2 * Copyright (C) 2008 The Android Open Source Project
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
8 * http://www.apache.org/licenses/LICENSE-2.0
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
18 * Fundamental synchronization mechanisms.
20 * The top part of the file has operations on "monitor" structs; the
21 * next part has the native calls on objects.
23 * The current implementation uses "thin locking" to avoid allocating
24 * an Object's full Monitor struct until absolutely necessary (i.e.,
25 * during contention or a call to wait()).
27 * TODO: make improvements to thin locking
28 * We may be able to improve performance and reduce memory requirements by:
29 * - reverting to a thin lock once the Monitor is no longer necessary
30 * - using a pool of monitor objects, with some sort of recycling scheme
32 * TODO: recycle native-level monitors when objects are garbage collected.
45 #ifdef WITH_DEADLOCK_PREDICTION /* fwd */
46 static const char* kStartBanner =
47 "<-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#";
48 static const char* kEndBanner =
49 "#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#-#->";
52 * Unsorted, expanding list of objects.
54 * This is very similar to PointerSet (which came into existence after this),
55 * but these are unsorted, uniqueness is not enforced by the "add" function,
56 * and the base object isn't allocated on the heap.
58 typedef struct ExpandingObjectList {
62 } ExpandingObjectList;
65 static void updateDeadlockPrediction(Thread* self, Object* obj);
66 static void removeCollectedObject(Object* obj);
67 static void expandObjClear(ExpandingObjectList* pList);
71 * Every Object has a monitor associated with it, but not every Object is
72 * actually locked. Even the ones that are locked do not need a
73 * full-fledged monitor until a) there is actual contention or b) wait()
74 * is called on the Object.
76 * For Dalvik, we have implemented a scheme similar to the one described
77 * in Bacon et al.'s "Thin locks: featherweight synchronization for Java"
78 * (ACM 1998). Things are even easier for us, though, because we have
79 * a full 32 bits to work with.
81 * The two states of an Object's lock are referred to as "thin" and
82 * "fat". A lock may transition from the "thin" state to the "fat"
83 * state and this transition is referred to as inflation. Once a lock
84 * has been inflated it remains in the "fat" state indefinitely.
86 * The lock value itself is stored in Object.lock. The LSB of the
87 * lock encodes its state. When cleared, the lock is in the "thin"
88 * state and its bits are formatted as follows:
90 * [31 ---- 19] [18 ---- 3] [2 ---- 1] [0]
91 * lock count thread id hash state 0
93 * When set, the lock is in the "fat" state and its bits are formatted
96 * [31 ---- 3] [2 ---- 1] [0]
97 * pointer hash state 1
99 * For an in-depth description of the mechanics of thin-vs-fat locking,
100 * read the paper referred to above.
105 * - mutually exclusive access to resources
106 * - a way for multiple threads to wait for notification
108 * In effect, they fill the role of both mutexes and condition variables.
110 * Only one thread can own the monitor at any time. There may be several
111 * threads waiting on it (the wait call unlocks it). One or more waiting
112 * threads may be getting interrupted or notified at any given time.
115 Thread* owner; /* which thread currently owns the lock? */
116 int lockCount; /* owner's recursive lock depth */
117 Object* obj; /* what object are we part of [debug only] */
119 Thread* waitSet; /* threads currently waiting on this monitor */
121 pthread_mutex_t lock;
125 #ifdef WITH_DEADLOCK_PREDICTION
127 * Objects that have been locked immediately after this one in the
128 * past. We use an expanding flat array, allocated on first use, to
129 * minimize allocations. Deletions from the list, expected to be
130 * infrequent, are crunched down.
132 ExpandingObjectList historyChildren;
135 * We also track parents. This isn't strictly necessary, but it makes
136 * the cleanup at GC time significantly faster.
138 ExpandingObjectList historyParents;
140 /* used during cycle detection */
143 /* stack trace, established the first time we locked the object */
144 int historyStackDepth;
145 int* historyRawStackTrace;
151 * Create and initialize a monitor.
153 Monitor* dvmCreateMonitor(Object* obj)
157 mon = (Monitor*) calloc(1, sizeof(Monitor));
159 LOGE("Unable to allocate monitor\n");
162 if (((u4)mon & 7) != 0) {
163 LOGE("Misaligned monitor: %p\n", mon);
167 dvmInitMutex(&mon->lock);
169 /* replace the head of the list with the new monitor */
171 mon->next = gDvm.monitorList;
172 } while (!ATOMIC_CMP_SWAP((int32_t*)(void*)&gDvm.monitorList,
173 (int32_t)mon->next, (int32_t)mon));
179 * Free the monitor list. Only used when shutting the VM down.
181 void dvmFreeMonitorList(void)
186 mon = gDvm.monitorList;
187 while (mon != NULL) {
190 #ifdef WITH_DEADLOCK_PREDICTION
191 expandObjClear(&mon->historyChildren);
192 expandObjClear(&mon->historyParents);
193 free(mon->historyRawStackTrace);
201 * Log some info about our monitors.
203 void dvmDumpMonitorInfo(const char* msg)
205 #if QUIET_ZYGOTE_MONITOR
214 totalCount = liveCount = 0;
215 Monitor* mon = gDvm.monitorList;
216 while (mon != NULL) {
218 if (mon->obj != NULL)
223 LOGD("%s: monitor list has %d entries (%d live)\n",
224 msg, totalCount, liveCount);
228 * Get the object that a monitor is part of.
230 Object* dvmGetMonitorObject(Monitor* mon)
239 * Returns the thread id of the thread owning the given lock.
241 static u4 lockOwner(Object* obj)
248 * Since we're reading the lock value multiple times, latch it so
249 * that it doesn't change out from under us if we get preempted.
252 if (LW_SHAPE(lock) == LW_SHAPE_THIN) {
253 return LW_LOCK_OWNER(lock);
255 owner = LW_MONITOR(lock)->owner;
256 return owner ? owner->threadId : 0;
261 * Checks whether the given thread holds the given
264 bool dvmHoldsLock(Thread* thread, Object* obj)
266 if (thread == NULL || obj == NULL) {
269 return thread->threadId == lockOwner(obj);
274 * Free the monitor associated with an object and make the object's lock
275 * thin again. This is called during garbage collection.
277 static void freeObjectMonitor(Object* obj)
281 assert(LW_SHAPE(obj->lock) == LW_SHAPE_FAT);
283 #ifdef WITH_DEADLOCK_PREDICTION
284 if (gDvm.deadlockPredictMode != kDPOff)
285 removeCollectedObject(obj);
288 mon = LW_MONITOR(obj->lock);
289 obj->lock = DVM_LOCK_INITIAL_THIN_VALUE;
291 /* This lock is associated with an object
292 * that's being swept. The only possible way
293 * anyone could be holding this lock would be
294 * if some JNI code locked but didn't unlock
295 * the object, in which case we've got some bad
296 * native code somewhere.
