X-Git-Url: https://git.saurik.com/apple/xnu.git/blobdiff_plain/0a7de7458d150b5d4dffc935ba399be265ef0a1a..cb3231590a3c94ab4375e2228bd5e86b0cf1ad7e:/osfmk/kern/locks.c diff --git a/osfmk/kern/locks.c b/osfmk/kern/locks.c index 04106709b..78aee369c 100644 --- a/osfmk/kern/locks.c +++ b/osfmk/kern/locks.c @@ -1,5 +1,5 @@ /* - * Copyright (c) 2000-2016 Apple Inc. All rights reserved. + * Copyright (c) 2000-2019 Apple Inc. All rights reserved. * * @APPLE_OSREFERENCE_LICENSE_HEADER_START@ * @@ -54,7 +54,6 @@ * the rights to redistribute these changes. */ -#define ATOMIC_PRIVATE 1 #define LOCK_PRIVATE 1 #include @@ -90,25 +89,16 @@ #define ALIGN_TEST(p, t) do{}while(0) #endif -/* Silence the volatile to _Atomic cast warning */ -#define ATOMIC_CAST(t, p) ((_Atomic t*)(uintptr_t)(p)) - -/* Enforce program order of loads and stores. */ -#define ordered_load(target, type) \ - __c11_atomic_load((_Atomic type *)(target), memory_order_relaxed) -#define ordered_store(target, type, value) \ - __c11_atomic_store((_Atomic type *)(target), value, memory_order_relaxed) - -#define ordered_load_hw(lock) ordered_load(&(lock)->lock_data, uintptr_t) -#define ordered_store_hw(lock, value) ordered_store(&(lock)->lock_data, uintptr_t, (value)) - #define NOINLINE __attribute__((noinline)) +#define ordered_load_hw(lock) os_atomic_load(&(lock)->lock_data, compiler_acq_rel) +#define ordered_store_hw(lock, value) os_atomic_store(&(lock)->lock_data, (value), compiler_acq_rel) + queue_head_t lck_grp_queue; unsigned int lck_grp_cnt; -decl_lck_mtx_data(, lck_grp_lock) +decl_lck_mtx_data(, lck_grp_lock); static lck_mtx_ext_t lck_grp_lock_ext; SECURITY_READ_ONLY_LATE(boolean_t) spinlock_timeout_panic = TRUE; @@ -175,7 +165,7 @@ lck_mod_init( LockCompatGroup.lck_grp_attr |= LCK_GRP_ATTR_TIME_STAT; } - LockCompatGroup.lck_grp_refcnt = 1; + os_ref_init(&LockCompatGroup.lck_grp_refcnt, NULL); enqueue_tail(&lck_grp_queue, (queue_entry_t)&LockCompatGroup); lck_grp_cnt = 1; @@ -228,7 +218,7 @@ void lck_grp_attr_setstat( lck_grp_attr_t *attr) { - (void)hw_atomic_or(&attr->grp_attr_val, LCK_GRP_ATTR_STAT); + os_atomic_or(&attr->grp_attr_val, LCK_GRP_ATTR_STAT, relaxed); } @@ -307,7 +297,7 @@ lck_grp_init(lck_grp_t * grp, const char * grp_name, lck_grp_attr_t * attr) #endif /* LOCK_STATS */ } - grp->lck_grp_refcnt = 1; + os_ref_init(&grp->lck_grp_refcnt, NULL); lck_mtx_lock(&lck_grp_lock); enqueue_tail(&lck_grp_queue, (queue_entry_t)grp); @@ -339,7 +329,7 @@ void lck_grp_reference( lck_grp_t *grp) { - (void)hw_atomic_add(&grp->lck_grp_refcnt, 1); + os_ref_retain(&grp->lck_grp_refcnt); } @@ -351,9 +341,11 @@ void lck_grp_deallocate( lck_grp_t *grp) { - if (hw_atomic_sub(&grp->lck_grp_refcnt, 1) == 0) { - kfree(grp, sizeof(lck_grp_t)); + if (os_ref_release(&grp->lck_grp_refcnt) != 0) { + return; } + + kfree(grp, sizeof(lck_grp_t)); } /* @@ -381,7 +373,7 @@ lck_grp_lckcnt_incr( return panic("lck_grp_lckcnt_incr(): invalid lock type: %d\n", lck_type); } - (void)hw_atomic_add(lckcnt, 1); + os_atomic_inc(lckcnt, relaxed); } /* @@ -411,7 +403,7 @@ lck_grp_lckcnt_decr( return; } - updated = (int)hw_atomic_sub(lckcnt, 1); + updated = os_atomic_dec(lckcnt, relaxed); assert(updated >= 0); } @@ -467,7 +459,7 @@ void lck_attr_setdebug( lck_attr_t *attr) { - (void)hw_atomic_or(&attr->lck_attr_val, LCK_ATTR_DEBUG); + os_atomic_or(&attr->lck_attr_val, LCK_ATTR_DEBUG, relaxed); } /* @@ -477,7 +469,7 @@ void lck_attr_cleardebug( lck_attr_t *attr) { - (void)hw_atomic_and(&attr->lck_attr_val, ~LCK_ATTR_DEBUG); + os_atomic_andnot(&attr->lck_attr_val, LCK_ATTR_DEBUG, relaxed); } @@ -488,7 +480,7 @@ void lck_attr_rw_shared_priority( lck_attr_t *attr) { - (void)hw_atomic_or(&attr->lck_attr_val, LCK_ATTR_RW_SHARED_PRIORITY); + os_atomic_or(&attr->lck_attr_val, LCK_ATTR_RW_SHARED_PRIORITY, relaxed); } @@ -513,6 +505,31 @@ hw_lock_init(hw_lock_t lock) ordered_store_hw(lock, 0); } +#if __SMP__ +static inline bool +hw_lock_trylock_contended(hw_lock_t lock, uintptr_t newval) +{ +#if OS_ATOMIC_USE_LLSC + uintptr_t oldval; + os_atomic_rmw_loop(&lock->lock_data, oldval, newval, acquire, { + if (oldval != 0) { + wait_for_event(); // clears the monitor so we don't need give_up() + return false; + } + }); + return true; +#else // !OS_ATOMIC_USE_LLSC +#if OS_ATOMIC_HAS_LLSC + uintptr_t oldval = os_atomic_load_exclusive(&lock->lock_data, relaxed); + if (oldval != 0) { + wait_for_event(); // clears the monitor so we don't need give_up() + return false; + } +#endif // OS_ATOMIC_HAS_LLSC + return os_atomic_cmpxchg(&lock->lock_data, 0, newval, acquire); +#endif // !OS_ATOMIC_USE_LLSC +} + /* * Routine: hw_lock_lock_contended * @@ -520,8 +537,6 @@ hw_lock_init(hw_lock_t lock) * timeout is in mach_absolute_time ticks. Called with * preemption disabled. */ - -#if __SMP__ static unsigned int NOINLINE hw_lock_lock_contended(hw_lock_t lock, uintptr_t data, uint64_t timeout, boolean_t do_panic LCK_GRP_ARG(lck_grp_t *grp)) { @@ -551,8 +566,7 @@ hw_lock_lock_contended(hw_lock_t lock, uintptr_t data, uint64_t timeout, boolean continue; } #endif - if (atomic_compare_exchange(&lock->lock_data, 0, data, - memory_order_acquire_smp, TRUE)) { + if (hw_lock_trylock_contended(lock, data)) { #if CONFIG_DTRACE || LOCK_STATS if (__improbable(stat_enabled)) { lck_grp_spin_update_spin(lock LCK_GRP_ARG(grp), mach_absolute_time() - begin); @@ -578,6 +592,42 @@ hw_lock_lock_contended(hw_lock_t lock, uintptr_t data, uint64_t timeout, boolean } #endif // __SMP__ +void * +hw_wait_while_equals(void **address, void *current) +{ +#if __SMP__ + void *v; + uint64_t end = 0; + + for (;;) { + for (int i = 0; i < LOCK_SNOOP_SPINS; i++) { + cpu_pause(); +#if OS_ATOMIC_HAS_LLSC + v = os_atomic_load_exclusive(address, relaxed); + if (__probable(v != current)) { + os_atomic_clear_exclusive(); + return v; + } + wait_for_event(); +#else + v = os_atomic_load(address, relaxed); + if (__probable(v != current)) { + return v; + } +#endif // OS_ATOMIC_HAS_LLSC + } + if (end == 0) { + end = ml_get_timebase() + LOCK_PANIC_TIMEOUT; + } else if (ml_get_timebase() >= end) { + panic("Wait while equals timeout @ *%p == %p", address, v); + } + } +#else // !__SMP__ + panic("Value at %p is %p", address, current); + __builtin_unreachable(); +#endif // !__SMP__ +} + static inline void hw_lock_lock_internal(hw_lock_t lock, thread_t thread LCK_GRP_ARG(lck_grp_t *grp)) { @@ -585,14 +635,12 @@ hw_lock_lock_internal(hw_lock_t lock, thread_t thread LCK_GRP_ARG(lck_grp_t *grp state = LCK_MTX_THREAD_TO_STATE(thread) | PLATFORM_LCK_ILOCK; #if __SMP__ - #if LOCK_PRETEST if (ordered_load_hw(lock)) { goto contended; } #endif // LOCK_PRETEST - if (atomic_compare_exchange(&lock->lock_data, 0, state, - memory_order_acquire_smp, TRUE)) { + if (hw_lock_trylock_contended(lock, state)) { goto end; } #if LOCK_PRETEST @@ -659,14 +707,12 @@ int disable_preemption_for_thread(thread); state = LCK_MTX_THREAD_TO_STATE(thread) | PLATFORM_LCK_ILOCK; #if __SMP__ - #if LOCK_PRETEST if (ordered_load_hw(lock)) { goto contended; } #endif // LOCK_PRETEST - if (atomic_compare_exchange(&lock->lock_data, 0, state, - memory_order_acquire_smp, TRUE)) { + if (hw_lock_trylock_contended(lock, state)) { success = 1; goto end; } @@ -704,8 +750,8 @@ hw_lock_try_internal(hw_lock_t lock, thread_t thread LCK_GRP_ARG(lck_grp_t *grp) goto failed; } #endif // LOCK_PRETEST - success = atomic_compare_exchange(&lock->lock_data, 0, LCK_MTX_THREAD_TO_STATE(thread) | PLATFORM_LCK_ILOCK, - memory_order_acquire_smp, FALSE); + success = os_atomic_cmpxchg(&lock->lock_data, 0, + LCK_MTX_THREAD_TO_STATE(thread) | PLATFORM_LCK_ILOCK, acquire); #else if (lock->lock_data == 0) { lock->lock_data = LCK_MTX_THREAD_TO_STATE(thread) | PLATFORM_LCK_ILOCK; @@ -754,7 +800,7 @@ int static inline void hw_lock_unlock_internal(hw_lock_t lock) { - __c11_atomic_store((_Atomic uintptr_t *)&lock->lock_data, 0, memory_order_release_smp); + os_atomic_store(&lock->lock_data, 0, release); #if __arm__ || __arm64__ // ARM tests are only for open-source exclusion set_event(); @@ -790,6 +836,198 @@ hw_lock_held(hw_lock_t lock) return ordered_load_hw(lock) != 0; } +#if __SMP__ +static unsigned int +hw_lock_bit_to_contended(hw_lock_bit_t *lock, uint32_t mask, uint32_t timeout LCK_GRP_ARG(lck_grp_t *grp)); +#endif + +static inline unsigned int +hw_lock_bit_to_internal(hw_lock_bit_t *lock, unsigned int bit, uint32_t timeout LCK_GRP_ARG(lck_grp_t *grp)) +{ + unsigned int success = 0; + uint32_t mask = (1 << bit); +#if !