/*
* This function is called if the first try at releasing a write lock failed.
* This means that one of the 2 waiter bits must be set indicating that at
* least one thread is waiting on this lock.
*/
void
_rw_wunlock_hard(struct rwlock *rw, uintptr_t tid, const char *file, int line)
{
struct turnstile *ts;
uintptr_t v;
int queue;
if (SCHEDULER_STOPPED())
return;
if (rw_wlocked(rw) && rw_recursed(rw)) {
rw->rw_recurse--;
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p unrecursing", __func__, rw);
return;
}
KASSERT(rw->rw_lock & (RW_LOCK_READ_WAITERS | RW_LOCK_WRITE_WAITERS),
("%s: neither of the waiter flags are set", __func__));
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p contested", __func__, rw);
turnstile_chain_lock(&rw->lock_object);
ts = turnstile_lookup(&rw->lock_object);
MPASS(ts != NULL);
/*
* Use the same algo as sx locks for now. Prefer waking up shared
* waiters if we have any over writers. This is probably not ideal.
*
* 'v' is the value we are going to write back to rw_lock. If we
* have waiters on both queues, we need to preserve the state of
* the waiter flag for the queue we don't wake up. For now this is
* hardcoded for the algorithm mentioned above.
*
* In the case of both readers and writers waiting we wakeup the
* readers but leave the RW_LOCK_WRITE_WAITERS flag set. If a
* new writer comes in before a reader it will claim the lock up
* above. There is probably a potential priority inversion in
* there that could be worked around either by waking both queues
* of waiters or doing some complicated lock handoff gymnastics.
*/
v = RW_UNLOCKED;
if (rw->rw_lock & RW_LOCK_WRITE_WAITERS) {
queue = TS_EXCLUSIVE_QUEUE;
v |= (rw->rw_lock & RW_LOCK_READ_WAITERS);
} else
queue = TS_SHARED_QUEUE;
/* Wake up all waiters for the specific queue. */
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR3(KTR_LOCK, "%s: %p waking up %s waiters", __func__, rw,
queue == TS_SHARED_QUEUE ? "read" : "write");
turnstile_broadcast(ts, queue);
atomic_store_rel_ptr(&rw->rw_lock, v);
turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
turnstile_chain_unlock(&rw->lock_object);
}
/*
* 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();
}
/*
* _mtx_lock_spin: the tougher part of acquiring an MTX_SPIN lock.
*
* This is only called if we need to actually spin for the lock. Recursion
* is handled inline.
*/
void
_mtx_lock_spin(struct mtx *m, uintptr_t tid, int opts, const char *file,
int line)
{
int i = 0;
#ifdef LOCK_PROFILING
int contested = 0;
uint64_t waittime = 0;
#endif
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
while (!_obtain_lock(m, tid)) {
/* Give interrupts a chance while we spin. */
spinlock_exit();
while (m->mtx_lock != MTX_UNOWNED) {
if (i++ < 10000000) {
cpu_spinwait();
continue;
}
if (i < 60000000 || kdb_active || panicstr != NULL)
DELAY(1);
else
_mtx_lock_spin_failed(m);
cpu_spinwait();
}
spinlock_enter();
}
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_SPIN_LOCK_ACQUIRE, m,
contested, waittime, (file), (line));
LOCKSTAT_RECORD1(LS_MTX_SPIN_LOCK_SPIN, m, i);
}
/*
* _mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
*
* We call this if the lock is either contested (i.e. we need to go to
* sleep waiting for it), or if we need to recurse on it.
*/
void
_mtx_lock_sleep(struct mtx *m, uintptr_t tid, int opts, const char *file,
int line)
{
struct turnstile *ts;
uintptr_t v;
#ifdef ADAPTIVE_MUTEXES
volatile struct thread *owner;
#endif
#ifdef KTR
int cont_logged = 0;
#endif
#ifdef LOCK_PROFILING
int contested = 0;
uint64_t waittime = 0;
#endif
#ifdef KDTRACE_HOOKS
uint64_t spin_cnt = 0;
uint64_t sleep_cnt = 0;
int64_t sleep_time = 0;
#endif
if (mtx_owned(m)) {
KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0,
("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
m->lock_object.lo_name, file, line));
m->mtx_recurse++;
atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
return;
}
lock_profile_obtain_lock_failed(&m->lock_object,
&contested, &waittime);
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR4(KTR_LOCK,
"_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
while (!_obtain_lock(m, tid)) {
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
#ifdef ADAPTIVE_MUTEXES
/*
* If the owner is running on another CPU, spin until the
* owner stops running or the state of the lock changes.
