185 } ; |
411 } ; |
186 |
412 |
187 static SharedGlobals GVars ; |
413 static SharedGlobals GVars ; |
188 static int MonitorScavengeThreshold = 1000000 ; |
414 static int MonitorScavengeThreshold = 1000000 ; |
189 static volatile int ForceMonitorScavenge = 0 ; // Scavenge required and pending |
415 static volatile int ForceMonitorScavenge = 0 ; // Scavenge required and pending |
190 |
|
191 |
|
192 // Tunables ... |
|
193 // The knob* variables are effectively final. Once set they should |
|
194 // never be modified hence. Consider using __read_mostly with GCC. |
|
195 |
|
196 static int Knob_LogSpins = 0 ; // enable jvmstat tally for spins |
|
197 static int Knob_HandOff = 0 ; |
|
198 static int Knob_Verbose = 0 ; |
|
199 static int Knob_ReportSettings = 0 ; |
|
200 |
|
201 static int Knob_SpinLimit = 5000 ; // derived by an external tool - |
|
202 static int Knob_SpinBase = 0 ; // Floor AKA SpinMin |
|
203 static int Knob_SpinBackOff = 0 ; // spin-loop backoff |
|
204 static int Knob_CASPenalty = -1 ; // Penalty for failed CAS |
|
205 static int Knob_OXPenalty = -1 ; // Penalty for observed _owner change |
|
206 static int Knob_SpinSetSucc = 1 ; // spinners set the _succ field |
|
207 static int Knob_SpinEarly = 1 ; |
|
208 static int Knob_SuccEnabled = 1 ; // futile wake throttling |
|
209 static int Knob_SuccRestrict = 0 ; // Limit successors + spinners to at-most-one |
|
210 static int Knob_MaxSpinners = -1 ; // Should be a function of # CPUs |
|
211 static int Knob_Bonus = 100 ; // spin success bonus |
|
212 static int Knob_BonusB = 100 ; // spin success bonus |
|
213 static int Knob_Penalty = 200 ; // spin failure penalty |
|
214 static int Knob_Poverty = 1000 ; |
|
215 static int Knob_SpinAfterFutile = 1 ; // Spin after returning from park() |
|
216 static int Knob_FixedSpin = 0 ; |
|
217 static int Knob_OState = 3 ; // Spinner checks thread state of _owner |
|
218 static int Knob_UsePause = 1 ; |
|
219 static int Knob_ExitPolicy = 0 ; |
|
220 static int Knob_PreSpin = 10 ; // 20-100 likely better |
|
221 static int Knob_ResetEvent = 0 ; |
|
222 static int BackOffMask = 0 ; |
|
223 |
|
224 static int Knob_FastHSSEC = 0 ; |
|
225 static int Knob_MoveNotifyee = 2 ; // notify() - disposition of notifyee |
|
226 static int Knob_QMode = 0 ; // EntryList-cxq policy - queue discipline |
|
227 static volatile int InitDone = 0 ; |
|
228 |
|
229 |
|
230 // hashCode() generation : |
|
231 // |
|
232 // Possibilities: |
|
233 // * MD5Digest of {obj,stwRandom} |
|
234 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. |
|
235 // * A DES- or AES-style SBox[] mechanism |
|
236 // * One of the Phi-based schemes, such as: |
|
237 // 2654435761 = 2^32 * Phi (golden ratio) |
|
238 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; |
|
239 // * A variation of Marsaglia's shift-xor RNG scheme. |
|
240 // * (obj ^ stwRandom) is appealing, but can result |
|
241 // in undesirable regularity in the hashCode values of adjacent objects |
|
242 // (objects allocated back-to-back, in particular). This could potentially |
|
243 // result in hashtable collisions and reduced hashtable efficiency. |
|
244 // There are simple ways to "diffuse" the middle address bits over the |
|
245 // generated hashCode values: |
|
246 // |
|
247 |
|
248 static inline intptr_t get_next_hash(Thread * Self, oop obj) { |
|
249 intptr_t value = 0 ; |
|
250 if (hashCode == 0) { |
|
251 // This form uses an unguarded global Park-Miller RNG, |
|
252 // so it's possible for two threads to race and generate the same RNG. |
|
253 // On MP system we'll have lots of RW access to a global, so the |
|
254 // mechanism induces lots of coherency traffic. |
|
255 value = os::random() ; |
|
256 } else |
|
257 if (hashCode == 1) { |
|
258 // This variation has the property of being stable (idempotent) |
|
259 // between STW operations. This can be useful in some of the 1-0 |
|
260 // synchronization schemes. |
|
261 intptr_t addrBits = intptr_t(obj) >> 3 ; |
|
262 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom ; |
|
263 } else |
|
264 if (hashCode == 2) { |
|
265 value = 1 ; // for sensitivity testing |
|
266 } else |
|
267 if (hashCode == 3) { |
|
268 value = ++GVars.hcSequence ; |
|
269 } else |
|
270 if (hashCode == 4) { |
|
271 value = intptr_t(obj) ; |
|
272 } else { |
|
273 // Marsaglia's xor-shift scheme with thread-specific state |
|
274 // This is probably the best overall implementation -- we'll |
|
275 // likely make this the default in future releases. |
|
276 unsigned t = Self->_hashStateX ; |
|
277 t ^= (t << 11) ; |
|
278 Self->_hashStateX = Self->_hashStateY ; |
|
279 Self->_hashStateY = Self->_hashStateZ ; |
|
280 Self->_hashStateZ = Self->_hashStateW ; |
|
281 unsigned v = Self->_hashStateW ; |
|
282 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)) ; |
|
283 Self->_hashStateW = v ; |
|
284 value = v ; |
|
285 } |
|
286 |
|
287 value &= markOopDesc::hash_mask; |
|
288 if (value == 0) value = 0xBAD ; |
|
289 assert (value != markOopDesc::no_hash, "invariant") ; |
|
290 TEVENT (hashCode: GENERATE) ; |
|
291 return value; |
|
292 } |
|
293 |
|
294 void BasicLock::print_on(outputStream* st) const { |
|
295 st->print("monitor"); |
|
296 } |
|
297 |
|
298 void BasicLock::move_to(oop obj, BasicLock* dest) { |
|
299 // Check to see if we need to inflate the lock. This is only needed |
|
300 // if an object is locked using "this" lightweight monitor. In that |
|
301 // case, the displaced_header() is unlocked, because the |
|
302 // displaced_header() contains the header for the originally unlocked |
|
303 // object. However the object could have already been inflated. But it |
|
304 // does not matter, the inflation will just a no-op. For other cases, |
|
305 // the displaced header will be either 0x0 or 0x3, which are location |
|
306 // independent, therefore the BasicLock is free to move. |
|
307 // |
|
308 // During OSR we may need to relocate a BasicLock (which contains a |
|
309 // displaced word) from a location in an interpreter frame to a |
|
310 // new location in a compiled frame. "this" refers to the source |
|
311 // basiclock in the interpreter frame. "dest" refers to the destination |
|
312 // basiclock in the new compiled frame. We *always* inflate in move_to(). |
|
313 // The always-Inflate policy works properly, but in 1.5.0 it can sometimes |
|
314 // cause performance problems in code that makes heavy use of a small # of |
|
315 // uncontended locks. (We'd inflate during OSR, and then sync performance |
|
316 // would subsequently plummet because the thread would be forced thru the slow-path). |
|
317 // This problem has been made largely moot on IA32 by inlining the inflated fast-path |
|
318 // operations in Fast_Lock and Fast_Unlock in i486.ad. |
|
319 // |
|
320 // Note that there is a way to safely swing the object's markword from |
|
321 // one stack location to another. This avoids inflation. Obviously, |
|
322 // we need to ensure that both locations refer to the current thread's stack. |
|
323 // There are some subtle concurrency issues, however, and since the benefit is |
|
324 // is small (given the support for inflated fast-path locking in the fast_lock, etc) |
|
325 // we'll leave that optimization for another time. |
|
326 |
|
327 if (displaced_header()->is_neutral()) { |
|
328 ObjectSynchronizer::inflate_helper(obj); |
|
329 // WARNING: We can not put check here, because the inflation |
|
330 // will not update the displaced header. Once BasicLock is inflated, |
|
331 // no one should ever look at its content. |
|
332 } else { |
|
333 // Typically the displaced header will be 0 (recursive stack lock) or |
|
334 // unused_mark. Naively we'd like to assert that the displaced mark |
|
335 // value is either 0, neutral, or 3. But with the advent of the |
|
336 // store-before-CAS avoidance in fast_lock/compiler_lock_object |
|
337 // we can find any flavor mark in the displaced mark. |
|
338 } |
|
339 // [RGV] The next line appears to do nothing! |
|
340 intptr_t dh = (intptr_t) displaced_header(); |
|
341 dest->set_displaced_header(displaced_header()); |
|
342 } |
|
343 |
|
344 // ----------------------------------------------------------------------------- |
|
345 |
|
346 // standard constructor, allows locking failures |
|
347 ObjectLocker::ObjectLocker(Handle obj, Thread* thread, bool doLock) { |
|
348 _dolock = doLock; |
|
349 _thread = thread; |
|
350 debug_only(if (StrictSafepointChecks) _thread->check_for_valid_safepoint_state(false);) |
|
351 _obj = obj; |
|
352 |
|
353 if (_dolock) { |
|
354 TEVENT (ObjectLocker) ; |
|
355 |
|
356 ObjectSynchronizer::fast_enter(_obj, &_lock, false, _thread); |
|
357 } |
|
358 } |
|
359 |
|
360 ObjectLocker::~ObjectLocker() { |
|
361 if (_dolock) { |
|
362 ObjectSynchronizer::fast_exit(_obj(), &_lock, _thread); |
|
363 } |
|
364 } |
|
365 |
|
366 // ----------------------------------------------------------------------------- |
|
367 |
|
368 |
|
369 PerfCounter * ObjectSynchronizer::_sync_Inflations = NULL ; |
|
370 PerfCounter * ObjectSynchronizer::_sync_Deflations = NULL ; |
|
371 PerfCounter * ObjectSynchronizer::_sync_ContendedLockAttempts = NULL ; |
|
372 PerfCounter * ObjectSynchronizer::_sync_FutileWakeups = NULL ; |
|
373 PerfCounter * ObjectSynchronizer::_sync_Parks = NULL ; |
|
374 PerfCounter * ObjectSynchronizer::_sync_EmptyNotifications = NULL ; |
|
375 PerfCounter * ObjectSynchronizer::_sync_Notifications = NULL ; |
|
376 PerfCounter * ObjectSynchronizer::_sync_PrivateA = NULL ; |
|
377 PerfCounter * ObjectSynchronizer::_sync_PrivateB = NULL ; |
|
378 PerfCounter * ObjectSynchronizer::_sync_SlowExit = NULL ; |
|
379 PerfCounter * ObjectSynchronizer::_sync_SlowEnter = NULL ; |
|
380 PerfCounter * ObjectSynchronizer::_sync_SlowNotify = NULL ; |
|
381 PerfCounter * ObjectSynchronizer::_sync_SlowNotifyAll = NULL ; |
|
382 PerfCounter * ObjectSynchronizer::_sync_FailedSpins = NULL ; |
|
383 PerfCounter * ObjectSynchronizer::_sync_SuccessfulSpins = NULL ; |
|
384 PerfCounter * ObjectSynchronizer::_sync_MonInCirculation = NULL ; |
|
385 PerfCounter * ObjectSynchronizer::_sync_MonScavenged = NULL ; |
|
386 PerfLongVariable * ObjectSynchronizer::_sync_MonExtant = NULL ; |
|
387 |
|
388 // One-shot global initialization for the sync subsystem. |
|
389 // We could also defer initialization and initialize on-demand |
|
390 // the first time we call inflate(). Initialization would |
|
391 // be protected - like so many things - by the MonitorCache_lock. |
|
392 |
|
393 void ObjectSynchronizer::Initialize () { |
|
394 static int InitializationCompleted = 0 ; |
|
395 assert (InitializationCompleted == 0, "invariant") ; |
|
396 InitializationCompleted = 1 ; |
|
397 if (UsePerfData) { |
|
398 EXCEPTION_MARK ; |
|
399 #define NEWPERFCOUNTER(n) {n = PerfDataManager::create_counter(SUN_RT, #n, PerfData::U_Events,CHECK); } |
|
400 #define NEWPERFVARIABLE(n) {n = PerfDataManager::create_variable(SUN_RT, #n, PerfData::U_Events,CHECK); } |
|
401 NEWPERFCOUNTER(_sync_Inflations) ; |
|
402 NEWPERFCOUNTER(_sync_Deflations) ; |
|
403 NEWPERFCOUNTER(_sync_ContendedLockAttempts) ; |
|
404 NEWPERFCOUNTER(_sync_FutileWakeups) ; |
|
405 NEWPERFCOUNTER(_sync_Parks) ; |
|
406 NEWPERFCOUNTER(_sync_EmptyNotifications) ; |
|
407 NEWPERFCOUNTER(_sync_Notifications) ; |
|
408 NEWPERFCOUNTER(_sync_SlowEnter) ; |
|
409 NEWPERFCOUNTER(_sync_SlowExit) ; |
|
410 NEWPERFCOUNTER(_sync_SlowNotify) ; |
|
411 NEWPERFCOUNTER(_sync_SlowNotifyAll) ; |
|
412 NEWPERFCOUNTER(_sync_FailedSpins) ; |
|
413 NEWPERFCOUNTER(_sync_SuccessfulSpins) ; |
|
414 NEWPERFCOUNTER(_sync_PrivateA) ; |
|
415 NEWPERFCOUNTER(_sync_PrivateB) ; |
|
416 NEWPERFCOUNTER(_sync_MonInCirculation) ; |
|
417 NEWPERFCOUNTER(_sync_MonScavenged) ; |
|
418 NEWPERFVARIABLE(_sync_MonExtant) ; |
|
419 #undef NEWPERFCOUNTER |
|
420 } |
|
421 } |
|
422 |
|
423 // Compile-time asserts |
|
424 // When possible, it's better to catch errors deterministically at |
|
425 // compile-time than at runtime. The down-side to using compile-time |
|
426 // asserts is that error message -- often something about negative array |
|
427 // indices -- is opaque. |
|
428 |
|
429 #define CTASSERT(x) { int tag[1-(2*!(x))]; printf ("Tag @" INTPTR_FORMAT "\n", (intptr_t)tag); } |
|
430 |
|
431 void ObjectMonitor::ctAsserts() { |
|
432 CTASSERT(offset_of (ObjectMonitor, _header) == 0); |
|
433 } |
|
434 |
|
435 static int Adjust (volatile int * adr, int dx) { |
|
436 int v ; |
|
437 for (v = *adr ; Atomic::cmpxchg (v + dx, adr, v) != v; v = *adr) ; |
|
438 return v ; |
|
439 } |
|
440 |
|
441 // Ad-hoc mutual exclusion primitives: SpinLock and Mux |
|
442 // |
|
443 // We employ SpinLocks _only for low-contention, fixed-length |
|
444 // short-duration critical sections where we're concerned |
|
445 // about native mutex_t or HotSpot Mutex:: latency. |
|
446 // The mux construct provides a spin-then-block mutual exclusion |
|
447 // mechanism. |
|
448 // |
|
449 // Testing has shown that contention on the ListLock guarding gFreeList |
|
450 // is common. If we implement ListLock as a simple SpinLock it's common |
|
451 // for the JVM to devolve to yielding with little progress. This is true |
|
452 // despite the fact that the critical sections protected by ListLock are |
|
453 // extremely short. |
|
454 // |
|
455 // TODO-FIXME: ListLock should be of type SpinLock. |
|
456 // We should make this a 1st-class type, integrated into the lock |
|
457 // hierarchy as leaf-locks. Critically, the SpinLock structure |
|
458 // should have sufficient padding to avoid false-sharing and excessive |
|
459 // cache-coherency traffic. |
|
460 |
|
461 |
|
462 typedef volatile int SpinLockT ; |
|
463 |
|
464 void Thread::SpinAcquire (volatile int * adr, const char * LockName) { |
|
465 if (Atomic::cmpxchg (1, adr, 0) == 0) { |
|
466 return ; // normal fast-path return |
|
467 } |
|
468 |
|
469 // Slow-path : We've encountered contention -- Spin/Yield/Block strategy. |
|
470 TEVENT (SpinAcquire - ctx) ; |
|
471 int ctr = 0 ; |
|
472 int Yields = 0 ; |
|
473 for (;;) { |
|
474 while (*adr != 0) { |
|
475 ++ctr ; |
|
476 if ((ctr & 0xFFF) == 0 || !os::is_MP()) { |
|
477 if (Yields > 5) { |
|
478 // Consider using a simple NakedSleep() instead. |
|
479 // Then SpinAcquire could be called by non-JVM threads |
|
480 Thread::current()->_ParkEvent->park(1) ; |
|
481 } else { |
|
482 os::NakedYield() ; |
|
483 ++Yields ; |
|
484 } |
|
485 } else { |
|
486 SpinPause() ; |
|
487 } |
|
488 } |
|
489 if (Atomic::cmpxchg (1, adr, 0) == 0) return ; |
|
490 } |
|
491 } |
|
492 |
|
493 void Thread::SpinRelease (volatile int * adr) { |
|
494 assert (*adr != 0, "invariant") ; |
|
495 OrderAccess::fence() ; // guarantee at least release consistency. |
|
496 // Roach-motel semantics. |
|
497 // It's safe if subsequent LDs and STs float "up" into the critical section, |
|
498 // but prior LDs and STs within the critical section can't be allowed |
|
499 // to reorder or float past the ST that releases the lock. |
|
500 *adr = 0 ; |
|
501 } |
|
502 |
|
503 // muxAcquire and muxRelease: |
|
504 // |
|
505 // * muxAcquire and muxRelease support a single-word lock-word construct. |
|
506 // The LSB of the word is set IFF the lock is held. |
|
507 // The remainder of the word points to the head of a singly-linked list |
|
508 // of threads blocked on the lock. |
|
509 // |
|
510 // * The current implementation of muxAcquire-muxRelease uses its own |
|
511 // dedicated Thread._MuxEvent instance. If we're interested in |
|
512 // minimizing the peak number of extant ParkEvent instances then |
|
513 // we could eliminate _MuxEvent and "borrow" _ParkEvent as long |
|
514 // as certain invariants were satisfied. Specifically, care would need |
|
515 // to be taken with regards to consuming unpark() "permits". |
|
516 // A safe rule of thumb is that a thread would never call muxAcquire() |
|
517 // if it's enqueued (cxq, EntryList, WaitList, etc) and will subsequently |
|
518 // park(). Otherwise the _ParkEvent park() operation in muxAcquire() could |
|
519 // consume an unpark() permit intended for monitorenter, for instance. |
|
520 // One way around this would be to widen the restricted-range semaphore |
|
521 // implemented in park(). Another alternative would be to provide |
|
522 // multiple instances of the PlatformEvent() for each thread. One |
|
523 // instance would be dedicated to muxAcquire-muxRelease, for instance. |
|
524 // |
|
525 // * Usage: |
|
526 // -- Only as leaf locks |
|
527 // -- for short-term locking only as muxAcquire does not perform |
|
528 // thread state transitions. |
|
529 // |
|
530 // Alternatives: |
|
531 // * We could implement muxAcquire and muxRelease with MCS or CLH locks |
|
532 // but with parking or spin-then-park instead of pure spinning. |
|
533 // * Use Taura-Oyama-Yonenzawa locks. |
|
534 // * It's possible to construct a 1-0 lock if we encode the lockword as |
|
535 // (List,LockByte). Acquire will CAS the full lockword while Release |
|
536 // will STB 0 into the LockByte. The 1-0 scheme admits stranding, so |
|
537 // acquiring threads use timers (ParkTimed) to detect and recover from |
|
538 // the stranding window. Thread/Node structures must be aligned on 256-byte |
|
539 // boundaries by using placement-new. |
|
540 // * Augment MCS with advisory back-link fields maintained with CAS(). |
|
541 // Pictorially: LockWord -> T1 <-> T2 <-> T3 <-> ... <-> Tn <-> Owner. |
|
542 // The validity of the backlinks must be ratified before we trust the value. |
|
543 // If the backlinks are invalid the exiting thread must back-track through the |
|
544 // the forward links, which are always trustworthy. |
|
545 // * Add a successor indication. The LockWord is currently encoded as |
|
546 // (List, LOCKBIT:1). We could also add a SUCCBIT or an explicit _succ variable |
|
547 // to provide the usual futile-wakeup optimization. |
|
548 // See RTStt for details. |
|
549 // * Consider schedctl.sc_nopreempt to cover the critical section. |
|
550 // |
|
551 |
|
552 |
|
553 typedef volatile intptr_t MutexT ; // Mux Lock-word |
|
554 enum MuxBits { LOCKBIT = 1 } ; |
|
555 |
|
556 void Thread::muxAcquire (volatile intptr_t * Lock, const char * LockName) { |
|
557 intptr_t w = Atomic::cmpxchg_ptr (LOCKBIT, Lock, 0) ; |
|
558 if (w == 0) return ; |
|
559 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { |
|
560 return ; |
|
561 } |
|
562 |
|
563 TEVENT (muxAcquire - Contention) ; |
|
564 ParkEvent * const Self = Thread::current()->_MuxEvent ; |
|
565 assert ((intptr_t(Self) & LOCKBIT) == 0, "invariant") ; |
|
566 for (;;) { |
|
567 int its = (os::is_MP() ? 100 : 0) + 1 ; |
|
568 |
|
569 // Optional spin phase: spin-then-park strategy |
|
570 while (--its >= 0) { |
|
571 w = *Lock ; |
|
572 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { |
|
573 return ; |
|
574 } |
|
575 } |
|
576 |
|
577 Self->reset() ; |
|
578 Self->OnList = intptr_t(Lock) ; |
|
579 // The following fence() isn't _strictly necessary as the subsequent |
|
580 // CAS() both serializes execution and ratifies the fetched *Lock value. |
|
581 OrderAccess::fence(); |
|
582 for (;;) { |
|
583 w = *Lock ; |
|
584 if ((w & LOCKBIT) == 0) { |
|
585 if (Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { |
|
586 Self->OnList = 0 ; // hygiene - allows stronger asserts |
|
587 return ; |
|
588 } |
|
589 continue ; // Interference -- *Lock changed -- Just retry |
|
590 } |
|
591 assert (w & LOCKBIT, "invariant") ; |
|
592 Self->ListNext = (ParkEvent *) (w & ~LOCKBIT ); |
|
593 if (Atomic::cmpxchg_ptr (intptr_t(Self)|LOCKBIT, Lock, w) == w) break ; |
|
594 } |
|
595 |
|
596 while (Self->OnList != 0) { |
|
597 Self->park() ; |
|
598 } |
|
599 } |
|
600 } |
|
601 |
|
602 void Thread::muxAcquireW (volatile intptr_t * Lock, ParkEvent * ev) { |
|
603 intptr_t w = Atomic::cmpxchg_ptr (LOCKBIT, Lock, 0) ; |
|
604 if (w == 0) return ; |
|
605 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { |
|
606 return ; |
|
607 } |
|
608 |
|
609 TEVENT (muxAcquire - Contention) ; |
|
610 ParkEvent * ReleaseAfter = NULL ; |
|
611 if (ev == NULL) { |
|
612 ev = ReleaseAfter = ParkEvent::Allocate (NULL) ; |
|
613 } |
|
614 assert ((intptr_t(ev) & LOCKBIT) == 0, "invariant") ; |
|
615 for (;;) { |
|
616 guarantee (ev->OnList == 0, "invariant") ; |
|
617 int its = (os::is_MP() ? 100 : 0) + 1 ; |
|
618 |
|
619 // Optional spin phase: spin-then-park strategy |
|
620 while (--its >= 0) { |
|
621 w = *Lock ; |
|
622 if ((w & LOCKBIT) == 0 && Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { |
|
623 if (ReleaseAfter != NULL) { |
|
624 ParkEvent::Release (ReleaseAfter) ; |
|
625 } |
|
626 return ; |
|
627 } |
|
628 } |
|
629 |
|
630 ev->reset() ; |
|
631 ev->OnList = intptr_t(Lock) ; |
|
632 // The following fence() isn't _strictly necessary as the subsequent |
|
633 // CAS() both serializes execution and ratifies the fetched *Lock value. |
|
634 OrderAccess::fence(); |
|
635 for (;;) { |
|
636 w = *Lock ; |
|
637 if ((w & LOCKBIT) == 0) { |
|
638 if (Atomic::cmpxchg_ptr (w|LOCKBIT, Lock, w) == w) { |
|
639 ev->OnList = 0 ; |
|
640 // We call ::Release while holding the outer lock, thus |
|
641 // artificially lengthening the critical section. |
|
642 // Consider deferring the ::Release() until the subsequent unlock(), |
|
643 // after we've dropped the outer lock. |
|
644 if (ReleaseAfter != NULL) { |
|
645 ParkEvent::Release (ReleaseAfter) ; |
|
646 } |
|
647 return ; |
|
648 } |
|
649 continue ; // Interference -- *Lock changed -- Just retry |
|
650 } |
|
651 assert (w & LOCKBIT, "invariant") ; |
|
652 ev->ListNext = (ParkEvent *) (w & ~LOCKBIT ); |
|
653 if (Atomic::cmpxchg_ptr (intptr_t(ev)|LOCKBIT, Lock, w) == w) break ; |
|
654 } |
|
655 |
|
656 while (ev->OnList != 0) { |
|
657 ev->park() ; |
|
658 } |
|
659 } |
|
660 } |
|
661 |
|
662 // Release() must extract a successor from the list and then wake that thread. |
|
663 // It can "pop" the front of the list or use a detach-modify-reattach (DMR) scheme |
|
664 // similar to that used by ParkEvent::Allocate() and ::Release(). DMR-based |
|
665 // Release() would : |
|
666 // (A) CAS() or swap() null to *Lock, releasing the lock and detaching the list. |
|
667 // (B) Extract a successor from the private list "in-hand" |
|
668 // (C) attempt to CAS() the residual back into *Lock over null. |
|
669 // If there were any newly arrived threads and the CAS() would fail. |
|
670 // In that case Release() would detach the RATs, re-merge the list in-hand |
|
671 // with the RATs and repeat as needed. Alternately, Release() might |
|
672 // detach and extract a successor, but then pass the residual list to the wakee. |
|
673 // The wakee would be responsible for reattaching and remerging before it |
|
674 // competed for the lock. |
|
675 // |
|
676 // Both "pop" and DMR are immune from ABA corruption -- there can be |
|
677 // multiple concurrent pushers, but only one popper or detacher. |
|
678 // This implementation pops from the head of the list. This is unfair, |
|
679 // but tends to provide excellent throughput as hot threads remain hot. |
|
680 // (We wake recently run threads first). |
|
681 |
|
682 void Thread::muxRelease (volatile intptr_t * Lock) { |
|
683 for (;;) { |
|
684 const intptr_t w = Atomic::cmpxchg_ptr (0, Lock, LOCKBIT) ; |
|
685 assert (w & LOCKBIT, "invariant") ; |
|
686 if (w == LOCKBIT) return ; |
|
687 ParkEvent * List = (ParkEvent *) (w & ~LOCKBIT) ; |
|
688 assert (List != NULL, "invariant") ; |
|
689 assert (List->OnList == intptr_t(Lock), "invariant") ; |
|
690 ParkEvent * nxt = List->ListNext ; |
|
691 |
|
692 // The following CAS() releases the lock and pops the head element. |
|
693 if (Atomic::cmpxchg_ptr (intptr_t(nxt), Lock, w) != w) { |
|
694 continue ; |
|
695 } |
|
696 List->OnList = 0 ; |
|
697 OrderAccess::fence() ; |
|
698 List->unpark () ; |
|
699 return ; |
|
700 } |
|
701 } |
|
702 |
|
703 // ObjectMonitor Lifecycle |
|
704 // ----------------------- |
|
705 // Inflation unlinks monitors from the global gFreeList and |
|
706 // associates them with objects. Deflation -- which occurs at |
|
707 // STW-time -- disassociates idle monitors from objects. Such |
|
708 // scavenged monitors are returned to the gFreeList. |
|
709 // |
|
710 // The global list is protected by ListLock. All the critical sections |
|
711 // are short and operate in constant-time. |
|
712 // |
|
713 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. |
|
714 // |
|
715 // Lifecycle: |
|
716 // -- unassigned and on the global free list |
|
717 // -- unassigned and on a thread's private omFreeList |
|
718 // -- assigned to an object. The object is inflated and the mark refers |
|
719 // to the objectmonitor. |
|
720 // |
|
721 // TODO-FIXME: |
|
722 // |
|
723 // * We currently protect the gFreeList with a simple lock. |
|
724 // An alternate lock-free scheme would be to pop elements from the gFreeList |
|
725 // with CAS. This would be safe from ABA corruption as long we only |
|
726 // recycled previously appearing elements onto the list in deflate_idle_monitors() |
|
727 // at STW-time. Completely new elements could always be pushed onto the gFreeList |
|
728 // with CAS. Elements that appeared previously on the list could only |
|
729 // be installed at STW-time. |
|
730 // |
|
731 // * For efficiency and to help reduce the store-before-CAS penalty |
|
732 // the objectmonitors on gFreeList or local free lists should be ready to install |
|
733 // with the exception of _header and _object. _object can be set after inflation. |
|
734 // In particular, keep all objectMonitors on a thread's private list in ready-to-install |
|
735 // state with m.Owner set properly. |
|
736 // |
|
737 // * We could all diffuse contention by using multiple global (FreeList, Lock) |
|
738 // pairs -- threads could use trylock() and a cyclic-scan strategy to search for |
|
739 // an unlocked free list. |
|
740 // |
|
741 // * Add lifecycle tags and assert()s. |
|
742 // |
|
743 // * Be more consistent about when we clear an objectmonitor's fields: |
|
744 // A. After extracting the objectmonitor from a free list. |
|
745 // B. After adding an objectmonitor to a free list. |
|
746 // |
|
747 |
|
748 ObjectMonitor * ObjectSynchronizer::gBlockList = NULL ; |
|
749 ObjectMonitor * volatile ObjectSynchronizer::gFreeList = NULL ; |
|
750 ObjectMonitor * volatile ObjectSynchronizer::gOmInUseList = NULL ; |
|
751 int ObjectSynchronizer::gOmInUseCount = 0; |
|
752 static volatile intptr_t ListLock = 0 ; // protects global monitor free-list cache |
|
753 static volatile int MonitorFreeCount = 0 ; // # on gFreeList |
|
754 static volatile int MonitorPopulation = 0 ; // # Extant -- in circulation |
|
755 #define CHAINMARKER ((oop)-1) |
|
756 |
|
757 // Constraining monitor pool growth via MonitorBound ... |
|
758 // |
|
759 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the |
|
760 // the rate of scavenging is driven primarily by GC. As such, we can find |
|
761 // an inordinate number of monitors in circulation. |
|
762 // To avoid that scenario we can artificially induce a STW safepoint |
|
763 // if the pool appears to be growing past some reasonable bound. |
|
764 // Generally we favor time in space-time tradeoffs, but as there's no |
|
765 // natural back-pressure on the # of extant monitors we need to impose some |
|
766 // type of limit. Beware that if MonitorBound is set to too low a value |
|
767 // we could just loop. In addition, if MonitorBound is set to a low value |
|
768 // we'll incur more safepoints, which are harmful to performance. |
|
769 // See also: GuaranteedSafepointInterval |
|
770 // |
|
771 // As noted elsewhere, the correct long-term solution is to deflate at |
|
772 // monitorexit-time, in which case the number of inflated objects is bounded |
|
773 // by the number of threads. That policy obviates the need for scavenging at |
|
774 // STW safepoint time. As an aside, scavenging can be time-consuming when the |
|
775 // # of extant monitors is large. Unfortunately there's a day-1 assumption baked |
|
776 // into much HotSpot code that the object::monitor relationship, once established |
