void
mutex_destroy(kmutex_t *mp)
{
mutex_impl_t *lp = (mutex_impl_t *)mp;
if (lp->m_owner == 0 && !MUTEX_HAS_WAITERS(lp)) {
MUTEX_DESTROY(lp);
} else if (MUTEX_TYPE_SPIN(lp)) {
LOCKSTAT_RECORD0(LS_MUTEX_DESTROY_RELEASE, lp);
MUTEX_DESTROY(lp);
} else if (MUTEX_TYPE_ADAPTIVE(lp)) {
LOCKSTAT_RECORD0(LS_MUTEX_DESTROY_RELEASE, lp);
if (MUTEX_OWNER(lp) != curthread)
mutex_panic("mutex_destroy: not owner", lp);
if (MUTEX_HAS_WAITERS(lp)) {
turnstile_t *ts = turnstile_lookup(lp);
turnstile_exit(lp);
if (ts != NULL)
mutex_panic("mutex_destroy: has waiters", lp);
}
MUTEX_DESTROY(lp);
} else {
mutex_panic("mutex_destroy: bad mutex", lp);
}
}
/*
* Try to do a non-blocking upgrade from a shared lock to an exclusive lock.
* This will only succeed if this thread holds a single shared lock.
* Return 1 if if the upgrade succeed, 0 otherwise.
*/
int
sx_try_upgrade_(struct sx *sx, const char *file, int line)
{
uintptr_t x;
int success;
if (SCHEDULER_STOPPED())
return (1);
KASSERT(sx->sx_lock != SX_LOCK_DESTROYED,
("sx_try_upgrade() of destroyed sx @ %s:%d", file, line));
_sx_assert(sx, SA_SLOCKED, file, line);
/*
* Try to switch from one shared lock to an exclusive lock. We need
* to maintain the SX_LOCK_EXCLUSIVE_WAITERS flag if set so that
* we will wake up the exclusive waiters when we drop the lock.
*/
x = sx->sx_lock & SX_LOCK_EXCLUSIVE_WAITERS;
success = atomic_cmpset_ptr(&sx->sx_lock, SX_SHARERS_LOCK(1) | x,
(uintptr_t)curthread | x);
LOCK_LOG_TRY("XUPGRADE", &sx->lock_object, 0, success, file, line);
if (success) {
WITNESS_UPGRADE(&sx->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
file, line);
LOCKSTAT_RECORD0(sx__upgrade, sx);
}
return (success);
}
/*
* 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();
}
/*
* Simple C support for the cases where spin locks miss on the first try.
*/
void
lock_set_spin(lock_t *lp)
{
int spin_count = 1;
int backoff; /* current backoff */
int backctr; /* ctr for backoff */
if (panicstr)
return;
if (ncpus == 1)
panic("lock_set: %p lock held and only one CPU", lp);
if (&plat_lock_delay) {
backoff = 0;
} else {
backoff = BACKOFF_BASE;
}
while (LOCK_HELD(lp) || !lock_spin_try(lp)) {
if (panicstr)
return;
spin_count++;
/*
* Add an exponential backoff delay before trying again
* to touch the mutex data structure.
* the spin_count test and call to nulldev are to prevent
* the compiler optimizer from eliminating the delay loop.
*/
if (&plat_lock_delay) {
plat_lock_delay(&backoff);
} else {
/* delay */
for (backctr = backoff; backctr; backctr--) {
if (!spin_count) (void) nulldev();
}
backoff = backoff << 1; /* double it */
if (backoff > BACKOFF_CAP) {
backoff = BACKOFF_CAP;
}
SMT_PAUSE();
}
}
if (spin_count) {
LOCKSTAT_RECORD(LS_LOCK_SET_SPIN, lp, spin_count);
}
LOCKSTAT_RECORD0(LS_LOCK_SET_ACQUIRE, lp);
}
/*
* mutex_vector_exit() is called from mutex_exit() if the lock is not
* adaptive, has waiters, or is not owned by the current thread (panic).
*/
void
mutex_vector_exit(mutex_impl_t *lp)
{
turnstile_t *ts;
if (MUTEX_TYPE_SPIN(lp)) {
lock_clear_splx(&lp->m_spin.m_spinlock, lp->m_spin.m_oldspl);
return;
}
if (MUTEX_OWNER(lp) != curthread) {
mutex_panic("mutex_exit: not owner", lp);
return;
}
ts = turnstile_lookup(lp);
MUTEX_CLEAR_LOCK_AND_WAITERS(lp);
if (ts == NULL)
turnstile_exit(lp);
else
turnstile_wakeup(ts, TS_WRITER_Q, ts->ts_waiters, NULL);
LOCKSTAT_RECORD0(LS_MUTEX_EXIT_RELEASE, lp);
}
/*
* Downgrade a write lock into a single read lock.