*/
v = m->mtx_lock;
if (v != MTX_UNOWNED) {
owner = (struct thread *)(v & ~MTX_FLAGMASK);
if (TD_IS_RUNNING(owner)) {
if (LOCK_LOG_TEST(&m->lock_object, 0))
CTR3(KTR_LOCK,
"%s: spinning on %p held by %p",
__func__, m, owner);
while (mtx_owner(m) == owner &&
TD_IS_RUNNING(owner)) {
cpu_spinwait();
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
}
continue;
}
}
#endif
ts = turnstile_trywait(&m->lock_object);
v = m->mtx_lock;
/*
* Check if the lock has been released while spinning for
* the turnstile chain lock.
*/
if (v == MTX_UNOWNED) {
turnstile_cancel(ts);
continue;
}
#ifdef ADAPTIVE_MUTEXES
/*
* The current lock owner might have started executing
* on another CPU (or the lock could have changed
* owners) while we were waiting on the turnstile
* chain lock. If so, drop the turnstile lock and try
* again.
*/
owner = (struct thread *)(v & ~MTX_FLAGMASK);
if (TD_IS_RUNNING(owner)) {
turnstile_cancel(ts);
continue;
}
#endif
/*
* If the mutex isn't already contested and a failure occurs
* setting the contested bit, the mutex was either released
* or the state of the MTX_RECURSED bit changed.
*/
if ((v & MTX_CONTESTED) == 0 &&
!atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
turnstile_cancel(ts);
continue;
}
/*
* We definitely must sleep for this lock.
*/
mtx_assert(m, MA_NOTOWNED);
#ifdef KTR
if (!cont_logged) {
CTR6(KTR_CONTENTION,
"contention: %p at %s:%d wants %s, taken by %s:%d",
(void *)tid, file, line, m->lock_object.lo_name,
WITNESS_FILE(&m->lock_object),
WITNESS_LINE(&m->lock_object));
cont_logged = 1;
}
#endif
/*
* Block on the turnstile.
*/
#ifdef KDTRACE_HOOKS
sleep_time -= lockstat_nsecs();
#endif
turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
#ifdef KDTRACE_HOOKS
sleep_time += lockstat_nsecs();
sleep_cnt++;
#endif
}
#ifdef KTR
if (cont_logged) {
CTR4(KTR_CONTENTION,
"contention end: %s acquired by %p at %s:%d",
m->lock_object.lo_name, (void *)tid, file, line);
}
#endif
LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(LS_MTX_LOCK_ACQUIRE, m, contested,
waittime, file, line);
#ifdef KDTRACE_HOOKS
if (sleep_time)
LOCKSTAT_RECORD1(LS_MTX_LOCK_BLOCK, m, sleep_time);
/*
* Only record the loops spinning and not sleeping.
*/
if (spin_cnt > sleep_cnt)
LOCKSTAT_RECORD1(LS_MTX_LOCK_SPIN, m, (spin_cnt - sleep_cnt));
#endif
}
/*
* This function is called when we are unable to obtain a write lock on the
* first try. This means that at least one other thread holds either a
* read or write lock.
*/
void
__rw_wlock_hard(volatile uintptr_t *c, uintptr_t tid, const char *file,
int line)
{
struct rwlock *rw;
struct turnstile *ts;
#ifdef ADAPTIVE_RWLOCKS
volatile struct thread *owner;
int spintries = 0;
int i;
#endif
uintptr_t v, x;
#ifdef LOCK_PROFILING
uint64_t waittime = 0;
int contested = 0;
#endif
#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;
rw = rwlock2rw(c);
if (rw_wlocked(rw)) {
KASSERT(rw->lock_object.lo_flags & LO_RECURSABLE,
("%s: recursing but non-recursive rw %s @ %s:%d\n",
__func__, rw->lock_object.lo_name, file, line));
rw->rw_recurse++;
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p recursing", __func__, rw);
return;
}
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR5(KTR_LOCK, "%s: %s contested (lock=%p) at %s:%d", __func__,
rw->lock_object.lo_name, (void *)rw->rw_lock, file, line);
#ifdef KDTRACE_HOOKS
all_time -= lockstat_nsecs(&rw->lock_object);
state = rw->rw_lock;
#endif
while (!_rw_write_lock(rw, tid)) {
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
#ifdef HWPMC_HOOKS
PMC_SOFT_CALL( , , lock, failed);
#endif
lock_profile_obtain_lock_failed(&rw->lock_object,
&contested, &waittime);
#ifdef ADAPTIVE_RWLOCKS
/*
* 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.