298 assert(pthread_mutex_trylock(&mon->lock) == 0);
299 pthread_mutex_destroy(&mon->lock);
300 #ifdef WITH_DEADLOCK_PREDICTION
301 expandObjClear(&mon->historyChildren);
302 expandObjClear(&mon->historyParents);
303 free(mon->historyRawStackTrace);
309 * Frees monitor objects belonging to unmarked objects.
311 void dvmSweepMonitorList(Monitor** mon, int (*isUnmarkedObject)(void*))
314 Monitor *prev, *curr;
318 assert(*mon != NULL);
319 assert(isUnmarkedObject != NULL);
321 prev->next = curr = *mon;
322 while (curr != NULL) {
324 if (obj != NULL && (*isUnmarkedObject)(obj) != 0) {
325 prev->next = curr = curr->next;
326 freeObjectMonitor(obj);
338 static void lockMonitor(Thread* self, Monitor* mon)
342 if (mon->owner == self) {
345 ThreadStatus oldStatus;
347 if (pthread_mutex_trylock(&mon->lock) != 0) {
348 /* mutex is locked, switch to wait status and sleep on it */
349 oldStatus = dvmChangeStatus(self, THREAD_MONITOR);
350 cc = pthread_mutex_lock(&mon->lock);
352 dvmChangeStatus(self, oldStatus);
356 assert(mon->lockCount == 0);
361 * Try to lock a monitor.
363 * Returns "true" on success.
365 static bool tryLockMonitor(Thread* self, Monitor* mon)
369 if (mon->owner == self) {
373 cc = pthread_mutex_trylock(&mon->lock);
376 assert(mon->lockCount == 0);
388 * Returns true if the unlock succeeded.
389 * If the unlock failed, an exception will be pending.
391 static bool unlockMonitor(Thread* self, Monitor* mon)
393 assert(self != NULL);
394 assert(mon != NULL); // can this happen?
396 if (mon->owner == self) {
398 * We own the monitor, so nobody else can be in here.
400 if (mon->lockCount == 0) {
403 cc = pthread_mutex_unlock(&mon->lock);
410 * We don't own this, so we're not allowed to unlock it.
411 * The JNI spec says that we should throw IllegalMonitorStateException
414 dvmThrowExceptionFmt("Ljava/lang/IllegalMonitorStateException;",
415 "unlock of unowned monitor, self=%d owner=%d",
417 mon->owner ? mon->owner->threadId : 0);
424 * Checks the wait set for circular structure. Returns 0 if the list
425 * is not circular. Otherwise, returns 1. Used only by asserts.
427 static int waitSetCheck(Monitor *mon)
433 fast = slow = mon->waitSet;
436 if (fast == NULL) return 0;
437 if (fast->waitNext == NULL) return 0;
438 if (fast == slow && n > 0) return 1;
440 fast = fast->waitNext->waitNext;
441 slow = slow->waitNext;
446 * Links a thread into a monitor's wait set. The monitor lock must be
447 * held by the caller of this routine.
449 static void waitSetAppend(Monitor *mon, Thread *thread)
454 assert(mon->owner == dvmThreadSelf());
455 assert(thread != NULL);
456 assert(thread->waitNext == NULL);
457 assert(waitSetCheck(mon) == 0);
458 if (mon->waitSet == NULL) {
459 mon->waitSet = thread;
463 while (elt->waitNext != NULL) {
466 elt->waitNext = thread;
470 * Unlinks a thread from a monitor's wait set. The monitor lock must
471 * be held by the caller of this routine.
473 static void waitSetRemove(Monitor *mon, Thread *thread)
478 assert(mon->owner == dvmThreadSelf());
479 assert(thread != NULL);
480 assert(waitSetCheck(mon) == 0);
481 if (mon->waitSet == NULL) {
484 if (mon->waitSet == thread) {
485 mon->waitSet = thread->waitNext;
486 thread->waitNext = NULL;
490 while (elt->waitNext != NULL) {
491 if (elt->waitNext == thread) {
492 elt->waitNext = thread->waitNext;
493 thread->waitNext = NULL;
501 * Converts the given relative waiting time into an absolute time.
503 void absoluteTime(s8 msec, s4 nsec, struct timespec *ts)
507 #ifdef HAVE_TIMEDWAIT_MONOTONIC
508 clock_gettime(CLOCK_MONOTONIC, ts);
512 gettimeofday(&tv, NULL);
513 ts->tv_sec = tv.tv_sec;
514 ts->tv_nsec = tv.tv_usec * 1000;
517 endSec = ts->tv_sec + msec / 1000;
518 if (endSec >= 0x7fffffff) {
519 LOGV("NOTE: end time exceeds epoch\n");
523 ts->tv_nsec = (ts->tv_nsec + (msec % 1000) * 1000000) + nsec;
526 if (ts->tv_nsec >= 1000000000L) {
528 ts->tv_nsec -= 1000000000L;
532 int dvmRelativeCondWait(pthread_cond_t* cond, pthread_mutex_t* mutex,
537 absoluteTime(msec, nsec, &ts);
538 #if defined(HAVE_TIMEDWAIT_MONOTONIC)
539 ret = pthread_cond_timedwait_monotonic(cond, mutex, &ts);
541 ret = pthread_cond_timedwait(cond, mutex, &ts);
543 assert(ret == 0 || ret == ETIMEDOUT);
548 * Wait on a monitor until timeout, interrupt, or notification. Used for
549 * Object.wait() and (somewhat indirectly) Thread.sleep() and Thread.join().
551 * If another thread calls Thread.interrupt(), we throw InterruptedException
552 * and return immediately if one of the following are true:
553 * - blocked in wait(), wait(long), or wait(long, int) methods of Object
554 * - blocked in join(), join(long), or join(long, int) methods of Thread
555 * - blocked in sleep(long), or sleep(long, int) methods of Thread
556 * Otherwise, we set the "interrupted" flag.
558 * Checks to make sure that "nsec" is in the range 0-999999
559 * (i.e. fractions of a millisecond) and throws the appropriate
560 * exception if it isn't.
562 * The spec allows "spurious wakeups", and recommends that all code using
563 * Object.wait() do so in a loop. This appears to derive from concerns
564 * about pthread_cond_wait() on multiprocessor systems. Some commentary
565 * on the web casts doubt on whether these can/should occur.
567 * Since we're allowed to wake up "early", we clamp extremely long durations
568 * to return at the end of the 32-bit time epoch.
570 static void waitMonitor(Thread* self, Monitor* mon, s8 msec, s4 nsec,
571 bool interruptShouldThrow)
574 bool wasInterrupted = false;
578 assert(self != NULL);
581 /* Make sure that we hold the lock. */
582 if (mon->owner != self) {
583 dvmThrowException("Ljava/lang/IllegalMonitorStateException;",
584 "object not locked by thread before wait()");
589 * Enforce the timeout range.
591 if (msec < 0 || nsec < 0 || nsec > 999999) {
592 dvmThrowException("Ljava/lang/IllegalArgumentException;",
593 "timeout arguments out of range");
598 * Compute absolute wakeup time, if necessary.
