Mercurial > hg > truffle
annotate src/share/vm/runtime/objectMonitor.cpp @ 3187:6ff76a1b8339
The benchmark tool should now print zero values to the csv file, if a benchmark fails
author | Josef Haider <josef.haider@khg.jku.at> |
---|---|
date | Thu, 07 Jul 2011 19:43:17 +0200 |
parents | 1d1603768966 |
children | f08d439fab8c |
rev | line source |
---|---|
1878 | 1 /* |
2426
1d1603768966
7010070: Update all 2010 Oracle-changed OpenJDK files to have the proper copyright dates - second pass
trims
parents:
2177
diff
changeset
|
2 * Copyright (c) 1998, 2011, Oracle and/or its affiliates. All rights reserved. |
1878 | 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 | |
1972 | 25 #include "precompiled.hpp" |
26 #include "classfile/vmSymbols.hpp" | |
27 #include "memory/resourceArea.hpp" | |
28 #include "oops/markOop.hpp" | |
29 #include "oops/oop.inline.hpp" | |
30 #include "runtime/handles.inline.hpp" | |
31 #include "runtime/interfaceSupport.hpp" | |
32 #include "runtime/mutexLocker.hpp" | |
33 #include "runtime/objectMonitor.hpp" | |
34 #include "runtime/objectMonitor.inline.hpp" | |
35 #include "runtime/osThread.hpp" | |
36 #include "runtime/stubRoutines.hpp" | |
37 #include "runtime/thread.hpp" | |
38 #include "services/threadService.hpp" | |
39 #include "utilities/dtrace.hpp" | |
40 #include "utilities/preserveException.hpp" | |
41 #ifdef TARGET_OS_FAMILY_linux | |
42 # include "os_linux.inline.hpp" | |
43 # include "thread_linux.inline.hpp" | |
44 #endif | |
45 #ifdef TARGET_OS_FAMILY_solaris | |
46 # include "os_solaris.inline.hpp" | |
47 # include "thread_solaris.inline.hpp" | |
48 #endif | |
49 #ifdef TARGET_OS_FAMILY_windows | |
50 # include "os_windows.inline.hpp" | |
51 # include "thread_windows.inline.hpp" | |
52 #endif | |
1878 | 53 |
54 #if defined(__GNUC__) && !defined(IA64) | |
55 // Need to inhibit inlining for older versions of GCC to avoid build-time failures | |
56 #define ATTR __attribute__((noinline)) | |
57 #else | |
58 #define ATTR | |
59 #endif | |
60 | |
61 | |
62 #ifdef DTRACE_ENABLED | |
63 | |
64 // Only bother with this argument setup if dtrace is available | |
65 // TODO-FIXME: probes should not fire when caller is _blocked. assert() accordingly. | |
66 | |
67 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notify, | |
68 jlong, uintptr_t, char*, int); | |
69 HS_DTRACE_PROBE_DECL4(hotspot, monitor__notifyAll, | |
70 jlong, uintptr_t, char*, int); | |
71 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__enter, | |
72 jlong, uintptr_t, char*, int); | |
73 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__entered, | |
74 jlong, uintptr_t, char*, int); | |
75 HS_DTRACE_PROBE_DECL4(hotspot, monitor__contended__exit, | |
76 jlong, uintptr_t, char*, int); | |
77 | |
78 #define DTRACE_MONITOR_PROBE_COMMON(klassOop, thread) \ | |
79 char* bytes = NULL; \ | |
80 int len = 0; \ | |
81 jlong jtid = SharedRuntime::get_java_tid(thread); \ | |
2177
3582bf76420e
6990754: Use native memory and reference counting to implement SymbolTable
coleenp
parents:
1972
diff
changeset
|
82 Symbol* klassname = ((oop)(klassOop))->klass()->klass_part()->name(); \ |
1878 | 83 if (klassname != NULL) { \ |
84 bytes = (char*)klassname->bytes(); \ | |
85 len = klassname->utf8_length(); \ | |
86 } | |
87 | |
88 #define DTRACE_MONITOR_WAIT_PROBE(monitor, klassOop, thread, millis) \ | |
89 { \ | |
90 if (DTraceMonitorProbes) { \ | |
91 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ | |
92 HS_DTRACE_PROBE5(hotspot, monitor__wait, jtid, \ | |
93 (monitor), bytes, len, (millis)); \ | |
94 } \ | |
95 } | |
96 | |
97 #define DTRACE_MONITOR_PROBE(probe, monitor, klassOop, thread) \ | |
98 { \ | |
99 if (DTraceMonitorProbes) { \ | |
100 DTRACE_MONITOR_PROBE_COMMON(klassOop, thread); \ | |
101 HS_DTRACE_PROBE4(hotspot, monitor__##probe, jtid, \ | |
102 (uintptr_t)(monitor), bytes, len); \ | |
103 } \ | |
104 } | |
105 | |
106 #else // ndef DTRACE_ENABLED | |
107 | |
108 #define DTRACE_MONITOR_WAIT_PROBE(klassOop, thread, millis, mon) {;} | |
109 #define DTRACE_MONITOR_PROBE(probe, klassOop, thread, mon) {;} | |
110 | |
111 #endif // ndef DTRACE_ENABLED | |
112 | |
113 // Tunables ... | |
114 // The knob* variables are effectively final. Once set they should | |
115 // never be modified hence. Consider using __read_mostly with GCC. | |
116 | |
117 int ObjectMonitor::Knob_Verbose = 0 ; | |
118 int ObjectMonitor::Knob_SpinLimit = 5000 ; // derived by an external tool - | |
119 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins | |
120 static int Knob_HandOff = 0 ; | |
121 static int Knob_ReportSettings = 0 ; | |
122 | |
123 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin | |
124 static int Knob_SpinBackOff = 0 ; // spin-loop backoff | |
125 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS | |
126 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change | |
127 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field | |
128 static int Knob_SpinEarly = 1 ; | |
129 static int Knob_SuccEnabled = 1 ; // futile wake throttling | |
130 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one | |
131 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs | |
132 static int Knob_Bonus = 100 ; // spin success bonus | |
133 static int Knob_BonusB = 100 ; // spin success bonus | |
134 static int Knob_Penalty = 200 ; // spin failure penalty | |
135 static int Knob_Poverty = 1000 ; | |
136 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() | |
137 static int Knob_FixedSpin = 0 ; | |
138 static int Knob_OState = 3 ; // Spinner checks thread state of _owner | |
139 static int Knob_UsePause = 1 ; | |
140 static int Knob_ExitPolicy = 0 ; | |
141 static int Knob_PreSpin = 10 ; // 20-100 likely better | |
142 static int Knob_ResetEvent = 0 ; | |
143 static int BackOffMask = 0 ; | |
144 | |
145 static int Knob_FastHSSEC = 0 ; | |
146 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee | |
147 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline | |
148 static volatile int InitDone = 0 ; | |
149 | |
150 #define TrySpin TrySpin_VaryDuration | |
151 | |
152 // ----------------------------------------------------------------------------- | |
153 // Theory of operations -- Monitors lists, thread residency, etc: | |
154 // | |
155 // * A thread acquires ownership of a monitor by successfully | |
156 // CAS()ing the _owner field from null to non-null. | |
157 // | |
158 // * Invariant: A thread appears on at most one monitor list -- | |
159 // cxq, EntryList or WaitSet -- at any one time. | |
160 // | |
161 // * Contending threads "push" themselves onto the cxq with CAS | |
162 // and then spin/park. | |
163 // | |
164 // * After a contending thread eventually acquires the lock it must | |
165 // dequeue itself from either the EntryList or the cxq. | |
166 // | |
167 // * The exiting thread identifies and unparks an "heir presumptive" | |
168 // tentative successor thread on the EntryList. Critically, the | |
169 // exiting thread doesn't unlink the successor thread from the EntryList. | |
170 // After having been unparked, the wakee will recontend for ownership of | |
171 // the monitor. The successor (wakee) will either acquire the lock or | |
172 // re-park itself. | |
173 // | |
174 // Succession is provided for by a policy of competitive handoff. | |
175 // The exiting thread does _not_ grant or pass ownership to the | |
176 // successor thread. (This is also referred to as "handoff" succession"). | |
177 // Instead the exiting thread releases ownership and possibly wakes | |
178 // a successor, so the successor can (re)compete for ownership of the lock. | |
179 // If the EntryList is empty but the cxq is populated the exiting | |
180 // thread will drain the cxq into the EntryList. It does so by | |
181 // by detaching the cxq (installing null with CAS) and folding | |
182 // the threads from the cxq into the EntryList. The EntryList is | |
183 // doubly linked, while the cxq is singly linked because of the | |
184 // CAS-based "push" used to enqueue recently arrived threads (RATs). | |
185 // | |
186 // * Concurrency invariants: | |
187 // | |
188 // -- only the monitor owner may access or mutate the EntryList. | |
189 // The mutex property of the monitor itself protects the EntryList | |
190 // from concurrent interference. | |
191 // -- Only the monitor owner may detach the cxq. | |
192 // | |
193 // * The monitor entry list operations avoid locks, but strictly speaking | |
194 // they're not lock-free. Enter is lock-free, exit is not. | |
195 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html | |
196 // | |
197 // * The cxq can have multiple concurrent "pushers" but only one concurrent | |
198 // detaching thread. This mechanism is immune from the ABA corruption. | |
199 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. | |
200 // | |
201 // * Taken together, the cxq and the EntryList constitute or form a | |
202 // single logical queue of threads stalled trying to acquire the lock. | |
203 // We use two distinct lists to improve the odds of a constant-time | |
204 // dequeue operation after acquisition (in the ::enter() epilog) and | |
205 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). | |
206 // A key desideratum is to minimize queue & monitor metadata manipulation | |
207 // that occurs while holding the monitor lock -- that is, we want to | |
208 // minimize monitor lock holds times. Note that even a small amount of | |
209 // fixed spinning will greatly reduce the # of enqueue-dequeue operations | |
210 // on EntryList|cxq. That is, spinning relieves contention on the "inner" | |
211 // locks and monitor metadata. | |
212 // | |
213 // Cxq points to the the set of Recently Arrived Threads attempting entry. | |
214 // Because we push threads onto _cxq with CAS, the RATs must take the form of | |
215 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when | |
216 // the unlocking thread notices that EntryList is null but _cxq is != null. | |
217 // | |
218 // The EntryList is ordered by the prevailing queue discipline and | |
219 // can be organized in any convenient fashion, such as a doubly-linked list or | |
220 // a circular doubly-linked list. Critically, we want insert and delete operations | |
221 // to operate in constant-time. If we need a priority queue then something akin | |
222 // to Solaris' sleepq would work nicely. Viz., | |
223 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. | |
224 // Queue discipline is enforced at ::exit() time, when the unlocking thread | |
225 // drains the cxq into the EntryList, and orders or reorders the threads on the | |
226 // EntryList accordingly. | |
227 // | |
228 // Barring "lock barging", this mechanism provides fair cyclic ordering, | |
229 // somewhat similar to an elevator-scan. | |
230 // | |
231 // * The monitor synchronization subsystem avoids the use of native | |
232 // synchronization primitives except for the narrow platform-specific | |
233 // park-unpark abstraction. See the comments in os_solaris.cpp regarding | |
234 // the semantics of park-unpark. Put another way, this monitor implementation | |
235 // depends only on atomic operations and park-unpark. The monitor subsystem | |
236 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the | |
237 // underlying OS manages the READY<->RUN transitions. | |
238 // | |
239 // * Waiting threads reside on the WaitSet list -- wait() puts | |
240 // the caller onto the WaitSet. | |
241 // | |
242 // * notify() or notifyAll() simply transfers threads from the WaitSet to | |
243 // either the EntryList or cxq. Subsequent exit() operations will | |
244 // unpark the notifyee. Unparking a notifee in notify() is inefficient - | |
245 // it's likely the notifyee would simply impale itself on the lock held | |
246 // by the notifier. | |
247 // | |
248 // * An interesting alternative is to encode cxq as (List,LockByte) where | |
249 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary | |
250 // variable, like _recursions, in the scheme. The threads or Events that form | |
251 // the list would have to be aligned in 256-byte addresses. A thread would | |
252 // try to acquire the lock or enqueue itself with CAS, but exiting threads | |
253 // could use a 1-0 protocol and simply STB to set the LockByte to 0. | |
254 // Note that is is *not* word-tearing, but it does presume that full-word | |
255 // CAS operations are coherent with intermix with STB operations. That's true | |
256 // on most common processors. | |
257 // | |
258 // * See also http://blogs.sun.com/dave | |
259 | |
260 | |
261 // ----------------------------------------------------------------------------- | |
262 // Enter support | |
263 | |
264 bool ObjectMonitor::try_enter(Thread* THREAD) { | |
265 if (THREAD != _owner) { | |
266 if (THREAD->is_lock_owned ((address)_owner)) { | |
267 assert(_recursions == 0, "internal state error"); | |
268 _owner = THREAD ; | |
269 _recursions = 1 ; | |
270 OwnerIsThread = 1 ; | |
271 return true; | |
272 } | |
273 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { | |
274 return false; | |
275 } | |
276 return true; | |
277 } else { | |
278 _recursions++; | |
279 return true; | |
280 } | |
281 } | |
282 | |
283 void ATTR ObjectMonitor::enter(TRAPS) { | |