__SMP__ + uint32_t state; +#endif + +#if __SMP__ + if (__improbable(!hw_atomic_test_and_set32(lock, mask, mask, memory_order_acquire, FALSE))) { + success = hw_lock_bit_to_contended(lock, mask, timeout LCK_GRP_ARG(grp)); + } else { + success = 1; + } +#else // __SMP__ + (void)timeout; + state = ordered_load_bit(lock); + if (!(mask & state)) { + ordered_store_bit(lock, state | mask); + success = 1; + } +#endif // __SMP__ + + if (success) { + lck_grp_spin_update_held(lock LCK_GRP_ARG(grp)); + } + + return success; +} + +unsigned +int +(hw_lock_bit_to)(hw_lock_bit_t * lock, unsigned int bit, uint32_t timeout LCK_GRP_ARG(lck_grp_t *grp)) +{ + _disable_preemption(); + return hw_lock_bit_to_internal(lock, bit, timeout LCK_GRP_ARG(grp)); +} + +#if __SMP__ +static unsigned int NOINLINE +hw_lock_bit_to_contended(hw_lock_bit_t *lock, uint32_t mask, uint32_t timeout LCK_GRP_ARG(lck_grp_t *grp)) +{ + uint64_t end = 0; + int i; +#if CONFIG_DTRACE || LOCK_STATS + uint64_t begin = 0; + boolean_t stat_enabled = lck_grp_spin_spin_enabled(lock LCK_GRP_ARG(grp)); +#endif /* CONFIG_DTRACE || LOCK_STATS */ + +#if LOCK_STATS || CONFIG_DTRACE + if (__improbable(stat_enabled)) { + begin = mach_absolute_time(); + } +#endif /* LOCK_STATS || CONFIG_DTRACE */ + for (;;) { + for (i = 0; i < LOCK_SNOOP_SPINS; i++) { + // Always load-exclusive before wfe + // This grabs the monitor and wakes up on a release event + if (hw_atomic_test_and_set32(lock, mask, mask, memory_order_acquire, TRUE)) { + goto end; + } + } + if (end == 0) { + end = ml_get_timebase() + timeout; + } else if (ml_get_timebase() >= end) { + break; + } + } + return 0; +end: +#if CONFIG_DTRACE || LOCK_STATS + if (__improbable(stat_enabled)) { + lck_grp_spin_update_spin(lock LCK_GRP_ARG(grp), mach_absolute_time() - begin); + } + lck_grp_spin_update_miss(lock LCK_GRP_ARG(grp)); +#endif /* CONFIG_DTRACE || LCK_GRP_STAT */ + + return 1; +} +#endif // __SMP__ + +void +(hw_lock_bit)(hw_lock_bit_t * lock, unsigned int bit LCK_GRP_ARG(lck_grp_t *grp)) +{ + if (hw_lock_bit_to(lock, bit, LOCK_PANIC_TIMEOUT, LCK_GRP_PROBEARG(grp))) { + return; + } +#if __SMP__ + panic("hw_lock_bit(): timed out (%p)", lock); +#else + panic("hw_lock_bit(): interlock held (%p)", lock); +#endif +} + +void +(hw_lock_bit_nopreempt)(hw_lock_bit_t * lock, unsigned int bit LCK_GRP_ARG(lck_grp_t *grp)) +{ + if (__improbable(get_preemption_level() == 0)) { + panic("Attempt to take no-preempt bitlock %p in preemptible context", lock); + } + if (hw_lock_bit_to_internal(lock, bit, LOCK_PANIC_TIMEOUT LCK_GRP_ARG(grp))) { + return; + } +#if __SMP__ + panic("hw_lock_bit_nopreempt(): timed out (%p)", lock); +#else + panic("hw_lock_bit_nopreempt(): interlock held (%p)", lock); +#endif +} + +unsigned +int +(hw_lock_bit_try)(hw_lock_bit_t * lock, unsigned int bit LCK_GRP_ARG(lck_grp_t *grp)) +{ + uint32_t mask = (1 << bit); +#if !__SMP__ + uint32_t state; +#endif + boolean_t success = FALSE; + + _disable_preemption(); +#if __SMP__ + // TODO: consider weak (non-looping) atomic test-and-set + success = hw_atomic_test_and_set32(lock, mask, mask, memory_order_acquire, FALSE); +#else + state = ordered_load_bit(lock); + if (!(mask & state)) { + ordered_store_bit(lock, state | mask); + success = TRUE; + } +#endif // __SMP__ + if (!success) { + _enable_preemption(); + } + + if (success) { + lck_grp_spin_update_held(lock LCK_GRP_ARG(grp)); + } + + return success; +} + +static inline void +hw_unlock_bit_internal(hw_lock_bit_t *lock, unsigned int bit) +{ + uint32_t mask = (1 << bit); +#if !__SMP__ + uint32_t state; +#endif + +#if __SMP__ + os_atomic_andnot(lock, mask, release); +#if __arm__ + set_event(); +#endif +#else // __SMP__ + state = ordered_load_bit(lock); + ordered_store_bit(lock, state & ~mask); +#endif // __SMP__ +#if CONFIG_DTRACE + LOCKSTAT_RECORD(LS_LCK_SPIN_UNLOCK_RELEASE, lock, bit); +#endif +} + +/* + * Routine: hw_unlock_bit + * + * Release spin-lock. The second parameter is the bit number to test and set. + * Decrement the preemption level. + */ +void +hw_unlock_bit(hw_lock_bit_t * lock, unsigned int bit) +{ + hw_unlock_bit_internal(lock, bit); + _enable_preemption(); +} + +void +hw_unlock_bit_nopreempt(hw_lock_bit_t * lock, unsigned int bit) +{ + if (__improbable(get_preemption_level() == 0)) { + panic("Attempt to release no-preempt bitlock %p in preemptible context", lock); + } + hw_unlock_bit_internal(lock, bit); +} + /* * Routine: lck_spin_sleep */ @@ -983,37 +1221,9 @@ lck_mtx_sleep_deadline( * The lock owner is always promoted to the max priority of all its waiters. * Max priority is capped at MAXPRI_PROMOTE. * - * lck_mtx_pri being set implies that the lock owner is promoted to at least lck_mtx_pri - * This prevents the thread from dropping in priority while holding a mutex - * (note: Intel locks currently don't do this, to avoid thread lock churn) - * - * thread->promotions has a +1 for every mutex currently promoting the thread - * and 1 for was_promoted_on_wakeup being set. - * TH_SFLAG_PROMOTED is set on a thread whenever it has any promotions - * from any mutex (i.e. thread->promotions != 0) - * - * was_promoted_on_wakeup is set on a thread which is woken up by a mutex when - * it raises the priority of the woken thread to match lck_mtx_pri. - * It can be set for multiple iterations of wait, fail to acquire, re-wait, etc - * was_promoted_on_wakeup being set always implies a +1 promotions count. - * * The last waiter is not given a promotion when it wakes up or acquires the lock. * When the last waiter is waking up, a new contender can always come in and * steal the lock without having to wait for the last waiter to make forward progress. - * - * lck_mtx_waiters has a +1 for every waiter currently between wait and acquire - * This prevents us from asserting that every wakeup wakes up a thread. - * This also causes excess thread_wakeup calls in the unlock path. - * It can only be fooled into thinking there are more waiters than are - * actually blocked, not less. - * It does allows us to reduce the complexity of the lock state. - * - * This also means that a starved bg thread as the last waiter could end up - * keeping the lock in the contended state for a long period of time, which - * may keep lck_mtx_pri artificially high for a very long time even though - * it is not participating or blocking anyone else. - * Intel locks don't have this problem because they can go uncontended - * as soon as there are no blocked threads involved. */ /* @@ -1034,9 +1244,10 @@ lck_mtx_sleep_deadline( void lck_mtx_lock_wait( lck_mtx_t *lck, - thread_t holder) + thread_t holder, + struct turnstile **ts) { - thread_t self = current_thread(); + thread_t thread = current_thread(); lck_mtx_t *mutex; __kdebug_only uintptr_t trace_lck = unslide_for_kdebug(lck); @@ -1057,64 +1268,27 @@ lck_mtx_lock_wait( KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_MTX_LCK_WAIT_CODE) | DBG_FUNC_START, trace_lck, (uintptr_t)thread_tid(thread), 0, 0, 0); - spl_t s = splsched(); - thread_lock(holder); - - assert_promotions_invariant(holder); - - if ((holder->sched_flags & TH_SFLAG_DEPRESS) == 0) { - assert(holder->sched_pri >= mutex->lck_mtx_pri); - } - - integer_t priority = self->sched_pri; - priority = MAX(priority, self->base_pri); - priority = MAX(priority, BASEPRI_DEFAULT); - priority = MIN(priority, MAXPRI_PROMOTE); - - if (mutex->lck_mtx_pri == 0) { - /* This is the first promotion for this mutex */ - if (holder->promotions++ == 0) { - /* This is the first promotion for holder */ - sched_thread_promote_to_pri(holder, priority, trace_lck); - } else { - /* Holder was previously promoted due to a different mutex, raise to match