1 /*
2 * Copyright (c) 1998, 2020, 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 "jfr/jfrEvents.hpp"
28 #include "jfr/support/jfrThreadId.hpp"
29 #include "logging/log.hpp"
30 #include "logging/logStream.hpp"
31 #include "memory/allocation.inline.hpp"
32 #include "memory/resourceArea.hpp"
33 #include "oops/markWord.hpp"
34 #include "oops/oop.inline.hpp"
35 #include "runtime/atomic.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/orderAccess.hpp"
42 #include "runtime/osThread.hpp"
43 #include "runtime/safepointMechanism.inline.hpp"
44 #include "runtime/sharedRuntime.hpp"
45 #include "runtime/stubRoutines.hpp"
46 #include "runtime/thread.inline.hpp"
47 #include "services/threadService.hpp"
48 #include "utilities/dtrace.hpp"
49 #include "utilities/macros.hpp"
50 #include "utilities/preserveException.hpp"
51 #if INCLUDE_JFR
52 #include "jfr/support/jfrFlush.hpp"
53 #endif
54
55 #ifdef DTRACE_ENABLED
56
57 // Only bother with this argument setup if dtrace is available
58 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly.
59
60
61 #define DTRACE_MONITOR_PROBE_COMMON(obj, thread) \
62 char* bytes = NULL; \
63 int len = 0; \
64 jlong jtid = SharedRuntime::get_java_tid(thread); \
65 Symbol* klassname = ((oop)obj)->klass()->name(); \
66 if (klassname != NULL) { \
67 bytes = (char*)klassname->bytes(); \
68 len = klassname->utf8_length(); \
69 }
70
71 #define DTRACE_MONITOR_WAIT_PROBE(monitor, obj, thread, millis) \
72 { \
73 if (DTraceMonitorProbes) { \
74 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
75 HOTSPOT_MONITOR_WAIT(jtid, \
76 (monitor), bytes, len, (millis)); \
77 } \
78 }
79
80 #define HOTSPOT_MONITOR_contended__enter HOTSPOT_MONITOR_CONTENDED_ENTER
81 #define HOTSPOT_MONITOR_contended__entered HOTSPOT_MONITOR_CONTENDED_ENTERED
82 #define HOTSPOT_MONITOR_contended__exit HOTSPOT_MONITOR_CONTENDED_EXIT
83 #define HOTSPOT_MONITOR_notify HOTSPOT_MONITOR_NOTIFY
84 #define HOTSPOT_MONITOR_notifyAll HOTSPOT_MONITOR_NOTIFYALL
85
86 #define DTRACE_MONITOR_PROBE(probe, monitor, obj, thread) \
87 { \
88 if (DTraceMonitorProbes) { \
89 DTRACE_MONITOR_PROBE_COMMON(obj, thread); \
90 HOTSPOT_MONITOR_##probe(jtid, \
91 (uintptr_t)(monitor), bytes, len); \
92 } \
93 }
94
95 #else // ndef DTRACE_ENABLED
96
97 #define DTRACE_MONITOR_WAIT_PROBE(obj, thread, millis, mon) {;}
98 #define DTRACE_MONITOR_PROBE(probe, obj, thread, mon) {;}
99
100 #endif // ndef DTRACE_ENABLED
101
102 // Tunables ...
103 // The knob* variables are effectively final. Once set they should
104 // never be modified hence. Consider using __read_mostly with GCC.
105
106 int ObjectMonitor::Knob_SpinLimit = 5000; // derived by an external tool -
107
108 static int Knob_Bonus = 100; // spin success bonus
109 static int Knob_BonusB = 100; // spin success bonus
110 static int Knob_Penalty = 200; // spin failure penalty
111 static int Knob_Poverty = 1000;
112 static int Knob_FixedSpin = 0;
113 static int Knob_PreSpin = 10; // 20-100 likely better
114
115 DEBUG_ONLY(static volatile bool InitDone = false;)
116
117 // -----------------------------------------------------------------------------
118 // Theory of operations -- Monitors lists, thread residency, etc:
119 //
120 // * A thread acquires ownership of a monitor by successfully
121 // CAS()ing the _owner field from null to non-null.
122 //
123 // * Invariant: A thread appears on at most one monitor list --
124 // cxq, EntryList or WaitSet -- at any one time.
125 //
126 // * Contending threads "push" themselves onto the cxq with CAS
127 // and then spin/park.
128 //
129 // * After a contending thread eventually acquires the lock it must
130 // dequeue itself from either the EntryList or the cxq.
131 //
132 // * The exiting thread identifies and unparks an "heir presumptive"
133 // tentative successor thread on the EntryList. Critically, the
134 // exiting thread doesn't unlink the successor thread from the EntryList.
135 // After having been unparked, the wakee will recontend for ownership of
136 // the monitor. The successor (wakee) will either acquire the lock or
137 // re-park itself.
138 //
139 // Succession is provided for by a policy of competitive handoff.
140 // The exiting thread does _not_ grant or pass ownership to the
141 // successor thread. (This is also referred to as "handoff" succession").
142 // Instead the exiting thread releases ownership and possibly wakes
143 // a successor, so the successor can (re)compete for ownership of the lock.
144 // If the EntryList is empty but the cxq is populated the exiting
145 // thread will drain the cxq into the EntryList. It does so by
146 // by detaching the cxq (installing null with CAS) and folding
147 // the threads from the cxq into the EntryList. The EntryList is
148 // doubly linked, while the cxq is singly linked because of the
149 // CAS-based "push" used to enqueue recently arrived threads (RATs).
150 //
151 // * Concurrency invariants:
152 //
153 // -- only the monitor owner may access or mutate the EntryList.
154 // The mutex property of the monitor itself protects the EntryList
155 // from concurrent interference.
156 // -- Only the monitor owner may detach the cxq.
157 //
158 // * The monitor entry list operations avoid locks, but strictly speaking
159 // they're not lock-free. Enter is lock-free, exit is not.
160 // For a description of 'Methods and apparatus providing non-blocking access
161 // to a resource,' see U.S. Pat. No. 7844973.
162 //
163 // * The cxq can have multiple concurrent "pushers" but only one concurrent
164 // detaching thread. This mechanism is immune from the ABA corruption.
165 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious.
166 //
167 // * Taken together, the cxq and the EntryList constitute or form a
168 // single logical queue of threads stalled trying to acquire the lock.
169 // We use two distinct lists to improve the odds of a constant-time
170 // dequeue operation after acquisition (in the ::enter() epilogue) and
171 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm).
172 // A key desideratum is to minimize queue & monitor metadata manipulation
173 // that occurs while holding the monitor lock -- that is, we want to
174 // minimize monitor lock holds times. Note that even a small amount of
175 // fixed spinning will greatly reduce the # of enqueue-dequeue operations
176 // on EntryList|cxq. That is, spinning relieves contention on the "inner"
177 // locks and monitor metadata.
178 //
179 // Cxq points to the set of Recently Arrived Threads attempting entry.
180 // Because we push threads onto _cxq with CAS, the RATs must take the form of
181 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when
182 // the unlocking thread notices that EntryList is null but _cxq is != null.
183 //
184 // The EntryList is ordered by the prevailing queue discipline and
185 // can be organized in any convenient fashion, such as a doubly-linked list or
186 // a circular doubly-linked list. Critically, we want insert and delete operations
187 // to operate in constant-time. If we need a priority queue then something akin
188 // to Solaris' sleepq would work nicely. Viz.,
189 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c.
190 // Queue discipline is enforced at ::exit() time, when the unlocking thread
191 // drains the cxq into the EntryList, and orders or reorders the threads on the
192 // EntryList accordingly.
193 //
194 // Barring "lock barging", this mechanism provides fair cyclic ordering,
195 // somewhat similar to an elevator-scan.
196 //
197 // * The monitor synchronization subsystem avoids the use of native
198 // synchronization primitives except for the narrow platform-specific
199 // park-unpark abstraction. See the comments in os_solaris.cpp regarding
200 // the semantics of park-unpark. Put another way, this monitor implementation
201 // depends only on atomic operations and park-unpark. The monitor subsystem
202 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the
203 // underlying OS manages the READY<->RUN transitions.
204 //
205 // * Waiting threads reside on the WaitSet list -- wait() puts
206 // the caller onto the WaitSet.
207 //
208 // * notify() or notifyAll() simply transfers threads from the WaitSet to
209 // either the EntryList or cxq. Subsequent exit() operations will
210 // unpark the notifyee. Unparking a notifee in notify() is inefficient -
211 // it's likely the notifyee would simply impale itself on the lock held
212 // by the notifier.
213 //
214 // * An interesting alternative is to encode cxq as (List,LockByte) where
215 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary
216 // variable, like _recursions, in the scheme. The threads or Events that form
217 // the list would have to be aligned in 256-byte addresses. A thread would
218 // try to acquire the lock or enqueue itself with CAS, but exiting threads
219 // could use a 1-0 protocol and simply STB to set the LockByte to 0.
220 // Note that is is *not* word-tearing, but it does presume that full-word
221 // CAS operations are coherent with intermix with STB operations. That's true
222 // on most common processors.
223 //
224 // * See also http://blogs.sun.com/dave
225
226
227 void* ObjectMonitor::operator new (size_t size) throw() {
228 return AllocateHeap(size, mtInternal);
229 }
230 void* ObjectMonitor::operator new[] (size_t size) throw() {
231 return operator new (size);
232 }
233 void ObjectMonitor::operator delete(void* p) {
234 FreeHeap(p);
235 }
236 void ObjectMonitor::operator delete[] (void *p) {
237 operator delete(p);
238 }
239
240 // -----------------------------------------------------------------------------
241 // Enter support
242
243 bool ObjectMonitor::enter(TRAPS) {
244 // The following code is ordered to check the most common cases first
245 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors.
246 Thread * const Self = THREAD;
247
248 void* cur = try_set_owner_from(NULL, Self);
249 if (cur == NULL) {
250 assert(_recursions == 0, "invariant");
251 return true;
252 }
253
254 if (cur == Self) {
255 // TODO-FIXME: check for integer overflow! BUGID 6557169.
256 _recursions++;
257 return true;
258 }
259
260 if (Self->is_lock_owned((address)cur)) {
261 assert(_recursions == 0, "internal state error");
262 _recursions = 1;
263 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*.
264 return true;
265 }
266
267 // We've encountered genuine contention.
268 assert(Self->_Stalled == 0, "invariant");
269 Self->_Stalled = intptr_t(this);
270
271 // Try one round of spinning *before* enqueueing Self
272 // and before going through the awkward and expensive state
273 // transitions. The following spin is strictly optional ...
274 // Note that if we acquire the monitor from an initial spin
275 // we forgo posting JVMTI events and firing DTRACE probes.
276 if (TrySpin(Self) > 0) {
277 assert(_owner == Self, "must be Self: owner=" INTPTR_FORMAT, p2i(_owner));
278 assert(_recursions == 0, "must be 0: recursions=" INTX_FORMAT, _recursions);
279 assert(((oop)object())->mark() == markWord::encode(this),
280 "object mark must match encoded this: mark=" INTPTR_FORMAT
281 ", encoded this=" INTPTR_FORMAT, ((oop)object())->mark().value(),
282 markWord::encode(this).value());
283 Self->_Stalled = 0;
284 return true;
285 }
286
287 assert(_owner != Self, "invariant");
288 assert(_succ != Self, "invariant");
289 assert(Self->is_Java_thread(), "invariant");
290 JavaThread * jt = (JavaThread *) Self;
291 assert(!SafepointSynchronize::is_at_safepoint(), "invariant");
292 assert(jt->thread_state() != _thread_blocked, "invariant");
293
294 // Keep track of contention for JVM/TI and M&M queries.
295 add_to_contentions(1);
296 if (is_being_async_deflated()) {
297 // Async deflation is in progress and our contentions increment
298 // above lost the race to async deflation. Undo the work and
299 // force the caller to retry.
300 const oop l_object = (oop)object();
301 if (l_object != NULL) {
302 // Attempt to restore the header/dmw to the object's header so that
303 // we only retry once if the deflater thread happens to be slow.
304 install_displaced_markword_in_object(l_object);
305 }
306 Self->_Stalled = 0;
307 add_to_contentions(-1);
308 return false;
309 }
310
311 JFR_ONLY(JfrConditionalFlushWithStacktrace<EventJavaMonitorEnter> flush(jt);)
312 EventJavaMonitorEnter event;
313 if (event.should_commit()) {
314 event.set_monitorClass(((oop)this->object())->klass());
315 event.set_address((uintptr_t)(this->object_addr()));
316 }
317
318 { // Change java thread status to indicate blocked on monitor enter.
319 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this);
320
321 Self->set_current_pending_monitor(this);
322
323 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt);
324 if (JvmtiExport::should_post_monitor_contended_enter()) {
325 JvmtiExport::post_monitor_contended_enter(jt, this);
326
327 // The current thread does not yet own the monitor and does not
328 // yet appear on any queues that would get it made the successor.
329 // This means that the JVMTI_EVENT_MONITOR_CONTENDED_ENTER event
330 // handler cannot accidentally consume an unpark() meant for the
331 // ParkEvent associated with this ObjectMonitor.
332 }
333
334 OSThreadContendState osts(Self->osthread());
335 ThreadBlockInVM tbivm(jt);
336
337 // TODO-FIXME: change the following for(;;) loop to straight-line code.
338 for (;;) {
339 jt->set_suspend_equivalent();
340 // cleared by handle_special_suspend_equivalent_condition()
341 // or java_suspend_self()
342
343 EnterI(THREAD);
344
345 if (!ExitSuspendEquivalent(jt)) break;
346
347 // We have acquired the contended monitor, but while we were
348 // waiting another thread suspended us. We don't want to enter
349 // the monitor while suspended because that would surprise the
350 // thread that suspended us.
351 //
352 _recursions = 0;
353 _succ = NULL;
354 exit(false, Self);
355
356 jt->java_suspend_self();
357 }
358 Self->set_current_pending_monitor(NULL);
359
360 // We cleared the pending monitor info since we've just gotten past
361 // the enter-check-for-suspend dance and we now own the monitor free
362 // and clear, i.e., it is no longer pending. The ThreadBlockInVM
363 // destructor can go to a safepoint at the end of this block. If we
364 // do a thread dump during that safepoint, then this thread will show
365 // as having "-locked" the monitor, but the OS and java.lang.Thread
366 // states will still report that the thread is blocked trying to
367 // acquire it.
368 }
369
370 add_to_contentions(-1);
371 assert(contentions() >= 0, "must not be negative: contentions=%d", contentions());
372 Self->_Stalled = 0;
373
374 // Must either set _recursions = 0 or ASSERT _recursions == 0.
375 assert(_recursions == 0, "invariant");
376 assert(_owner == Self, "invariant");
377 assert(_succ != Self, "invariant");
378 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant");
379
380 // The thread -- now the owner -- is back in vm mode.
381 // Report the glorious news via TI,DTrace and jvmstat.
382 // The probe effect is non-trivial. All the reportage occurs
383 // while we hold the monitor, increasing the length of the critical
384 // section. Amdahl's parallel speedup law comes vividly into play.
385 //
386 // Another option might be to aggregate the events (thread local or
387 // per-monitor aggregation) and defer reporting until a more opportune
388 // time -- such as next time some thread encounters contention but has
389 // yet to acquire the lock. While spinning that thread could
390 // spinning we could increment JVMStat counters, etc.
391
392 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt);
393 if (JvmtiExport::should_post_monitor_contended_entered()) {
394 JvmtiExport::post_monitor_contended_entered(jt, this);
395
396 // The current thread already owns the monitor and is not going to
397 // call park() for the remainder of the monitor enter protocol. So
398 // it doesn't matter if the JVMTI_EVENT_MONITOR_CONTENDED_ENTERED
399 // event handler consumed an unpark() issued by the thread that
400 // just exited the monitor.
401 }
402 if (event.should_commit()) {
403 event.set_previousOwner((uintptr_t)_previous_owner_tid);
404 event.commit();
405 }
406 OM_PERFDATA_OP(ContendedLockAttempts, inc());
407 return true;
408 }
409
410 // Caveat: TryLock() is not necessarily serializing if it returns failure.
411 // Callers must compensate as needed.
412
413 int ObjectMonitor::TryLock(Thread * Self) {
414 void * own = _owner;
415 if (own != NULL) return 0;
416 if (try_set_owner_from(NULL, Self) == NULL) {
417 assert(_recursions == 0, "invariant");
418 return 1;
419 }
420 // The lock had been free momentarily, but we lost the race to the lock.
421 // Interference -- the CAS failed.
422 // We can either return -1 or retry.
423 // Retry doesn't make as much sense because the lock was just acquired.
424 return -1;
425 }
426
427 // Install the displaced mark word (dmw) of a deflating ObjectMonitor
428 // into the header of the object associated with the monitor. This
429 // idempotent method is called by a thread that is deflating a
430 // monitor and by other threads that have detected a race with the
431 // deflation process.
432 void ObjectMonitor::install_displaced_markword_in_object(const oop obj) {
433 // This function must only be called when (owner == DEFLATER_MARKER
434 // && contentions <= 0), but we can't guarantee that here because
435 // those values could change when the ObjectMonitor gets moved from
436 // the global free list to a per-thread free list.
437
438 guarantee(obj != NULL, "must be non-NULL");
439
440 // Separate loads in is_being_async_deflated(), which is almost always
441 // called before this function, from the load of dmw/header below.
442 if (support_IRIW_for_not_multiple_copy_atomic_cpu) {
443 // A non-multiple copy atomic (nMCA) machine needs a bigger
444 // hammer to separate the loads before and the load below.
445 OrderAccess::fence();
446 } else {
447 OrderAccess::loadload();
448 }
449
450 const oop l_object = (oop)object();
451 if (l_object == NULL) {
452 // ObjectMonitor's object ref has already been cleared by async
453 // deflation so we're done here.
454 return;
455 }
456 assert(l_object == obj, "object=" INTPTR_FORMAT " must equal obj="
457 INTPTR_FORMAT, p2i(l_object), p2i(obj));
458
459 markWord dmw = header();
460 // The dmw has to be neutral (not NULL, not locked and not marked).
461 assert(dmw.is_neutral(), "must be neutral: dmw=" INTPTR_FORMAT, dmw.value());
462
463 // Install displaced mark word if the object's header still points
464 // to this ObjectMonitor. More than one racing caller to this function
465 // can rarely reach this point, but only one can win.
466 markWord res = obj->cas_set_mark(dmw, markWord::encode(this));
467 if (res != markWord::encode(this)) {
468 // This should be rare so log at the Info level when it happens.
469 log_info(monitorinflation)("install_displaced_markword_in_object: "
470 "failed cas_set_mark: new_mark=" INTPTR_FORMAT
471 ", old_mark=" INTPTR_FORMAT ", res=" INTPTR_FORMAT,
472 dmw.value(), markWord::encode(this).value(),
473 res.value());
474 }
475
476 // Note: It does not matter which thread restored the header/dmw
477 // into the object's header. The thread deflating the monitor just
478 // wanted the object's header restored and it is. The threads that
479 // detected a race with the deflation process also wanted the
480 // object's header restored before they retry their operation and
481 // because it is restored they will only retry once.
482 }
483
484 // Convert the fields used by is_busy() to a string that can be
485 // used for diagnostic output.
486 const char* ObjectMonitor::is_busy_to_string(stringStream* ss) {
487 ss->print("is_busy: waiters=%d, ", _waiters);
488 if (contentions() > 0) {
489 ss->print("contentions=%d, ", contentions());
490 } else {
491 ss->print("contentions=0");
492 }
493 if (_owner != DEFLATER_MARKER) {
494 ss->print("owner=" INTPTR_FORMAT, p2i(_owner));
495 } else {
496 // We report NULL instead of DEFLATER_MARKER here because is_busy()
497 // ignores DEFLATER_MARKER values.
498 ss->print("owner=" INTPTR_FORMAT, NULL);
499 }
500 ss->print(", cxq=" INTPTR_FORMAT ", EntryList=" INTPTR_FORMAT, p2i(_cxq),
501 p2i(_EntryList));
502 return ss->base();
503 }
504
505 #define MAX_RECHECK_INTERVAL 1000
506
507 void ObjectMonitor::EnterI(TRAPS) {
508 Thread * const Self = THREAD;
509 assert(Self->is_Java_thread(), "invariant");
510 assert(((JavaThread *) Self)->thread_state() == _thread_blocked, "invariant");
511
512 // Try the lock - TATAS
513 if (TryLock (Self) > 0) {
514 assert(_succ != Self, "invariant");
515 assert(_owner == Self, "invariant");
516 assert(_Responsible != Self, "invariant");
517 return;
518 }
519
520 if (try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) {
521 // Cancelled the in-progress async deflation by changing owner from
522 // DEFLATER_MARKER to Self. As part of the contended enter protocol,
523 // contentions was incremented to a positive value before EnterI()
524 // was called and that prevents the deflater thread from winning the
525 // last part of the 2-part async deflation protocol. After EnterI()
526 // returns to enter(), contentions is decremented because the caller
527 // now owns the monitor. We bump contentions an extra time here to
528 // prevent the deflater thread from winning the last part of the
529 // 2-part async deflation protocol after the regular decrement
530 // occurs in enter(). The deflater thread will decrement contentions
531 // after it recognizes that the async deflation was cancelled.
532 add_to_contentions(1);
533 assert(_succ != Self, "invariant");
534 assert(_Responsible != Self, "invariant");
535 return;
536 }
537
538 assert(InitDone, "Unexpectedly not initialized");
539
540 // We try one round of spinning *before* enqueueing Self.
541 //
542 // If the _owner is ready but OFFPROC we could use a YieldTo()
543 // operation to donate the remainder of this thread's quantum
544 // to the owner. This has subtle but beneficial affinity
545 // effects.
546
547 if (TrySpin(Self) > 0) {
548 assert(_owner == Self, "invariant");
549 assert(_succ != Self, "invariant");
550 assert(_Responsible != Self, "invariant");
551 return;
552 }
553
554 // The Spin failed -- Enqueue and park the thread ...
555 assert(_succ != Self, "invariant");
556 assert(_owner != Self, "invariant");
557 assert(_Responsible != Self, "invariant");
558
559 // Enqueue "Self" on ObjectMonitor's _cxq.