284 // The following code is ordered to check the most common cases first | |
285 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. | |
286 Thread * const Self = THREAD ; | |
287 void * cur ; | |
288 | |
289 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; | |
290 if (cur == NULL) { | |
291 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. | |
292 assert (_recursions == 0 , "invariant") ; | |
293 assert (_owner == Self, "invariant") ; | |
294 // CONSIDER: set or assert OwnerIsThread == 1 | |
295 return ; | |
296 } | |
297 | |
298 if (cur == Self) { | |
299 // TODO-FIXME: check for integer overflow! BUGID 6557169. | |
300 _recursions ++ ; | |
301 return ; | |
302 } | |
303 | |
304 if (Self->is_lock_owned ((address)cur)) { | |
305 assert (_recursions == 0, "internal state error"); | |
306 _recursions = 1 ; | |
307 // Commute owner from a thread-specific on-stack BasicLockObject address to | |
308 // a full-fledged "Thread *". | |
309 _owner = Self ; | |
310 OwnerIsThread = 1 ; | |
311 return ; | |
312 } | |
313 | |
314 // We've encountered genuine contention. | |
315 assert (Self->_Stalled == 0, "invariant") ; | |
316 Self->_Stalled = intptr_t(this) ; | |
317 | |
318 // Try one round of spinning *before* enqueueing Self | |
319 // and before going through the awkward and expensive state | |
320 // transitions. The following spin is strictly optional ... | |
321 // Note that if we acquire the monitor from an initial spin | |
322 // we forgo posting JVMTI events and firing DTRACE probes. | |
323 if (Knob_SpinEarly && TrySpin (Self) > 0) { | |
324 assert (_owner == Self , "invariant") ; | |
325 assert (_recursions == 0 , "invariant") ; | |
326 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; | |
327 Self->_Stalled = 0 ; | |
328 return ; | |
329 } | |
330 | |
331 assert (_owner != Self , "invariant") ; | |
332 assert (_succ != Self , "invariant") ; | |
333 assert (Self->is_Java_thread() , "invariant") ; | |
334 JavaThread * jt = (JavaThread *) Self ; | |
335 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; | |
336 assert (jt->thread_state() != _thread_blocked , "invariant") ; | |
337 assert (this->object() != NULL , "invariant") ; | |
338 assert (_count >= 0, "invariant") ; | |
339 | |
340 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). | |
341 // Ensure the object-monitor relationship remains stable while there's contention. | |
342 Atomic::inc_ptr(&_count); | |
343 | |
344 { // Change java thread status to indicate blocked on monitor enter. | |
345 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); | |
346 | |
347 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); | |
348 if (JvmtiExport::should_post_monitor_contended_enter()) { | |
349 JvmtiExport::post_monitor_contended_enter(jt, this); | |
350 } | |
351 | |
352 OSThreadContendState osts(Self->osthread()); | |
353 ThreadBlockInVM tbivm(jt); | |
354 | |
355 Self->set_current_pending_monitor(this); | |
356 | |
357 // TODO-FIXME: change the following for(;;) loop to straight-line code. | |
358 for (;;) { | |
359 jt->set_suspend_equivalent(); | |
360 // cleared by handle_special_suspend_equivalent_condition() | |
361 // or java_suspend_self() | |
362 | |
363 EnterI (THREAD) ; | |
364 | |
365 if (!ExitSuspendEquivalent(jt)) break ; | |
366 | |
367 // | |
368 // We have acquired the contended monitor, but while we were | |
369 // waiting another thread suspended us. We don't want to enter | |
370 // the monitor while suspended because that would surprise the | |
371 // thread that suspended us. | |
372 // | |
373 _recursions = 0 ; | |
374 _succ = NULL ; | |
375 exit (Self) ; | |
376 | |
377 jt->java_suspend_self(); | |
378 } | |
379 Self->set_current_pending_monitor(NULL); | |
380 } | |
381 | |
382 Atomic::dec_ptr(&_count); | |
383 assert (_count >= 0, "invariant") ; | |
384 Self->_Stalled = 0 ; | |
385 | |
386 // Must either set _recursions = 0 or ASSERT _recursions == 0. | |
387 assert (_recursions == 0 , "invariant") ; | |
388 assert (_owner == Self , "invariant") ; | |
389 assert (_succ != Self , "invariant") ; | |
390 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; | |
391 | |
392 // The thread -- now the owner -- is back in vm mode. | |
393 // Report the glorious news via TI,DTrace and jvmstat. | |
394 // The probe effect is non-trivial. All the reportage occurs | |
395 // while we hold the monitor, increasing the length of the critical | |
396 // section. Amdahl's parallel speedup law comes vividly into play. | |
397 // | |
398 // Another option might be to aggregate the events (thread local or | |
399 // per-monitor aggregation) and defer reporting until a more opportune | |
400 // time -- such as next time some thread encounters contention but has | |
401 // yet to acquire the lock. While spinning that thread could | |
402 // spinning we could increment JVMStat counters, etc. | |
403 | |
404 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); | |
405 if (JvmtiExport::should_post_monitor_contended_entered()) { | |
406 JvmtiExport::post_monitor_contended_entered(jt, this); | |
407 } | |
408 if (ObjectMonitor::_sync_ContendedLockAttempts != NULL) { | |
409 ObjectMonitor::_sync_ContendedLockAttempts->inc() ; | |
410 } | |
411 } | |
412 | |
413 | |
414 // Caveat: TryLock() is not necessarily serializing if it returns failure. | |
415 // Callers must compensate as needed. | |
416 | |
417 int ObjectMonitor::TryLock (Thread * Self) { | |
418 for (;;) { | |
419 void * own = _owner ; | |
420 if (own != NULL) return 0 ; | |
421 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { | |
422 // Either guarantee _recursions == 0 or set _recursions = 0. | |
423 assert (_recursions == 0, "invariant") ; | |
424 assert (_owner == Self, "invariant") ; | |
425 // CONSIDER: set or assert that OwnerIsThread == 1 | |
426 return 1 ; | |
427 } | |
428 // The lock had been free momentarily, but we lost the race to the lock. | |
429 // Interference -- the CAS failed. | |
430 // We can either return -1 or retry. | |
431 // Retry doesn't make as much sense because the lock was just acquired. | |
432 if (true) return -1 ; | |
433 } | |
434 } | |
435 | |
436 void ATTR ObjectMonitor::EnterI (TRAPS) { | |
437 Thread * Self = THREAD ; | |
438 assert (Self->is_Java_thread(), "invariant") ; | |
439 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; | |
440 | |
441 // Try the lock - TATAS | |
442 if (TryLock (Self) > 0) { | |
443 assert (_succ != Self , "invariant") ; | |
444 assert (_owner == Self , "invariant") ; | |
445 assert (_Responsible != Self , "invariant") ; | |
446 return ; | |
447 } | |
448 | |
449 DeferredInitialize () ; | |
450 | |
451 // We try one round of spinning *before* enqueueing Self. | |
452 // | |
453 // If the _owner is ready but OFFPROC we could use a YieldTo() | |
454 // operation to donate the remainder of this thread's quantum | |
455 // to the owner. This has subtle but beneficial affinity | |
456 // effects. | |
457 | |
458 if (TrySpin (Self) > 0) { | |
459 assert (_owner == Self , "invariant") ; | |
460 assert (_succ != Self , "invariant") ; | |
461 assert (_Responsible != Self , "invariant") ; | |
462 return ; | |
463 } | |
464 | |
465 // The Spin failed -- Enqueue and park the thread ... | |
466 assert (_succ != Self , "invariant") ; | |
467 assert (_owner != Self , "invariant") ; | |
468 assert (_Responsible != Self , "invariant") ; | |
469 | |
470 // Enqueue "Self" on ObjectMonitor's _cxq. | |
471 // | |
472 // Node acts as a proxy for Self. | |
473 // As an aside, if were to ever rewrite the synchronization code mostly | |
474 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class | |
475 // Java objects. This would avoid awkward lifecycle and liveness issues, | |
476 // as well as eliminate a subset of ABA issues. | |
477 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. | |
478 // | |
479 | |
480 ObjectWaiter node(Self) ; | |
481 Self->_ParkEvent->reset() ; | |
482 node._prev = (ObjectWaiter *) 0xBAD ; | |
483 node.TState = ObjectWaiter::TS_CXQ ; | |
484 | |
485 // Push "Self" onto the front of the _cxq. | |
486 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. | |
487 // Note that spinning tends to reduce the rate at which threads | |
488 // enqueue and dequeue on EntryList|cxq. | |
489 ObjectWaiter * nxt ; | |
490 for (;;) { | |
491 node._next = nxt = _cxq ; | |
492 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; | |
493 | |
494 // Interference - the CAS failed because _cxq changed. Just retry. | |
495 // As an optional optimization we retry the lock. | |
496 if (TryLock (Self) > 0) { | |
497 assert (_succ != Self , "invariant") ; | |
498 assert (_owner == Self , "invariant") ; | |
499 assert (_Responsible != Self , "invariant") ; | |
500 return ; | |
501 } | |
502 } | |
503 | |
504 // Check for cxq|EntryList edge transition to non-null. This indicates | |
505 // the onset of contention. While contention persists exiting threads | |
506 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit | |
507 // operations revert to the faster 1-0 mode. This enter operation may interleave | |
508 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we | |
509 // arrange for one of the contending thread to use a timed park() operations | |
510 // to detect and recover from the race. (Stranding is form of progress failure | |
511 // where the monitor is unlocked but all the contending threads remain parked). | |
512 // That is, at least one of the contended threads will periodically poll _owner. | |
513 // One of the contending threads will become the designated "Responsible" thread. | |
514 // The Responsible thread uses a timed park instead of a normal indefinite park | |
515 // operation -- it periodically wakes and checks for and recovers from potential | |
516 // strandings admitted by 1-0 exit operations. We need at most one Responsible | |
517 // thread per-monitor at any given moment. Only threads on cxq|EntryList may | |
518 // be responsible for a monitor. | |
519 // | |
520 // Currently, one of the contended threads takes on the added role of "Responsible". | |
521 // A viable alternative would be to use a dedicated "stranding checker" thread | |
522 // that periodically iterated over all the threads (or active monitors) and unparked | |
523 // successors where there was risk of stranding. This would help eliminate the | |
524 // timer scalability issues we see on some platforms as we'd only have one thread | |
525 // -- the checker -- parked on a timer. | |
526 | |
527 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { | |
528 // Try to assume the role of responsible thread for the monitor. | |
529 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } | |
530 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; | |
531 } | |
532 | |
533 // The lock have been released while this thread was occupied queueing | |
534 // itself onto _cxq. To close the race and avoid "stranding" and | |
535 // progress-liveness failure we must resample-retry _owner before parking. | |
536 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. | |
537 // In this case the ST-MEMBAR is accomplished with CAS(). | |
538 // | |
539 // TODO: Defer all thread state transitions until park-time. | |
540 // Since state transitions are heavy and inefficient we'd like | |
541 // to defer the state transitions until absolutely necessary, | |
542 // and in doing so avoid some transitions ... | |
543 | |
544 TEVENT (Inflated enter - Contention) ; | |
545 int nWakeups = 0 ; | |
546 int RecheckInterval = 1 ; | |
547 | |
548 for (;;) { | |
549 | |
550 if (TryLock (Self) > 0) break ; | |
551 assert (_owner != Self, "invariant") ; | |
552 | |
553 if ((SyncFlags & 2) && _Responsible == NULL) { | |
554 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; | |
555 } | |
556 | |
557 // park self | |
558 if (_Responsible == Self || (SyncFlags & 1)) { | |
559 TEVENT (Inflated enter - park TIMED) ; | |
560 Self->_ParkEvent->park ((jlong) RecheckInterval) ; | |
561 // Increase the RecheckInterval, but clamp the value. | |
562 RecheckInterval *= 8 ; | |
563 if (RecheckInterval > 1000) RecheckInterval = 1000 ; | |
564 } else { | |
565 TEVENT (Inflated enter - park UNTIMED) ; | |
566 Self->_ParkEvent->park() ; | |
567 } | |
568 | |
569 if (TryLock(Self) > 0) break ; | |
570 | |
571 // The lock is still contested. | |
572 // Keep a tally of the # of futile wakeups. | |
573 // Note that the counter is not protected by a lock or updated by atomics. | |
574 // That is by design - we trade "lossy" counters which are exposed to | |
575 // races during updates for a lower probe effect. | |
576 TEVENT (Inflated enter - Futile wakeup) ; | |
577 if (ObjectMonitor::_sync_FutileWakeups != NULL) { | |
578 ObjectMonitor::_sync_FutileWakeups->inc() ; | |
579 } | |
580 ++ nWakeups ; | |
581 | |
582 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. | |
583 // We can defer clearing _succ until after the spin completes | |
584 // TrySpin() must tolerate being called with _succ == Self. | |
585 // Try yet another round of adaptive spinning. | |
586 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; | |
587 | |
588 // We can find that we were unpark()ed and redesignated _succ while | |
589 // we were spinning. That's harmless. If we iterate and call park(), | |
