1890
inflate_cause_name(const InflateCause cause)1891 const char* ObjectSynchronizer::inflate_cause_name(const InflateCause cause) {
1892 switch (cause) {
1893 case inflate_cause_vm_internal: return "VM Internal";
1894 case inflate_cause_monitor_enter: return "Monitor Enter";
1895 case inflate_cause_wait: return "Monitor Wait";
1896 case inflate_cause_notify: return "Monitor Notify";
1897 case inflate_cause_hash_code: return "Monitor Hash Code";
1898 case inflate_cause_jni_enter: return "JNI Monitor Enter";
1899 case inflate_cause_jni_exit: return "JNI Monitor Exit";
1900 default:
1901 ShouldNotReachHere();
1902 }
1903 return "Unknown";
1904 }
1905
1906 //------------------------------------------------------------------------------
1907 // Debugging code
1908
sanity_checks(const bool verbose,const uint cache_line_size,int * error_cnt_ptr,int * warning_cnt_ptr)1909 void ObjectSynchronizer::sanity_checks(const bool verbose,
1910 const uint cache_line_size,
1911 int *error_cnt_ptr,
1912 int *warning_cnt_ptr) {
1913 u_char *addr_begin = (u_char*)&GVars;
1914 u_char *addr_stwRandom = (u_char*)&GVars.stwRandom;
1915 u_char *addr_hcSequence = (u_char*)&GVars.hcSequence;
1916
1917 if (verbose) {
1918 tty->print_cr("INFO: sizeof(SharedGlobals)=" SIZE_FORMAT,
1919 sizeof(SharedGlobals));
1920 }
1921
1922 uint offset_stwRandom = (uint)(addr_stwRandom - addr_begin);
1923 if (verbose) tty->print_cr("INFO: offset(stwRandom)=%u", offset_stwRandom);
1924
1925 uint offset_hcSequence = (uint)(addr_hcSequence - addr_begin);
1926 if (verbose) {
1927 tty->print_cr("INFO: offset(_hcSequence)=%u", offset_hcSequence);
1928 }
1929
1930 if (cache_line_size != 0) {
1931 // We were able to determine the L1 data cache line size so
1932 // do some cache line specific sanity checks
1933
1934 if (offset_stwRandom < cache_line_size) {
1935 tty->print_cr("WARNING: the SharedGlobals.stwRandom field is closer "
1936 "to the struct beginning than a cache line which permits "
1937 "false sharing.");
1938 (*warning_cnt_ptr)++;
1939 }
1940
1941 if ((offset_hcSequence - offset_stwRandom) < cache_line_size) {
1942 tty->print_cr("WARNING: the SharedGlobals.stwRandom and "
1943 "SharedGlobals.hcSequence fields are closer than a cache "
1944 "line which permits false sharing.");
1945 (*warning_cnt_ptr)++;
1946 }
1947
1948 if ((sizeof(SharedGlobals) - offset_hcSequence) < cache_line_size) {
1949 tty->print_cr("WARNING: the SharedGlobals.hcSequence field is closer "
1950 "to the struct end than a cache line which permits false "
1951 "sharing.");
1952 (*warning_cnt_ptr)++;
1953 }
1954 }
1955 }
1956
1957 #ifndef PRODUCT
1958
1959 // Check if monitor belongs to the monitor cache
1960 // The list is grow-only so it's *relatively* safe to traverse
1961 // the list of extant blocks without taking a lock.
1962
verify_objmon_isinpool(ObjectMonitor * monitor)1963 int ObjectSynchronizer::verify_objmon_isinpool(ObjectMonitor *monitor) {
1964 PaddedEnd<ObjectMonitor> * block = OrderAccess::load_acquire(&gBlockList);
1965 while (block != NULL) {
1966 assert(block->object() == CHAINMARKER, "must be a block header");
1967 if (monitor > &block[0] && monitor < &block[_BLOCKSIZE]) {
1968 address mon = (address)monitor;
1969 address blk = (address)block;
1970 size_t diff = mon - blk;
1971 assert((diff % sizeof(PaddedEnd<ObjectMonitor>)) == 0, "must be aligned");
1972 return 1;
1973 }
1974 block = (PaddedEnd<ObjectMonitor> *)block->FreeNext;
1975 }
1976 return 0;
1977 }
1978
1979 #endif
1980