*/
void
__rw_downgrade(volatile uintptr_t *c, const char *file, int line)
{
struct rwlock *rw;
struct turnstile *ts;
uintptr_t tid, v;
int rwait, wwait;
if (SCHEDULER_STOPPED())
return;
rw = rwlock2rw(c);
KASSERT(rw->rw_lock != RW_DESTROYED,
("rw_downgrade() of destroyed rwlock @ %s:%d", file, line));
__rw_assert(c, RA_WLOCKED | RA_NOTRECURSED, file, line);
#ifndef INVARIANTS
if (rw_recursed(rw))
panic("downgrade of a recursed lock");
#endif
WITNESS_DOWNGRADE(&rw->lock_object, 0, file, line);
/*
* Convert from a writer to a single reader. First we handle
* the easy case with no waiters. If there are any waiters, we
* lock the turnstile and "disown" the lock.
*/
tid = (uintptr_t)curthread;
if (atomic_cmpset_rel_ptr(&rw->rw_lock, tid, RW_READERS_LOCK(1)))
goto out;
/*
* Ok, we think we have waiters, so lock the turnstile so we can
* read the waiter flags without any races.
*/
turnstile_chain_lock(&rw->lock_object);
v = rw->rw_lock & RW_LOCK_WAITERS;
rwait = v & RW_LOCK_READ_WAITERS;
wwait = v & RW_LOCK_WRITE_WAITERS;
MPASS(rwait | wwait);
/*
* Downgrade from a write lock while preserving waiters flag
* and give up ownership of the turnstile.
*/
ts = turnstile_lookup(&rw->lock_object);
MPASS(ts != NULL);
if (!wwait)
v &= ~RW_LOCK_READ_WAITERS;
atomic_store_rel_ptr(&rw->rw_lock, RW_READERS_LOCK(1) | v);
/*
* Wake other readers if there are no writers pending. Otherwise they
* won't be able to acquire the lock anyway.
*/
if (rwait && !wwait) {
turnstile_broadcast(ts, TS_SHARED_QUEUE);
turnstile_unpend(ts, TS_EXCLUSIVE_LOCK);
} else
turnstile_disown(ts);
turnstile_chain_unlock(&rw->lock_object);
out:
curthread->td_rw_rlocks++;
LOCK_LOG_LOCK("WDOWNGRADE", &rw->lock_object, 0, 0, file, line);
LOCKSTAT_RECORD0(rw__downgrade, rw);
}
/*
* Attempt to do a non-blocking upgrade from a read lock to a write
* lock. This will only succeed if this thread holds a single read
* lock. Returns true if the upgrade succeeded and false otherwise.
*/
int
__rw_try_upgrade(volatile uintptr_t *c, const char *file, int line)
{
struct rwlock *rw;
uintptr_t v, x, tid;
struct turnstile *ts;
int success;
if (SCHEDULER_STOPPED())
return (1);
rw = rwlock2rw(c);
KASSERT(rw->rw_lock != RW_DESTROYED,
("rw_try_upgrade() of destroyed rwlock @ %s:%d", file, line));
__rw_assert(c, RA_RLOCKED, file, line);
/*
* Attempt to switch from one reader to a writer. If there
* are any write waiters, then we will have to lock the
* turnstile first to prevent races with another writer
* calling turnstile_wait() before we have claimed this
* turnstile. So, do the simple case of no waiters first.
*/
tid = (uintptr_t)curthread;
success = 0;
for (;;) {
v = rw->rw_lock;
if (RW_READERS(v) > 1)
break;
if (!(v & RW_LOCK_WAITERS)) {
success = atomic_cmpset_ptr(&rw->rw_lock, v, tid);
if (!success)
continue;
break;
}
/*
* Ok, we think we have waiters, so lock the turnstile.
*/
ts = turnstile_trywait(&rw->lock_object);
v = rw->rw_lock;
if (RW_READERS(v) > 1) {
turnstile_cancel(ts);
break;
}
/*
* Try to switch from one reader to a writer again. This time
* we honor the current state of the waiters flags.
* If we obtain the lock with the flags set, then claim
* ownership of the turnstile.
*/
x = rw->rw_lock & RW_LOCK_WAITERS;
success = atomic_cmpset_ptr(&rw->rw_lock, v, tid | x);
if (success) {
if (x)
turnstile_claim(ts);
else
turnstile_cancel(ts);
break;
}
turnstile_cancel(ts);
}
LOCK_LOG_TRY("WUPGRADE", &rw->lock_object, 0, success, file, line);
if (success) {
curthread->td_rw_rlocks--;
WITNESS_UPGRADE(&rw->lock_object, LOP_EXCLUSIVE | LOP_TRYLOCK,
file, line);
LOCKSTAT_RECORD0(rw__upgrade, rw);
}
return (success);
}
/*
* mutex_vector_enter() is called from the assembly mutex_enter() routine
* if the lock is held or is not of type MUTEX_ADAPTIVE.