590 // park() will consume the event and return immediately and we'll | |
591 // just spin again. This pattern can repeat, leaving _succ to simply | |
592 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). | |
593 // Alternately, we can sample fired() here, and if set, forgo spinning | |
594 // in the next iteration. | |
595 | |
596 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { | |
597 Self->_ParkEvent->reset() ; | |
598 OrderAccess::fence() ; | |
599 } | |
600 if (_succ == Self) _succ = NULL ; | |
601 | |
602 // Invariant: after clearing _succ a thread *must* retry _owner before parking. | |
603 OrderAccess::fence() ; | |
604 } | |
605 | |
606 // Egress : | |
607 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. | |
608 // Normally we'll find Self on the EntryList . | |
609 // From the perspective of the lock owner (this thread), the | |
610 // EntryList is stable and cxq is prepend-only. | |
611 // The head of cxq is volatile but the interior is stable. | |
612 // In addition, Self.TState is stable. | |
613 | |
614 assert (_owner == Self , "invariant") ; | |
615 assert (object() != NULL , "invariant") ; | |
616 // I'd like to write: | |
617 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; | |
618 // but as we're at a safepoint that's not safe. | |
619 | |
620 UnlinkAfterAcquire (Self, &node) ; | |
621 if (_succ == Self) _succ = NULL ; | |
622 | |
623 assert (_succ != Self, "invariant") ; | |
624 if (_Responsible == Self) { | |
625 _Responsible = NULL ; | |
626 // Dekker pivot-point. | |
627 // Consider OrderAccess::storeload() here | |
628 | |
629 // We may leave threads on cxq|EntryList without a designated | |
630 // "Responsible" thread. This is benign. When this thread subsequently | |
631 // exits the monitor it can "see" such preexisting "old" threads -- | |
632 // threads that arrived on the cxq|EntryList before the fence, above -- | |
633 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads | |
634 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible | |
635 // non-null and elect a new "Responsible" timer thread. | |
636 // | |
637 // This thread executes: | |
638 // ST Responsible=null; MEMBAR (in enter epilog - here) | |
639 // LD cxq|EntryList (in subsequent exit) | |
640 // | |
641 // Entering threads in the slow/contended path execute: | |
642 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) | |
643 // The (ST cxq; MEMBAR) is accomplished with CAS(). | |
644 // | |
645 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent | |
646 // exit operation from floating above the ST Responsible=null. | |
647 // | |
648 // In *practice* however, EnterI() is always followed by some atomic | |
649 // operation such as the decrement of _count in ::enter(). Those atomics | |
650 // obviate the need for the explicit MEMBAR, above. | |
651 } | |
652 | |
653 // We've acquired ownership with CAS(). | |
654 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. | |
655 // But since the CAS() this thread may have also stored into _succ, | |
656 // EntryList, cxq or Responsible. These meta-data updates must be | |
657 // visible __before this thread subsequently drops the lock. | |
658 // Consider what could occur if we didn't enforce this constraint -- | |
659 // STs to monitor meta-data and user-data could reorder with (become | |
660 // visible after) the ST in exit that drops ownership of the lock. | |
661 // Some other thread could then acquire the lock, but observe inconsistent | |
662 // or old monitor meta-data and heap data. That violates the JMM. | |
663 // To that end, the 1-0 exit() operation must have at least STST|LDST | |
664 // "release" barrier semantics. Specifically, there must be at least a | |
665 // STST|LDST barrier in exit() before the ST of null into _owner that drops | |
666 // the lock. The barrier ensures that changes to monitor meta-data and data | |
667 // protected by the lock will be visible before we release the lock, and | |
668 // therefore before some other thread (CPU) has a chance to acquire the lock. | |
669 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. | |
670 // | |
671 // Critically, any prior STs to _succ or EntryList must be visible before | |
672 // the ST of null into _owner in the *subsequent* (following) corresponding | |
673 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily | |
674 // execute a serializing instruction. | |
675 | |
676 if (SyncFlags & 8) { | |
677 OrderAccess::fence() ; | |
678 } | |
679 return ; | |
680 } | |
681 | |
682 // ReenterI() is a specialized inline form of the latter half of the | |
683 // contended slow-path from EnterI(). We use ReenterI() only for | |
684 // monitor reentry in wait(). | |
685 // | |
686 // In the future we should reconcile EnterI() and ReenterI(), adding | |
687 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the | |
688 // loop accordingly. | |
689 | |
690 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { | |
691 assert (Self != NULL , "invariant") ; | |
692 assert (SelfNode != NULL , "invariant") ; | |
693 assert (SelfNode->_thread == Self , "invariant") ; | |
694 assert (_waiters > 0 , "invariant") ; | |
695 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; | |
696 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; | |
697 JavaThread * jt = (JavaThread *) Self ; | |
698 | |
699 int nWakeups = 0 ; | |
700 for (;;) { | |
701 ObjectWaiter::TStates v = SelfNode->TState ; | |
702 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; | |
703 assert (_owner != Self, "invariant") ; | |
704 | |
705 if (TryLock (Self) > 0) break ; | |
706 if (TrySpin (Self) > 0) break ; | |
707 | |
708 TEVENT (Wait Reentry - parking) ; | |
709 | |
710 // State transition wrappers around park() ... | |
711 // ReenterI() wisely defers state transitions until | |
712 // it's clear we must park the thread. | |
713 { | |
714 OSThreadContendState osts(Self->osthread()); | |
715 ThreadBlockInVM tbivm(jt); | |
716 | |
717 // cleared by handle_special_suspend_equivalent_condition() | |
718 // or java_suspend_self() | |
719 jt->set_suspend_equivalent(); | |
720 if (SyncFlags & 1) { | |
721 Self->_ParkEvent->park ((jlong)1000) ; | |
722 } else { | |
723 Self->_ParkEvent->park () ; | |
724 } | |
725 | |
726 // were we externally suspended while we were waiting? | |
727 for (;;) { | |
728 if (!ExitSuspendEquivalent (jt)) break ; | |
729 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } | |
730 jt->java_suspend_self(); | |
731 jt->set_suspend_equivalent(); | |
732 } | |
733 } | |
734 | |
735 // Try again, but just so we distinguish between futile wakeups and | |
736 // successful wakeups. The following test isn't algorithmically | |
737 // necessary, but it helps us maintain sensible statistics. | |
738 if (TryLock(Self) > 0) break ; | |
739 | |
740 // The lock is still contested. | |
741 // Keep a tally of the # of futile wakeups. | |
742 // Note that the counter is not protected by a lock or updated by atomics. | |
743 // That is by design - we trade "lossy" counters which are exposed to | |
744 // races during updates for a lower probe effect. | |
745 TEVENT (Wait Reentry - futile wakeup) ; | |
746 ++ nWakeups ; | |
747 | |
748 // Assuming this is not a spurious wakeup we'll normally | |
749 // find that _succ == Self. | |
750 if (_succ == Self) _succ = NULL ; | |
751 | |
752 // Invariant: after clearing _succ a contending thread | |
753 // *must* retry _owner before parking. | |
754 OrderAccess::fence() ; | |
755 | |
756 if (ObjectMonitor::_sync_FutileWakeups != NULL) { | |
757 ObjectMonitor::_sync_FutileWakeups->inc() ; | |
758 } | |
759 } | |
760 | |
761 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . | |
762 // Normally we'll find Self on the EntryList. | |
763 // Unlinking from the EntryList is constant-time and atomic-free. | |
764 // From the perspective of the lock owner (this thread), the | |
765 // EntryList is stable and cxq is prepend-only. | |
766 // The head of cxq is volatile but the interior is stable. | |
767 // In addition, Self.TState is stable. | |
768 | |
769 assert (_owner == Self, "invariant") ; | |
770 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; | |
771 UnlinkAfterAcquire (Self, SelfNode) ; | |
772 if (_succ == Self) _succ = NULL ; | |
773 assert (_succ != Self, "invariant") ; | |
774 SelfNode->TState = ObjectWaiter::TS_RUN ; | |
775 OrderAccess::fence() ; // see comments at the end of EnterI() | |
776 } | |
777 | |
778 // after the thread acquires the lock in ::enter(). Equally, we could defer | |
779 // unlinking the thread until ::exit()-time. | |
780 | |
781 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) | |
782 { | |
783 assert (_owner == Self, "invariant") ; | |
784 assert (SelfNode->_thread == Self, "invariant") ; | |
785 | |
786 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { | |
787 // Normal case: remove Self from the DLL EntryList . | |
788 // This is a constant-time operation. | |
789 ObjectWaiter * nxt = SelfNode->_next ; | |
790 ObjectWaiter * prv = SelfNode->_prev ; | |
791 if (nxt != NULL) nxt->_prev = prv ; | |
792 if (prv != NULL) prv->_next = nxt ; | |
793 if (SelfNode == _EntryList ) _EntryList = nxt ; | |
794 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
795 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
796 TEVENT (Unlink from EntryList) ; | |
797 } else { | |
798 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; | |
799 // Inopportune interleaving -- Self is still on the cxq. | |
800 // This usually means the enqueue of self raced an exiting thread. | |
801 // Normally we'll find Self near the front of the cxq, so | |
802 // dequeueing is typically fast. If needbe we can accelerate | |
803 // this with some MCS/CHL-like bidirectional list hints and advisory | |
804 // back-links so dequeueing from the interior will normally operate | |
805 // in constant-time. | |
806 // Dequeue Self from either the head (with CAS) or from the interior | |
807 // with a linear-time scan and normal non-atomic memory operations. | |
808 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList | |
809 // and then unlink Self from EntryList. We have to drain eventually, | |
810 // so it might as well be now. | |
811 | |
812 ObjectWaiter * v = _cxq ; | |
813 assert (v != NULL, "invariant") ; | |
814 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { | |
815 // The CAS above can fail from interference IFF a "RAT" arrived. | |
816 // In that case Self must be in the interior and can no longer be | |
817 // at the head of cxq. | |
818 if (v == SelfNode) { | |
819 assert (_cxq != v, "invariant") ; | |
820 v = _cxq ; // CAS above failed - start scan at head of list | |
821 } | |
822 ObjectWaiter * p ; | |
823 ObjectWaiter * q = NULL ; | |
824 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { | |
825 q = p ; | |
826 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; | |
827 } | |
828 assert (v != SelfNode, "invariant") ; | |
829 assert (p == SelfNode, "Node not found on cxq") ; | |
830 assert (p != _cxq, "invariant") ; | |
831 assert (q != NULL, "invariant") ; | |
832 assert (q->_next == p, "invariant") ; | |
833 q->_next = p->_next ; | |
834 } | |
835 TEVENT (Unlink from cxq) ; | |
836 } | |
837 | |
838 // Diagnostic hygiene ... | |
839 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; | |
840 SelfNode->_next = (ObjectWaiter *) 0xBAD ; | |
841 SelfNode->TState = ObjectWaiter::TS_RUN ; | |
842 } | |
843 | |
844 // ----------------------------------------------------------------------------- | |
845 // Exit support | |
846 // | |
847 // exit() | |
848 // ~~~~~~ | |
849 // Note that the collector can't reclaim the objectMonitor or deflate | |
850 // the object out from underneath the thread calling ::exit() as the | |
851 // thread calling ::exit() never transitions to a stable state. | |
852 // This inhibits GC, which in turn inhibits asynchronous (and | |
853 // inopportune) reclamation of "this". | |
854 // | |
855 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; | |
856 // There's one exception to the claim above, however. EnterI() can call | |
857 // exit() to drop a lock if the acquirer has been externally suspended. | |
858 // In that case exit() is called with _thread_state as _thread_blocked, | |
859 // but the monitor's _count field is > 0, which inhibits reclamation. | |
860 // | |
861 // 1-0 exit | |
862 // ~~~~~~~~ | |
863 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of | |
864 // the fast-path operators have been optimized so the common ::exit() | |
865 // operation is 1-0. See i486.ad fast_unlock(), for instance. | |
866 // The code emitted by fast_unlock() elides the usual MEMBAR. This | |
867 // greatly improves latency -- MEMBAR and CAS having considerable local | |
868 // latency on modern processors -- but at the cost of "stranding". Absent the | |
869 // MEMBAR, a thread in fast_unlock() can race a thread in the slow | |
870 // ::enter() path, resulting in the entering thread being stranding | |
871 // and a progress-liveness failure. Stranding is extremely rare. | |
872 // We use timers (timed park operations) & periodic polling to detect | |
873 // and recover from stranding. Potentially stranded threads periodically | |