this one */ - sched_thread_update_promotion_to_pri(holder, priority, trace_lck); - } - } else { - /* Holder was previously promoted due to this mutex, check if the pri needs to go up */ - sched_thread_update_promotion_to_pri(holder, priority, trace_lck); - } - - assert(holder->promotions > 0); - assert(holder->promotion_priority >= priority); - - if ((holder->sched_flags & TH_SFLAG_DEPRESS) == 0) { - assert(holder->sched_pri >= mutex->lck_mtx_pri); - } - - assert_promotions_invariant(holder); - - thread_unlock(holder); - splx(s); + assert(thread->waiting_for_mutex == NULL); + thread->waiting_for_mutex = mutex; + mutex->lck_mtx_waiters++; - if (mutex->lck_mtx_pri < priority) { - mutex->lck_mtx_pri = priority; + if (*ts == NULL) { + *ts = turnstile_prepare((uintptr_t)mutex, NULL, TURNSTILE_NULL, TURNSTILE_KERNEL_MUTEX); } - if (self->waiting_for_mutex == NULL) { - self->waiting_for_mutex = mutex; - mutex->lck_mtx_waiters++; - } + struct turnstile *turnstile = *ts; + thread_set_pending_block_hint(thread, kThreadWaitKernelMutex); + turnstile_update_inheritor(turnstile, holder, (TURNSTILE_DELAYED_UPDATE | TURNSTILE_INHERITOR_THREAD)); - assert(self->waiting_for_mutex == mutex); + waitq_assert_wait64(&turnstile->ts_waitq, CAST_EVENT64_T(LCK_MTX_EVENT(mutex)), THREAD_UNINT | THREAD_WAIT_NOREPORT_USER, TIMEOUT_WAIT_FOREVER); - thread_set_pending_block_hint(self, kThreadWaitKernelMutex); - assert_wait(LCK_MTX_EVENT(mutex), THREAD_UNINT | THREAD_WAIT_NOREPORT_USER); lck_mtx_ilk_unlock(mutex); + turnstile_update_inheritor_complete(turnstile, TURNSTILE_INTERLOCK_NOT_HELD); + thread_block(THREAD_CONTINUE_NULL); - assert(mutex->lck_mtx_waiters > 0); + thread->waiting_for_mutex = NULL; KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_MTX_LCK_WAIT_CODE) | DBG_FUNC_END, 0, 0, 0, 0, 0); #if CONFIG_DTRACE @@ -1146,11 +1320,11 @@ lck_mtx_lock_wait( */ int lck_mtx_lock_acquire( - lck_mtx_t *lck) + lck_mtx_t *lck, + struct turnstile *ts) { thread_t thread = current_thread(); lck_mtx_t *mutex; - integer_t priority; if (lck->lck_mtx_tag != LCK_MTX_TAG_INDIRECT) { mutex = lck; @@ -1158,79 +1332,19 @@ lck_mtx_lock_acquire( mutex = &lck->lck_mtx_ptr->lck_mtx; } - /* - * If waiting_for_mutex is set, then this thread was previously blocked waiting on this lock - * If it's un-set, then this thread stole the lock from another waiter. - */ - if (thread->waiting_for_mutex == mutex) { - assert(mutex->lck_mtx_waiters > 0); - - thread->waiting_for_mutex = NULL; - mutex->lck_mtx_waiters--; - } - assert(thread->waiting_for_mutex == NULL); if (mutex->lck_mtx_waiters > 0) { - priority = mutex->lck_mtx_pri; - } else { - /* I was the last waiter, so the mutex is no longer promoted or contended */ - mutex->lck_mtx_pri = 0; - priority = 0; - } - - if (priority || thread->was_promoted_on_wakeup) { - __kdebug_only uintptr_t trace_lck = unslide_for_kdebug(lck); - - /* - * Note: was_promoted_on_wakeup can happen for multiple wakeups in a row without - * an intervening acquire if a thread keeps failing to acquire the lock - * - * If priority is true but not promoted on wakeup, - * then this is a lock steal of a promoted mutex, so it needs a ++ of promotions. - * - * If promoted on wakeup is true, but priority is not, - * then this is the last owner, and the last owner does not need a promotion. - */ - - spl_t s = splsched(); - thread_lock(thread); - - assert_promotions_invariant(thread); - - if (thread->was_promoted_on_wakeup) { - assert(thread->promotions > 0); + if (ts == NULL) { + ts = turnstile_prepare((uintptr_t)mutex, NULL, TURNSTILE_NULL, TURNSTILE_KERNEL_MUTEX); } - if (priority) { - if (thread->promotions++ == 0) { - /* This is the first promotion for holder */ - sched_thread_promote_to_pri(thread, priority, trace_lck); - } else { - /* - * Holder was previously promoted due to a different mutex, raise to match this one - * Or, this thread was promoted on wakeup but someone else later contended on mutex - * at higher priority before we got here - */ - sched_thread_update_promotion_to_pri(thread, priority, trace_lck); - } - } - - if (thread->was_promoted_on_wakeup) { - thread->was_promoted_on_wakeup = 0; - if (--thread->promotions == 0) { - sched_thread_unpromote(thread, trace_lck); - } - } - - assert_promotions_invariant(thread); - - if (priority && (thread->sched_flags & TH_SFLAG_DEPRESS) == 0) { - assert(thread->sched_pri >= priority); - } + turnstile_update_inheritor(ts, thread, (TURNSTILE_IMMEDIATE_UPDATE | TURNSTILE_INHERITOR_THREAD)); + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_HELD); + } - thread_unlock(thread); - splx(s); + if (ts != NULL) { + turnstile_complete((uintptr_t)mutex, NULL, NULL, TURNSTILE_KERNEL_MUTEX); } return mutex->lck_mtx_waiters; @@ -1243,11 +1357,10 @@ lck_mtx_lock_acquire( * * Called with the interlock locked. * - * TODO: the 'waiters' flag does not indicate waiters exist on the waitqueue, - * it indicates waiters exist between wait and acquire. - * This means that here we may do extra unneeded wakeups. + * NOTE: callers should call turnstile_clenup after + * dropping the interlock. */ -void +boolean_t lck_mtx_unlock_wakeup( lck_mtx_t *lck, thread_t holder) @@ -1255,6 +1368,8 @@ lck_mtx_unlock_wakeup( thread_t thread = current_thread(); lck_mtx_t *mutex; __kdebug_only uintptr_t trace_lck = unslide_for_kdebug(lck); + struct turnstile *ts; + kern_return_t did_wake; if (lck->lck_mtx_tag != LCK_MTX_TAG_INDIRECT) { mutex = lck; @@ -1270,88 +1385,29 @@ lck_mtx_unlock_wakeup( trace_lck, (uintptr_t)thread_tid(thread), 0, 0, 0); assert(mutex->lck_mtx_waiters > 0); - assert(thread->was_promoted_on_wakeup == 0); assert(thread->waiting_for_mutex == NULL); - /* - * The waiters count does not precisely match the number of threads on the waitqueue, - * therefore we cannot assert that we actually wake up a thread here - */ + ts = turnstile_prepare((uintptr_t)mutex, NULL, TURNSTILE_NULL, TURNSTILE_KERNEL_MUTEX); + if (mutex->lck_mtx_waiters > 1) { - thread_wakeup_one_with_pri(LCK_MTX_EVENT(lck), lck->lck_mtx_pri); + /* WAITQ_PROMOTE_ON_WAKE will call turnstile_update_inheritor on the wokenup thread */ + did_wake = waitq_wakeup64_one(&ts->ts_waitq, CAST_EVENT64_T(LCK_MTX_EVENT(mutex)), THREAD_AWAKENED, WAITQ_PROMOTE_ON_WAKE); } else { - thread_wakeup_one(LCK_MTX_EVENT(lck)); + did_wake = waitq_wakeup64_one(&ts->ts_waitq, CAST_EVENT64_T(LCK_MTX_EVENT(mutex)), THREAD_AWAKENED, WAITQ_ALL_PRIORITIES); + turnstile_update_inheritor(ts, NULL, TURNSTILE_IMMEDIATE_UPDATE); } + assert(did_wake == KERN_SUCCESS); - /* When mutex->lck_mtx_pri is set, it means means I as the owner have a promotion. */ - if (mutex->lck_mtx_pri) { - spl_t s = splsched(); - thread_lock(thread); - - assert(thread->promotions > 0); - - assert_promotions_invariant(thread); + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_HELD); + turnstile_complete((uintptr_t)mutex, NULL, NULL, TURNSTILE_KERNEL_MUTEX); - if (--thread->promotions == 0) { - sched_thread_unpromote(thread, trace_lck); - } - - assert_promotions_invariant(thread); - - thread_unlock(thread); - splx(s); - } + mutex->lck_mtx_waiters--; KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_MTX_UNLCK_WAKEUP_CODE) | DBG_FUNC_END, 0, 0, 0, 0, 0); -} - -/* - * Callout from the waitqueue code from inside thread_wakeup_one_with_pri - * At splsched, thread is pulled from waitq, still locked, not on runqueue yet - * - * We always make sure to set the promotion flag, even if the thread is already at this priority, - * so that it doesn't go down. - */ -void -lck_mtx_wakeup_adjust_pri(thread_t thread, integer_t priority) -{ - assert(priority <= MAXPRI_PROMOTE); - assert(thread->waiting_for_mutex != NULL); - - __kdebug_only uintptr_t trace_lck = unslide_for_kdebug(thread->waiting_for_mutex); - - assert_promotions_invariant(thread); - - if (thread->was_promoted_on_wakeup) { - /* Thread was previously promoted, but contended again */ - sched_thread_update_promotion_to_pri(thread, priority, trace_lck); - return; - } - - if (thread->promotions > 0 && priority <= thread->promotion_priority) { - /* - * Thread