|
777 // or observed, will remain stable except over potential safepoints. |
|
778 // |
|
779 // We can use either a blocking synchronous VM operation or an async VM operation. |
|
780 // -- If we use a blocking VM operation : |
|
781 // Calls to ScavengeCheck() should be inserted only into 'safe' locations in paths |
|
782 // that lead to ::inflate() or ::omAlloc(). |
|
783 // Even though the safepoint will not directly induce GC, a GC might |
|
784 // piggyback on the safepoint operation, so the caller should hold no naked oops. |
|
785 // Furthermore, monitor::object relationships are NOT necessarily stable over this call |
|
786 // unless the caller has made provisions to "pin" the object to the monitor, say |
|
787 // by incrementing the monitor's _count field. |
|
788 // -- If we use a non-blocking asynchronous VM operation : |
|
789 // the constraints above don't apply. The safepoint will fire in the future |
|
790 // at a more convenient time. On the other hand the latency between posting and |
|
791 // running the safepoint introduces or admits "slop" or laxity during which the |
|
792 // monitor population can climb further above the threshold. The monitor population, |
|
793 // however, tends to converge asymptotically over time to a count that's slightly |
|
794 // above the target value specified by MonitorBound. That is, we avoid unbounded |
|
795 // growth, albeit with some imprecision. |
|
796 // |
|
797 // The current implementation uses asynchronous VM operations. |
|
798 // |
|
799 // Ideally we'd check if (MonitorPopulation > MonitorBound) in omAlloc() |
|
800 // immediately before trying to grow the global list via allocation. |
|
801 // If the predicate was true then we'd induce a synchronous safepoint, wait |
|
802 // for the safepoint to complete, and then again to allocate from the global |
|
803 // free list. This approach is much simpler and precise, admitting no "slop". |
|
804 // Unfortunately we can't safely safepoint in the midst of omAlloc(), so |
|
805 // instead we use asynchronous safepoints. |
|
806 |
|
807 static void InduceScavenge (Thread * Self, const char * Whence) { |
|
808 // Induce STW safepoint to trim monitors |
|
809 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. |
|
810 // More precisely, trigger an asynchronous STW safepoint as the number |
|
811 // of active monitors passes the specified threshold. |
|
812 // TODO: assert thread state is reasonable |
|
813 |
|
814 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { |
|
815 if (Knob_Verbose) { |
|
816 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ; |
|
817 ::fflush(stdout) ; |
|
818 } |
|
819 // Induce a 'null' safepoint to scavenge monitors |
|
820 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted |
|
821 // to the VMthread and have a lifespan longer than that of this activation record. |
|
822 // The VMThread will delete the op when completed. |
|
823 VMThread::execute (new VM_ForceAsyncSafepoint()) ; |
|
824 |
|
825 if (Knob_Verbose) { |
|
826 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ; |
|
827 ::fflush(stdout) ; |
|
828 } |
|
829 } |
|
830 } |
|
831 /* Too slow for general assert or debug |
|
832 void ObjectSynchronizer::verifyInUse (Thread *Self) { |
|
833 ObjectMonitor* mid; |
|
834 int inusetally = 0; |
|
835 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { |
|
836 inusetally ++; |
|
837 } |
|
838 assert(inusetally == Self->omInUseCount, "inuse count off"); |
|
839 |
|
840 int freetally = 0; |
|
841 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { |
|
842 freetally ++; |
|
843 } |
|
844 assert(freetally == Self->omFreeCount, "free count off"); |
|
845 } |
|
846 */ |
|
847 |
|
848 ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) { |
|
849 // A large MAXPRIVATE value reduces both list lock contention |
|
850 // and list coherency traffic, but also tends to increase the |
|
851 // number of objectMonitors in circulation as well as the STW |
|
852 // scavenge costs. As usual, we lean toward time in space-time |
|
853 // tradeoffs. |
|
854 const int MAXPRIVATE = 1024 ; |
|
855 for (;;) { |
|
856 ObjectMonitor * m ; |
|
857 |
|
858 // 1: try to allocate from the thread's local omFreeList. |
|
859 // Threads will attempt to allocate first from their local list, then |
|
860 // from the global list, and only after those attempts fail will the thread |
|
861 // attempt to instantiate new monitors. Thread-local free lists take |
|
862 // heat off the ListLock and improve allocation latency, as well as reducing |
|
863 // coherency traffic on the shared global list. |
|
864 m = Self->omFreeList ; |
|
865 if (m != NULL) { |
|
866 Self->omFreeList = m->FreeNext ; |
|
867 Self->omFreeCount -- ; |
|
868 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene |
|
869 guarantee (m->object() == NULL, "invariant") ; |
|
870 if (MonitorInUseLists) { |
|
871 m->FreeNext = Self->omInUseList; |
|
872 Self->omInUseList = m; |
|
873 Self->omInUseCount ++; |
|
874 // verifyInUse(Self); |
|
875 } else { |
|
876 m->FreeNext = NULL; |
|
877 } |
|
878 return m ; |
|
879 } |
|
880 |
|
881 // 2: try to allocate from the global gFreeList |
|
882 // CONSIDER: use muxTry() instead of muxAcquire(). |
|
883 // If the muxTry() fails then drop immediately into case 3. |
|
884 // If we're using thread-local free lists then try |
|
885 // to reprovision the caller's free list. |
|
886 if (gFreeList != NULL) { |
|
887 // Reprovision the thread's omFreeList. |
|
888 // Use bulk transfers to reduce the allocation rate and heat |
|
889 // on various locks. |
|
890 Thread::muxAcquire (&ListLock, "omAlloc") ; |
|
891 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL; ) { |
|
892 MonitorFreeCount --; |
|
893 ObjectMonitor * take = gFreeList ; |
|
894 gFreeList = take->FreeNext ; |
|
895 guarantee (take->object() == NULL, "invariant") ; |
|
896 guarantee (!take->is_busy(), "invariant") ; |
|
897 take->Recycle() ; |
|
898 omRelease (Self, take, false) ; |
|
899 } |
|
900 Thread::muxRelease (&ListLock) ; |
|
901 Self->omFreeProvision += 1 + (Self->omFreeProvision/2) ; |
|
902 if (Self->omFreeProvision > MAXPRIVATE ) Self->omFreeProvision = MAXPRIVATE ; |
|
903 TEVENT (omFirst - reprovision) ; |
|
904 |
|
905 const int mx = MonitorBound ; |
|
906 if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) { |
|
907 // We can't safely induce a STW safepoint from omAlloc() as our thread |
|
908 // state may not be appropriate for such activities and callers may hold |
|
909 // naked oops, so instead we defer the action. |
|
910 InduceScavenge (Self, "omAlloc") ; |
|
911 } |
|
912 continue; |
|
913 } |
|
914 |
|
915 // 3: allocate a block of new ObjectMonitors |
|
916 // Both the local and global free lists are empty -- resort to malloc(). |
|
917 // In the current implementation objectMonitors are TSM - immortal. |
|
918 assert (_BLOCKSIZE > 1, "invariant") ; |
|
919 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE]; |
|
920 |
|
921 // NOTE: (almost) no way to recover if allocation failed. |
|
922 // We might be able to induce a STW safepoint and scavenge enough |
|
923 // objectMonitors to permit progress. |
|
924 if (temp == NULL) { |
|
925 vm_exit_out_of_memory (sizeof (ObjectMonitor[_BLOCKSIZE]), "Allocate ObjectMonitors") ; |
|
926 } |
|
927 |
|
928 // Format the block. |
|
929 // initialize the linked list, each monitor points to its next |
|
930 // forming the single linked free list, the very first monitor |
|
931 // will points to next block, which forms the block list. |
|
932 // The trick of using the 1st element in the block as gBlockList |
|
933 // linkage should be reconsidered. A better implementation would |
|
934 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } |
|
935 |
|
936 for (int i = 1; i < _BLOCKSIZE ; i++) { |
|
937 temp[i].FreeNext = &temp[i+1]; |
|
938 } |
|
939 |
|
940 // terminate the last monitor as the end of list |
|
941 temp[_BLOCKSIZE - 1].FreeNext = NULL ; |
|
942 |
|
943 // Element [0] is reserved for global list linkage |
|
944 temp[0].set_object(CHAINMARKER); |
|
945 |
|
946 // Consider carving out this thread's current request from the |
|
947 // block in hand. This avoids some lock traffic and redundant |
|
948 // list activity. |
|
949 |
|
950 // Acquire the ListLock to manipulate BlockList and FreeList. |
|
951 // An Oyama-Taura-Yonezawa scheme might be more efficient. |
|
952 Thread::muxAcquire (&ListLock, "omAlloc [2]") ; |
|
953 MonitorPopulation += _BLOCKSIZE-1; |
|
954 MonitorFreeCount += _BLOCKSIZE-1; |
|
955 |
|
956 // Add the new block to the list of extant blocks (gBlockList). |
|
957 // The very first objectMonitor in a block is reserved and dedicated. |
|
958 // It serves as blocklist "next" linkage. |
|
959 temp[0].FreeNext = gBlockList; |
|
960 gBlockList = temp; |
|
961 |
|
962 // Add the new string of objectMonitors to the global free list |
|
963 temp[_BLOCKSIZE - 1].FreeNext = gFreeList ; |
|
964 gFreeList = temp + 1; |
|
965 Thread::muxRelease (&ListLock) ; |
|
966 TEVENT (Allocate block of monitors) ; |
|
967 } |
|
968 } |
|
969 |
|
970 // Place "m" on the caller's private per-thread omFreeList. |
|
971 // In practice there's no need to clamp or limit the number of |
|
972 // monitors on a thread's omFreeList as the only time we'll call |
|
973 // omRelease is to return a monitor to the free list after a CAS |
|
974 // attempt failed. This doesn't allow unbounded #s of monitors to |
|
975 // accumulate on a thread's free list. |
|
976 // |
|
977 // In the future the usage of omRelease() might change and monitors |
|
978 // could migrate between free lists. In that case to avoid excessive |
|
979 // accumulation we could limit omCount to (omProvision*2), otherwise return |
|
980 // the objectMonitor to the global list. We should drain (return) in reasonable chunks. |
|
981 // That is, *not* one-at-a-time. |
|
982 |
|
983 |
|
984 void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m, bool fromPerThreadAlloc) { |
|
985 guarantee (m->object() == NULL, "invariant") ; |
|
986 |
|
987 // Remove from omInUseList |
|
988 if (MonitorInUseLists && fromPerThreadAlloc) { |
|
989 ObjectMonitor* curmidinuse = NULL; |
|
990 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; ) { |
|
991 if (m == mid) { |
|
992 // extract from per-thread in-use-list |
|
993 if (mid == Self->omInUseList) { |
|
994 Self->omInUseList = mid->FreeNext; |
|
995 } else if (curmidinuse != NULL) { |
|
996 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist |
|
997 } |
|
998 Self->omInUseCount --; |
|
999 // verifyInUse(Self); |
|
1000 break; |
|
1001 } else { |
|
1002 curmidinuse = mid; |
|
1003 mid = mid->FreeNext; |
|
1004 } |
|
1005 } |
|
1006 } |
|
1007 |
|
1008 // FreeNext is used for both onInUseList and omFreeList, so clear old before setting new |
|
1009 m->FreeNext = Self->omFreeList ; |
|
1010 Self->omFreeList = m ; |
|
1011 Self->omFreeCount ++ ; |
|
1012 } |
|
1013 |
|
1014 // Return the monitors of a moribund thread's local free list to |
|
1015 // the global free list. Typically a thread calls omFlush() when |
|
1016 // it's dying. We could also consider having the VM thread steal |
|
1017 // monitors from threads that have not run java code over a few |
|
1018 // consecutive STW safepoints. Relatedly, we might decay |
|
1019 // omFreeProvision at STW safepoints. |
|
1020 // |
|
1021 // Also return the monitors of a moribund thread"s omInUseList to |
|
1022 // a global gOmInUseList under the global list lock so these |
|
1023 // will continue to be scanned. |
|
1024 // |
|
1025 // We currently call omFlush() from the Thread:: dtor _after the thread |
|
1026 // has been excised from the thread list and is no longer a mutator. |
|
1027 // That means that omFlush() can run concurrently with a safepoint and |
|
1028 // the scavenge operator. Calling omFlush() from JavaThread::exit() might |
|
1029 // be a better choice as we could safely reason that that the JVM is |
|
1030 // not at a safepoint at the time of the call, and thus there could |
|
1031 // be not inopportune interleavings between omFlush() and the scavenge |
|
1032 // operator. |
|
1033 |
|
1034 void ObjectSynchronizer::omFlush (Thread * Self) { |
|
1035 ObjectMonitor * List = Self->omFreeList ; // Null-terminated SLL |
|
1036 Self->omFreeList = NULL ; |
|
1037 ObjectMonitor * Tail = NULL ; |
|
1038 int Tally = 0; |
|
1039 if (List != NULL) { |
|
1040 ObjectMonitor * s ; |
|
1041 for (s = List ; s != NULL ; s = s->FreeNext) { |
|
1042 Tally ++ ; |
|
1043 Tail = s ; |
|
1044 guarantee (s->object() == NULL, "invariant") ; |
|
1045 guarantee (!s->is_busy(), "invariant") ; |
|
1046 s->set_owner (NULL) ; // redundant but good hygiene |
|
1047 TEVENT (omFlush - Move one) ; |
|
1048 } |
|
1049 guarantee (Tail != NULL && List != NULL, "invariant") ; |
|
1050 } |
|
1051 |
|
1052 ObjectMonitor * InUseList = Self->omInUseList; |
|
1053 ObjectMonitor * InUseTail = NULL ; |
|
1054 int InUseTally = 0; |
|
1055 if (InUseList != NULL) { |
|
1056 Self->omInUseList = NULL; |
|
1057 ObjectMonitor *curom; |
|
1058 for (curom = InUseList; curom != NULL; curom = curom->FreeNext) { |
|
1059 InUseTail = curom; |
|
1060 InUseTally++; |
|
1061 } |
|
1062 // TODO debug |
|
1063 assert(Self->omInUseCount == InUseTally, "inuse count off"); |
|
1064 Self->omInUseCount = 0; |
|
1065 guarantee (InUseTail != NULL && InUseList != NULL, "invariant"); |
|
1066 } |
|
1067 |
|
1068 Thread::muxAcquire (&ListLock, "omFlush") ; |
|
1069 if (Tail != NULL) { |
|
1070 Tail->FreeNext = gFreeList ; |
|
1071 gFreeList = List ; |
|
1072 MonitorFreeCount += Tally; |
|
1073 } |
|
1074 |
|
1075 if (InUseTail != NULL) { |
|
1076 InUseTail->FreeNext = gOmInUseList; |
|
1077 gOmInUseList = InUseList; |
|
1078 gOmInUseCount += InUseTally; |
|
1079 } |
|
1080 |
|
1081 Thread::muxRelease (&ListLock) ; |
|
1082 TEVENT (omFlush) ; |
|
1083 } |
|
1084 |
|
1085 |
|
1086 // Get the next block in the block list. |
|
1087 static inline ObjectMonitor* next(ObjectMonitor* block) { |
|
1088 assert(block->object() == CHAINMARKER, "must be a block header"); |
|
1089 block = block->FreeNext ; |
|
1090 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); |
|
1091 return block; |
|
1092 } |
|
1093 |
|
1094 // Fast path code shared by multiple functions |
|
1095 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { |
|
1096 markOop mark = obj->mark(); |
|
1097 if (mark->has_monitor()) { |
|
1098 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); |
|
1099 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); |
|
1100 return mark->monitor(); |
|
1101 } |
|
1102 return ObjectSynchronizer::inflate(Thread::current(), obj); |
|
1103 } |
|
1104 |
|
1105 // Note that we could encounter some performance loss through false-sharing as |
|
1106 // multiple locks occupy the same $ line. Padding might be appropriate. |
|
1107 |
|
1108 #define NINFLATIONLOCKS 256 |
|
1109 static volatile intptr_t InflationLocks [NINFLATIONLOCKS] ; |
|
1110 |
416 |
1111 static markOop ReadStableMark (oop obj) { |
417 static markOop ReadStableMark (oop obj) { |
1112 markOop mark = obj->mark() ; |
418 markOop mark = obj->mark() ; |
1113 if (!mark->is_being_inflated()) { |
419 if (!mark->is_being_inflated()) { |
1114 return mark ; // normal fast-path return |
420 return mark ; // normal fast-path return |
1174 SpinPause() ; // SMP-polite spinning |
480 SpinPause() ; // SMP-polite spinning |
1175 } |
481 } |
1176 } |
482 } |
1177 } |
483 } |
1178 |
484 |
|
485 // hashCode() generation : |
|
486 // |
|
487 // Possibilities: |
|
488 // * MD5Digest of {obj,stwRandom} |
|
489 // * CRC32 of {obj,stwRandom} or any linear-feedback shift register function. |
|
490 // * A DES- or AES-style SBox[] mechanism |
|
491 // * One of the Phi-based schemes, such as: |
|
492 // 2654435761 = 2^32 * Phi (golden ratio) |
|
493 // HashCodeValue = ((uintptr_t(obj) >> 3) * 2654435761) ^ GVars.stwRandom ; |
|
494 // * A variation of Marsaglia's shift-xor RNG scheme. |
|
495 // * (obj ^ stwRandom) is appealing, but can result |
|
496 // in undesirable regularity in the hashCode values of adjacent objects |
|
497 // (objects allocated back-to-back, in particular). This could potentially |
|
498 // result in hashtable collisions and reduced hashtable efficiency. |
|
499 // There are simple ways to "diffuse" the middle address bits over the |
|
500 // generated hashCode values: |
|
501 // |
|
502 |
|
503 static inline intptr_t get_next_hash(Thread * Self, oop obj) { |
|
504 intptr_t value = 0 ; |
|
505 if (hashCode == 0) { |
|
506 // This form uses an unguarded global Park-Miller RNG, |
|
507 // so it's possible for two threads to race and generate the same RNG. |
|
508 // On MP system we'll have lots of RW access to a global, so the |
|
509 // mechanism induces lots of coherency traffic. |
|
510 value = os::random() ; |
|
511 } else |
|
512 if (hashCode == 1) { |
|
513 // This variation has the property of being stable (idempotent) |
|
514 // between STW operations. This can be useful in some of the 1-0 |
|
515 // synchronization schemes. |
|
516 intptr_t addrBits = intptr_t(obj) >> 3 ; |
|
517 value = addrBits ^ (addrBits >> 5) ^ GVars.stwRandom ; |
|
518 } else |
|
519 if (hashCode == 2) { |
|
520 value = 1 ; // for sensitivity testing |
|
521 } else |
|
522 if (hashCode == 3) { |
|
523 value = ++GVars.hcSequence ; |
|
524 } else |
|
525 if (hashCode == 4) { |
|
526 value = intptr_t(obj) ; |
|
527 } else { |
|
528 // Marsaglia's xor-shift scheme with thread-specific state |
|
529 // This is probably the best overall implementation -- we'll |
|
530 // likely make this the default in future releases. |
|
531 unsigned t = Self->_hashStateX ; |
|
532 t ^= (t << 11) ; |
|
533 Self->_hashStateX = Self->_hashStateY ; |
|
534 Self->_hashStateY = Self->_hashStateZ ; |
|
535 Self->_hashStateZ = Self->_hashStateW ; |
|
536 unsigned v = Self->_hashStateW ; |
|
537 v = (v ^ (v >> 19)) ^ (t ^ (t >> 8)) ; |
|
538 Self->_hashStateW = v ; |
|
539 value = v ; |
|
540 } |
|
541 |
|
542 value &= markOopDesc::hash_mask; |
|
543 if (value == 0) value = 0xBAD ; |
|
544 assert (value != markOopDesc::no_hash, "invariant") ; |
|
545 TEVENT (hashCode: GENERATE) ; |
|
546 return value; |
|
547 } |
|
548 // |
|
549 intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) { |
|
550 if (UseBiasedLocking) { |
|
551 // NOTE: many places throughout the JVM do not expect a safepoint |
|
552 // to be taken here, in particular most operations on perm gen |
|
553 // objects. However, we only ever bias Java instances and all of |
|
554 // the call sites of identity_hash that might revoke biases have |
|
555 // been checked to make sure they can handle a safepoint. The |
|
556 // added check of the bias pattern is to avoid useless calls to |
|
557 // thread-local storage. |
|
558 if (obj->mark()->has_bias_pattern()) { |
|
559 // Box and unbox the raw reference just in case we cause a STW safepoint. |
|
560 Handle hobj (Self, obj) ; |
|
561 // Relaxing assertion for bug 6320749. |
|
562 assert (Universe::verify_in_progress() || |
|
563 !SafepointSynchronize::is_at_safepoint(), |
|
564 "biases should not be seen by VM thread here"); |
|
565 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); |
|
566 obj = hobj() ; |
|
567 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
568 } |
|
569 } |
|
570 |
|
571 // hashCode() is a heap mutator ... |
|
572 // Relaxing assertion for bug 6320749. |
|
573 assert (Universe::verify_in_progress() || |
|
574 !SafepointSynchronize::is_at_safepoint(), "invariant") ; |
|
575 assert (Universe::verify_in_progress() || |
|
576 Self->is_Java_thread() , "invariant") ; |
|
577 assert (Universe::verify_in_progress() || |
|
578 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; |
|
579 |
|
580 ObjectMonitor* monitor = NULL; |
|
581 markOop temp, test; |
|
582 intptr_t hash; |
|
583 markOop mark = ReadStableMark (obj); |
|
584 |
|
585 // object should remain ineligible for biased locking |
|
586 assert (!mark->has_bias_pattern(), "invariant") ; |
|
587 |
|
588 if (mark->is_neutral()) { |
|
589 hash = mark->hash(); // this is a normal header |
|
590 if (hash) { // if it has hash, just return it |
|
591 return hash; |
|
592 } |
|
593 hash = get_next_hash(Self, obj); // allocate a new hash code |
|
594 temp = mark->copy_set_hash(hash); // merge the hash code into header |
|
595 // use (machine word version) atomic operation to install the hash |
|
596 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); |
|
597 if (test == mark) { |
|
598 return hash; |
|
599 } |
|
600 // If atomic operation failed, we must inflate the header |
|
601 // into heavy weight monitor. We could add more code here |
|
602 // for fast path, but it does not worth the complexity. |
|
603 } else if (mark->has_monitor()) { |
|
604 monitor = mark->monitor(); |
|
605 temp = monitor->header(); |
|
606 assert (temp->is_neutral(), "invariant") ; |
|
607 hash = temp->hash(); |
|
608 if (hash) { |
|
609 return hash; |
|
610 } |
|
611 // Skip to the following code to reduce code size |
|
612 } else if (Self->is_lock_owned((address)mark->locker())) { |
|
613 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned |
|
614 assert (temp->is_neutral(), "invariant") ; |
|
615 hash = temp->hash(); // by current thread, check if the displaced |
|
616 if (hash) { // header contains hash code |
|
617 return hash; |
|
618 } |
|
619 // WARNING: |
|
620 // The displaced header is strictly immutable. |
|
621 // It can NOT be changed in ANY cases. So we have |
|
622 // to inflate the header into heavyweight monitor |
|
623 // even the current thread owns the lock. The reason |
|
624 // is the BasicLock (stack slot) will be asynchronously |
|
625 // read by other threads during the inflate() function. |
|
626 // Any change to stack may not propagate to other threads |
|
627 // correctly. |
|
628 } |
|
629 |
|
630 // Inflate the monitor to set hash code |
|
631 monitor = ObjectSynchronizer::inflate(Self, obj); |
|
632 // Load displaced header and check it has hash code |
|
633 mark = monitor->header(); |
|
634 assert (mark->is_neutral(), "invariant") ; |
|
635 hash = mark->hash(); |
|
636 if (hash == 0) { |
|
637 hash = get_next_hash(Self, obj); |
|
638 temp = mark->copy_set_hash(hash); // merge hash code into header |
|
639 assert (temp->is_neutral(), "invariant") ; |
|
640 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); |
|
641 if (test != mark) { |
|
642 // The only update to the header in the monitor (outside GC) |
|
643 // is install the hash code. If someone add new usage of |
|
644 // displaced header, please update this code |
|
645 hash = test->hash(); |
|
646 assert (test->is_neutral(), "invariant") ; |
|
647 assert (hash != 0, "Trivial unexpected object/monitor header usage."); |
|
648 } |
|
649 } |
|
650 // We finally get the hash |
|
651 return hash; |
|
652 } |
|
653 |
|
654 // Deprecated -- use FastHashCode() instead. |
|
655 |
|
656 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { |
|
657 return FastHashCode (Thread::current(), obj()) ; |
|
658 } |
|
659 |
|
660 |
|
661 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, |
|
662 Handle h_obj) { |
|
663 if (UseBiasedLocking) { |
|
664 BiasedLocking::revoke_and_rebias(h_obj, false, thread); |
|
665 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
666 } |
|
667 |
|
668 assert(thread == JavaThread::current(), "Can only be called on current thread"); |
|
669 oop obj = h_obj(); |
|
670 |
|
671 markOop mark = ReadStableMark (obj) ; |
|
672 |
|
673 // Uncontended case, header points to stack |
|
674 if (mark->has_locker()) { |
|
675 return thread->is_lock_owned((address)mark->locker()); |
|
676 } |
|
677 // Contended case, header points to ObjectMonitor (tagged pointer) |
|
678 if (mark->has_monitor()) { |
|
679 ObjectMonitor* monitor = mark->monitor(); |
|
680 return monitor->is_entered(thread) != 0 ; |
|
681 } |
|
682 // Unlocked case, header in place |
|
683 assert(mark->is_neutral(), "sanity check"); |
|
684 return false; |
|
685 } |
|
686 |
|
687 // Be aware of this method could revoke bias of the lock object. |
|
688 // This method querys the ownership of the lock handle specified by 'h_obj'. |
|
689 // If the current thread owns the lock, it returns owner_self. If no |
|
690 // thread owns the lock, it returns owner_none. Otherwise, it will return |
|
691 // ower_other. |
|
692 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership |
|
693 (JavaThread *self, Handle h_obj) { |
|
694 // The caller must beware this method can revoke bias, and |
|
695 // revocation can result in a safepoint. |
|
696 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; |
|
697 assert (self->thread_state() != _thread_blocked , "invariant") ; |
|
698 |
|
699 // Possible mark states: neutral, biased, stack-locked, inflated |
|
700 |
|
701 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { |
|
702 // CASE: biased |
|
703 BiasedLocking::revoke_and_rebias(h_obj, false, self); |
|
704 assert(!h_obj->mark()->has_bias_pattern(), |
|
705 "biases should be revoked by now"); |
|
706 } |
|
707 |
|
708 assert(self == JavaThread::current(), "Can only be called on current thread"); |
|
709 oop obj = h_obj(); |
|
710 markOop mark = ReadStableMark (obj) ; |
|
711 |
|
712 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. |
|
713 if (mark->has_locker()) { |
|
714 return self->is_lock_owned((address)mark->locker()) ? |
|
715 owner_self : owner_other; |
|
716 } |
|
717 |
|
718 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. |
|
719 // The Object:ObjectMonitor relationship is stable as long as we're |
|
720 // not at a safepoint. |
|
721 if (mark->has_monitor()) { |
|
722 void * owner = mark->monitor()->_owner ; |
|
723 if (owner == NULL) return owner_none ; |
|
724 return (owner == self || |
|
725 self->is_lock_owned((address)owner)) ? owner_self : owner_other; |
|
726 } |
|
727 |
|
728 // CASE: neutral |
|
729 assert(mark->is_neutral(), "sanity check"); |
|
730 return owner_none ; // it's unlocked |
|
731 } |
|
732 |
|
733 // FIXME: jvmti should call this |
|
734 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { |
|
735 if (UseBiasedLocking) { |
|
736 if (SafepointSynchronize::is_at_safepoint()) { |
|
737 BiasedLocking::revoke_at_safepoint(h_obj); |
|
738 } else { |
|
739 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); |
|
740 } |
|
741 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
742 } |
|
743 |
|
744 oop obj = h_obj(); |
|
745 address owner = NULL; |
|
746 |
|
747 markOop mark = ReadStableMark (obj) ; |
|
748 |
|
749 // Uncontended case, header points to stack |
|
750 if (mark->has_locker()) { |
|
751 owner = (address) mark->locker(); |
|
752 } |
|
753 |
|
754 // Contended case, header points to ObjectMonitor (tagged pointer) |
|
755 if (mark->has_monitor()) { |
|
756 ObjectMonitor* monitor = mark->monitor(); |
|
757 assert(monitor != NULL, "monitor should be non-null"); |
|
758 owner = (address) monitor->owner(); |
|
759 } |
|
760 |
|
761 if (owner != NULL) { |
|
762 return Threads::owning_thread_from_monitor_owner(owner, doLock); |
|
763 } |
|
764 |
|
765 // Unlocked case, header in place |
|
766 // Cannot have assertion since this object may have been |
|
767 // locked by another thread when reaching here. |
|
768 // assert(mark->is_neutral(), "sanity check"); |
|
769 |
|
770 return NULL; |
|
771 } |
|
772 // Visitors ... |
|
773 |
|
774 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { |
|
775 ObjectMonitor* block = gBlockList; |
|
776 ObjectMonitor* mid; |
|
777 while (block) { |
|
778 assert(block->object() == CHAINMARKER, "must be a block header"); |
|
779 for (int i = _BLOCKSIZE - 1; i > 0; i--) { |
|
780 mid = block + i; |
|
781 oop object = (oop) mid->object(); |
|
782 if (object != NULL) { |
|
783 closure->do_monitor(mid); |
|
784 } |
|
785 } |
|
786 block = (ObjectMonitor*) block->FreeNext; |
|
787 } |
|
788 } |
|
789 |
|
790 // Get the next block in the block list. |
|
791 static inline ObjectMonitor* next(ObjectMonitor* block) { |
|
792 assert(block->object() == CHAINMARKER, "must be a block header"); |
|
793 block = block->FreeNext ; |
|
794 assert(block == NULL || block->object() == CHAINMARKER, "must be a block header"); |
|
795 return block; |
|
796 } |
|
797 |
|
798 |
|
799 void ObjectSynchronizer::oops_do(OopClosure* f) { |
|
800 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); |
|
801 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { |
|
802 assert(block->object() == CHAINMARKER, "must be a block header"); |
|
803 for (int i = 1; i < _BLOCKSIZE; i++) { |
|
804 ObjectMonitor* mid = &block[i]; |
|
805 if (mid->object() != NULL) { |
|
806 f->do_oop((oop*)mid->object_addr()); |
|
807 } |
|
808 } |
|
809 } |
|
810 } |
|
811 |
|
812 |
|
813 // ----------------------------------------------------------------------------- |
|
814 // ObjectMonitor Lifecycle |
|
815 // ----------------------- |
|
816 // Inflation unlinks monitors from the global gFreeList and |
|
817 // associates them with objects. Deflation -- which occurs at |
|
818 // STW-time -- disassociates idle monitors from objects. Such |
|
819 // scavenged monitors are returned to the gFreeList. |
|
820 // |
|
821 // The global list is protected by ListLock. All the critical sections |
|
822 // are short and operate in constant-time. |
|
823 // |
|
824 // ObjectMonitors reside in type-stable memory (TSM) and are immortal. |
|
825 // |
|
826 // Lifecycle: |
|
827 // -- unassigned and on the global free list |
|
828 // -- unassigned and on a thread's private omFreeList |
|
829 // -- assigned to an object. The object is inflated and the mark refers |
|
830 // to the objectmonitor. |
|
831 // |
|
832 |
|
833 |
|
834 // Constraining monitor pool growth via MonitorBound ... |
|
835 // |
|
836 // The monitor pool is grow-only. We scavenge at STW safepoint-time, but the |
|
837 // the rate of scavenging is driven primarily by GC. As such, we can find |
|
838 // an inordinate number of monitors in circulation. |
|
839 // To avoid that scenario we can artificially induce a STW safepoint |
|
840 // if the pool appears to be growing past some reasonable bound. |
|
841 // Generally we favor time in space-time tradeoffs, but as there's no |
|
842 // natural back-pressure on the # of extant monitors we need to impose some |
|
843 // type of limit. Beware that if MonitorBound is set to too low a value |
|
844 // we could just loop. In addition, if MonitorBound is set to a low value |
|
845 // we'll incur more safepoints, which are harmful to performance. |
|
846 // See also: GuaranteedSafepointInterval |
|
847 // |
|
848 // The current implementation uses asynchronous VM operations. |
|
849 // |
|
850 |
|
851 static void InduceScavenge (Thread * Self, const char * Whence) { |
|
852 // Induce STW safepoint to trim monitors |
|
853 // Ultimately, this results in a call to deflate_idle_monitors() in the near future. |
|
854 // More precisely, trigger an asynchronous STW safepoint as the number |
|
855 // of active monitors passes the specified threshold. |
|
856 // TODO: assert thread state is reasonable |
|
857 |
|
858 if (ForceMonitorScavenge == 0 && Atomic::xchg (1, &ForceMonitorScavenge) == 0) { |
|
859 if (ObjectMonitor::Knob_Verbose) { |
|
860 ::printf ("Monitor scavenge - Induced STW @%s (%d)\n", Whence, ForceMonitorScavenge) ; |
|
861 ::fflush(stdout) ; |
|
862 } |
|
863 // Induce a 'null' safepoint to scavenge monitors |
|
864 // Must VM_Operation instance be heap allocated as the op will be enqueue and posted |
|
865 // to the VMthread and have a lifespan longer than that of this activation record. |
|
866 // The VMThread will delete the op when completed. |
|
867 VMThread::execute (new VM_ForceAsyncSafepoint()) ; |
|
868 |
|
869 if (ObjectMonitor::Knob_Verbose) { |
|
870 ::printf ("Monitor scavenge - STW posted @%s (%d)\n", Whence, ForceMonitorScavenge) ; |
|
871 ::fflush(stdout) ; |
|
872 } |
|
873 } |
|
874 } |
|
875 /* Too slow for general assert or debug |
|
876 void ObjectSynchronizer::verifyInUse (Thread *Self) { |
|
877 ObjectMonitor* mid; |
|
878 int inusetally = 0; |
|
879 for (mid = Self->omInUseList; mid != NULL; mid = mid->FreeNext) { |
|
880 inusetally ++; |
|
881 } |
|
882 assert(inusetally == Self->omInUseCount, "inuse count off"); |
|
883 |
|
884 int freetally = 0; |
|
885 for (mid = Self->omFreeList; mid != NULL; mid = mid->FreeNext) { |
|
886 freetally ++; |
|
887 } |
|
888 assert(freetally == Self->omFreeCount, "free count off"); |
|
889 } |
|
890 */ |
|
891 ObjectMonitor * ATTR ObjectSynchronizer::omAlloc (Thread * Self) { |
|
892 // A large MAXPRIVATE value reduces both list lock contention |
|
893 // and list coherency traffic, but also tends to increase the |
|
894 // number of objectMonitors in circulation as well as the STW |
|
895 // scavenge costs. As usual, we lean toward time in space-time |
|
896 // tradeoffs. |
|
897 const int MAXPRIVATE = 1024 ; |
|
898 for (;;) { |
|
899 ObjectMonitor * m ; |
|
900 |
|
901 // 1: try to allocate from the thread's local omFreeList. |
|
902 // Threads will attempt to allocate first from their local list, then |
|
903 // from the global list, and only after those attempts fail will the thread |
|
904 // attempt to instantiate new monitors. Thread-local free lists take |
|
905 // heat off the ListLock and improve allocation latency, as well as reducing |
|
906 // coherency traffic on the shared global list. |
|
907 m = Self->omFreeList ; |
|
908 if (m != NULL) { |
|
909 Self->omFreeList = m->FreeNext ; |
|
910 Self->omFreeCount -- ; |
|
911 // CONSIDER: set m->FreeNext = BAD -- diagnostic hygiene |
|
912 guarantee (m->object() == NULL, "invariant") ; |
|
913 if (MonitorInUseLists) { |
|
914 m->FreeNext = Self->omInUseList; |
|
915 Self->omInUseList = m; |
|
916 Self->omInUseCount ++; |
|
917 // verifyInUse(Self); |
|
918 } else { |
|
919 m->FreeNext = NULL; |
|
920 } |
|
921 return m ; |
|
922 } |
|
923 |
|
924 // 2: try to allocate from the global gFreeList |
|
925 // CONSIDER: use muxTry() instead of muxAcquire(). |
|
926 // If the muxTry() fails then drop immediately into case 3. |
|
927 // If we're using thread-local free lists then try |
|
928 // to reprovision the caller's free list. |
|
929 if (gFreeList != NULL) { |
|
930 // Reprovision the thread's omFreeList. |
|
931 // Use bulk transfers to reduce the allocation rate and heat |
|
932 // on various locks. |
|
933 Thread::muxAcquire (&ListLock, "omAlloc") ; |
|
934 for (int i = Self->omFreeProvision; --i >= 0 && gFreeList != NULL; ) { |
|
935 MonitorFreeCount --; |
|
936 ObjectMonitor * take = gFreeList ; |
|
937 gFreeList = take->FreeNext ; |
|
938 guarantee (take->object() == NULL, "invariant") ; |
|
939 guarantee (!take->is_busy(), "invariant") ; |
|
940 take->Recycle() ; |
|
941 omRelease (Self, take, false) ; |
|
942 } |
|
943 Thread::muxRelease (&ListLock) ; |
|
944 Self->omFreeProvision += 1 + (Self->omFreeProvision/2) ; |
|
945 if (Self->omFreeProvision > MAXPRIVATE ) Self->omFreeProvision = MAXPRIVATE ; |
|
946 TEVENT (omFirst - reprovision) ; |
|
947 |
|
948 const int mx = MonitorBound ; |
|
949 if (mx > 0 && (MonitorPopulation-MonitorFreeCount) > mx) { |
|
950 // We can't safely induce a STW safepoint from omAlloc() as our thread |
|
951 // state may not be appropriate for such activities and callers may hold |
|
952 // naked oops, so instead we defer the action. |
|
953 InduceScavenge (Self, "omAlloc") ; |
|
954 } |
|
955 continue; |
|
956 } |
|
957 |
|
958 // 3: allocate a block of new ObjectMonitors |
|
959 // Both the local and global free lists are empty -- resort to malloc(). |
|
960 // In the current implementation objectMonitors are TSM - immortal. |
|
961 assert (_BLOCKSIZE > 1, "invariant") ; |
|
962 ObjectMonitor * temp = new ObjectMonitor[_BLOCKSIZE]; |
|
963 |
|
964 // NOTE: (almost) no way to recover if allocation failed. |
|
965 // We might be able to induce a STW safepoint and scavenge enough |
|
966 // objectMonitors to permit progress. |
|
967 if (temp == NULL) { |
|
968 vm_exit_out_of_memory (sizeof (ObjectMonitor[_BLOCKSIZE]), "Allocate ObjectMonitors") ; |
|
969 } |
|
970 |
|
971 // Format the block. |
|
972 // initialize the linked list, each monitor points to its next |
|
973 // forming the single linked free list, the very first monitor |
|
974 // will points to next block, which forms the block list. |
|
975 // The trick of using the 1st element in the block as gBlockList |
|
976 // linkage should be reconsidered. A better implementation would |
|
977 // look like: class Block { Block * next; int N; ObjectMonitor Body [N] ; } |
|
978 |
|
979 for (int i = 1; i < _BLOCKSIZE ; i++) { |
|
980 temp[i].FreeNext = &temp[i+1]; |
|
981 } |
|
982 |
|
983 // terminate the last monitor as the end of list |
|
984 temp[_BLOCKSIZE - 1].FreeNext = NULL ; |
|
985 |
|
986 // Element [0] is reserved for global list linkage |
|
987 temp[0].set_object(CHAINMARKER); |
|
988 |
|
989 // Consider carving out this thread's current request from the |
|
990 // block in hand. This avoids some lock traffic and redundant |
|
991 // list activity. |
|
992 |
|
993 // Acquire the ListLock to manipulate BlockList and FreeList. |
|
994 // An Oyama-Taura-Yonezawa scheme might be more efficient. |
|
995 Thread::muxAcquire (&ListLock, "omAlloc [2]") ; |
|
996 MonitorPopulation += _BLOCKSIZE-1; |
|
997 MonitorFreeCount += _BLOCKSIZE-1; |
|
998 |
|
999 // Add the new block to the list of extant blocks (gBlockList). |
|
1000 // The very first objectMonitor in a block is reserved and dedicated. |
|
1001 // It serves as blocklist "next" linkage. |
|
1002 temp[0].FreeNext = gBlockList; |
|
1003 gBlockList = temp; |
|
1004 |
|
1005 // Add the new string of objectMonitors to the global free list |
|
1006 temp[_BLOCKSIZE - 1].FreeNext = gFreeList ; |
|
1007 gFreeList = temp + 1; |
|
1008 Thread::muxRelease (&ListLock) ; |
|
1009 TEVENT (Allocate block of monitors) ; |
|
1010 } |
|
1011 } |
|
1012 |
|
1013 // Place "m" on the caller's private per-thread omFreeList. |
|
1014 // In practice there's no need to clamp or limit the number of |
|
1015 // monitors on a thread's omFreeList as the only time we'll call |
|
1016 // omRelease is to return a monitor to the free list after a CAS |
|
1017 // attempt failed. This doesn't allow unbounded #s of monitors to |
|
1018 // accumulate on a thread's free list. |
|
1019 // |
|
1020 |
|
1021 void ObjectSynchronizer::omRelease (Thread * Self, ObjectMonitor * m, bool fromPerThreadAlloc) { |
|
1022 guarantee (m->object() == NULL, "invariant") ; |
|
1023 |
|
1024 // Remove from omInUseList |
|
1025 if (MonitorInUseLists && fromPerThreadAlloc) { |
|
1026 ObjectMonitor* curmidinuse = NULL; |
|
1027 for (ObjectMonitor* mid = Self->omInUseList; mid != NULL; ) { |
|
1028 if (m == mid) { |
|
1029 // extract from per-thread in-use-list |
|
1030 if (mid == Self->omInUseList) { |
|
1031 Self->omInUseList = mid->FreeNext; |
|
1032 } else if (curmidinuse != NULL) { |
|
1033 curmidinuse->FreeNext = mid->FreeNext; // maintain the current thread inuselist |
|
1034 } |
|
1035 Self->omInUseCount --; |
|
1036 // verifyInUse(Self); |
|
1037 break; |
|
1038 } else { |
|
1039 curmidinuse = mid; |
|
1040 mid = mid->FreeNext; |
|
1041 } |
|
1042 } |
|
1043 } |
|
1044 |
|
1045 // FreeNext is used for both onInUseList and omFreeList, so clear old before setting new |
|
1046 m->FreeNext = Self->omFreeList ; |
|
1047 Self->omFreeList = m ; |
|
1048 Self->omFreeCount ++ ; |
|
1049 } |
|
1050 |
|
1051 // Return the monitors of a moribund thread's local free list to |
|
1052 // the global free list. Typically a thread calls omFlush() when |
|
1053 // it's dying. We could also consider having the VM thread steal |
|
1054 // monitors from threads that have not run java code over a few |
|
1055 // consecutive STW safepoints. Relatedly, we might decay |
|
1056 // omFreeProvision at STW safepoints. |
|
1057 // |
|
1058 // Also return the monitors of a moribund thread"s omInUseList to |
|
1059 // a global gOmInUseList under the global list lock so these |
|
1060 // will continue to be scanned. |
|
1061 // |
|
1062 // We currently call omFlush() from the Thread:: dtor _after the thread |
|
1063 // has been excised from the thread list and is no longer a mutator. |
|
1064 // That means that omFlush() can run concurrently with a safepoint and |
|
1065 // the scavenge operator. Calling omFlush() from JavaThread::exit() might |
|
1066 // be a better choice as we could safely reason that that the JVM is |
|
1067 // not at a safepoint at the time of the call, and thus there could |
|
1068 // be not inopportune interleavings between omFlush() and the scavenge |
|
1069 // operator. |
|
1070 |
|
1071 void ObjectSynchronizer::omFlush (Thread * Self) { |
|
1072 ObjectMonitor * List = Self->omFreeList ; // Null-terminated SLL |
|
1073 Self->omFreeList = NULL ; |
|
1074 ObjectMonitor * Tail = NULL ; |
|
1075 int Tally = 0; |
|
1076 if (List != NULL) { |
|
1077 ObjectMonitor * s ; |
|
1078 for (s = List ; s != NULL ; s = s->FreeNext) { |
|
1079 Tally ++ ; |
|
1080 Tail = s ; |
|
1081 guarantee (s->object() == NULL, "invariant") ; |
|
1082 guarantee (!s->is_busy(), "invariant") ; |
|
1083 s->set_owner (NULL) ; // redundant but good hygiene |
|
1084 TEVENT (omFlush - Move one) ; |
|
1085 } |
|
1086 guarantee (Tail != NULL && List != NULL, "invariant") ; |
|
1087 } |
|
1088 |
|
1089 ObjectMonitor * InUseList = Self->omInUseList; |
|
1090 ObjectMonitor * InUseTail = NULL ; |
|
1091 int InUseTally = 0; |
|
1092 if (InUseList != NULL) { |
|
1093 Self->omInUseList = NULL; |
|
1094 ObjectMonitor *curom; |
|
1095 for (curom = InUseList; curom != NULL; curom = curom->FreeNext) { |
|
1096 InUseTail = curom; |
|
1097 InUseTally++; |
|
1098 } |
|
1099 // TODO debug |
|
1100 assert(Self->omInUseCount == InUseTally, "inuse count off"); |
|
1101 Self->omInUseCount = 0; |
|
1102 guarantee (InUseTail != NULL && InUseList != NULL, "invariant"); |
|
1103 } |
|
1104 |
|
1105 Thread::muxAcquire (&ListLock, "omFlush") ; |
|
1106 if (Tail != NULL) { |
|
1107 Tail->FreeNext = gFreeList ; |
|
1108 gFreeList = List ; |
|
1109 MonitorFreeCount += Tally; |
|
1110 } |
|
1111 |
|
1112 if (InUseTail != NULL) { |
|
1113 InUseTail->FreeNext = gOmInUseList; |
|
1114 gOmInUseList = InUseList; |
|
1115 gOmInUseCount += InUseTally; |
|
1116 } |
|
1117 |
|
1118 Thread::muxRelease (&ListLock) ; |
|
1119 TEVENT (omFlush) ; |
|
1120 } |
|
1121 |
|
1122 // Fast path code shared by multiple functions |
|
1123 ObjectMonitor* ObjectSynchronizer::inflate_helper(oop obj) { |
|
1124 markOop mark = obj->mark(); |
|
1125 if (mark->has_monitor()) { |
|
1126 assert(ObjectSynchronizer::verify_objmon_isinpool(mark->monitor()), "monitor is invalid"); |
|
1127 assert(mark->monitor()->header()->is_neutral(), "monitor must record a good object header"); |
|
1128 return mark->monitor(); |
|
1129 } |
|
1130 return ObjectSynchronizer::inflate(Thread::current(), obj); |
|
1131 } |
|
1132 |
|
1133 |
|
1134 // Note that we could encounter some performance loss through false-sharing as |
|
1135 // multiple locks occupy the same $ line. Padding might be appropriate. |
|
1136 |
|
1137 |
1179 ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) { |
1138 ObjectMonitor * ATTR ObjectSynchronizer::inflate (Thread * Self, oop object) { |
1180 // Inflate mutates the heap ... |
1139 // Inflate mutates the heap ... |
1181 // Relaxing assertion for bug 6320749. |
1140 // Relaxing assertion for bug 6320749. |
1182 assert (Universe::verify_in_progress() || |
1141 assert (Universe::verify_in_progress() || |
1183 !SafepointSynchronize::is_at_safepoint(), "invariant") ; |
1142 !SafepointSynchronize::is_at_safepoint(), "invariant") ; |
1364 } |
1323 } |
1365 return m ; |
1324 return m ; |
1366 } |
1325 } |
1367 } |
1326 } |
1368 |
1327 |
1369 |
1328 // Note that we could encounter some performance loss through false-sharing as |
1370 // This the fast monitor enter. The interpreter and compiler use |
1329 // multiple locks occupy the same $ line. Padding might be appropriate. |
1371 // some assembly copies of this code. Make sure update those code |
1330 |
1372 // if the following function is changed. The implementation is |
|
1373 // extremely sensitive to race condition. Be careful. |
|
1374 |
|
1375 void ObjectSynchronizer::fast_enter(Handle obj, BasicLock* lock, bool attempt_rebias, TRAPS) { |
|
1376 if (UseBiasedLocking) { |
|
1377 if (!SafepointSynchronize::is_at_safepoint()) { |
|
1378 BiasedLocking::Condition cond = BiasedLocking::revoke_and_rebias(obj, attempt_rebias, THREAD); |
|
1379 if (cond == BiasedLocking::BIAS_REVOKED_AND_REBIASED) { |
|
1380 return; |
|
1381 } |
|
1382 } else { |
|
1383 assert(!attempt_rebias, "can not rebias toward VM thread"); |
|
1384 BiasedLocking::revoke_at_safepoint(obj); |
|
1385 } |
|
1386 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1387 } |
|
1388 |
|
1389 slow_enter (obj, lock, THREAD) ; |
|
1390 } |
|
1391 |
|
1392 void ObjectSynchronizer::fast_exit(oop object, BasicLock* lock, TRAPS) { |
|
1393 assert(!object->mark()->has_bias_pattern(), "should not see bias pattern here"); |
|
1394 // if displaced header is null, the previous enter is recursive enter, no-op |
|
1395 markOop dhw = lock->displaced_header(); |
|
1396 markOop mark ; |
|
1397 if (dhw == NULL) { |
|
1398 // Recursive stack-lock. |
|
1399 // Diagnostics -- Could be: stack-locked, inflating, inflated. |
|
1400 mark = object->mark() ; |
|
1401 assert (!mark->is_neutral(), "invariant") ; |
|
1402 if (mark->has_locker() && mark != markOopDesc::INFLATING()) { |
|
1403 assert(THREAD->is_lock_owned((address)mark->locker()), "invariant") ; |
|
1404 } |
|
1405 if (mark->has_monitor()) { |
|
1406 ObjectMonitor * m = mark->monitor() ; |
|
1407 assert(((oop)(m->object()))->mark() == mark, "invariant") ; |
|
1408 assert(m->is_entered(THREAD), "invariant") ; |
|
1409 } |
|
1410 return ; |
|
1411 } |
|
1412 |
|
1413 mark = object->mark() ; |
|
1414 |
|
1415 // If the object is stack-locked by the current thread, try to |
|
1416 // swing the displaced header from the box back to the mark. |
|
1417 if (mark == (markOop) lock) { |
|
1418 assert (dhw->is_neutral(), "invariant") ; |
|
1419 if ((markOop) Atomic::cmpxchg_ptr (dhw, object->mark_addr(), mark) == mark) { |
|
1420 TEVENT (fast_exit: release stacklock) ; |
|
1421 return; |
|
1422 } |
|
1423 } |
|
1424 |
|
1425 ObjectSynchronizer::inflate(THREAD, object)->exit (THREAD) ; |
|
1426 } |
|
1427 |
|
1428 // This routine is used to handle interpreter/compiler slow case |
|
1429 // We don't need to use fast path here, because it must have been |
|
1430 // failed in the interpreter/compiler code. |
|
1431 void ObjectSynchronizer::slow_enter(Handle obj, BasicLock* lock, TRAPS) { |
|
1432 markOop mark = obj->mark(); |
|
1433 assert(!mark->has_bias_pattern(), "should not see bias pattern here"); |
|
1434 |
|
1435 if (mark->is_neutral()) { |
|
1436 // Anticipate successful CAS -- the ST of the displaced mark must |
|
1437 // be visible <= the ST performed by the CAS. |
|
1438 lock->set_displaced_header(mark); |
|
1439 if (mark == (markOop) Atomic::cmpxchg_ptr(lock, obj()->mark_addr(), mark)) { |
|
1440 TEVENT (slow_enter: release stacklock) ; |
|
1441 return ; |
|
1442 } |
|
1443 // Fall through to inflate() ... |
|
1444 } else |
|
1445 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { |
|
1446 assert(lock != mark->locker(), "must not re-lock the same lock"); |
|
1447 assert(lock != (BasicLock*)obj->mark(), "don't relock with same BasicLock"); |
|
1448 lock->set_displaced_header(NULL); |
|
1449 return; |
|
1450 } |
|
1451 |
|
1452 #if 0 |
|
1453 // The following optimization isn't particularly useful. |
|
1454 if (mark->has_monitor() && mark->monitor()->is_entered(THREAD)) { |
|
1455 lock->set_displaced_header (NULL) ; |
|
1456 return ; |
|
1457 } |
|
1458 #endif |
|
1459 |
|
1460 // The object header will never be displaced to this lock, |
|
1461 // so it does not matter what the value is, except that it |
|
1462 // must be non-zero to avoid looking like a re-entrant lock, |
|
1463 // and must not look locked either. |
|
1464 lock->set_displaced_header(markOopDesc::unused_mark()); |
|
1465 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); |
|
1466 } |
|
1467 |
|
1468 // This routine is used to handle interpreter/compiler slow case |
|
1469 // We don't need to use fast path here, because it must have |
|
1470 // failed in the interpreter/compiler code. Simply use the heavy |
|
1471 // weight monitor should be ok, unless someone find otherwise. |
|
1472 void ObjectSynchronizer::slow_exit(oop object, BasicLock* lock, TRAPS) { |
|
1473 fast_exit (object, lock, THREAD) ; |
|
1474 } |
|
1475 |
|
1476 // NOTE: must use heavy weight monitor to handle jni monitor enter |
|
1477 void ObjectSynchronizer::jni_enter(Handle obj, TRAPS) { // possible entry from jni enter |
|
1478 // the current locking is from JNI instead of Java code |
|
1479 TEVENT (jni_enter) ; |
|
1480 if (UseBiasedLocking) { |
|
1481 BiasedLocking::revoke_and_rebias(obj, false, THREAD); |
|
1482 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1483 } |
|
1484 THREAD->set_current_pending_monitor_is_from_java(false); |
|
1485 ObjectSynchronizer::inflate(THREAD, obj())->enter(THREAD); |
|
1486 THREAD->set_current_pending_monitor_is_from_java(true); |
|
1487 } |
|
1488 |
|
1489 // NOTE: must use heavy weight monitor to handle jni monitor enter |
|
1490 bool ObjectSynchronizer::jni_try_enter(Handle obj, Thread* THREAD) { |
|
1491 if (UseBiasedLocking) { |
|
1492 BiasedLocking::revoke_and_rebias(obj, false, THREAD); |
|
1493 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1494 } |
|
1495 |
|
1496 ObjectMonitor* monitor = ObjectSynchronizer::inflate_helper(obj()); |
|
1497 return monitor->try_enter(THREAD); |
|
1498 } |
|
1499 |
|
1500 |
|
1501 // NOTE: must use heavy weight monitor to handle jni monitor exit |
|
1502 void ObjectSynchronizer::jni_exit(oop obj, Thread* THREAD) { |
|
1503 TEVENT (jni_exit) ; |
|
1504 if (UseBiasedLocking) { |
|
1505 BiasedLocking::revoke_and_rebias(obj, false, THREAD); |
|
1506 } |
|
1507 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1508 |
|
1509 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj); |
|
1510 // If this thread has locked the object, exit the monitor. Note: can't use |
|
1511 // monitor->check(CHECK); must exit even if an exception is pending. |
|
1512 if (monitor->check(THREAD)) { |
|
1513 monitor->exit(THREAD); |
|
1514 } |
|
1515 } |
|
1516 |
|
1517 // complete_exit()/reenter() are used to wait on a nested lock |
|
1518 // i.e. to give up an outer lock completely and then re-enter |
|
1519 // Used when holding nested locks - lock acquisition order: lock1 then lock2 |
|
1520 // 1) complete_exit lock1 - saving recursion count |
|
1521 // 2) wait on lock2 |
|
1522 // 3) when notified on lock2, unlock lock2 |
|
1523 // 4) reenter lock1 with original recursion count |
|
1524 // 5) lock lock2 |
|
1525 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() |
|
1526 intptr_t ObjectSynchronizer::complete_exit(Handle obj, TRAPS) { |
|
1527 TEVENT (complete_exit) ; |
|
1528 if (UseBiasedLocking) { |
|
1529 BiasedLocking::revoke_and_rebias(obj, false, THREAD); |
|
1530 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1531 } |
|
1532 |
|
1533 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); |
|
1534 |
|
1535 return monitor->complete_exit(THREAD); |
|
1536 } |
|
1537 |
|
1538 // NOTE: must use heavy weight monitor to handle complete_exit/reenter() |
|
1539 void ObjectSynchronizer::reenter(Handle obj, intptr_t recursion, TRAPS) { |
|
1540 TEVENT (reenter) ; |
|
1541 if (UseBiasedLocking) { |
|
1542 BiasedLocking::revoke_and_rebias(obj, false, THREAD); |
|
1543 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1544 } |
|
1545 |
|
1546 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); |
|
1547 |
|
1548 monitor->reenter(recursion, THREAD); |
|
1549 } |
|
1550 |
|
1551 // This exists only as a workaround of dtrace bug 6254741 |
|
1552 int dtrace_waited_probe(ObjectMonitor* monitor, Handle obj, Thread* thr) { |
|
1553 DTRACE_MONITOR_PROBE(waited, monitor, obj(), thr); |
|
1554 return 0; |
|
1555 } |
|
1556 |
|
1557 // NOTE: must use heavy weight monitor to handle wait() |
|
1558 void ObjectSynchronizer::wait(Handle obj, jlong millis, TRAPS) { |
|
1559 if (UseBiasedLocking) { |
|
1560 BiasedLocking::revoke_and_rebias(obj, false, THREAD); |
|
1561 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1562 } |
|
1563 if (millis < 0) { |
|
1564 TEVENT (wait - throw IAX) ; |
|
1565 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); |
|
1566 } |
|
1567 ObjectMonitor* monitor = ObjectSynchronizer::inflate(THREAD, obj()); |
|
1568 DTRACE_MONITOR_WAIT_PROBE(monitor, obj(), THREAD, millis); |
|
1569 monitor->wait(millis, true, THREAD); |
|
1570 |
|
1571 /* This dummy call is in place to get around dtrace bug 6254741. Once |
|
1572 that's fixed we can uncomment the following line and remove the call */ |
|
1573 // DTRACE_MONITOR_PROBE(waited, monitor, obj(), THREAD); |
|
1574 dtrace_waited_probe(monitor, obj, THREAD); |
|
1575 } |
|
1576 |
|
1577 void ObjectSynchronizer::waitUninterruptibly (Handle obj, jlong millis, TRAPS) { |
|
1578 if (UseBiasedLocking) { |
|
1579 BiasedLocking::revoke_and_rebias(obj, false, THREAD); |
|
1580 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1581 } |
|
1582 if (millis < 0) { |
|
1583 TEVENT (wait - throw IAX) ; |
|
1584 THROW_MSG(vmSymbols::java_lang_IllegalArgumentException(), "timeout value is negative"); |
|
1585 } |
|
1586 ObjectSynchronizer::inflate(THREAD, obj()) -> wait(millis, false, THREAD) ; |
|
1587 } |
|
1588 |
|
1589 void ObjectSynchronizer::notify(Handle obj, TRAPS) { |
|
1590 if (UseBiasedLocking) { |
|
1591 BiasedLocking::revoke_and_rebias(obj, false, THREAD); |
|
1592 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1593 } |
|
1594 |
|
1595 markOop mark = obj->mark(); |
|
1596 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { |
|
1597 return; |
|
1598 } |
|
1599 ObjectSynchronizer::inflate(THREAD, obj())->notify(THREAD); |
|
1600 } |
|
1601 |
|
1602 // NOTE: see comment of notify() |
|
1603 void ObjectSynchronizer::notifyall(Handle obj, TRAPS) { |
|
1604 if (UseBiasedLocking) { |
|
1605 BiasedLocking::revoke_and_rebias(obj, false, THREAD); |
|
1606 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1607 } |
|
1608 |
|
1609 markOop mark = obj->mark(); |
|
1610 if (mark->has_locker() && THREAD->is_lock_owned((address)mark->locker())) { |
|
1611 return; |
|
1612 } |
|
1613 ObjectSynchronizer::inflate(THREAD, obj())->notifyAll(THREAD); |
|
1614 } |
|
1615 |
|
1616 intptr_t ObjectSynchronizer::FastHashCode (Thread * Self, oop obj) { |
|
1617 if (UseBiasedLocking) { |
|
1618 // NOTE: many places throughout the JVM do not expect a safepoint |
|
1619 // to be taken here, in particular most operations on perm gen |
|
1620 // objects. However, we only ever bias Java instances and all of |
|
1621 // the call sites of identity_hash that might revoke biases have |
|
1622 // been checked to make sure they can handle a safepoint. The |
|
1623 // added check of the bias pattern is to avoid useless calls to |
|
1624 // thread-local storage. |
|
1625 if (obj->mark()->has_bias_pattern()) { |
|
1626 // Box and unbox the raw reference just in case we cause a STW safepoint. |
|
1627 Handle hobj (Self, obj) ; |
|
1628 // Relaxing assertion for bug 6320749. |
|
1629 assert (Universe::verify_in_progress() || |
|
1630 !SafepointSynchronize::is_at_safepoint(), |
|
1631 "biases should not be seen by VM thread here"); |
|
1632 BiasedLocking::revoke_and_rebias(hobj, false, JavaThread::current()); |
|
1633 obj = hobj() ; |
|
1634 assert(!obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1635 } |
|
1636 } |
|
1637 |
|
1638 // hashCode() is a heap mutator ... |
|
1639 // Relaxing assertion for bug 6320749. |
|
1640 assert (Universe::verify_in_progress() || |
|
1641 !SafepointSynchronize::is_at_safepoint(), "invariant") ; |
|
1642 assert (Universe::verify_in_progress() || |
|
1643 Self->is_Java_thread() , "invariant") ; |
|
1644 assert (Universe::verify_in_progress() || |
|
1645 ((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; |