874 // wake up and poll the lock. See the usage of the _Responsible variable. | |
875 // | |
876 // The CAS() in enter provides for safety and exclusion, while the CAS or | |
877 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking | |
878 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. | |
879 // We detect and recover from stranding with timers. | |
880 // | |
881 // If a thread transiently strands it'll park until (a) another | |
882 // thread acquires the lock and then drops the lock, at which time the | |
883 // exiting thread will notice and unpark the stranded thread, or, (b) | |
884 // the timer expires. If the lock is high traffic then the stranding latency | |
885 // will be low due to (a). If the lock is low traffic then the odds of | |
886 // stranding are lower, although the worst-case stranding latency | |
887 // is longer. Critically, we don't want to put excessive load in the | |
888 // platform's timer subsystem. We want to minimize both the timer injection | |
889 // rate (timers created/sec) as well as the number of timers active at | |
890 // any one time. (more precisely, we want to minimize timer-seconds, which is | |
891 // the integral of the # of active timers at any instant over time). | |
892 // Both impinge on OS scalability. Given that, at most one thread parked on | |
893 // a monitor will use a timer. | |
894 | |
895 void ATTR ObjectMonitor::exit(TRAPS) { | |
896 Thread * Self = THREAD ; | |
897 if (THREAD != _owner) { | |
898 if (THREAD->is_lock_owned((address) _owner)) { | |
899 // Transmute _owner from a BasicLock pointer to a Thread address. | |
900 // We don't need to hold _mutex for this transition. | |
901 // Non-null to Non-null is safe as long as all readers can | |
902 // tolerate either flavor. | |
903 assert (_recursions == 0, "invariant") ; | |
904 _owner = THREAD ; | |
905 _recursions = 0 ; | |
906 OwnerIsThread = 1 ; | |
907 } else { | |
908 // NOTE: we need to handle unbalanced monitor enter/exit | |
909 // in native code by throwing an exception. | |
910 // TODO: Throw an IllegalMonitorStateException ? | |
911 TEVENT (Exit - Throw IMSX) ; | |
912 assert(false, "Non-balanced monitor enter/exit!"); | |
913 if (false) { | |
914 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); | |
915 } | |
916 return; | |
917 } | |
918 } | |
919 | |
920 if (_recursions != 0) { | |
921 _recursions--; // this is simple recursive enter | |
922 TEVENT (Inflated exit - recursive) ; | |
923 return ; | |
924 } | |
925 | |
926 // Invariant: after setting Responsible=null an thread must execute | |
927 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. | |
928 if ((SyncFlags & 4) == 0) { | |
929 _Responsible = NULL ; | |
930 } | |
931 | |
932 for (;;) { | |
933 assert (THREAD == _owner, "invariant") ; | |
934 | |
935 | |
936 if (Knob_ExitPolicy == 0) { | |
937 // release semantics: prior loads and stores from within the critical section | |
938 // must not float (reorder) past the following store that drops the lock. | |
939 // On SPARC that requires MEMBAR #loadstore|#storestore. | |
940 // But of course in TSO #loadstore|#storestore is not required. | |
941 // I'd like to write one of the following: | |
942 // A. OrderAccess::release() ; _owner = NULL | |
943 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; | |
944 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both | |
945 // store into a _dummy variable. That store is not needed, but can result | |
946 // in massive wasteful coherency traffic on classic SMP systems. | |
947 // Instead, I use release_store(), which is implemented as just a simple | |
948 // ST on x64, x86 and SPARC. | |
949 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock | |
950 OrderAccess::storeload() ; // See if we need to wake a successor | |
951 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { | |
952 TEVENT (Inflated exit - simple egress) ; | |
953 return ; | |
954 } | |
955 TEVENT (Inflated exit - complex egress) ; | |
956 | |
957 // Normally the exiting thread is responsible for ensuring succession, | |
958 // but if other successors are ready or other entering threads are spinning | |
959 // then this thread can simply store NULL into _owner and exit without | |
960 // waking a successor. The existence of spinners or ready successors | |
961 // guarantees proper succession (liveness). Responsibility passes to the | |
962 // ready or running successors. The exiting thread delegates the duty. | |
963 // More precisely, if a successor already exists this thread is absolved | |
964 // of the responsibility of waking (unparking) one. | |
965 // | |
966 // The _succ variable is critical to reducing futile wakeup frequency. | |
967 // _succ identifies the "heir presumptive" thread that has been made | |
968 // ready (unparked) but that has not yet run. We need only one such | |
969 // successor thread to guarantee progress. | |
970 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf | |
971 // section 3.3 "Futile Wakeup Throttling" for details. | |
972 // | |
973 // Note that spinners in Enter() also set _succ non-null. | |
974 // In the current implementation spinners opportunistically set | |
975 // _succ so that exiting threads might avoid waking a successor. | |
976 // Another less appealing alternative would be for the exiting thread | |
977 // to drop the lock and then spin briefly to see if a spinner managed | |
978 // to acquire the lock. If so, the exiting thread could exit | |
979 // immediately without waking a successor, otherwise the exiting | |
980 // thread would need to dequeue and wake a successor. | |
981 // (Note that we'd need to make the post-drop spin short, but no | |
982 // shorter than the worst-case round-trip cache-line migration time. | |
983 // The dropped lock needs to become visible to the spinner, and then | |
984 // the acquisition of the lock by the spinner must become visible to | |
985 // the exiting thread). | |
986 // | |
987 | |
988 // It appears that an heir-presumptive (successor) must be made ready. | |
989 // Only the current lock owner can manipulate the EntryList or | |
990 // drain _cxq, so we need to reacquire the lock. If we fail | |
991 // to reacquire the lock the responsibility for ensuring succession | |
992 // falls to the new owner. | |
993 // | |
994 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { | |
995 return ; | |
996 } | |
997 TEVENT (Exit - Reacquired) ; | |
998 } else { | |
999 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { | |
1000 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock | |
1001 OrderAccess::storeload() ; | |
1002 // Ratify the previously observed values. | |
1003 if (_cxq == NULL || _succ != NULL) { | |
1004 TEVENT (Inflated exit - simple egress) ; | |
1005 return ; | |
1006 } | |
1007 | |
1008 // inopportune interleaving -- the exiting thread (this thread) | |
1009 // in the fast-exit path raced an entering thread in the slow-enter | |
1010 // path. | |
1011 // We have two choices: | |
1012 // A. Try to reacquire the lock. | |
1013 // If the CAS() fails return immediately, otherwise | |
1014 // we either restart/rerun the exit operation, or simply | |
1015 // fall-through into the code below which wakes a successor. | |
1016 // B. If the elements forming the EntryList|cxq are TSM | |
1017 // we could simply unpark() the lead thread and return | |
1018 // without having set _succ. | |
1019 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { | |
1020 TEVENT (Inflated exit - reacquired succeeded) ; | |
1021 return ; | |
1022 } | |
1023 TEVENT (Inflated exit - reacquired failed) ; | |
1024 } else { | |
1025 TEVENT (Inflated exit - complex egress) ; | |
1026 } | |
1027 } | |
1028 | |
1029 guarantee (_owner == THREAD, "invariant") ; | |
1030 | |
1031 ObjectWaiter * w = NULL ; | |
1032 int QMode = Knob_QMode ; | |
1033 | |
1034 if (QMode == 2 && _cxq != NULL) { | |
1035 // QMode == 2 : cxq has precedence over EntryList. | |
1036 // Try to directly wake a successor from the cxq. | |
1037 // If successful, the successor will need to unlink itself from cxq. | |
1038 w = _cxq ; | |
1039 assert (w != NULL, "invariant") ; | |
1040 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; | |
1041 ExitEpilog (Self, w) ; | |
1042 return ; | |
1043 } | |
1044 | |
1045 if (QMode == 3 && _cxq != NULL) { | |
1046 // Aggressively drain cxq into EntryList at the first opportunity. | |
1047 // This policy ensure that recently-run threads live at the head of EntryList. | |
1048 // Drain _cxq into EntryList - bulk transfer. | |
1049 // First, detach _cxq. | |
1050 // The following loop is tantamount to: w = swap (&cxq, NULL) | |
1051 w = _cxq ; | |
1052 for (;;) { | |
1053 assert (w != NULL, "Invariant") ; | |
1054 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; | |
1055 if (u == w) break ; | |
1056 w = u ; | |
1057 } | |
1058 assert (w != NULL , "invariant") ; | |
1059 | |
1060 ObjectWaiter * q = NULL ; | |
1061 ObjectWaiter * p ; | |
1062 for (p = w ; p != NULL ; p = p->_next) { | |
1063 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; | |
1064 p->TState = ObjectWaiter::TS_ENTER ; | |
1065 p->_prev = q ; | |
1066 q = p ; | |
1067 } | |
1068 | |
1069 // Append the RATs to the EntryList | |
1070 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. | |
1071 ObjectWaiter * Tail ; | |
1072 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; | |
1073 if (Tail == NULL) { | |
1074 _EntryList = w ; | |
1075 } else { | |
1076 Tail->_next = w ; | |
1077 w->_prev = Tail ; | |
1078 } | |
1079 | |
1080 // Fall thru into code that tries to wake a successor from EntryList | |
1081 } | |
1082 | |
1083 if (QMode == 4 && _cxq != NULL) { | |
1084 // Aggressively drain cxq into EntryList at the first opportunity. | |
1085 // This policy ensure that recently-run threads live at the head of EntryList. | |
1086 | |
1087 // Drain _cxq into EntryList - bulk transfer. | |
1088 // First, detach _cxq. | |
1089 // The following loop is tantamount to: w = swap (&cxq, NULL) | |
1090 w = _cxq ; | |
1091 for (;;) { | |
1092 assert (w != NULL, "Invariant") ; | |
1093 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; | |
1094 if (u == w) break ; | |
1095 w = u ; | |
1096 } | |
1097 assert (w != NULL , "invariant") ; | |
1098 | |
1099 ObjectWaiter * q = NULL ; | |
1100 ObjectWaiter * p ; | |
1101 for (p = w ; p != NULL ; p = p->_next) { | |
1102 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; | |
1103 p->TState = ObjectWaiter::TS_ENTER ; | |
1104 p->_prev = q ; | |
1105 q = p ; | |
1106 } | |
1107 | |
1108 // Prepend the RATs to the EntryList | |
1109 if (_EntryList != NULL) { | |
1110 q->_next = _EntryList ; | |
1111 _EntryList->_prev = q ; | |
1112 } | |
1113 _EntryList = w ; | |
1114 | |
1115 // Fall thru into code that tries to wake a successor from EntryList | |
1116 } | |
1117 | |
1118 w = _EntryList ; | |
1119 if (w != NULL) { | |
1120 // I'd like to write: guarantee (w->_thread != Self). | |
1121 // But in practice an exiting thread may find itself on the EntryList. | |
1122 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and | |
1123 // then calls exit(). Exit release the lock by setting O._owner to NULL. | |
1124 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The | |
1125 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then | |
1126 // release the lock "O". T2 resumes immediately after the ST of null into | |
1127 // _owner, above. T2 notices that the EntryList is populated, so it | |
1128 // reacquires the lock and then finds itself on the EntryList. | |
1129 // Given all that, we have to tolerate the circumstance where "w" is | |
1130 // associated with Self. | |
1131 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
1132 ExitEpilog (Self, w) ; | |
1133 return ; | |
1134 } | |
1135 | |
1136 // If we find that both _cxq and EntryList are null then just | |
1137 // re-run the exit protocol from the top. | |
1138 w = _cxq ; | |
1139 if (w == NULL) continue ; | |
1140 | |
1141 // Drain _cxq into EntryList - bulk transfer. | |
1142 // First, detach _cxq. | |
1143 // The following loop is tantamount to: w = swap (&cxq, NULL) | |
1144 for (;;) { | |
1145 assert (w != NULL, "Invariant") ; | |
1146 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; | |
1147 if (u == w) break ; | |
1148 w = u ; | |
1149 } | |
1150 TEVENT (Inflated exit - drain cxq into EntryList) ; | |
1151 | |
1152 assert (w != NULL , "invariant") ; | |
1153 assert (_EntryList == NULL , "invariant") ; | |
1154 | |
1155 // Convert the LIFO SLL anchored by _cxq into a DLL. | |
1156 // The list reorganization step operates in O(LENGTH(w)) time. | |
1157 // It's critical that this step operate quickly as | |
1158 // "Self" still holds the outer-lock, restricting parallelism | |
1159 // and effectively lengthening the critical section. | |
1160 // Invariant: s chases t chases u. | |
1161 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so | |
1162 // we have faster access to the tail. | |
1163 | |
1164 if (QMode == 1) { | |
1165 // QMode == 1 : drain cxq to EntryList, reversing order | |
1166 // We also reverse the order of the list. | |
1167 ObjectWaiter * s = NULL ; | |
1168 ObjectWaiter * t = w ; | |
1169 ObjectWaiter * u = NULL ; | |
1170 while (t != NULL) { | |