is already promoted to the right level, no need to do more - * I can draft off of another promotion here, which is OK - * because I know the thread will soon run acquire to get its own promotion - */ - assert((thread->sched_flags & TH_SFLAG_PROMOTED) == TH_SFLAG_PROMOTED); - return; - } - - thread->was_promoted_on_wakeup = 1; - - if (thread->promotions++ == 0) { - /* This is the first promotion for this thread */ - sched_thread_promote_to_pri(thread, priority, trace_lck); - } else { - /* Holder was previously promoted due to a different mutex, raise to match this one */ - sched_thread_update_promotion_to_pri(thread, priority, trace_lck); - } - assert_promotions_invariant(thread); + return mutex->lck_mtx_waiters > 0; } - /* * Routine: mutex_pause * @@ -1703,56 +1759,1495 @@ host_lockgroup_info( } /* - * Atomic primitives, prototyped in kern/simple_lock.h - * Noret versions are more efficient on some architectures + * sleep_with_inheritor and wakeup_with_inheritor KPI + * + * Functions that allow to sleep on an event and use turnstile to propagate the priority of the sleeping threads to + * the latest thread specified as inheritor. + * + * The inheritor management is delegated to the caller, the caller needs to store a thread identifier to provide to this functions to specified upon whom + * direct the push. The inheritor cannot run in user space while holding a push from an event. Therefore is the caller responsibility to call a + * wakeup_with_inheritor from inheritor before running in userspace or specify another inheritor before letting the old inheritor run in userspace. + * + * sleep_with_inheritor requires to hold a locking primitive while invoked, but wakeup_with_inheritor and change_sleep_inheritor don't require it. + * + * Turnstile requires a non blocking primitive as interlock to synchronize the turnstile data structure manipulation, threfore sleep_with_inheritor, change_sleep_inheritor and + * wakeup_with_inheritor will require the same interlock to manipulate turnstiles. + * If sleep_with_inheritor is associated with a locking primitive that can block (like lck_mtx_t or lck_rw_t), an handoff to a non blocking primitive is required before + * invoking any turnstile operation. + * + * All functions will save the turnstile associated with the event on the turnstile kernel hash table and will use the the turnstile kernel hash table bucket + * spinlock as the turnstile interlock. Because we do not want to hold interrupt disabled while holding the bucket interlock a new turnstile kernel hash table + * is instantiated for this KPI to manage the hash without interrupt disabled. + * Also: + * - all events on the system that hash on the same bucket will contend on the same spinlock. + * - every event will have a dedicated wait_queue. + * + * Different locking primitives can be associated with sleep_with_inheritor as long as the primitive_lock() and primitive_unlock() functions are provided to + * sleep_with_inheritor_turnstile to perform the handoff with the bucket spinlock. */ -uint32_t -hw_atomic_add(volatile uint32_t *dest, uint32_t delt) +kern_return_t +wakeup_with_inheritor_and_turnstile_type(event_t event, turnstile_type_t type, wait_result_t result, bool wake_one, lck_wake_action_t action, thread_t *thread_wokenup) { - ALIGN_TEST(dest, uint32_t); - return __c11_atomic_fetch_add(ATOMIC_CAST(uint32_t, dest), delt, memory_order_relaxed) + delt; -} + uint32_t index; + struct turnstile *ts = NULL; + kern_return_t ret = KERN_NOT_WAITING; + int priority; + thread_t wokeup; -uint32_t -hw_atomic_sub(volatile uint32_t *dest, uint32_t delt) -{ - ALIGN_TEST(dest, uint32_t); - return __c11_atomic_fetch_sub(ATOMIC_CAST(uint32_t, dest), delt, memory_order_relaxed) - delt; -} + /* + * the hash bucket spinlock is used as turnstile interlock + */ + turnstile_hash_bucket_lock((uintptr_t)event, &index, type); -uint32_t -hw_atomic_or(volatile uint32_t *dest, uint32_t mask) -{ - ALIGN_TEST(dest, uint32_t); - return __c11_atomic_fetch_or(ATOMIC_CAST(uint32_t, dest), mask, memory_order_relaxed) | mask; -} + ts = turnstile_prepare((uintptr_t)event, NULL, TURNSTILE_NULL, type); + if (wake_one) { + if (action == LCK_WAKE_DEFAULT) { + priority = WAITQ_PROMOTE_ON_WAKE; + } else { + assert(action == LCK_WAKE_DO_NOT_TRANSFER_PUSH); + priority = WAITQ_ALL_PRIORITIES; + } + + /* + * WAITQ_PROMOTE_ON_WAKE will call turnstile_update_inheritor + * if it finds a thread + */ + wokeup = waitq_wakeup64_identify(&ts->ts_waitq, CAST_EVENT64_T(event), result, priority); + if (wokeup != NULL) { + if (thread_wokenup != NULL) { + *thread_wokenup = wokeup; + } else { + thread_deallocate_safe(wokeup); + } + ret = KERN_SUCCESS; + if (action == LCK_WAKE_DO_NOT_TRANSFER_PUSH) { + goto complete; + } + } else { + if (thread_wokenup != NULL) { + *thread_wokenup = NULL; + } + turnstile_update_inheritor(ts, TURNSTILE_INHERITOR_NULL, TURNSTILE_IMMEDIATE_UPDATE); + ret = KERN_NOT_WAITING; + } + } else { + ret = waitq_wakeup64_all(&ts->ts_waitq, CAST_EVENT64_T(event), result, WAITQ_ALL_PRIORITIES); + turnstile_update_inheritor(ts, TURNSTILE_INHERITOR_NULL, TURNSTILE_IMMEDIATE_UPDATE); + } + + /* + * turnstile_update_inheritor_complete could be called while holding the interlock. + * In this case the new inheritor or is null, or is a thread that is just been woken up + * and have not blocked because it is racing with the same interlock used here + * after the wait. + * So there is no chain to update for the new inheritor. + * + * However unless the current thread is the old inheritor, + * old inheritor can be blocked and requires a chain update. + * + * The chain should be short because kernel turnstiles cannot have user turnstiles + * chained after them. + * + * We can anyway optimize this by asking turnstile to tell us + * if old inheritor needs an update and drop the lock + * just in that case. + */ + turnstile_hash_bucket_unlock((uintptr_t)NULL, &index, type, 0); + + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_NOT_HELD); + + turnstile_hash_bucket_lock((uintptr_t)NULL, &index, type); + +complete: + turnstile_complete((uintptr_t)event, NULL, NULL, type); + + turnstile_hash_bucket_unlock((uintptr_t)NULL, &index, type, 0); + + turnstile_cleanup(); + + return ret; +} + +static wait_result_t +sleep_with_inheritor_and_turnstile_type(event_t event, + thread_t inheritor, + wait_interrupt_t interruptible, + uint64_t deadline, + turnstile_type_t type, + void (^primitive_lock)(void), + void (^primitive_unlock)(void)) +{ + wait_result_t ret; + uint32_t index; + struct turnstile *ts = NULL; + + /* + * the hash bucket spinlock is used as turnstile interlock, + * lock it before releasing the primitive lock + */ + turnstile_hash_bucket_lock((uintptr_t)event, &index, type); + + primitive_unlock(); + + ts = turnstile_prepare((uintptr_t)event, NULL, TURNSTILE_NULL, type); + + thread_set_pending_block_hint(current_thread(), kThreadWaitSleepWithInheritor); + /* + * We need TURNSTILE_DELAYED_UPDATE because we will call + * waitq_assert_wait64 after. + */ + turnstile_update_inheritor(ts, inheritor, (TURNSTILE_DELAYED_UPDATE | TURNSTILE_INHERITOR_THREAD)); + + ret = waitq_assert_wait64(&ts->ts_waitq, CAST_EVENT64_T(event), interruptible, deadline); + + turnstile_hash_bucket_unlock((uintptr_t)NULL, &index, type, 0); + + /* + * Update new and old inheritor chains outside the interlock; + */ + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_NOT_HELD); + + if (ret == THREAD_WAITING) { + ret = thread_block(THREAD_CONTINUE_NULL); + } + + turnstile_hash_bucket_lock((uintptr_t)NULL, &index, type); + + turnstile_complete((uintptr_t)event, NULL, NULL, type); + + turnstile_hash_bucket_unlock((uintptr_t)NULL, &index, type, 0); + + turnstile_cleanup(); + + primitive_lock(); + + return ret; +} + +kern_return_t +change_sleep_inheritor_and_turnstile_type(event_t event, + thread_t inheritor, + turnstile_type_t type) +{ + uint32_t index; + struct turnstile *ts = NULL; + kern_return_t ret = KERN_SUCCESS; + /* + * the hash bucket spinlock is used