/* This common inlined function is used to increment the counter of an
* errorcheck or recursive mutex.
*
* For errorcheck mutexes, it will return EDEADLK
* If the counter overflows, it will return EAGAIN
* Otherwise, it atomically increments the counter and returns 0
* after providing an acquire barrier.
*
* mtype is the current mutex type
* mvalue is the current mutex value (already loaded)
* mutex pointers to the mutex.
*/
static __inline__ __attribute__((always_inline)) int
_recursive_increment(pthread_mutex_t* mutex, int mvalue, int mtype)
{
if (mtype == MUTEX_TYPE_BITS_ERRORCHECK) {
/* trying to re-lock a mutex we already acquired */
return EDEADLK;
}
/* Detect recursive lock overflow and return EAGAIN.
* This is safe because only the owner thread can modify the
* counter bits in the mutex value.
*/
if (MUTEX_COUNTER_BITS_WILL_OVERFLOW(mvalue)) {
return EAGAIN;
}
/* We own the mutex, but other threads are able to change
* the lower bits (e.g. promoting it to "contended"), so we
* need to use an atomic cmpxchg loop to update the counter.
*/
for (;;) {
/* increment counter, overflow was already checked */
int newval = mvalue + MUTEX_COUNTER_BITS_ONE;
if (__predict_true(__bionic_cmpxchg(mvalue, newval, &mutex->value) == 0)) {
/* mutex is still locked, not need for a memory barrier */
return 0;
}
/* the value was changed, this happens when another thread changes
* the lower state bits from 1 to 2 to indicate contention. This
* cannot change the counter, so simply reload and try again.
*/
mvalue = mutex->value;
}
}
// This function is used by pthread_cond_broadcast and
// pthread_cond_signal to atomically decrement the counter
// then wake up 'counter' threads.
static int __pthread_cond_pulse(pthread_cond_t* cond, int counter) {
if (__predict_false(cond == NULL)) {
return EINVAL;
}
long flags = (cond->value & ~COND_COUNTER_MASK);
while (true) {
long old_value = cond->value;
long new_value = ((old_value - COND_COUNTER_INCREMENT) & COND_COUNTER_MASK) | flags;
if (__bionic_cmpxchg(old_value, new_value, &cond->value) == 0) {
break;
}
}
// Ensure that all memory accesses previously made by this thread are
// visible to the woken thread(s). On the other side, the "wait"
// code will issue any necessary barriers when locking the mutex.
//
// This may not strictly be necessary -- if the caller follows
// recommended practice and holds the mutex before signaling the cond
// var, the mutex ops will provide correct semantics. If they don't
// hold the mutex, they're subject to race conditions anyway.
ANDROID_MEMBAR_FULL();
__futex_wake_ex(&cond->value, COND_IS_SHARED(cond), counter);
return 0;
}
/*
* Lock a non-recursive mutex.
*
* As noted above, there are three states:
* 0 (unlocked, no contention)
* 1 (locked, no contention)
* 2 (locked, contention)
*
* Non-recursive mutexes don't use the thread-id or counter fields, and the
* "type" value is zero, so the only bits that will be set are the ones in
* the lock state field.
*/
static __inline__ void
_normal_lock(pthread_mutex_t* mutex, int shared)
{
/* convenience shortcuts */
const int unlocked = shared | MUTEX_STATE_BITS_UNLOCKED;
const int locked_uncontended = shared | MUTEX_STATE_BITS_LOCKED_UNCONTENDED;
/*
* The common case is an unlocked mutex, so we begin by trying to
* change the lock's state from 0 (UNLOCKED) to 1 (LOCKED).
* __bionic_cmpxchg() returns 0 if it made the swap successfully.
* If the result is nonzero, this lock is already held by another thread.
*/
if (__bionic_cmpxchg(unlocked, locked_uncontended, &mutex->value) != 0) {
const int locked_contended = shared | MUTEX_STATE_BITS_LOCKED_CONTENDED;
/*
* We want to go to sleep until the mutex is available, which
* requires promoting it to state 2 (CONTENDED). We need to
* swap in the new state value and then wait until somebody wakes us up.
*
* __bionic_swap() returns the previous value. We swap 2 in and
* see if we got zero back; if so, we have acquired the lock. If
* not, another thread still holds the lock and we wait again.
*
* The second argument to the __futex_wait() call is compared
* against the current value. If it doesn't match, __futex_wait()
* returns immediately (otherwise, it sleeps for a time specified
* by the third argument; 0 means sleep forever). This ensures
* that the mutex is in state 2 when we go to sleep on it, which
* guarantees a wake-up call.
*/
while (__bionic_swap(locked_contended, &mutex->value) != unlocked)
__futex_wait_ex(&mutex->value, shared, locked_contended, 0);
}
ANDROID_MEMBAR_FULL();
}
/* This function is used by pthread_cond_broadcast and
* pthread_cond_signal to atomically decrement the counter
* then wake-up 'counter' threads.
*/
static int
__pthread_cond_pulse(pthread_cond_t *cond, int counter)
{
long flags;
if (__unlikely(cond == NULL))
return EINVAL;
flags = (cond->value & ~COND_COUNTER_MASK);
for (;;) {
long oldval = cond->value;
long newval = ((oldval - COND_COUNTER_INCREMENT) & COND_COUNTER_MASK)
| flags;
if (__bionic_cmpxchg(oldval, newval, &cond->value) == 0)
break;
}
/*
* Ensure that all memory accesses previously made by this thread are
* visible to the woken thread(s). On the other side, the "wait"
* code will issue any necessary barriers when locking the mutex.
*
* This may not strictly be necessary -- if the caller follows
* recommended practice and holds the mutex before signaling the cond
* var, the mutex ops will provide correct semantics. If they don't
* hold the mutex, they're subject to race conditions anyway.
*/
ANDROID_MEMBAR_FULL();
__futex_wake_ex(&cond->value, COND_IS_SHARED(cond), counter);
return 0;
}
__LIBC_HIDDEN__
int pthread_mutex_trylock_impl(pthread_mutex_t *mutex)
{
int mvalue, mtype, tid, shared;
if (__unlikely(mutex == NULL))
return EINVAL;
mvalue = mutex->value;
mtype = (mvalue & MUTEX_TYPE_MASK);
shared = (mvalue & MUTEX_SHARED_MASK);
/* Handle common case first */
if ( __likely(mtype == MUTEX_TYPE_BITS_NORMAL) )
{
if (__bionic_cmpxchg(shared|MUTEX_STATE_BITS_UNLOCKED,
shared|MUTEX_STATE_BITS_LOCKED_UNCONTENDED,
&mutex->value) == 0) {
ANDROID_MEMBAR_FULL();
return 0;
}
return EBUSY;
}
/* Do we already own this recursive or error-check mutex ? */
tid = __get_thread()->tid;
if ( tid == MUTEX_OWNER_FROM_BITS(mvalue) )
return _recursive_increment(mutex, mvalue, mtype);
/* Same as pthread_mutex_lock, except that we don't want to wait, and
* the only operation that can succeed is a single cmpxchg to acquire the
* lock if it is released / not owned by anyone. No need for a complex loop.
*/
mtype |= shared | MUTEX_STATE_BITS_UNLOCKED;
mvalue = MUTEX_OWNER_TO_BITS(tid) | mtype | MUTEX_STATE_BITS_LOCKED_UNCONTENDED;
if (__likely(__bionic_cmpxchg(mtype, mvalue, &mutex->value) == 0)) {
ANDROID_MEMBAR_FULL();
return 0;
}
return EBUSY;
}
__LIBC_HIDDEN__
int pthread_mutex_unlock_impl(pthread_mutex_t *mutex)
{
int mvalue, mtype, tid, shared;
if (__unlikely(mutex == NULL))
return EINVAL;
mvalue = mutex->value;
mtype = (mvalue & MUTEX_TYPE_MASK);
shared = (mvalue & MUTEX_SHARED_MASK);
/* Handle common case first */
if (__likely(mtype == MUTEX_TYPE_BITS_NORMAL)) {
_normal_unlock(mutex, shared);
return 0;
}
/* Do we already own this recursive or error-check mutex ? */
tid = __get_thread()->tid;
if ( tid != MUTEX_OWNER_FROM_BITS(mvalue) )
return EPERM;
/* If the counter is > 0, we can simply decrement it atomically.
* Since other threads can mutate the lower state bits (and only the
* lower state bits), use a cmpxchg to do it.
*/
if (!MUTEX_COUNTER_BITS_IS_ZERO(mvalue)) {
for (;;) {
int newval = mvalue - MUTEX_COUNTER_BITS_ONE;
if (__likely(__bionic_cmpxchg(mvalue, newval, &mutex->value) == 0)) {
/* success: we still own the mutex, so no memory barrier */
return 0;
}
/* the value changed, so reload and loop */
mvalue = mutex->value;
}
}
/* the counter is 0, so we're going to unlock the mutex by resetting
* its value to 'unlocked'. We need to perform a swap in order
* to read the current state, which will be 2 if there are waiters
* to awake.
*
* TODO: Change this to __bionic_swap_release when we implement it
* to get rid of the explicit memory barrier below.
*/
ANDROID_MEMBAR_FULL(); /* RELEASE BARRIER */
mvalue = __bionic_swap(mtype | shared | MUTEX_STATE_BITS_UNLOCKED, &mutex->value);
/* Wake one waiting thread, if any */
if (MUTEX_STATE_BITS_IS_LOCKED_CONTENDED(mvalue)) {
__futex_wake_ex(&mutex->value, shared, 1);
}
return 0;
}
extern "C" int __cxa_guard_acquire(_guard_t* gv)
{
// 0 -> pending, return 1
// pending -> waiting, wait and return 0
// waiting: untouched, wait and return 0
// ready: untouched, return 0
retry:
if (__bionic_cmpxchg(0, pending, &gv->state) == 0) {
ANDROID_MEMBAR_FULL();
return 1;
}
__bionic_cmpxchg(pending, waiting, &gv->state); // Indicate there is a waiter
__futex_wait(&gv->state, waiting, NULL);
if (gv->state != ready) // __cxa_guard_abort was called, let every thread try since there is no return code for this condition
goto retry;
ANDROID_MEMBAR_FULL();
return 0;
}
extern "C" void __cxa_guard_release(_guard_t* gv)
{
// pending -> ready
// waiting -> ready, and wake
ANDROID_MEMBAR_FULL();
if (__bionic_cmpxchg(pending, ready, &gv->state) == 0) {
return;
}
gv->state = ready;
__futex_wake(&gv->state, 0x7fffffff);
}
static int __pthread_mutex_timedlock(pthread_mutex_t* mutex, const timespec* abs_timeout, clockid_t clock) {
timespec ts;
int mvalue = mutex->value;
int mtype = (mvalue & MUTEX_TYPE_MASK);
int shared = (mvalue & MUTEX_SHARED_MASK);
// Handle common case first.
if (__predict_true(mtype == MUTEX_TYPE_BITS_NORMAL)) {
const int unlocked = shared | MUTEX_STATE_BITS_UNLOCKED;
const int locked_uncontended = shared | MUTEX_STATE_BITS_LOCKED_UNCONTENDED;
const int locked_contended = shared | MUTEX_STATE_BITS_LOCKED_CONTENDED;
// Fast path for uncontended lock. Note: MUTEX_TYPE_BITS_NORMAL is 0.
if (__bionic_cmpxchg(unlocked, locked_uncontended, &mutex->value) == 0) {
ANDROID_MEMBAR_FULL();
return 0;
}
// Loop while needed.
while (__bionic_swap(locked_contended, &mutex->value) != unlocked) {
if (__timespec_from_absolute(&ts, abs_timeout, clock) < 0) {
return ETIMEDOUT;
}
__futex_wait_ex(&mutex->value, shared, locked_contended, &ts);
}
ANDROID_MEMBAR_FULL();
return 0;
}
// Do we already own this recursive or error-check mutex?