|
1646 |
|
1647 ObjectMonitor* monitor = NULL; |
|
1648 markOop temp, test; |
|
1649 intptr_t hash; |
|
1650 markOop mark = ReadStableMark (obj); |
|
1651 |
|
1652 // object should remain ineligible for biased locking |
|
1653 assert (!mark->has_bias_pattern(), "invariant") ; |
|
1654 |
|
1655 if (mark->is_neutral()) { |
|
1656 hash = mark->hash(); // this is a normal header |
|
1657 if (hash) { // if it has hash, just return it |
|
1658 return hash; |
|
1659 } |
|
1660 hash = get_next_hash(Self, obj); // allocate a new hash code |
|
1661 temp = mark->copy_set_hash(hash); // merge the hash code into header |
|
1662 // use (machine word version) atomic operation to install the hash |
|
1663 test = (markOop) Atomic::cmpxchg_ptr(temp, obj->mark_addr(), mark); |
|
1664 if (test == mark) { |
|
1665 return hash; |
|
1666 } |
|
1667 // If atomic operation failed, we must inflate the header |
|
1668 // into heavy weight monitor. We could add more code here |
|
1669 // for fast path, but it does not worth the complexity. |
|
1670 } else if (mark->has_monitor()) { |
|
1671 monitor = mark->monitor(); |
|
1672 temp = monitor->header(); |
|
1673 assert (temp->is_neutral(), "invariant") ; |
|
1674 hash = temp->hash(); |
|
1675 if (hash) { |
|
1676 return hash; |
|
1677 } |
|
1678 // Skip to the following code to reduce code size |
|
1679 } else if (Self->is_lock_owned((address)mark->locker())) { |
|
1680 temp = mark->displaced_mark_helper(); // this is a lightweight monitor owned |
|
1681 assert (temp->is_neutral(), "invariant") ; |
|
1682 hash = temp->hash(); // by current thread, check if the displaced |
|
1683 if (hash) { // header contains hash code |
|
1684 return hash; |
|
1685 } |
|
1686 // WARNING: |
|
1687 // The displaced header is strictly immutable. |
|
1688 // It can NOT be changed in ANY cases. So we have |
|
1689 // to inflate the header into heavyweight monitor |
|
1690 // even the current thread owns the lock. The reason |
|
1691 // is the BasicLock (stack slot) will be asynchronously |
|
1692 // read by other threads during the inflate() function. |
|
1693 // Any change to stack may not propagate to other threads |
|
1694 // correctly. |
|
1695 } |
|
1696 |
|
1697 // Inflate the monitor to set hash code |
|
1698 monitor = ObjectSynchronizer::inflate(Self, obj); |
|
1699 // Load displaced header and check it has hash code |
|
1700 mark = monitor->header(); |
|
1701 assert (mark->is_neutral(), "invariant") ; |
|
1702 hash = mark->hash(); |
|
1703 if (hash == 0) { |
|
1704 hash = get_next_hash(Self, obj); |
|
1705 temp = mark->copy_set_hash(hash); // merge hash code into header |
|
1706 assert (temp->is_neutral(), "invariant") ; |
|
1707 test = (markOop) Atomic::cmpxchg_ptr(temp, monitor, mark); |
|
1708 if (test != mark) { |
|
1709 // The only update to the header in the monitor (outside GC) |
|
1710 // is install the hash code. If someone add new usage of |
|
1711 // displaced header, please update this code |
|
1712 hash = test->hash(); |
|
1713 assert (test->is_neutral(), "invariant") ; |
|
1714 assert (hash != 0, "Trivial unexpected object/monitor header usage."); |
|
1715 } |
|
1716 } |
|
1717 // We finally get the hash |
|
1718 return hash; |
|
1719 } |
|
1720 |
|
1721 // Deprecated -- use FastHashCode() instead. |
|
1722 |
|
1723 intptr_t ObjectSynchronizer::identity_hash_value_for(Handle obj) { |
|
1724 return FastHashCode (Thread::current(), obj()) ; |
|
1725 } |
|
1726 |
|
1727 bool ObjectSynchronizer::current_thread_holds_lock(JavaThread* thread, |
|
1728 Handle h_obj) { |
|
1729 if (UseBiasedLocking) { |
|
1730 BiasedLocking::revoke_and_rebias(h_obj, false, thread); |
|
1731 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1732 } |
|
1733 |
|
1734 assert(thread == JavaThread::current(), "Can only be called on current thread"); |
|
1735 oop obj = h_obj(); |
|
1736 |
|
1737 markOop mark = ReadStableMark (obj) ; |
|
1738 |
|
1739 // Uncontended case, header points to stack |
|
1740 if (mark->has_locker()) { |
|
1741 return thread->is_lock_owned((address)mark->locker()); |
|
1742 } |
|
1743 // Contended case, header points to ObjectMonitor (tagged pointer) |
|
1744 if (mark->has_monitor()) { |
|
1745 ObjectMonitor* monitor = mark->monitor(); |
|
1746 return monitor->is_entered(thread) != 0 ; |
|
1747 } |
|
1748 // Unlocked case, header in place |
|
1749 assert(mark->is_neutral(), "sanity check"); |
|
1750 return false; |
|
1751 } |
|
1752 |
|
1753 // Be aware of this method could revoke bias of the lock object. |
|
1754 // This method querys the ownership of the lock handle specified by 'h_obj'. |
|
1755 // If the current thread owns the lock, it returns owner_self. If no |
|
1756 // thread owns the lock, it returns owner_none. Otherwise, it will return |
|
1757 // ower_other. |
|
1758 ObjectSynchronizer::LockOwnership ObjectSynchronizer::query_lock_ownership |
|
1759 (JavaThread *self, Handle h_obj) { |
|
1760 // The caller must beware this method can revoke bias, and |
|
1761 // revocation can result in a safepoint. |
|
1762 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; |
|
1763 assert (self->thread_state() != _thread_blocked , "invariant") ; |
|
1764 |
|
1765 // Possible mark states: neutral, biased, stack-locked, inflated |
|
1766 |
|
1767 if (UseBiasedLocking && h_obj()->mark()->has_bias_pattern()) { |
|
1768 // CASE: biased |
|
1769 BiasedLocking::revoke_and_rebias(h_obj, false, self); |
|
1770 assert(!h_obj->mark()->has_bias_pattern(), |
|
1771 "biases should be revoked by now"); |
|
1772 } |
|
1773 |
|
1774 assert(self == JavaThread::current(), "Can only be called on current thread"); |
|
1775 oop obj = h_obj(); |
|
1776 markOop mark = ReadStableMark (obj) ; |
|
1777 |
|
1778 // CASE: stack-locked. Mark points to a BasicLock on the owner's stack. |
|
1779 if (mark->has_locker()) { |
|
1780 return self->is_lock_owned((address)mark->locker()) ? |
|
1781 owner_self : owner_other; |
|
1782 } |
|
1783 |
|
1784 // CASE: inflated. Mark (tagged pointer) points to an objectMonitor. |
|
1785 // The Object:ObjectMonitor relationship is stable as long as we're |
|
1786 // not at a safepoint. |
|
1787 if (mark->has_monitor()) { |
|
1788 void * owner = mark->monitor()->_owner ; |
|
1789 if (owner == NULL) return owner_none ; |
|
1790 return (owner == self || |
|
1791 self->is_lock_owned((address)owner)) ? owner_self : owner_other; |
|
1792 } |
|
1793 |
|
1794 // CASE: neutral |
|
1795 assert(mark->is_neutral(), "sanity check"); |
|
1796 return owner_none ; // it's unlocked |
|
1797 } |
|
1798 |
|
1799 // FIXME: jvmti should call this |
|
1800 JavaThread* ObjectSynchronizer::get_lock_owner(Handle h_obj, bool doLock) { |
|
1801 if (UseBiasedLocking) { |
|
1802 if (SafepointSynchronize::is_at_safepoint()) { |
|
1803 BiasedLocking::revoke_at_safepoint(h_obj); |
|
1804 } else { |
|
1805 BiasedLocking::revoke_and_rebias(h_obj, false, JavaThread::current()); |
|
1806 } |
|
1807 assert(!h_obj->mark()->has_bias_pattern(), "biases should be revoked by now"); |
|
1808 } |
|
1809 |
|
1810 oop obj = h_obj(); |
|
1811 address owner = NULL; |
|
1812 |
|
1813 markOop mark = ReadStableMark (obj) ; |
|
1814 |
|
1815 // Uncontended case, header points to stack |
|
1816 if (mark->has_locker()) { |
|
1817 owner = (address) mark->locker(); |
|
1818 } |
|
1819 |
|
1820 // Contended case, header points to ObjectMonitor (tagged pointer) |
|
1821 if (mark->has_monitor()) { |
|
1822 ObjectMonitor* monitor = mark->monitor(); |
|
1823 assert(monitor != NULL, "monitor should be non-null"); |
|
1824 owner = (address) monitor->owner(); |
|
1825 } |
|
1826 |
|
1827 if (owner != NULL) { |
|
1828 return Threads::owning_thread_from_monitor_owner(owner, doLock); |
|
1829 } |
|
1830 |
|
1831 // Unlocked case, header in place |
|
1832 // Cannot have assertion since this object may have been |
|
1833 // locked by another thread when reaching here. |
|
1834 // assert(mark->is_neutral(), "sanity check"); |
|
1835 |
|
1836 return NULL; |
|
1837 } |
|
1838 |
|
1839 // Iterate through monitor cache and attempt to release thread's monitors |
|
1840 // Gives up on a particular monitor if an exception occurs, but continues |
|
1841 // the overall iteration, swallowing the exception. |
|
1842 class ReleaseJavaMonitorsClosure: public MonitorClosure { |
|
1843 private: |
|
1844 TRAPS; |
|
1845 |
|
1846 public: |
|
1847 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} |
|
1848 void do_monitor(ObjectMonitor* mid) { |
|
1849 if (mid->owner() == THREAD) { |
|
1850 (void)mid->complete_exit(CHECK); |
|
1851 } |
|
1852 } |
|
1853 }; |
|
1854 |
|
1855 // Release all inflated monitors owned by THREAD. Lightweight monitors are |
|
1856 // ignored. This is meant to be called during JNI thread detach which assumes |
|
1857 // all remaining monitors are heavyweight. All exceptions are swallowed. |
|
1858 // Scanning the extant monitor list can be time consuming. |
|
1859 // A simple optimization is to add a per-thread flag that indicates a thread |
|
1860 // called jni_monitorenter() during its lifetime. |
|
1861 // |
|
1862 // Instead of No_Savepoint_Verifier it might be cheaper to |
|
1863 // use an idiom of the form: |
|
1864 // auto int tmp = SafepointSynchronize::_safepoint_counter ; |
|
1865 // <code that must not run at safepoint> |
|
1866 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; |
|
1867 // Since the tests are extremely cheap we could leave them enabled |
|
1868 // for normal product builds. |
|
1869 |
|
1870 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { |
|
1871 assert(THREAD == JavaThread::current(), "must be current Java thread"); |
|
1872 No_Safepoint_Verifier nsv ; |
|
1873 ReleaseJavaMonitorsClosure rjmc(THREAD); |
|
1874 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread"); |
|
1875 ObjectSynchronizer::monitors_iterate(&rjmc); |
|
1876 Thread::muxRelease(&ListLock); |
|
1877 THREAD->clear_pending_exception(); |
|
1878 } |
|
1879 |
|
1880 // Visitors ... |
|
1881 |
|
1882 void ObjectSynchronizer::monitors_iterate(MonitorClosure* closure) { |
|
1883 ObjectMonitor* block = gBlockList; |
|
1884 ObjectMonitor* mid; |
|
1885 while (block) { |
|
1886 assert(block->object() == CHAINMARKER, "must be a block header"); |
|
1887 for (int i = _BLOCKSIZE - 1; i > 0; i--) { |
|
1888 mid = block + i; |
|
1889 oop object = (oop) mid->object(); |
|
1890 if (object != NULL) { |
|
1891 closure->do_monitor(mid); |
|
1892 } |
|
1893 } |
|
1894 block = (ObjectMonitor*) block->FreeNext; |
|
1895 } |
|
1896 } |
|
1897 |
|
1898 void ObjectSynchronizer::oops_do(OopClosure* f) { |
|
1899 assert(SafepointSynchronize::is_at_safepoint(), "must be at safepoint"); |
|
1900 for (ObjectMonitor* block = gBlockList; block != NULL; block = next(block)) { |
|
1901 assert(block->object() == CHAINMARKER, "must be a block header"); |
|
1902 for (int i = 1; i < _BLOCKSIZE; i++) { |
|
1903 ObjectMonitor* mid = &block[i]; |
|
1904 if (mid->object() != NULL) { |
|
1905 f->do_oop((oop*)mid->object_addr()); |
|
1906 } |
|
1907 } |
|
1908 } |
|
1909 } |
|
1910 |
1331 |
1911 // Deflate_idle_monitors() is called at all safepoints, immediately |
1332 // Deflate_idle_monitors() is called at all safepoints, immediately |
1912 // after all mutators are stopped, but before any objects have moved. |
1333 // after all mutators are stopped, but before any objects have moved. |
1913 // It traverses the list of known monitors, deflating where possible. |
1334 // It traverses the list of known monitors, deflating where possible. |
1914 // The scavenged monitor are returned to the monitor free list. |
1335 // The scavenged monitor are returned to the monitor free list. |
2105 FreeTail->FreeNext = gFreeList ; |
1525 FreeTail->FreeNext = gFreeList ; |
2106 gFreeList = FreeHead ; |
1526 gFreeList = FreeHead ; |
2107 } |
1527 } |
2108 Thread::muxRelease (&ListLock) ; |
1528 Thread::muxRelease (&ListLock) ; |
2109 |
1529 |
2110 if (_sync_Deflations != NULL) _sync_Deflations->inc(nScavenged) ; |
1530 if (ObjectMonitor::_sync_Deflations != NULL) ObjectMonitor::_sync_Deflations->inc(nScavenged) ; |
2111 if (_sync_MonExtant != NULL) _sync_MonExtant ->set_value(nInCirculation); |
1531 if (ObjectMonitor::_sync_MonExtant != NULL) ObjectMonitor::_sync_MonExtant ->set_value(nInCirculation); |
2112 |
1532 |
2113 // TODO: Add objectMonitor leak detection. |
1533 // TODO: Add objectMonitor leak detection. |
2114 // Audit/inventory the objectMonitors -- make sure they're all accounted for. |
1534 // Audit/inventory the objectMonitors -- make sure they're all accounted for. |
2115 GVars.stwRandom = os::random() ; |
1535 GVars.stwRandom = os::random() ; |
2116 GVars.stwCycle ++ ; |
1536 GVars.stwCycle ++ ; |
2117 } |
1537 } |
2118 |
1538 |
2119 // A macro is used below because there may already be a pending |
1539 // Monitor cleanup on JavaThread::exit |
2120 // exception which should not abort the execution of the routines |
1540 |
2121 // which use this (which is why we don't put this into check_slow and |
1541 // Iterate through monitor cache and attempt to release thread's monitors |
2122 // call it with a CHECK argument). |
1542 // Gives up on a particular monitor if an exception occurs, but continues |
2123 |
1543 // the overall iteration, swallowing the exception. |
2124 #define CHECK_OWNER() \ |
1544 class ReleaseJavaMonitorsClosure: public MonitorClosure { |
2125 do { \ |
1545 private: |
2126 if (THREAD != _owner) { \ |
1546 TRAPS; |
2127 if (THREAD->is_lock_owned((address) _owner)) { \ |
1547 |
2128 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ \ |
1548 public: |
2129 _recursions = 0; \ |
1549 ReleaseJavaMonitorsClosure(Thread* thread) : THREAD(thread) {} |
2130 OwnerIsThread = 1 ; \ |
1550 void do_monitor(ObjectMonitor* mid) { |
2131 } else { \ |
1551 if (mid->owner() == THREAD) { |
2132 TEVENT (Throw IMSX) ; \ |
1552 (void)mid->complete_exit(CHECK); |
2133 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); \ |
1553 } |
2134 } \ |
1554 } |
2135 } \ |
1555 }; |
2136 } while (false) |
1556 |
2137 |
1557 // Release all inflated monitors owned by THREAD. Lightweight monitors are |
2138 // TODO-FIXME: eliminate ObjectWaiters. Replace this visitor/enumerator |
1558 // ignored. This is meant to be called during JNI thread detach which assumes |
2139 // interface with a simple FirstWaitingThread(), NextWaitingThread() interface. |
1559 // all remaining monitors are heavyweight. All exceptions are swallowed. |
2140 |
1560 // Scanning the extant monitor list can be time consuming. |
2141 ObjectWaiter* ObjectMonitor::first_waiter() { |
1561 // A simple optimization is to add a per-thread flag that indicates a thread |
2142 return _WaitSet; |
1562 // called jni_monitorenter() during its lifetime. |
2143 } |
1563 // |
2144 |
1564 // Instead of No_Savepoint_Verifier it might be cheaper to |
2145 ObjectWaiter* ObjectMonitor::next_waiter(ObjectWaiter* o) { |
1565 // use an idiom of the form: |
2146 return o->_next; |
1566 // auto int tmp = SafepointSynchronize::_safepoint_counter ; |
2147 } |
1567 // <code that must not run at safepoint> |
2148 |
1568 // guarantee (((tmp ^ _safepoint_counter) | (tmp & 1)) == 0) ; |
2149 Thread* ObjectMonitor::thread_of_waiter(ObjectWaiter* o) { |
1569 // Since the tests are extremely cheap we could leave them enabled |
2150 return o->_thread; |
1570 // for normal product builds. |
2151 } |
1571 |
2152 |
1572 void ObjectSynchronizer::release_monitors_owned_by_thread(TRAPS) { |
2153 // initialize the monitor, exception the semaphore, all other fields |
1573 assert(THREAD == JavaThread::current(), "must be current Java thread"); |
2154 // are simple integers or pointers |
1574 No_Safepoint_Verifier nsv ; |
2155 ObjectMonitor::ObjectMonitor() { |
1575 ReleaseJavaMonitorsClosure rjmc(THREAD); |
2156 _header = NULL; |
1576 Thread::muxAcquire(&ListLock, "release_monitors_owned_by_thread"); |
2157 _count = 0; |
1577 ObjectSynchronizer::monitors_iterate(&rjmc); |
2158 _waiters = 0, |
1578 Thread::muxRelease(&ListLock); |
2159 _recursions = 0; |
1579 THREAD->clear_pending_exception(); |
2160 _object = NULL; |
1580 } |
2161 _owner = NULL; |
|
2162 _WaitSet = NULL; |
|
2163 _WaitSetLock = 0 ; |
|
2164 _Responsible = NULL ; |
|
2165 _succ = NULL ; |
|
2166 _cxq = NULL ; |
|
2167 FreeNext = NULL ; |
|
2168 _EntryList = NULL ; |
|
2169 _SpinFreq = 0 ; |
|
2170 _SpinClock = 0 ; |
|
2171 OwnerIsThread = 0 ; |
|
2172 } |
|
2173 |
|
2174 ObjectMonitor::~ObjectMonitor() { |
|
2175 // TODO: Add asserts ... |
|
2176 // _cxq == 0 _succ == NULL _owner == NULL _waiters == 0 |
|
2177 // _count == 0 _EntryList == NULL etc |
|
2178 } |
|
2179 |
|
2180 intptr_t ObjectMonitor::is_busy() const { |
|
2181 // TODO-FIXME: merge _count and _waiters. |
|
2182 // TODO-FIXME: assert _owner == null implies _recursions = 0 |
|
2183 // TODO-FIXME: assert _WaitSet != null implies _count > 0 |
|
2184 return _count|_waiters|intptr_t(_owner)|intptr_t(_cxq)|intptr_t(_EntryList ) ; |
|
2185 } |
|
2186 |
|
2187 void ObjectMonitor::Recycle () { |
|
2188 // TODO: add stronger asserts ... |
|
2189 // _cxq == 0 _succ == NULL _owner == NULL _waiters == 0 |
|
2190 // _count == 0 EntryList == NULL |
|
2191 // _recursions == 0 _WaitSet == NULL |
|
2192 // TODO: assert (is_busy()|_recursions) == 0 |
|
2193 _succ = NULL ; |
|
2194 _EntryList = NULL ; |
|
2195 _cxq = NULL ; |
|
2196 _WaitSet = NULL ; |
|
2197 _recursions = 0 ; |
|
2198 _SpinFreq = 0 ; |
|
2199 _SpinClock = 0 ; |
|
2200 OwnerIsThread = 0 ; |
|
2201 } |
|
2202 |
|
2203 // WaitSet management ... |
|
2204 |
|
2205 inline void ObjectMonitor::AddWaiter(ObjectWaiter* node) { |
|
2206 assert(node != NULL, "should not dequeue NULL node"); |
|
2207 assert(node->_prev == NULL, "node already in list"); |
|
2208 assert(node->_next == NULL, "node already in list"); |
|
2209 // put node at end of queue (circular doubly linked list) |
|
2210 if (_WaitSet == NULL) { |
|
2211 _WaitSet = node; |
|
2212 node->_prev = node; |
|
2213 node->_next = node; |
|
2214 } else { |
|
2215 ObjectWaiter* head = _WaitSet ; |
|
2216 ObjectWaiter* tail = head->_prev; |
|
2217 assert(tail->_next == head, "invariant check"); |
|
2218 tail->_next = node; |
|
2219 head->_prev = node; |
|
2220 node->_next = head; |
|
2221 node->_prev = tail; |
|
2222 } |
|
2223 } |
|
2224 |
|
2225 inline ObjectWaiter* ObjectMonitor::DequeueWaiter() { |
|
2226 // dequeue the very first waiter |
|
2227 ObjectWaiter* waiter = _WaitSet; |
|
2228 if (waiter) { |
|
2229 DequeueSpecificWaiter(waiter); |
|
2230 } |
|
2231 return waiter; |
|
2232 } |
|
2233 |
|
2234 inline void ObjectMonitor::DequeueSpecificWaiter(ObjectWaiter* node) { |
|
2235 assert(node != NULL, "should not dequeue NULL node"); |
|
2236 assert(node->_prev != NULL, "node already removed from list"); |
|
2237 assert(node->_next != NULL, "node already removed from list"); |
|
2238 // when the waiter has woken up because of interrupt, |
|
2239 // timeout or other spurious wake-up, dequeue the |
|
2240 // waiter from waiting list |
|
2241 ObjectWaiter* next = node->_next; |
|
2242 if (next == node) { |
|
2243 assert(node->_prev == node, "invariant check"); |
|
2244 _WaitSet = NULL; |
|
2245 } else { |
|
2246 ObjectWaiter* prev = node->_prev; |
|
2247 assert(prev->_next == node, "invariant check"); |
|
2248 assert(next->_prev == node, "invariant check"); |
|
2249 next->_prev = prev; |
|
2250 prev->_next = next; |
|
2251 if (_WaitSet == node) { |
|
2252 _WaitSet = next; |
|
2253 } |
|
2254 } |
|
2255 node->_next = NULL; |
|
2256 node->_prev = NULL; |
|
2257 } |
|
2258 |
|
2259 static char * kvGet (char * kvList, const char * Key) { |
|
2260 if (kvList == NULL) return NULL ; |
|
2261 size_t n = strlen (Key) ; |
|
2262 char * Search ; |
|
2263 for (Search = kvList ; *Search ; Search += strlen(Search) + 1) { |
|
2264 if (strncmp (Search, Key, n) == 0) { |
|
2265 if (Search[n] == '=') return Search + n + 1 ; |
|
2266 if (Search[n] == 0) return (char *) "1" ; |
|
2267 } |
|
2268 } |
|
2269 return NULL ; |
|
2270 } |
|
2271 |
|
2272 static int kvGetInt (char * kvList, const char * Key, int Default) { |
|
2273 char * v = kvGet (kvList, Key) ; |
|
2274 int rslt = v ? ::strtol (v, NULL, 0) : Default ; |
|
2275 if (Knob_ReportSettings && v != NULL) { |
|
2276 ::printf (" SyncKnob: %s %d(%d)\n", Key, rslt, Default) ; |
|
2277 ::fflush (stdout) ; |
|
2278 } |
|
2279 return rslt ; |
|
2280 } |
|
2281 |
|
2282 // By convention we unlink a contending thread from EntryList|cxq immediately |
|
2283 // after the thread acquires the lock in ::enter(). Equally, we could defer |
|
2284 // unlinking the thread until ::exit()-time. |
|
2285 |
|
2286 void ObjectMonitor::UnlinkAfterAcquire (Thread * Self, ObjectWaiter * SelfNode) |
|
2287 { |
|
2288 assert (_owner == Self, "invariant") ; |
|
2289 assert (SelfNode->_thread == Self, "invariant") ; |
|
2290 |
|
2291 if (SelfNode->TState == ObjectWaiter::TS_ENTER) { |
|
2292 // Normal case: remove Self from the DLL EntryList . |
|
2293 // This is a constant-time operation. |
|
2294 ObjectWaiter * nxt = SelfNode->_next ; |
|
2295 ObjectWaiter * prv = SelfNode->_prev ; |
|
2296 if (nxt != NULL) nxt->_prev = prv ; |
|
2297 if (prv != NULL) prv->_next = nxt ; |
|
2298 if (SelfNode == _EntryList ) _EntryList = nxt ; |
|
2299 assert (nxt == NULL || nxt->TState == ObjectWaiter::TS_ENTER, "invariant") ; |
|
2300 assert (prv == NULL || prv->TState == ObjectWaiter::TS_ENTER, "invariant") ; |
|
2301 TEVENT (Unlink from EntryList) ; |
|
2302 } else { |
|
2303 guarantee (SelfNode->TState == ObjectWaiter::TS_CXQ, "invariant") ; |
|
2304 // Inopportune interleaving -- Self is still on the cxq. |
|
2305 // This usually means the enqueue of self raced an exiting thread. |
|
2306 // Normally we'll find Self near the front of the cxq, so |
|
2307 // dequeueing is typically fast. If needbe we can accelerate |
|
2308 // this with some MCS/CHL-like bidirectional list hints and advisory |
|
2309 // back-links so dequeueing from the interior will normally operate |
|
2310 // in constant-time. |
|
2311 // Dequeue Self from either the head (with CAS) or from the interior |
|
2312 // with a linear-time scan and normal non-atomic memory operations. |
|
2313 // CONSIDER: if Self is on the cxq then simply drain cxq into EntryList |
|
2314 // and then unlink Self from EntryList. We have to drain eventually, |
|
2315 // so it might as well be now. |
|
2316 |
|
2317 ObjectWaiter * v = _cxq ; |
|
2318 assert (v != NULL, "invariant") ; |
|
2319 if (v != SelfNode || Atomic::cmpxchg_ptr (SelfNode->_next, &_cxq, v) != v) { |
|
2320 // The CAS above can fail from interference IFF a "RAT" arrived. |
|
2321 // In that case Self must be in the interior and can no longer be |
|
2322 // at the head of cxq. |
|
2323 if (v == SelfNode) { |
|
2324 assert (_cxq != v, "invariant") ; |
|
2325 v = _cxq ; // CAS above failed - start scan at head of list |
|
2326 } |
|
2327 ObjectWaiter * p ; |
|
2328 ObjectWaiter * q = NULL ; |
|
2329 for (p = v ; p != NULL && p != SelfNode; p = p->_next) { |
|
2330 q = p ; |
|
2331 assert (p->TState == ObjectWaiter::TS_CXQ, "invariant") ; |
|
2332 } |
|
2333 assert (v != SelfNode, "invariant") ; |
|
2334 assert (p == SelfNode, "Node not found on cxq") ; |
|
2335 assert (p != _cxq, "invariant") ; |
|
2336 assert (q != NULL, "invariant") ; |
|
2337 assert (q->_next == p, "invariant") ; |
|
2338 q->_next = p->_next ; |
|
2339 } |
|
2340 TEVENT (Unlink from cxq) ; |
|
2341 } |
|
2342 |
|
2343 // Diagnostic hygiene ... |
|
2344 SelfNode->_prev = (ObjectWaiter *) 0xBAD ; |
|
2345 SelfNode->_next = (ObjectWaiter *) 0xBAD ; |
|
2346 SelfNode->TState = ObjectWaiter::TS_RUN ; |
|
2347 } |
|
2348 |
|
2349 // Caveat: TryLock() is not necessarily serializing if it returns failure. |
|
2350 // Callers must compensate as needed. |
|
2351 |
|
2352 int ObjectMonitor::TryLock (Thread * Self) { |
|
2353 for (;;) { |
|
2354 void * own = _owner ; |
|
2355 if (own != NULL) return 0 ; |
|
2356 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { |
|
2357 // Either guarantee _recursions == 0 or set _recursions = 0. |
|
2358 assert (_recursions == 0, "invariant") ; |
|
2359 assert (_owner == Self, "invariant") ; |
|
2360 // CONSIDER: set or assert that OwnerIsThread == 1 |
|
2361 return 1 ; |
|
2362 } |
|
2363 // The lock had been free momentarily, but we lost the race to the lock. |
|
2364 // Interference -- the CAS failed. |
|
2365 // We can either return -1 or retry. |
|
2366 // Retry doesn't make as much sense because the lock was just acquired. |
|
2367 if (true) return -1 ; |
|
2368 } |
|
2369 } |
|
2370 |
|
2371 // NotRunnable() -- informed spinning |
|
2372 // |
|
2373 // Don't bother spinning if the owner is not eligible to drop the lock. |
|
2374 // Peek at the owner's schedctl.sc_state and Thread._thread_values and |
|
2375 // spin only if the owner thread is _thread_in_Java or _thread_in_vm. |
|
2376 // The thread must be runnable in order to drop the lock in timely fashion. |
|
2377 // If the _owner is not runnable then spinning will not likely be |
|
2378 // successful (profitable). |
|
2379 // |
|
2380 // Beware -- the thread referenced by _owner could have died |
|
2381 // so a simply fetch from _owner->_thread_state might trap. |
|
2382 // Instead, we use SafeFetchXX() to safely LD _owner->_thread_state. |
|
2383 // Because of the lifecycle issues the schedctl and _thread_state values |
|
2384 // observed by NotRunnable() might be garbage. NotRunnable must |
|
2385 // tolerate this and consider the observed _thread_state value |
|
2386 // as advisory. |
|
2387 // |
|
2388 // Beware too, that _owner is sometimes a BasicLock address and sometimes |
|
2389 // a thread pointer. We differentiate the two cases with OwnerIsThread. |
|
2390 // Alternately, we might tag the type (thread pointer vs basiclock pointer) |
|
2391 // with the LSB of _owner. Another option would be to probablistically probe |
|
2392 // the putative _owner->TypeTag value. |
|
2393 // |
|
2394 // Checking _thread_state isn't perfect. Even if the thread is |
|
2395 // in_java it might be blocked on a page-fault or have been preempted |
|
2396 // and sitting on a ready/dispatch queue. _thread state in conjunction |
|
2397 // with schedctl.sc_state gives us a good picture of what the |
|
2398 // thread is doing, however. |
|
2399 // |
|
2400 // TODO: check schedctl.sc_state. |
|
2401 // We'll need to use SafeFetch32() to read from the schedctl block. |
|
2402 // See RFE #5004247 and http://sac.sfbay.sun.com/Archives/CaseLog/arc/PSARC/2005/351/ |
|
2403 // |
|
2404 // The return value from NotRunnable() is *advisory* -- the |
|
2405 // result is based on sampling and is not necessarily coherent. |
|
2406 // The caller must tolerate false-negative and false-positive errors. |
|
2407 // Spinning, in general, is probabilistic anyway. |
|
2408 |
|
2409 |
|
2410 int ObjectMonitor::NotRunnable (Thread * Self, Thread * ox) { |
|
2411 // Check either OwnerIsThread or ox->TypeTag == 2BAD. |
|
2412 if (!OwnerIsThread) return 0 ; |
|
2413 |
|
2414 if (ox == NULL) return 0 ; |
|
2415 |
|
2416 // Avoid transitive spinning ... |
|
2417 // Say T1 spins or blocks trying to acquire L. T1._Stalled is set to L. |
|
2418 // Immediately after T1 acquires L it's possible that T2, also |
|
2419 // spinning on L, will see L.Owner=T1 and T1._Stalled=L. |
|
2420 // This occurs transiently after T1 acquired L but before |
|
2421 // T1 managed to clear T1.Stalled. T2 does not need to abort |
|
2422 // its spin in this circumstance. |
|
2423 intptr_t BlockedOn = SafeFetchN ((intptr_t *) &ox->_Stalled, intptr_t(1)) ; |
|
2424 |
|
2425 if (BlockedOn == 1) return 1 ; |
|
2426 if (BlockedOn != 0) { |
|
2427 return BlockedOn != intptr_t(this) && _owner == ox ; |
|
2428 } |
|
2429 |
|
2430 assert (sizeof(((JavaThread *)ox)->_thread_state == sizeof(int)), "invariant") ; |
|
2431 int jst = SafeFetch32 ((int *) &((JavaThread *) ox)->_thread_state, -1) ; ; |
|
2432 // consider also: jst != _thread_in_Java -- but that's overspecific. |
|
2433 return jst == _thread_blocked || jst == _thread_in_native ; |
|
2434 } |
|
2435 |
|
2436 |
|
2437 // Adaptive spin-then-block - rational spinning |
|
2438 // |
|
2439 // Note that we spin "globally" on _owner with a classic SMP-polite TATAS |
|
2440 // algorithm. On high order SMP systems it would be better to start with |
|
2441 // a brief global spin and then revert to spinning locally. In the spirit of MCS/CLH, |
|
2442 // a contending thread could enqueue itself on the cxq and then spin locally |
|
2443 // on a thread-specific variable such as its ParkEvent._Event flag. |
|
2444 // That's left as an exercise for the reader. Note that global spinning is |
|
2445 // not problematic on Niagara, as the L2$ serves the interconnect and has both |
|
2446 // low latency and massive bandwidth. |
|
2447 // |
|
2448 // Broadly, we can fix the spin frequency -- that is, the % of contended lock |
|
2449 // acquisition attempts where we opt to spin -- at 100% and vary the spin count |
|
2450 // (duration) or we can fix the count at approximately the duration of |
|
2451 // a context switch and vary the frequency. Of course we could also |
|
2452 // vary both satisfying K == Frequency * Duration, where K is adaptive by monitor. |
|
2453 // See http://j2se.east/~dice/PERSIST/040824-AdaptiveSpinning.html. |
|
2454 // |
|
2455 // This implementation varies the duration "D", where D varies with |
|
2456 // the success rate of recent spin attempts. (D is capped at approximately |
|
2457 // length of a round-trip context switch). The success rate for recent |
|
2458 // spin attempts is a good predictor of the success rate of future spin |
|
2459 // attempts. The mechanism adapts automatically to varying critical |
|
2460 // section length (lock modality), system load and degree of parallelism. |
|
2461 // D is maintained per-monitor in _SpinDuration and is initialized |
|
2462 // optimistically. Spin frequency is fixed at 100%. |
|
2463 // |
|
2464 // Note that _SpinDuration is volatile, but we update it without locks |
|
2465 // or atomics. The code is designed so that _SpinDuration stays within |
|
2466 // a reasonable range even in the presence of races. The arithmetic |
|
2467 // operations on _SpinDuration are closed over the domain of legal values, |
|
2468 // so at worst a race will install and older but still legal value. |
|
2469 // At the very worst this introduces some apparent non-determinism. |
|
2470 // We might spin when we shouldn't or vice-versa, but since the spin |
|
2471 // count are relatively short, even in the worst case, the effect is harmless. |
|
2472 // |
|
2473 // Care must be taken that a low "D" value does not become an |
|
2474 // an absorbing state. Transient spinning failures -- when spinning |
|
2475 // is overall profitable -- should not cause the system to converge |
|
2476 // on low "D" values. We want spinning to be stable and predictable |
|
2477 // and fairly responsive to change and at the same time we don't want |
|
2478 // it to oscillate, become metastable, be "too" non-deterministic, |
|
2479 // or converge on or enter undesirable stable absorbing states. |
|
2480 // |
|
2481 // We implement a feedback-based control system -- using past behavior |
|
2482 // to predict future behavior. We face two issues: (a) if the |
|
2483 // input signal is random then the spin predictor won't provide optimal |
|
2484 // results, and (b) if the signal frequency is too high then the control |
|
2485 // system, which has some natural response lag, will "chase" the signal. |
|
2486 // (b) can arise from multimodal lock hold times. Transient preemption |
|
2487 // can also result in apparent bimodal lock hold times. |
|
2488 // Although sub-optimal, neither condition is particularly harmful, as |
|
2489 // in the worst-case we'll spin when we shouldn't or vice-versa. |
|
2490 // The maximum spin duration is rather short so the failure modes aren't bad. |
|
2491 // To be conservative, I've tuned the gain in system to bias toward |
|
2492 // _not spinning. Relatedly, the system can sometimes enter a mode where it |
|
2493 // "rings" or oscillates between spinning and not spinning. This happens |
|
2494 // when spinning is just on the cusp of profitability, however, so the |
|
2495 // situation is not dire. The state is benign -- there's no need to add |
|
2496 // hysteresis control to damp the transition rate between spinning and |
|
2497 // not spinning. |
|
2498 // |
|
2499 // - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - |
|
2500 // |
|
2501 // Spin-then-block strategies ... |
|
2502 // |
|
2503 // Thoughts on ways to improve spinning : |
|
2504 // |
|
2505 // * Periodically call {psr_}getloadavg() while spinning, and |
|
2506 // permit unbounded spinning if the load average is < |
|
2507 // the number of processors. Beware, however, that getloadavg() |
|
2508 // is exceptionally fast on solaris (about 1/10 the cost of a full |
|
2509 // spin cycle, but quite expensive on linux. Beware also, that |
|
2510 // multiple JVMs could "ring" or oscillate in a feedback loop. |
|
2511 // Sufficient damping would solve that problem. |
|
2512 // |
|
2513 // * We currently use spin loops with iteration counters to approximate |
|
2514 // spinning for some interval. Given the availability of high-precision |
|
2515 // time sources such as gethrtime(), %TICK, %STICK, RDTSC, etc., we should |
|
2516 // someday reimplement the spin loops to duration-based instead of iteration-based. |
|
2517 // |
|
2518 // * Don't spin if there are more than N = (CPUs/2) threads |
|
2519 // currently spinning on the monitor (or globally). |
|
2520 // That is, limit the number of concurrent spinners. |
|
2521 // We might also limit the # of spinners in the JVM, globally. |
|
2522 // |
|
2523 // * If a spinning thread observes _owner change hands it should |
|
2524 // abort the spin (and park immediately) or at least debit |
|
2525 // the spin counter by a large "penalty". |
|
2526 // |
|
2527 // * Classically, the spin count is either K*(CPUs-1) or is a |
|
2528 // simple constant that approximates the length of a context switch. |
|
2529 // We currently use a value -- computed by a special utility -- that |
|
2530 // approximates round-trip context switch times. |
|
2531 // |
|
2532 // * Normally schedctl_start()/_stop() is used to advise the kernel |
|
2533 // to avoid preempting threads that are running in short, bounded |
|
2534 // critical sections. We could use the schedctl hooks in an inverted |
|
2535 // sense -- spinners would set the nopreempt flag, but poll the preempt |
|
2536 // pending flag. If a spinner observed a pending preemption it'd immediately |
|
2537 // abort the spin and park. As such, the schedctl service acts as |
|
2538 // a preemption warning mechanism. |
|
2539 // |
|
2540 // * In lieu of spinning, if the system is running below saturation |
|
2541 // (that is, loadavg() << #cpus), we can instead suppress futile |
|
2542 // wakeup throttling, or even wake more than one successor at exit-time. |
|
2543 // The net effect is largely equivalent to spinning. In both cases, |
|
2544 // contending threads go ONPROC and opportunistically attempt to acquire |
|
2545 // the lock, decreasing lock handover latency at the expense of wasted |
|
2546 // cycles and context switching. |
|
2547 // |
|
2548 // * We might to spin less after we've parked as the thread will |
|
2549 // have less $ and TLB affinity with the processor. |
|
2550 // Likewise, we might spin less if we come ONPROC on a different |
|
2551 // processor or after a long period (>> rechose_interval). |
|
2552 // |
|
2553 // * A table-driven state machine similar to Solaris' dispadmin scheduling |
|
2554 // tables might be a better design. Instead of encoding information in |
|
2555 // _SpinDuration, _SpinFreq and _SpinClock we'd just use explicit, |
|
2556 // discrete states. Success or failure during a spin would drive |
|
2557 // state transitions, and each state node would contain a spin count. |
|
2558 // |
|
2559 // * If the processor is operating in a mode intended to conserve power |
|
2560 // (such as Intel's SpeedStep) or to reduce thermal output (thermal |
|
2561 // step-down mode) then the Java synchronization subsystem should |
|
2562 // forgo spinning. |
|
2563 // |
|
2564 // * The minimum spin duration should be approximately the worst-case |
|
2565 // store propagation latency on the platform. That is, the time |
|
2566 // it takes a store on CPU A to become visible on CPU B, where A and |
|
2567 // B are "distant". |
|
2568 // |
|
2569 // * We might want to factor a thread's priority in the spin policy. |
|
2570 // Threads with a higher priority might spin for slightly longer. |
|
2571 // Similarly, if we use back-off in the TATAS loop, lower priority |
|
2572 // threads might back-off longer. We don't currently use a |
|
2573 // thread's priority when placing it on the entry queue. We may |
|
2574 // want to consider doing so in future releases. |
|
2575 // |
|
2576 // * We might transiently drop a thread's scheduling priority while it spins. |
|
2577 // SCHED_BATCH on linux and FX scheduling class at priority=0 on Solaris |
|
2578 // would suffice. We could even consider letting the thread spin indefinitely at |
|
2579 // a depressed or "idle" priority. This brings up fairness issues, however -- |
|
2580 // in a saturated system a thread would with a reduced priority could languish |
|
2581 // for extended periods on the ready queue. |
|
2582 // |
|
2583 // * While spinning try to use the otherwise wasted time to help the VM make |
|
2584 // progress: |
|
2585 // |
|
2586 // -- YieldTo() the owner, if the owner is OFFPROC but ready |
|
2587 // Done our remaining quantum directly to the ready thread. |
|
2588 // This helps "push" the lock owner through the critical section. |
|
2589 // It also tends to improve affinity/locality as the lock |
|
2590 // "migrates" less frequently between CPUs. |
|
2591 // -- Walk our own stack in anticipation of blocking. Memoize the roots. |
|
2592 // -- Perform strand checking for other thread. Unpark potential strandees. |
|
2593 // -- Help GC: trace or mark -- this would need to be a bounded unit of work. |
|
2594 // Unfortunately this will pollute our $ and TLBs. Recall that we |
|
2595 // spin to avoid context switching -- context switching has an |
|
2596 // immediate cost in latency, a disruptive cost to other strands on a CMT |
|
2597 // processor, and an amortized cost because of the D$ and TLB cache |
|
2598 // reload transient when the thread comes back ONPROC and repopulates |
|
2599 // $s and TLBs. |
|
2600 // -- call getloadavg() to see if the system is saturated. It'd probably |
|
2601 // make sense to call getloadavg() half way through the spin. |
|
2602 // If the system isn't at full capacity the we'd simply reset |
|
2603 // the spin counter to and extend the spin attempt. |
|
2604 // -- Doug points out that we should use the same "helping" policy |
|
2605 // in thread.yield(). |
|
2606 // |
|
2607 // * Try MONITOR-MWAIT on systems that support those instructions. |
|
2608 // |
|
2609 // * The spin statistics that drive spin decisions & frequency are |
|
2610 // maintained in the objectmonitor structure so if we deflate and reinflate |
|
2611 // we lose spin state. In practice this is not usually a concern |
|
2612 // as the default spin state after inflation is aggressive (optimistic) |
|
2613 // and tends toward spinning. So in the worst case for a lock where |
|
2614 // spinning is not profitable we may spin unnecessarily for a brief |
|
2615 // period. But then again, if a lock is contended it'll tend not to deflate |
|
2616 // in the first place. |
|
2617 |
|
2618 |
|
2619 intptr_t ObjectMonitor::SpinCallbackArgument = 0 ; |
|
2620 int (*ObjectMonitor::SpinCallbackFunction)(intptr_t, int) = NULL ; |
|
2621 |
|
2622 // Spinning: Fixed frequency (100%), vary duration |
|
2623 |
|
2624 int ObjectMonitor::TrySpin_VaryDuration (Thread * Self) { |
|
2625 |
|
2626 // Dumb, brutal spin. Good for comparative measurements against adaptive spinning. |
|
2627 int ctr = Knob_FixedSpin ; |
|
2628 if (ctr != 0) { |
|
2629 while (--ctr >= 0) { |
|
2630 if (TryLock (Self) > 0) return 1 ; |
|
2631 SpinPause () ; |
|
2632 } |
|
2633 return 0 ; |
|
2634 } |
|
2635 |
|
2636 for (ctr = Knob_PreSpin + 1; --ctr >= 0 ; ) { |
|
2637 if (TryLock(Self) > 0) { |
|
2638 // Increase _SpinDuration ... |
|
2639 // Note that we don't clamp SpinDuration precisely at SpinLimit. |
|
2640 // Raising _SpurDuration to the poverty line is key. |
|
2641 int x = _SpinDuration ; |
|
2642 if (x < Knob_SpinLimit) { |
|
2643 if (x < Knob_Poverty) x = Knob_Poverty ; |
|
2644 _SpinDuration = x + Knob_BonusB ; |
|
2645 } |
|
2646 return 1 ; |
|
2647 } |
|
2648 SpinPause () ; |
|
2649 } |
|
2650 |
|
2651 // Admission control - verify preconditions for spinning |
|
2652 // |
|
2653 // We always spin a little bit, just to prevent _SpinDuration == 0 from |
|
2654 // becoming an absorbing state. Put another way, we spin briefly to |
|
2655 // sample, just in case the system load, parallelism, contention, or lock |
|
2656 // modality changed. |
|
2657 // |
|
2658 // Consider the following alternative: |
|
2659 // Periodically set _SpinDuration = _SpinLimit and try a long/full |
|
2660 // spin attempt. "Periodically" might mean after a tally of |
|
2661 // the # of failed spin attempts (or iterations) reaches some threshold. |
|
2662 // This takes us into the realm of 1-out-of-N spinning, where we |
|
2663 // hold the duration constant but vary the frequency. |
|
2664 |
|
2665 ctr = _SpinDuration ; |
|
2666 if (ctr < Knob_SpinBase) ctr = Knob_SpinBase ; |
|
2667 if (ctr <= 0) return 0 ; |
|
2668 |
|
2669 if (Knob_SuccRestrict && _succ != NULL) return 0 ; |
|
2670 if (Knob_OState && NotRunnable (Self, (Thread *) _owner)) { |
|
2671 TEVENT (Spin abort - notrunnable [TOP]); |
|
2672 return 0 ; |
|
2673 } |
|
2674 |
|
2675 int MaxSpin = Knob_MaxSpinners ; |
|
2676 if (MaxSpin >= 0) { |
|
2677 if (_Spinner > MaxSpin) { |
|
2678 TEVENT (Spin abort -- too many spinners) ; |
|
2679 return 0 ; |
|
2680 } |
|
2681 // Slighty racy, but benign ... |
|
2682 Adjust (&_Spinner, 1) ; |
|
2683 } |
|
2684 |
|
2685 // We're good to spin ... spin ingress. |
|
2686 // CONSIDER: use Prefetch::write() to avoid RTS->RTO upgrades |
|
2687 // when preparing to LD...CAS _owner, etc and the CAS is likely |
|
2688 // to succeed. |
|
2689 int hits = 0 ; |
|
2690 int msk = 0 ; |
|
2691 int caspty = Knob_CASPenalty ; |
|
2692 int oxpty = Knob_OXPenalty ; |
|
2693 int sss = Knob_SpinSetSucc ; |
|
2694 if (sss && _succ == NULL ) _succ = Self ; |
|
2695 Thread * prv = NULL ; |
|
2696 |
|
2697 // There are three ways to exit the following loop: |
|
2698 // 1. A successful spin where this thread has acquired the lock. |
|
2699 // 2. Spin failure with prejudice |
|
2700 // 3. Spin failure without prejudice |
|
2701 |
|
2702 while (--ctr >= 0) { |
|
2703 |
|
2704 // Periodic polling -- Check for pending GC |
|
2705 // Threads may spin while they're unsafe. |
|
2706 // We don't want spinning threads to delay the JVM from reaching |
|
2707 // a stop-the-world safepoint or to steal cycles from GC. |
|
2708 // If we detect a pending safepoint we abort in order that |
|
2709 // (a) this thread, if unsafe, doesn't delay the safepoint, and (b) |
|
2710 // this thread, if safe, doesn't steal cycles from GC. |
|
2711 // This is in keeping with the "no loitering in runtime" rule. |
|
2712 // We periodically check to see if there's a safepoint pending. |
|
2713 if ((ctr & 0xFF) == 0) { |
|
2714 if (SafepointSynchronize::do_call_back()) { |
|
2715 TEVENT (Spin: safepoint) ; |
|
2716 goto Abort ; // abrupt spin egress |
|
2717 } |
|
2718 if (Knob_UsePause & 1) SpinPause () ; |
|
2719 |
|
2720 int (*scb)(intptr_t,int) = SpinCallbackFunction ; |
|
2721 if (hits > 50 && scb != NULL) { |
|
2722 int abend = (*scb)(SpinCallbackArgument, 0) ; |
|
2723 } |
|
2724 } |
|
2725 |
|
2726 if (Knob_UsePause & 2) SpinPause() ; |
|
2727 |
|
2728 // Exponential back-off ... Stay off the bus to reduce coherency traffic. |
|
2729 // This is useful on classic SMP systems, but is of less utility on |
|
2730 // N1-style CMT platforms. |
|
2731 // |
|
2732 // Trade-off: lock acquisition latency vs coherency bandwidth. |
|
2733 // Lock hold times are typically short. A histogram |
|
2734 // of successful spin attempts shows that we usually acquire |
|
2735 // the lock early in the spin. That suggests we want to |
|
2736 // sample _owner frequently in the early phase of the spin, |
|
2737 // but then back-off and sample less frequently as the spin |
|
2738 // progresses. The back-off makes a good citizen on SMP big |
|
2739 // SMP systems. Oversampling _owner can consume excessive |
|
2740 // coherency bandwidth. Relatedly, if we _oversample _owner we |
|
2741 // can inadvertently interfere with the the ST m->owner=null. |
|
2742 // executed by the lock owner. |
|
2743 if (ctr & msk) continue ; |
|
2744 ++hits ; |
|
2745 if ((hits & 0xF) == 0) { |
|
2746 // The 0xF, above, corresponds to the exponent. |
|
2747 // Consider: (msk+1)|msk |
|
2748 msk = ((msk << 2)|3) & BackOffMask ; |
|
2749 } |
|
2750 |
|
2751 // Probe _owner with TATAS |
|
2752 // If this thread observes the monitor transition or flicker |
|
2753 // from locked to unlocked to locked, then the odds that this |
|
2754 // thread will acquire the lock in this spin attempt go down |
|
2755 // considerably. The same argument applies if the CAS fails |
|
2756 // or if we observe _owner change from one non-null value to |
|
2757 // another non-null value. In such cases we might abort |
|
2758 // the spin without prejudice or apply a "penalty" to the |
|
2759 // spin count-down variable "ctr", reducing it by 100, say. |
|
2760 |
|
2761 Thread * ox = (Thread *) _owner ; |
|
2762 if (ox == NULL) { |
|
2763 ox = (Thread *) Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; |
|
2764 if (ox == NULL) { |
|
2765 // The CAS succeeded -- this thread acquired ownership |
|
2766 // Take care of some bookkeeping to exit spin state. |
|
2767 if (sss && _succ == Self) { |
|
2768 _succ = NULL ; |
|
2769 } |
|
2770 if (MaxSpin > 0) Adjust (&_Spinner, -1) ; |
|
2771 |
|
2772 // Increase _SpinDuration : |
|
2773 // The spin was successful (profitable) so we tend toward |
|
2774 // longer spin attempts in the future. |
|
2775 // CONSIDER: factor "ctr" into the _SpinDuration adjustment. |
|
2776 // If we acquired the lock early in the spin cycle it |
|
2777 // makes sense to increase _SpinDuration proportionally. |
|
2778 // Note that we don't clamp SpinDuration precisely at SpinLimit. |
|
2779 int x = _SpinDuration ; |
|
2780 if (x < Knob_SpinLimit) { |
|
2781 if (x < Knob_Poverty) x = Knob_Poverty ; |
|
2782 _SpinDuration = x + Knob_Bonus ; |
|
2783 } |
|
2784 return 1 ; |
|
2785 } |
|
2786 |
|
2787 // The CAS failed ... we can take any of the following actions: |
|
2788 // * penalize: ctr -= Knob_CASPenalty |
|
2789 // * exit spin with prejudice -- goto Abort; |
|
2790 // * exit spin without prejudice. |
|
2791 // * Since CAS is high-latency, retry again immediately. |
|
2792 prv = ox ; |
|
2793 TEVENT (Spin: cas failed) ; |
|
2794 if (caspty == -2) break ; |
|
2795 if (caspty == -1) goto Abort ; |
|
2796 ctr -= caspty ; |
|
2797 continue ; |
|
2798 } |
|
2799 |
|
2800 // Did lock ownership change hands ? |
|
2801 if (ox != prv && prv != NULL ) { |
|
2802 TEVENT (spin: Owner changed) |
|
2803 if (oxpty == -2) break ; |
|
2804 if (oxpty == -1) goto Abort ; |
|
2805 ctr -= oxpty ; |
|
2806 } |
|
2807 prv = ox ; |
|
2808 |
|
2809 // Abort the spin if the owner is not executing. |
|
2810 // The owner must be executing in order to drop the lock. |
|
2811 // Spinning while the owner is OFFPROC is idiocy. |
|
2812 // Consider: ctr -= RunnablePenalty ; |
|
2813 if (Knob_OState && NotRunnable (Self, ox)) { |
|
2814 TEVENT (Spin abort - notrunnable); |
|
2815 goto Abort ; |
|
2816 } |
|
2817 if (sss && _succ == NULL ) _succ = Self ; |
|
2818 } |
|
2819 |
|
2820 // Spin failed with prejudice -- reduce _SpinDuration. |
|
2821 // TODO: Use an AIMD-like policy to adjust _SpinDuration. |
|
2822 // AIMD is globally stable. |
|
2823 TEVENT (Spin failure) ; |
|
2824 { |
|
2825 int x = _SpinDuration ; |
|
2826 if (x > 0) { |
|
2827 // Consider an AIMD scheme like: x -= (x >> 3) + 100 |
|
2828 // This is globally sample and tends to damp the response. |
|
2829 x -= Knob_Penalty ; |
|
2830 if (x < 0) x = 0 ; |
|
2831 _SpinDuration = x ; |
|
2832 } |
|
2833 } |
|
2834 |
|
2835 Abort: |
|
2836 if (MaxSpin >= 0) Adjust (&_Spinner, -1) ; |
|
2837 if (sss && _succ == Self) { |
|
2838 _succ = NULL ; |
|
2839 // Invariant: after setting succ=null a contending thread |
|
2840 // must recheck-retry _owner before parking. This usually happens |
|
2841 // in the normal usage of TrySpin(), but it's safest |
|
2842 // to make TrySpin() as foolproof as possible. |
|
2843 OrderAccess::fence() ; |
|
2844 if (TryLock(Self) > 0) return 1 ; |
|
2845 } |
|
2846 return 0 ; |
|
2847 } |
|
2848 |
|
2849 #define TrySpin TrySpin_VaryDuration |
|
2850 |
|
2851 static void DeferredInitialize () { |
|
2852 if (InitDone > 0) return ; |
|
2853 if (Atomic::cmpxchg (-1, &InitDone, 0) != 0) { |
|
2854 while (InitDone != 1) ; |
|
2855 return ; |
|
2856 } |
|
2857 |
|
2858 // One-shot global initialization ... |
|
2859 // The initialization is idempotent, so we don't need locks. |
|
2860 // In the future consider doing this via os::init_2(). |
|
2861 // SyncKnobs consist of <Key>=<Value> pairs in the style |
|
2862 // of environment variables. Start by converting ':' to NUL. |
|
2863 |
|
2864 if (SyncKnobs == NULL) SyncKnobs = "" ; |
|
2865 |
|
2866 size_t sz = strlen (SyncKnobs) ; |
|
2867 char * knobs = (char *) malloc (sz + 2) ; |
|
2868 if (knobs == NULL) { |
|
2869 vm_exit_out_of_memory (sz + 2, "Parse SyncKnobs") ; |
|
2870 guarantee (0, "invariant") ; |
|
2871 } |
|
2872 strcpy (knobs, SyncKnobs) ; |
|
2873 knobs[sz+1] = 0 ; |
|
2874 for (char * p = knobs ; *p ; p++) { |
|
2875 if (*p == ':') *p = 0 ; |
|
2876 } |
|
2877 |
|
2878 #define SETKNOB(x) { Knob_##x = kvGetInt (knobs, #x, Knob_##x); } |
|
2879 SETKNOB(ReportSettings) ; |
|
2880 SETKNOB(Verbose) ; |
|
2881 SETKNOB(FixedSpin) ; |
|
2882 SETKNOB(SpinLimit) ; |
|
2883 SETKNOB(SpinBase) ; |
|
2884 SETKNOB(SpinBackOff); |
|
2885 SETKNOB(CASPenalty) ; |
|
2886 SETKNOB(OXPenalty) ; |
|
2887 SETKNOB(LogSpins) ; |
|
2888 SETKNOB(SpinSetSucc) ; |
|
2889 SETKNOB(SuccEnabled) ; |
|
2890 SETKNOB(SuccRestrict) ; |
|
2891 SETKNOB(Penalty) ; |
|
2892 SETKNOB(Bonus) ; |
|
2893 SETKNOB(BonusB) ; |
|
2894 SETKNOB(Poverty) ; |
|
2895 SETKNOB(SpinAfterFutile) ; |
|
2896 SETKNOB(UsePause) ; |
|
2897 SETKNOB(SpinEarly) ; |
|
2898 SETKNOB(OState) ; |
|
2899 SETKNOB(MaxSpinners) ; |
|
2900 SETKNOB(PreSpin) ; |
|
2901 SETKNOB(ExitPolicy) ; |
|
2902 SETKNOB(QMode); |
|
2903 SETKNOB(ResetEvent) ; |
|
2904 SETKNOB(MoveNotifyee) ; |
|
2905 SETKNOB(FastHSSEC) ; |
|
2906 #undef SETKNOB |
|
2907 |
|
2908 if (os::is_MP()) { |
|
2909 BackOffMask = (1 << Knob_SpinBackOff) - 1 ; |
|
2910 if (Knob_ReportSettings) ::printf ("BackOffMask=%X\n", BackOffMask) ; |
|
2911 // CONSIDER: BackOffMask = ROUNDUP_NEXT_POWER2 (ncpus-1) |
|
2912 } else { |
|
2913 Knob_SpinLimit = 0 ; |
|
2914 Knob_SpinBase = 0 ; |
|
2915 Knob_PreSpin = 0 ; |
|
2916 Knob_FixedSpin = -1 ; |
|
2917 } |
|
2918 |
|
2919 if (Knob_LogSpins == 0) { |
|
2920 ObjectSynchronizer::_sync_FailedSpins = NULL ; |
|
2921 } |
|
2922 |
|
2923 free (knobs) ; |
|
2924 OrderAccess::fence() ; |
|
2925 InitDone = 1 ; |
|
2926 } |
|
2927 |
|
2928 // Theory of operations -- Monitors lists, thread residency, etc: |
|
2929 // |
|
2930 // * A thread acquires ownership of a monitor by successfully |
|
2931 // CAS()ing the _owner field from null to non-null. |
|
2932 // |
|
2933 // * Invariant: A thread appears on at most one monitor list -- |
|
2934 // cxq, EntryList or WaitSet -- at any one time. |
|
2935 // |
|
2936 // * Contending threads "push" themselves onto the cxq with CAS |
|
2937 // and then spin/park. |
|
2938 // |
|
2939 // * After a contending thread eventually acquires the lock it must |
|
2940 // dequeue itself from either the EntryList or the cxq. |
|
2941 // |
|
2942 // * The exiting thread identifies and unparks an "heir presumptive" |
|
2943 // tentative successor thread on the EntryList. Critically, the |
|
2944 // exiting thread doesn't unlink the successor thread from the EntryList. |
|
2945 // After having been unparked, the wakee will recontend for ownership of |
|
2946 // the monitor. The successor (wakee) will either acquire the lock or |
|
2947 // re-park itself. |
|
2948 // |
|
2949 // Succession is provided for by a policy of competitive handoff. |
|
2950 // The exiting thread does _not_ grant or pass ownership to the |
|
2951 // successor thread. (This is also referred to as "handoff" succession"). |
|
2952 // Instead the exiting thread releases ownership and possibly wakes |
|
2953 // a successor, so the successor can (re)compete for ownership of the lock. |
|
2954 // If the EntryList is empty but the cxq is populated the exiting |
|
2955 // thread will drain the cxq into the EntryList. It does so by |
|
2956 // by detaching the cxq (installing null with CAS) and folding |
|
2957 // the threads from the cxq into the EntryList. The EntryList is |
|
2958 // doubly linked, while the cxq is singly linked because of the |
|
2959 // CAS-based "push" used to enqueue recently arrived threads (RATs). |
|
2960 // |
|
2961 // * Concurrency invariants: |
|
2962 // |
|
2963 // -- only the monitor owner may access or mutate the EntryList. |
|
2964 // The mutex property of the monitor itself protects the EntryList |
|
2965 // from concurrent interference. |
|
2966 // -- Only the monitor owner may detach the cxq. |
|
2967 // |
|
2968 // * The monitor entry list operations avoid locks, but strictly speaking |
|
2969 // they're not lock-free. Enter is lock-free, exit is not. |
|
2970 // See http://j2se.east/~dice/PERSIST/040825-LockFreeQueues.html |
|
2971 // |
|
2972 // * The cxq can have multiple concurrent "pushers" but only one concurrent |
|
2973 // detaching thread. This mechanism is immune from the ABA corruption. |
|
2974 // More precisely, the CAS-based "push" onto cxq is ABA-oblivious. |
|
2975 // |
|
2976 // * Taken together, the cxq and the EntryList constitute or form a |
|
2977 // single logical queue of threads stalled trying to acquire the lock. |
|
2978 // We use two distinct lists to improve the odds of a constant-time |
|
2979 // dequeue operation after acquisition (in the ::enter() epilog) and |
|
2980 // to reduce heat on the list ends. (c.f. Michael Scott's "2Q" algorithm). |
|
2981 // A key desideratum is to minimize queue & monitor metadata manipulation |
|
2982 // that occurs while holding the monitor lock -- that is, we want to |
|
2983 // minimize monitor lock holds times. Note that even a small amount of |
|
2984 // fixed spinning will greatly reduce the # of enqueue-dequeue operations |
|
2985 // on EntryList|cxq. That is, spinning relieves contention on the "inner" |
|
2986 // locks and monitor metadata. |
|
2987 // |
|
2988 // Cxq points to the the set of Recently Arrived Threads attempting entry. |
|
2989 // Because we push threads onto _cxq with CAS, the RATs must take the form of |
|
2990 // a singly-linked LIFO. We drain _cxq into EntryList at unlock-time when |
|
2991 // the unlocking thread notices that EntryList is null but _cxq is != null. |
|
2992 // |
|
2993 // The EntryList is ordered by the prevailing queue discipline and |
|
2994 // can be organized in any convenient fashion, such as a doubly-linked list or |
|
2995 // a circular doubly-linked list. Critically, we want insert and delete operations |
|
2996 // to operate in constant-time. If we need a priority queue then something akin |
|
2997 // to Solaris' sleepq would work nicely. Viz., |
|
2998 // http://agg.eng/ws/on10_nightly/source/usr/src/uts/common/os/sleepq.c. |
|
2999 // Queue discipline is enforced at ::exit() time, when the unlocking thread |
|
3000 // drains the cxq into the EntryList, and orders or reorders the threads on the |
|
3001 // EntryList accordingly. |