as turnstile interlock + */ + turnstile_hash_bucket_lock((uintptr_t)event, &index, type); + + ts = turnstile_prepare((uintptr_t)event, NULL, TURNSTILE_NULL, type); + + if (!turnstile_has_waiters(ts)) { + ret = KERN_NOT_WAITING; + } + + /* + * We will not call an assert_wait later so use TURNSTILE_IMMEDIATE_UPDATE + */ + turnstile_update_inheritor(ts, inheritor, (TURNSTILE_IMMEDIATE_UPDATE | TURNSTILE_INHERITOR_THREAD)); + + turnstile_hash_bucket_unlock((uintptr_t)NULL, &index, type, 0); + + /* + * update the chains outside the interlock + */ + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_NOT_HELD); + + turnstile_hash_bucket_lock((uintptr_t)NULL, &index, type); + + turnstile_complete((uintptr_t)event, NULL, NULL, type); + + turnstile_hash_bucket_unlock((uintptr_t)NULL, &index, type, 0); + + turnstile_cleanup(); + + return ret; +} + +typedef void (^void_block_void)(void); + +/* + * sleep_with_inheritor functions with lck_mtx_t as locking primitive. + */ + +wait_result_t +lck_mtx_sleep_with_inheritor_and_turnstile_type(lck_mtx_t *lock, lck_sleep_action_t lck_sleep_action, event_t event, thread_t inheritor, wait_interrupt_t interruptible, uint64_t deadline, turnstile_type_t type) +{ + LCK_MTX_ASSERT(lock, LCK_MTX_ASSERT_OWNED); + + if (lck_sleep_action & LCK_SLEEP_UNLOCK) { + return sleep_with_inheritor_and_turnstile_type(event, + inheritor, + interruptible, + deadline, + type, + ^{;}, + ^{lck_mtx_unlock(lock);}); + } else if (lck_sleep_action & LCK_SLEEP_SPIN) { + return sleep_with_inheritor_and_turnstile_type(event, + inheritor, + interruptible, + deadline, + type, + ^{lck_mtx_lock_spin(lock);}, + ^{lck_mtx_unlock(lock);}); + } else if (lck_sleep_action & LCK_SLEEP_SPIN_ALWAYS) { + return sleep_with_inheritor_and_turnstile_type(event, + inheritor, + interruptible, + deadline, + type, + ^{lck_mtx_lock_spin_always(lock);}, + ^{lck_mtx_unlock(lock);}); + } else { + return sleep_with_inheritor_and_turnstile_type(event, + inheritor, + interruptible, + deadline, + type, + ^{lck_mtx_lock(lock);}, + ^{lck_mtx_unlock(lock);}); + } +} + +/* + * Name: lck_spin_sleep_with_inheritor + * + * Description: deschedule the current thread and wait on the waitq associated with event to be woken up. + * While waiting, the sched priority of the waiting thread will contribute to the push of the event that will + * be directed to the inheritor specified. + * An interruptible mode and deadline can be specified to return earlier from the wait. + * + * Args: + * Arg1: lck_spin_t lock used to protect the sleep. The lock will be dropped while sleeping and reaquired before returning according to the sleep action specified. + * Arg2: sleep action. LCK_SLEEP_DEFAULT, LCK_SLEEP_UNLOCK. + * Arg3: event to wait on. + * Arg4: thread to propagate the event push to. + * Arg5: interruptible flag for wait. + * Arg6: deadline for wait. + * + * Conditions: Lock must be held. Returns with the lock held according to the sleep action specified. + * Lock will be dropped while waiting. + * The inheritor specified cannot run in user space until another inheritor is specified for the event or a + * wakeup for the event is called. + * + * Returns: result of the wait. + */ +wait_result_t +lck_spin_sleep_with_inheritor( + lck_spin_t *lock, + lck_sleep_action_t lck_sleep_action, + event_t event, + thread_t inheritor, + wait_interrupt_t interruptible, + uint64_t deadline) +{ + if (lck_sleep_action & LCK_SLEEP_UNLOCK) { + return sleep_with_inheritor_and_turnstile_type(event, inheritor, + interruptible, deadline, TURNSTILE_SLEEP_INHERITOR, + ^{}, ^{ lck_spin_unlock(lock); }); + } else { + return sleep_with_inheritor_and_turnstile_type(event, inheritor, + interruptible, deadline, TURNSTILE_SLEEP_INHERITOR, + ^{ lck_spin_lock(lock); }, ^{ lck_spin_unlock(lock); }); + } +} + +/* + * Name: lck_mtx_sleep_with_inheritor + * + * Description: deschedule the current thread and wait on the waitq associated with event to be woken up. + * While waiting, the sched priority of the waiting thread will contribute to the push of the event that will + * be directed to the inheritor specified. + * An interruptible mode and deadline can be specified to return earlier from the wait. + * + * Args: + * Arg1: lck_mtx_t lock used to protect the sleep. The lock will be dropped while sleeping and reaquired before returning according to the sleep action specified. + * Arg2: sleep action. LCK_SLEEP_DEFAULT, LCK_SLEEP_UNLOCK, LCK_SLEEP_SPIN, LCK_SLEEP_SPIN_ALWAYS. + * Arg3: event to wait on. + * Arg4: thread to propagate the event push to. + * Arg5: interruptible flag for wait. + * Arg6: deadline for wait. + * + * Conditions: Lock must be held. Returns with the lock held according to the sleep action specified. + * Lock will be dropped while waiting. + * The inheritor specified cannot run in user space until another inheritor is specified for the event or a + * wakeup for the event is called. + * + * Returns: result of the wait. + */ +wait_result_t +lck_mtx_sleep_with_inheritor(lck_mtx_t *lock, lck_sleep_action_t lck_sleep_action, event_t event, thread_t inheritor, wait_interrupt_t interruptible, uint64_t deadline) +{ + return lck_mtx_sleep_with_inheritor_and_turnstile_type(lock, lck_sleep_action, event, inheritor, interruptible, deadline, TURNSTILE_SLEEP_INHERITOR); +} + +/* + * sleep_with_inheritor functions with lck_rw_t as locking primitive. + */ + +wait_result_t +lck_rw_sleep_with_inheritor_and_turnstile_type(lck_rw_t *lock, lck_sleep_action_t lck_sleep_action, event_t event, thread_t inheritor, wait_interrupt_t interruptible, uint64_t deadline, turnstile_type_t type) +{ + __block lck_rw_type_t lck_rw_type = LCK_RW_TYPE_EXCLUSIVE; + + LCK_RW_ASSERT(lock, LCK_RW_ASSERT_HELD); + + if (lck_sleep_action & LCK_SLEEP_UNLOCK) { + return sleep_with_inheritor_and_turnstile_type(event, + inheritor, + interruptible, + deadline, + type, + ^{;}, + ^{lck_rw_type = lck_rw_done(lock);}); + } else if (!(lck_sleep_action & (LCK_SLEEP_SHARED | LCK_SLEEP_EXCLUSIVE))) { + return sleep_with_inheritor_and_turnstile_type(event, + inheritor, + interruptible, + deadline, + type, + ^{lck_rw_lock(lock, lck_rw_type);}, + ^{lck_rw_type = lck_rw_done(lock);}); + } else if (lck_sleep_action & LCK_SLEEP_EXCLUSIVE) { + return sleep_with_inheritor_and_turnstile_type(event, + inheritor, + interruptible, + deadline, + type, + ^{lck_rw_lock_exclusive(lock);}, + ^{lck_rw_type = lck_rw_done(lock);}); + } else { + return sleep_with_inheritor_and_turnstile_type(event, + inheritor, + interruptible, + deadline, + type, + ^{lck_rw_lock_shared(lock);}, + ^{lck_rw_type = lck_rw_done(lock);}); + } +} + +/* + * Name: lck_rw_sleep_with_inheritor + * + * Description: deschedule the current thread and wait on the waitq associated with event to be woken up. + * While waiting, the sched priority of the waiting thread will contribute to the push of the event that will + * be directed to the inheritor specified. + * An interruptible mode and deadline can be specified to return earlier from the wait. + * + * Args: + * Arg1: lck_rw_t lock used to protect the sleep. The lock will be dropped while sleeping and reaquired before returning according to the sleep action specified. + * Arg2: sleep action. LCK_SLEEP_DEFAULT, LCK_SLEEP_SHARED, LCK_SLEEP_EXCLUSIVE. + * Arg3: event to wait on. + * Arg4: thread to propagate the event push to. + * Arg5: interruptible flag for wait. + * Arg6: deadline for wait. + * + * Conditions: Lock must be held. Returns with the lock held according to the sleep action specified. + * Lock will be dropped while waiting. + * The inheritor specified cannot run in user space until another inheritor is specified for the event or a + * wakeup for the event is called. + * + * Returns: result of the wait. + */ +wait_result_t +lck_rw_sleep_with_inheritor(lck_rw_t *lock, lck_sleep_action_t lck_sleep_action, event_t event, thread_t inheritor, wait_interrupt_t interruptible, uint64_t deadline) +{ + return lck_rw_sleep_with_inheritor_and_turnstile_type(lock, lck_sleep_action, event, inheritor, interruptible, deadline, TURNSTILE_SLEEP_INHERITOR); +} + +/* + * wakeup_with_inheritor