1171 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; | |
1172 t->TState = ObjectWaiter::TS_ENTER ; | |
1173 u = t->_next ; | |
1174 t->_prev = u ; | |
1175 t->_next = s ; | |
1176 s = t; | |
1177 t = u ; | |
1178 } | |
1179 _EntryList = s ; | |
1180 assert (s != NULL, "invariant") ; | |
1181 } else { | |
1182 // QMode == 0 or QMode == 2 | |
1183 _EntryList = w ; | |
1184 ObjectWaiter * q = NULL ; | |
1185 ObjectWaiter * p ; | |
1186 for (p = w ; p != NULL ; p = p->_next) { | |
1187 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; | |
1188 p->TState = ObjectWaiter::TS_ENTER ; | |
1189 p->_prev = q ; | |
1190 q = p ; | |
1191 } | |
1192 } | |
1193 | |
1194 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL | |
1195 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). | |
1196 | |
1197 // See if we can abdicate to a spinner instead of waking a thread. | |
1198 // A primary goal of the implementation is to reduce the | |
1199 // context-switch rate. | |
1200 if (_succ != NULL) continue; | |
1201 | |
1202 w = _EntryList ; | |
1203 if (w != NULL) { | |
1204 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
1205 ExitEpilog (Self, w) ; | |
1206 return ; | |
1207 } | |
1208 } | |
1209 } | |
1210 | |
1211 // ExitSuspendEquivalent: | |
1212 // A faster alternate to handle_special_suspend_equivalent_condition() | |
1213 // | |
1214 // handle_special_suspend_equivalent_condition() unconditionally | |
1215 // acquires the SR_lock. On some platforms uncontended MutexLocker() | |
1216 // operations have high latency. Note that in ::enter() we call HSSEC | |
1217 // while holding the monitor, so we effectively lengthen the critical sections. | |
1218 // | |
1219 // There are a number of possible solutions: | |
1220 // | |
1221 // A. To ameliorate the problem we might also defer state transitions | |
1222 // to as late as possible -- just prior to parking. | |
1223 // Given that, we'd call HSSEC after having returned from park(), | |
1224 // but before attempting to acquire the monitor. This is only a | |
1225 // partial solution. It avoids calling HSSEC while holding the | |
1226 // monitor (good), but it still increases successor reacquisition latency -- | |
1227 // the interval between unparking a successor and the time the successor | |
1228 // resumes and retries the lock. See ReenterI(), which defers state transitions. | |
1229 // If we use this technique we can also avoid EnterI()-exit() loop | |
1230 // in ::enter() where we iteratively drop the lock and then attempt | |
1231 // to reacquire it after suspending. | |
1232 // | |
1233 // B. In the future we might fold all the suspend bits into a | |
1234 // composite per-thread suspend flag and then update it with CAS(). | |
1235 // Alternately, a Dekker-like mechanism with multiple variables | |
1236 // would suffice: | |
1237 // ST Self->_suspend_equivalent = false | |
1238 // MEMBAR | |
1239 // LD Self_>_suspend_flags | |
1240 // | |
1241 | |
1242 | |
1243 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { | |
1244 int Mode = Knob_FastHSSEC ; | |
1245 if (Mode && !jSelf->is_external_suspend()) { | |
1246 assert (jSelf->is_suspend_equivalent(), "invariant") ; | |
1247 jSelf->clear_suspend_equivalent() ; | |
1248 if (2 == Mode) OrderAccess::storeload() ; | |
1249 if (!jSelf->is_external_suspend()) return false ; | |
1250 // We raced a suspension -- fall thru into the slow path | |
1251 TEVENT (ExitSuspendEquivalent - raced) ; | |
1252 jSelf->set_suspend_equivalent() ; | |
1253 } | |
1254 return jSelf->handle_special_suspend_equivalent_condition() ; | |
1255 } | |
1256 | |
1257 | |
1258 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { | |
1259 assert (_owner == Self, "invariant") ; | |
1260 | |
1261 // Exit protocol: | |
1262 // 1. ST _succ = wakee | |
1263 // 2. membar #loadstore|#storestore; | |
1264 // 2. ST _owner = NULL | |
1265 // 3. unpark(wakee) | |
1266 | |
1267 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; | |
1268 ParkEvent * Trigger = Wakee->_event ; | |
1269 | |
1270 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. | |
1271 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be | |
1272 // out-of-scope (non-extant). | |
1273 Wakee = NULL ; | |
1274 | |
1275 // Drop the lock | |
1276 OrderAccess::release_store_ptr (&_owner, NULL) ; | |
1277 OrderAccess::fence() ; // ST _owner vs LD in unpark() | |
1278 | |
1279 if (SafepointSynchronize::do_call_back()) { | |
1280 TEVENT (unpark before SAFEPOINT) ; | |
1281 } | |
1282 | |
1283 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); | |
1284 Trigger->unpark() ; | |
1285 | |
1286 // Maintain stats and report events to JVMTI | |
1287 if (ObjectMonitor::_sync_Parks != NULL) { | |
1288 ObjectMonitor::_sync_Parks->inc() ; | |
1289 } | |
1290 } | |
1291 | |
1292 | |
1293 // ----------------------------------------------------------------------------- | |
1294 // Class Loader deadlock handling. | |
1295 // | |
1296 // complete_exit exits a lock returning recursion count | |
1297 // complete_exit/reenter operate as a wait without waiting | |
1298 // complete_exit requires an inflated monitor | |
1299 // The _owner field is not always the Thread addr even with an | |
1300 // inflated monitor, e.g. the monitor can be inflated by a non-owning | |
1301 // thread due to contention. | |
1302 intptr_t ObjectMonitor::complete_exit(TRAPS) { | |
1303 Thread * const Self = THREAD; | |
1304 assert(Self->is_Java_thread(), "Must be Java thread!"); | |
1305 JavaThread *jt = (JavaThread *)THREAD; | |
1306 | |
1307 DeferredInitialize(); | |
1308 | |
1309 if (THREAD != _owner) { | |
1310 if (THREAD->is_lock_owned ((address)_owner)) { | |
1311 assert(_recursions == 0, "internal state error"); | |
1312 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ | |
1313 _recursions = 0 ; | |
1314 OwnerIsThread = 1 ; | |
1315 } | |
1316 } | |
1317 | |
1318 guarantee(Self == _owner, "complete_exit not owner"); | |
1319 intptr_t save = _recursions; // record the old recursion count | |
1320 _recursions = 0; // set the recursion level to be 0 | |
1321 exit (Self) ; // exit the monitor | |
1322 guarantee (_owner != Self, "invariant"); | |
1323 return save; | |
1324 } | |
1325 | |
1326 // reenter() enters a lock and sets recursion count | |
1327 // complete_exit/reenter operate as a wait without waiting | |
1328 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { | |
1329 Thread * const Self = THREAD; | |
1330 assert(Self->is_Java_thread(), "Must be Java thread!"); | |
1331 JavaThread *jt = (JavaThread *)THREAD; | |
1332 | |
1333 guarantee(_owner != Self, "reenter already owner"); | |
1334 enter (THREAD); // enter the monitor | |
1335 guarantee (_recursions == 0, "reenter recursion"); | |
1336 _recursions = recursions; | |
1337 return; | |
1338 } | |
1339 | |
1340 | |
1341 // ----------------------------------------------------------------------------- | |
1342 // A macro is used below because there may already be a pending | |
1343 // exception which should not abort the execution of the routines | |
1344 // which use this (which is why we don't put this into check_slow and | |
1345 // call it with a CHECK argument). | |
1346 | |
1347 #define CHECK_OWNER() \ | |
1348 do { \ | |
1349 if (THREAD != _owner) { \ | |
1350 if (THREAD->is_lock_owned((address) _owner)) { \ | |
1351 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ | |
1352 _recursions = 0; \ | |
1353 OwnerIsThread = 1 ; \ | |
1354 } else { \ | |
1355 TEVENT (Throw IMSX) ; \ | |
1356 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ | |
1357 } \ | |
1358 } \ | |
1359 } while (false) | |
1360 | |
1361 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. | |
1362 // TODO-FIXME: remove check_slow() -- it's likely dead. | |
1363 | |
1364 void ObjectMonitor::check_slow(TRAPS) { | |
1365 TEVENT (check_slow - throw IMSX) ; | |
1366 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); | |
1367 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); | |
1368 } | |
1369 | |
1370 static int Adjust (volatile int * adr, int dx) { | |
1371 int v ; | |
1372 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; | |
1373 return v ; | |
1374 } | |
1375 // ----------------------------------------------------------------------------- | |
1376 // Wait/Notify/NotifyAll | |
1377 // | |
1378 // Note: a subset of changes to ObjectMonitor::wait() | |
1379 // will need to be replicated in complete_exit above | |
1380 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { | |
1381 Thread * const Self = THREAD ; | |
1382 assert(Self->is_Java_thread(), "Must be Java thread!"); | |
1383 JavaThread *jt = (JavaThread *)THREAD; | |
1384 | |
1385 DeferredInitialize () ; | |
1386 | |
1387 // Throw IMSX or IEX. | |
1388 CHECK_OWNER(); | |
1389 | |
1390 // check for a pending interrupt | |
1391 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { | |
1392 // post monitor waited event. Note that this is past-tense, we are done waiting. | |
1393 if (JvmtiExport::should_post_monitor_waited()) { | |
1394 // Note: 'false' parameter is passed here because the | |
1395 // wait was not timed out due to thread interrupt. | |
1396 JvmtiExport::post_monitor_waited(jt, this, false); | |
1397 } | |
1398 TEVENT (Wait - Throw IEX) ; | |
1399 THROW(vmSymbols::java_lang_InterruptedException()); | |
1400 return ; | |
1401 } | |
1402 TEVENT (Wait) ; | |
1403 | |
1404 assert (Self->_Stalled == 0, "invariant") ; | |
1405 Self->_Stalled = intptr_t(this) ; | |
1406 jt->set_current_waiting_monitor(this); | |
1407 | |
1408 // create a node to be put into the queue | |
1409 // Critically, after we reset() the event but prior to park(), we must check | |
1410 // for a pending interrupt. | |
1411 ObjectWaiter node(Self); | |
1412 node.TState = ObjectWaiter::TS_WAIT ; | |
1413 Self->_ParkEvent->reset() ; | |
1414 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag | |
1415 | |
1416 // Enter the waiting queue, which is a circular doubly linked list in this case | |
1417 // but it could be a priority queue or any data structure. | |
1418 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only | |
1419 // by the the owner of the monitor *except* in the case where park() | |
1420 // returns because of a timeout of interrupt. Contention is exceptionally rare | |
1421 // so we use a simple spin-lock instead of a heavier-weight blocking lock. | |
1422 | |
1423 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; | |
1424 AddWaiter (&node) ; | |
1425 Thread::SpinRelease (&_WaitSetLock) ; | |
1426 | |
1427 if ((SyncFlags & 4) == 0) { | |
1428 _Responsible = NULL ; | |
1429 } | |
1430 intptr_t save = _recursions; // record the old recursion count | |
1431 _waiters++; // increment the number of waiters | |
1432 _recursions = 0; // set the recursion level to be 1 | |
1433 exit (Self) ; // exit the monitor | |
1434 guarantee (_owner != Self, "invariant") ; | |
1435 | |
1436 // As soon as the ObjectMonitor's ownership is dropped in the exit() | |
1437 // call above, another thread can enter() the ObjectMonitor, do the | |
1438 // notify(), and exit() the ObjectMonitor. If the other thread's | |
1439 // exit() call chooses this thread as the successor and the unpark() | |
1440 // call happens to occur while this thread is posting a | |
1441 // MONITOR_CONTENDED_EXIT event, then we run the risk of the event | |
1442 // handler using RawMonitors and consuming the unpark(). | |
1443 // | |
1444 // To avoid the problem, we re-post the event. This does no harm | |
1445 // even if the original unpark() was not consumed because we are the | |
1446 // chosen successor for this monitor. | |
1447 if (node._notified != 0 && _succ == Self) { | |
1448 node._event->unpark(); | |
1449 } | |
1450 | |
1451 // The thread is on the WaitSet list - now park() it. | |
1452 // On MP systems it's conceivable that a brief spin before we park | |
1453 // could be profitable. | |
1454 // | |
1455 // TODO-FIXME: change the following logic to a loop of the form | |
1456 // while (!timeout && !interrupted && _notified == 0) park() | |
1457 | |
1458 int ret = OS_OK ; | |
1459 int WasNotified = 0 ; | |
1460 { // State transition wrappers | |
1461 OSThread* osthread = Self->osthread(); | |
1462 OSThreadWaitState osts(osthread, true); | |
1463 { | |
1464 ThreadBlockInVM tbivm(jt); | |
1465 // Thread is in thread_blocked state and oop access is unsafe. | |
1466 jt->set_suspend_equivalent(); | |
1467 | |
1468 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { | |
1469 // Intentionally empty | |
1470 } else | |
1471 if (node._notified == 0) { | |
1472 if (millis <= 0) { | |
1473 Self->_ParkEvent->park () ; | |
1474 } else { | |
1475 ret = Self->_ParkEvent->park (millis) ; | |
1476 } | |
1477 } | |
1478 | |
1479 // were we externally suspended while we were waiting? | |
1480 if (ExitSuspendEquivalent (jt)) { | |
1481 // TODO-FIXME: add -- if succ == Self then succ = null. | |
1482 jt->java_suspend_self(); | |
1483 } | |
1484 | |
1485 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm | |
1486 | |
1487 | |
1488 // Node may be on the WaitSet, the EntryList (or cxq), or in transition | |
1489 // from the WaitSet to the EntryList. | |
1490 // See if we need to remove Node from the WaitSet. | |
1491 // We use double-checked locking to avoid grabbing _WaitSetLock | |