*/
v = rw->rw_lock;
owner = (struct thread *)RW_OWNER(v);
if (!(v & RW_LOCK_READ) && TD_IS_RUNNING(owner)) {
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR3(KTR_LOCK, "%s: spinning on %p held by %p",
__func__, rw, owner);
KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
"spinning", "lockname:\"%s\"",
rw->lock_object.lo_name);
while ((struct thread*)RW_OWNER(rw->rw_lock) == owner &&
TD_IS_RUNNING(owner)) {
cpu_spinwait();
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
}
KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
"running");
continue;
}
if ((v & RW_LOCK_READ) && RW_READERS(v) &&
spintries < rowner_retries) {
if (!(v & RW_LOCK_WRITE_SPINNER)) {
if (!atomic_cmpset_ptr(&rw->rw_lock, v,
v | RW_LOCK_WRITE_SPINNER)) {
continue;
}
}
spintries++;
KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
"spinning", "lockname:\"%s\"",
rw->lock_object.lo_name);
for (i = 0; i < rowner_loops; i++) {
if ((rw->rw_lock & RW_LOCK_WRITE_SPINNER) == 0)
break;
cpu_spinwait();
}
KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
"running");
#ifdef KDTRACE_HOOKS
spin_cnt += rowner_loops - i;
#endif
if (i != rowner_loops)
continue;
}
#endif
ts = turnstile_trywait(&rw->lock_object);
v = rw->rw_lock;
#ifdef ADAPTIVE_RWLOCKS
/*
* The current lock owner might have started executing
* on another CPU (or the lock could have changed
* owners) while we were waiting on the turnstile
* chain lock. If so, drop the turnstile lock and try
* again.
*/
if (!(v & RW_LOCK_READ)) {
owner = (struct thread *)RW_OWNER(v);
if (TD_IS_RUNNING(owner)) {
turnstile_cancel(ts);
continue;
}
}
#endif
/*
* Check for the waiters flags about this rwlock.
* If the lock was released, without maintain any pending
* waiters queue, simply try to acquire it.
* If a pending waiters queue is present, claim the lock
* ownership and maintain the pending queue.
*/
x = v & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
if ((v & ~x) == RW_UNLOCKED) {
x &= ~RW_LOCK_WRITE_SPINNER;
if (atomic_cmpset_acq_ptr(&rw->rw_lock, v, tid | x)) {
if (x)
turnstile_claim(ts);
else
turnstile_cancel(ts);
break;
}
turnstile_cancel(ts);
continue;
}
/*
* If the RW_LOCK_WRITE_WAITERS flag isn't set, then try to
* set it. If we fail to set it, then loop back and try
* again.
*/
if (!(v & RW_LOCK_WRITE_WAITERS)) {
if (!atomic_cmpset_ptr(&rw->rw_lock, v,
v | RW_LOCK_WRITE_WAITERS)) {
turnstile_cancel(ts);
continue;
}
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p set write waiters flag",
__func__, rw);
}
/*
* We were unable to acquire the lock and the write waiters
* flag is set, so we must block on the turnstile.
*/
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
rw);
#ifdef KDTRACE_HOOKS
sleep_time -= lockstat_nsecs(&rw->lock_object);
#endif
turnstile_wait(ts, rw_owner(rw), TS_EXCLUSIVE_QUEUE);
#ifdef KDTRACE_HOOKS
sleep_time += lockstat_nsecs(&rw->lock_object);
sleep_cnt++;
#endif
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
__func__, rw);
#ifdef ADAPTIVE_RWLOCKS
spintries = 0;
#endif
}
#ifdef KDTRACE_HOOKS
all_time += lockstat_nsecs(&rw->lock_object);
if (sleep_time)
LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
LOCKSTAT_WRITER, (state & RW_LOCK_READ) == 0,
(state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
/* Record only the loops spinning and not sleeping. */
if (spin_cnt > sleep_cnt)
LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
(state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
#endif
LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
waittime, file, line, LOCKSTAT_WRITER);
}
void
_rw_runlock_cookie(volatile uintptr_t *c, const char *file, int line)
{
struct rwlock *rw;
struct turnstile *ts;
uintptr_t x, v, queue;
if (SCHEDULER_STOPPED())
return;
rw = rwlock2rw(c);
KASSERT(rw->rw_lock != RW_DESTROYED,
("rw_runlock() of destroyed rwlock @ %s:%d", file, line));
__rw_assert(c, RA_RLOCKED, file, line);
WITNESS_UNLOCK(&rw->lock_object, 0, file, line);
LOCK_LOG_LOCK("RUNLOCK", &rw->lock_object, 0, 0, file, line);
/* TODO: drop "owner of record" here. */
for (;;) {
/*
* See if there is more than one read lock held. If so,
* just drop one and return.
*/
x = rw->rw_lock;
if (RW_READERS(x) > 1) {
if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
x - RW_ONE_READER)) {
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR4(KTR_LOCK,
"%s: %p succeeded %p -> %p",
__func__, rw, (void *)x,
(void *)(x - RW_ONE_READER));
break;
}
continue;
}
/*
* If there aren't any waiters for a write lock, then try
* to drop it quickly.
*/
if (!(x & RW_LOCK_WAITERS)) {
MPASS((x & ~RW_LOCK_WRITE_SPINNER) ==
RW_READERS_LOCK(1));
if (atomic_cmpset_rel_ptr(&rw->rw_lock, x,
RW_UNLOCKED)) {
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p last succeeded",
__func__, rw);
break;
}
continue;
}
/*
* Ok, we know we have waiters and we think we are the
* last reader, so grab the turnstile lock.
*/
turnstile_chain_lock(&rw->lock_object);
v = rw->rw_lock & (RW_LOCK_WAITERS | RW_LOCK_WRITE_SPINNER);
MPASS(v & RW_LOCK_WAITERS);
/*
* Try to drop our lock leaving the lock in a unlocked
* state.
*
* If you wanted to do explicit lock handoff you'd have to
* do it here. You'd also want to use turnstile_signal()
* and you'd have to handle the race where a higher
* priority thread blocks on the write lock before the
* thread you wakeup actually runs and have the new thread
* "steal" the lock. For now it's a lot simpler to just
* wakeup all of the waiters.
*
* As above, if we fail, then another thread might have
* acquired a read lock, so drop the turnstile lock and
* restart.
*/
x = RW_UNLOCKED;
if (v & RW_LOCK_WRITE_WAITERS) {
queue = TS_EXCLUSIVE_QUEUE;
x |= (v & RW_LOCK_READ_WAITERS);
} else
queue = TS_SHARED_QUEUE;
if (!atomic_cmpset_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v,
x)) {
turnstile_chain_unlock(&rw->lock_object);
continue;
}
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p last succeeded with waiters",
__func__, rw);
/*
* Ok. The lock is released and all that's left is to
* wake up the waiters. Note that the lock might not be
* free anymore, but in that case the writers will just
* block again if they run before the new lock holder(s)
* release the lock.
*/
ts = turnstile_lookup(&rw->lock_object);
MPASS(ts != NULL);
turnstile_broadcast(ts, queue);
turnstile_unpend(ts, TS_SHARED_LOCK);
turnstile_chain_unlock(&rw->lock_object);
break;
}
LOCKSTAT_PROFILE_RELEASE_RWLOCK(rw__release, rw, LOCKSTAT_READER);
curthread->td_locks--;
curthread->td_rw_rlocks--;
}
void
__rw_rlock(volatile uintptr_t *c, const char *file, int line)
{
struct rwlock *rw;
struct turnstile *ts;
#ifdef ADAPTIVE_RWLOCKS
volatile struct thread *owner;
int spintries = 0;
int i;
#endif
#ifdef LOCK_PROFILING
uint64_t waittime = 0;
int contested = 0;
#endif
uintptr_t v;
#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;
rw = rwlock2rw(c);
KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread),
("rw_rlock() by idle thread %p on rwlock %s @ %s:%d",
curthread, rw->lock_object.lo_name, file, line));
KASSERT(rw->rw_lock != RW_DESTROYED,
("rw_rlock() of destroyed rwlock @ %s:%d", file, line));
KASSERT(rw_wowner(rw) != curthread,
("rw_rlock: wlock already held for %s @ %s:%d",
rw->lock_object.lo_name, file, line));
WITNESS_CHECKORDER(&rw->lock_object, LOP_NEWORDER, file, line, NULL);
#ifdef KDTRACE_HOOKS
all_time -= lockstat_nsecs(&rw->lock_object);
state = rw->rw_lock;
#endif
for (;;) {
/*
* Handle the easy case. If no other thread has a write
* lock, then try to bump up the count of read locks. Note
* that we have to preserve the current state of the
* RW_LOCK_WRITE_WAITERS flag. If we fail to acquire a
* read lock, then rw_lock must have changed, so restart
* the loop. Note that this handles the case of a
* completely unlocked rwlock since such a lock is encoded
* as a read lock with no waiters.