functions are independent from the locking primitive. + */ + +/* + * Name: wakeup_one_with_inheritor + * + * Description: wake up one waiter for event if any. The thread woken up will be the one with the higher sched priority waiting on event. + * The push for the event will be transferred from the last inheritor to the woken up thread if LCK_WAKE_DEFAULT is specified. + * If LCK_WAKE_DO_NOT_TRANSFER_PUSH is specified the push will not be transferred. + * + * Args: + * Arg1: event to wake from. + * Arg2: wait result to pass to the woken up thread. + * Arg3: wake flag. LCK_WAKE_DEFAULT or LCK_WAKE_DO_NOT_TRANSFER_PUSH. + * Arg4: pointer for storing the thread wokenup. + * + * Returns: KERN_NOT_WAITING if no threads were waiting, KERN_SUCCESS otherwise. + * + * Conditions: The new inheritor wokenup cannot run in user space until another inheritor is specified for the event or a + * wakeup for the event is called. + * A reference for the wokenup thread is acquired. + * NOTE: this cannot be called from interrupt context. + */ +kern_return_t +wakeup_one_with_inheritor(event_t event, wait_result_t result, lck_wake_action_t action, thread_t *thread_wokenup) +{ + return wakeup_with_inheritor_and_turnstile_type(event, + TURNSTILE_SLEEP_INHERITOR, + result, + TRUE, + action, + thread_wokenup); +} + +/* + * Name: wakeup_all_with_inheritor + * + * Description: wake up all waiters waiting for event. The old inheritor will lose the push. + * + * Args: + * Arg1: event to wake from. + * Arg2: wait result to pass to the woken up threads. + * + * Returns: KERN_NOT_WAITING if no threads were waiting, KERN_SUCCESS otherwise. + * + * Conditions: NOTE: this cannot be called from interrupt context. + */ +kern_return_t +wakeup_all_with_inheritor(event_t event, wait_result_t result) +{ + return wakeup_with_inheritor_and_turnstile_type(event, + TURNSTILE_SLEEP_INHERITOR, + result, + FALSE, + 0, + NULL); +} + +/* + * change_sleep_inheritor is independent from the locking primitive. + */ + +/* + * Name: change_sleep_inheritor + * + * Description: Redirect the push of the waiting threads of event to the new inheritor specified. + * + * Args: + * Arg1: event to redirect the push. + * Arg2: new inheritor for event. + * + * Returns: KERN_NOT_WAITING if no threads were waiting, KERN_SUCCESS otherwise. + * + * Conditions: In case of success, the new inheritor cannot run in user space until another inheritor is specified for the event or a + * wakeup for the event is called. + * NOTE: this cannot be called from interrupt context. + */ +kern_return_t +change_sleep_inheritor(event_t event, thread_t inheritor) +{ + return change_sleep_inheritor_and_turnstile_type(event, + inheritor, + TURNSTILE_SLEEP_INHERITOR); +} + +void +kdp_sleep_with_inheritor_find_owner(struct waitq * waitq, __unused event64_t event, thread_waitinfo_t * waitinfo) +{ + assert(waitinfo->wait_type == kThreadWaitSleepWithInheritor); + assert(waitq_is_turnstile_queue(waitq)); + waitinfo->owner = 0; + waitinfo->context = 0; + + if (waitq_held(waitq)) { + return; + } + + struct turnstile *turnstile = waitq_to_turnstile(waitq); + assert(turnstile->ts_inheritor_flags & TURNSTILE_INHERITOR_THREAD); + waitinfo->owner = thread_tid(turnstile->ts_inheritor); +} + +typedef void (*void_func_void)(void); + +static kern_return_t +gate_try_close(gate_t *gate) +{ + uintptr_t state; + thread_t holder; + kern_return_t ret; + __assert_only bool waiters; + thread_t thread = current_thread(); + + if (os_atomic_cmpxchg(&gate->gate_data, 0, GATE_THREAD_TO_STATE(thread), acquire)) { + return KERN_SUCCESS; + } + + gate_ilock(gate); + state = ordered_load_gate(gate); + holder = GATE_STATE_TO_THREAD(state); + + if (holder == NULL) { + waiters = gate_has_waiters(state); + assert(waiters == FALSE); + + state = GATE_THREAD_TO_STATE(current_thread()); + state |= GATE_ILOCK; + ordered_store_gate(gate, state); + ret = KERN_SUCCESS; + } else { + if (holder == current_thread()) { + panic("Trying to close a gate already owned by current thread %p", current_thread()); + } + ret = KERN_FAILURE; + } + + gate_iunlock(gate); + return ret; +} + +static void +gate_close(gate_t* gate) +{ + uintptr_t state; + thread_t holder; + __assert_only bool waiters; + thread_t thread = current_thread(); + + if (os_atomic_cmpxchg(&gate->gate_data, 0, GATE_THREAD_TO_STATE(thread), acquire)) { + return; + } + + gate_ilock(gate); + state = ordered_load_gate(gate); + holder = GATE_STATE_TO_THREAD(state); + + if (holder != NULL) { + panic("Closing a gate already owned by %p from current thread %p", holder, current_thread()); + } + + waiters = gate_has_waiters(state); + assert(waiters == FALSE); + + state = GATE_THREAD_TO_STATE(thread); + state |= GATE_ILOCK; + ordered_store_gate(gate, state); + + gate_iunlock(gate); +} + +static void +gate_open_turnstile(gate_t *gate) +{ + struct turnstile *ts = NULL; + + ts = turnstile_prepare((uintptr_t)gate, &gate->turnstile, TURNSTILE_NULL, TURNSTILE_KERNEL_MUTEX); + waitq_wakeup64_all(&ts->ts_waitq, CAST_EVENT64_T(GATE_EVENT(gate)), THREAD_AWAKENED, WAITQ_ALL_PRIORITIES); + turnstile_update_inheritor(ts, TURNSTILE_INHERITOR_NULL, TURNSTILE_IMMEDIATE_UPDATE); + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_HELD); + turnstile_complete((uintptr_t)gate, &gate->turnstile, NULL, TURNSTILE_KERNEL_MUTEX); + /* + * We can do the cleanup while holding the interlock. + * It is ok because: + * 1. current_thread is the previous inheritor and it is running + * 2. new inheritor is NULL. + * => No chain of turnstiles needs to be updated. + */ + turnstile_cleanup(); +} + +static void +gate_open(gate_t *gate) +{ + uintptr_t state; + thread_t holder; + bool waiters; + thread_t thread = current_thread(); + + if (os_atomic_cmpxchg(&gate->gate_data, GATE_THREAD_TO_STATE(thread), 0, release)) { + return; + } + + gate_ilock(gate); + state = ordered_load_gate(gate); + holder = GATE_STATE_TO_THREAD(state); + waiters = gate_has_waiters(state); + + if (holder != thread) { + panic("Opening gate owned by %p from current thread %p", holder, thread); + } + + if (waiters) { + gate_open_turnstile(gate); + } + + state = GATE_ILOCK; + ordered_store_gate(gate, state); + + gate_iunlock(gate); +} + +static kern_return_t +gate_handoff_turnstile(gate_t *gate, + int flags, + thread_t *thread_woken_up, + bool *waiters) +{ + struct turnstile *ts = NULL; + kern_return_t ret = KERN_FAILURE; + thread_t hp_thread; + + ts = turnstile_prepare((uintptr_t)gate, &gate->turnstile, TURNSTILE_NULL, TURNSTILE_KERNEL_MUTEX); + /* + * Wake up the higest priority thread waiting on the gate + */ + hp_thread = waitq_wakeup64_identify(&ts->ts_waitq, CAST_EVENT64_T(GATE_EVENT(gate)), THREAD_AWAKENED, WAITQ_PROMOTE_ON_WAKE); + + if (hp_thread != NULL) { + /* + * In this case waitq_wakeup64_identify has called turnstile_update_inheritor for us + */ + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_HELD); + *thread_woken_up = hp_thread; + *waiters = turnstile_has_waiters(ts); + /* + * Note: hp_thread is the new holder and the new inheritor. + * In case there are no more waiters, it doesn't need to be the inheritor + * and it shouldn't be it by the time it finishes the wait, so that its next open or + * handoff can go through the fast path. + * We could set the inheritor to NULL here, or the new holder itself can set it + * on its way back from the sleep. In the latter case there are more chanses that + * new waiters will come by, avoiding to do the opearation at all. + */ + ret = KERN_SUCCESS; + } else { + /* + * waiters can have been woken up by an interrupt and still not + * have updated gate->waiters, so we couldn't find them on the waitq. + * Update the inheritor to NULL here, so that the current thread can return to userspace + * indipendently from when the interrupted waiters will finish the wait. + */ + if (flags == GATE_HANDOFF_OPEN_IF_NO_WAITERS) { + turnstile_update_inheritor(ts, TURNSTILE_INHERITOR_NULL, TURNSTILE_IMMEDIATE_UPDATE); + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_HELD); + } + // there