*/
void
mutex_vector_enter(mutex_impl_t *lp)
{
kthread_id_t owner;
hrtime_t sleep_time = 0; /* how long we slept */
uint_t spin_count = 0; /* how many times we spun */
cpu_t *cpup, *last_cpu;
extern cpu_t *cpu_list;
turnstile_t *ts;
volatile mutex_impl_t *vlp = (volatile mutex_impl_t *)lp;
int backoff; /* current backoff */
int backctr; /* ctr for backoff */
int sleep_count = 0;
ASSERT_STACK_ALIGNED();
if (MUTEX_TYPE_SPIN(lp)) {
lock_set_spl(&lp->m_spin.m_spinlock, lp->m_spin.m_minspl,
&lp->m_spin.m_oldspl);
return;
}
if (!MUTEX_TYPE_ADAPTIVE(lp)) {
mutex_panic("mutex_enter: bad mutex", lp);
return;
}
/*
* Adaptive mutexes must not be acquired from above LOCK_LEVEL.
* We can migrate after loading CPU but before checking CPU_ON_INTR,
* so we must verify by disabling preemption and loading CPU again.
*/
cpup = CPU;
if (CPU_ON_INTR(cpup) && !panicstr) {
kpreempt_disable();
if (CPU_ON_INTR(CPU))
mutex_panic("mutex_enter: adaptive at high PIL", lp);
kpreempt_enable();
}
CPU_STATS_ADDQ(cpup, sys, mutex_adenters, 1);
if (&plat_lock_delay) {
backoff = 0;
} else {
backoff = BACKOFF_BASE;
}
for (;;) {
spin:
spin_count++;
/*
* Add an exponential backoff delay before trying again
* to touch the mutex data structure.
* the spin_count test and call to nulldev are to prevent
* the compiler optimizer from eliminating the delay loop.
*/
if (&plat_lock_delay) {
plat_lock_delay(&backoff);
} else {
for (backctr = backoff; backctr; backctr--) {
if (!spin_count) (void) nulldev();
}; /* delay */
backoff = backoff << 1; /* double it */
if (backoff > BACKOFF_CAP) {
backoff = BACKOFF_CAP;
}
SMT_PAUSE();
}
if (panicstr)
return;
if ((owner = MUTEX_OWNER(vlp)) == NULL) {
if (mutex_adaptive_tryenter(lp))
break;
continue;
}
if (owner == curthread)
mutex_panic("recursive mutex_enter", lp);
/*
* If lock is held but owner is not yet set, spin.
* (Only relevant for platforms that don't have cas.)
*/
if (owner == MUTEX_NO_OWNER)
continue;
/*
* When searching the other CPUs, start with the one where
* we last saw the owner thread. If owner is running, spin.
*
* We must disable preemption at this point to guarantee
* that the list doesn't change while we traverse it
* without the cpu_lock mutex. While preemption is
* disabled, we must revalidate our cached cpu pointer.
*/
kpreempt_disable();
if (cpup->cpu_next == NULL)
cpup = cpu_list;
last_cpu = cpup; /* mark end of search */
do {
if (cpup->cpu_thread == owner) {
kpreempt_enable();
goto spin;
}
} while ((cpup = cpup->cpu_next) != last_cpu);
kpreempt_enable();
/*
* The owner appears not to be running, so block.
* See the Big Theory Statement for memory ordering issues.
*/
ts = turnstile_lookup(lp);
MUTEX_SET_WAITERS(lp);
membar_enter();
/*
* Recheck whether owner is running after waiters bit hits
* global visibility (above). If owner is running, spin.
*
* Since we are at ipl DISP_LEVEL, kernel preemption is
* disabled, however we still need to revalidate our cached
* cpu pointer to make sure the cpu hasn't been deleted.
*/
if (cpup->cpu_next == NULL)
last_cpu = cpup = cpu_list;
do {
if (cpup->cpu_thread == owner) {
turnstile_exit(lp);
goto spin;
}
} while ((cpup = cpup->cpu_next) != last_cpu);
membar_consumer();
/*
* If owner and waiters bit are unchanged, block.
*/
if (MUTEX_OWNER(vlp) == owner && MUTEX_HAS_WAITERS(vlp)) {
sleep_time -= gethrtime();
(void) turnstile_block(ts, TS_WRITER_Q, lp,
&mutex_sobj_ops, NULL, NULL);
sleep_time += gethrtime();
sleep_count++;
} else {
turnstile_exit(lp);
}
}
ASSERT(MUTEX_OWNER(lp) == curthread);
if (sleep_time != 0) {
/*
* Note, sleep time is the sum of all the sleeping we
* did.
*/
LOCKSTAT_RECORD(LS_MUTEX_ENTER_BLOCK, lp, sleep_time);
}
/*
* We do not count a sleep as a spin.
*/
if (spin_count > sleep_count)
LOCKSTAT_RECORD(LS_MUTEX_ENTER_SPIN, lp,
spin_count - sleep_count);
LOCKSTAT_RECORD0(LS_MUTEX_ENTER_ACQUIRE, lp);
}