*/
v = rw->rw_lock;
if (RW_CAN_READ(v)) {
/*
* The RW_LOCK_READ_WAITERS flag should only be set
* if the lock has been unlocked and write waiters
* were present.
*/
if (atomic_cmpset_acq_ptr(&rw->rw_lock, v,
v + RW_ONE_READER)) {
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR4(KTR_LOCK,
"%s: %p succeed %p -> %p", __func__,
rw, (void *)v,
(void *)(v + RW_ONE_READER));
break;
}
continue;
}
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
#ifdef HWPMC_HOOKS
PMC_SOFT_CALL( , , lock, failed);
#endif
lock_profile_obtain_lock_failed(&rw->lock_object,
&contested, &waittime);
#ifdef ADAPTIVE_RWLOCKS
/*
* If the owner is running on another CPU, spin until
* the owner stops running or the state of the lock
* changes.
*/
if ((v & RW_LOCK_READ) == 0) {
owner = (struct thread *)RW_OWNER(v);
if (TD_IS_RUNNING(owner)) {
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR3(KTR_LOCK,
"%s: spinning on %p held by %p",
__func__, rw, owner);
KTR_STATE1(KTR_SCHED, "thread",
sched_tdname(curthread), "spinning",
"lockname:\"%s\"", rw->lock_object.lo_name);
while ((struct thread*)RW_OWNER(rw->rw_lock) ==
owner && TD_IS_RUNNING(owner)) {
cpu_spinwait();
#ifdef KDTRACE_HOOKS
spin_cnt++;
#endif
}
KTR_STATE0(KTR_SCHED, "thread",
sched_tdname(curthread), "running");
continue;
}
} else if (spintries < rowner_retries) {
spintries++;
KTR_STATE1(KTR_SCHED, "thread", sched_tdname(curthread),
"spinning", "lockname:\"%s\"",
rw->lock_object.lo_name);
for (i = 0; i < rowner_loops; i++) {
v = rw->rw_lock;
if ((v & RW_LOCK_READ) == 0 || RW_CAN_READ(v))
break;
cpu_spinwait();
}
#ifdef KDTRACE_HOOKS
spin_cnt += rowner_loops - i;
#endif
KTR_STATE0(KTR_SCHED, "thread", sched_tdname(curthread),
"running");
if (i != rowner_loops)
continue;
}
#endif
/*
* Okay, now it's the hard case. Some other thread already
* has a write lock or there are write waiters present,
* acquire the turnstile lock so we can begin the process
* of blocking.
*/
ts = turnstile_trywait(&rw->lock_object);
/*
* The lock might have been released while we spun, so
* recheck its state and restart the loop if needed.
*/
v = rw->rw_lock;
if (RW_CAN_READ(v)) {
turnstile_cancel(ts);
continue;
}
#ifdef ADAPTIVE_RWLOCKS
/*
* The current lock owner might have started executing
* on another CPU (or the lock could have changed
* owners) while we were waiting on the turnstile
* chain lock. If so, drop the turnstile lock and try
* again.
*/
if ((v & RW_LOCK_READ) == 0) {
owner = (struct thread *)RW_OWNER(v);
if (TD_IS_RUNNING(owner)) {
turnstile_cancel(ts);
continue;
}
}
#endif
/*
* The lock is held in write mode or it already has waiters.
*/
MPASS(!RW_CAN_READ(v));
/*
* If the RW_LOCK_READ_WAITERS flag is already set, then
* we can go ahead and block. If it is not set then try
* to set it. If we fail to set it drop the turnstile
* lock and restart the loop.