are no waiters. + ret = KERN_NOT_WAITING; + } + + turnstile_complete((uintptr_t)gate, &gate->turnstile, NULL, TURNSTILE_KERNEL_MUTEX); + + /* + * We can do the cleanup while holding the interlock. + * It is ok because: + * 1. current_thread is the previous inheritor and it is running + * 2. new inheritor is NULL or it is a just wokenup thread that will race acquiring the lock + * of the gate before trying to sleep. + * => No chain of turnstiles needs to be updated. + */ + turnstile_cleanup(); + + return ret; +} + +static kern_return_t +gate_handoff(gate_t *gate, + int flags) +{ + kern_return_t ret; + thread_t new_holder = NULL; + uintptr_t state; + thread_t holder; + bool waiters; + thread_t thread = current_thread(); + + assert(flags == GATE_HANDOFF_OPEN_IF_NO_WAITERS || flags == GATE_HANDOFF_DEFAULT); + + if (flags == GATE_HANDOFF_OPEN_IF_NO_WAITERS) { + if (os_atomic_cmpxchg(&gate->gate_data, GATE_THREAD_TO_STATE(thread), 0, release)) { + //gate opened but there were no waiters, so return KERN_NOT_WAITING. + return KERN_NOT_WAITING; + } + } + + gate_ilock(gate); + state = ordered_load_gate(gate); + holder = GATE_STATE_TO_THREAD(state); + waiters = gate_has_waiters(state); + + if (holder != current_thread()) { + panic("Handing off gate owned by %p from current thread %p", holder, current_thread()); + } + + if (waiters) { + ret = gate_handoff_turnstile(gate, flags, &new_holder, &waiters); + if (ret == KERN_SUCCESS) { + state = GATE_THREAD_TO_STATE(new_holder); + if (waiters) { + state |= GATE_WAITERS; + } + } else { + if (flags == GATE_HANDOFF_OPEN_IF_NO_WAITERS) { + state = 0; + } + } + } else { + if (flags == GATE_HANDOFF_OPEN_IF_NO_WAITERS) { + state = 0; + } + ret = KERN_NOT_WAITING; + } + state |= GATE_ILOCK; + ordered_store_gate(gate, state); + + gate_iunlock(gate); + + if (new_holder) { + thread_deallocate(new_holder); + } + return ret; +} + +static void_func_void +gate_steal_turnstile(gate_t *gate, + thread_t new_inheritor) +{ + struct turnstile *ts = NULL; + + ts = turnstile_prepare((uintptr_t)gate, &gate->turnstile, TURNSTILE_NULL, TURNSTILE_KERNEL_MUTEX); + + turnstile_update_inheritor(ts, new_inheritor, (TURNSTILE_IMMEDIATE_UPDATE | TURNSTILE_INHERITOR_THREAD)); + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_HELD); + turnstile_complete((uintptr_t)gate, &gate->turnstile, NULL, TURNSTILE_KERNEL_MUTEX); + + /* + * turnstile_cleanup might need to update the chain of the old holder. + * This operation should happen without the turnstile interlock held. + */ + return turnstile_cleanup; +} + +static void +gate_steal(gate_t *gate) +{ + uintptr_t state; + thread_t holder; + thread_t thread = current_thread(); + bool waiters; + + void_func_void func_after_interlock_unlock; + + gate_ilock(gate); + state = ordered_load_gate(gate); + holder = GATE_STATE_TO_THREAD(state); + waiters = gate_has_waiters(state); + + assert(holder != NULL); + state = GATE_THREAD_TO_STATE(thread) | GATE_ILOCK; + if (waiters) { + state |= GATE_WAITERS; + ordered_store_gate(gate, state); + func_after_interlock_unlock = gate_steal_turnstile(gate, thread); + gate_iunlock(gate); + + func_after_interlock_unlock(); + } else { + ordered_store_gate(gate, state); + gate_iunlock(gate); + } +} + +static void_func_void +gate_wait_turnstile(gate_t *gate, + wait_interrupt_t interruptible, + uint64_t deadline, + thread_t holder, + wait_result_t* wait, + bool* waiters) +{ + struct turnstile *ts; + uintptr_t state; + + ts = turnstile_prepare((uintptr_t)gate, &gate->turnstile, TURNSTILE_NULL, TURNSTILE_KERNEL_MUTEX); + + turnstile_update_inheritor(ts, holder, (TURNSTILE_DELAYED_UPDATE | TURNSTILE_INHERITOR_THREAD)); + waitq_assert_wait64(&ts->ts_waitq, CAST_EVENT64_T(GATE_EVENT(gate)), interruptible, deadline); + + gate_iunlock(gate); + + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_NOT_HELD); + + *wait = thread_block(THREAD_CONTINUE_NULL); + + gate_ilock(gate); + + *waiters = turnstile_has_waiters(ts); + + if (!*waiters) { + /* + * We want to enable the fast path as soon as we see that there are no more waiters. + * On the fast path the holder will not do any turnstile operations. + * Set the inheritor as NULL here. + * + * NOTE: if it was an open operation that woke this thread up, the inheritor has + * already been set to NULL. + */ + state = ordered_load_gate(gate); + holder = GATE_STATE_TO_THREAD(state); + if (holder && + ((*wait != THREAD_AWAKENED) || // thread interrupted or timedout + holder == current_thread())) { // thread was woken up and it is the new holder + turnstile_update_inheritor(ts, TURNSTILE_INHERITOR_NULL, TURNSTILE_IMMEDIATE_UPDATE); + turnstile_update_inheritor_complete(ts, TURNSTILE_INTERLOCK_NOT_HELD); + } + } + + turnstile_complete((uintptr_t)gate, &gate->turnstile, NULL, TURNSTILE_KERNEL_MUTEX); + + /* + * turnstile_cleanup might need to update the chain of the old holder. + * This operation should happen without the turnstile primitive interlock held. + */ + return turnstile_cleanup; +} + +static gate_wait_result_t +gate_wait(gate_t* gate, + wait_interrupt_t interruptible, + uint64_t deadline, + void (^primitive_unlock)(void), + void (^primitive_lock)(void)) +{ + gate_wait_result_t ret; + void_func_void func_after_interlock_unlock; + wait_result_t wait_result; + uintptr_t state; + thread_t holder; + bool waiters; + + + gate_ilock(gate); + state = ordered_load_gate(gate); + holder = GATE_STATE_TO_THREAD(state); + + if (holder == NULL) { + panic("Trying to wait on open gate thread %p gate %p", current_thread(), gate); + } + + state |= GATE_WAITERS; + ordered_store_gate(gate, state); + + /* + * Release the primitive lock before any + * turnstile operation. Turnstile + * does not support a blocking primitive as + * interlock. + * + * In this way, concurrent threads will be + * able to acquire the primitive lock + * but still will wait for me through the + * gate interlock. + */ + primitive_unlock(); + + func_after_interlock_unlock = gate_wait_turnstile( gate, + interruptible, + deadline, + holder, + &wait_result, + &waiters); + + state = ordered_load_gate(gate); + holder = GATE_STATE_TO_THREAD(state); + + switch (wait_result) { + case THREAD_INTERRUPTED: + case THREAD_TIMED_OUT: + assert(holder != current_thread()); + + if (waiters) { + state |= GATE_WAITERS; + } else { + state &= ~GATE_WAITERS; + } + ordered_store_gate(gate, state); + + if (wait_result == THREAD_INTERRUPTED) { + ret = GATE_INTERRUPTED; + } else { + ret = GATE_TIMED_OUT; + } + break; + default: + /* + * Note it is possible that even if the gate was handed off to + * me, someone called gate_steal() before I woke up. + * + * As well as it is possible that the gate was opened, but someone + * closed it while I was waking up. + * + * In both cases we return GATE_OPENED, as the gate was opened to me + * at one point, it is the caller responsibility to check again if + * the gate is open. + */ + if (holder == current_thread()) { + ret = GATE_HANDOFF; + } else { + ret = GATE_OPENED; + } + break; + } + + gate_iunlock(gate); + + /* + * turnstile func that needs to be executed without + * holding the primitive interlock + */ + func_after_interlock_unlock(); + + primitive_lock(); + + return ret; +} +static void +gate_assert(gate_t *gate, int flags) +{ + uintptr_t state; + thread_t holder; + + gate_ilock(gate); + state = ordered_load_gate(gate); + holder = GATE_STATE_TO_THREAD(state); + + switch (flags) { + case GATE_ASSERT_CLOSED: + assert(holder != NULL); + break; + case GATE_ASSERT_OPEN: + assert(holder == NULL); + break; + case GATE_ASSERT_HELD: + assert(holder == current_thread()); + break; + default: + panic("invalid %s flag %d", __func__, flags); + } + + gate_iunlock(gate); +} + +static void +gate_init(gate_t *gate) +{ + gate->gate_data = 0; + gate->turnstile = NULL; +} + +static void +gate_destroy(__assert_only gate_t *gate) +{ + assert(gate->gate_data == 0); + assert(gate->turnstile == NULL); +} + +/* + * Name: lck_rw_gate_init + * + * Description: initializes a variable