|
3002 // |
|
3003 // Barring "lock barging", this mechanism provides fair cyclic ordering, |
|
3004 // somewhat similar to an elevator-scan. |
|
3005 // |
|
3006 // * The monitor synchronization subsystem avoids the use of native |
|
3007 // synchronization primitives except for the narrow platform-specific |
|
3008 // park-unpark abstraction. See the comments in os_solaris.cpp regarding |
|
3009 // the semantics of park-unpark. Put another way, this monitor implementation |
|
3010 // depends only on atomic operations and park-unpark. The monitor subsystem |
|
3011 // manages all RUNNING->BLOCKED and BLOCKED->READY transitions while the |
|
3012 // underlying OS manages the READY<->RUN transitions. |
|
3013 // |
|
3014 // * Waiting threads reside on the WaitSet list -- wait() puts |
|
3015 // the caller onto the WaitSet. |
|
3016 // |
|
3017 // * notify() or notifyAll() simply transfers threads from the WaitSet to |
|
3018 // either the EntryList or cxq. Subsequent exit() operations will |
|
3019 // unpark the notifyee. Unparking a notifee in notify() is inefficient - |
|
3020 // it's likely the notifyee would simply impale itself on the lock held |
|
3021 // by the notifier. |
|
3022 // |
|
3023 // * An interesting alternative is to encode cxq as (List,LockByte) where |
|
3024 // the LockByte is 0 iff the monitor is owned. _owner is simply an auxiliary |
|
3025 // variable, like _recursions, in the scheme. The threads or Events that form |
|
3026 // the list would have to be aligned in 256-byte addresses. A thread would |
|
3027 // try to acquire the lock or enqueue itself with CAS, but exiting threads |
|
3028 // could use a 1-0 protocol and simply STB to set the LockByte to 0. |
|
3029 // Note that is is *not* word-tearing, but it does presume that full-word |
|
3030 // CAS operations are coherent with intermix with STB operations. That's true |
|
3031 // on most common processors. |
|
3032 // |
|
3033 // * See also http://blogs.sun.com/dave |
|
3034 |
|
3035 |
|
3036 void ATTR ObjectMonitor::EnterI (TRAPS) { |
|
3037 Thread * Self = THREAD ; |
|
3038 assert (Self->is_Java_thread(), "invariant") ; |
|
3039 assert (((JavaThread *) Self)->thread_state() == _thread_blocked , "invariant") ; |
|
3040 |
|
3041 // Try the lock - TATAS |
|
3042 if (TryLock (Self) > 0) { |
|
3043 assert (_succ != Self , "invariant") ; |
|
3044 assert (_owner == Self , "invariant") ; |
|
3045 assert (_Responsible != Self , "invariant") ; |
|
3046 return ; |
|
3047 } |
|
3048 |
|
3049 DeferredInitialize () ; |
|
3050 |
|
3051 // We try one round of spinning *before* enqueueing Self. |
|
3052 // |
|
3053 // If the _owner is ready but OFFPROC we could use a YieldTo() |
|
3054 // operation to donate the remainder of this thread's quantum |
|
3055 // to the owner. This has subtle but beneficial affinity |
|
3056 // effects. |
|
3057 |
|
3058 if (TrySpin (Self) > 0) { |
|
3059 assert (_owner == Self , "invariant") ; |
|
3060 assert (_succ != Self , "invariant") ; |
|
3061 assert (_Responsible != Self , "invariant") ; |
|
3062 return ; |
|
3063 } |
|
3064 |
|
3065 // The Spin failed -- Enqueue and park the thread ... |
|
3066 assert (_succ != Self , "invariant") ; |
|
3067 assert (_owner != Self , "invariant") ; |
|
3068 assert (_Responsible != Self , "invariant") ; |
|
3069 |
|
3070 // Enqueue "Self" on ObjectMonitor's _cxq. |
|
3071 // |
|
3072 // Node acts as a proxy for Self. |
|
3073 // As an aside, if were to ever rewrite the synchronization code mostly |
|
3074 // in Java, WaitNodes, ObjectMonitors, and Events would become 1st-class |
|
3075 // Java objects. This would avoid awkward lifecycle and liveness issues, |
|
3076 // as well as eliminate a subset of ABA issues. |
|
3077 // TODO: eliminate ObjectWaiter and enqueue either Threads or Events. |
|
3078 // |
|
3079 |
|
3080 ObjectWaiter node(Self) ; |
|
3081 Self->_ParkEvent->reset() ; |
|
3082 node._prev = (ObjectWaiter *) 0xBAD ; |
|
3083 node.TState = ObjectWaiter::TS_CXQ ; |
|
3084 |
|
3085 // Push "Self" onto the front of the _cxq. |
|
3086 // Once on cxq/EntryList, Self stays on-queue until it acquires the lock. |
|
3087 // Note that spinning tends to reduce the rate at which threads |
|
3088 // enqueue and dequeue on EntryList|cxq. |
|
3089 ObjectWaiter * nxt ; |
|
3090 for (;;) { |
|
3091 node._next = nxt = _cxq ; |
|
3092 if (Atomic::cmpxchg_ptr (&node, &_cxq, nxt) == nxt) break ; |
|
3093 |
|
3094 // Interference - the CAS failed because _cxq changed. Just retry. |
|
3095 // As an optional optimization we retry the lock. |
|
3096 if (TryLock (Self) > 0) { |
|
3097 assert (_succ != Self , "invariant") ; |
|
3098 assert (_owner == Self , "invariant") ; |
|
3099 assert (_Responsible != Self , "invariant") ; |
|
3100 return ; |
|
3101 } |
|
3102 } |
|
3103 |
|
3104 // Check for cxq|EntryList edge transition to non-null. This indicates |
|
3105 // the onset of contention. While contention persists exiting threads |
|
3106 // will use a ST:MEMBAR:LD 1-1 exit protocol. When contention abates exit |
|
3107 // operations revert to the faster 1-0 mode. This enter operation may interleave |
|
3108 // (race) a concurrent 1-0 exit operation, resulting in stranding, so we |
|
3109 // arrange for one of the contending thread to use a timed park() operations |
|
3110 // to detect and recover from the race. (Stranding is form of progress failure |
|
3111 // where the monitor is unlocked but all the contending threads remain parked). |
|
3112 // That is, at least one of the contended threads will periodically poll _owner. |
|
3113 // One of the contending threads will become the designated "Responsible" thread. |
|
3114 // The Responsible thread uses a timed park instead of a normal indefinite park |
|
3115 // operation -- it periodically wakes and checks for and recovers from potential |
|
3116 // strandings admitted by 1-0 exit operations. We need at most one Responsible |
|
3117 // thread per-monitor at any given moment. Only threads on cxq|EntryList may |
|
3118 // be responsible for a monitor. |
|
3119 // |
|
3120 // Currently, one of the contended threads takes on the added role of "Responsible". |
|
3121 // A viable alternative would be to use a dedicated "stranding checker" thread |
|
3122 // that periodically iterated over all the threads (or active monitors) and unparked |
|
3123 // successors where there was risk of stranding. This would help eliminate the |
|
3124 // timer scalability issues we see on some platforms as we'd only have one thread |
|
3125 // -- the checker -- parked on a timer. |
|
3126 |
|
3127 if ((SyncFlags & 16) == 0 && nxt == NULL && _EntryList == NULL) { |
|
3128 // Try to assume the role of responsible thread for the monitor. |
|
3129 // CONSIDER: ST vs CAS vs { if (Responsible==null) Responsible=Self } |
|
3130 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; |
|
3131 } |
|
3132 |
|
3133 // The lock have been released while this thread was occupied queueing |
|
3134 // itself onto _cxq. To close the race and avoid "stranding" and |
|
3135 // progress-liveness failure we must resample-retry _owner before parking. |
|
3136 // Note the Dekker/Lamport duality: ST cxq; MEMBAR; LD Owner. |
|
3137 // In this case the ST-MEMBAR is accomplished with CAS(). |
|
3138 // |
|
3139 // TODO: Defer all thread state transitions until park-time. |
|
3140 // Since state transitions are heavy and inefficient we'd like |
|
3141 // to defer the state transitions until absolutely necessary, |
|
3142 // and in doing so avoid some transitions ... |
|
3143 |
|
3144 TEVENT (Inflated enter - Contention) ; |
|
3145 int nWakeups = 0 ; |
|
3146 int RecheckInterval = 1 ; |
|
3147 |
|
3148 for (;;) { |
|
3149 |
|
3150 if (TryLock (Self) > 0) break ; |
|
3151 assert (_owner != Self, "invariant") ; |
|
3152 |
|
3153 if ((SyncFlags & 2) && _Responsible == NULL) { |
|
3154 Atomic::cmpxchg_ptr (Self, &_Responsible, NULL) ; |
|
3155 } |
|
3156 |
|
3157 // park self |
|
3158 if (_Responsible == Self || (SyncFlags & 1)) { |
|
3159 TEVENT (Inflated enter - park TIMED) ; |
|
3160 Self->_ParkEvent->park ((jlong) RecheckInterval) ; |
|
3161 // Increase the RecheckInterval, but clamp the value. |
|
3162 RecheckInterval *= 8 ; |
|
3163 if (RecheckInterval > 1000) RecheckInterval = 1000 ; |
|
3164 } else { |
|
3165 TEVENT (Inflated enter - park UNTIMED) ; |
|
3166 Self->_ParkEvent->park() ; |
|
3167 } |
|
3168 |
|
3169 if (TryLock(Self) > 0) break ; |
|
3170 |
|
3171 // The lock is still contested. |
|
3172 // Keep a tally of the # of futile wakeups. |
|
3173 // Note that the counter is not protected by a lock or updated by atomics. |
|
3174 // That is by design - we trade "lossy" counters which are exposed to |
|
3175 // races during updates for a lower probe effect. |
|
3176 TEVENT (Inflated enter - Futile wakeup) ; |
|
3177 if (ObjectSynchronizer::_sync_FutileWakeups != NULL) { |
|
3178 ObjectSynchronizer::_sync_FutileWakeups->inc() ; |
|
3179 } |
|
3180 ++ nWakeups ; |
|
3181 |
|
3182 // Assuming this is not a spurious wakeup we'll normally find _succ == Self. |
|
3183 // We can defer clearing _succ until after the spin completes |
|
3184 // TrySpin() must tolerate being called with _succ == Self. |
|
3185 // Try yet another round of adaptive spinning. |
|
3186 if ((Knob_SpinAfterFutile & 1) && TrySpin (Self) > 0) break ; |
|
3187 |
|
3188 // We can find that we were unpark()ed and redesignated _succ while |
|
3189 // we were spinning. That's harmless. If we iterate and call park(), |
|
3190 // park() will consume the event and return immediately and we'll |
|
3191 // just spin again. This pattern can repeat, leaving _succ to simply |
|
3192 // spin on a CPU. Enable Knob_ResetEvent to clear pending unparks(). |
|
3193 // Alternately, we can sample fired() here, and if set, forgo spinning |
|
3194 // in the next iteration. |
|
3195 |
|
3196 if ((Knob_ResetEvent & 1) && Self->_ParkEvent->fired()) { |
|
3197 Self->_ParkEvent->reset() ; |
|
3198 OrderAccess::fence() ; |
|
3199 } |
|
3200 if (_succ == Self) _succ = NULL ; |
|
3201 |
|
3202 // Invariant: after clearing _succ a thread *must* retry _owner before parking. |
|
3203 OrderAccess::fence() ; |
|
3204 } |
|
3205 |
|
3206 // Egress : |
|
3207 // Self has acquired the lock -- Unlink Self from the cxq or EntryList. |
|
3208 // Normally we'll find Self on the EntryList . |
|
3209 // From the perspective of the lock owner (this thread), the |
|
3210 // EntryList is stable and cxq is prepend-only. |
|
3211 // The head of cxq is volatile but the interior is stable. |
|
3212 // In addition, Self.TState is stable. |
|
3213 |
|
3214 assert (_owner == Self , "invariant") ; |
|
3215 assert (object() != NULL , "invariant") ; |
|
3216 // I'd like to write: |
|
3217 // guarantee (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; |
|
3218 // but as we're at a safepoint that's not safe. |
|
3219 |
|
3220 UnlinkAfterAcquire (Self, &node) ; |
|
3221 if (_succ == Self) _succ = NULL ; |
|
3222 |
|
3223 assert (_succ != Self, "invariant") ; |
|
3224 if (_Responsible == Self) { |
|
3225 _Responsible = NULL ; |
|
3226 // Dekker pivot-point. |
|
3227 // Consider OrderAccess::storeload() here |
|
3228 |
|
3229 // We may leave threads on cxq|EntryList without a designated |
|
3230 // "Responsible" thread. This is benign. When this thread subsequently |
|
3231 // exits the monitor it can "see" such preexisting "old" threads -- |
|
3232 // threads that arrived on the cxq|EntryList before the fence, above -- |
|
3233 // by LDing cxq|EntryList. Newly arrived threads -- that is, threads |
|
3234 // that arrive on cxq after the ST:MEMBAR, above -- will set Responsible |
|
3235 // non-null and elect a new "Responsible" timer thread. |
|
3236 // |
|
3237 // This thread executes: |
|
3238 // ST Responsible=null; MEMBAR (in enter epilog - here) |
|
3239 // LD cxq|EntryList (in subsequent exit) |
|
3240 // |
|
3241 // Entering threads in the slow/contended path execute: |
|
3242 // ST cxq=nonnull; MEMBAR; LD Responsible (in enter prolog) |
|
3243 // The (ST cxq; MEMBAR) is accomplished with CAS(). |
|
3244 // |
|
3245 // The MEMBAR, above, prevents the LD of cxq|EntryList in the subsequent |
|
3246 // exit operation from floating above the ST Responsible=null. |
|
3247 // |
|
3248 // In *practice* however, EnterI() is always followed by some atomic |
|
3249 // operation such as the decrement of _count in ::enter(). Those atomics |
|
3250 // obviate the need for the explicit MEMBAR, above. |
|
3251 } |
|
3252 |
|
3253 // We've acquired ownership with CAS(). |
|
3254 // CAS is serializing -- it has MEMBAR/FENCE-equivalent semantics. |
|
3255 // But since the CAS() this thread may have also stored into _succ, |
|
3256 // EntryList, cxq or Responsible. These meta-data updates must be |
|
3257 // visible __before this thread subsequently drops the lock. |
|
3258 // Consider what could occur if we didn't enforce this constraint -- |
|
3259 // STs to monitor meta-data and user-data could reorder with (become |
|
3260 // visible after) the ST in exit that drops ownership of the lock. |
|
3261 // Some other thread could then acquire the lock, but observe inconsistent |
|
3262 // or old monitor meta-data and heap data. That violates the JMM. |
|
3263 // To that end, the 1-0 exit() operation must have at least STST|LDST |
|
3264 // "release" barrier semantics. Specifically, there must be at least a |
|
3265 // STST|LDST barrier in exit() before the ST of null into _owner that drops |
|
3266 // the lock. The barrier ensures that changes to monitor meta-data and data |
|
3267 // protected by the lock will be visible before we release the lock, and |
|
3268 // therefore before some other thread (CPU) has a chance to acquire the lock. |
|
3269 // See also: http://gee.cs.oswego.edu/dl/jmm/cookbook.html. |
|
3270 // |
|
3271 // Critically, any prior STs to _succ or EntryList must be visible before |
|
3272 // the ST of null into _owner in the *subsequent* (following) corresponding |
|
3273 // monitorexit. Recall too, that in 1-0 mode monitorexit does not necessarily |
|
3274 // execute a serializing instruction. |
|
3275 |
|
3276 if (SyncFlags & 8) { |
|
3277 OrderAccess::fence() ; |
|
3278 } |
|
3279 return ; |
|
3280 } |
|
3281 |
|
3282 // ExitSuspendEquivalent: |
|
3283 // A faster alternate to handle_special_suspend_equivalent_condition() |
|
3284 // |
|
3285 // handle_special_suspend_equivalent_condition() unconditionally |
|
3286 // acquires the SR_lock. On some platforms uncontended MutexLocker() |
|
3287 // operations have high latency. Note that in ::enter() we call HSSEC |
|
3288 // while holding the monitor, so we effectively lengthen the critical sections. |
|
3289 // |
|
3290 // There are a number of possible solutions: |
|
3291 // |
|
3292 // A. To ameliorate the problem we might also defer state transitions |
|
3293 // to as late as possible -- just prior to parking. |
|
3294 // Given that, we'd call HSSEC after having returned from park(), |
|
3295 // but before attempting to acquire the monitor. This is only a |
|
3296 // partial solution. It avoids calling HSSEC while holding the |
|
3297 // monitor (good), but it still increases successor reacquisition latency -- |
|
3298 // the interval between unparking a successor and the time the successor |
|
3299 // resumes and retries the lock. See ReenterI(), which defers state transitions. |
|
3300 // If we use this technique we can also avoid EnterI()-exit() loop |
|
3301 // in ::enter() where we iteratively drop the lock and then attempt |
|
3302 // to reacquire it after suspending. |
|
3303 // |
|
3304 // B. In the future we might fold all the suspend bits into a |
|
3305 // composite per-thread suspend flag and then update it with CAS(). |
|
3306 // Alternately, a Dekker-like mechanism with multiple variables |
|
3307 // would suffice: |
|
3308 // ST Self->_suspend_equivalent = false |
|
3309 // MEMBAR |
|
3310 // LD Self_>_suspend_flags |
|
3311 // |
|
3312 |
|
3313 |
|
3314 bool ObjectMonitor::ExitSuspendEquivalent (JavaThread * jSelf) { |
|
3315 int Mode = Knob_FastHSSEC ; |
|
3316 if (Mode && !jSelf->is_external_suspend()) { |
|
3317 assert (jSelf->is_suspend_equivalent(), "invariant") ; |
|
3318 jSelf->clear_suspend_equivalent() ; |
|
3319 if (2 == Mode) OrderAccess::storeload() ; |
|
3320 if (!jSelf->is_external_suspend()) return false ; |
|
3321 // We raced a suspension -- fall thru into the slow path |
|
3322 TEVENT (ExitSuspendEquivalent - raced) ; |
|
3323 jSelf->set_suspend_equivalent() ; |
|
3324 } |
|
3325 return jSelf->handle_special_suspend_equivalent_condition() ; |
|
3326 } |
|
3327 |
|
3328 |
|
3329 // ReenterI() is a specialized inline form of the latter half of the |
|
3330 // contended slow-path from EnterI(). We use ReenterI() only for |
|
3331 // monitor reentry in wait(). |
|
3332 // |
|
3333 // In the future we should reconcile EnterI() and ReenterI(), adding |
|
3334 // Knob_Reset and Knob_SpinAfterFutile support and restructuring the |
|
3335 // loop accordingly. |
|
3336 |
|
3337 void ATTR ObjectMonitor::ReenterI (Thread * Self, ObjectWaiter * SelfNode) { |
|
3338 assert (Self != NULL , "invariant") ; |
|
3339 assert (SelfNode != NULL , "invariant") ; |
|
3340 assert (SelfNode->_thread == Self , "invariant") ; |
|
3341 assert (_waiters > 0 , "invariant") ; |
|
3342 assert (((oop)(object()))->mark() == markOopDesc::encode(this) , "invariant") ; |
|
3343 assert (((JavaThread *)Self)->thread_state() != _thread_blocked, "invariant") ; |
|
3344 JavaThread * jt = (JavaThread *) Self ; |
|
3345 |
|
3346 int nWakeups = 0 ; |
|
3347 for (;;) { |
|
3348 ObjectWaiter::TStates v = SelfNode->TState ; |
|
3349 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; |
|
3350 assert (_owner != Self, "invariant") ; |
|
3351 |
|
3352 if (TryLock (Self) > 0) break ; |
|
3353 if (TrySpin (Self) > 0) break ; |
|
3354 |
|
3355 TEVENT (Wait Reentry - parking) ; |
|
3356 |
|
3357 // State transition wrappers around park() ... |
|
3358 // ReenterI() wisely defers state transitions until |
|
3359 // it's clear we must park the thread. |
|
3360 { |
|
3361 OSThreadContendState osts(Self->osthread()); |
|
3362 ThreadBlockInVM tbivm(jt); |
|
3363 |
|
3364 // cleared by handle_special_suspend_equivalent_condition() |
|
3365 // or java_suspend_self() |
|
3366 jt->set_suspend_equivalent(); |
|
3367 if (SyncFlags & 1) { |
|
3368 Self->_ParkEvent->park ((jlong)1000) ; |
|
3369 } else { |
|
3370 Self->_ParkEvent->park () ; |
|
3371 } |
|
3372 |
|
3373 // were we externally suspended while we were waiting? |
|
3374 for (;;) { |
|
3375 if (!ExitSuspendEquivalent (jt)) break ; |
|
3376 if (_succ == Self) { _succ = NULL; OrderAccess::fence(); } |
|
3377 jt->java_suspend_self(); |
|
3378 jt->set_suspend_equivalent(); |
|
3379 } |
|
3380 } |
|
3381 |
|
3382 // Try again, but just so we distinguish between futile wakeups and |
|
3383 // successful wakeups. The following test isn't algorithmically |
|
3384 // necessary, but it helps us maintain sensible statistics. |
|
3385 if (TryLock(Self) > 0) break ; |
|
3386 |
|
3387 // The lock is still contested. |
|
3388 // Keep a tally of the # of futile wakeups. |
|
3389 // Note that the counter is not protected by a lock or updated by atomics. |
|
3390 // That is by design - we trade "lossy" counters which are exposed to |
|
3391 // races during updates for a lower probe effect. |
|
3392 TEVENT (Wait Reentry - futile wakeup) ; |
|
3393 ++ nWakeups ; |
|
3394 |
|
3395 // Assuming this is not a spurious wakeup we'll normally |
|
3396 // find that _succ == Self. |
|
3397 if (_succ == Self) _succ = NULL ; |
|
3398 |
|
3399 // Invariant: after clearing _succ a contending thread |
|
3400 // *must* retry _owner before parking. |
|
3401 OrderAccess::fence() ; |
|
3402 |
|
3403 if (ObjectSynchronizer::_sync_FutileWakeups != NULL) { |
|
3404 ObjectSynchronizer::_sync_FutileWakeups->inc() ; |
|
3405 } |
|
3406 } |
|
3407 |
|
3408 // Self has acquired the lock -- Unlink Self from the cxq or EntryList . |
|
3409 // Normally we'll find Self on the EntryList. |
|
3410 // Unlinking from the EntryList is constant-time and atomic-free. |
|
3411 // From the perspective of the lock owner (this thread), the |
|
3412 // EntryList is stable and cxq is prepend-only. |
|
3413 // The head of cxq is volatile but the interior is stable. |
|
3414 // In addition, Self.TState is stable. |
|
3415 |
|
3416 assert (_owner == Self, "invariant") ; |
|
3417 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; |
|
3418 UnlinkAfterAcquire (Self, SelfNode) ; |
|
3419 if (_succ == Self) _succ = NULL ; |
|
3420 assert (_succ != Self, "invariant") ; |
|
3421 SelfNode->TState = ObjectWaiter::TS_RUN ; |
|
3422 OrderAccess::fence() ; // see comments at the end of EnterI() |
|
3423 } |
|
3424 |
|
3425 bool ObjectMonitor::try_enter(Thread* THREAD) { |
|
3426 if (THREAD != _owner) { |
|
3427 if (THREAD->is_lock_owned ((address)_owner)) { |
|
3428 assert(_recursions == 0, "internal state error"); |
|
3429 _owner = THREAD ; |
|
3430 _recursions = 1 ; |
|
3431 OwnerIsThread = 1 ; |
|
3432 return true; |
|
3433 } |
|
3434 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { |
|
3435 return false; |
|
3436 } |
|
3437 return true; |
|
3438 } else { |
|
3439 _recursions++; |
|
3440 return true; |
|
3441 } |
|
3442 } |
|
3443 |
|
3444 void ATTR ObjectMonitor::enter(TRAPS) { |
|
3445 // The following code is ordered to check the most common cases first |
|
3446 // and to reduce RTS->RTO cache line upgrades on SPARC and IA32 processors. |
|
3447 Thread * const Self = THREAD ; |
|
3448 void * cur ; |
|
3449 |
|
3450 cur = Atomic::cmpxchg_ptr (Self, &_owner, NULL) ; |
|
3451 if (cur == NULL) { |
|
3452 // Either ASSERT _recursions == 0 or explicitly set _recursions = 0. |
|
3453 assert (_recursions == 0 , "invariant") ; |
|
3454 assert (_owner == Self, "invariant") ; |
|
3455 // CONSIDER: set or assert OwnerIsThread == 1 |
|
3456 return ; |
|
3457 } |
|
3458 |
|
3459 if (cur == Self) { |
|
3460 // TODO-FIXME: check for integer overflow! BUGID 6557169. |
|
3461 _recursions ++ ; |
|
3462 return ; |
|
3463 } |
|
3464 |
|
3465 if (Self->is_lock_owned ((address)cur)) { |
|
3466 assert (_recursions == 0, "internal state error"); |
|
3467 _recursions = 1 ; |
|
3468 // Commute owner from a thread-specific on-stack BasicLockObject address to |
|
3469 // a full-fledged "Thread *". |
|
3470 _owner = Self ; |
|
3471 OwnerIsThread = 1 ; |
|
3472 return ; |
|
3473 } |
|
3474 |
|
3475 // We've encountered genuine contention. |
|
3476 assert (Self->_Stalled == 0, "invariant") ; |
|
3477 Self->_Stalled = intptr_t(this) ; |
|
3478 |
|
3479 // Try one round of spinning *before* enqueueing Self |
|
3480 // and before going through the awkward and expensive state |
|
3481 // transitions. The following spin is strictly optional ... |
|
3482 // Note that if we acquire the monitor from an initial spin |
|
3483 // we forgo posting JVMTI events and firing DTRACE probes. |
|
3484 if (Knob_SpinEarly && TrySpin (Self) > 0) { |
|
3485 assert (_owner == Self , "invariant") ; |
|
3486 assert (_recursions == 0 , "invariant") ; |
|
3487 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; |
|
3488 Self->_Stalled = 0 ; |
|
3489 return ; |
|
3490 } |
|
3491 |
|
3492 assert (_owner != Self , "invariant") ; |
|
3493 assert (_succ != Self , "invariant") ; |
|
3494 assert (Self->is_Java_thread() , "invariant") ; |
|
3495 JavaThread * jt = (JavaThread *) Self ; |
|
3496 assert (!SafepointSynchronize::is_at_safepoint(), "invariant") ; |
|
3497 assert (jt->thread_state() != _thread_blocked , "invariant") ; |
|
3498 assert (this->object() != NULL , "invariant") ; |
|
3499 assert (_count >= 0, "invariant") ; |
|
3500 |
|
3501 // Prevent deflation at STW-time. See deflate_idle_monitors() and is_busy(). |
|
3502 // Ensure the object-monitor relationship remains stable while there's contention. |
|
3503 Atomic::inc_ptr(&_count); |
|
3504 |
|
3505 { // Change java thread status to indicate blocked on monitor enter. |
|
3506 JavaThreadBlockedOnMonitorEnterState jtbmes(jt, this); |
|
3507 |
|
3508 DTRACE_MONITOR_PROBE(contended__enter, this, object(), jt); |
|
3509 if (JvmtiExport::should_post_monitor_contended_enter()) { |
|
3510 JvmtiExport::post_monitor_contended_enter(jt, this); |
|
3511 } |
|
3512 |
|
3513 OSThreadContendState osts(Self->osthread()); |
|
3514 ThreadBlockInVM tbivm(jt); |
|
3515 |
|
3516 Self->set_current_pending_monitor(this); |
|
3517 |
|
3518 // TODO-FIXME: change the following for(;;) loop to straight-line code. |
|
3519 for (;;) { |
|
3520 jt->set_suspend_equivalent(); |
|
3521 // cleared by handle_special_suspend_equivalent_condition() |
|
3522 // or java_suspend_self() |
|
3523 |
|
3524 EnterI (THREAD) ; |
|
3525 |
|
3526 if (!ExitSuspendEquivalent(jt)) break ; |
|
3527 |
|
3528 // |
|
3529 // We have acquired the contended monitor, but while we were |
|
3530 // waiting another thread suspended us. We don't want to enter |
|
3531 // the monitor while suspended because that would surprise the |
|
3532 // thread that suspended us. |
|
3533 // |
|
3534 _recursions = 0 ; |
|
3535 _succ = NULL ; |
|
3536 exit (Self) ; |
|
3537 |
|
3538 jt->java_suspend_self(); |
|
3539 } |
|
3540 Self->set_current_pending_monitor(NULL); |
|
3541 } |
|
3542 |
|
3543 Atomic::dec_ptr(&_count); |
|
3544 assert (_count >= 0, "invariant") ; |
|
3545 Self->_Stalled = 0 ; |
|
3546 |
|
3547 // Must either set _recursions = 0 or ASSERT _recursions == 0. |
|
3548 assert (_recursions == 0 , "invariant") ; |
|
3549 assert (_owner == Self , "invariant") ; |
|
3550 assert (_succ != Self , "invariant") ; |
|
3551 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; |
|
3552 |
|
3553 // The thread -- now the owner -- is back in vm mode. |
|
3554 // Report the glorious news via TI,DTrace and jvmstat. |
|
3555 // The probe effect is non-trivial. All the reportage occurs |
|
3556 // while we hold the monitor, increasing the length of the critical |
|
3557 // section. Amdahl's parallel speedup law comes vividly into play. |
|
3558 // |
|
3559 // Another option might be to aggregate the events (thread local or |
|
3560 // per-monitor aggregation) and defer reporting until a more opportune |
|
3561 // time -- such as next time some thread encounters contention but has |
|
3562 // yet to acquire the lock. While spinning that thread could |
|
3563 // spinning we could increment JVMStat counters, etc. |
|
3564 |
|
3565 DTRACE_MONITOR_PROBE(contended__entered, this, object(), jt); |
|
3566 if (JvmtiExport::should_post_monitor_contended_entered()) { |
|
3567 JvmtiExport::post_monitor_contended_entered(jt, this); |
|
3568 } |
|
3569 if (ObjectSynchronizer::_sync_ContendedLockAttempts != NULL) { |
|
3570 ObjectSynchronizer::_sync_ContendedLockAttempts->inc() ; |
|
3571 } |
|
3572 } |
|
3573 |
|
3574 void ObjectMonitor::ExitEpilog (Thread * Self, ObjectWaiter * Wakee) { |
|
3575 assert (_owner == Self, "invariant") ; |
|
3576 |
|
3577 // Exit protocol: |
|
3578 // 1. ST _succ = wakee |
|
3579 // 2. membar #loadstore|#storestore; |
|
3580 // 2. ST _owner = NULL |
|
3581 // 3. unpark(wakee) |
|
3582 |
|
3583 _succ = Knob_SuccEnabled ? Wakee->_thread : NULL ; |
|
3584 ParkEvent * Trigger = Wakee->_event ; |
|
3585 |
|
3586 // Hygiene -- once we've set _owner = NULL we can't safely dereference Wakee again. |
|
3587 // The thread associated with Wakee may have grabbed the lock and "Wakee" may be |
|
3588 // out-of-scope (non-extant). |
|
3589 Wakee = NULL ; |
|
3590 |
|
3591 // Drop the lock |
|
3592 OrderAccess::release_store_ptr (&_owner, NULL) ; |
|
3593 OrderAccess::fence() ; // ST _owner vs LD in unpark() |
|
3594 |
|
3595 // TODO-FIXME: |
|
3596 // If there's a safepoint pending the best policy would be to |
|
3597 // get _this thread to a safepoint and only wake the successor |
|
3598 // after the safepoint completed. monitorexit uses a "leaf" |
|
3599 // state transition, however, so this thread can't become |
|
3600 // safe at this point in time. (Its stack isn't walkable). |
|
3601 // The next best thing is to defer waking the successor by |
|
3602 // adding to a list of thread to be unparked after at the |
|
3603 // end of the forthcoming STW). |
|
3604 if (SafepointSynchronize::do_call_back()) { |
|
3605 TEVENT (unpark before SAFEPOINT) ; |
|
3606 } |
|
3607 |
|
3608 // Possible optimizations ... |
|
3609 // |
|
3610 // * Consider: set Wakee->UnparkTime = timeNow() |
|
3611 // When the thread wakes up it'll compute (timeNow() - Self->UnparkTime()). |
|
3612 // By measuring recent ONPROC latency we can approximate the |
|
3613 // system load. In turn, we can feed that information back |
|
3614 // into the spinning & succession policies. |
|
3615 // (ONPROC latency correlates strongly with load). |
|
3616 // |
|
3617 // * Pull affinity: |
|
3618 // If the wakee is cold then transiently setting it's affinity |
|
3619 // to the current CPU is a good idea. |
|
3620 // See http://j2se.east/~dice/PERSIST/050624-PullAffinity.txt |
|
3621 DTRACE_MONITOR_PROBE(contended__exit, this, object(), Self); |
|
3622 Trigger->unpark() ; |
|
3623 |
|
3624 // Maintain stats and report events to JVMTI |
|
3625 if (ObjectSynchronizer::_sync_Parks != NULL) { |
|
3626 ObjectSynchronizer::_sync_Parks->inc() ; |
|
3627 } |
|
3628 } |
|
3629 |
|
3630 |
|
3631 // exit() |
|
3632 // ~~~~~~ |
|
3633 // Note that the collector can't reclaim the objectMonitor or deflate |
|
3634 // the object out from underneath the thread calling ::exit() as the |
|
3635 // thread calling ::exit() never transitions to a stable state. |
|
3636 // This inhibits GC, which in turn inhibits asynchronous (and |
|
3637 // inopportune) reclamation of "this". |
|
3638 // |
|
3639 // We'd like to assert that: (THREAD->thread_state() != _thread_blocked) ; |
|
3640 // There's one exception to the claim above, however. EnterI() can call |
|
3641 // exit() to drop a lock if the acquirer has been externally suspended. |
|
3642 // In that case exit() is called with _thread_state as _thread_blocked, |
|
3643 // but the monitor's _count field is > 0, which inhibits reclamation. |
|
3644 // |
|
3645 // 1-0 exit |
|
3646 // ~~~~~~~~ |
|
3647 // ::exit() uses a canonical 1-1 idiom with a MEMBAR although some of |
|
3648 // the fast-path operators have been optimized so the common ::exit() |
|
3649 // operation is 1-0. See i486.ad fast_unlock(), for instance. |
|
3650 // The code emitted by fast_unlock() elides the usual MEMBAR. This |
|
3651 // greatly improves latency -- MEMBAR and CAS having considerable local |
|
3652 // latency on modern processors -- but at the cost of "stranding". Absent the |
|
3653 // MEMBAR, a thread in fast_unlock() can race a thread in the slow |
|
3654 // ::enter() path, resulting in the entering thread being stranding |
|
3655 // and a progress-liveness failure. Stranding is extremely rare. |
|
3656 // We use timers (timed park operations) & periodic polling to detect |
|
3657 // and recover from stranding. Potentially stranded threads periodically |
|
3658 // wake up and poll the lock. See the usage of the _Responsible variable. |
|
3659 // |
|
3660 // The CAS() in enter provides for safety and exclusion, while the CAS or |
|
3661 // MEMBAR in exit provides for progress and avoids stranding. 1-0 locking |
|
3662 // eliminates the CAS/MEMBAR from the exist path, but it admits stranding. |
|
3663 // We detect and recover from stranding with timers. |
|
3664 // |
|
3665 // If a thread transiently strands it'll park until (a) another |
|
3666 // thread acquires the lock and then drops the lock, at which time the |
|
3667 // exiting thread will notice and unpark the stranded thread, or, (b) |
|
3668 // the timer expires. If the lock is high traffic then the stranding latency |
|
3669 // will be low due to (a). If the lock is low traffic then the odds of |
|
3670 // stranding are lower, although the worst-case stranding latency |
|
3671 // is longer. Critically, we don't want to put excessive load in the |
|
3672 // platform's timer subsystem. We want to minimize both the timer injection |
|
3673 // rate (timers created/sec) as well as the number of timers active at |
|
3674 // any one time. (more precisely, we want to minimize timer-seconds, which is |
|
3675 // the integral of the # of active timers at any instant over time). |
|
3676 // Both impinge on OS scalability. Given that, at most one thread parked on |
|
3677 // a monitor will use a timer. |
|
3678 |
|
3679 void ATTR ObjectMonitor::exit(TRAPS) { |
|
3680 Thread * Self = THREAD ; |
|
3681 if (THREAD != _owner) { |
|
3682 if (THREAD->is_lock_owned((address) _owner)) { |
|
3683 // Transmute _owner from a BasicLock pointer to a Thread address. |
|
3684 // We don't need to hold _mutex for this transition. |
|
3685 // Non-null to Non-null is safe as long as all readers can |
|
3686 // tolerate either flavor. |
|
3687 assert (_recursions == 0, "invariant") ; |
|
3688 _owner = THREAD ; |
|
3689 _recursions = 0 ; |
|
3690 OwnerIsThread = 1 ; |
|
3691 } else { |
|
3692 // NOTE: we need to handle unbalanced monitor enter/exit |
|
3693 // in native code by throwing an exception. |
|
3694 // TODO: Throw an IllegalMonitorStateException ? |
|
3695 TEVENT (Exit - Throw IMSX) ; |
|
3696 assert(false, "Non-balanced monitor enter/exit!"); |
|
3697 if (false) { |
|
3698 THROW(vmSymbols::java_lang_IllegalMonitorStateException()); |
|
3699 } |
|
3700 return; |
|
3701 } |
|
3702 } |
|
3703 |
|
3704 if (_recursions != 0) { |
|
3705 _recursions--; // this is simple recursive enter |
|
3706 TEVENT (Inflated exit - recursive) ; |
|
3707 return ; |
|
3708 } |
|
3709 |
|
3710 // Invariant: after setting Responsible=null an thread must execute |
|
3711 // a MEMBAR or other serializing instruction before fetching EntryList|cxq. |
|
3712 if ((SyncFlags & 4) == 0) { |
|
3713 _Responsible = NULL ; |
|
3714 } |
|
3715 |
|
3716 for (;;) { |
|
3717 assert (THREAD == _owner, "invariant") ; |
|
3718 |
|
3719 // Fast-path monitor exit: |
|
3720 // |
|
3721 // Observe the Dekker/Lamport duality: |
|
3722 // A thread in ::exit() executes: |
|
3723 // ST Owner=null; MEMBAR; LD EntryList|cxq. |
|
3724 // A thread in the contended ::enter() path executes the complementary: |
|
3725 // ST EntryList|cxq = nonnull; MEMBAR; LD Owner. |
|
3726 // |
|
3727 // Note that there's a benign race in the exit path. We can drop the |
|
3728 // lock, another thread can reacquire the lock immediately, and we can |
|
3729 // then wake a thread unnecessarily (yet another flavor of futile wakeup). |
|
3730 // This is benign, and we've structured the code so the windows are short |
|
3731 // and the frequency of such futile wakeups is low. |
|
3732 // |
|
3733 // We could eliminate the race by encoding both the "LOCKED" state and |
|
3734 // the queue head in a single word. Exit would then use either CAS to |
|
3735 // clear the LOCKED bit/byte. This precludes the desirable 1-0 optimization, |
|
3736 // however. |
|
3737 // |
|
3738 // Possible fast-path ::exit() optimization: |
|
3739 // The current fast-path exit implementation fetches both cxq and EntryList. |
|
3740 // See also i486.ad fast_unlock(). Testing has shown that two LDs |
|
3741 // isn't measurably slower than a single LD on any platforms. |
|
3742 // Still, we could reduce the 2 LDs to one or zero by one of the following: |
|
3743 // |
|
3744 // - Use _count instead of cxq|EntryList |
|
3745 // We intend to eliminate _count, however, when we switch |
|
3746 // to on-the-fly deflation in ::exit() as is used in |
|
3747 // Metalocks and RelaxedLocks. |
|
3748 // |
|
3749 // - Establish the invariant that cxq == null implies EntryList == null. |
|
3750 // set cxq == EMPTY (1) to encode the state where cxq is empty |
|
3751 // by EntryList != null. EMPTY is a distinguished value. |
|
3752 // The fast-path exit() would fetch cxq but not EntryList. |
|
3753 // |
|
3754 // - Encode succ as follows: |
|
3755 // succ = t : Thread t is the successor -- t is ready or is spinning. |
|
3756 // Exiting thread does not need to wake a successor. |
|
3757 // succ = 0 : No successor required -> (EntryList|cxq) == null |
|
3758 // Exiting thread does not need to wake a successor |
|
3759 // succ = 1 : Successor required -> (EntryList|cxq) != null and |
|
3760 // logically succ == null. |
|
3761 // Exiting thread must wake a successor. |
|
3762 // |
|
3763 // The 1-1 fast-exit path would appear as : |
|
3764 // _owner = null ; membar ; |
|
3765 // if (_succ == 1 && CAS (&_owner, null, Self) == null) goto SlowPath |
|
3766 // goto FastPathDone ; |
|
3767 // |
|
3768 // and the 1-0 fast-exit path would appear as: |
|
3769 // if (_succ == 1) goto SlowPath |
|
3770 // Owner = null ; |
|
3771 // goto FastPathDone |
|
3772 // |
|
3773 // - Encode the LSB of _owner as 1 to indicate that exit() |
|
3774 // must use the slow-path and make a successor ready. |
|
3775 // (_owner & 1) == 0 IFF succ != null || (EntryList|cxq) == null |
|
3776 // (_owner & 1) == 0 IFF succ == null && (EntryList|cxq) != null (obviously) |
|
3777 // The 1-0 fast exit path would read: |
|
3778 // if (_owner != Self) goto SlowPath |
|
3779 // _owner = null |
|
3780 // goto FastPathDone |
|
3781 |
|
3782 if (Knob_ExitPolicy == 0) { |
|
3783 // release semantics: prior loads and stores from within the critical section |
|
3784 // must not float (reorder) past the following store that drops the lock. |
|
3785 // On SPARC that requires MEMBAR #loadstore|#storestore. |
|
3786 // But of course in TSO #loadstore|#storestore is not required. |
|
3787 // I'd like to write one of the following: |
|
3788 // A. OrderAccess::release() ; _owner = NULL |
|
3789 // B. OrderAccess::loadstore(); OrderAccess::storestore(); _owner = NULL; |
|
3790 // Unfortunately OrderAccess::release() and OrderAccess::loadstore() both |
|
3791 // store into a _dummy variable. That store is not needed, but can result |
|
3792 // in massive wasteful coherency traffic on classic SMP systems. |
|
3793 // Instead, I use release_store(), which is implemented as just a simple |
|
3794 // ST on x64, x86 and SPARC. |
|
3795 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock |
|
3796 OrderAccess::storeload() ; // See if we need to wake a successor |
|
3797 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { |
|
3798 TEVENT (Inflated exit - simple egress) ; |
|
3799 return ; |
|
3800 } |
|
3801 TEVENT (Inflated exit - complex egress) ; |
|
3802 |
|
3803 // Normally the exiting thread is responsible for ensuring succession, |
|
3804 // but if other successors are ready or other entering threads are spinning |
|
3805 // then this thread can simply store NULL into _owner and exit without |
|
3806 // waking a successor. The existence of spinners or ready successors |
|
3807 // guarantees proper succession (liveness). Responsibility passes to the |
|
3808 // ready or running successors. The exiting thread delegates the duty. |
|
3809 // More precisely, if a successor already exists this thread is absolved |
|
3810 // of the responsibility of waking (unparking) one. |
|
3811 // |
|
3812 // The _succ variable is critical to reducing futile wakeup frequency. |
|
3813 // _succ identifies the "heir presumptive" thread that has been made |
|
3814 // ready (unparked) but that has not yet run. We need only one such |
|
3815 // successor thread to guarantee progress. |
|
3816 // See http://www.usenix.org/events/jvm01/full_papers/dice/dice.pdf |
|
3817 // section 3.3 "Futile Wakeup Throttling" for details. |
|
3818 // |
|
3819 // Note that spinners in Enter() also set _succ non-null. |
|
3820 // In the current implementation spinners opportunistically set |
|
3821 // _succ so that exiting threads might avoid waking a successor. |
|
3822 // Another less appealing alternative would be for the exiting thread |
|
3823 // to drop the lock and then spin briefly to see if a spinner managed |
|
3824 // to acquire the lock. If so, the exiting thread could exit |
|
3825 // immediately without waking a successor, otherwise the exiting |
|
3826 // thread would need to dequeue and wake a successor. |
|
3827 // (Note that we'd need to make the post-drop spin short, but no |
|
3828 // shorter than the worst-case round-trip cache-line migration time. |
|
3829 // The dropped lock needs to become visible to the spinner, and then |
|
3830 // the acquisition of the lock by the spinner must become visible to |
|
3831 // the exiting thread). |
|
3832 // |
|
3833 |
|
3834 // It appears that an heir-presumptive (successor) must be made ready. |
|
3835 // Only the current lock owner can manipulate the EntryList or |
|
3836 // drain _cxq, so we need to reacquire the lock. If we fail |
|
3837 // to reacquire the lock the responsibility for ensuring succession |
|
3838 // falls to the new owner. |
|
3839 // |
|
3840 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { |
|
3841 return ; |
|
3842 } |
|
3843 TEVENT (Exit - Reacquired) ; |
|
3844 } else { |
|
3845 if ((intptr_t(_EntryList)|intptr_t(_cxq)) == 0 || _succ != NULL) { |
|
3846 OrderAccess::release_store_ptr (&_owner, NULL) ; // drop the lock |
|
3847 OrderAccess::storeload() ; |
|
3848 // Ratify the previously observed values. |
|
3849 if (_cxq == NULL || _succ != NULL) { |
|
3850 TEVENT (Inflated exit - simple egress) ; |
|
3851 return ; |
|
3852 } |
|
3853 |
|
3854 // inopportune interleaving -- the exiting thread (this thread) |
|
3855 // in the fast-exit path raced an entering thread in the slow-enter |
|
3856 // path. |
|
3857 // We have two choices: |
|
3858 // A. Try to reacquire the lock. |
|
3859 // If the CAS() fails return immediately, otherwise |
|
3860 // we either restart/rerun the exit operation, or simply |
|
3861 // fall-through into the code below which wakes a successor. |
|
3862 // B. If the elements forming the EntryList|cxq are TSM |
|
3863 // we could simply unpark() the lead thread and return |
|
3864 // without having set _succ. |
|
3865 if (Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) != NULL) { |
|
3866 TEVENT (Inflated exit - reacquired succeeded) ; |
|
3867 return ; |
|
3868 } |
|
3869 TEVENT (Inflated exit - reacquired failed) ; |
|
3870 } else { |
|
3871 TEVENT (Inflated exit - complex egress) ; |
|
3872 } |
|
3873 } |
|
3874 |
|
3875 guarantee (_owner == THREAD, "invariant") ; |
|
3876 |
|
3877 // Select an appropriate successor ("heir presumptive") from the EntryList |
|
3878 // and make it ready. Generally we just wake the head of EntryList . |
|
3879 // There's no algorithmic constraint that we use the head - it's just |
|
3880 // a policy decision. Note that the thread at head of the EntryList |
|
3881 // remains at the head until it acquires the lock. This means we'll |
|
3882 // repeatedly wake the same thread until it manages to grab the lock. |
|
3883 // This is generally a good policy - if we're seeing lots of futile wakeups |
|
3884 // at least we're waking/rewaking a thread that's like to be hot or warm |
|
3885 // (have residual D$ and TLB affinity). |
|
3886 // |
|
3887 // "Wakeup locality" optimization: |
|
3888 // http://j2se.east/~dice/PERSIST/040825-WakeLocality.txt |
|
3889 // In the future we'll try to bias the selection mechanism |
|
3890 // to preferentially pick a thread that recently ran on |
|
3891 // a processor element that shares cache with the CPU on which |
|
3892 // the exiting thread is running. We need access to Solaris' |
|
3893 // schedctl.sc_cpu to make that work. |
|
3894 // |
|
3895 ObjectWaiter * w = NULL ; |
|
3896 int QMode = Knob_QMode ; |
|
3897 |
|
3898 if (QMode == 2 && _cxq != NULL) { |
|
3899 // QMode == 2 : cxq has precedence over EntryList. |
|
3900 // Try to directly wake a successor from the cxq. |
|
3901 // If successful, the successor will need to unlink itself from cxq. |
|
3902 w = _cxq ; |
|
3903 assert (w != NULL, "invariant") ; |
|
3904 assert (w->TState == ObjectWaiter::TS_CXQ, "Invariant") ; |
|
3905 ExitEpilog (Self, w) ; |
|
3906 return ; |
|
3907 } |
|
3908 |
|
3909 if (QMode == 3 && _cxq != NULL) { |
|
3910 // Aggressively drain cxq into EntryList at the first opportunity. |
|
3911 // This policy ensure that recently-run threads live at the head of EntryList. |
|
3912 // Drain _cxq into EntryList - bulk transfer. |
|
3913 // First, detach _cxq. |
|
3914 // The following loop is tantamount to: w = swap (&cxq, NULL) |
|
3915 w = _cxq ; |
|
3916 for (;;) { |
|
3917 assert (w != NULL, "Invariant") ; |
|
3918 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; |
|
3919 if (u == w) break ; |
|
3920 w = u ; |
|
3921 } |
|
3922 assert (w != NULL , "invariant") ; |
|
3923 |
|
3924 ObjectWaiter * q = NULL ; |
|
3925 ObjectWaiter * p ; |
|
3926 for (p = w ; p != NULL ; p = p->_next) { |
|
3927 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; |
|
3928 p->TState = ObjectWaiter::TS_ENTER ; |
|
3929 p->_prev = q ; |
|
3930 q = p ; |
|
3931 } |
|
3932 |
|
3933 // Append the RATs to the EntryList |
|
3934 // TODO: organize EntryList as a CDLL so we can locate the tail in constant-time. |
|
3935 ObjectWaiter * Tail ; |
|
3936 for (Tail = _EntryList ; Tail != NULL && Tail->_next != NULL ; Tail = Tail->_next) ; |
|
3937 if (Tail == NULL) { |
|
3938 _EntryList = w ; |
|
3939 } else { |
|
3940 Tail->_next = w ; |
|
3941 w->_prev = Tail ; |
|
3942 } |
|
3943 |
|
3944 // Fall thru into code that tries to wake a successor from EntryList |
|
3945 } |
|
3946 |
|
3947 if (QMode == 4 && _cxq != NULL) { |
|
3948 // Aggressively drain cxq into EntryList at the first opportunity. |
|
3949 // This policy ensure that recently-run threads live at the head of EntryList. |
|
3950 |
|
3951 // Drain _cxq into EntryList - bulk transfer. |
|
3952 // First, detach _cxq. |
|
3953 // The following loop is tantamount to: w = swap (&cxq, NULL) |
|
3954 w = _cxq ; |
|
3955 for (;;) { |
|
3956 assert (w != NULL, "Invariant") ; |
|
3957 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; |
|
3958 if (u == w) break ; |
|
3959 w = u ; |
|
3960 } |
|
3961 assert (w != NULL , "invariant") ; |
|
3962 |
|
3963 ObjectWaiter * q = NULL ; |
|
3964 ObjectWaiter * p ; |
|
3965 for (p = w ; p != NULL ; p = p->_next) { |
|
3966 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; |
|
3967 p->TState = ObjectWaiter::TS_ENTER ; |
|
3968 p->_prev = q ; |
|
3969 q = p ; |
|
3970 } |
|
3971 |
|
3972 // Prepend the RATs to the EntryList |
|
3973 if (_EntryList != NULL) { |
|
3974 q->_next = _EntryList ; |
|
3975 _EntryList->_prev = q ; |
|
3976 } |
|
3977 _EntryList = w ; |
|
3978 |
|
3979 // Fall thru into code that tries to wake a successor from EntryList |
|
3980 } |
|
3981 |
|
3982 w = _EntryList ; |
|
3983 if (w != NULL) { |
|
3984 // I'd like to write: guarantee (w->_thread != Self). |
|
3985 // But in practice an exiting thread may find itself on the EntryList. |
|
3986 // Lets say thread T1 calls O.wait(). Wait() enqueues T1 on O's waitset and |
|
3987 // then calls exit(). Exit release the lock by setting O._owner to NULL. |
|
3988 // Lets say T1 then stalls. T2 acquires O and calls O.notify(). The |
|
3989 // notify() operation moves T1 from O's waitset to O's EntryList. T2 then |
|
3990 // release the lock "O". T2 resumes immediately after the ST of null into |
|
3991 // _owner, above. T2 notices that the EntryList is populated, so it |
|
3992 // reacquires the lock and then finds itself on the EntryList. |
|
3993 // Given all that, we have to tolerate the circumstance where "w" is |
|
3994 // associated with Self. |
|
3995 assert (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; |
|
3996 ExitEpilog (Self, w) ; |
|
3997 return ; |
|
3998 } |
|
3999 |
|
4000 // If we find that both _cxq and EntryList are null then just |
|
4001 // re-run the exit protocol from the top. |
|
4002 w = _cxq ; |
|
4003 if (w == NULL) continue ; |
|
4004 |
|
4005 // Drain _cxq into EntryList - bulk transfer. |
|
4006 // First, detach _cxq. |
|
4007 // The following loop is tantamount to: w = swap (&cxq, NULL) |
|
4008 for (;;) { |
|
4009 assert (w != NULL, "Invariant") ; |
|
4010 ObjectWaiter * u = (ObjectWaiter *) Atomic::cmpxchg_ptr (NULL, &_cxq, w) ; |
|
4011 if (u == w) break ; |
|
4012 w = u ; |
|
4013 } |
|
4014 TEVENT (Inflated exit - drain cxq into EntryList) ; |
|
4015 |
|
4016 assert (w != NULL , "invariant") ; |
|
4017 assert (_EntryList == NULL , "invariant") ; |
|
4018 |
|
4019 // Convert the LIFO SLL anchored by _cxq into a DLL. |
|
4020 // The list reorganization step operates in O(LENGTH(w)) time. |
|
4021 // It's critical that this step operate quickly as |
|
4022 // "Self" still holds the outer-lock, restricting parallelism |
|
4023 // and effectively lengthening the critical section. |
|
4024 // Invariant: s chases t chases u. |
|
4025 // TODO-FIXME: consider changing EntryList from a DLL to a CDLL so |
|
4026 // we have faster access to the tail. |
|
4027 |
|
4028 if (QMode == 1) { |
|
4029 // QMode == 1 : drain cxq to EntryList, reversing order |
|
4030 // We also reverse the order of the list. |
|
4031 ObjectWaiter * s = NULL ; |
|
4032 ObjectWaiter * t = w ; |
|
4033 ObjectWaiter * u = NULL ; |
|
4034 while (t != NULL) { |
|
4035 guarantee (t->TState == ObjectWaiter::TS_CXQ, "invariant") ; |
|
4036 t->TState = ObjectWaiter::TS_ENTER ; |
|
4037 u = t->_next ; |
|
4038 t->_prev = u ; |
|
4039 t->_next = s ; |
|
4040 s = t; |
|
4041 t = u ; |
|
4042 } |
|
4043 _EntryList = s ; |
|
4044 assert (s != NULL, "invariant") ; |
|
4045 } else { |
|
4046 // QMode == 0 or QMode == 2 |
|
4047 _EntryList = w ; |
|
4048 ObjectWaiter * q = NULL ; |
|
4049 ObjectWaiter * p ; |
|
4050 for (p = w ; p != NULL ; p = p->_next) { |
|
4051 guarantee (p->TState == ObjectWaiter::TS_CXQ, "Invariant") ; |
|
4052 p->TState = ObjectWaiter::TS_ENTER ; |
|
4053 p->_prev = q ; |
|
4054 q = p ; |
|
4055 } |
|
4056 } |
|
4057 |
|
4058 // In 1-0 mode we need: ST EntryList; MEMBAR #storestore; ST _owner = NULL |
|
4059 // The MEMBAR is satisfied by the release_store() operation in ExitEpilog(). |
|
4060 |
|
4061 // See if we can abdicate to a spinner instead of waking a thread. |
|
4062 // A primary goal of the implementation is to reduce the |
|
4063 // context-switch rate. |
|
4064 if (_succ != NULL) continue; |
|
4065 |
|
4066 w = _EntryList ; |
|
4067 if (w != NULL) { |
|
4068 guarantee (w->TState == ObjectWaiter::TS_ENTER, "invariant") ; |
|
4069 ExitEpilog (Self, w) ; |
|
4070 return ; |
|
4071 } |
|
4072 } |
|
4073 } |
|
4074 // complete_exit exits a lock returning recursion count |
|
4075 // complete_exit/reenter operate as a wait without waiting |
|
4076 // complete_exit requires an inflated monitor |
|
4077 // The _owner field is not always the Thread addr even with an |
|
4078 // inflated monitor, e.g. the monitor can be inflated by a non-owning |
|
4079 // thread due to contention. |
|
4080 intptr_t ObjectMonitor::complete_exit(TRAPS) { |
|
4081 Thread * const Self = THREAD; |
|
4082 assert(Self->is_Java_thread(), "Must be Java thread!"); |
|
4083 JavaThread *jt = (JavaThread *)THREAD; |
|
4084 |
|
4085 DeferredInitialize(); |
|
4086 |
|
4087 if (THREAD != _owner) { |
|
4088 if (THREAD->is_lock_owned ((address)_owner)) { |
|
4089 assert(_recursions == 0, "internal state error"); |
|
4090 _owner = THREAD ; /* Convert from basiclock addr to Thread addr */ |
|
4091 _recursions = 0 ; |
|
4092 OwnerIsThread = 1 ; |
|
4093 } |
|
4094 } |
|
4095 |
|
4096 guarantee(Self == _owner, "complete_exit not owner"); |
|
4097 intptr_t save = _recursions; // record the old recursion count |
|
4098 _recursions = 0; // set the recursion level to be 0 |
|
4099 exit (Self) ; // exit the monitor |
|
4100 guarantee (_owner != Self, "invariant"); |
|
4101 return save; |
|
4102 } |
|
4103 |
|
4104 // reenter() enters a lock and sets recursion count |
|
4105 // complete_exit/reenter operate as a wait without waiting |
|
4106 void ObjectMonitor::reenter(intptr_t recursions, TRAPS) { |
|
4107 Thread * const Self = THREAD; |
|
4108 assert(Self->is_Java_thread(), "Must be Java thread!"); |
|
4109 JavaThread *jt = (JavaThread *)THREAD; |
|
4110 |
|
4111 guarantee(_owner != Self, "reenter already owner"); |
|
4112 enter (THREAD); // enter the monitor |
|
4113 guarantee (_recursions == 0, "reenter recursion"); |
|
4114 _recursions = recursions; |
|
4115 return; |
|
4116 } |
|
4117 |
|
4118 // Note: a subset of changes to ObjectMonitor::wait() |
|
4119 // will need to be replicated in complete_exit above |
|
4120 void ObjectMonitor::wait(jlong millis, bool interruptible, TRAPS) { |
|
4121 Thread * const Self = THREAD ; |
|
4122 assert(Self->is_Java_thread(), "Must be Java thread!"); |
|
4123 JavaThread *jt = (JavaThread *)THREAD; |
|
4124 |
|
4125 DeferredInitialize () ; |
|
4126 |
|
4127 // Throw IMSX or IEX. |
|
4128 CHECK_OWNER(); |
|
4129 |
|
4130 // check for a pending interrupt |
|
4131 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { |
|
4132 // post monitor waited event. Note that this is past-tense, we are done waiting. |
|
4133 if (JvmtiExport::should_post_monitor_waited()) { |
|
4134 // Note: 'false' parameter is passed here because the |
|
4135 // wait was not timed out due to thread interrupt. |
|
4136 JvmtiExport::post_monitor_waited(jt, this, false); |
|
4137 } |
|
4138 TEVENT (Wait - Throw IEX) ; |
|
4139 THROW(vmSymbols::java_lang_InterruptedException()); |
|
4140 return ; |
|
4141 } |
|
4142 TEVENT (Wait) ; |
|
4143 |
|
4144 assert (Self->_Stalled == 0, "invariant") ; |
|
4145 Self->_Stalled = intptr_t(this) ; |
|
4146 jt->set_current_waiting_monitor(this); |
|
4147 |
|
4148 // create a node to be put into the queue |
|
4149 // Critically, after we reset() the event but prior to park(), we must check |
|
4150 // for a pending interrupt. |
|
4151 ObjectWaiter node(Self); |
|
4152 node.TState = ObjectWaiter::TS_WAIT ; |
|
4153 Self->_ParkEvent->reset() ; |
|
4154 OrderAccess::fence(); // ST into Event; membar ; LD interrupted-flag |
|
4155 |
|
4156 // Enter the waiting queue, which is a circular doubly linked list in this case |
|
4157 // but it could be a priority queue or any data structure. |
|
4158 // _WaitSetLock protects the wait queue. Normally the wait queue is accessed only |
|
4159 // by the the owner of the monitor *except* in the case where park() |
|
4160 // returns because of a timeout of interrupt. Contention is exceptionally rare |
|
4161 // so we use a simple spin-lock instead of a heavier-weight blocking lock. |
|
4162 |
|
4163 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - add") ; |
|
4164 AddWaiter (&node) ; |
|
4165 Thread::SpinRelease (&_WaitSetLock) ; |
|
4166 |
|
4167 if ((SyncFlags & 4) == 0) { |
|
4168 _Responsible = NULL ; |
|
4169 } |
|
4170 intptr_t save = _recursions; // record the old recursion count |
|
4171 _waiters++; // increment the number of waiters |
|
4172 _recursions = 0; // set the recursion level to be 1 |
|
4173 exit (Self) ; // exit the monitor |
|
4174 guarantee (_owner != Self, "invariant") ; |
|
4175 |
|
4176 // As soon as the ObjectMonitor's ownership is dropped in the exit() |
|
4177 // call above, another thread can enter() the ObjectMonitor, do the |
|
4178 // notify(), and exit() the ObjectMonitor. If the other thread's |
|
4179 // exit() call chooses this thread as the successor and the unpark() |
|
4180 // call happens to occur while this thread is posting a |
|
4181 // MONITOR_CONTENDED_EXIT event, then we run the risk of the event |
|
4182 // handler using RawMonitors and consuming the unpark(). |
|
4183 // |
|
4184 // To avoid the problem, we re-post the event. This does no harm |
|
4185 // even if the original unpark() was not consumed because we are the |
|
4186 // chosen successor for this monitor. |
|
4187 if (node._notified != 0 && _succ == Self) { |
|
4188 node._event->unpark(); |
|
4189 } |
|
4190 |
|
4191 // The thread is on the WaitSet list - now park() it. |
|
4192 // On MP systems it's conceivable that a brief spin before we park |
|
4193 // could be profitable. |
|
4194 // |
|
4195 // TODO-FIXME: change the following logic to a loop of the form |
|
4196 // while (!timeout && !interrupted && _notified == 0) park() |
|
4197 |
|
4198 int ret = OS_OK ; |
|
4199 int WasNotified = 0 ; |
|
4200 { // State transition wrappers |
|
4201 OSThread* osthread = Self->osthread(); |
|
4202 OSThreadWaitState osts(osthread, true); |
|
4203 { |
|