*/
if (!(v & RW_LOCK_READ_WAITERS)) {
if (!atomic_cmpset_ptr(&rw->rw_lock, v,
v | RW_LOCK_READ_WAITERS)) {
turnstile_cancel(ts);
continue;
}
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p set read waiters flag",
__func__, rw);
}
/*
* We were unable to acquire the lock and the read waiters
* flag is set, so we must block on the turnstile.
*/
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p blocking on turnstile", __func__,
rw);
#ifdef KDTRACE_HOOKS
sleep_time -= lockstat_nsecs(&rw->lock_object);
#endif
turnstile_wait(ts, rw_owner(rw), TS_SHARED_QUEUE);
#ifdef KDTRACE_HOOKS
sleep_time += lockstat_nsecs(&rw->lock_object);
sleep_cnt++;
#endif
if (LOCK_LOG_TEST(&rw->lock_object, 0))
CTR2(KTR_LOCK, "%s: %p resuming from turnstile",
__func__, rw);
}
#ifdef KDTRACE_HOOKS
all_time += lockstat_nsecs(&rw->lock_object);
if (sleep_time)
LOCKSTAT_RECORD4(rw__block, rw, sleep_time,
LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
(state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
/* Record only the loops spinning and not sleeping. */
if (spin_cnt > sleep_cnt)
LOCKSTAT_RECORD4(rw__spin, rw, all_time - sleep_time,
LOCKSTAT_READER, (state & RW_LOCK_READ) == 0,
(state & RW_LOCK_READ) == 0 ? 0 : RW_READERS(state));
#endif
/*
* TODO: acquire "owner of record" here. Here be turnstile dragons
* however. turnstiles don't like owners changing between calls to
* turnstile_wait() currently.
*/
LOCKSTAT_PROFILE_OBTAIN_RWLOCK_SUCCESS(rw__acquire, rw, contested,
waittime, file, line, LOCKSTAT_READER);
LOCK_LOG_LOCK("RLOCK", &rw->lock_object, 0, 0, file, line);
WITNESS_LOCK(&rw->lock_object, 0, file, line);
curthread->td_locks++;
curthread->td_rw_rlocks++;
}
/*
* 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;
}
}
/*
* _mtx_lock_spin_cookie: the tougher part of acquiring an MTX_SPIN lock.
*
* This is only called if we need to actually spin for the lock. Recursion
* is handled inline.
*/
void
_mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t tid, int opts,
const char *file, int line)
{
struct mtx *m;
struct lock_delay_arg lda;
#ifdef LOCK_PROFILING
int contested = 0;
uint64_t waittime = 0;
#endif
#ifdef KDTRACE_HOOKS
int64_t spin_time = 0;
#endif
if (SCHEDULER_STOPPED())
return;
lock_delay_arg_init(&lda, &mtx_spin_delay);
m = mtxlock2mtx(c);
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
"spinning", "lockname:\"%s\"", m->lock_object.lo_name);
#ifdef HWPMC_HOOKS
PMC_SOFT_CALL( , , lock, failed);
#endif
lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
#ifdef KDTRACE_HOOKS
spin_time -= lockstat_nsecs(&m->lock_object);
#endif
for (;;) {
if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid))
break;
/* Give interrupts a chance while we spin. */
spinlock_exit();
while (m->mtx_lock != MTX_UNOWNED) {
if (lda.spin_cnt < 10000000) {
lock_delay(&lda);
continue;
}
lda.spin_cnt++;
if (lda.spin_cnt < 60000000 || kdb_active ||
panicstr != NULL)
DELAY(1);
else
_mtx_lock_spin_failed(m);
cpu_spinwait();
}
spinlock_enter();
}
#ifdef KDTRACE_HOOKS
spin_time += lockstat_nsecs(&m->lock_object);
#endif
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
"running");
#ifdef KDTRACE_HOOKS
LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m,
contested, waittime, file, line);
if (spin_time != 0)
LOCKSTAT_RECORD1(spin__spin, m, spin_time);
#endif
}
/*
* __mtx_lock_sleep: the tougher part of acquiring an MTX_DEF lock.
*
* We call this if the lock is either contested (i.e. we need to go to
* sleep waiting for it), or if we need to recurse on it.
*/
void
__mtx_lock_sleep(volatile uintptr_t *c, uintptr_t tid, int opts,
const char *file, int line)
{
struct mtx *m;
struct turnstile *ts;
uintptr_t v;
#ifdef ADAPTIVE_MUTEXES
volatile struct thread *owner;
#endif
#ifdef KTR
int cont_logged = 0;
#endif
#ifdef LOCK_PROFILING
int contested = 0;
uint64_t waittime = 0;
#endif
#if defined(ADAPTIVE_MUTEXES) || defined(KDTRACE_HOOKS)
struct lock_delay_arg lda;
#endif
#ifdef KDTRACE_HOOKS
u_int sleep_cnt = 0;
int64_t sleep_time = 0;
int64_t all_time = 0;
#endif
if (SCHEDULER_STOPPED())
return;
#if defined(ADAPTIVE_MUTEXES)
lock_delay_arg_init(&lda, &mtx_delay);
#elif defined(KDTRACE_HOOKS)
lock_delay_arg_init(&lda, NULL);
#endif
m = mtxlock2mtx(c);
if (mtx_owned(m)) {
KASSERT((m->lock_object.lo_flags & LO_RECURSABLE) != 0 ||
(opts & MTX_RECURSE) != 0,
("_mtx_lock_sleep: recursed on non-recursive mutex %s @ %s:%d\n",
m->lock_object.lo_name, file, line));
opts &= ~MTX_RECURSE;
m->mtx_recurse++;
atomic_set_ptr(&m->mtx_lock, MTX_RECURSED);
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR1(KTR_LOCK, "_mtx_lock_sleep: %p recursing", m);
return;
}
opts &= ~MTX_RECURSE;
#ifdef HWPMC_HOOKS
PMC_SOFT_CALL( , , lock, failed);
#endif
lock_profile_obtain_lock_failed(&m->lock_object,
&contested, &waittime);
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR4(KTR_LOCK,
"_mtx_lock_sleep: %s contested (lock=%p) at %s:%d",
m->lock_object.lo_name, (void *)m->mtx_lock, file, line);
#ifdef KDTRACE_HOOKS
all_time -= lockstat_nsecs(&m->lock_object);
#endif
for (;;) {
if (m->mtx_lock == MTX_UNOWNED && _mtx_obtain_lock(m, tid))
break;
#ifdef KDTRACE_HOOKS
lda.spin_cnt++;
#endif
#ifdef ADAPTIVE_MUTEXES
/*
* If the owner is running on another CPU, spin until the
* owner stops running or the state of the lock changes.
*/
v = m->mtx_lock;
if (v != MTX_UNOWNED) {
owner = (struct thread *)(v & ~MTX_FLAGMASK);
if (TD_IS_RUNNING(owner)) {
if (LOCK_LOG_TEST(&m->lock_object, 0))
CTR3(KTR_LOCK,
"%s: spinning on %p held by %p",
__func__, m, owner);
KTR_STATE1(KTR_SCHED, "thread",
sched_tdname((struct thread *)tid),
"spinning", "lockname:\"%s\"",
m->lock_object.lo_name);
while (mtx_owner(m) == owner &&
TD_IS_RUNNING(owner))
lock_delay(&lda);
KTR_STATE0(KTR_SCHED, "thread",
sched_tdname((struct thread *)tid),
"running");
continue;
}
}
#endif
ts = turnstile_trywait(&m->lock_object);
v = m->mtx_lock;
/*
* Check if the lock has been released while spinning for
* the turnstile chain lock.
*/
if (v == MTX_UNOWNED) {
turnstile_cancel(ts);
continue;
}
#ifdef ADAPTIVE_MUTEXES
/*
* The current lock owner might have started executing
* on another CPU (or the lock could have changed
* owners) while we were waiting on the turnstile
* chain lock. If so, drop the turnstile lock and try
* again.
*/
owner = (struct thread *)(v & ~MTX_FLAGMASK);
if (TD_IS_RUNNING(owner)) {
turnstile_cancel(ts);
continue;
}
#endif
/*
* If the mutex isn't already contested and a failure occurs
* setting the contested bit, the mutex was either released
* or the state of the MTX_RECURSED bit changed.
*/
if ((v & MTX_CONTESTED) == 0 &&
!atomic_cmpset_ptr(&m->mtx_lock, v, v | MTX_CONTESTED)) {
turnstile_cancel(ts);
continue;
}
/*
* We definitely must sleep for this lock.
*/
mtx_assert(m, MA_NOTOWNED);
#ifdef KTR
if (!cont_logged) {
CTR6(KTR_CONTENTION,
"contention: %p at %s:%d wants %s, taken by %s:%d",
(void *)tid, file, line, m->lock_object.lo_name,
WITNESS_FILE(&m->lock_object),
WITNESS_LINE(&m->lock_object));
cont_logged = 1;
}
#endif
/*
* Block on the turnstile.
*/
#ifdef KDTRACE_HOOKS
sleep_time -= lockstat_nsecs(&m->lock_object);
#endif
turnstile_wait(ts, mtx_owner(m), TS_EXCLUSIVE_QUEUE);
#ifdef KDTRACE_HOOKS
sleep_time += lockstat_nsecs(&m->lock_object);
sleep_cnt++;
#endif
}
#ifdef KDTRACE_HOOKS
all_time += lockstat_nsecs(&m->lock_object);
#endif
#ifdef KTR
if (cont_logged) {
CTR4(KTR_CONTENTION,
"contention end: %s acquired by %p at %s:%d",
m->lock_object.lo_name, (void *)tid, file, line);
}
#endif
LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(adaptive__acquire, m, contested,
waittime, file, line);
#ifdef KDTRACE_HOOKS
if (sleep_time)
LOCKSTAT_RECORD1(adaptive__block, m, sleep_time);
/*
* Only record the loops spinning and not sleeping.
*/
if (lda.spin_cnt > sleep_cnt)
LOCKSTAT_RECORD1(adaptive__spin, m, all_time - sleep_time);
#endif
}
void
_mtx_lock_spin_cookie(volatile uintptr_t *c, uintptr_t v)
#endif
{
struct mtx *m;
struct lock_delay_arg lda;
uintptr_t tid;
#ifdef LOCK_PROFILING
int contested = 0;
uint64_t waittime = 0;
#endif
#ifdef KDTRACE_HOOKS
int64_t spin_time = 0;
#endif
#if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
int doing_lockprof;
#endif
tid = (uintptr_t)curthread;
m = mtxlock2mtx(c);
if (__predict_false(v == MTX_UNOWNED))
v = MTX_READ_VALUE(m);
if (__predict_false(v == tid)) {
m->mtx_recurse++;
return;
}
if (SCHEDULER_STOPPED())
return;
lock_delay_arg_init(&lda, &mtx_spin_delay);
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR1(KTR_LOCK, "_mtx_lock_spin: %p spinning", m);
KTR_STATE1(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
"spinning", "lockname:\"%s\"", m->lock_object.lo_name);
#ifdef HWPMC_HOOKS
PMC_SOFT_CALL( , , lock, failed);
#endif
lock_profile_obtain_lock_failed(&m->lock_object, &contested, &waittime);
#ifdef LOCK_PROFILING
doing_lockprof = 1;
#elif defined(KDTRACE_HOOKS)
doing_lockprof = lockstat_enabled;
if (__predict_false(doing_lockprof))
spin_time -= lockstat_nsecs(&m->lock_object);
#endif
for (;;) {
if (v == MTX_UNOWNED) {
if (_mtx_obtain_lock_fetch(m, &v, tid))
break;
continue;
}
/* Give interrupts a chance while we spin. */
spinlock_exit();
do {
if (lda.spin_cnt < 10000000) {
lock_delay(&lda);
} else {
lda.spin_cnt++;
if (lda.spin_cnt < 60000000 || kdb_active ||
panicstr != NULL)
DELAY(1);
else
_mtx_lock_spin_failed(m);
cpu_spinwait();
}
v = MTX_READ_VALUE(m);
} while (v != MTX_UNOWNED);
spinlock_enter();
}
if (LOCK_LOG_TEST(&m->lock_object, opts))
CTR1(KTR_LOCK, "_mtx_lock_spin: %p spin done", m);
KTR_STATE0(KTR_SCHED, "thread", sched_tdname((struct thread *)tid),
"running");
#if defined(KDTRACE_HOOKS) || defined(LOCK_PROFILING)
if (__predict_true(!doing_lockprof))
return;
#endif
#ifdef KDTRACE_HOOKS
spin_time += lockstat_nsecs(&m->lock_object);
#endif
LOCKSTAT_PROFILE_OBTAIN_LOCK_SUCCESS(spin__acquire, m,
contested, waittime, file, line);
#ifdef KDTRACE_HOOKS
if (lda.spin_cnt != 0)
LOCKSTAT_RECORD1(spin__spin, m, spin_time);
#endif
}