pid_t tid = __get_thread()->tid;
if (tid == MUTEX_OWNER_FROM_BITS(mvalue)) {
return _recursive_increment(mutex, mvalue, mtype);
}
// The following implements the same loop as pthread_mutex_lock_impl
// but adds checks to ensure that the operation never exceeds the
// absolute expiration time.
mtype |= shared;
// First try a quick lock.
if (mvalue == mtype) {
mvalue = MUTEX_OWNER_TO_BITS(tid) | mtype | MUTEX_STATE_BITS_LOCKED_UNCONTENDED;
if (__predict_true(__bionic_cmpxchg(mtype, mvalue, &mutex->value) == 0)) {
ANDROID_MEMBAR_FULL();
return 0;
}
mvalue = mutex->value;
}
while (true) {
// If the value is 'unlocked', try to acquire it directly.
// NOTE: put state to 2 since we know there is contention.
if (mvalue == mtype) { // Unlocked.
mvalue = MUTEX_OWNER_TO_BITS(tid) | mtype | MUTEX_STATE_BITS_LOCKED_CONTENDED;
if (__bionic_cmpxchg(mtype, mvalue, &mutex->value) == 0) {
ANDROID_MEMBAR_FULL();
return 0;
}
// The value changed before we could lock it. We need to check
// the time to avoid livelocks, reload the value, then loop again.
if (__timespec_from_absolute(&ts, abs_timeout, clock) < 0) {
return ETIMEDOUT;
}
mvalue = mutex->value;
continue;
}
// The value is locked. If 'uncontended', try to switch its state
// to 'contented' to ensure we get woken up later.
if (MUTEX_STATE_BITS_IS_LOCKED_UNCONTENDED(mvalue)) {
int newval = MUTEX_STATE_BITS_FLIP_CONTENTION(mvalue);
if (__bionic_cmpxchg(mvalue, newval, &mutex->value) != 0) {
// This failed because the value changed, reload it.
mvalue = mutex->value;
} else {
// This succeeded, update mvalue.
mvalue = newval;
}
}
// Check time and update 'ts'.
if (__timespec_from_absolute(&ts, abs_timeout, clock) < 0) {
return ETIMEDOUT;
}
// Only wait to be woken up if the state is '2', otherwise we'll
// simply loop right now. This can happen when the second cmpxchg
// in our loop failed because the mutex was unlocked by another thread.
if (MUTEX_STATE_BITS_IS_LOCKED_CONTENDED(mvalue)) {
if (__futex_wait_ex(&mutex->value, shared, mvalue, &ts) == -ETIMEDOUT) {
return ETIMEDOUT;
}
mvalue = mutex->value;
}
}
/* NOTREACHED */
}
__LIBC_HIDDEN__
int pthread_mutex_lock_impl(pthread_mutex_t *mutex)
{
int mvalue, mtype, tid, shared;
mvalue = mutex->value;
mtype = (mvalue & MUTEX_TYPE_MASK);
shared = (mvalue & MUTEX_SHARED_MASK);
/* Handle normal case first */
if ( __predict_true(mtype == MUTEX_TYPE_BITS_NORMAL) ) {
_normal_lock(mutex, shared);
return 0;
}
/* Do we already own this recursive or error-check mutex ? */
tid = __get_thread()->tid;
if ( tid == MUTEX_OWNER_FROM_BITS(mvalue) )
return _recursive_increment(mutex, mvalue, mtype);
/* Add in shared state to avoid extra 'or' operations below */
mtype |= shared;
/* First, if the mutex is unlocked, try to quickly acquire it.
* In the optimistic case where this works, set the state to 1 to
* indicate locked with no contention */
if (mvalue == mtype) {
int newval = MUTEX_OWNER_TO_BITS(tid) | mtype | MUTEX_STATE_BITS_LOCKED_UNCONTENDED;
if (__bionic_cmpxchg(mvalue, newval, &mutex->value) == 0) {
ANDROID_MEMBAR_FULL();
return 0;
}
/* argh, the value changed, reload before entering the loop */
mvalue = mutex->value;
}
for (;;) {
int newval;
/* if the mutex is unlocked, its value should be 'mtype' and
* we try to acquire it by setting its owner and state atomically.
* NOTE: We put the state to 2 since we _know_ there is contention
* when we are in this loop. This ensures all waiters will be
* unlocked.
*/
if (mvalue == mtype) {
newval = MUTEX_OWNER_TO_BITS(tid) | mtype | MUTEX_STATE_BITS_LOCKED_CONTENDED;
/* TODO: Change this to __bionic_cmpxchg_acquire when we
* implement it to get rid of the explicit memory
* barrier below.
*/
if (__predict_false(__bionic_cmpxchg(mvalue, newval, &mutex->value) != 0)) {
mvalue = mutex->value;
continue;
}
ANDROID_MEMBAR_FULL();
return 0;
}
/* the mutex is already locked by another thread, if its state is 1
* we will change it to 2 to indicate contention. */
if (MUTEX_STATE_BITS_IS_LOCKED_UNCONTENDED(mvalue)) {
newval = MUTEX_STATE_BITS_FLIP_CONTENTION(mvalue); /* locked state 1 => state 2 */
if (__predict_false(__bionic_cmpxchg(mvalue, newval, &mutex->value) != 0)) {
mvalue = mutex->value;
continue;
}
mvalue = newval;
}
/* wait until the mutex is unlocked */
__futex_wait_ex(&mutex->value, shared, mvalue, NULL);
mvalue = mutex->value;
}
/* NOTREACHED */
}
__LIBC_HIDDEN__
int pthread_mutex_lock_timeout_np_impl(pthread_mutex_t *mutex, unsigned msecs)
{
clockid_t clock = CLOCK_MONOTONIC;
struct timespec abstime;
struct timespec ts;
int mvalue, mtype, tid, shared;
/* compute absolute expiration time */
__timespec_to_relative_msec(&abstime, msecs, clock);
if (__unlikely(mutex == NULL))
return EINVAL;
mvalue = mutex->value;
mtype = (mvalue & MUTEX_TYPE_MASK);
shared = (mvalue & MUTEX_SHARED_MASK);
/* Handle common case first */
if ( __likely(mtype == MUTEX_TYPE_BITS_NORMAL) )
{
const int unlocked = shared | MUTEX_STATE_BITS_UNLOCKED;
const int locked_uncontended = shared | MUTEX_STATE_BITS_LOCKED_UNCONTENDED;
const int locked_contended = shared | MUTEX_STATE_BITS_LOCKED_CONTENDED;
/* fast path for uncontended lock. Note: MUTEX_TYPE_BITS_NORMAL is 0 */
if (__bionic_cmpxchg(unlocked, locked_uncontended, &mutex->value) == 0) {
ANDROID_MEMBAR_FULL();
return 0;
}
/* loop while needed */
while (__bionic_swap(locked_contended, &mutex->value) != unlocked) {
if (__timespec_to_absolute(&ts, &abstime, clock) < 0)
return EBUSY;
__futex_wait_ex(&mutex->value, shared, locked_contended, &ts);
}
ANDROID_MEMBAR_FULL();
return 0;
}
/* Do we already own this recursive or error-check mutex ? */
tid = __get_thread()->tid;
if ( tid == MUTEX_OWNER_FROM_BITS(mvalue) )
return _recursive_increment(mutex, mvalue, mtype);
/* the following implements the same loop than pthread_mutex_lock_impl
* but adds checks to ensure that the operation never exceeds the
* absolute expiration time.
*/
mtype |= shared;
/* first try a quick lock */
if (mvalue == mtype) {
mvalue = MUTEX_OWNER_TO_BITS(tid) | mtype | MUTEX_STATE_BITS_LOCKED_UNCONTENDED;
if (__likely(__bionic_cmpxchg(mtype, mvalue, &mutex->value) == 0)) {
ANDROID_MEMBAR_FULL();
return 0;
}
mvalue = mutex->value;
}
for (;;) {
struct timespec ts;
/* if the value is 'unlocked', try to acquire it directly */
/* NOTE: put state to 2 since we know there is contention */
if (mvalue == mtype) { /* unlocked */
mvalue = MUTEX_OWNER_TO_BITS(tid) | mtype | MUTEX_STATE_BITS_LOCKED_CONTENDED;
if (__bionic_cmpxchg(mtype, mvalue, &mutex->value) == 0) {
ANDROID_MEMBAR_FULL();
return 0;
}
/* the value changed before we could lock it. We need to check
* the time to avoid livelocks, reload the value, then loop again. */
if (__timespec_to_absolute(&ts, &abstime, clock) < 0)
return EBUSY;
mvalue = mutex->value;
continue;
}
/* The value is locked. If 'uncontended', try to switch its state
* to 'contented' to ensure we get woken up later. */
if (MUTEX_STATE_BITS_IS_LOCKED_UNCONTENDED(mvalue)) {
int newval = MUTEX_STATE_BITS_FLIP_CONTENTION(mvalue);
if (__bionic_cmpxchg(mvalue, newval, &mutex->value) != 0) {
/* this failed because the value changed, reload it */
mvalue = mutex->value;
} else {
/* this succeeded, update mvalue */
mvalue = newval;
}
}
/* check time and update 'ts' */
if (__timespec_to_absolute(&ts, &abstime, clock) < 0)
return EBUSY;
/* Only wait to be woken up if the state is '2', otherwise we'll
* simply loop right now. This can happen when the second cmpxchg
* in our loop failed because the mutex was unlocked by another
* thread.
*/
if (MUTEX_STATE_BITS_IS_LOCKED_CONTENDED(mvalue)) {
if (__futex_wait_ex(&mutex->value, shared, mvalue, &ts) == ETIMEDOUT) {
return EBUSY;
}
mvalue = mutex->value;
}
}
/* NOTREACHED */
}
/* NOTE: this implementation doesn't support a init function that throws a C++ exception
* or calls fork()
*/
int pthread_once( pthread_once_t* once_control, void (*init_routine)(void) )
{
volatile pthread_once_t* ocptr = once_control;
/* PTHREAD_ONCE_INIT is 0, we use the following bit flags
*
* bit 0 set -> initialization is under way
* bit 1 set -> initialization is complete
*/
#define ONCE_INITIALIZING (1 << 0)
#define ONCE_COMPLETED (1 << 1)
/* First check if the once is already initialized. This will be the common
* case and we want to make this as fast as possible. Note that this still
* requires a load_acquire operation here to ensure that all the
* stores performed by the initialization function are observable on
* this CPU after we exit.
*/
if (__likely((*ocptr & ONCE_COMPLETED) != 0)) {
ANDROID_MEMBAR_FULL();
return 0;
}
for (;;) {
/* Try to atomically set the INITIALIZING flag.
* This requires a cmpxchg loop, and we may need
* to exit prematurely if we detect that
* COMPLETED is now set.
*/
int32_t oldval, newval;
do {
oldval = *ocptr;
if ((oldval & ONCE_COMPLETED) != 0)
break;
newval = oldval | ONCE_INITIALIZING;
} while (__bionic_cmpxchg(oldval, newval, ocptr) != 0);
if ((oldval & ONCE_COMPLETED) != 0) {
/* We detected that COMPLETED was set while in our loop */
ANDROID_MEMBAR_FULL();
return 0;
}
if ((oldval & ONCE_INITIALIZING) == 0) {
/* We got there first, we can jump out of the loop to
* handle the initialization */
break;
}
/* Another thread is running the initialization and hasn't completed
* yet, so wait for it, then try again. */
__futex_wait_ex(ocptr, 0, oldval, NULL);
}
/* call the initialization function. */
(*init_routine)();
/* Do a store_release indicating that initialization is complete */
ANDROID_MEMBAR_FULL();
*ocptr = ONCE_COMPLETED;
/* Wake up any waiters, if any */
__futex_wake_ex(ocptr, 0, INT_MAX);
return 0;
}