600 if (msec == 0 && nsec == 0) {
603 absoluteTime(msec, nsec, &ts);
608 * Add ourselves to the set of threads waiting on this monitor, and
609 * release our hold. We need to let it go even if we're a few levels
610 * deep in a recursive lock, and we need to restore that later.
612 * We append to the wait set ahead of clearing the count and owner
613 * fields so the subroutine can check that the calling thread owns
614 * the monitor. Aside from that, the order of member updates is
615 * not order sensitive as we hold the pthread mutex.
617 waitSetAppend(mon, self);
618 int prevLockCount = mon->lockCount;
623 * Update thread status. If the GC wakes up, it'll ignore us, knowing
624 * that we won't touch any references in this state, and we'll check
625 * our suspend mode before we transition out.
628 dvmChangeStatus(self, THREAD_TIMED_WAIT);
630 dvmChangeStatus(self, THREAD_WAIT);
632 ret = pthread_mutex_lock(&self->waitMutex);
636 * Set waitMonitor to the monitor object we will be waiting on.
637 * When waitMonitor is non-NULL a notifying or interrupting thread
638 * must signal the thread's waitCond to wake it up.
640 assert(self->waitMonitor == NULL);
641 self->waitMonitor = mon;
644 * Handle the case where the thread was interrupted before we called
647 if (self->interrupted) {
648 wasInterrupted = true;
649 self->waitMonitor = NULL;
650 pthread_mutex_unlock(&self->waitMutex);
655 * Release the monitor lock and wait for a notification or
656 * a timeout to occur.
658 pthread_mutex_unlock(&mon->lock);
661 ret = pthread_cond_wait(&self->waitCond, &self->waitMutex);
664 #ifdef HAVE_TIMEDWAIT_MONOTONIC
665 ret = pthread_cond_timedwait_monotonic(&self->waitCond, &self->waitMutex, &ts);
667 ret = pthread_cond_timedwait(&self->waitCond, &self->waitMutex, &ts);
669 assert(ret == 0 || ret == ETIMEDOUT);
671 if (self->interrupted) {
672 wasInterrupted = true;
675 self->interrupted = false;
676 self->waitMonitor = NULL;
678 pthread_mutex_unlock(&self->waitMutex);
680 /* Reacquire the monitor lock. */
681 lockMonitor(self, mon);
685 * We remove our thread from wait set after restoring the count
686 * and owner fields so the subroutine can check that the calling
687 * thread owns the monitor. Aside from that, the order of member
688 * updates is not order sensitive as we hold the pthread mutex.
691 mon->lockCount = prevLockCount;
692 waitSetRemove(mon, self);
694 /* set self->status back to THREAD_RUNNING, and self-suspend if needed */
695 dvmChangeStatus(self, THREAD_RUNNING);
697 if (wasInterrupted) {
699 * We were interrupted while waiting, or somebody interrupted an
700 * un-interruptible thread earlier and we're bailing out immediately.
702 * The doc sayeth: "The interrupted status of the current thread is
703 * cleared when this exception is thrown."
705 self->interrupted = false;
706 if (interruptShouldThrow)
707 dvmThrowException("Ljava/lang/InterruptedException;", NULL);
712 * Notify one thread waiting on this monitor.
714 static void notifyMonitor(Thread* self, Monitor* mon)
718 assert(self != NULL);
721 /* Make sure that we hold the lock. */
722 if (mon->owner != self) {
723 dvmThrowException("Ljava/lang/IllegalMonitorStateException;",
724 "object not locked by thread before notify()");
727 /* Signal the first waiting thread in the wait set. */
728 while (mon->waitSet != NULL) {
729 thread = mon->waitSet;
730 mon->waitSet = thread->waitNext;
731 thread->waitNext = NULL;
732 pthread_mutex_lock(&thread->waitMutex);
733 /* Check to see if the thread is still waiting. */
734 if (thread->waitMonitor != NULL) {
735 pthread_cond_signal(&thread->waitCond);
736 pthread_mutex_unlock(&thread->waitMutex);
739 pthread_mutex_unlock(&thread->waitMutex);
744 * Notify all threads waiting on this monitor.
746 static void notifyAllMonitor(Thread* self, Monitor* mon)
750 assert(self != NULL);
753 /* Make sure that we hold the lock. */
754 if (mon->owner != self) {
755 dvmThrowException("Ljava/lang/IllegalMonitorStateException;",
756 "object not locked by thread before notifyAll()");
759 /* Signal all threads in the wait set. */
760 while (mon->waitSet != NULL) {
761 thread = mon->waitSet;
762 mon->waitSet = thread->waitNext;
763 thread->waitNext = NULL;
764 pthread_mutex_lock(&thread->waitMutex);
765 /* Check to see if the thread is still waiting. */
766 if (thread->waitMonitor != NULL) {
767 pthread_cond_signal(&thread->waitCond);
769 pthread_mutex_unlock(&thread->waitMutex);
774 * Implements monitorenter for "synchronized" stuff.
776 * This does not fail or throw an exception (unless deadlock prediction
777 * is enabled and set to "err" mode).
779 void dvmLockObject(Thread* self, Object *obj)
783 ThreadStatus oldStatus;
784 useconds_t sleepDelay;
785 const useconds_t maxSleepDelay = 1 << 20;
786 u4 thin, newThin, threadId;
788 assert(self != NULL);
790 threadId = self->threadId;
794 if (LW_SHAPE(thin) == LW_SHAPE_THIN) {
796 * The lock is a thin lock. The owner field is used to
797 * determine the acquire method, ordered by cost.
799 if (LW_LOCK_OWNER(thin) == threadId) {
801 * The calling thread owns the lock. Increment the
802 * value of the recursion count field.
804 obj->lock += 1 << LW_LOCK_COUNT_SHIFT;
805 } else if (LW_LOCK_OWNER(thin) == 0) {
807 * The lock is unowned. Install the thread id of the
808 * calling thread into the owner field. This is the
809 * common case. In performance critical code the JIT
810 * will have tried this before calling out to the VM.
812 newThin = thin | (threadId << LW_LOCK_OWNER_SHIFT);
813 if (!ATOMIC_CMP_SWAP((int32_t *)thinp, thin, newThin)) {
815 * The acquire failed. Try again.
820 LOG_THIN("(%d) spin on lock %p: %#x (%#x) %#x",
821 threadId, &obj->lock, 0, *thinp, thin);
823 * The lock is owned by another thread. Notify the VM
824 * that we are about to wait.
826 oldStatus = dvmChangeStatus(self, THREAD_MONITOR);
828 * Spin until the thin lock is released or inflated.
834 * Check the shape of the lock word. Another thread
835 * may have inflated the lock while we were waiting.
837 if (LW_SHAPE(thin) == LW_SHAPE_THIN) {
838 if (LW_LOCK_OWNER(thin) == 0) {
840 * The lock has been released. Install the
841 * thread id of the calling thread into the
844 newThin = thin | (threadId << LW_LOCK_OWNER_SHIFT);
845 if (ATOMIC_CMP_SWAP((int32_t *)thinp,
848 * The acquire succeed. Break out of the
849 * loop and proceed to inflate the lock.
855 * The lock has not been released. Yield so
856 * the owning thread can run.
858 if (sleepDelay == 0) {
863 if (sleepDelay < maxSleepDelay / 2) {
870 * The thin lock was inflated by another thread.
871 * Let the VM know we are no longer waiting and
874 LOG_THIN("(%d) lock %p surprise-fattened",
875 threadId, &obj->lock);
876 dvmChangeStatus(self, oldStatus);
880 LOG_THIN("(%d) spin on lock done %p: %#x (%#x) %#x",
881 threadId, &obj->lock, 0, *thinp, thin);
883 * We have acquired the thin lock. Let the VM know that
884 * we are no longer waiting.
886 dvmChangeStatus(self, oldStatus);
890 mon = dvmCreateMonitor(obj);
891 lockMonitor(self, mon);
893 thin &= LW_HASH_STATE_MASK << LW_HASH_STATE_SHIFT;
894 thin |= (u4)mon | LW_SHAPE_FAT;
897 LOG_THIN("(%d) lock %p fattened", threadId, &obj->lock);
901 * The lock is a fat lock.
903 assert(LW_MONITOR(obj->lock) != NULL);
904 lockMonitor(self, LW_MONITOR(obj->lock));
906 #ifdef WITH_DEADLOCK_PREDICTION
908 * See if we were allowed to grab the lock at this time. We do it
909 * *after* acquiring the lock, rather than before, so that we can
910 * freely update the Monitor struct. This seems counter-intuitive,
911 * but our goal is deadlock *prediction* not deadlock *prevention*.
912 * (If we actually deadlock, the situation is easy to diagnose from
913 * a thread dump, so there's no point making a special effort to do
914 * the checks before the lock is held.)
916 * This needs to happen before we add the object to the thread's
917 * monitor list, so we can tell the difference between first-lock and
920 * It's also important that we do this while in THREAD_RUNNING, so
921 * that we don't interfere with cleanup operations in the GC.
923 if (gDvm.deadlockPredictMode != kDPOff) {
924 if (self->status != THREAD_RUNNING) {
925 LOGE("Bad thread status (%d) in DP\n", self->status);
926 dvmDumpThread(self, false);
929 assert(!dvmCheckException(self));
930 updateDeadlockPrediction(self, obj);
931 if (dvmCheckException(self)) {
933 * If we're throwing an exception here, we need to free the
934 * lock. We add the object to the thread's monitor list so the
935 * "unlock" code can remove it.
937 dvmAddToMonitorList(self, obj, false);
938 dvmUnlockObject(self, obj);
939 LOGV("--- unlocked, pending is '%s'\n",
940 dvmGetException(self)->clazz->descriptor);
945 * Add the locked object, and the current stack trace, to the list
946 * held by the Thread object. If deadlock prediction isn't on,
947 * don't capture the stack trace.
949 dvmAddToMonitorList(self, obj, gDvm.deadlockPredictMode != kDPOff);
950 #elif defined(WITH_MONITOR_TRACKING)
952 * Add the locked object to the list held by the Thread object.
954 dvmAddToMonitorList(self, obj, false);
959 * Implements monitorexit for "synchronized" stuff.
961 * On failure, throws an exception and returns "false".
963 bool dvmUnlockObject(Thread* self, Object *obj)
967 assert(self != NULL);
968 assert(self->status == THREAD_RUNNING);
971 * Cache the lock word as its value can change while we are
972 * examining its state.
975 if (LW_SHAPE(thin) == LW_SHAPE_THIN) {
977 * The lock is thin. We must ensure that the lock is owned
978 * by the given thread before unlocking it.
980 if (LW_LOCK_OWNER(thin) == self->threadId) {
982 * We are the lock owner. It is safe to update the lock
983 * without CAS as lock ownership guards the lock itself.
985 if (LW_LOCK_COUNT(thin) == 0) {
987 * The lock was not recursively acquired, the common
988 * case. Unlock by clearing all bits except for the
991 obj->lock &= (LW_HASH_STATE_MASK << LW_HASH_STATE_SHIFT);
994 * The object was recursively acquired. Decrement the
995 * lock recursion count field.
997 obj->lock -= 1 << LW_LOCK_COUNT_SHIFT;
1001 * We do not own the lock. The JVM spec requires that we
1002 * throw an exception in this case.
1004 dvmThrowException("Ljava/lang/IllegalMonitorStateException;",
1005 "unlock of unowned monitor");
1010 * The lock is fat. We must check to see if unlockMonitor has
1011 * raised any exceptions before continuing.
1013 assert(LW_MONITOR(obj->lock) != NULL);
1014 if (!unlockMonitor(self, LW_MONITOR(obj->lock))) {
1016 * An exception has been raised. Do not fall through.
1022 #ifdef WITH_MONITOR_TRACKING
1024 * Remove the object from the Thread's list.
1026 dvmRemoveFromMonitorList(self, obj);
1033 * Object.wait(). Also called for class init.
1035 void dvmObjectWait(Thread* self, Object *obj, s8 msec, s4 nsec,
1036 bool interruptShouldThrow)
1038 Monitor* mon = LW_MONITOR(obj->lock);
1040 u4 thin = obj->lock;
1042 /* If the lock is still thin, we need to fatten it.
1044 if (LW_SHAPE(thin) == LW_SHAPE_THIN) {
1045 /* Make sure that 'self' holds the lock.
1047 if (LW_LOCK_OWNER(thin) != self->threadId) {
1048 dvmThrowException("Ljava/lang/IllegalMonitorStateException;",
1049 "object not locked by thread before wait()");
1053 /* This thread holds the lock. We need to fatten the lock
1054 * so 'self' can block on it. Don't update the object lock
1055 * field yet, because 'self' needs to acquire the lock before
1056 * any other thread gets a chance.
1058 mon = dvmCreateMonitor(obj);
1060 /* 'self' has actually locked the object one or more times;
1061 * make sure that the monitor reflects this.
1063 lockMonitor(self, mon);
1064 mon->lockCount = LW_LOCK_COUNT(thin);
1065 LOG_THIN("(%d) lock 0x%08x fattened by wait() to count %d\n",
1066 self->threadId, (uint)&obj->lock, mon->lockCount);
1069 /* Make the monitor public now that it's in the right state.
1071 thin &= LW_HASH_STATE_MASK << LW_HASH_STATE_SHIFT;
1072 thin |= (u4)mon | LW_SHAPE_FAT;
1077 waitMonitor(self, mon, msec, nsec, interruptShouldThrow);
1083 void dvmObjectNotify(Thread* self, Object *obj)
1085 u4 thin = obj->lock;
1087 /* If the lock is still thin, there aren't any waiters;
1088 * waiting on an object forces lock fattening.
1090 if (LW_SHAPE(thin) == LW_SHAPE_THIN) {
1091 /* Make sure that 'self' holds the lock.
