/*
* Downgrade an unrecursed exclusive lock into a single shared lock.
*/
void
_sx_downgrade(struct sx *sx, const char *file, int line)
{
uintptr_t x;
int wakeup_swapper;
KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
("sx_downgrade() of destroyed sx @ %s:%d", file, line));
_sx_assert(sx, SA_XLOCKED | SA_NOTRECURSED, file, line);
#ifndef INVARIANTS
if (sx_recursed(sx))
panic("downgrade of a recursed lock");
#endif
WITNESS_DOWNGRADE(&sx->lock_object, 0, file, line);
/*
* Try to switch from an exclusive lock with no shared waiters
* to one sharer with no shared waiters. If there are
* exclusive waiters, we don't need to lock the sleep queue so
* long as we preserve the flag. We do one quick try and if
* that fails we grab the sleepq lock to keep the flags from
* changing and do it the slow way.
*
* We have to lock the sleep queue if there are shared waiters
* so we can wake them up.
*/
x = sx->sx_lock;
if (!(x & SX_LOCK_SHARED_WAITERS) &&
atomic_cmpset_rel_ptr(&sx->sx_lock, x, SX_SHARERS_LOCK(1) |
(x & SX_LOCK_EXCLUSIVE_WAITERS))) {
LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
return;
}
/*
* Lock the sleep queue so we can read the waiters bits
* without any races and wakeup any shared waiters.
*/
sleepq_lock(&sx->lock_object);
/*
* Preserve SX_LOCK_EXCLUSIVE_WAITERS while downgraded to a single
* shared lock. If there are any shared waiters, wake them up.
*/
wakeup_swapper = 0;
x = sx->sx_lock;
atomic_store_rel_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) |
(x & SX_LOCK_EXCLUSIVE_WAITERS));
if (x & SX_LOCK_SHARED_WAITERS)
wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
0, SQ_SHARED_QUEUE);
sleepq_release(&sx->lock_object);
LOCK_LOG_LOCK("XDOWNGRADE", &sx->lock_object, 0, 0, file, line);
LOCKSTAT_RECORD0(LS_SX_DOWNGRADE_DOWNGRADE, sx);
if (wakeup_swapper)
kick_proc0();
}
/*
* This function represents the so-called 'hard case' for sx_xunlock
* operation. All 'easy case' failures are redirected to this. Note
* that ideally this would be a static function, but it needs to be
* accessible from at least sx.h.
*/
void
_sx_xunlock_hard(struct sx *sx, uintptr_t tid, const char *file, int line)
{
uintptr_t x;
int queue, wakeup_swapper;
if (SCHEDULER_STOPPED())
return;
MPASS(!(sx->sx_lock & SX_LOCK_SHARED));
/* If the lock is recursed, then unrecurse one level. */
if (sx_xlocked(sx) && sx_recursed(sx)) {
if ((--sx->sx_recurse) == 0)
atomic_clear_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, sx);
return;
}
MPASS(sx->sx_lock & (SX_LOCK_SHARED_WAITERS |
SX_LOCK_EXCLUSIVE_WAITERS));
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p contested", __func__, sx);
sleepq_lock(&sx->lock_object);
x = SX_LOCK_UNLOCKED;
/*
* The wake up algorithm here is quite simple and probably not
* ideal. It gives precedence to shared waiters if they are
* present. For this condition, we have to preserve the
* state of the exclusive waiters flag.
* If interruptible sleeps left the shared queue empty avoid a
* starvation for the threads sleeping on the exclusive queue by giving
* them precedence and cleaning up the shared waiters bit anyway.