560 //
561 // Node acts as a proxy for Self.
562 // As an aside, if were to ever rewrite the synchronization code mostly
563 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class
564 // Java objects. This would avoid awkward lifecycle and liveness issues,
565 // as well as eliminate a subset of ABA issues.
566 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events.
567
568 ObjectWaiter node(Self);
569 Self->_ParkEvent->reset();
570 node._prev = (ObjectWaiter *) 0xBAD;
571 node.TState = ObjectWaiter::TS_CXQ;
572
573 // Push "Self" onto the front of the _cxq.
574 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock.
575 // Note that spinning tends to reduce the rate at which threads
576 // enqueue and dequeue on EntryList|cxq.
577 ObjectWaiter * nxt;
578 for (;;) {
579 node._next = nxt = _cxq;
580 if (Atomic::cmpxchg(&_cxq, nxt, &node) == nxt) break;
581
582 // Interference - the CAS failed because _cxq changed. Just retry.
583 // As an optional optimization we retry the lock.
584 if (TryLock (Self) > 0) {
585 assert(_succ != Self, "invariant");
586 assert(_owner == Self, "invariant");
587 assert(_Responsible != Self, "invariant");
588 return;
589 }
590 }
591
592 // Check for cxq|EntryList edge transition to non-null. This indicates
593 // the onset of contention. While contention persists exiting threads
594 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit
595 // operations revert to the faster 1-0 mode. This enter operation may interleave
596 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we
597 // arrange for one of the contending thread to use a timed park() operations
598 // to detect and recover from the race. (Stranding is form of progress failure
599 // where the monitor is unlocked but all the contending threads remain parked).
600 // That is, at least one of the contended threads will periodically poll _owner.
601 // One of the contending threads will become the designated "Responsible" thread.
602 // The Responsible thread uses a timed park instead of a normal indefinite park
603 // operation -- it periodically wakes and checks for and recovers from potential
604 // strandings admitted by 1-0 exit operations. We need at most one Responsible
605 // thread per-monitor at any given moment. Only threads on cxq|EntryList may
606 // be responsible for a monitor.
607 //
608 // Currently, one of the contended threads takes on the added role of "Responsible".
609 // A viable alternative would be to use a dedicated "stranding checker" thread
610 // that periodically iterated over all the threads (or active monitors) and unparked
611 // successors where there was risk of stranding. This would help eliminate the
612 // timer scalability issues we see on some platforms as we'd only have one thread
613 // -- the checker -- parked on a timer.
614
615 if (nxt == NULL && _EntryList == NULL) {
616 // Try to assume the role of responsible thread for the monitor.
617 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self }
618 Atomic::replace_if_null(&_Responsible, Self);
619 }
620
621 // The lock might have been released while this thread was occupied queueing
622 // itself onto _cxq. To close the race and avoid "stranding" and
623 // progress-liveness failure we must resample-retry _owner before parking.
624 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner.
625 // In this case the ST-MEMBAR is accomplished with CAS().
626 //
627 // TODO: Defer all thread state transitions until park-time.
628 // Since state transitions are heavy and inefficient we'd like
629 // to defer the state transitions until absolutely necessary,
630 // and in doing so avoid some transitions ...
631
632 int nWakeups = 0;
633 int recheckInterval = 1;
634
635 for (;;) {
636
637 if (TryLock(Self) > 0) break;
638 assert(_owner != Self, "invariant");
639
640 // park self
641 if (_Responsible == Self) {
642 Self->_ParkEvent->park((jlong) recheckInterval);
643 // Increase the recheckInterval, but clamp the value.
644 recheckInterval *= 8;
645 if (recheckInterval > MAX_RECHECK_INTERVAL) {
646 recheckInterval = MAX_RECHECK_INTERVAL;
647 }
648 } else {
649 Self->_ParkEvent->park();
650 }
651
652 if (TryLock(Self) > 0) break;
653
654 if (try_set_owner_from(DEFLATER_MARKER, Self) == DEFLATER_MARKER) {
655 // Cancelled the in-progress async deflation by changing owner from
656 // DEFLATER_MARKER to Self. As part of the contended enter protocol,
657 // contentions was incremented to a positive value before EnterI()
658 // was called and that prevents the deflater thread from winning the
659 // last part of the 2-part async deflation protocol. After EnterI()
660 // returns to enter(), contentions is decremented because the caller
661 // now owns the monitor. We bump contentions an extra time here to
662 // prevent the deflater thread from winning the last part of the
663 // 2-part async deflation protocol after the regular decrement
664 // occurs in enter(). The deflater thread will decrement contentions
665 // after it recognizes that the async deflation was cancelled.
666 add_to_contentions(1);
667 break;
668 }
669
670 // The lock is still contested.
671 // Keep a tally of the # of futile wakeups.
672 // Note that the counter is not protected by a lock or updated by atomics.
673 // That is by design - we trade "lossy" counters which are exposed to
674 // races during updates for a lower probe effect.
675
676 // This PerfData object can be used in parallel with a safepoint.
677 // See the work around in PerfDataManager::destroy().
678 OM_PERFDATA_OP(FutileWakeups, inc());
679 ++nWakeups;
680
681 // Assuming this is not a spurious wakeup we'll normally find _succ == Self.
682 // We can defer clearing _succ until after the spin completes
683 // TrySpin() must tolerate being called with _succ == Self.
684 // Try yet another round of adaptive spinning.
685 if (TrySpin(Self) > 0) break;
686
687 // We can find that we were unpark()ed and redesignated _succ while
688 // we were spinning. That's harmless. If we iterate and call park(),
689 // park() will consume the event and return immediately and we'll
690 // just spin again. This pattern can repeat, leaving _succ to simply
691 // spin on a CPU.
692
693 if (_succ == Self) _succ = NULL;
694
695 // Invariant: after clearing _succ a thread *must* retry _owner before parking.
696 OrderAccess::fence();
697 }
698
699 // Egress :
700 // Self has acquired the lock -- Unlink Self from the cxq or EntryList.
701 // Normally we'll find Self on the EntryList .
702 // From the perspective of the lock owner (this thread), the
703 // EntryList is stable and cxq is prepend-only.
704 // The head of cxq is volatile but the interior is stable.
705 // In addition, Self.TState is stable.
706
707 assert(_owner == Self, "invariant");
708 assert(object() != NULL, "invariant");
709 // I'd like to write:
710 // guarantee (((oop)(object()))->mark() == markWord::encode(this), "invariant") ;
711 // but as we're at a safepoint that's not safe.
712
713 UnlinkAfterAcquire(Self, &node);
714 if (_succ == Self) _succ = NULL;
715
716 assert(_succ != Self, "invariant");
717 if (_Responsible == Self) {
718 _Responsible = NULL;
719 OrderAccess::fence(); // Dekker pivot-point
720
721 // We may leave threads on cxq|EntryList without a designated
722 // "Responsible" thread. This is benign. When this thread subsequently
723 // exits the monitor it can "see" such preexisting "old" threads --
724 // threads that arrived on the cxq|EntryList before the fence, above --
725 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads
726 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible
727 // non-null and elect a new "Responsible" timer thread.
728 //
729 // This thread executes:
730 // ST Responsible=null; MEMBAR (in enter epilogue - here)
731 // LD cxq|EntryList (in subsequent exit)
732 //
733 // Entering threads in the slow/contended path execute:
734 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog)
735 // The (ST cxq; MEMBAR) is accomplished with CAS().
736 //
737 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent
738 // exit operation from floating above the ST Responsible=null.
739 }
740
741 // We've acquired ownership with CAS().
742 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics.
743 // But since the CAS() this thread may have also stored into _succ,
744 // EntryList, cxq or Responsible. These meta-data updates must be
745 // visible __before this thread subsequently drops the lock.
746 // Consider what could occur if we didn't enforce this constraint --
747 // STs to monitor meta-data and user-data could reorder with (become
748 // visible after) the ST in exit that drops ownership of the lock.
749 // Some other thread could then acquire the lock, but observe inconsistent
750 // or old monitor meta-data and heap data. That violates the JMM.
751 // To that end, the 1-0 exit() operation must have at least STST|LDST
752 // "release" barrier semantics. Specifically, there must be at least a
753 // STST|LDST barrier in exit() before the ST of null into _owner that drops
754 // the lock. The barrier ensures that changes to monitor meta-data and data
755 // protected by the lock will be visible before we release the lock, and
756 // therefore before some other thread (CPU) has a chance to acquire the lock.
757 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html.
758 //
759 // Critically, any prior STs to _succ or EntryList must be visible before
760 // the ST of null into _owner in the *subsequent* (following) corresponding
761 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily
762 // execute a serializing instruction.
763
764 return;
765 }
766
767 // ReenterI() is a specialized inline form of the latter half of the
768 // contended slow-path from EnterI(). We use ReenterI() only for
769 // monitor reentry in wait().
770 //
771 // In the future we should reconcile EnterI() and ReenterI().
772
773 void ObjectMonitor::ReenterI(Thread * Self, ObjectWaiter * SelfNode) {
774 assert(Self != NULL, "invariant");
775 assert(SelfNode != NULL, "invariant");
776 assert(SelfNode->_thread == Self, "invariant");
777 assert(_waiters > 0, "invariant");
778 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant");
779 assert(((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant");
780 JavaThread * jt = (JavaThread *) Self;
781
782 int nWakeups = 0;
783 for (;;) {
784 ObjectWaiter::TStates v = SelfNode->TState;
785 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
786 assert(_owner != Self, "invariant");
787
788 if (TryLock(Self) > 0) break;
789 if (TrySpin(Self) > 0) break;
790
791 // State transition wrappers around park() ...
792 // ReenterI() wisely defers state transitions until
793 // it's clear we must park the thread.
794 {
795 OSThreadContendState osts(Self->osthread());
796 ThreadBlockInVM tbivm(jt);
797
798 // cleared by handle_special_suspend_equivalent_condition()
799 // or java_suspend_self()
800 jt->set_suspend_equivalent();
801 Self->_ParkEvent->park();
802
803 // were we externally suspended while we were waiting?
804 for (;;) {
805 if (!ExitSuspendEquivalent(jt)) break;
806 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); }
807 jt->java_suspend_self();
808 jt->set_suspend_equivalent();
809 }
810 }
811
812 // Try again, but just so we distinguish between futile wakeups and
813 // successful wakeups. The following test isn't algorithmically
814 // necessary, but it helps us maintain sensible statistics.
815 if (TryLock(Self) > 0) break;
816
817 // The lock is still contested.
818 // Keep a tally of the # of futile wakeups.
819 // Note that the counter is not protected by a lock or updated by atomics.
820 // That is by design - we trade "lossy" counters which are exposed to
821 // races during updates for a lower probe effect.
822 ++nWakeups;
823
824 // Assuming this is not a spurious wakeup we'll normally
825 // find that _succ == Self.
826 if (_succ == Self) _succ = NULL;
827
828 // Invariant: after clearing _succ a contending thread
829 // *must* retry _owner before parking.
830 OrderAccess::fence();
831
832 // This PerfData object can be used in parallel with a safepoint.
833 // See the work around in PerfDataManager::destroy().
834 OM_PERFDATA_OP(FutileWakeups, inc());
835 }
836
837 // Self has acquired the lock -- Unlink Self from the cxq or EntryList .
838 // Normally we'll find Self on the EntryList.
839 // Unlinking from the EntryList is constant-time and atomic-free.
840 // From the perspective of the lock owner (this thread), the
841 // EntryList is stable and cxq is prepend-only.
842 // The head of cxq is volatile but the interior is stable.
843 // In addition, Self.TState is stable.
844
845 assert(_owner == Self, "invariant");
846 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant");
847 UnlinkAfterAcquire(Self, SelfNode);
848 if (_succ == Self) _succ = NULL;
849 assert(_succ != Self, "invariant");
850 SelfNode->TState = ObjectWaiter::TS_RUN;
851 OrderAccess::fence(); // see comments at the end of EnterI()
852 }
853
854 // By convention we unlink a contending thread from EntryList|cxq immediately
855 // after the thread acquires the lock in ::enter(). Equally, we could defer
856 // unlinking the thread until ::exit()-time.
857
858 void ObjectMonitor::UnlinkAfterAcquire(Thread *Self, ObjectWaiter *SelfNode) {
859 assert(_owner == Self, "invariant");
860 assert(SelfNode->_thread == Self, "invariant");
861
862 if (SelfNode->TState == ObjectWaiter::TS_ENTER) {
863 // Normal case: remove Self from the DLL EntryList .
864 // This is a constant-time operation.
865 ObjectWaiter * nxt = SelfNode->_next;
866 ObjectWaiter * prv = SelfNode->_prev;
867 if (nxt != NULL) nxt->_prev = prv;
868 if (prv != NULL) prv->_next = nxt;
869 if (SelfNode == _EntryList) _EntryList = nxt;
870 assert(nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant");
871 assert(prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant");
872 } else {
873 assert(SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant");
874 // Inopportune interleaving -- Self is still on the cxq.
875 // This usually means the enqueue of self raced an exiting thread.
876 // Normally we'll find Self near the front of the cxq, so
877 // dequeueing is typically fast. If needbe we can accelerate
878 // this with some MCS/CHL-like bidirectional list hints and advisory
879 // back-links so dequeueing from the interior will normally operate
880 // in constant-time.
881 // Dequeue Self from either the head (with CAS) or from the interior
882 // with a linear-time scan and normal non-atomic memory operations.
883 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList
884 // and then unlink Self from EntryList. We have to drain eventually,
885 // so it might as well be now.
886
887 ObjectWaiter * v = _cxq;
888 assert(v != NULL, "invariant");
889 if (v != SelfNode || Atomic::cmpxchg(&_cxq, v, SelfNode->_next) != v) {
890 // The CAS above can fail from interference IFF a "RAT" arrived.
891 // In that case Self must be in the interior and can no longer be
892 // at the head of cxq.
893 if (v == SelfNode) {
894 assert(_cxq != v, "invariant");
895 v = _cxq; // CAS above failed - start scan at head of list
896 }
897 ObjectWaiter * p;
898 ObjectWaiter * q = NULL;
899 for (p = v; p != NULL && p != SelfNode; p = p->_next) {
900 q = p;
901 assert(p->TState == ObjectWaiter::TS_CXQ, "invariant");
902 }
903 assert(v != SelfNode, "invariant");
904 assert(p == SelfNode, "Node not found on cxq");
905 assert(p != _cxq, "invariant");
906 assert(q != NULL, "invariant");
907 assert(q->_next == p, "invariant");
908 q->_next = p->_next;
909 }
910 }
911
912 #ifdef ASSERT
913 // Diagnostic hygiene ...
914 SelfNode->_prev = (ObjectWaiter *) 0xBAD;
915 SelfNode->_next = (ObjectWaiter *) 0xBAD;
916 SelfNode->TState = ObjectWaiter::TS_RUN;
917 #endif
918 }
919
920 // -----------------------------------------------------------------------------
921 // Exit support
922 //
923 // exit()
924 // ~~~~~~
925 // Note that the collector can't reclaim the objectMonitor or deflate
926 // the object out from underneath the thread calling ::exit() as the
927 // thread calling ::exit() never transitions to a stable state.
928 // This inhibits GC, which in turn inhibits asynchronous (and
929 // inopportune) reclamation of "this".
930 //
931 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ;
932 // There's one exception to the claim above, however. EnterI() can call
933 // exit() to drop a lock if the acquirer has been externally suspended.
934 // In that case exit() is called with _thread_state == _thread_blocked,
935 // but the monitor's _contentions field is > 0, which inhibits reclamation.
936 //
937 // 1-0 exit
938 // ~~~~~~~~
939 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of
940 // the fast-path operators have been optimized so the common ::exit()
941 // operation is 1-0, e.g., see macroAssembler_x86.cpp: fast_unlock().
942 // The code emitted by fast_unlock() elides the usual MEMBAR. This
943 // greatly improves latency -- MEMBAR and CAS having considerable local
944 // latency on modern processors -- but at the cost of "stranding". Absent the
945 // MEMBAR, a thread in fast_unlock() can race a thread in the slow
946 // ::enter() path, resulting in the entering thread being stranding
947 // and a progress-liveness failure. Stranding is extremely rare.
948 // We use timers (timed park operations) & periodic polling to detect
949 // and recover from stranding. Potentially stranded threads periodically
950 // wake up and poll the lock. See the usage of the _Responsible variable.
951 //
952 // The CAS() in enter provides for safety and exclusion, while the CAS or
953 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking
954 // eliminates the CAS/MEMBAR from the exit path, but it admits stranding.
955 // We detect and recover from stranding with timers.
956 //
957 // If a thread transiently strands it'll park until (a) another
958 // thread acquires the lock and then drops the lock, at which time the
959 // exiting thread will notice and unpark the stranded thread, or, (b)
960 // the timer expires. If the lock is high traffic then the stranding latency
961 // will be low due to (a). If the lock is low traffic then the odds of
962 // stranding are lower, although the worst-case stranding latency
963 // is longer. Critically, we don't want to put excessive load in the
964 // platform's timer subsystem. We want to minimize both the timer injection
965 // rate (timers created/sec) as well as the number of timers active at
966 // any one time. (more precisely, we want to minimize timer-seconds, which is
967 // the integral of the # of active timers at any instant over time).
968 // Both impinge on OS scalability. Given that, at most one thread parked on
969 // a monitor will use a timer.
970 //
971 // There is also the risk of a futile wake-up. If we drop the lock
972 // another thread can reacquire the lock immediately, and we can
973 // then wake a thread unnecessarily. This is benign, and we've
974 // structured the code so the windows are short and the frequency
975 // of such futile wakups is low.
976
977 void ObjectMonitor::exit(bool not_suspended, TRAPS) {
978 Thread* const Self = THREAD;
979 void* cur = Atomic::load(&_owner);
980 if (THREAD != cur) {
981 if (THREAD->is_lock_owned((address)cur)) {
982 assert(_recursions == 0, "invariant");
983 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*.
984 _recursions = 0;
985 } else {
986 // Apparent unbalanced locking ...
987 // Naively we'd like to throw IllegalMonitorStateException.
988 // As a practical matter we can neither allocate nor throw an
989 // exception as ::exit() can be called from leaf routines.
990 // see x86_32.ad Fast_Unlock() and the I1 and I2 properties.
991 // Upon deeper reflection, however, in a properly run JVM the only
992 // way we should encounter this situation is in the presence of
993 // unbalanced JNI locking. TODO: CheckJNICalls.
994 // See also: CR4414101
995 #ifdef ASSERT
996 LogStreamHandle(Error, monitorinflation) lsh;
997 lsh.print_cr("ERROR: ObjectMonitor::exit(): thread=" INTPTR_FORMAT
998 " is exiting an ObjectMonitor it does not own.", p2i(THREAD));
999 lsh.print_cr("The imbalance is possibly caused by JNI locking.");
1000 print_debug_style_on(&lsh);
1001 #endif
1002 assert(false, "Non-balanced monitor enter/exit!");
1003 return;
1004 }
1005 }
1006
1007 if (_recursions != 0) {
1008 _recursions--; // this is simple recursive enter
1009 return;
1010 }
1011
1012 // Invariant: after setting Responsible=null an thread must execute
1013 // a MEMBAR or other serializing instruction before fetching EntryList|cxq.
1014 _Responsible = NULL;
1015
1016 #if INCLUDE_JFR
1017 // get the owner's thread id for the MonitorEnter event
1018 // if it is enabled and the thread isn't suspended
1019 if (not_suspended && EventJavaMonitorEnter::is_enabled()) {
1020 _previous_owner_tid = JFR_THREAD_ID(Self);
1021 }
1022 #endif
1023
1024 for (;;) {
1025 assert(THREAD == _owner, "invariant");
1026
1027 // Drop the lock.
1028 // release semantics: prior loads and stores from within the critical section
1029 // must not float (reorder) past the following store that drops the lock.
1030 // Uses a storeload to separate release_store(owner) from the
1031 // successor check. The try_set_owner() below uses cmpxchg() so
1032 // we get the fence down there.
1033 release_clear_owner(Self);
1034 OrderAccess::storeload();
1035
1036 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) {
1037 return;
1038 }
1039 // Other threads are blocked trying to acquire the lock.
1040
1041 // Normally the exiting thread is responsible for ensuring succession,
1042 // but if other successors are ready or other entering threads are spinning
1043 // then this thread can simply store NULL into _owner and exit without
1044 // waking a successor. The existence of spinners or ready successors
1045 // guarantees proper succession (liveness). Responsibility passes to the
1046 // ready or running successors. The exiting thread delegates the duty.
1047 // More precisely, if a successor already exists this thread is absolved
1048 // of the responsibility of waking (unparking) one.
1049 //
1050 // The _succ variable is critical to reducing futile wakeup frequency.
1051 // _succ identifies the "heir presumptive" thread that has been made
1052 // ready (unparked) but that has not yet run. We need only one such
1053 // successor thread to guarantee progress.
1054 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf
1055 // section 3.3 "Futile Wakeup Throttling" for details.
1056 //
1057 // Note that spinners in Enter() also set _succ non-null.
1058 // In the current implementation spinners opportunistically set
1059 // _succ so that exiting threads might avoid waking a successor.
1060 // Another less appealing alternative would be for the exiting thread
1061 // to drop the lock and then spin briefly to see if a spinner managed
1062 // to acquire the lock. If so, the exiting thread could exit
1063 // immediately without waking a successor, otherwise the exiting
1064 // thread would need to dequeue and wake a successor.
1065 // (Note that we'd need to make the post-drop spin short, but no
1066 // shorter than the worst-case round-trip cache-line migration time.
1067 // The dropped lock needs to become visible to the spinner, and then
1068 // the acquisition of the lock by the spinner must become visible to
1069 // the exiting thread).
1070
1071 // It appears that an heir-presumptive (successor) must be made ready.
1072 // Only the current lock owner can manipulate the EntryList or
1073 // drain _cxq, so we need to reacquire the lock. If we fail
1074 // to reacquire the lock the responsibility for ensuring succession
1075 // falls to the new owner.
1076 //
1077 if (try_set_owner_from(NULL, Self) != NULL) {
1078 return;
1079 }
1080
1081 guarantee(_owner == THREAD, "invariant");
1082
1083 ObjectWaiter * w = NULL;
1084
1085 w = _EntryList;
1086 if (w != NULL) {
1087 // I'd like to write: guarantee (w->_thread != Self).
1088 // But in practice an exiting thread may find itself on the EntryList.
1089 // Let's say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and
1090 // then calls exit(). Exit release the lock by setting O._owner to NULL.
1091 // Let's say T1 then stalls. T2 acquires O and calls O.notify(). The
1092 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then
1093 // release the lock "O". T2 resumes immediately after the ST of null into
1094 // _owner, above. T2 notices that the EntryList is populated, so it
1095 // reacquires the lock and then finds itself on the EntryList.
1096 // Given all that, we have to tolerate the circumstance where "w" is
1097 // associated with Self.
1098 assert(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1099 ExitEpilog(Self, w);
1100 return;
1101 }
1102
1103 // If we find that both _cxq and EntryList are null then just
1104 // re-run the exit protocol from the top.
1105 w = _cxq;
1106 if (w == NULL) continue;
1107
1108 // Drain _cxq into EntryList - bulk transfer.
1109 // First, detach _cxq.
1110 // The following loop is tantamount to: w = swap(&cxq, NULL)
1111 for (;;) {
1112 assert(w != NULL, "Invariant");
1113 ObjectWaiter * u = Atomic::cmpxchg(&_cxq, w, (ObjectWaiter*)NULL);
1114 if (u == w) break;
1115 w = u;
1116 }
1117
1118 assert(w != NULL, "invariant");
1119 assert(_EntryList == NULL, "invariant");
1120
1121 // Convert the LIFO SLL anchored by _cxq into a DLL.
1122 // The list reorganization step operates in O(LENGTH(w)) time.
1123 // It's critical that this step operate quickly as
1124 // "Self" still holds the outer-lock, restricting parallelism
1125 // and effectively lengthening the critical section.
1126 // Invariant: s chases t chases u.
1127 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so
1128 // we have faster access to the tail.
1129
1130 _EntryList = w;
1131 ObjectWaiter * q = NULL;
1132 ObjectWaiter * p;
1133 for (p = w; p != NULL; p = p->_next) {
1134 guarantee(p->TState == ObjectWaiter::TS_CXQ, "Invariant");
1135 p->TState = ObjectWaiter::TS_ENTER;
1136 p->_prev = q;
1137 q = p;
1138 }
1139
1140 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL
1141 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog().
1142
1143 // See if we can abdicate to a spinner instead of waking a thread.
1144 // A primary goal of the implementation is to reduce the
1145 // context-switch rate.
1146 if (_succ != NULL) continue;
1147
1148 w = _EntryList;
1149 if (w != NULL) {
1150 guarantee(w->TState == ObjectWaiter::TS_ENTER, "invariant");
1151 ExitEpilog(Self, w);
1152 return;
1153 }
1154 }
1155 }
1156
1157 // ExitSuspendEquivalent:
1158 // A faster alternate to handle_special_suspend_equivalent_condition()
1159 //
1160 // handle_special_suspend_equivalent_condition() unconditionally
1161 // acquires the SR_lock. On some platforms uncontended MutexLocker()
1162 // operations have high latency. Note that in ::enter() we call HSSEC
1163 // while holding the monitor, so we effectively lengthen the critical sections.
1164 //
1165 // There are a number of possible solutions:
1166 //
1167 // A. To ameliorate the problem we might also defer state transitions
1168 // to as late as possible -- just prior to parking.
1169 // Given that, we'd call HSSEC after having returned from park(),
1170 // but before attempting to acquire the monitor. This is only a
1171 // partial solution. It avoids calling HSSEC while holding the
1172 // monitor (good), but it still increases successor reacquisition latency --
1173 // the interval between unparking a successor and the time the successor
1174 // resumes and retries the lock. See ReenterI(), which defers state transitions.
1175 // If we use this technique we can also avoid EnterI()-exit() loop
1176 // in ::enter() where we iteratively drop the lock and then attempt
1177 // to reacquire it after suspending.
1178 //
1179 // B. In the future we might fold all the suspend bits into a
1180 // composite per-thread suspend flag and then update it with CAS().
1181 // Alternately, a Dekker-like mechanism with multiple variables
1182 // would suffice:
1183 // ST Self->_suspend_equivalent = false
1184 // MEMBAR
1185 // LD Self_>_suspend_flags
1186
1187 bool ObjectMonitor::ExitSuspendEquivalent(JavaThread * jSelf) {
1188 return jSelf->handle_special_suspend_equivalent_condition();
1189 }
1190
1191
1192 void ObjectMonitor::ExitEpilog(Thread * Self, ObjectWaiter * Wakee) {
1193 assert(_owner == Self, "invariant");
1194
1195 // Exit protocol:
1196 // 1. ST _succ = wakee
1197 // 2. membar #loadstore|#storestore;
1198 // 2. ST _owner = NULL
1199 // 3. unpark(wakee)
1200
1201 _succ = Wakee->_thread;
1202 ParkEvent * Trigger = Wakee->_event;
1203
1204 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again.
1205 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be
1206 // out-of-scope (non-extant).
1207 Wakee = NULL;
1208
1209 // Drop the lock.
1210 // Uses a fence to separate release_store(owner) from the LD in unpark().
1211 release_clear_owner(Self);
1212 OrderAccess::fence();
1213
1214 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self);
1215 Trigger->unpark();
1216
1217 // Maintain stats and report events to JVMTI
1218 OM_PERFDATA_OP(Parks, inc());
1219 }
1220
1221
1222 // -----------------------------------------------------------------------------
1223 // Class Loader deadlock handling.
1224 //
1225 // complete_exit exits a lock returning recursion count
1226 // complete_exit/reenter operate as a wait without waiting
1227 // complete_exit requires an inflated monitor
1228 // The _owner field is not always the Thread addr even with an
1229 // inflated monitor, e.g. the monitor can be inflated by a non-owning
1230 // thread due to contention.
1231 intx ObjectMonitor::complete_exit(TRAPS) {
1232 Thread * const Self = THREAD;
1233 assert(Self->is_Java_thread(), "Must be Java thread!");
1234 JavaThread *jt = (JavaThread *)THREAD;
1235
1236 assert(InitDone, "Unexpectedly not initialized");
1237
1238 void* cur = Atomic::load(&_owner);
1239 if (THREAD != cur) {
1240 if (THREAD->is_lock_owned((address)cur)) {
1241 assert(_recursions == 0, "internal state error");
1242 set_owner_from_BasicLock(cur, Self); // Convert from BasicLock* to Thread*.
1243 _recursions = 0;
1244 }
1245 }
1246
1247 guarantee(Self == _owner, "complete_exit not owner");
1248 intx save = _recursions; // record the old recursion count
1249 _recursions = 0; // set the recursion level to be 0
1250 exit(true, Self); // exit the monitor
1251 guarantee(_owner != Self, "invariant");
1252 return save;
1253 }
1254
1255 // reenter() enters a lock and sets recursion count
1256 // complete_exit/reenter operate as a wait without waiting
1257 bool ObjectMonitor::reenter(intx recursions, TRAPS) {
1258 Thread * const Self = THREAD;
1259 assert(Self->is_Java_thread(), "Must be Java thread!");
1260 JavaThread *jt = (JavaThread *)THREAD;
1261
1262 guarantee(_owner != Self, "reenter already owner");
1263 if (!enter(THREAD)) {
1264 return false;
1265 }
1266 // Entered the monitor.
1267 guarantee(_recursions == 0, "reenter recursion");
1268 _recursions = recursions;
1269 return true;
1270 }
1271
1272 // Checks that the current THREAD owns this monitor and causes an
1273 // immediate return if it doesn't. We don't use the CHECK macro
1274 // because we want the IMSE to be the only exception that is thrown
1275 // from the call site when false is returned. Any other pending
1276 // exception is ignored.
1277 #define CHECK_OWNER() \
1278 do { \
1279 if (!check_owner(THREAD)) { \
1280 assert(HAS_PENDING_EXCEPTION, "expected a pending IMSE here."); \
1281 return; \
1282 } \
1283 } while (false)
1284
1285 // Returns true if the specified thread owns the ObjectMonitor.
1286 // Otherwise returns false and throws IllegalMonitorStateException
1287 // (IMSE). If there is a pending exception and the specified thread
1288 // is not the owner, that exception will be replaced by the IMSE.
1289 bool ObjectMonitor::check_owner(Thread* THREAD) {
1290 void* cur = Atomic::load(&_owner);
1291 if (cur == THREAD) {
1292 return true;
1293 }
1294 if (THREAD->is_lock_owned((address)cur)) {
1295 set_owner_from_BasicLock(cur, THREAD); // Convert from BasicLock* to Thread*.
1296 _recursions = 0;
1297 return true;
1298 }
1299 THROW_MSG_(vmSymbols::java_lang_IllegalMonitorStateException(),
1300 "current thread is not owner", false);
1301 }
1302
1303 static void post_monitor_wait_event(EventJavaMonitorWait* event,
1304 ObjectMonitor* monitor,
1305 jlong notifier_tid,
1306 jlong timeout,
1307 bool timedout) {
1308 assert(event != NULL, "invariant");
1309 assert(monitor != NULL, "invariant");
1310 event->set_monitorClass(((oop)monitor->object())->klass());
1311 event->set_timeout(timeout);
1312 event->set_address((uintptr_t)monitor->object_addr());
1313 event->set_notifier(notifier_tid);
1314 event->set_timedOut(timedout);
1315 event->commit();
1316 }
1317
1318 // -----------------------------------------------------------------------------
1319 // Wait/Notify/NotifyAll
1320 //
1321 // Note: a subset of changes to ObjectMonitor::wait()
1322 // will need to be replicated in complete_exit
1323 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) {
1324 Thread * const Self = THREAD;
1325 assert(Self->is_Java_thread(), "Must be Java thread!");
1326 JavaThread *jt = (JavaThread *)THREAD;
1327
1328 assert(InitDone, "Unexpectedly not initialized");
1329
1330 CHECK_OWNER(); // Throws IMSE if not owner.
1331
1332 EventJavaMonitorWait event;
1333
1334 // check for a pending interrupt
1335 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1336 // post monitor waited event. Note that this is past-tense, we are done waiting.
1337 if (JvmtiExport::should_post_monitor_waited()) {
1338 // Note: 'false' parameter is passed here because the
1339 // wait was not timed out due to thread interrupt.
1340 JvmtiExport::post_monitor_waited(jt, this, false);
1341
1342 // In this short circuit of the monitor wait protocol, the
1343 // current thread never drops ownership of the monitor and
1344 // never gets added to the wait queue so the current thread
1345 // cannot be made the successor. This means that the
1346 // JVMTI_EVENT_MONITOR_WAITED event handler cannot accidentally
1347 // consume an unpark() meant for the ParkEvent associated with
1348 // this ObjectMonitor.
1349 }
1350 if (event.should_commit()) {
1351 post_monitor_wait_event(&event, this, 0, millis, false);
1352 }
1353 THROW(vmSymbols::java_lang_InterruptedException());
1354 return;
1355 }
1356
1357 assert(Self->_Stalled == 0, "invariant");
1358 Self->_Stalled = intptr_t(this);
1359 jt->set_current_waiting_monitor(this);
1360
1361 // create a node to be put into the queue
1362 // Critically, after we reset() the event but prior to park(), we must check
1363 // for a pending interrupt.
1364 ObjectWaiter node(Self);
1365 node.TState = ObjectWaiter::TS_WAIT;
1366 Self->_ParkEvent->reset();
1367 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag
1368
1369 // Enter the waiting queue, which is a circular doubly linked list in this case
1370 // but it could be a priority queue or any data structure.
1371 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only
1372 // by the the owner of the monitor *except* in the case where park()
1373 // returns because of a timeout of interrupt. Contention is exceptionally rare
1374 // so we use a simple spin-lock instead of a heavier-weight blocking lock.
1375
1376 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - add");
1377 AddWaiter(&node);
1378 Thread::SpinRelease(&_WaitSetLock);
1379
1380 _Responsible = NULL;
1381
1382 intx save = _recursions; // record the old recursion count
1383 _waiters++; // increment the number of waiters
1384 _recursions = 0; // set the recursion level to be 1
1385 exit(true, Self); // exit the monitor
1386 guarantee(_owner != Self, "invariant");
1387
1388 // The thread is on the WaitSet list - now park() it.
1389 // On MP systems it's conceivable that a brief spin before we park
1390 // could be profitable.
1391 //
1392 // TODO-FIXME: change the following logic to a loop of the form
1393 // while (!timeout && !interrupted && _notified == 0) park()
1394
1395 int ret = OS_OK;
1396 int WasNotified = 0;
1397
1398 // Need to check interrupt state whilst still _thread_in_vm
1399 bool interrupted = interruptible && jt->is_interrupted(false);
1400
1401 { // State transition wrappers
1402 OSThread* osthread = Self->osthread();
1403 OSThreadWaitState osts(osthread, true);
1404 {
1405 ThreadBlockInVM tbivm(jt);
1406 // Thread is in thread_blocked state and oop access is unsafe.
1407 jt->set_suspend_equivalent();
1408
1409 if (interrupted || HAS_PENDING_EXCEPTION) {
1410 // Intentionally empty
1411 } else if (node._notified == 0) {
1412 if (millis <= 0) {
1413 Self->_ParkEvent->park();
1414 } else {
1415 ret = Self->_ParkEvent->park(millis);
1416 }
1417 }
1418
1419 // were we externally suspended while we were waiting?
1420 if (ExitSuspendEquivalent (jt)) {
1421 // TODO-FIXME: add -- if succ == Self then succ = null.
1422 jt->java_suspend_self();
1423 }
1424
1425 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm
1426
1427 // Node may be on the WaitSet, the EntryList (or cxq), or in transition
1428 // from the WaitSet to the EntryList.
1429 // See if we need to remove Node from the WaitSet.
1430 // We use double-checked locking to avoid grabbing _WaitSetLock
1431 // if the thread is not on the wait queue.
1432 //
1433 // Note that we don't need a fence before the fetch of TState.
1434 // In the worst case we'll fetch a old-stale value of TS_WAIT previously
1435 // written by the is thread. (perhaps the fetch might even be satisfied
1436 // by a look-aside into the processor's own store buffer, although given
1437 // the length of the code path between the prior ST and this load that's
1438 // highly unlikely). If the following LD fetches a stale TS_WAIT value
1439 // then we'll acquire the lock and then re-fetch a fresh TState value.
1440 // That is, we fail toward safety.
1441
1442 if (node.TState == ObjectWaiter::TS_WAIT) {
1443 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - unlink");
1444 if (node.TState == ObjectWaiter::TS_WAIT) {
1445 DequeueSpecificWaiter(&node); // unlink from WaitSet
1446 assert(node._notified == 0, "invariant");
1447 node.TState = ObjectWaiter::TS_RUN;
1448 }
1449 Thread::SpinRelease(&_WaitSetLock);
1450 }
1451
1452 // The thread is now either on off-list (TS_RUN),
1453 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ).
1454 // The Node's TState variable is stable from the perspective of this thread.
1455 // No other threads will asynchronously modify TState.
1456 guarantee(node.TState != ObjectWaiter::TS_WAIT, "invariant");
1457 OrderAccess::loadload();
1458 if (_succ == Self) _succ = NULL;
1459 WasNotified = node._notified;
1460
1461 // Reentry phase -- reacquire the monitor.
1462 // re-enter contended monitor after object.wait().
1463 // retain OBJECT_WAIT state until re-enter successfully completes
1464 // Thread state is thread_in_vm and oop access is again safe,
1465 // although the raw address of the object may have changed.
1466 // (Don't cache naked oops over safepoints, of course).
1467
1468 // post monitor waited event. Note that this is past-tense, we are done waiting.
1469 if (JvmtiExport::should_post_monitor_waited()) {
1470 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT);
1471
1472 if (node._notified != 0 && _succ == Self) {
1473 // In this part of the monitor wait-notify-reenter protocol it
1474 // is possible (and normal) for another thread to do a fastpath
1475 // monitor enter-exit while this thread is still trying to get
1476 // to the reenter portion of the protocol.
1477 //
1478 // The ObjectMonitor was notified and the current thread is
1479 // the successor which also means that an unpark() has already
1480 // been done. The JVMTI_EVENT_MONITOR_WAITED event handler can
1481 // consume the unpark() that was done when the successor was
1482 // set because the same ParkEvent is shared between Java
1483 // monitors and JVM/TI RawMonitors (for now).
1484 //
1485 // We redo the unpark() to ensure forward progress, i.e., we
1486 // don't want all pending threads hanging (parked) with none
1487 // entering the unlocked monitor.
1488 node._event->unpark();
1489 }
1490 }
1491
1492 if (event.should_commit()) {
1493 post_monitor_wait_event(&event, this, node._notifier_tid, millis, ret == OS_TIMEOUT);
1494 }
1495
1496 OrderAccess::fence();
1497
1498 assert(Self->_Stalled != 0, "invariant");
1499 Self->_Stalled = 0;
1500
1501 assert(_owner != Self, "invariant");
1502 ObjectWaiter::TStates v = node.TState;
1503 if (v == ObjectWaiter::TS_RUN) {
1504 enter(Self);
1505 } else {
1506 guarantee(v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant");
1507 ReenterI(Self, &node);
1508 node.wait_reenter_end(this);
1509 }
1510
1511 // Self has reacquired the lock.
1512 // Lifecycle - the node representing Self must not appear on any queues.
1513 // Node is about to go out-of-scope, but even if it were immortal we wouldn't
1514 // want residual elements associated with this thread left on any lists.
1515 guarantee(node.TState == ObjectWaiter::TS_RUN, "invariant");
1516 assert(_owner == Self, "invariant");
1517 assert(_succ != Self, "invariant");
1518 } // OSThreadWaitState()
1519
1520 jt->set_current_waiting_monitor(NULL);
1521
1522 guarantee(_recursions == 0, "invariant");
1523 _recursions = save; // restore the old recursion count
1524 _waiters--; // decrement the number of waiters
1525
1526 // Verify a few postconditions
1527 assert(_owner == Self, "invariant");
1528 assert(_succ != Self, "invariant");
1529 assert(((oop)(object()))->mark() == markWord::encode(this), "invariant");
1530
1531 // check if the notification happened
1532 if (!WasNotified) {
1533 // no, it could be timeout or Thread.interrupt() or both
1534 // check for interrupt event, otherwise it is timeout
1535 if (interruptible && jt->is_interrupted(true) && !HAS_PENDING_EXCEPTION) {
1536 THROW(vmSymbols::java_lang_InterruptedException());
1537 }
1538 }
1539
1540 // NOTE: Spurious wake up will be consider as timeout.
1541 // Monitor notify has precedence over thread interrupt.
1542 }
1543
1544
1545 // Consider:
1546 // If the lock is cool (cxq == null && succ == null) and we're on an MP system
1547 // then instead of transferring a thread from the WaitSet to the EntryList
1548 // we might just dequeue a thread from the WaitSet and directly unpark() it.
1549
1550 void ObjectMonitor::INotify(Thread * Self) {
1551 Thread::SpinAcquire(&_WaitSetLock, "WaitSet - notify");
1552 ObjectWaiter * iterator = DequeueWaiter();
1553 if (iterator != NULL) {
1554 guarantee(iterator->TState == ObjectWaiter::TS_WAIT, "invariant");
1555 guarantee(iterator->_notified == 0, "invariant");
1556 // Disposition - what might we do with iterator ?
1557 // a. add it directly to the EntryList - either tail (policy == 1)
1558 // or head (policy == 0).
1559 // b. push it onto the front of the _cxq (policy == 2).
1560 // For now we use (b).
1561
1562 iterator->TState = ObjectWaiter::TS_ENTER;
1563
1564 iterator->_notified = 1;
1565 iterator->_notifier_tid = JFR_THREAD_ID(Self);
1566
1567 ObjectWaiter * list = _EntryList;
1568 if (list != NULL) {
1569 assert(list->_prev == NULL, "invariant");
1570 assert(list->TState == ObjectWaiter::TS_ENTER, "invariant");
1571 assert(list != iterator, "invariant");
1572 }
1573
1574 // prepend to cxq
1575 if (list == NULL) {
1576 iterator->_next = iterator->_prev = NULL;
1577 _EntryList = iterator;
1578 } else {
1579 iterator->TState = ObjectWaiter::TS_CXQ;
1580 for (;;) {
1581 ObjectWaiter * front = _cxq;
1582 iterator->_next = front;
1583 if (Atomic::cmpxchg(&_cxq, front, iterator) == front) {
1584 break;
1585 }
1586 }
1587 }
1588
1589 // _WaitSetLock protects the wait queue, not the EntryList. We could
1590 // move the add-to-EntryList operation, above, outside the critical section
1591 // protected by _WaitSetLock. In practice that's not useful. With the
1592 // exception of wait() timeouts and interrupts the monitor owner
1593 // is the only thread that grabs _WaitSetLock. There's almost no contention
1594 // on _WaitSetLock so it's not profitable to reduce the length of the
1595 // critical section.
1596
1597 iterator->wait_reenter_begin(this);
1598 }
1599 Thread::SpinRelease(&_WaitSetLock);
1600 }
1601
1602 // Consider: a not-uncommon synchronization bug is to use notify() when
1603 // notifyAll() is more appropriate, potentially resulting in stranded
1604 // threads; this is one example of a lost wakeup. A useful diagnostic
1605 // option is to force all notify() operations to behave as notifyAll().
1606 //
1607 // Note: We can also detect many such problems with a "minimum wait".
1608 // When the "minimum wait" is set to a small non-zero timeout value
1609 // and the program does not hang whereas it did absent "minimum wait",
1610 // that suggests a lost wakeup bug.
1611
1612 void ObjectMonitor::notify(TRAPS) {
1613 CHECK_OWNER(); // Throws IMSE if not owner.
1614 if (_WaitSet == NULL) {
1615 return;
1616 }
1617 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD);
1618 INotify(THREAD);
1619 OM_PERFDATA_OP(Notifications, inc(1));
1620 }
1621
1622
1623 // The current implementation of notifyAll() transfers the waiters one-at-a-time
1624 // from the waitset to the EntryList. This could be done more efficiently with a
1625 // single bulk transfer but in practice it's not time-critical. Beware too,
1626 // that in prepend-mode we invert the order of the waiters. Let's say that the
1627 // waitset is "ABCD" and the EntryList is "XYZ". After a notifyAll() in prepend
1628 // mode the waitset will be empty and the EntryList will be "DCBAXYZ".
1629
1630 void ObjectMonitor::notifyAll(TRAPS) {
1631 CHECK_OWNER(); // Throws IMSE if not owner.
1632 if (_WaitSet == NULL) {
1633 return;
1634 }
1635
1636 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD);
1637 int tally = 0;
1638 while (_WaitSet != NULL) {
1639 tally++;
1640 INotify(THREAD);
1641 }
1642
1643 OM_PERFDATA_OP(Notifications, inc(tally));
1644 }
1645
1646 // -----------------------------------------------------------------------------
1647 // Adaptive Spinning Support
1648 //
1649 // Adaptive spin-then-block - rational spinning
1650 //
1651 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS
1652 // algorithm. On high order SMP systems it would be better to start with
1653 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH,
1654 // a contending thread could enqueue itself on the cxq and then spin locally
1655 // on a thread-specific variable such as its ParkEvent._Event flag.
1656 // That's left as an exercise for the reader. Note that global spinning is
1657 // not problematic on Niagara, as the L2 cache serves the interconnect and
1658 // has both low latency and massive bandwidth.
1659 //
1660 // Broadly, we can fix the spin frequency -- that is, the % of contended lock
1661 // acquisition attempts where we opt to spin -- at 100% and vary the spin count
1662 // (duration) or we can fix the count at approximately the duration of
1663 // a context switch and vary the frequency. Of course we could also
1664 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor.
1665 // For a description of 'Adaptive spin-then-block mutual exclusion in
1666 // multi-threaded processing,' see U.S. Pat. No. 8046758.
1667 //
1668 // This implementation varies the duration "D", where D varies with
1669 // the success rate of recent spin attempts. (D is capped at approximately
1670 // length of a round-trip context switch). The success rate for recent
1671 // spin attempts is a good predictor of the success rate of future spin
1672 // attempts. The mechanism adapts automatically to varying critical
1673 // section length (lock modality), system load and degree of parallelism.
1674 // D is maintained per-monitor in _SpinDuration and is initialized
1675 // optimistically. Spin frequency is fixed at 100%.
1676 //
1677 // Note that _SpinDuration is volatile, but we update it without locks
1678 // or atomics. The code is designed so that _SpinDuration stays within
1679 // a reasonable range even in the presence of races. The arithmetic
1680 // operations on _SpinDuration are closed over the domain of legal values,
1681 // so at worst a race will install and older but still legal value.
1682 // At the very worst this introduces some apparent non-determinism.
1683 // We might spin when we shouldn't or vice-versa, but since the spin
1684 // count are relatively short, even in the worst case, the effect is harmless.
1685 //
1686 // Care must be taken that a low "D" value does not become an
1687 // an absorbing state. Transient spinning failures -- when spinning
1688 // is overall profitable -- should not cause the system to converge
1689 // on low "D" values. We want spinning to be stable and predictable
1690 // and fairly responsive to change and at the same time we don't want
1691 // it to oscillate, become metastable, be "too" non-deterministic,
1692 // or converge on or enter undesirable stable absorbing states.
1693 //
1694 // We implement a feedback-based control system -- using past behavior
1695 // to predict future behavior. We face two issues: (a) if the
1696 // input signal is random then the spin predictor won't provide optimal
1697 // results, and (b) if the signal frequency is too high then the control
1698 // system, which has some natural response lag, will "chase" the signal.
1699 // (b) can arise from multimodal lock hold times. Transient preemption
1700 // can also result in apparent bimodal lock hold times.
1701 // Although sub-optimal, neither condition is particularly harmful, as
1702 // in the worst-case we'll spin when we shouldn't or vice-versa.
1703 // The maximum spin duration is rather short so the failure modes aren't bad.
1704 // To be conservative, I've tuned the gain in system to bias toward
1705 // _not spinning. Relatedly, the system can sometimes enter a mode where it
1706 // "rings" or oscillates between spinning and not spinning. This happens
1707 // when spinning is just on the cusp of profitability, however, so the
1708 // situation is not dire. The state is benign -- there's no need to add
1709 // hysteresis control to damp the transition rate between spinning and
1710 // not spinning.
1711
1712 // Spinning: Fixed frequency (100%), vary duration
1713 int ObjectMonitor::TrySpin(Thread * Self) {
1714 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning.
1715 int ctr = Knob_FixedSpin;
1716 if (ctr != 0) {
1717 while (--ctr >= 0) {
1718 if (TryLock(Self) > 0) return 1;
1719 SpinPause();
1720 }
1721 return 0;
1722 }
1723
1724 for (ctr = Knob_PreSpin + 1; --ctr >= 0;) {
1725 if (TryLock(Self) > 0) {
1726 // Increase _SpinDuration ...
1727 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1728 // Raising _SpurDuration to the poverty line is key.
1729 int x = _SpinDuration;
1730 if (x < Knob_SpinLimit) {
1731 if (x < Knob_Poverty) x = Knob_Poverty;
1732 _SpinDuration = x + Knob_BonusB;
1733 }
1734 return 1;
1735 }
1736 SpinPause();
1737 }
1738
1739 // Admission control - verify preconditions for spinning
1740 //
1741 // We always spin a little bit, just to prevent _SpinDuration == 0 from
1742 // becoming an absorbing state. Put another way, we spin briefly to
1743 // sample, just in case the system load, parallelism, contention, or lock
1744 // modality changed.
1745 //
1746 // Consider the following alternative:
1747 // Periodically set _SpinDuration = _SpinLimit and try a long/full
1748 // spin attempt. "Periodically" might mean after a tally of
1749 // the # of failed spin attempts (or iterations) reaches some threshold.
1750 // This takes us into the realm of 1-out-of-N spinning, where we
1751 // hold the duration constant but vary the frequency.
1752
1753 ctr = _SpinDuration;
1754 if (ctr <= 0) return 0;
1755
1756 if (NotRunnable(Self, (Thread *) _owner)) {
1757 return 0;
1758 }
1759
1760 // We're good to spin ... spin ingress.
1761 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades
1762 // when preparing to LD...CAS _owner, etc and the CAS is likely
1763 // to succeed.
1764 if (_succ == NULL) {
1765 _succ = Self;
1766 }
1767 Thread * prv = NULL;
1768
1769 // There are three ways to exit the following loop:
1770 // 1. A successful spin where this thread has acquired the lock.
1771 // 2. Spin failure with prejudice
1772 // 3. Spin failure without prejudice
1773
1774 while (--ctr >= 0) {
1775
1776 // Periodic polling -- Check for pending GC
1777 // Threads may spin while they're unsafe.
1778 // We don't want spinning threads to delay the JVM from reaching
1779 // a stop-the-world safepoint or to steal cycles from GC.
1780 // If we detect a pending safepoint we abort in order that
1781 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b)
1782 // this thread, if safe, doesn't steal cycles from GC.
1783 // This is in keeping with the "no loitering in runtime" rule.
1784 // We periodically check to see if there's a safepoint pending.
1785 if ((ctr & 0xFF) == 0) {
1786 if (SafepointMechanism::should_block(Self)) {
1787 goto Abort; // abrupt spin egress
1788 }
1789 SpinPause();
1790 }
1791
1792 // Probe _owner with TATAS
1793 // If this thread observes the monitor transition or flicker
1794 // from locked to unlocked to locked, then the odds that this
1795 // thread will acquire the lock in this spin attempt go down
1796 // considerably. The same argument applies if the CAS fails
1797 // or if we observe _owner change from one non-null value to
1798 // another non-null value. In such cases we might abort
1799 // the spin without prejudice or apply a "penalty" to the
1800 // spin count-down variable "ctr", reducing it by 100, say.
1801
1802 Thread * ox = (Thread *) _owner;
1803 if (ox == NULL) {
1804 ox = (Thread*)try_set_owner_from(NULL, Self);
1805 if (ox == NULL) {
1806 // The CAS succeeded -- this thread acquired ownership
1807 // Take care of some bookkeeping to exit spin state.
1808 if (_succ == Self) {
1809 _succ = NULL;
1810 }
1811
1812 // Increase _SpinDuration :
1813 // The spin was successful (profitable) so we tend toward
1814 // longer spin attempts in the future.
1815 // CONSIDER: factor "ctr" into the _SpinDuration adjustment.
1816 // If we acquired the lock early in the spin cycle it
1817 // makes sense to increase _SpinDuration proportionally.
1818 // Note that we don't clamp SpinDuration precisely at SpinLimit.
1819 int x = _SpinDuration;
1820 if (x < Knob_SpinLimit) {
1821 if (x < Knob_Poverty) x = Knob_Poverty;
1822 _SpinDuration = x + Knob_Bonus;
1823 }
1824 return 1;
1825 }
1826
1827 // The CAS failed ... we can take any of the following actions:
1828 // * penalize: ctr -= CASPenalty
1829 // * exit spin with prejudice -- goto Abort;
1830 // * exit spin without prejudice.
1831 // * Since CAS is high-latency, retry again immediately.
1832 prv = ox;
1833 goto Abort;
1834 }
1835
1836 // Did lock ownership change hands ?
1837 if (ox != prv && prv != NULL) {
1838 goto Abort;
1839 }
1840 prv = ox;
1841
1842 // Abort the spin if the owner is not executing.
1843 // The owner must be executing in order to drop the lock.
1844 // Spinning while the owner is OFFPROC is idiocy.
1845 // Consider: ctr -= RunnablePenalty ;
1846 if (NotRunnable(Self, ox)) {
1847 goto Abort;
1848 }
1849 if (_succ == NULL) {
1850 _succ = Self;
1851 }
1852 }
1853
1854 // Spin failed with prejudice -- reduce _SpinDuration.
1855 // TODO: Use an AIMD-like policy to adjust _SpinDuration.
1856 // AIMD is globally stable.
1857 {
1858 int x = _SpinDuration;
1859 if (x > 0) {
1860 // Consider an AIMD scheme like: x -= (x >> 3) + 100
1861 // This is globally sample and tends to damp the response.
1862 x -= Knob_Penalty;
1863 if (x < 0) x = 0;
1864 _SpinDuration = x;
1865 }
1866 }
1867
1868 Abort:
1869 if (_succ == Self) {
1870 _succ = NULL;
1871 // Invariant: after setting succ=null a contending thread
1872 // must recheck-retry _owner before parking. This usually happens
1873 // in the normal usage of TrySpin(), but it's safest
1874 // to make TrySpin() as foolproof as possible.
1875 OrderAccess::fence();
1876 if (TryLock(Self) > 0) return 1;
1877 }
1878 return 0;
1879 }
1880
1881 // NotRunnable() -- informed spinning
1882 //
1883 // Don't bother spinning if the owner is not eligible to drop the lock.
1884 // Spin only if the owner thread is _thread_in_Java or _thread_in_vm.
1885 // The thread must be runnable in order to drop the lock in timely fashion.
1886 // If the _owner is not runnable then spinning will not likely be
1887 // successful (profitable).
1888 //
1889 // Beware -- the thread referenced by _owner could have died
1890 // so a simply fetch from _owner->_thread_state might trap.
1891 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state.
1892 // Because of the lifecycle issues, the _thread_state values
1893 // observed by NotRunnable() might be garbage. NotRunnable must
1894 // tolerate this and consider the observed _thread_state value
1895 // as advisory.
1896 //
1897 // Beware too, that _owner is sometimes a BasicLock address and sometimes
1898 // a thread pointer.
1899 // Alternately, we might tag the type (thread pointer vs basiclock pointer)
1900 // with the LSB of _owner. Another option would be to probabilistically probe
1901 // the putative _owner->TypeTag value.
1902 //
1903 // Checking _thread_state isn't perfect. Even if the thread is
1904 // in_java it might be blocked on a page-fault or have been preempted
1905 // and sitting on a ready/dispatch queue.
1906 //
1907 // The return value from NotRunnable() is *advisory* -- the
1908 // result is based on sampling and is not necessarily coherent.
1909 // The caller must tolerate false-negative and false-positive errors.
1910 // Spinning, in general, is probabilistic anyway.
1911
1912
1913 int ObjectMonitor::NotRunnable(Thread * Self, Thread * ox) {
1914 // Check ox->TypeTag == 2BAD.
1915 if (ox == NULL) return 0;
1916
1917 // Avoid transitive spinning ...
1918 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L.
1919 // Immediately after T1 acquires L it's possible that T2, also
1920 // spinning on L, will see L.Owner=T1 and T1._Stalled=L.
1921 // This occurs transiently after T1 acquired L but before
1922 // T1 managed to clear T1.Stalled. T2 does not need to abort
1923 // its spin in this circumstance.
1924 intptr_t BlockedOn = SafeFetchN((intptr_t *) &ox->_Stalled, intptr_t(1));
1925
1926 if (BlockedOn == 1) return 1;
1927 if (BlockedOn != 0) {
1928 return BlockedOn != intptr_t(this) && _owner == ox;
1929 }
1930
1931 assert(sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant");
1932 int jst = SafeFetch32((int *) &((JavaThread *) ox)->_thread_state, -1);;
1933 // consider also: jst != _thread_in_Java -- but that's overspecific.
1934 return jst == _thread_blocked || jst == _thread_in_native;
1935 }
1936
1937
1938 // -----------------------------------------------------------------------------
1939 // WaitSet management ...
1940
1941 ObjectWaiter::ObjectWaiter(Thread* thread) {
1942 _next = NULL;
1943 _prev = NULL;
1944 _notified = 0;
1945 _notifier_tid = 0;
1946 TState = TS_RUN;
1947 _thread = thread;
1948 _event = thread->_ParkEvent;
1949 _active = false;
1950 assert(_event != NULL, "invariant");
1951 }
1952
1953 void ObjectWaiter::wait_reenter_begin(ObjectMonitor * const mon) {
1954 JavaThread *jt = (JavaThread *)this->_thread;
1955 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon);
1956 }
1957
1958 void ObjectWaiter::wait_reenter_end(ObjectMonitor * const mon) {
1959 JavaThread *jt = (JavaThread *)this->_thread;
1960 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active);
1961 }
1962
1963 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) {
1964 assert(node != NULL, "should not add NULL node");
1965 assert(node->_prev == NULL, "node already in list");
1966 assert(node->_next == NULL, "node already in list");
1967 // put node at end of queue (circular doubly linked list)
1968 if (_WaitSet == NULL) {
1969 _WaitSet = node;
1970 node->_prev = node;
1971 node->_next = node;
1972 } else {
1973 ObjectWaiter* head = _WaitSet;
1974 ObjectWaiter* tail = head->_prev;
1975 assert(tail->_next == head, "invariant check");
1976 tail->_next = node;
1977 head->_prev = node;
1978 node->_next = head;
1979 node->_prev = tail;
1980 }
1981 }
1982
1983 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() {
1984 // dequeue the very first waiter
1985 ObjectWaiter* waiter = _WaitSet;
1986 if (waiter) {
1987 DequeueSpecificWaiter(waiter);
1988 }
1989 return waiter;
1990 }
1991
1992 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) {
1993 assert(node != NULL, "should not dequeue NULL node");
1994 assert(node->_prev != NULL, "node already removed from list");
1995 assert(node->_next != NULL, "node already removed from list");
1996 // when the waiter has woken up because of interrupt,
1997 // timeout or other spurious wake-up, dequeue the
1998 // waiter from waiting list
1999 ObjectWaiter* next = node->_next;
2000 if (next == node) {
2001 assert(node->_prev == node, "invariant check");
2002 _WaitSet = NULL;
2003 } else {
2004 ObjectWaiter* prev = node->_prev;
2005 assert(prev->_next == node, "invariant check");
2006 assert(next->_prev == node, "invariant check");
2007 next->_prev = prev;
2008 prev->_next = next;
2009 if (_WaitSet == node) {
2010 _WaitSet = next;
2011 }
2012 }
2013 node->_next = NULL;
2014 node->_prev = NULL;
2015 }
2016
2017 // -----------------------------------------------------------------------------
2018 // PerfData support
2019 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL;
2020 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL;
2021 PerfCounter * ObjectMonitor::_sync_Parks = NULL;
2022 PerfCounter * ObjectMonitor::_sync_Notifications = NULL;
2023 PerfCounter * ObjectMonitor::_sync_Inflations = NULL;
2024 PerfCounter * ObjectMonitor::_sync_Deflations = NULL;
2025 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL;
2026
2027 // One-shot global initialization for the sync subsystem.
2028 // We could also defer initialization and initialize on-demand
2029 // the first time we call ObjectSynchronizer::inflate().
2030 // Initialization would be protected - like so many things - by
2031 // the MonitorCache_lock.
2032
2033 void ObjectMonitor::Initialize() {
2034 assert(!InitDone, "invariant");
2035
2036 if (!os::is_MP()) {
2037 Knob_SpinLimit = 0;
2038 Knob_PreSpin = 0;
2039 Knob_FixedSpin = -1;
2040 }
2041
2042 if (UsePerfData) {
2043 EXCEPTION_MARK;
2044 #define NEWPERFCOUNTER(n) \
2045 { \
2046 n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events, \
2047 CHECK); \
2048 }
2049 #define NEWPERFVARIABLE(n) \
2050 { \
2051 n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events, \
2052 CHECK); \
2053 }
2054 NEWPERFCOUNTER(_sync_Inflations);
2055 NEWPERFCOUNTER(_sync_Deflations);
2056 NEWPERFCOUNTER(_sync_ContendedLockAttempts);
2057 NEWPERFCOUNTER(_sync_FutileWakeups);
2058 NEWPERFCOUNTER(_sync_Parks);
2059 NEWPERFCOUNTER(_sync_Notifications);
2060 NEWPERFVARIABLE(_sync_MonExtant);
2061 #undef NEWPERFCOUNTER
2062 #undef NEWPERFVARIABLE
2063 }
2064
2065 DEBUG_ONLY(InitDone = true;)
2066 }
2067
2068 void ObjectMonitor::print_on(outputStream* st) const {
2069 // The minimal things to print for markWord printing, more can be added for debugging and logging.
2070 st->print("{contentions=0x%08x,waiters=0x%08x"
2071 ",recursions=" INTX_FORMAT ",owner=" INTPTR_FORMAT "}",
2072 contentions(), waiters(), recursions(),
2073 p2i(owner()));
2074 }
2075 void ObjectMonitor::print() const { print_on(tty); }
2076
2077 #ifdef ASSERT
2078 // Print the ObjectMonitor like a debugger would:
2079 //
2080 // (ObjectMonitor) 0x00007fdfb6012e40 = {
2081 // _header = 0x0000000000000001
2082 // _object = 0x000000070ff45fd0
2083 // _allocation_state = Old
2084 // _pad_buf0 = {
2085 // [0] = '\0'
2086 // ...
2087 // [43] = '\0'
2088 // }
2089 // _owner = 0x0000000000000000
2090 // _previous_owner_tid = 0
2091 // _pad_buf1 = {
2092 // [0] = '\0'
2093 // ...
2094 // [47] = '\0'
2095 // }
2096 // _next_om = 0x0000000000000000
2097 // _recursions = 0
2098 // _EntryList = 0x0000000000000000
2099 // _cxq = 0x0000000000000000
2100 // _succ = 0x0000000000000000
2101 // _Responsible = 0x0000000000000000
2102 // _Spinner = 0
2103 // _SpinDuration = 5000
2104 // _contentions = 0
2105 // _WaitSet = 0x0000700009756248
2106 // _waiters = 1
2107 // _WaitSetLock = 0
2108 // }
2109 //
2110 void ObjectMonitor::print_debug_style_on(outputStream* st) const {
2111 st->print_cr("(ObjectMonitor*) " INTPTR_FORMAT " = {", p2i(this));
2112 st->print_cr(" _header = " INTPTR_FORMAT, header().value());
2113 st->print_cr(" _object = " INTPTR_FORMAT, p2i(_object));
2114 st->print(" _allocation_state = ");
2115 if (is_free()) {
2116 st->print("Free");
2117 } else if (is_old()) {
2118 st->print("Old");
2119 } else if (is_new()) {
2120 st->print("New");
2121 } else {
2122 st->print("unknown=%d", _allocation_state);
2123 }
2124 st->cr();
2125 st->print_cr(" _pad_buf0 = {");
2126 st->print_cr(" [0] = '\\0'");
2127 st->print_cr(" ...");
2128 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf0) - 1);
2129 st->print_cr(" }");
2130 st->print_cr(" _owner = " INTPTR_FORMAT, p2i(_owner));
2131 st->print_cr(" _previous_owner_tid = " JLONG_FORMAT, _previous_owner_tid);
2132 st->print_cr(" _pad_buf1 = {");
2133 st->print_cr(" [0] = '\\0'");
2134 st->print_cr(" ...");
2135 st->print_cr(" [%d] = '\\0'", (int)sizeof(_pad_buf1) - 1);
2136 st->print_cr(" }");
2137 st->print_cr(" _next_om = " INTPTR_FORMAT, p2i(next_om()));
2138 st->print_cr(" _recursions = " INTX_FORMAT, _recursions);
2139 st->print_cr(" _EntryList = " INTPTR_FORMAT, p2i(_EntryList));
2140 st->print_cr(" _cxq = " INTPTR_FORMAT, p2i(_cxq));
2141 st->print_cr(" _succ = " INTPTR_FORMAT, p2i(_succ));
2142 st->print_cr(" _Responsible = " INTPTR_FORMAT, p2i(_Responsible));
2143 st->print_cr(" _Spinner = %d", _Spinner);
2144 st->print_cr(" _SpinDuration = %d", _SpinDuration);
2145 st->print_cr(" _contentions = %d", contentions());
2146 st->print_cr(" _WaitSet = " INTPTR_FORMAT, p2i(_WaitSet));
2147 st->print_cr(" _waiters = %d", _waiters);
2148 st->print_cr(" _WaitSetLock = %d", _WaitSetLock);
2149 st->print_cr("}");
2150 }
2151 #endif