1093 if (LW_LOCK_OWNER(thin) != self->threadId) {
1094 dvmThrowException("Ljava/lang/IllegalMonitorStateException;",
1095 "object not locked by thread before notify()");
1099 /* no-op; there are no waiters to notify.
1104 notifyMonitor(self, LW_MONITOR(thin));
1109 * Object.notifyAll().
1111 void dvmObjectNotifyAll(Thread* self, Object *obj)
1113 u4 thin = obj->lock;
1115 /* If the lock is still thin, there aren't any waiters;
1116 * waiting on an object forces lock fattening.
1118 if (LW_SHAPE(thin) == LW_SHAPE_THIN) {
1119 /* Make sure that 'self' holds the lock.
1121 if (LW_LOCK_OWNER(thin) != self->threadId) {
1122 dvmThrowException("Ljava/lang/IllegalMonitorStateException;",
1123 "object not locked by thread before notifyAll()");
1127 /* no-op; there are no waiters to notify.
1132 notifyAllMonitor(self, LW_MONITOR(thin));
1137 * This implements java.lang.Thread.sleep(long msec, int nsec).
1139 * The sleep is interruptible by other threads, which means we can't just
1140 * plop into an OS sleep call. (We probably could if we wanted to send
1141 * signals around and rely on EINTR, but that's inefficient and relies
1142 * on native code respecting our signal mask.)
1144 * We have to do all of this stuff for Object.wait() as well, so it's
1145 * easiest to just sleep on a private Monitor.
1147 * It appears that we want sleep(0,0) to go through the motions of sleeping
1148 * for a very short duration, rather than just returning.
1150 void dvmThreadSleep(u8 msec, u4 nsec)
1152 Thread* self = dvmThreadSelf();
1153 Monitor* mon = gDvm.threadSleepMon;
1155 /* sleep(0,0) wakes up immediately, wait(0,0) means wait forever; adjust */
1156 if (msec == 0 && nsec == 0)
1159 lockMonitor(self, mon);
1160 waitMonitor(self, mon, msec, nsec, true);
1161 unlockMonitor(self, mon);
1165 * Implement java.lang.Thread.interrupt().
1167 void dvmThreadInterrupt(Thread* thread)
1169 assert(thread != NULL);
1171 pthread_mutex_lock(&thread->waitMutex);
1174 * If the interrupted flag is already set no additional action is
1177 if (thread->interrupted == true) {
1178 pthread_mutex_unlock(&thread->waitMutex);
1183 * Raise the "interrupted" flag. This will cause it to bail early out
1184 * of the next wait() attempt, if it's not currently waiting on
1187 thread->interrupted = true;
1191 * Is the thread waiting?
1193 * Note that fat vs. thin doesn't matter here; waitMonitor
1194 * is only set when a thread actually waits on a monitor,
1195 * which implies that the monitor has already been fattened.
1197 if (thread->waitMonitor != NULL) {
1198 pthread_cond_signal(&thread->waitCond);
1201 pthread_mutex_unlock(&thread->waitMutex);
1204 #ifndef WITH_COPYING_GC
1205 u4 dvmIdentityHashCode(Object *obj)
1210 static size_t arrayElementWidth(const ArrayObject *array)
1212 const char *descriptor;
1214 if (dvmIsObjectArray(array)) {
1215 return sizeof(Object *);
1217 descriptor = array->obj.clazz->descriptor;
1218 switch (descriptor[1]) {
1219 case 'B': return 1; /* byte */
1220 case 'C': return 2; /* char */
1221 case 'D': return 8; /* double */
1222 case 'F': return 4; /* float */
1223 case 'I': return 4; /* int */
1224 case 'J': return 8; /* long */
1225 case 'S': return 2; /* short */
1226 case 'Z': return 1; /* boolean */
1229 LOGE("object %p has an unhandled descriptor '%s'", array, descriptor);
1230 dvmDumpThread(dvmThreadSelf(), false);
1232 return 0; /* Quiet the compiler. */
1235 static size_t arrayObjectLength(const ArrayObject *array)
1239 length = offsetof(ArrayObject, contents);
1240 length += array->length * arrayElementWidth(array);
1245 * Returns the identity hash code of the given object.
1247 u4 dvmIdentityHashCode(Object *obj)
1249 Thread *self, *thread;
1252 u4 lock, owner, hashState;
1256 * Null is defined to have an identity hash code of 0.
1262 hashState = LW_HASH_STATE(*lw);
1263 if (hashState == LW_HASH_STATE_HASHED) {
1265 * The object has been hashed but has not had its hash code
1266 * relocated by the garbage collector. Use the raw object
1269 return (u4)obj >> 3;
1270 } else if (hashState == LW_HASH_STATE_HASHED_AND_MOVED) {
1272 * The object has been hashed and its hash code has been
1273 * relocated by the collector. Use the value of the naturally
1274 * aligned word following the instance data.
1276 if (IS_CLASS_FLAG_SET(obj->clazz, CLASS_ISARRAY)) {
1277 length = arrayObjectLength((ArrayObject *)obj);
1278 length = (length + 3) & ~3;
1280 length = obj->clazz->objectSize;
1282 return *(u4 *)(((char *)obj) + length);
1283 } else if (hashState == LW_HASH_STATE_UNHASHED) {
1285 * The object has never been hashed. Change the hash state to
1286 * hashed and use the raw object address.
1288 self = dvmThreadSelf();
1289 if (self->threadId == lockOwner(obj)) {
1291 * We already own the lock so we can update the hash state
1294 *lw |= (LW_HASH_STATE_HASHED << LW_HASH_STATE_SHIFT);
1295 return (u4)obj >> 3;
1298 * We do not own the lock. Try acquiring the lock. Should
1299 * this fail, we must suspend the owning thread.
1301 if (LW_SHAPE(*lw) == LW_SHAPE_THIN) {
1303 * If the lock is thin assume it is unowned. We simulate
1304 * an acquire, update, and release with a single CAS.
1306 lock = DVM_LOCK_INITIAL_THIN_VALUE;
1307 lock |= (LW_HASH_STATE_HASHED << LW_HASH_STATE_SHIFT);
1308 if (ATOMIC_CMP_SWAP((int32_t *)lw,
1309 (int32_t)DVM_LOCK_INITIAL_THIN_VALUE,
1312 * A new lockword has been installed with a hash state
1313 * of hashed. Use the raw object address.
1315 return (u4)obj >> 3;
1318 if (tryLockMonitor(self, LW_MONITOR(*lw))) {
1320 * The monitor lock has been acquired. Change the
1321 * hash state to hashed and use the raw object
1324 *lw |= (LW_HASH_STATE_HASHED << LW_HASH_STATE_SHIFT);
1325 unlockMonitor(self, LW_MONITOR(*lw));
1326 return (u4)obj >> 3;
1330 * At this point we have failed to acquire the lock. We must
1331 * identify the owning thread and suspend it.
1333 dvmLockThreadList(self);
1335 * Cache the lock word as its value can change between
1336 * determining its shape and retrieving its owner.
1339 if (LW_SHAPE(lock) == LW_SHAPE_THIN) {
1341 * Find the thread with the corresponding thread id.