1492 // if the thread is not on the wait queue. | |
1493 // | |
1494 // Note that we don't need a fence before the fetch of TState. | |
1495 // In the worst case we'll fetch a old-stale value of TS_WAIT previously | |
1496 // written by the is thread. (perhaps the fetch might even be satisfied | |
1497 // by a look-aside into the processor's own store buffer, although given | |
1498 // the length of the code path between the prior ST and this load that's | |
1499 // highly unlikely). If the following LD fetches a stale TS_WAIT value | |
1500 // then we'll acquire the lock and then re-fetch a fresh TState value. | |
1501 // That is, we fail toward safety. | |
1502 | |
1503 if (node.TState == ObjectWaiter::TS_WAIT) { | |
1504 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; | |
1505 if (node.TState == ObjectWaiter::TS_WAIT) { | |
1506 DequeueSpecificWaiter (&node) ; // unlink from WaitSet | |
1507 assert(node._notified == 0, "invariant"); | |
1508 node.TState = ObjectWaiter::TS_RUN ; | |
1509 } | |
1510 Thread::SpinRelease (&_WaitSetLock) ; | |
1511 } | |
1512 | |
1513 // The thread is now either on off-list (TS_RUN), | |
1514 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). | |
1515 // The Node's TState variable is stable from the perspective of this thread. | |
1516 // No other threads will asynchronously modify TState. | |
1517 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; | |
1518 OrderAccess::loadload() ; | |
1519 if (_succ == Self) _succ = NULL ; | |
1520 WasNotified = node._notified ; | |
1521 | |
1522 // Reentry phase -- reacquire the monitor. | |
1523 // re-enter contended monitor after object.wait(). | |
1524 // retain OBJECT_WAIT state until re-enter successfully completes | |
1525 // Thread state is thread_in_vm and oop access is again safe, | |
1526 // although the raw address of the object may have changed. | |
1527 // (Don't cache naked oops over safepoints, of course). | |
1528 | |
1529 // post monitor waited event. Note that this is past-tense, we are done waiting. | |
1530 if (JvmtiExport::should_post_monitor_waited()) { | |
1531 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); | |
1532 } | |
1533 OrderAccess::fence() ; | |
1534 | |
1535 assert (Self->_Stalled != 0, "invariant") ; | |
1536 Self->_Stalled = 0 ; | |
1537 | |
1538 assert (_owner != Self, "invariant") ; | |
1539 ObjectWaiter::TStates v = node.TState ; | |
1540 if (v == ObjectWaiter::TS_RUN) { | |
1541 enter (Self) ; | |
1542 } else { | |
1543 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; | |
1544 ReenterI (Self, &node) ; | |
1545 node.wait_reenter_end(this); | |
1546 } | |
1547 | |
1548 // Self has reacquired the lock. | |
1549 // Lifecycle - the node representing Self must not appear on any queues. | |
1550 // Node is about to go out-of-scope, but even if it were immortal we wouldn't | |
1551 // want residual elements associated with this thread left on any lists. | |
1552 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; | |
1553 assert (_owner == Self, "invariant") ; | |
1554 assert (_succ != Self , "invariant") ; | |
1555 } // OSThreadWaitState() | |
1556 | |
1557 jt->set_current_waiting_monitor(NULL); | |
1558 | |
1559 guarantee (_recursions == 0, "invariant") ; | |
1560 _recursions = save; // restore the old recursion count | |
1561 _waiters--; // decrement the number of waiters | |
1562 | |
1563 // Verify a few postconditions | |
1564 assert (_owner == Self , "invariant") ; | |
1565 assert (_succ != Self , "invariant") ; | |
1566 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; | |
1567 | |
1568 if (SyncFlags & 32) { | |
1569 OrderAccess::fence() ; | |
1570 } | |
1571 | |
1572 // check if the notification happened | |
1573 if (!WasNotified) { | |
1574 // no, it could be timeout or Thread.interrupt() or both | |
1575 // check for interrupt event, otherwise it is timeout | |
1576 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { | |
1577 TEVENT (Wait - throw IEX from epilog) ; | |
1578 THROW(vmSymbols::java_lang_InterruptedException()); | |
1579 } | |
1580 } | |
1581 | |
1582 // NOTE: Spurious wake up will be consider as timeout. | |
1583 // Monitor notify has precedence over thread interrupt. | |
1584 } | |
1585 | |
1586 | |
1587 // Consider: | |
1588 // If the lock is cool (cxq == null && succ == null) and we're on an MP system | |
1589 // then instead of transferring a thread from the WaitSet to the EntryList | |
1590 // we might just dequeue a thread from the WaitSet and directly unpark() it. | |
1591 | |
1592 void ObjectMonitor::notify(TRAPS) { | |
1593 CHECK_OWNER(); | |
1594 if (_WaitSet == NULL) { | |
1595 TEVENT (Empty-Notify) ; | |
1596 return ; | |
1597 } | |
1598 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); | |
1599 | |
1600 int Policy = Knob_MoveNotifyee ; | |
1601 | |
1602 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; | |
1603 ObjectWaiter * iterator = DequeueWaiter() ; | |
1604 if (iterator != NULL) { | |
1605 TEVENT (Notify1 - Transfer) ; | |
1606 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; | |
1607 guarantee (iterator->_notified == 0, "invariant") ; | |
1608 if (Policy != 4) { | |
1609 iterator->TState = ObjectWaiter::TS_ENTER ; | |
1610 } | |
1611 iterator->_notified = 1 ; | |
1612 | |
1613 ObjectWaiter * List = _EntryList ; | |
1614 if (List != NULL) { | |
1615 assert (List->_prev == NULL, "invariant") ; | |
1616 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
1617 assert (List != iterator, "invariant") ; | |
1618 } | |
1619 | |
1620 if (Policy == 0) { // prepend to EntryList | |
1621 if (List == NULL) { | |
1622 iterator->_next = iterator->_prev = NULL ; | |
1623 _EntryList = iterator ; | |
1624 } else { | |
1625 List->_prev = iterator ; | |
1626 iterator->_next = List ; | |
1627 iterator->_prev = NULL ; | |
1628 _EntryList = iterator ; | |
1629 } | |
1630 } else | |
1631 if (Policy == 1) { // append to EntryList | |
1632 if (List == NULL) { | |
1633 iterator->_next = iterator->_prev = NULL ; | |
1634 _EntryList = iterator ; | |
1635 } else { | |
1636 // CONSIDER: finding the tail currently requires a linear-time walk of | |
1637 // the EntryList. We can make tail access constant-time by converting to | |
1638 // a CDLL instead of using our current DLL. | |
1639 ObjectWaiter * Tail ; | |
1640 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; | |
1641 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; | |
1642 Tail->_next = iterator ; | |
1643 iterator->_prev = Tail ; | |
1644 iterator->_next = NULL ; | |
1645 } | |
1646 } else | |
1647 if (Policy == 2) { // prepend to cxq | |
1648 // prepend to cxq | |
1649 if (List == NULL) { | |
1650 iterator->_next = iterator->_prev = NULL ; | |
1651 _EntryList = iterator ; | |
1652 } else { | |
1653 iterator->TState = ObjectWaiter::TS_CXQ ; | |
1654 for (;;) { | |
1655 ObjectWaiter * Front = _cxq ; | |
1656 iterator->_next = Front ; | |
1657 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { | |
1658 break ; | |
1659 } | |
1660 } | |
1661 } | |
1662 } else | |
1663 if (Policy == 3) { // append to cxq | |
1664 iterator->TState = ObjectWaiter::TS_CXQ ; | |
1665 for (;;) { | |
1666 ObjectWaiter * Tail ; | |
1667 Tail = _cxq ; | |
1668 if (Tail == NULL) { | |
1669 iterator->_next = NULL ; | |
1670 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { | |
1671 break ; | |
1672 } | |
1673 } else { | |
1674 while (Tail->_next != NULL) Tail = Tail->_next ; | |
1675 Tail->_next = iterator ; | |
1676 iterator->_prev = Tail ; | |
1677 iterator->_next = NULL ; | |
1678 break ; | |
1679 } | |
1680 } | |
1681 } else { | |
1682 ParkEvent * ev = iterator->_event ; | |
1683 iterator->TState = ObjectWaiter::TS_RUN ; | |
1684 OrderAccess::fence() ; | |
1685 ev->unpark() ; | |
1686 } | |
1687 | |
1688 if (Policy < 4) { | |
1689 iterator->wait_reenter_begin(this); | |
1690 } | |
1691 | |
1692 // _WaitSetLock protects the wait queue, not the EntryList. We could | |
1693 // move the add-to-EntryList operation, above, outside the critical section | |
1694 // protected by _WaitSetLock. In practice that's not useful. With the | |
1695 // exception of wait() timeouts and interrupts the monitor owner | |
1696 // is the only thread that grabs _WaitSetLock. There's almost no contention | |
1697 // on _WaitSetLock so it's not profitable to reduce the length of the | |
1698 // critical section. | |
1699 } | |
1700 | |
1701 Thread::SpinRelease (&_WaitSetLock) ; | |
1702 | |
1703 if (iterator != NULL && ObjectMonitor::_sync_Notifications != NULL) { | |
1704 ObjectMonitor::_sync_Notifications->inc() ; | |
1705 } | |
1706 } | |
1707 | |
1708 | |
1709 void ObjectMonitor::notifyAll(TRAPS) { | |
1710 CHECK_OWNER(); | |
1711 ObjectWaiter* iterator; | |
1712 if (_WaitSet == NULL) { | |
1713 TEVENT (Empty-NotifyAll) ; | |
1714 return ; | |
1715 } | |
1716 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); | |
1717 | |
1718 int Policy = Knob_MoveNotifyee ; | |
1719 int Tally = 0 ; | |
1720 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; | |
1721 | |
1722 for (;;) { | |
1723 iterator = DequeueWaiter () ; | |
1724 if (iterator == NULL) break ; | |
1725 TEVENT (NotifyAll - Transfer1) ; | |
1726 ++Tally ; | |
1727 | |
1728 // Disposition - what might we do with iterator ? | |
1729 // a. add it directly to the EntryList - either tail or head. | |
1730 // b. push it onto the front of the _cxq. | |
1731 // For now we use (a). | |
1732 | |
1733 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; | |
1734 guarantee (iterator->_notified == 0, "invariant") ; | |
1735 iterator->_notified = 1 ; | |
1736 if (Policy != 4) { | |
1737 iterator->TState = ObjectWaiter::TS_ENTER ; | |
1738 } | |
1739 | |
1740 ObjectWaiter * List = _EntryList ; | |
1741 if (List != NULL) { | |
1742 assert (List->_prev == NULL, "invariant") ; | |
1743 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; | |
1744 assert (List != iterator, "invariant") ; | |
1745 } | |
1746 | |
1747 if (Policy == 0) { // prepend to EntryList | |
1748 if (List == NULL) { | |
1749 iterator->_next = iterator->_prev = NULL ; | |
1750 _EntryList = iterator ; | |
1751 } else { | |
1752 List->_prev = iterator ; | |
1753 iterator->_next = List ; | |
1754 iterator->_prev = NULL ; | |
1755 _EntryList = iterator ; | |
1756 } | |
1757 } else | |
1758 if (Policy == 1) { // append to EntryList | |
1759 if (List == NULL) { | |
1760 iterator->_next = iterator->_prev = NULL ; | |
1761 _EntryList = iterator ; | |
1762 } else { | |
1763 // CONSIDER: finding the tail currently requires a linear-time walk of | |
1764 // the EntryList. We can make tail access constant-time by converting to | |
1765 // a CDLL instead of using our current DLL. | |
1766 ObjectWaiter * Tail ; | |
1767 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; | |
1768 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; | |
1769 Tail->_next = iterator ; | |
1770 iterator->_prev = Tail ; | |
1771 iterator->_next = NULL ; | |
1772 } | |
1773 } else | |
1774 if (Policy == 2) { // prepend to cxq | |
1775 // prepend to cxq | |
1776 iterator->TState = ObjectWaiter::TS_CXQ ; | |
1777 for (;;) { | |
1778 ObjectWaiter * Front = _cxq ; | |
1779 iterator->_next = Front ; | |
1780 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { | |
1781 break ; | |
1782 } | |
1783 } | |
1784 } else | |
1785 if (Policy == 3) { // append to cxq | |
1786 iterator->TState = ObjectWaiter::TS_CXQ ; | |
1787 for (;;) { | |
1788 ObjectWaiter * Tail ; | |
1789 Tail = _cxq ; | |
1790 if (Tail == NULL) { | |
1791 iterator->_next = NULL ; | |
1792 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { | |
1793 break ; | |
1794 } | |
1795 } else { | |
1796 while (Tail->_next != NULL) Tail = Tail->_next ; | |
1797 Tail->_next = iterator ; | |
1798 iterator->_prev = Tail ; | |
1799 iterator->_next = NULL ; | |
1800 break ; | |
1801 } | |
1802 } | |
1803 } else { | |
1804 ParkEvent * ev = iterator->_event ; | |
1805 iterator->TState = ObjectWaiter::TS_RUN ; | |
1806 OrderAccess::fence() ; | |
1807 ev->unpark() ; | |
1808 } | |
1809 | |
1810 if (Policy < 4) { | |
1811 iterator->wait_reenter_begin(this); | |
1812 } | |
1813 | |
1814 // _WaitSetLock protects the wait queue, not the EntryList. We could | |
1815 // move the add-to-EntryList operation, above, outside the critical section | |
1816 // protected by _WaitSetLock. In practice that's not useful. With the | |
1817 // exception of wait() timeouts and interrupts the monitor owner | |
1818 // is the only thread that grabs _WaitSetLock. There's almost no contention | |
1819 // on _WaitSetLock so it's not profitable to reduce the length of the | |
1820 // critical section. | |
1821 } | |
1822 | |
1823 Thread::SpinRelease (&_WaitSetLock) ; | |
1824 | |
1825 if (Tally != 0 && ObjectMonitor::_sync_Notifications != NULL) { | |
1826 ObjectMonitor::_sync_Notifications->inc(Tally) ; | |
1827 } | |
1828 } | |
1829 | |
1830 // ----------------------------------------------------------------------------- | |
1831 // Adaptive Spinning Support | |
1832 // | |
1833 // Adaptive spin-then-block - rational spinning | |
1834 // | |
1835 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS | |
1836 // algorithm. On high order SMP systems it would be better to start with | |