*/
if ((sx->sx_lock & SX_LOCK_SHARED_WAITERS) != 0 &&
sleepq_sleepcnt(&sx->lock_object, SQ_SHARED_QUEUE) != 0) {
queue = SQ_SHARED_QUEUE;
x |= (sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS);
} else
queue = SQ_EXCLUSIVE_QUEUE;
/* Wake up all the waiters for the specific queue. */
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR3(KTR_LOCK, "%s: %p waking up all threads on %s queue",
__func__, sx, queue == SQ_SHARED_QUEUE ? "shared" :
"exclusive");
atomic_store_rel_ptr(&sx->sx_lock, x);
wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX, 0,
queue);
sleepq_release(&sx->lock_object);
if (wakeup_swapper)
kick_proc0();
}
/*
* This function represents the so-called 'hard case' for sx_slock
* operation. All 'easy case' failures are redirected to this. Note
* that ideally this would be a static function, but it needs to be
* accessible from at least sx.h.
*/
int
_sx_slock_hard(struct sx *sx, int opts, const char *file, int line)
{
GIANT_DECLARE;
#ifdef ADAPTIVE_SX
volatile struct thread *owner;
#endif
#ifdef LOCK_PROFILING
uint64_t waittime = 0;
int contested = 0;
#endif
uintptr_t x;
int error = 0;
#ifdef KDTRACE_HOOKS
uintptr_t state;
uint64_t spin_cnt = 0;
uint64_t sleep_cnt = 0;
int64_t sleep_time = 0;
int64_t all_time = 0;
#endif
if (SCHEDULER_STOPPED())
return (0);
#ifdef KDTRACE_HOOKS
state = sx->sx_lock;
all_time -= lockstat_nsecs(&sx->lock_object);
#endif
/*
* As with rwlocks, we don't make any attempt to try to block
* shared locks once there is an exclusive waiter.
*/
for (;;) {
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
x = sx->sx_lock;
/*
* If no other thread has an exclusive lock then try to bump up
* the count of sharers. Since we have to preserve the state
* of SX_LOCK_EXCLUSIVE_WAITERS, if we fail to acquire the
* shared lock loop back and retry.
*/
if (x & SX_LOCK_SHARED) {
MPASS(!(x & SX_LOCK_SHARED_WAITERS));
if (atomic_cmpset_acq_ptr(&sx->sx_lock, x,
x + SX_ONE_SHARER)) {
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR4(KTR_LOCK,
"%s: %p succeed %p -> %p", __func__,
sx, (void *)x,
(void *)(x + SX_ONE_SHARER));
break;
}
continue;
}
#ifdef HWPMC_HOOKS
PMC_SOFT_CALL( , , lock, failed);
#endif
lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
&waittime);
#ifdef ADAPTIVE_SX
/*
* If the owner is running on another CPU, spin until
* the owner stops running or the state of the lock
* changes.
*/
if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
x = SX_OWNER(x);
owner = (struct thread *)x;
if (TD_IS_RUNNING(owner)) {
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR3(KTR_LOCK,
"%s: spinning on %p held by %p",
__func__, sx, owner);
KTR_STATE1(KTR_SCHED, "thread",
sched_tdname(curthread), "spinning",
"lockname:\"%s\"", sx->lock_object.lo_name);
GIANT_SAVE();
while (SX_OWNER(sx->sx_lock) == x &&
TD_IS_RUNNING(owner)) {
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
cpu_spinwait();
}
KTR_STATE0(KTR_SCHED, "thread",
sched_tdname(curthread), "running");
continue;
}
}
#endif
/*
* Some other thread already has an exclusive lock, so
* start the process of blocking.
*/
sleepq_lock(&sx->lock_object);
x = sx->sx_lock;
/*
* The lock could have been released while we spun.
* In this case loop back and retry.
*/
if (x & SX_LOCK_SHARED) {
sleepq_release(&sx->lock_object);
continue;
}
#ifdef ADAPTIVE_SX
/*
* If the owner is running on another CPU, spin until
* the owner stops running or the state of the lock
* changes.
*/
if (!(x & SX_LOCK_SHARED) &&
(sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
owner = (struct thread *)SX_OWNER(x);
if (TD_IS_RUNNING(owner)) {
sleepq_release(&sx->lock_object);
continue;
}
}
#endif
/*
* Try to set the SX_LOCK_SHARED_WAITERS flag. If we
* fail to set it drop the sleep queue lock and loop
* back.
*/
if (!(x & SX_LOCK_SHARED_WAITERS)) {
if (!atomic_cmpset_ptr(&sx->sx_lock, x,
x | SX_LOCK_SHARED_WAITERS)) {
sleepq_release(&sx->lock_object);
continue;
}
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p set shared waiters flag",
__func__, sx);
}
/*
* Since we have been unable to acquire the shared lock,
* we have to sleep.
*/
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
__func__, sx);
#ifdef KDTRACE_HOOKS
sleep_time -= lockstat_nsecs(&sx->lock_object);
#endif
GIANT_SAVE();
sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
SLEEPQ_INTERRUPTIBLE : 0), SQ_SHARED_QUEUE);
if (!(opts & SX_INTERRUPTIBLE))
sleepq_wait(&sx->lock_object, 0);
else
error = sleepq_wait_sig(&sx->lock_object, 0);
#ifdef KDTRACE_HOOKS
sleep_time += lockstat_nsecs(&sx->lock_object);
sleep_cnt++;
#endif
if (error) {
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK,
"%s: interruptible sleep by %p suspended by signal",
__func__, sx);
break;
}
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
__func__, sx);
}
#ifdef KDTRACE_HOOKS
all_time += lockstat_nsecs(&sx->lock_object);
if (sleep_time)
LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
(state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
if (spin_cnt > sleep_cnt)
LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
LOCKSTAT_READER, (state & SX_LOCK_SHARED) == 0,
(state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
#endif
if (error == 0)
LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
contested, waittime, file, line, LOCKSTAT_READER);
GIANT_RESTORE();
return (error);
}
/*
* This function represents the so-called 'hard case' for sx_xlock
* operation. All 'easy case' failures are redirected to this. Note
* that ideally this would be a static function, but it needs to be
* accessible from at least sx.h.
*/
int
_sx_xlock_hard(struct sx *sx, uintptr_t tid, int opts, const char *file,
int line)
{
GIANT_DECLARE;
#ifdef ADAPTIVE_SX
volatile struct thread *owner;
u_int i, spintries = 0;
#endif
uintptr_t x;
#ifdef LOCK_PROFILING
uint64_t waittime = 0;
int contested = 0;
#endif
int error = 0;
#ifdef KDTRACE_HOOKS
uintptr_t state;
uint64_t spin_cnt = 0;
uint64_t sleep_cnt = 0;
int64_t sleep_time = 0;
int64_t all_time = 0;
#endif
if (SCHEDULER_STOPPED())
return (0);
/* If we already hold an exclusive lock, then recurse. */
if (sx_xlocked(sx)) {
KASSERT((sx->lock_object.lo_flags & LO_RECURSABLE) != 0,
("_sx_xlock_hard: recursed on non-recursive sx %s @ %s:%d\n",
sx->lock_object.lo_name, file, line));
sx->sx_recurse++;
atomic_set_ptr(&sx->sx_lock, SX_LOCK_RECURSED);
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p recursing", __func__, sx);
return (0);
}
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
sx->lock_object.lo_name, (void *)sx->sx_lock, file, line);
#ifdef KDTRACE_HOOKS
all_time -= lockstat_nsecs(&sx->lock_object);
state = sx->sx_lock;
#endif
while (!atomic_cmpset_acq_ptr(&sx->sx_lock, SX_LOCK_UNLOCKED, tid)) {
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
#ifdef HWPMC_HOOKS
PMC_SOFT_CALL( , , lock, failed);
#endif
lock_profile_obtain_lock_failed(&sx->lock_object, &contested,
&waittime);
#ifdef ADAPTIVE_SX
/*
* If the lock is write locked and the owner is
* running on another CPU, spin until the owner stops
* running or the state of the lock changes.
*/
x = sx->sx_lock;
if ((sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
if ((x & SX_LOCK_SHARED) == 0) {
x = SX_OWNER(x);
owner = (struct thread *)x;
if (TD_IS_RUNNING(owner)) {
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR3(KTR_LOCK,
"%s: spinning on %p held by %p",
__func__, sx, owner);
KTR_STATE1(KTR_SCHED, "thread",
sched_tdname(curthread), "spinning",
"lockname:\"%s\"",
sx->lock_object.lo_name);
GIANT_SAVE();
while (SX_OWNER(sx->sx_lock) == x &&
TD_IS_RUNNING(owner)) {
cpu_spinwait();
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
}
KTR_STATE0(KTR_SCHED, "thread",
sched_tdname(curthread), "running");
continue;
}
} else if (SX_SHARERS(x) && spintries < asx_retries) {
KTR_STATE1(KTR_SCHED, "thread",
sched_tdname(curthread), "spinning",
"lockname:\"%s\"", sx->lock_object.lo_name);
GIANT_SAVE();
spintries++;
for (i = 0; i < asx_loops; i++) {
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR4(KTR_LOCK,
"%s: shared spinning on %p with %u and %u",
__func__, sx, spintries, i);
x = sx->sx_lock;
if ((x & SX_LOCK_SHARED) == 0 ||
SX_SHARERS(x) == 0)
break;
cpu_spinwait();
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
}
KTR_STATE0(KTR_SCHED, "thread",
sched_tdname(curthread), "running");
if (i != asx_loops)
continue;
}
}
#endif
sleepq_lock(&sx->lock_object);
x = sx->sx_lock;
/*
* If the lock was released while spinning on the
* sleep queue chain lock, try again.
*/
if (x == SX_LOCK_UNLOCKED) {
sleepq_release(&sx->lock_object);
continue;
}
#ifdef ADAPTIVE_SX
/*
* The current lock owner might have started executing
* on another CPU (or the lock could have changed
* owners) while we were waiting on the sleep queue
* chain lock. If so, drop the sleep queue lock and try
* again.
*/
if (!(x & SX_LOCK_SHARED) &&
(sx->lock_object.lo_flags & SX_NOADAPTIVE) == 0) {
owner = (struct thread *)SX_OWNER(x);
if (TD_IS_RUNNING(owner)) {
sleepq_release(&sx->lock_object);
continue;
}
}
#endif
/*
* If an exclusive lock was released with both shared
* and exclusive waiters and a shared waiter hasn't
* woken up and acquired the lock yet, sx_lock will be
* set to SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS.
* If we see that value, try to acquire it once. Note
* that we have to preserve SX_LOCK_EXCLUSIVE_WAITERS
* as there are other exclusive waiters still. If we
* fail, restart the loop.
*/
if (x == (SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS)) {
if (atomic_cmpset_acq_ptr(&sx->sx_lock,
SX_LOCK_UNLOCKED | SX_LOCK_EXCLUSIVE_WAITERS,
tid | SX_LOCK_EXCLUSIVE_WAITERS)) {
sleepq_release(&sx->lock_object);
CTR2(KTR_LOCK, "%s: %p claimed by new writer",
__func__, sx);
break;
}
sleepq_release(&sx->lock_object);
continue;
}
/*
* Try to set the SX_LOCK_EXCLUSIVE_WAITERS. If we fail,
* than loop back and retry.
*/
if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
if (!atomic_cmpset_ptr(&sx->sx_lock, x,
x | SX_LOCK_EXCLUSIVE_WAITERS)) {
sleepq_release(&sx->lock_object);
continue;
}
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p set excl waiters flag",
__func__, sx);
}
/*
* Since we have been unable to acquire the exclusive
* lock and the exclusive waiters flag is set, we have
* to sleep.
*/
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p blocking on sleep queue",
__func__, sx);
#ifdef KDTRACE_HOOKS
sleep_time -= lockstat_nsecs(&sx->lock_object);
#endif
GIANT_SAVE();
sleepq_add(&sx->lock_object, NULL, sx->lock_object.lo_name,
SLEEPQ_SX | ((opts & SX_INTERRUPTIBLE) ?