1343 owner = LW_LOCK_OWNER(lock);
1344 assert(owner != self->threadId);
1346 * If the lock has no owner do not bother scanning the
1347 * thread list and fall through to the failure handler.
1349 thread = owner ? gDvm.threadList : NULL;
1350 while (thread != NULL) {
1351 if (thread->threadId == owner) {
1354 thread = thread->next;
1357 thread = LW_MONITOR(lock)->owner;
1360 * If thread is NULL the object has been released since the
1361 * thread list lock was acquired. Try again.
1363 if (thread == NULL) {
1364 dvmUnlockThreadList();
1368 * Wait for the owning thread to suspend.
1370 dvmSuspendThread(thread);
1371 if (dvmHoldsLock(thread, obj)) {
1373 * The owning thread has been suspended. We can safely
1374 * change the hash state to hashed.
1376 *lw |= (LW_HASH_STATE_HASHED << LW_HASH_STATE_SHIFT);
1377 dvmResumeThread(thread);
1378 dvmUnlockThreadList();
1379 return (u4)obj >> 3;
1382 * The wrong thread has been suspended. Try again.
1384 dvmResumeThread(thread);
1385 dvmUnlockThreadList();
1388 LOGE("object %p has an unknown hash state %#x", obj, hashState);
1389 dvmDumpThread(dvmThreadSelf(), false);
1391 return 0; /* Quiet the compiler. */
1393 #endif /* WITH_COPYING_GC */
1395 #ifdef WITH_DEADLOCK_PREDICTION
1397 * ===========================================================================
1398 * Deadlock prediction
1399 * ===========================================================================
1402 The idea is to predict the possibility of deadlock by recording the order
1403 in which monitors are acquired. If we see an attempt to acquire a lock
1404 out of order, we can identify the locks and offending code.
1406 To make this work, we need to keep track of the locks held by each thread,
1407 and create history trees for each lock. When a thread tries to acquire
1408 a new lock, we walk through the "history children" of the lock, looking
1409 for a match with locks the thread already holds. If we find a match,
1410 it means the thread has made a request that could result in a deadlock.
1412 To support recursive locks, we always allow re-locking a currently-held
1413 lock, and maintain a recursion depth count.
1415 An ASCII-art example, where letters represent Objects:
1425 The above is the tree we'd have after handling Object synchronization
1426 sequences "ABC", "AC", "AD". A has three children, {B, C, D}. C is also
1427 a child of B. (The lines represent pointers between parent and child.
1428 Every node can have multiple parents and multiple children.)
1430 If we hold AC, and want to lock B, we recursively search through B's
1431 children to see if A or C appears. It does, so we reject the attempt.
1432 (A straightforward way to implement it: add a link from C to B, then
1433 determine whether the graph starting at B contains a cycle.)
1435 If we hold AC and want to lock D, we would succeed, creating a new link
1438 The lock history and a stack trace is attached to the Object's Monitor
1439 struct, which means we need to fatten every Object we lock (thin locking
1440 is effectively disabled). If we don't need the stack trace we can
1441 avoid fattening the leaf nodes, only fattening objects that need to hold
1444 Updates to Monitor structs are only allowed for the thread that holds
1445 the Monitor, so we actually do most of our deadlock prediction work after
1446 the lock has been acquired.
1448 When an object with a monitor is GCed, we need to remove it from the
1449 history trees. There are two basic approaches:
1450 (1) For through the entire set of known monitors, search all child
1451 lists for the object in question. This is rather slow, resulting
1452 in GC passes that take upwards of 10 seconds to complete.
1453 (2) Maintain "parent" pointers in each node. Remove the entries as
1454 required. This requires additional storage and maintenance for
1455 every operation, but is significantly faster at GC time.
1456 For each GCed object, we merge all of the object's children into each of
1457 the object's parents.
1460 #if !defined(WITH_MONITOR_TRACKING)
1461 # error "WITH_DEADLOCK_PREDICTION requires WITH_MONITOR_TRACKING"
1465 * Clear out the contents of an ExpandingObjectList, freeing any
1466 * dynamic allocations.
1468 static void expandObjClear(ExpandingObjectList* pList)
1470 if (pList->list != NULL) {
1474 pList->alloc = pList->count = 0;
1478 * Get the number of objects currently stored in the list.
1480 static inline int expandBufGetCount(const ExpandingObjectList* pList)
1482 return pList->count;
1486 * Get the Nth entry from the list.
1488 static inline Object* expandBufGetEntry(const ExpandingObjectList* pList,
1491 return pList->list[i];
1495 * Add a new entry to the list.
1497 * We don't check for or try to enforce uniqueness. It's expected that
1498 * the higher-level code does this for us.
1500 static void expandObjAddEntry(ExpandingObjectList* pList, Object* obj)
1502 if (pList->count == pList->alloc) {
1503 /* time to expand */
1506 if (pList->alloc == 0)
1510 LOGVV("expanding %p to %d\n", pList, pList->alloc);
1511 newList = realloc(pList->list, pList->alloc * sizeof(Object*));
1512 if (newList == NULL) {
1513 LOGE("Failed expanding DP object list (alloc=%d)\n", pList->alloc);
1516 pList->list = newList;
1519 pList->list[pList->count++] = obj;
1523 * Returns "true" if the element was successfully removed.
1525 static bool expandObjRemoveEntry(ExpandingObjectList* pList, Object* obj)
1529 for (i = pList->count-1; i >= 0; i--) {
1530 if (pList->list[i] == obj)
1536 if (i != pList->count-1) {
1538 * The order of elements is not important, so we just copy the
1539 * last entry into the new slot.
1541 //memmove(&pList->list[i], &pList->list[i+1],
1542 // (pList->count-1 - i) * sizeof(pList->list[0]));
1543 pList->list[i] = pList->list[pList->count-1];
1547 pList->list[pList->count] = (Object*) 0xdecadead;
1552 * Returns "true" if "obj" appears in the list.
1554 static bool expandObjHas(const ExpandingObjectList* pList, Object* obj)
1558 for (i = 0; i < pList->count; i++) {
1559 if (pList->list[i] == obj)
1566 * Print the list contents to stdout. For debugging.
1568 static void expandObjDump(const ExpandingObjectList* pList)
1571 for (i = 0; i < pList->count; i++)
1572 printf(" %p", pList->list[i]);
1576 * Check for duplicate entries. Returns the index of the first instance
1577 * of the duplicated value, or -1 if no duplicates were found.
1579 static int expandObjCheckForDuplicates(const ExpandingObjectList* pList)
1582 for (i = 0; i < pList->count-1; i++) {
1583 for (j = i + 1; j < pList->count; j++) {
1584 if (pList->list[i] == pList->list[j]) {
1595 * Determine whether "child" appears in the list of objects associated
1596 * with the Monitor in "parent". If "parent" is a thin lock, we return
1597 * false immediately.