1837 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, | |
1838 // a contending thread could enqueue itself on the cxq and then spin locally | |
1839 // on a thread-specific variable such as its ParkEvent._Event flag. | |
1840 // That's left as an exercise for the reader. Note that global spinning is | |
1841 // not problematic on Niagara, as the L2$ serves the interconnect and has both | |
1842 // low latency and massive bandwidth. | |
1843 // | |
1844 // Broadly, we can fix the spin frequency -- that is, the % of contended lock | |
1845 // acquisition attempts where we opt to spin -- at 100% and vary the spin count | |
1846 // (duration) or we can fix the count at approximately the duration of | |
1847 // a context switch and vary the frequency. Of course we could also | |
1848 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. | |
1849 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. | |
1850 // | |
1851 // This implementation varies the duration "D", where D varies with | |
1852 // the success rate of recent spin attempts. (D is capped at approximately | |
1853 // length of a round-trip context switch). The success rate for recent | |
1854 // spin attempts is a good predictor of the success rate of future spin | |
1855 // attempts. The mechanism adapts automatically to varying critical | |
1856 // section length (lock modality), system load and degree of parallelism. | |
1857 // D is maintained per-monitor in _SpinDuration and is initialized | |
1858 // optimistically. Spin frequency is fixed at 100%. | |
1859 // | |
1860 // Note that _SpinDuration is volatile, but we update it without locks | |
1861 // or atomics. The code is designed so that _SpinDuration stays within | |
1862 // a reasonable range even in the presence of races. The arithmetic | |
1863 // operations on _SpinDuration are closed over the domain of legal values, | |
1864 // so at worst a race will install and older but still legal value. | |
1865 // At the very worst this introduces some apparent non-determinism. | |
1866 // We might spin when we shouldn't or vice-versa, but since the spin | |
1867 // count are relatively short, even in the worst case, the effect is harmless. | |
1868 // | |
1869 // Care must be taken that a low "D" value does not become an | |
1870 // an absorbing state. Transient spinning failures -- when spinning | |
1871 // is overall profitable -- should not cause the system to converge | |
1872 // on low "D" values. We want spinning to be stable and predictable | |
1873 // and fairly responsive to change and at the same time we don't want | |
1874 // it to oscillate, become metastable, be "too" non-deterministic, | |
1875 // or converge on or enter undesirable stable absorbing states. | |
1876 // | |
1877 // We implement a feedback-based control system -- using past behavior | |
1878 // to predict future behavior. We face two issues: (a) if the | |
1879 // input signal is random then the spin predictor won't provide optimal | |
1880 // results, and (b) if the signal frequency is too high then the control | |
1881 // system, which has some natural response lag, will "chase" the signal. | |
1882 // (b) can arise from multimodal lock hold times. Transient preemption | |
1883 // can also result in apparent bimodal lock hold times. | |
1884 // Although sub-optimal, neither condition is particularly harmful, as | |
1885 // in the worst-case we'll spin when we shouldn't or vice-versa. | |
1886 // The maximum spin duration is rather short so the failure modes aren't bad. | |
1887 // To be conservative, I've tuned the gain in system to bias toward | |
1888 // _not spinning. Relatedly, the system can sometimes enter a mode where it | |
1889 // "rings" or oscillates between spinning and not spinning. This happens | |
1890 // when spinning is just on the cusp of profitability, however, so the | |
1891 // situation is not dire. The state is benign -- there's no need to add | |
1892 // hysteresis control to damp the transition rate between spinning and | |
1893 // not spinning. | |
1894 // | |
1895 | |
1896 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; | |
1897 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; | |
1898 | |
1899 // Spinning: Fixed frequency (100%), vary duration | |
1900 | |
1901 | |
1902 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { | |
1903 | |
1904 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. | |
1905 int ctr = Knob_FixedSpin ; | |
1906 if (ctr != 0) { | |
1907 while (--ctr >= 0) { | |
1908 if (TryLock (Self) > 0) return 1 ; | |
1909 SpinPause () ; | |
1910 } | |
1911 return 0 ; | |
1912 } | |
1913 | |
1914 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { | |
1915 if (TryLock(Self) > 0) { | |
1916 // Increase _SpinDuration ... | |
1917 // Note that we don't clamp SpinDuration precisely at SpinLimit. | |
1918 // Raising _SpurDuration to the poverty line is key. | |
1919 int x = _SpinDuration ; | |
1920 if (x < Knob_SpinLimit) { | |
1921 if (x < Knob_Poverty) x = Knob_Poverty ; | |
1922 _SpinDuration = x + Knob_BonusB ; | |
1923 } | |
1924 return 1 ; | |
1925 } | |
1926 SpinPause () ; | |
1927 } | |
1928 | |
1929 // Admission control - verify preconditions for spinning | |
1930 // | |
1931 // We always spin a little bit, just to prevent _SpinDuration == 0 from | |
1932 // becoming an absorbing state. Put another way, we spin briefly to | |
1933 // sample, just in case the system load, parallelism, contention, or lock | |
1934 // modality changed. | |
1935 // | |
1936 // Consider the following alternative: | |
1937 // Periodically set _SpinDuration = _SpinLimit and try a long/full | |
1938 // spin attempt. "Periodically" might mean after a tally of | |
1939 // the # of failed spin attempts (or iterations) reaches some threshold. | |
1940 // This takes us into the realm of 1-out-of-N spinning, where we | |
1941 // hold the duration constant but vary the frequency. | |
1942 | |
1943 ctr = _SpinDuration ; | |
1944 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; | |
1945 if (ctr <= 0) return 0 ; | |
1946 | |
1947 if (Knob_SuccRestrict && _succ != NULL) return 0 ; | |
1948 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { | |
1949 TEVENT (Spin abort - notrunnable [TOP]); | |
1950 return 0 ; | |
1951 } | |
1952 | |
1953 int MaxSpin = Knob_MaxSpinners ; | |
1954 if (MaxSpin >= 0) { | |
1955 if (_Spinner > MaxSpin) { | |
1956 TEVENT (Spin abort -- too many spinners) ; | |
1957 return 0 ; | |
1958 } | |
1959 // Slighty racy, but benign ... | |
1960 Adjust (&_Spinner, 1) ; | |
1961 } | |
1962 | |
1963 // We're good to spin ... spin ingress. | |
1964 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades | |
1965 // when preparing to LD...CAS _owner, etc and the CAS is likely | |
1966 // to succeed. | |
1967 int hits = 0 ; | |
1968 int msk = 0 ; | |
1969 int caspty = Knob_CASPenalty ; | |
1970 int oxpty = Knob_OXPenalty ; | |
1971 int sss = Knob_SpinSetSucc ; | |
1972 if (sss && _succ == NULL ) _succ = Self ; | |
1973 Thread * prv = NULL ; | |
1974 | |
1975 // There are three ways to exit the following loop: | |
1976 // 1. A successful spin where this thread has acquired the lock. | |
1977 // 2. Spin failure with prejudice | |
1978 // 3. Spin failure without prejudice | |
1979 | |
1980 while (--ctr >= 0) { | |
1981 | |
1982 // Periodic polling -- Check for pending GC | |
1983 // Threads may spin while they're unsafe. | |
1984 // We don't want spinning threads to delay the JVM from reaching | |
1985 // a stop-the-world safepoint or to steal cycles from GC. | |
1986 // If we detect a pending safepoint we abort in order that | |
1987 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) | |
1988 // this thread, if safe, doesn't steal cycles from GC. | |
1989 // This is in keeping with the "no loitering in runtime" rule. | |
1990 // We periodically check to see if there's a safepoint pending. | |
1991 if ((ctr & 0xFF) == 0) { | |
1992 if (SafepointSynchronize::do_call_back()) { | |
1993 TEVENT (Spin: safepoint) ; | |
1994 goto Abort ; // abrupt spin egress | |
1995 } | |
1996 if (Knob_UsePause & 1) SpinPause () ; | |
1997 | |
1998 int (*scb)(intptr_t,int) = SpinCallbackFunction ; | |
1999 if (hits > 50 && scb != NULL) { | |
2000 int abend = (*scb)(SpinCallbackArgument, 0) ; | |
2001 } | |
2002 } | |
2003 | |
2004 if (Knob_UsePause & 2) SpinPause() ; | |
2005 | |
2006 // Exponential back-off ... Stay off the bus to reduce coherency traffic. | |
2007 // This is useful on classic SMP systems, but is of less utility on | |
2008 // N1-style CMT platforms. | |
2009 // | |
2010 // Trade-off: lock acquisition latency vs coherency bandwidth. | |
2011 // Lock hold times are typically short. A histogram | |
2012 // of successful spin attempts shows that we usually acquire | |
2013 // the lock early in the spin. That suggests we want to | |
2014 // sample _owner frequently in the early phase of the spin, | |
2015 // but then back-off and sample less frequently as the spin | |
2016 // progresses. The back-off makes a good citizen on SMP big | |
2017 // SMP systems. Oversampling _owner can consume excessive | |
2018 // coherency bandwidth. Relatedly, if we _oversample _owner we | |
2019 // can inadvertently interfere with the the ST m->owner=null. | |
2020 // executed by the lock owner. | |
2021 if (ctr & msk) continue ; | |
2022 ++hits ; | |
2023 if ((hits & 0xF) == 0) { | |
2024 // The 0xF, above, corresponds to the exponent. | |
2025 // Consider: (msk+1)|msk | |
2026 msk = ((msk << 2)|3) & BackOffMask ; | |
2027 } | |
2028 | |
2029 // Probe _owner with TATAS | |
2030 // If this thread observes the monitor transition or flicker | |
2031 // from locked to unlocked to locked, then the odds that this | |
2032 // thread will acquire the lock in this spin attempt go down | |
2033 // considerably. The same argument applies if the CAS fails | |
2034 // or if we observe _owner change from one non-null value to | |
2035 // another non-null value. In such cases we might abort | |
2036 // the spin without prejudice or apply a "penalty" to the | |
2037 // spin count-down variable "ctr", reducing it by 100, say. | |
2038 | |
2039 Thread * ox = (Thread *) _owner ; | |
2040 if (ox == NULL) { | |
2041 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; | |
2042 if (ox == NULL) { | |
2043 // The CAS succeeded -- this thread acquired ownership | |
2044 // Take care of some bookkeeping to exit spin state. | |
2045 if (sss && _succ == Self) { | |
2046 _succ = NULL ; | |
2047 } | |
2048 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; | |
2049 | |
2050 // Increase _SpinDuration : | |
2051 // The spin was successful (profitable) so we tend toward | |
2052 // longer spin attempts in the future. | |
2053 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. | |
2054 // If we acquired the lock early in the spin cycle it | |
2055 // makes sense to increase _SpinDuration proportionally. | |
2056 // Note that we don't clamp SpinDuration precisely at SpinLimit. | |
2057 int x = _SpinDuration ; | |
2058 if (x < Knob_SpinLimit) { | |
2059 if (x < Knob_Poverty) x = Knob_Poverty ; | |
2060 _SpinDuration = x + Knob_Bonus ; | |
2061 } | |
2062 return 1 ; | |
2063 } | |
2064 | |
2065 // The CAS failed ... we can take any of the following actions: | |
2066 // * penalize: ctr -= Knob_CASPenalty | |
2067 // * exit spin with prejudice -- goto Abort; | |
2068 // * exit spin without prejudice. | |
2069 // * Since CAS is high-latency, retry again immediately. | |
2070 prv = ox ; | |
2071 TEVENT (Spin: cas failed) ; | |
2072 if (caspty == -2) break ; | |
2073 if (caspty == -1) goto Abort ; | |
2074 ctr -= caspty ; | |
2075 continue ; | |
2076 } | |
2077 | |
2078 // Did lock ownership change hands ? | |
2079 if (ox != prv && prv != NULL ) { | |
2080 TEVENT (spin: Owner changed) | |
2081 if (oxpty == -2) break ; | |
2082 if (oxpty == -1) goto Abort ; | |
2083 ctr -= oxpty ; | |
2084 } | |
2085 prv = ox ; | |
2086 | |
2087 // Abort the spin if the owner is not executing. | |
2088 // The owner must be executing in order to drop the lock. | |
2089 // Spinning while the owner is OFFPROC is idiocy. | |
2090 // Consider: ctr -= RunnablePenalty ; | |
2091 if (Knob_OState && NotRunnable (Self, ox)) { | |
2092 TEVENT (Spin abort - notrunnable); | |
2093 goto Abort ; | |
2094 } | |
2095 if (sss && _succ == NULL ) _succ = Self ; | |
2096 } | |
2097 | |
2098 // Spin failed with prejudice -- reduce _SpinDuration. | |
2099 // TODO: Use an AIMD-like policy to adjust _SpinDuration. | |
2100 // AIMD is globally stable. | |
2101 TEVENT (Spin failure) ; | |
2102 { | |
2103 int x = _SpinDuration ; | |
2104 if (x > 0) { | |
2105 // Consider an AIMD scheme like: x -= (x >> 3) + 100 | |
2106 // This is globally sample and tends to damp the response. | |
2107 x -= Knob_Penalty ; | |
2108 if (x < 0) x = 0 ; | |
2109 _SpinDuration = x ; | |
2110 } | |
2111 } | |
2112 | |
2113 Abort: | |
2114 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; | |
2115 if (sss && _succ == Self) { | |
2116 _succ = NULL ; | |
2117 // Invariant: after setting succ=null a contending thread | |
2118 // must recheck-retry _owner before parking. This usually happens | |
2119 // in the normal usage of TrySpin(), but it's safest | |
2120 // to make TrySpin() as foolproof as possible. | |