4204 ThreadBlockInVM tbivm(jt); |
|
4205 // Thread is in thread_blocked state and oop access is unsafe. |
|
4206 jt->set_suspend_equivalent(); |
|
4207 |
|
4208 if (interruptible && (Thread::is_interrupted(THREAD, false) || HAS_PENDING_EXCEPTION)) { |
|
4209 // Intentionally empty |
|
4210 } else |
|
4211 if (node._notified == 0) { |
|
4212 if (millis <= 0) { |
|
4213 Self->_ParkEvent->park () ; |
|
4214 } else { |
|
4215 ret = Self->_ParkEvent->park (millis) ; |
|
4216 } |
|
4217 } |
|
4218 |
|
4219 // were we externally suspended while we were waiting? |
|
4220 if (ExitSuspendEquivalent (jt)) { |
|
4221 // TODO-FIXME: add -- if succ == Self then succ = null. |
|
4222 jt->java_suspend_self(); |
|
4223 } |
|
4224 |
|
4225 } // Exit thread safepoint: transition _thread_blocked -> _thread_in_vm |
|
4226 |
|
4227 |
|
4228 // Node may be on the WaitSet, the EntryList (or cxq), or in transition |
|
4229 // from the WaitSet to the EntryList. |
|
4230 // See if we need to remove Node from the WaitSet. |
|
4231 // We use double-checked locking to avoid grabbing _WaitSetLock |
|
4232 // if the thread is not on the wait queue. |
|
4233 // |
|
4234 // Note that we don't need a fence before the fetch of TState. |
|
4235 // In the worst case we'll fetch a old-stale value of TS_WAIT previously |
|
4236 // written by the is thread. (perhaps the fetch might even be satisfied |
|
4237 // by a look-aside into the processor's own store buffer, although given |
|
4238 // the length of the code path between the prior ST and this load that's |
|
4239 // highly unlikely). If the following LD fetches a stale TS_WAIT value |
|
4240 // then we'll acquire the lock and then re-fetch a fresh TState value. |
|
4241 // That is, we fail toward safety. |
|
4242 |
|
4243 if (node.TState == ObjectWaiter::TS_WAIT) { |
|
4244 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - unlink") ; |
|
4245 if (node.TState == ObjectWaiter::TS_WAIT) { |
|
4246 DequeueSpecificWaiter (&node) ; // unlink from WaitSet |
|
4247 assert(node._notified == 0, "invariant"); |
|
4248 node.TState = ObjectWaiter::TS_RUN ; |
|
4249 } |
|
4250 Thread::SpinRelease (&_WaitSetLock) ; |
|
4251 } |
|
4252 |
|
4253 // The thread is now either on off-list (TS_RUN), |
|
4254 // on the EntryList (TS_ENTER), or on the cxq (TS_CXQ). |
|
4255 // The Node's TState variable is stable from the perspective of this thread. |
|
4256 // No other threads will asynchronously modify TState. |
|
4257 guarantee (node.TState != ObjectWaiter::TS_WAIT, "invariant") ; |
|
4258 OrderAccess::loadload() ; |
|
4259 if (_succ == Self) _succ = NULL ; |
|
4260 WasNotified = node._notified ; |
|
4261 |
|
4262 // Reentry phase -- reacquire the monitor. |
|
4263 // re-enter contended monitor after object.wait(). |
|
4264 // retain OBJECT_WAIT state until re-enter successfully completes |
|
4265 // Thread state is thread_in_vm and oop access is again safe, |
|
4266 // although the raw address of the object may have changed. |
|
4267 // (Don't cache naked oops over safepoints, of course). |
|
4268 |
|
4269 // post monitor waited event. Note that this is past-tense, we are done waiting. |
|
4270 if (JvmtiExport::should_post_monitor_waited()) { |
|
4271 JvmtiExport::post_monitor_waited(jt, this, ret == OS_TIMEOUT); |
|
4272 } |
|
4273 OrderAccess::fence() ; |
|
4274 |
|
4275 assert (Self->_Stalled != 0, "invariant") ; |
|
4276 Self->_Stalled = 0 ; |
|
4277 |
|
4278 assert (_owner != Self, "invariant") ; |
|
4279 ObjectWaiter::TStates v = node.TState ; |
|
4280 if (v == ObjectWaiter::TS_RUN) { |
|
4281 enter (Self) ; |
|
4282 } else { |
|
4283 guarantee (v == ObjectWaiter::TS_ENTER || v == ObjectWaiter::TS_CXQ, "invariant") ; |
|
4284 ReenterI (Self, &node) ; |
|
4285 node.wait_reenter_end(this); |
|
4286 } |
|
4287 |
|
4288 // Self has reacquired the lock. |
|
4289 // Lifecycle - the node representing Self must not appear on any queues. |
|
4290 // Node is about to go out-of-scope, but even if it were immortal we wouldn't |
|
4291 // want residual elements associated with this thread left on any lists. |
|
4292 guarantee (node.TState == ObjectWaiter::TS_RUN, "invariant") ; |
|
4293 assert (_owner == Self, "invariant") ; |
|
4294 assert (_succ != Self , "invariant") ; |
|
4295 } // OSThreadWaitState() |
|
4296 |
|
4297 jt->set_current_waiting_monitor(NULL); |
|
4298 |
|
4299 guarantee (_recursions == 0, "invariant") ; |
|
4300 _recursions = save; // restore the old recursion count |
|
4301 _waiters--; // decrement the number of waiters |
|
4302 |
|
4303 // Verify a few postconditions |
|
4304 assert (_owner == Self , "invariant") ; |
|
4305 assert (_succ != Self , "invariant") ; |
|
4306 assert (((oop)(object()))->mark() == markOopDesc::encode(this), "invariant") ; |
|
4307 |
|
4308 if (SyncFlags & 32) { |
|
4309 OrderAccess::fence() ; |
|
4310 } |
|
4311 |
|
4312 // check if the notification happened |
|
4313 if (!WasNotified) { |
|
4314 // no, it could be timeout or Thread.interrupt() or both |
|
4315 // check for interrupt event, otherwise it is timeout |
|
4316 if (interruptible && Thread::is_interrupted(Self, true) && !HAS_PENDING_EXCEPTION) { |
|
4317 TEVENT (Wait - throw IEX from epilog) ; |
|
4318 THROW(vmSymbols::java_lang_InterruptedException()); |
|
4319 } |
|
4320 } |
|
4321 |
|
4322 // NOTE: Spurious wake up will be consider as timeout. |
|
4323 // Monitor notify has precedence over thread interrupt. |
|
4324 } |
|
4325 |
|
4326 |
|
4327 // Consider: |
|
4328 // If the lock is cool (cxq == null && succ == null) and we're on an MP system |
|
4329 // then instead of transferring a thread from the WaitSet to the EntryList |
|
4330 // we might just dequeue a thread from the WaitSet and directly unpark() it. |
|
4331 |
|
4332 void ObjectMonitor::notify(TRAPS) { |
|
4333 CHECK_OWNER(); |
|
4334 if (_WaitSet == NULL) { |
|
4335 TEVENT (Empty-Notify) ; |
|
4336 return ; |
|
4337 } |
|
4338 DTRACE_MONITOR_PROBE(notify, this, object(), THREAD); |
|
4339 |
|
4340 int Policy = Knob_MoveNotifyee ; |
|
4341 |
|
4342 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notify") ; |
|
4343 ObjectWaiter * iterator = DequeueWaiter() ; |
|
4344 if (iterator != NULL) { |
|
4345 TEVENT (Notify1 - Transfer) ; |
|
4346 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; |
|
4347 guarantee (iterator->_notified == 0, "invariant") ; |
|
4348 // Disposition - what might we do with iterator ? |
|
4349 // a. add it directly to the EntryList - either tail or head. |
|
4350 // b. push it onto the front of the _cxq. |
|
4351 // For now we use (a). |
|
4352 if (Policy != 4) { |
|
4353 iterator->TState = ObjectWaiter::TS_ENTER ; |
|
4354 } |
|
4355 iterator->_notified = 1 ; |
|
4356 |
|
4357 ObjectWaiter * List = _EntryList ; |
|
4358 if (List != NULL) { |
|
4359 assert (List->_prev == NULL, "invariant") ; |
|
4360 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; |
|
4361 assert (List != iterator, "invariant") ; |
|
4362 } |
|
4363 |
|
4364 if (Policy == 0) { // prepend to EntryList |
|
4365 if (List == NULL) { |
|
4366 iterator->_next = iterator->_prev = NULL ; |
|
4367 _EntryList = iterator ; |
|
4368 } else { |
|
4369 List->_prev = iterator ; |
|
4370 iterator->_next = List ; |
|
4371 iterator->_prev = NULL ; |
|
4372 _EntryList = iterator ; |
|
4373 } |
|
4374 } else |
|
4375 if (Policy == 1) { // append to EntryList |
|
4376 if (List == NULL) { |
|
4377 iterator->_next = iterator->_prev = NULL ; |
|
4378 _EntryList = iterator ; |
|
4379 } else { |
|
4380 // CONSIDER: finding the tail currently requires a linear-time walk of |
|
4381 // the EntryList. We can make tail access constant-time by converting to |
|
4382 // a CDLL instead of using our current DLL. |
|
4383 ObjectWaiter * Tail ; |
|
4384 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; |
|
4385 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; |
|
4386 Tail->_next = iterator ; |
|
4387 iterator->_prev = Tail ; |
|
4388 iterator->_next = NULL ; |
|
4389 } |
|
4390 } else |
|
4391 if (Policy == 2) { // prepend to cxq |
|
4392 // prepend to cxq |
|
4393 if (List == NULL) { |
|
4394 iterator->_next = iterator->_prev = NULL ; |
|
4395 _EntryList = iterator ; |
|
4396 } else { |
|
4397 iterator->TState = ObjectWaiter::TS_CXQ ; |
|
4398 for (;;) { |
|
4399 ObjectWaiter * Front = _cxq ; |
|
4400 iterator->_next = Front ; |
|
4401 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { |
|
4402 break ; |
|
4403 } |
|
4404 } |
|
4405 } |
|
4406 } else |
|
4407 if (Policy == 3) { // append to cxq |
|
4408 iterator->TState = ObjectWaiter::TS_CXQ ; |
|
4409 for (;;) { |
|
4410 ObjectWaiter * Tail ; |
|
4411 Tail = _cxq ; |
|
4412 if (Tail == NULL) { |
|
4413 iterator->_next = NULL ; |
|
4414 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { |
|
4415 break ; |
|
4416 } |
|
4417 } else { |
|
4418 while (Tail->_next != NULL) Tail = Tail->_next ; |
|
4419 Tail->_next = iterator ; |
|
4420 iterator->_prev = Tail ; |
|
4421 iterator->_next = NULL ; |
|
4422 break ; |
|
4423 } |
|
4424 } |
|
4425 } else { |
|
4426 ParkEvent * ev = iterator->_event ; |
|
4427 iterator->TState = ObjectWaiter::TS_RUN ; |
|
4428 OrderAccess::fence() ; |
|
4429 ev->unpark() ; |
|
4430 } |
|
4431 |
|
4432 if (Policy < 4) { |
|
4433 iterator->wait_reenter_begin(this); |
|
4434 } |
|
4435 |
|
4436 // _WaitSetLock protects the wait queue, not the EntryList. We could |
|
4437 // move the add-to-EntryList operation, above, outside the critical section |
|
4438 // protected by _WaitSetLock. In practice that's not useful. With the |
|
4439 // exception of wait() timeouts and interrupts the monitor owner |
|
4440 // is the only thread that grabs _WaitSetLock. There's almost no contention |
|
4441 // on _WaitSetLock so it's not profitable to reduce the length of the |
|
4442 // critical section. |
|
4443 } |
|
4444 |
|
4445 Thread::SpinRelease (&_WaitSetLock) ; |
|
4446 |
|
4447 if (iterator != NULL && ObjectSynchronizer::_sync_Notifications != NULL) { |
|
4448 ObjectSynchronizer::_sync_Notifications->inc() ; |
|
4449 } |
|
4450 } |
|
4451 |
|
4452 |
|
4453 void ObjectMonitor::notifyAll(TRAPS) { |
|
4454 CHECK_OWNER(); |
|
4455 ObjectWaiter* iterator; |
|
4456 if (_WaitSet == NULL) { |
|
4457 TEVENT (Empty-NotifyAll) ; |
|
4458 return ; |
|
4459 } |
|
4460 DTRACE_MONITOR_PROBE(notifyAll, this, object(), THREAD); |
|
4461 |
|
4462 int Policy = Knob_MoveNotifyee ; |
|
4463 int Tally = 0 ; |
|
4464 Thread::SpinAcquire (&_WaitSetLock, "WaitSet - notifyall") ; |
|
4465 |
|
4466 for (;;) { |
|
4467 iterator = DequeueWaiter () ; |
|
4468 if (iterator == NULL) break ; |
|
4469 TEVENT (NotifyAll - Transfer1) ; |
|
4470 ++Tally ; |
|
4471 |
|
4472 // Disposition - what might we do with iterator ? |
|
4473 // a. add it directly to the EntryList - either tail or head. |
|
4474 // b. push it onto the front of the _cxq. |
|
4475 // For now we use (a). |
|
4476 // |
|
4477 // TODO-FIXME: currently notifyAll() transfers the waiters one-at-a-time from the waitset |
|
4478 // to the EntryList. This could be done more efficiently with a single bulk transfer, |
|
4479 // but in practice it's not time-critical. Beware too, that in prepend-mode we invert the |
|
4480 // order of the waiters. Lets say that the waitset is "ABCD" and the EntryList is "XYZ". |
|
4481 // After a notifyAll() in prepend mode the waitset will be empty and the EntryList will |
|
4482 // be "DCBAXYZ". |
|
4483 |
|
4484 guarantee (iterator->TState == ObjectWaiter::TS_WAIT, "invariant") ; |
|
4485 guarantee (iterator->_notified == 0, "invariant") ; |
|
4486 iterator->_notified = 1 ; |
|
4487 if (Policy != 4) { |
|
4488 iterator->TState = ObjectWaiter::TS_ENTER ; |
|
4489 } |
|
4490 |
|
4491 ObjectWaiter * List = _EntryList ; |
|
4492 if (List != NULL) { |
|
4493 assert (List->_prev == NULL, "invariant") ; |
|
4494 assert (List->TState == ObjectWaiter::TS_ENTER, "invariant") ; |
|
4495 assert (List != iterator, "invariant") ; |
|
4496 } |
|
4497 |
|
4498 if (Policy == 0) { // prepend to EntryList |
|
4499 if (List == NULL) { |
|
4500 iterator->_next = iterator->_prev = NULL ; |
|
4501 _EntryList = iterator ; |
|
4502 } else { |
|
4503 List->_prev = iterator ; |
|
4504 iterator->_next = List ; |
|
4505 iterator->_prev = NULL ; |
|
4506 _EntryList = iterator ; |
|
4507 } |
|
4508 } else |
|
4509 if (Policy == 1) { // append to EntryList |
|
4510 if (List == NULL) { |
|
4511 iterator->_next = iterator->_prev = NULL ; |
|
4512 _EntryList = iterator ; |
|
4513 } else { |
|
4514 // CONSIDER: finding the tail currently requires a linear-time walk of |
|
4515 // the EntryList. We can make tail access constant-time by converting to |
|
4516 // a CDLL instead of using our current DLL. |
|
4517 ObjectWaiter * Tail ; |
|
4518 for (Tail = List ; Tail->_next != NULL ; Tail = Tail->_next) ; |
|
4519 assert (Tail != NULL && Tail->_next == NULL, "invariant") ; |
|
4520 Tail->_next = iterator ; |
|
4521 iterator->_prev = Tail ; |
|
4522 iterator->_next = NULL ; |
|
4523 } |
|
4524 } else |
|
4525 if (Policy == 2) { // prepend to cxq |
|
4526 // prepend to cxq |
|
4527 iterator->TState = ObjectWaiter::TS_CXQ ; |
|
4528 for (;;) { |
|
4529 ObjectWaiter * Front = _cxq ; |
|
4530 iterator->_next = Front ; |
|
4531 if (Atomic::cmpxchg_ptr (iterator, &_cxq, Front) == Front) { |
|
4532 break ; |
|
4533 } |
|
4534 } |
|
4535 } else |
|
4536 if (Policy == 3) { // append to cxq |
|
4537 iterator->TState = ObjectWaiter::TS_CXQ ; |
|
4538 for (;;) { |
|
4539 ObjectWaiter * Tail ; |
|
4540 Tail = _cxq ; |
|
4541 if (Tail == NULL) { |
|
4542 iterator->_next = NULL ; |
|
4543 if (Atomic::cmpxchg_ptr (iterator, &_cxq, NULL) == NULL) { |
|
4544 break ; |
|
4545 } |
|
4546 } else { |
|
4547 while (Tail->_next != NULL) Tail = Tail->_next ; |
|
4548 Tail->_next = iterator ; |
|
4549 iterator->_prev = Tail ; |
|
4550 iterator->_next = NULL ; |
|
4551 break ; |
|
4552 } |
|
4553 } |
|
4554 } else { |
|
4555 ParkEvent * ev = iterator->_event ; |
|
4556 iterator->TState = ObjectWaiter::TS_RUN ; |
|
4557 OrderAccess::fence() ; |
|
4558 ev->unpark() ; |
|
4559 } |
|
4560 |
|
4561 if (Policy < 4) { |
|
4562 iterator->wait_reenter_begin(this); |
|
4563 } |
|
4564 |
|
4565 // _WaitSetLock protects the wait queue, not the EntryList. We could |
|
4566 // move the add-to-EntryList operation, above, outside the critical section |
|
4567 // protected by _WaitSetLock. In practice that's not useful. With the |
|
4568 // exception of wait() timeouts and interrupts the monitor owner |
|
4569 // is the only thread that grabs _WaitSetLock. There's almost no contention |
|
4570 // on _WaitSetLock so it's not profitable to reduce the length of the |
|
4571 // critical section. |
|
4572 } |
|
4573 |
|
4574 Thread::SpinRelease (&_WaitSetLock) ; |
|
4575 |
|
4576 if (Tally != 0 && ObjectSynchronizer::_sync_Notifications != NULL) { |
|
4577 ObjectSynchronizer::_sync_Notifications->inc(Tally) ; |
|
4578 } |
|
4579 } |
|
4580 |
|
4581 // check_slow() is a misnomer. It's called to simply to throw an IMSX exception. |
|
4582 // TODO-FIXME: remove check_slow() -- it's likely dead. |
|
4583 |
|
4584 void ObjectMonitor::check_slow(TRAPS) { |
|
4585 TEVENT (check_slow - throw IMSX) ; |
|
4586 assert(THREAD != _owner && !THREAD->is_lock_owned((address) _owner), "must not be owner"); |
|
4587 THROW_MSG(vmSymbols::java_lang_IllegalMonitorStateException(), "current thread not owner"); |
|
4588 } |
|
4589 |
|
4590 |
|
4591 // ------------------------------------------------------------------------- |
|
4592 // The raw monitor subsystem is entirely distinct from normal |
|
4593 // java-synchronization or jni-synchronization. raw monitors are not |
|
4594 // associated with objects. They can be implemented in any manner |
|
4595 // that makes sense. The original implementors decided to piggy-back |
|
4596 // the raw-monitor implementation on the existing Java objectMonitor mechanism. |
|
4597 // This flaw needs to fixed. We should reimplement raw monitors as sui-generis. |
|
4598 // Specifically, we should not implement raw monitors via java monitors. |
|
4599 // Time permitting, we should disentangle and deconvolve the two implementations |
|
4600 // and move the resulting raw monitor implementation over to the JVMTI directories. |
|
4601 // Ideally, the raw monitor implementation would be built on top of |
|
4602 // park-unpark and nothing else. |
|
4603 // |
|
4604 // raw monitors are used mainly by JVMTI |
|
4605 // The raw monitor implementation borrows the ObjectMonitor structure, |
|
4606 // but the operators are degenerate and extremely simple. |
|
4607 // |
|
4608 // Mixed use of a single objectMonitor instance -- as both a raw monitor |
|
4609 // and a normal java monitor -- is not permissible. |
|
4610 // |
|
4611 // Note that we use the single RawMonitor_lock to protect queue operations for |
|
4612 // _all_ raw monitors. This is a scalability impediment, but since raw monitor usage |
|
4613 // is deprecated and rare, this is not of concern. The RawMonitor_lock can not |
|
4614 // be held indefinitely. The critical sections must be short and bounded. |
|
4615 // |
|
4616 // ------------------------------------------------------------------------- |
|
4617 |
|
4618 int ObjectMonitor::SimpleEnter (Thread * Self) { |
|
4619 for (;;) { |
|
4620 if (Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { |
|
4621 return OS_OK ; |
|
4622 } |
|
4623 |
|
4624 ObjectWaiter Node (Self) ; |
|
4625 Self->_ParkEvent->reset() ; // strictly optional |
|
4626 Node.TState = ObjectWaiter::TS_ENTER ; |
|
4627 |
|
4628 RawMonitor_lock->lock_without_safepoint_check() ; |
|
4629 Node._next = _EntryList ; |
|
4630 _EntryList = &Node ; |
|
4631 OrderAccess::fence() ; |
|
4632 if (_owner == NULL && Atomic::cmpxchg_ptr (Self, &_owner, NULL) == NULL) { |
|
4633 _EntryList = Node._next ; |
|
4634 RawMonitor_lock->unlock() ; |
|
4635 return OS_OK ; |
|
4636 } |
|
4637 RawMonitor_lock->unlock() ; |
|
4638 while (Node.TState == ObjectWaiter::TS_ENTER) { |
|
4639 Self->_ParkEvent->park() ; |
|
4640 } |
|
4641 } |
|
4642 } |
|
4643 |
|
4644 int ObjectMonitor::SimpleExit (Thread * Self) { |
|
4645 guarantee (_owner == Self, "invariant") ; |
|
4646 OrderAccess::release_store_ptr (&_owner, NULL) ; |
|
4647 OrderAccess::fence() ; |
|
4648 if (_EntryList == NULL) return OS_OK ; |
|
4649 ObjectWaiter * w ; |
|
4650 |
|
4651 RawMonitor_lock->lock_without_safepoint_check() ; |
|
4652 w = _EntryList ; |
|
4653 if (w != NULL) { |
|
4654 _EntryList = w->_next ; |
|
4655 } |
|
4656 RawMonitor_lock->unlock() ; |
|
4657 if (w != NULL) { |
|
4658 guarantee (w ->TState == ObjectWaiter::TS_ENTER, "invariant") ; |
|
4659 ParkEvent * ev = w->_event ; |
|
4660 w->TState = ObjectWaiter::TS_RUN ; |
|
4661 OrderAccess::fence() ; |
|
4662 ev->unpark() ; |
|
4663 } |
|
4664 return OS_OK ; |
|
4665 } |
|
4666 |
|
4667 int ObjectMonitor::SimpleWait (Thread * Self, jlong millis) { |
|
4668 guarantee (_owner == Self , "invariant") ; |
|
4669 guarantee (_recursions == 0, "invariant") ; |
|
4670 |
|
4671 ObjectWaiter Node (Self) ; |
|
4672 Node._notified = 0 ; |
|
4673 Node.TState = ObjectWaiter::TS_WAIT ; |
|
4674 |
|
4675 RawMonitor_lock->lock_without_safepoint_check() ; |
|
4676 Node._next = _WaitSet ; |
|
4677 _WaitSet = &Node ; |
|
4678 RawMonitor_lock->unlock() ; |
|
4679 |
|
4680 SimpleExit (Self) ; |
|
4681 guarantee (_owner != Self, "invariant") ; |
|
4682 |
|
4683 int ret = OS_OK ; |
|
4684 if (millis <= 0) { |
|
4685 Self->_ParkEvent->park(); |
|
4686 } else { |
|
4687 ret = Self->_ParkEvent->park(millis); |
|
4688 } |
|
4689 |
|
4690 // If thread still resides on the waitset then unlink it. |
|
4691 // Double-checked locking -- the usage is safe in this context |
|
4692 // as we TState is volatile and the lock-unlock operators are |
|
4693 // serializing (barrier-equivalent). |
|
4694 |
|
4695 if (Node.TState == ObjectWaiter::TS_WAIT) { |
|
4696 RawMonitor_lock->lock_without_safepoint_check() ; |
|
4697 if (Node.TState == ObjectWaiter::TS_WAIT) { |
|
4698 // Simple O(n) unlink, but performance isn't critical here. |
|
4699 ObjectWaiter * p ; |
|
4700 ObjectWaiter * q = NULL ; |
|
4701 for (p = _WaitSet ; p != &Node; p = p->_next) { |
|
4702 q = p ; |
|
4703 } |
|
4704 guarantee (p == &Node, "invariant") ; |
|
4705 if (q == NULL) { |
|
4706 guarantee (p == _WaitSet, "invariant") ; |
|
4707 _WaitSet = p->_next ; |
|
4708 } else { |
|
4709 guarantee (p == q->_next, "invariant") ; |
|
4710 q->_next = p->_next ; |
|
4711 } |
|
4712 Node.TState = ObjectWaiter::TS_RUN ; |
|
4713 } |
|
4714 RawMonitor_lock->unlock() ; |
|
4715 } |
|
4716 |
|
4717 guarantee (Node.TState == ObjectWaiter::TS_RUN, "invariant") ; |
|
4718 SimpleEnter (Self) ; |
|
4719 |
|
4720 guarantee (_owner == Self, "invariant") ; |
|
4721 guarantee (_recursions == 0, "invariant") ; |
|
4722 return ret ; |
|
4723 } |
|
4724 |
|
4725 int ObjectMonitor::SimpleNotify (Thread * Self, bool All) { |
|
4726 guarantee (_owner == Self, "invariant") ; |
|
4727 if (_WaitSet == NULL) return OS_OK ; |
|
4728 |
|
4729 // We have two options: |
|
4730 // A. Transfer the threads from the WaitSet to the EntryList |
|
4731 // B. Remove the thread from the WaitSet and unpark() it. |
|
4732 // |
|
4733 // We use (B), which is crude and results in lots of futile |
|
4734 // context switching. In particular (B) induces lots of contention. |
|
4735 |
|
4736 ParkEvent * ev = NULL ; // consider using a small auto array ... |
|
4737 RawMonitor_lock->lock_without_safepoint_check() ; |
|
4738 for (;;) { |
|
4739 ObjectWaiter * w = _WaitSet ; |
|
4740 if (w == NULL) break ; |
|
4741 _WaitSet = w->_next ; |
|
4742 if (ev != NULL) { ev->unpark(); ev = NULL; } |
|
4743 ev = w->_event ; |
|
4744 OrderAccess::loadstore() ; |
|
4745 w->TState = ObjectWaiter::TS_RUN ; |
|
4746 OrderAccess::storeload(); |
|
4747 if (!All) break ; |
|
4748 } |
|
4749 RawMonitor_lock->unlock() ; |
|
4750 if (ev != NULL) ev->unpark(); |
|
4751 return OS_OK ; |
|
4752 } |
|
4753 |
|
4754 // Any JavaThread will enter here with state _thread_blocked |
|
4755 int ObjectMonitor::raw_enter(TRAPS) { |
|
4756 TEVENT (raw_enter) ; |
|
4757 void * Contended ; |
|
4758 |
|
4759 // don't enter raw monitor if thread is being externally suspended, it will |
|
4760 // surprise the suspender if a "suspended" thread can still enter monitor |
|
4761 JavaThread * jt = (JavaThread *)THREAD; |
|
4762 if (THREAD->is_Java_thread()) { |
|
4763 jt->SR_lock()->lock_without_safepoint_check(); |
|
4764 while (jt->is_external_suspend()) { |
|
4765 jt->SR_lock()->unlock(); |
|
4766 jt->java_suspend_self(); |
|
4767 jt->SR_lock()->lock_without_safepoint_check(); |
|
4768 } |
|
4769 // guarded by SR_lock to avoid racing with new external suspend requests. |
|
4770 Contended = Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) ; |
|
4771 jt->SR_lock()->unlock(); |
|
4772 } else { |
|
4773 Contended = Atomic::cmpxchg_ptr (THREAD, &_owner, NULL) ; |
|
4774 } |
|
4775 |
|
4776 if (Contended == THREAD) { |
|
4777 _recursions ++ ; |
|
4778 return OM_OK ; |
|
4779 } |
|
4780 |
|
4781 if (Contended == NULL) { |
|
4782 guarantee (_owner == THREAD, "invariant") ; |
|
4783 guarantee (_recursions == 0, "invariant") ; |
|
4784 return OM_OK ; |
|
4785 } |
|
4786 |
|
4787 THREAD->set_current_pending_monitor(this); |
|
4788 |
|
4789 if (!THREAD->is_Java_thread()) { |
|
4790 // No other non-Java threads besides VM thread would acquire |
|
4791 // a raw monitor. |
|
4792 assert(THREAD->is_VM_thread(), "must be VM thread"); |
|
4793 SimpleEnter (THREAD) ; |
|
4794 } else { |
|
4795 guarantee (jt->thread_state() == _thread_blocked, "invariant") ; |
|
4796 for (;;) { |
|
4797 jt->set_suspend_equivalent(); |
|
4798 // cleared by handle_special_suspend_equivalent_condition() or |
|
4799 // java_suspend_self() |
|
4800 SimpleEnter (THREAD) ; |
|
4801 |
|
4802 // were we externally suspended while we were waiting? |
|
4803 if (!jt->handle_special_suspend_equivalent_condition()) break ; |
|
4804 |
|
4805 // This thread was externally suspended |
|
4806 // |
|
4807 // This logic isn't needed for JVMTI raw monitors, |
|
4808 // but doesn't hurt just in case the suspend rules change. This |
|
4809 // logic is needed for the ObjectMonitor.wait() reentry phase. |
|
4810 // We have reentered the contended monitor, but while we were |
|
4811 // waiting another thread suspended us. We don't want to reenter |
|
4812 // the monitor while suspended because that would surprise the |
|
4813 // thread that suspended us. |
|
4814 // |
|
4815 // Drop the lock - |
|
4816 SimpleExit (THREAD) ; |
|
4817 |
|
4818 jt->java_suspend_self(); |
|
4819 } |
|
4820 |
|
4821 assert(_owner == THREAD, "Fatal error with monitor owner!"); |
|
4822 assert(_recursions == 0, "Fatal error with monitor recursions!"); |
|
4823 } |
|
4824 |
|
4825 THREAD->set_current_pending_monitor(NULL); |
|
4826 guarantee (_recursions == 0, "invariant") ; |
|
4827 return OM_OK; |
|
4828 } |
|
4829 |
|
4830 // Used mainly for JVMTI raw monitor implementation |
|
4831 // Also used for ObjectMonitor::wait(). |
|
4832 int ObjectMonitor::raw_exit(TRAPS) { |
|
4833 TEVENT (raw_exit) ; |
|
4834 if (THREAD != _owner) { |
|
4835 return OM_ILLEGAL_MONITOR_STATE; |
|
4836 } |
|
4837 if (_recursions > 0) { |
|
4838 --_recursions ; |
|
4839 return OM_OK ; |
|
4840 } |
|
4841 |
|
4842 void * List = _EntryList ; |
|
4843 SimpleExit (THREAD) ; |
|
4844 |
|
4845 return OM_OK; |
|
4846 } |
|
4847 |
|
4848 // Used for JVMTI raw monitor implementation. |
|
4849 // All JavaThreads will enter here with state _thread_blocked |
|
4850 |
|
4851 int ObjectMonitor::raw_wait(jlong millis, bool interruptible, TRAPS) { |
|
4852 TEVENT (raw_wait) ; |
|
4853 if (THREAD != _owner) { |
|
4854 return OM_ILLEGAL_MONITOR_STATE; |
|
4855 } |
|
4856 |
|
4857 // To avoid spurious wakeups we reset the parkevent -- This is strictly optional. |
|
4858 // The caller must be able to tolerate spurious returns from raw_wait(). |
|
4859 THREAD->_ParkEvent->reset() ; |
|
4860 OrderAccess::fence() ; |
|
4861 |
|
4862 // check interrupt event |
|
4863 if (interruptible && Thread::is_interrupted(THREAD, true)) { |
|
4864 return OM_INTERRUPTED; |
|
4865 } |
|
4866 |
|
4867 intptr_t save = _recursions ; |
|
4868 _recursions = 0 ; |
|
4869 _waiters ++ ; |
|
4870 if (THREAD->is_Java_thread()) { |
|
4871 guarantee (((JavaThread *) THREAD)->thread_state() == _thread_blocked, "invariant") ; |
|
4872 ((JavaThread *)THREAD)->set_suspend_equivalent(); |
|
4873 } |
|
4874 int rv = SimpleWait (THREAD, millis) ; |
|
4875 _recursions = save ; |
|
4876 _waiters -- ; |
|
4877 |
|
4878 guarantee (THREAD == _owner, "invariant") ; |
|
4879 if (THREAD->is_Java_thread()) { |
|
4880 JavaThread * jSelf = (JavaThread *) THREAD ; |
|
4881 for (;;) { |
|
4882 if (!jSelf->handle_special_suspend_equivalent_condition()) break ; |
|
4883 SimpleExit (THREAD) ; |
|
4884 jSelf->java_suspend_self(); |
|
4885 SimpleEnter (THREAD) ; |
|
4886 jSelf->set_suspend_equivalent() ; |
|
4887 } |
|
4888 } |
|
4889 guarantee (THREAD == _owner, "invariant") ; |
|
4890 |
|
4891 if (interruptible && Thread::is_interrupted(THREAD, true)) { |
|
4892 return OM_INTERRUPTED; |
|
4893 } |
|
4894 return OM_OK ; |
|
4895 } |
|
4896 |
|
4897 int ObjectMonitor::raw_notify(TRAPS) { |
|
4898 TEVENT (raw_notify) ; |
|
4899 if (THREAD != _owner) { |
|
4900 return OM_ILLEGAL_MONITOR_STATE; |
|
4901 } |
|
4902 SimpleNotify (THREAD, false) ; |
|
4903 return OM_OK; |
|
4904 } |
|
4905 |
|
4906 int ObjectMonitor::raw_notifyAll(TRAPS) { |
|
4907 TEVENT (raw_notifyAll) ; |
|
4908 if (THREAD != _owner) { |
|
4909 return OM_ILLEGAL_MONITOR_STATE; |
|
4910 } |
|
4911 SimpleNotify (THREAD, true) ; |
|
4912 return OM_OK; |
|
4913 } |
|
4914 |
|
4915 #ifndef PRODUCT |
|
4916 void ObjectMonitor::verify() { |
|
4917 } |
|
4918 |
|
4919 void ObjectMonitor::print() { |
|
4920 } |
|
4921 #endif |
|
4922 |
1581 |
4923 //------------------------------------------------------------------------------ |
1582 //------------------------------------------------------------------------------ |
4924 // Non-product code |
1583 // Non-product code |
4925 |
1584 |
4926 #ifndef PRODUCT |
1585 #ifndef PRODUCT |