declared with decl_lck_rw_gate_data. + * + * Args: + * Arg1: lck_rw_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_rw_gate_data. + */ +void +lck_rw_gate_init(lck_rw_t *lock, gate_t *gate) +{ + (void) lock; + gate_init(gate); +} + +/* + * Name: lck_rw_gate_destroy + * + * Description: destroys a variable previously initialized. + * + * Args: + * Arg1: lck_rw_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_rw_gate_data. + */ +void +lck_rw_gate_destroy(lck_rw_t *lock, gate_t *gate) +{ + (void) lock; + gate_destroy(gate); +} + +/* + * Name: lck_rw_gate_try_close + * + * Description: Tries to close the gate. + * In case of success the current thread will be set as + * the holder of the gate. + * + * Args: + * Arg1: lck_rw_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_rw_gate_data. + * + * Conditions: Lock must be held. Returns with the lock held. + * + * Returns: + * KERN_SUCCESS in case the gate was successfully closed. The current thread is the new holder + * of the gate. + * A matching lck_rw_gate_open() or lck_rw_gate_handoff() needs to be called later on + * to wake up possible waiters on the gate before returning to userspace. + * If the intent is to conditionally probe the gate before waiting, the lock must not be dropped + * between the calls to lck_rw_gate_try_close() and lck_rw_gate_wait(). + * + * KERN_FAILURE in case the gate was already closed. Will panic if the current thread was already the holder of the gate. + * lck_rw_gate_wait() should be called instead if the intent is to unconditionally wait on this gate. + * The calls to lck_rw_gate_try_close() and lck_rw_gate_wait() should + * be done without dropping the lock that is protecting the gate in between. + */ +int +lck_rw_gate_try_close(__assert_only lck_rw_t *lock, gate_t *gate) +{ + LCK_RW_ASSERT(lock, LCK_RW_ASSERT_HELD); + + return gate_try_close(gate); +} + +/* + * Name: lck_rw_gate_close + * + * Description: Closes the gate. The current thread will be set as + * the holder of the gate. Will panic if the gate is already closed. + * A matching lck_rw_gate_open() or lck_rw_gate_handoff() needs to be called later on + * to wake up possible waiters on the gate before returning to userspace. + * + * Args: + * Arg1: lck_rw_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_rw_gate_data. + * + * Conditions: Lock must be held. Returns with the lock held. + * The gate must be open. + * + */ +void +lck_rw_gate_close(__assert_only lck_rw_t *lock, gate_t *gate) +{ + LCK_RW_ASSERT(lock, LCK_RW_ASSERT_HELD); + + return gate_close(gate); +} + +/* + * Name: lck_rw_gate_open + * + * Description: Opens the gate and wakes up possible waiters. + * + * Args: + * Arg1: lck_rw_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_rw_gate_data. + * + * Conditions: Lock must be held. Returns with the lock held. + * The current thread must be the holder of the gate. + * + */ +void +lck_rw_gate_open(__assert_only lck_rw_t *lock, gate_t *gate) +{ + LCK_RW_ASSERT(lock, LCK_RW_ASSERT_HELD); + + gate_open(gate); +} + +/* + * Name: lck_rw_gate_handoff + * + * Description: Tries to transfer the ownership of the gate. The waiter with highest sched + * priority will be selected as the new holder of the gate, and woken up, + * with the gate remaining in the closed state throughout. + * If no waiters are present, the gate will be kept closed and KERN_NOT_WAITING + * will be returned. + * GATE_HANDOFF_OPEN_IF_NO_WAITERS flag can be used to specify if the gate should be opened in + * case no waiters were found. + * + * + * Args: + * Arg1: lck_rw_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_rw_gate_data. + * Arg3: flags - GATE_HANDOFF_DEFAULT or GATE_HANDOFF_OPEN_IF_NO_WAITERS + * + * Conditions: Lock must be held. Returns with the lock held. + * The current thread must be the holder of the gate. + * + * Returns: + * KERN_SUCCESS in case one of the waiters became the new holder. + * KERN_NOT_WAITING in case there were no waiters. + * + */ +kern_return_t +lck_rw_gate_handoff(__assert_only lck_rw_t *lock, gate_t *gate, int flags) +{ + LCK_RW_ASSERT(lock, LCK_RW_ASSERT_HELD); + + return gate_handoff(gate, flags); +} + +/* + * Name: lck_rw_gate_steal + * + * Description: Set the current ownership of the gate. It sets the current thread as the + * new holder of the gate. + * A matching lck_rw_gate_open() or lck_rw_gate_handoff() needs to be called later on + * to wake up possible waiters on the gate before returning to userspace. + * NOTE: the previous holder should not call lck_rw_gate_open() or lck_rw_gate_handoff() + * anymore. + * + * + * Args: + * Arg1: lck_rw_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_rw_gate_data. + * + * Conditions: Lock must be held. Returns with the lock held. + * The gate must be closed and the current thread must not already be the holder. + * + */ +void +lck_rw_gate_steal(__assert_only lck_rw_t *lock, gate_t *gate) +{ + LCK_RW_ASSERT(lock, LCK_RW_ASSERT_HELD); + + gate_steal(gate); +} + +/* + * Name: lck_rw_gate_wait + * + * Description: Waits for the current thread to become the holder of the gate or for the + * gate to become open. An interruptible mode and deadline can be specified + * to return earlier from the wait. + * + * Args: + * Arg1: lck_rw_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_rw_gate_data. + * Arg3: sleep action. LCK_SLEEP_DEFAULT, LCK_SLEEP_SHARED, LCK_SLEEP_EXCLUSIVE. + * Arg3: interruptible flag for wait. + * Arg4: deadline + * + * Conditions: Lock must be held. Returns with the lock held according to the sleep action specified. + * Lock will be dropped while waiting. + * The gate must be closed. + * + * Returns: Reason why the thread was woken up. + * GATE_HANDOFF - the current thread was handed off the ownership of the gate. + * A matching lck_rw_gate_open() or lck_rw_gate_handoff() needs to be called later on + * to wake up possible waiters on the gate before returning to userspace. + * GATE_OPENED - the gate was opened by the holder. + * GATE_TIMED_OUT - the thread was woken up by a timeout. + * GATE_INTERRUPTED - the thread was interrupted while sleeping. + * + */ +gate_wait_result_t +lck_rw_gate_wait(lck_rw_t *lock, gate_t *gate, lck_sleep_action_t lck_sleep_action, wait_interrupt_t interruptible, uint64_t deadline) +{ + __block lck_rw_type_t lck_rw_type = LCK_RW_TYPE_EXCLUSIVE; + + LCK_RW_ASSERT(lock, LCK_RW_ASSERT_HELD); + + if (lck_sleep_action & LCK_SLEEP_UNLOCK) { + return gate_wait(gate, + interruptible, + deadline, + ^{lck_rw_type = lck_rw_done(lock);}, + ^{;}); + } else if (!(lck_sleep_action & (LCK_SLEEP_SHARED | LCK_SLEEP_EXCLUSIVE))) { + return gate_wait(gate, + interruptible, + deadline, + ^{lck_rw_type = lck_rw_done(lock);}, + ^{lck_rw_lock(lock, lck_rw_type);}); + } else if (lck_sleep_action & LCK_SLEEP_EXCLUSIVE) { + return gate_wait(gate, + interruptible, + deadline, + ^{lck_rw_type = lck_rw_done(lock);}, + ^{lck_rw_lock_exclusive(lock);}); + } else { + return gate_wait(gate, + interruptible, + deadline, + ^{lck_rw_type = lck_rw_done(lock);}, + ^{lck_rw_lock_shared(lock);}); + } +} + +/* + * Name: lck_rw_gate_assert + * + * Description: asserts that the gate is in the specified state. + * + * Args: + * Arg1: lck_rw_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_rw_gate_data. + * Arg3: flags to specified assert type. + * GATE_ASSERT_CLOSED - the gate is currently closed + * GATE_ASSERT_OPEN - the gate is currently opened + * GATE_ASSERT_HELD - the gate is currently closed and the current thread is the holder + */ +void +lck_rw_gate_assert(__assert_only lck_rw_t *lock, gate_t *gate, int flags) +{ + LCK_RW_ASSERT(lock, LCK_RW_ASSERT_HELD); + + gate_assert(gate, flags); + return; +} + +/* + * Name: lck_mtx_gate_init + * + * Description: initializes a variable declared with decl_lck_mtx_gate_data. + * + * Args: + * Arg1: lck_mtx_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_mtx_gate_data. + */ +void +lck_mtx_gate_init(lck_mtx_t *lock, gate_t *gate) +{ + (void) lock; + gate_init(gate); +} + +/* + * Name: lck_mtx_gate_destroy + * + * