boolean_t
lck_rw_lock_shared_to_exclusive(
lck_rw_t *lck)
{
int i;
boolean_t do_wakeup = FALSE;
wait_result_t res;
#if MACH_LDEBUG
int decrementer;
#endif /* MACH_LDEBUG */
boolean_t istate;
#if CONFIG_DTRACE
uint64_t wait_interval = 0;
int slept = 0;
int readers_at_sleep = 0;
#endif
istate = lck_interlock_lock(lck);
lck->lck_rw_shared_count--;
if (lck->lck_rw_want_upgrade) {
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_RW_LCK_SH_TO_EX_CODE) | DBG_FUNC_START,
(int)lck, lck->lck_rw_shared_count, lck->lck_rw_want_upgrade, 0, 0);
/*
* Someone else has requested upgrade.
* Since we've released a read lock, wake
* him up.
*/
if (lck->lck_w_waiting && (lck->lck_rw_shared_count == 0)) {
lck->lck_w_waiting = FALSE;
do_wakeup = TRUE;
}
lck_interlock_unlock(lck, istate);
if (do_wakeup)
thread_wakeup(RW_LOCK_WRITER_EVENT(lck));
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_RW_LCK_SH_TO_EX_CODE) | DBG_FUNC_END,
(int)lck, lck->lck_rw_shared_count, lck->lck_rw_want_upgrade, 0, 0);
return (FALSE);
}
lck->lck_rw_want_upgrade = TRUE;
#if MACH_LDEBUG
decrementer = DECREMENTER_TIMEOUT;
#endif /* MACH_LDEBUG */
while (lck->lck_rw_shared_count != 0) {
#if CONFIG_DTRACE
if (lockstat_probemap[LS_LCK_RW_LOCK_SHARED_TO_EXCL_SPIN] && wait_interval == 0) {
wait_interval = mach_absolute_time();
readers_at_sleep = lck->lck_rw_shared_count;
} else {
wait_interval = -1;
}
#endif
i = lock_wait_time[lck->lck_rw_can_sleep ? 1 : 0];
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_RW_LCK_SH_TO_EX1_CODE) | DBG_FUNC_START,
(int)lck, lck->lck_rw_shared_count, i, 0, 0);
if (i != 0) {
lck_interlock_unlock(lck, istate);
#if MACH_LDEBUG
if (!--decrementer)
Debugger("timeout - lck_rw_shared_count");
#endif /* MACH_LDEBUG */
while (--i != 0 && lck->lck_rw_shared_count != 0)
lck_rw_lock_pause(istate);
istate = lck_interlock_lock(lck);
}
if (lck->lck_rw_can_sleep && lck->lck_rw_shared_count != 0) {
lck->lck_w_waiting = TRUE;
res = assert_wait(RW_LOCK_WRITER_EVENT(lck), THREAD_UNINT);
if (res == THREAD_WAITING) {
lck_interlock_unlock(lck, istate);
res = thread_block(THREAD_CONTINUE_NULL);
#if CONFIG_DTRACE
slept = 1;
#endif
istate = lck_interlock_lock(lck);
}
}
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_RW_LCK_SH_TO_EX1_CODE) | DBG_FUNC_END,
(int)lck, lck->lck_rw_shared_count, 0, 0, 0);
}
lck_interlock_unlock(lck, istate);
#if CONFIG_DTRACE
/*
* We infer whether we took the sleep/spin path above by checking readers_at_sleep.
*/
if (wait_interval != 0 && wait_interval != (unsigned) -1 && readers_at_sleep) {
if (slept == 0) {
LOCKSTAT_RECORD2(LS_LCK_RW_LOCK_SHARED_TO_EXCL_SPIN, lck, mach_absolute_time() - wait_interval, 0);
} else {
LOCKSTAT_RECORD4(LS_LCK_RW_LOCK_SHARED_TO_EXCL_BLOCK, lck,
mach_absolute_time() - wait_interval, 1,
(readers_at_sleep == 0 ? 1 : 0), readers_at_sleep);
}
}
LOCKSTAT_RECORD(LS_LCK_RW_LOCK_SHARED_TO_EXCL_UPGRADE, lck, 1);
#endif
return (TRUE);
}
/*
* Routine: lck_rw_lock_exclusive
*/
void
lck_rw_lock_exclusive(
lck_rw_t *lck)
{
int i;
wait_result_t res;
#if MACH_LDEBUG
int decrementer;
#endif /* MACH_LDEBUG */
boolean_t istate;
#if CONFIG_DTRACE
uint64_t wait_interval = 0;
int slept = 0;
int readers_at_sleep;
#endif
istate = lck_interlock_lock(lck);
#if CONFIG_DTRACE
readers_at_sleep = lck->lck_rw_shared_count;
#endif
#if MACH_LDEBUG
decrementer = DECREMENTER_TIMEOUT;
#endif /* MACH_LDEBUG */
/*
* Try to acquire the lck_rw_want_write bit.
*/
while (lck->lck_rw_want_write) {
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_RW_LCK_EXCLUSIVE_CODE) | DBG_FUNC_START, (int)lck, 0, 0, 0, 0);
/*
* Either sleeping or spinning is happening, start
* a timing of our delay interval now.
*/
#if CONFIG_DTRACE
if ((lockstat_probemap[LS_LCK_RW_LOCK_EXCL_SPIN] || lockstat_probemap[LS_LCK_RW_LOCK_EXCL_BLOCK]) && wait_interval == 0) {
wait_interval = mach_absolute_time();
} else {
wait_interval = -1;
}
#endif
i = lock_wait_time[lck->lck_rw_can_sleep ? 1 : 0];
if (i != 0) {
lck_interlock_unlock(lck, istate);
#if MACH_LDEBUG
if (!--decrementer)
Debugger("timeout - lck_rw_want_write");
#endif /* MACH_LDEBUG */
while (--i != 0 && lck->lck_rw_want_write)
lck_rw_lock_pause(istate);
istate = lck_interlock_lock(lck);
}
if (lck->lck_rw_can_sleep && lck->lck_rw_want_write) {
lck->lck_w_waiting = TRUE;
res = assert_wait(RW_LOCK_WRITER_EVENT(lck), THREAD_UNINT);
if (res == THREAD_WAITING) {
lck_interlock_unlock(lck, istate);
res = thread_block(THREAD_CONTINUE_NULL);
#if CONFIG_DTRACE
slept = 1;
#endif
istate = lck_interlock_lock(lck);
}
}
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_RW_LCK_EXCLUSIVE_CODE) | DBG_FUNC_END, (int)lck, res, 0, 0, 0);
}
lck->lck_rw_want_write = TRUE;
/* Wait for readers (and upgrades) to finish */
#if MACH_LDEBUG
decrementer = DECREMENTER_TIMEOUT;
#endif /* MACH_LDEBUG */
while ((lck->lck_rw_shared_count != 0) || lck->lck_rw_want_upgrade) {
i = lock_wait_time[lck->lck_rw_can_sleep ? 1 : 0];
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_RW_LCK_EXCLUSIVE1_CODE) | DBG_FUNC_START,
(int)lck, lck->lck_rw_shared_count, lck->lck_rw_want_upgrade, i, 0);
#if CONFIG_DTRACE
/*
* Either sleeping or spinning is happening, start
* a timing of our delay interval now. If we set it
* to -1 we don't have accurate data so we cannot later
* decide to record a dtrace spin or sleep event.
*/
if ((lockstat_probemap[LS_LCK_RW_LOCK_EXCL_SPIN] || lockstat_probemap[LS_LCK_RW_LOCK_EXCL_BLOCK]) && wait_interval == 0) {
wait_interval = mach_absolute_time();
} else {
wait_interval = (unsigned) -1;
}
#endif
if (i != 0) {
lck_interlock_unlock(lck, istate);
#if MACH_LDEBUG
if (!--decrementer)
Debugger("timeout - wait for readers");
#endif /* MACH_LDEBUG */
while (--i != 0 && (lck->lck_rw_shared_count != 0 ||
lck->lck_rw_want_upgrade))
lck_rw_lock_pause(istate);
istate = lck_interlock_lock(lck);
}
if (lck->lck_rw_can_sleep && (lck->lck_rw_shared_count != 0 || lck->lck_rw_want_upgrade)) {
lck->lck_w_waiting = TRUE;
res = assert_wait(RW_LOCK_WRITER_EVENT(lck), THREAD_UNINT);
if (res == THREAD_WAITING) {
lck_interlock_unlock(lck, istate);
res = thread_block(THREAD_CONTINUE_NULL);
#if CONFIG_DTRACE
slept = 1;
#endif
istate = lck_interlock_lock(lck);
}
}
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_RW_LCK_EXCLUSIVE1_CODE) | DBG_FUNC_END,
(int)lck, lck->lck_rw_shared_count, lck->lck_rw_want_upgrade, res, 0);
}
lck_interlock_unlock(lck, istate);
#if CONFIG_DTRACE
/*
* Decide what latencies we suffered that are Dtrace events.
* If we have set wait_interval, then we either spun or slept.
* At least we get out from under the interlock before we record
* which is the best we can do here to minimize the impact
* of the tracing.
* If we have set wait_interval to -1, then dtrace was not enabled when we
* started sleeping/spinning so we don't record this event.
*/
if (wait_interval != 0 && wait_interval != (unsigned) -1) {
if (slept == 0) {
LOCKSTAT_RECORD2(LS_LCK_RW_LOCK_EXCL_SPIN, lck,
mach_absolute_time() - wait_interval, 1);
} else {
/*
* For the blocking case, we also record if when we blocked
* it was held for read or write, and how many readers.
* Notice that above we recorded this before we dropped
* the interlock so the count is accurate.
*/
LOCKSTAT_RECORD4(LS_LCK_RW_LOCK_EXCL_BLOCK, lck,
mach_absolute_time() - wait_interval, 1,
(readers_at_sleep == 0 ? 1 : 0), readers_at_sleep);
}
}
LOCKSTAT_RECORD(LS_LCK_RW_LOCK_EXCL_ACQUIRE, lck, 1);
#endif
}
/*
* Routine: lck_rw_lock_shared_gen
*/
void
lck_rw_lock_shared_gen(
lck_rw_t *lck)
{
int i;
wait_result_t res;
#if MACH_LDEBUG
int decrementer;
#endif /* MACH_LDEBUG */
boolean_t istate;
#if CONFIG_DTRACE
uint64_t wait_interval = 0;
int slept = 0;
int readers_at_sleep;
#endif
istate = lck_interlock_lock(lck);
#if CONFIG_DTRACE
readers_at_sleep = lck->lck_rw_shared_count;
#endif
#if MACH_LDEBUG
decrementer = DECREMENTER_TIMEOUT;
#endif /* MACH_LDEBUG */
while ((lck->lck_rw_want_write || lck->lck_rw_want_upgrade) &&
((lck->lck_rw_shared_count == 0) || lck->lck_rw_priv_excl)) {
i = lock_wait_time[lck->lck_rw_can_sleep ? 1 : 0];
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_RW_LCK_SHARED_CODE) | DBG_FUNC_START,
(int)lck, lck->lck_rw_want_write, lck->lck_rw_want_upgrade, i, 0);
#if CONFIG_DTRACE
if ((lockstat_probemap[LS_LCK_RW_LOCK_SHARED_SPIN] || lockstat_probemap[LS_LCK_RW_LOCK_SHARED_BLOCK]) && wait_interval == 0) {
wait_interval = mach_absolute_time();
} else {
wait_interval = -1;
}
#endif
if (i != 0) {
lck_interlock_unlock(lck, istate);
#if MACH_LDEBUG
if (!--decrementer)
Debugger("timeout - wait no writers");
#endif /* MACH_LDEBUG */
while (--i != 0 &&
(lck->lck_rw_want_write || lck->lck_rw_want_upgrade) &&
((lck->lck_rw_shared_count == 0) || lck->lck_rw_priv_excl))
lck_rw_lock_pause(istate);
istate = lck_interlock_lock(lck);
}
if (lck->lck_rw_can_sleep &&
(lck->lck_rw_want_write || lck->lck_rw_want_upgrade) &&
((lck->lck_rw_shared_count == 0) || lck->lck_rw_priv_excl)) {
lck->lck_r_waiting = TRUE;
res = assert_wait(RW_LOCK_READER_EVENT(lck), THREAD_UNINT);
if (res == THREAD_WAITING) {
lck_interlock_unlock(lck, istate);
res = thread_block(THREAD_CONTINUE_NULL);
#if CONFIG_DTRACE
slept = 1;
#endif
istate = lck_interlock_lock(lck);
}
}
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_RW_LCK_SHARED_CODE) | DBG_FUNC_END,
(int)lck, lck->lck_rw_want_write, lck->lck_rw_want_upgrade, res, 0);
}
lck->lck_rw_shared_count++;
lck_interlock_unlock(lck, istate);
#if CONFIG_DTRACE
if (wait_interval != 0 && wait_interval != (unsigned) -1) {
if (slept == 0) {
LOCKSTAT_RECORD2(LS_LCK_RW_LOCK_SHARED_SPIN, lck, mach_absolute_time() - wait_interval, 0);
} else {
LOCKSTAT_RECORD4(LS_LCK_RW_LOCK_SHARED_BLOCK, lck,
mach_absolute_time() - wait_interval, 0,
(readers_at_sleep == 0 ? 1 : 0), readers_at_sleep);
}
}
LOCKSTAT_RECORD(LS_LCK_RW_LOCK_SHARED_ACQUIRE, lck, 0);
#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_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);
}