SLEEPQ_INTERRUPTIBLE : 0), SQ_EXCLUSIVE_QUEUE);
if (!(opts & SX_INTERRUPTIBLE))
sleepq_wait(&sx->lock_object, 0);
else
error = sleepq_wait_sig(&sx->lock_object, 0);
#ifdef KDTRACE_HOOKS
sleep_time += lockstat_nsecs(&sx->lock_object);
sleep_cnt++;
#endif
if (error) {
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK,
"%s: interruptible sleep by %p suspended by signal",
__func__, sx);
break;
}
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p resuming from sleep queue",
__func__, sx);
}
#ifdef KDTRACE_HOOKS
all_time += lockstat_nsecs(&sx->lock_object);
if (sleep_time)
LOCKSTAT_RECORD4(sx__block, sx, sleep_time,
LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
(state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
if (spin_cnt > sleep_cnt)
LOCKSTAT_RECORD4(sx__spin, sx, all_time - sleep_time,
LOCKSTAT_WRITER, (state & SX_LOCK_SHARED) == 0,
(state & SX_LOCK_SHARED) == 0 ? 0 : SX_SHARERS(state));
#endif
if (!error)
LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(sx__acquire, sx,
contested, waittime, file, line, LOCKSTAT_WRITER);
GIANT_RESTORE();
return (error);
}
/*
* This function represents the so-called 'hard case' for sx_sunlock
* operation. All 'easy case' failures are redirected to this. Note
* that ideally this would be a static function, but it needs to be
* accessible from at least sx.h.
*/
void
_sx_sunlock_hard(struct sx *sx, const char *file, int line)
{
uintptr_t x;
int wakeup_swapper;
if (SCHEDULER_STOPPED())
return;
for (;;) {
x = sx->sx_lock;
/*
* We should never have sharers while at least one thread
* holds a shared lock.
*/
KASSERT(!(x & SX_LOCK_SHARED_WAITERS),
("%s: waiting sharers", __func__));
/*
* See if there is more than one shared lock held. If
* so, just drop one and return.
*/
if (SX_SHARERS(x) > 1) {
if (atomic_cmpset_rel_ptr(&sx->sx_lock, x,
x - SX_ONE_SHARER)) {
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR4(KTR_LOCK,
"%s: %p succeeded %p -> %p",
__func__, sx, (void *)x,
(void *)(x - SX_ONE_SHARER));
break;
}
continue;
}
/*
* If there aren't any waiters for an exclusive lock,
* then try to drop it quickly.
*/
if (!(x & SX_LOCK_EXCLUSIVE_WAITERS)) {
MPASS(x == SX_SHARERS_LOCK(1));
if (atomic_cmpset_rel_ptr(&sx->sx_lock,
SX_SHARERS_LOCK(1), SX_LOCK_UNLOCKED)) {
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p last succeeded",
__func__, sx);
break;
}
continue;
}
/*
* At this point, there should just be one sharer with
* exclusive waiters.
*/
MPASS(x == (SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS));
sleepq_lock(&sx->lock_object);
/*
* Wake up semantic here is quite simple:
* Just wake up all the exclusive waiters.
* Note that the state of the lock could have changed,
* so if it fails loop back and retry.
*/
if (!atomic_cmpset_rel_ptr(&sx->sx_lock,
SX_SHARERS_LOCK(1) | SX_LOCK_EXCLUSIVE_WAITERS,
SX_LOCK_UNLOCKED)) {
sleepq_release(&sx->lock_object);
continue;
}
if (LOCK_LOG_TEST(&sx->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p waking up all thread on"
"exclusive queue", __func__, sx);
wakeup_swapper = sleepq_broadcast(&sx->lock_object, SLEEPQ_SX,
0, SQ_EXCLUSIVE_QUEUE);
sleepq_release(&sx->lock_object);
if (wakeup_swapper)
kick_proc0();
break;
}
}