2121 OrderAccess::fence() ; | |
2122 if (TryLock(Self) > 0) return 1 ; | |
2123 } | |
2124 return 0 ; | |
2125 } | |
2126 | |
2127 // NotRunnable() -- informed spinning | |
2128 // | |
2129 // Don't bother spinning if the owner is not eligible to drop the lock. | |
2130 // Peek at the owner's schedctl.sc_state and Thread._thread_values and | |
2131 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. | |
2132 // The thread must be runnable in order to drop the lock in timely fashion. | |
2133 // If the _owner is not runnable then spinning will not likely be | |
2134 // successful (profitable). | |
2135 // | |
2136 // Beware -- the thread referenced by _owner could have died | |
2137 // so a simply fetch from _owner->_thread_state might trap. | |
2138 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. | |
2139 // Because of the lifecycle issues the schedctl and _thread_state values | |
2140 // observed by NotRunnable() might be garbage. NotRunnable must | |
2141 // tolerate this and consider the observed _thread_state value | |
2142 // as advisory. | |
2143 // | |
2144 // Beware too, that _owner is sometimes a BasicLock address and sometimes | |
2145 // a thread pointer. We differentiate the two cases with OwnerIsThread. | |
2146 // Alternately, we might tag the type (thread pointer vs basiclock pointer) | |
2147 // with the LSB of _owner. Another option would be to probablistically probe | |
2148 // the putative _owner->TypeTag value. | |
2149 // | |
2150 // Checking _thread_state isn't perfect. Even if the thread is | |
2151 // in_java it might be blocked on a page-fault or have been preempted | |
2152 // and sitting on a ready/dispatch queue. _thread state in conjunction | |
2153 // with schedctl.sc_state gives us a good picture of what the | |
2154 // thread is doing, however. | |
2155 // | |
2156 // TODO: check schedctl.sc_state. | |
2157 // We'll need to use SafeFetch32() to read from the schedctl block. | |
2158 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ | |
2159 // | |
2160 // The return value from NotRunnable() is *advisory* -- the | |
2161 // result is based on sampling and is not necessarily coherent. | |
2162 // The caller must tolerate false-negative and false-positive errors. | |
2163 // Spinning, in general, is probabilistic anyway. | |
2164 | |
2165 | |
2166 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { | |
2167 // Check either OwnerIsThread or ox->TypeTag == 2BAD. | |
2168 if (!OwnerIsThread) return 0 ; | |
2169 | |
2170 if (ox == NULL) return 0 ; | |
2171 | |
2172 // Avoid transitive spinning ... | |
2173 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. | |
2174 // Immediately after T1 acquires L it's possible that T2, also | |
2175 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. | |
2176 // This occurs transiently after T1 acquired L but before | |
2177 // T1 managed to clear T1.Stalled. T2 does not need to abort | |
2178 // its spin in this circumstance. | |
2179 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; | |
2180 | |
2181 if (BlockedOn == 1) return 1 ; | |
2182 if (BlockedOn != 0) { | |
2183 return BlockedOn != intptr_t(this) && _owner == ox ; | |
2184 } | |
2185 | |
2186 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; | |
2187 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; | |
2188 // consider also: jst != _thread_in_Java -- but that's overspecific. | |
2189 return jst == _thread_blocked || jst == _thread_in_native ; | |
2190 } | |
2191 | |
2192 | |
2193 // ----------------------------------------------------------------------------- | |
2194 // WaitSet management ... | |
2195 | |
2196 ObjectWaiter::ObjectWaiter(Thread* thread) { | |
2197 _next = NULL; | |
2198 _prev = NULL; | |
2199 _notified = 0; | |
2200 TState = TS_RUN ; | |
2201 _thread = thread; | |
2202 _event = thread->_ParkEvent ; | |
2203 _active = false; | |
2204 assert (_event != NULL, "invariant") ; | |
2205 } | |
2206 | |
2207 void ObjectWaiter::wait_reenter_begin(ObjectMonitor *mon) { | |
2208 JavaThread *jt = (JavaThread *)this->_thread; | |
2209 _active = JavaThreadBlockedOnMonitorEnterState::wait_reenter_begin(jt, mon); | |
2210 } | |
2211 | |
2212 void ObjectWaiter::wait_reenter_end(ObjectMonitor *mon) { | |
2213 JavaThread *jt = (JavaThread *)this->_thread; | |
2214 JavaThreadBlockedOnMonitorEnterState::wait_reenter_end(jt, _active); | |
2215 } | |
2216 | |
2217 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { | |
2218 assert(node != NULL, "should not dequeue NULL node"); | |
2219 assert(node->_prev == NULL, "node already in list"); | |
2220 assert(node->_next == NULL, "node already in list"); | |
2221 // put node at end of queue (circular doubly linked list) | |
2222 if (_WaitSet == NULL) { | |
2223 _WaitSet = node; | |
2224 node->_prev = node; | |
2225 node->_next = node; | |
2226 } else { | |
2227 ObjectWaiter* head = _WaitSet ; | |
2228 ObjectWaiter* tail = head->_prev; | |
2229 assert(tail->_next == head, "invariant check"); | |
2230 tail->_next = node; | |
2231 head->_prev = node; | |
2232 node->_next = head; | |
2233 node->_prev = tail; | |
2234 } | |
2235 } | |
2236 | |
2237 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { | |
2238 // dequeue the very first waiter | |
2239 ObjectWaiter* waiter = _WaitSet; | |
2240 if (waiter) { | |
2241 DequeueSpecificWaiter(waiter); | |
2242 } | |
2243 return waiter; | |
2244 } | |
2245 | |
2246 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { | |
2247 assert(node != NULL, "should not dequeue NULL node"); | |
2248 assert(node->_prev != NULL, "node already removed from list"); | |
2249 assert(node->_next != NULL, "node already removed from list"); | |
2250 // when the waiter has woken up because of interrupt, | |
2251 // timeout or other spurious wake-up, dequeue the | |
2252 // waiter from waiting list | |
2253 ObjectWaiter* next = node->_next; | |
2254 if (next == node) { | |
2255 assert(node->_prev == node, "invariant check"); | |
2256 _WaitSet = NULL; | |
2257 } else { | |
2258 ObjectWaiter* prev = node->_prev; | |
2259 assert(prev->_next == node, "invariant check"); | |
2260 assert(next->_prev == node, "invariant check"); | |
2261 next->_prev = prev; | |
2262 prev->_next = next; | |
2263 if (_WaitSet == node) { | |
2264 _WaitSet = next; | |
2265 } | |
2266 } | |
2267 node->_next = NULL; | |
2268 node->_prev = NULL; | |
2269 } | |
2270 | |
2271 // ----------------------------------------------------------------------------- | |
2272 // PerfData support | |
2273 PerfCounter * ObjectMonitor::_sync_ContendedLockAttempts = NULL ; | |
2274 PerfCounter * ObjectMonitor::_sync_FutileWakeups = NULL ; | |
2275 PerfCounter * ObjectMonitor::_sync_Parks = NULL ; | |
2276 PerfCounter * ObjectMonitor::_sync_EmptyNotifications = NULL ; | |
2277 PerfCounter * ObjectMonitor::_sync_Notifications = NULL ; | |
2278 PerfCounter * ObjectMonitor::_sync_PrivateA = NULL ; | |
2279 PerfCounter * ObjectMonitor::_sync_PrivateB = NULL ; | |
2280 PerfCounter * ObjectMonitor::_sync_SlowExit = NULL ; | |
2281 PerfCounter * ObjectMonitor::_sync_SlowEnter = NULL ; | |
2282 PerfCounter * ObjectMonitor::_sync_SlowNotify = NULL ; | |
2283 PerfCounter * ObjectMonitor::_sync_SlowNotifyAll = NULL ; | |
2284 PerfCounter * ObjectMonitor::_sync_FailedSpins = NULL ; | |
2285 PerfCounter * ObjectMonitor::_sync_SuccessfulSpins = NULL ; | |
2286 PerfCounter * ObjectMonitor::_sync_MonInCirculation = NULL ; | |
2287 PerfCounter * ObjectMonitor::_sync_MonScavenged = NULL ; | |
2288 PerfCounter * ObjectMonitor::_sync_Inflations = NULL ; | |
2289 PerfCounter * ObjectMonitor::_sync_Deflations = NULL ; | |
2290 PerfLongVariable * ObjectMonitor::_sync_MonExtant = NULL ; | |
2291 | |
2292 // One-shot global initialization for the sync subsystem. | |
2293 // We could also defer initialization and initialize on-demand | |
2294 // the first time we call inflate(). Initialization would | |
2295 // be protected - like so many things - by the MonitorCache_lock. | |
2296 | |
2297 void ObjectMonitor::Initialize () { | |
2298 static int InitializationCompleted = 0 ; | |
2299 assert (InitializationCompleted == 0, "invariant") ; | |
2300 InitializationCompleted = 1 ; | |
2301 if (UsePerfData) { | |
2302 EXCEPTION_MARK ; | |
2303 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } | |
2304 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } | |
2305 NEWPERFCOUNTER(_sync_Inflations) ; | |
2306 NEWPERFCOUNTER(_sync_Deflations) ; | |
2307 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; | |
2308 NEWPERFCOUNTER(_sync_FutileWakeups) ; | |
2309 NEWPERFCOUNTER(_sync_Parks) ; | |
2310 NEWPERFCOUNTER(_sync_EmptyNotifications) ; | |
2311 NEWPERFCOUNTER(_sync_Notifications) ; | |
2312 NEWPERFCOUNTER(_sync_SlowEnter) ; | |
2313 NEWPERFCOUNTER(_sync_SlowExit) ; | |
2314 NEWPERFCOUNTER(_sync_SlowNotify) ; | |
2315 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; | |
2316 NEWPERFCOUNTER(_sync_FailedSpins) ; | |
2317 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; | |
2318 NEWPERFCOUNTER(_sync_PrivateA) ; | |
2319 NEWPERFCOUNTER(_sync_PrivateB) ; | |
2320 NEWPERFCOUNTER(_sync_MonInCirculation) ; | |
2321 NEWPERFCOUNTER(_sync_MonScavenged) ; | |
2322 NEWPERFVARIABLE(_sync_MonExtant) ; | |
2323 #undef NEWPERFCOUNTER | |
2324 } | |
2325 } | |
2326 | |
2327 | |
2328 // Compile-time asserts | |
2329 // When possible, it's better to catch errors deterministically at | |
2330 // compile-time than at runtime. The down-side to using compile-time | |
2331 // asserts is that error message -- often something about negative array | |
2332 // indices -- is opaque. | |
2333 | |
2334 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } | |
2335 | |
2336 void ObjectMonitor::ctAsserts() { | |
2337 CTASSERT(offset_of (ObjectMonitor, _header) == 0); | |
2338 } | |
2339 | |
2340 | |
2341 static char * kvGet (char * kvList, const char * Key) { | |
2342 if (kvList == NULL) return NULL ; | |
2343 size_t n = strlen (Key) ; | |
2344 char * Search ; | |
2345 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { | |
2346 if (strncmp (Search, Key, n) == 0) { | |
2347 if (Search[n] == '=') return Search + n + 1 ; | |
2348 if (Search[n] == 0) return (char *) "1" ; | |
2349 } | |
2350 } | |
2351 return NULL ; | |
2352 } | |
2353 | |
2354 static int kvGetInt (char * kvList, const char * Key, int Default) { | |
2355 char * v = kvGet (kvList, Key) ; | |
2356 int rslt = v ? ::strtol (v, NULL, 0) : Default ; | |
2357 if (Knob_ReportSettings && v != NULL) { | |
2358 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; | |
2359 ::fflush (stdout) ; | |
2360 } | |
2361 return rslt ; | |
2362 } | |
2363 | |
2364 void ObjectMonitor::DeferredInitialize () { | |
2365 if (InitDone > 0) return ; | |
2366 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { | |
2367 while (InitDone != 1) ; | |
2368 return ; | |
2369 } | |
2370 | |
2371 // One-shot global initialization ... | |
2372 // The initialization is idempotent, so we don't need locks. | |
2373 // In the future consider doing this via os::init_2(). | |
2374 // SyncKnobs consist of <Key>=<Value> pairs in the style | |
2375 // of environment variables. Start by converting ':' to NUL. | |
2376 | |
2377 if (SyncKnobs == NULL) SyncKnobs = "" ; | |
2378 | |
2379 size_t sz = strlen (SyncKnobs) ; | |
2380 char * knobs = (char *) malloc (sz + 2) ; | |
2381 if (knobs == NULL) { | |
2382 vm_exit_out_of_memory (sz + 2, "Parse SyncKnobs") ; | |
2383 guarantee (0, "invariant") ; | |
2384 } | |
2385 strcpy (knobs, SyncKnobs) ; | |
2386 knobs[sz+1] = 0 ; | |
2387 for (char * p = knobs ; *p ; p++) { | |
2388 if (*p == ':') *p = 0 ; | |
2389 } | |
2390 | |
2391 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } | |
2392 SETKNOB(ReportSettings) ; | |
2393 SETKNOB(Verbose) ; | |
2394 SETKNOB(FixedSpin) ; | |
2395 SETKNOB(SpinLimit) ; | |
2396 SETKNOB(SpinBase) ; | |
2397 SETKNOB(SpinBackOff); | |
2398 SETKNOB(CASPenalty) ; | |
2399 SETKNOB(OXPenalty) ; | |
2400 SETKNOB(LogSpins) ; | |
2401 SETKNOB(SpinSetSucc) ; | |
2402 SETKNOB(SuccEnabled) ; | |
2403 SETKNOB(SuccRestrict) ; | |
2404 SETKNOB(Penalty) ; | |
2405 SETKNOB(Bonus) ; | |
2406 SETKNOB(BonusB) ; | |
2407 SETKNOB(Poverty) ; | |
2408 SETKNOB(SpinAfterFutile) ; | |
2409 SETKNOB(UsePause) ; | |
2410 SETKNOB(SpinEarly) ; | |
2411 SETKNOB(OState) ; | |
2412 SETKNOB(MaxSpinners) ; | |
2413 SETKNOB(PreSpin) ; | |
2414 SETKNOB(ExitPolicy) ; | |
2415 SETKNOB(QMode); | |
2416 SETKNOB(ResetEvent) ; | |
2417 SETKNOB(MoveNotifyee) ; | |
2418 SETKNOB(FastHSSEC) ; | |
2419 #undef SETKNOB | |
2420 | |
2421 if (os::is_MP()) { | |
2422 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; | |
2423 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; | |
2424 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) | |
2425 } else { | |
2426 Knob_SpinLimit = 0 ; | |
2427 Knob_SpinBase = 0 ; | |
2428 Knob_PreSpin = 0 ; | |
2429 Knob_FixedSpin = -1 ; | |
2430 } | |
2431 | |
2432 if (Knob_LogSpins == 0) { | |
2433 ObjectMonitor::_sync_FailedSpins = NULL ; | |
2434 } | |
2435 | |
2436 free (knobs) ; | |
2437 OrderAccess::fence() ; | |
2438 InitDone = 1 ; | |
2439 } | |
2440 | |
2441 #ifndef PRODUCT | |
2442 void ObjectMonitor::verify() { | |
2443 } | |
2444 | |
2445 void ObjectMonitor::print() { | |
2446 } | |
2447 #endif |