1599 static bool objectInChildList(const Object* parent, Object* child)
1601 u4 lock = parent->lock;
1602 if (!IS_LOCK_FAT(&lock)) {
1603 //LOGI("on thin\n");
1607 return expandObjHas(&LW_MONITOR(lock)->historyChildren, child);
1611 * Print the child list.
1613 static void dumpKids(Object* parent)
1615 Monitor* mon = LW_MONITOR(parent->lock);
1617 printf("Children of %p:", parent);
1618 expandObjDump(&mon->historyChildren);
1623 * Add "child" to the list of children in "parent", and add "parent" to
1624 * the list of parents in "child".
1626 static void linkParentToChild(Object* parent, Object* child)
1628 //assert(LW_MONITOR(parent->lock)->owner == dvmThreadSelf()); // !owned for merge
1629 assert(IS_LOCK_FAT(&parent->lock));
1630 assert(IS_LOCK_FAT(&child->lock));
1631 assert(parent != child);
1634 mon = LW_MONITOR(parent->lock);
1635 assert(!expandObjHas(&mon->historyChildren, child));
1636 expandObjAddEntry(&mon->historyChildren, child);
1638 mon = LW_MONITOR(child->lock);
1639 assert(!expandObjHas(&mon->historyParents, parent));
1640 expandObjAddEntry(&mon->historyParents, parent);
1645 * Remove "child" from the list of children in "parent".
1647 static void unlinkParentFromChild(Object* parent, Object* child)
1649 //assert(LW_MONITOR(parent->lock)->owner == dvmThreadSelf()); // !owned for GC
1650 assert(IS_LOCK_FAT(&parent->lock));
1651 assert(IS_LOCK_FAT(&child->lock));
1652 assert(parent != child);
1655 mon = LW_MONITOR(parent->lock);
1656 if (!expandObjRemoveEntry(&mon->historyChildren, child)) {
1657 LOGW("WARNING: child %p not found in parent %p\n", child, parent);
1659 assert(!expandObjHas(&mon->historyChildren, child));
1660 assert(expandObjCheckForDuplicates(&mon->historyChildren) < 0);
1662 mon = LW_MONITOR(child->lock);
1663 if (!expandObjRemoveEntry(&mon->historyParents, parent)) {
1664 LOGW("WARNING: parent %p not found in child %p\n", parent, child);
1666 assert(!expandObjHas(&mon->historyParents, parent));
1667 assert(expandObjCheckForDuplicates(&mon->historyParents) < 0);
1672 * Log the monitors held by the current thread. This is done as part of
1673 * flagging an error.
1675 static void logHeldMonitors(Thread* self)
1679 name = dvmGetThreadName(self);
1680 LOGW("Monitors currently held by thread (threadid=%d '%s')\n",
1681 self->threadId, name);
1682 LOGW("(most-recently-acquired on top):\n");
1685 LockedObjectData* lod = self->pLockedObjects;
1686 while (lod != NULL) {
1687 LOGW("--- object %p[%d] (%s)\n",
1688 lod->obj, lod->recursionCount, lod->obj->clazz->descriptor);
1689 dvmLogRawStackTrace(lod->rawStackTrace, lod->stackDepth);
1696 * Recursively traverse the object hierarchy starting at "obj". We mark
1697 * ourselves on entry and clear the mark on exit. If we ever encounter
1698 * a marked object, we have a cycle.
1700 * Returns "true" if all is well, "false" if we found a cycle.
1702 static bool traverseTree(Thread* self, const Object* obj)
1704 assert(IS_LOCK_FAT(&obj->lock));
1705 Monitor* mon = LW_MONITOR(obj->lock);
1708 * Have we been here before?
1710 if (mon->historyMark) {
1714 LOGW("%s\n", kStartBanner);
1715 LOGW("Illegal lock attempt:\n");
1716 LOGW("--- object %p (%s)\n", obj, obj->clazz->descriptor);
1718 rawStackTrace = dvmFillInStackTraceRaw(self, &stackDepth);
1719 dvmLogRawStackTrace(rawStackTrace, stackDepth);
1720 free(rawStackTrace);
1723 logHeldMonitors(self);
1726 LOGW("Earlier, the following lock order (from last to first) was\n");
1727 LOGW("established -- stack trace is from first successful lock):\n");
1730 mon->historyMark = true;
1733 * Examine the children. We do NOT hold these locks, so they might
1734 * very well transition from thin to fat or change ownership while
1737 * NOTE: we rely on the fact that they cannot revert from fat to thin
1738 * while we work. This is currently a safe assumption.
1740 * We can safely ignore thin-locked children, because by definition
1741 * they have no history and are leaf nodes. In the current
1742 * implementation we always fatten the locks to provide a place to
1743 * hang the stack trace.
1745 ExpandingObjectList* pList = &mon->historyChildren;
1747 for (i = expandBufGetCount(pList)-1; i >= 0; i--) {
1748 const Object* child = expandBufGetEntry(pList, i);
1749 u4 lock = child->lock;
1750 if (!IS_LOCK_FAT(&lock))
1752 if (!traverseTree(self, child)) {
1753 LOGW("--- object %p (%s)\n", obj, obj->clazz->descriptor);
1754 dvmLogRawStackTrace(mon->historyRawStackTrace,
1755 mon->historyStackDepth);
1756 mon->historyMark = false;
1761 mon->historyMark = false;
1767 * Update the deadlock prediction tree, based on the current thread
1768 * acquiring "acqObj". This must be called before the object is added to
1769 * the thread's list of held monitors.
1771 * If the thread already holds the lock (recursion), or this is a known
1772 * lock configuration, we return without doing anything. Otherwise, we add
1773 * a link from the most-recently-acquired lock in this thread to "acqObj"
1774 * after ensuring that the parent lock is "fat".
1776 * This MUST NOT be called while a GC is in progress in another thread,
1777 * because we assume exclusive access to history trees in owned monitors.
1779 static void updateDeadlockPrediction(Thread* self, Object* acqObj)
1781 LockedObjectData* lod;
1782 LockedObjectData* mrl;
1785 * Quick check for recursive access.
1787 lod = dvmFindInMonitorList(self, acqObj);
1789 LOGV("+++ DP: recursive %p\n", acqObj);
1794 * Make the newly-acquired object's monitor "fat". In some ways this
1795 * isn't strictly necessary, but we need the GC to tell us when
1796 * "interesting" objects go away, and right now the only way to make
1797 * an object look interesting is to give it a monitor.
1799 * This also gives us a place to hang a stack trace.
1801 * Our thread holds the lock, so we're allowed to rewrite the lock
1802 * without worrying that something will change out from under us.
1804 if (!IS_LOCK_FAT(&acqObj->lock)) {
1805 LOGVV("fattening lockee %p (recur=%d)\n",
1806 acqObj, LW_LOCK_COUNT(acqObj->lock.thin));
1807 Monitor* newMon = dvmCreateMonitor(acqObj);
1808 lockMonitor(self, newMon); // can't stall, don't need VMWAIT
1809 newMon->lockCount += LW_LOCK_COUNT(acqObj->lock);
1810 u4 hashState = LW_HASH_STATE(acqObj->lock) << LW_HASH_STATE_SHIFT;
1811 acqObj->lock = (u4)newMon | hashState | LW_SHAPE_FAT;
1814 /* if we don't have a stack trace for this monitor, establish one */
1815 if (LW_MONITOR(acqObj->lock)->historyRawStackTrace == NULL) {
1816 Monitor* mon = LW_MONITOR(acqObj->lock);
1817 mon->historyRawStackTrace = dvmFillInStackTraceRaw(self,
1818 &mon->historyStackDepth);
1822 * We need to examine and perhaps modify the most-recently-locked
1823 * monitor. We own that, so there's no risk of another thread
1826 * Retrieve the most-recently-locked entry from our thread.
1828 mrl = self->pLockedObjects;
1830 return; /* no other locks held */
1833 * Do a quick check to see if "acqObj" is a direct descendant. We can do
1834 * this without holding the global lock because of our assertion that
1835 * a GC is not running in parallel -- nobody except the GC can
1836 * modify a history list in a Monitor they don't own, and we own "mrl".
1837 * (There might be concurrent *reads*, but no concurrent *writes.)
1839 * If we find it, this is a known good configuration, and we're done.
1841 if (objectInChildList(mrl->obj, acqObj))
1845 * "mrl" is going to need to have a history tree. If it's currently
1846 * a thin lock, we make it fat now. The thin lock might have a
1847 * nonzero recursive lock count, which we need to carry over.
1849 * Our thread holds the lock, so we're allowed to rewrite the lock
1850 * without worrying that something will change out from under us.
1852 if (!IS_LOCK_FAT(&mrl->obj->lock)) {
1853 LOGVV("fattening parent %p f/b/o child %p (recur=%d)\n",
1854 mrl->obj, acqObj, LW_LOCK_COUNT(mrl->obj->lock));
1855 Monitor* newMon = dvmCreateMonitor(mrl->obj);
1856 lockMonitor(self, newMon); // can't stall, don't need VMWAIT
1857 newMon->lockCount += LW_LOCK_COUNT(mrl->obj->lock);
1858 u4 hashState = LW_HASH_STATE(mrl->obj->lock) << LW_HASH_STATE_SHIFT;
1859 mrl->obj->lock = (u4)newMon | hashState | LW_SHAPE_FAT;
1863 * We haven't seen this configuration before. We need to scan down
1864 * acqObj's tree to see if any of the monitors in self->pLockedObjects
1865 * appear. We grab a global lock before traversing or updating the
1868 * If we find a match for any of our held locks, we know that the lock
1869 * has previously been acquired *after* acqObj, and we throw an error.
1871 * The easiest way to do this is to create a link from "mrl" to "acqObj"
1872 * and do a recursive traversal, marking nodes as we cross them. If
1873 * we cross one a second time, we have a cycle and can throw an error.
1874 * (We do the flag-clearing traversal before adding the new link, so
1875 * that we're guaranteed to terminate.)
1877 * If "acqObj" is a thin lock, it has no history, and we can create a
1878 * link to it without additional checks. [ We now guarantee that it's
1881 bool failed = false;
1882 dvmLockMutex(&gDvm.deadlockHistoryLock);
1883 linkParentToChild(mrl->obj, acqObj);
1884 if (!traverseTree(self, acqObj)) {
1885 LOGW("%s\n", kEndBanner);
1888 /* remove the entry so we're still okay when in "warning" mode */
1889 unlinkParentFromChild(mrl->obj, acqObj);
1891 dvmUnlockMutex(&gDvm.deadlockHistoryLock);
1894 switch (gDvm.deadlockPredictMode) {
1896 dvmThrowException("Ldalvik/system/PotentialDeadlockError;", NULL);
1899 LOGE("Aborting due to potential deadlock\n");
1910 * We're removing "child" from existence. We want to pull all of
1911 * child's children into "parent", filtering out duplicates. This is
1912 * called during the GC.
1914 * This does not modify "child", which might have multiple parents.
1916 static void mergeChildren(Object* parent, const Object* child)
1921 assert(IS_LOCK_FAT(&child->lock));
1922 mon = LW_MONITOR(child->lock);
1923 ExpandingObjectList* pList = &mon->historyChildren;
1925 for (i = expandBufGetCount(pList)-1; i >= 0; i--) {
1926 Object* grandChild = expandBufGetEntry(pList, i);
1928 if (!objectInChildList(parent, grandChild)) {
1929 LOGVV("+++ migrating %p link to %p\n", grandChild, parent);
1930 linkParentToChild(parent, grandChild);
1932 LOGVV("+++ parent %p already links to %p\n", parent, grandChild);
1938 * An object with a fat lock is being collected during a GC pass. We
1939 * want to remove it from any lock history trees that it is a part of.
1941 * This may require updating the history trees in several monitors. The
1942 * monitor semantics guarantee that no other thread will be accessing
1943 * the history trees at the same time.
1945 static void removeCollectedObject(Object* obj)
1949 LOGVV("+++ collecting %p\n", obj);
1953 * We're currently running through the entire set of known monitors.
1954 * This can be somewhat slow. We may want to keep lists of parents
1955 * in each child to speed up GC.
1957 mon = gDvm.monitorList;
1958 while (mon != NULL) {
1959 Object* parent = mon->obj;
1960 if (parent != NULL) { /* value nulled for deleted entries */
1961 if (objectInChildList(parent, obj)) {
1962 LOGVV("removing child %p from parent %p\n", obj, parent);
1963 unlinkParentFromChild(parent, obj);
1964 mergeChildren(parent, obj);
1972 * For every parent of this object:
1973 * - merge all of our children into the parent's child list (creates
1974 * a two-way link between parent and child)
1975 * - remove ourselves from the parent's child list
1977 ExpandingObjectList* pList;
1980 assert(IS_LOCK_FAT(&obj->lock));
1981 mon = LW_MONITOR(obj->lock);
1982 pList = &mon->historyParents;
1983 for (i = expandBufGetCount(pList)-1; i >= 0; i--) {
1984 Object* parent = expandBufGetEntry(pList, i);
1985 Monitor* parentMon = LW_MONITOR(parent->lock);
1987 if (!expandObjRemoveEntry(&parentMon->historyChildren, obj)) {
1988 LOGW("WARNING: child %p not found in parent %p\n", obj, parent);
1990 assert(!expandObjHas(&parentMon->historyChildren, obj));
1992 mergeChildren(parent, obj);
1996 * For every child of this object:
1997 * - remove ourselves from the child's parent list
1999 pList = &mon->historyChildren;
2000 for (i = expandBufGetCount(pList)-1; i >= 0; i--) {
2001 Object* child = expandBufGetEntry(pList, i);
2002 Monitor* childMon = LW_MONITOR(child->lock);
2004 if (!expandObjRemoveEntry(&childMon->historyParents, obj)) {
2005 LOGW("WARNING: parent %p not found in child %p\n", obj, child);
2007 assert(!expandObjHas(&childMon->historyParents, obj));
2011 #endif /*WITH_DEADLOCK_PREDICTION*/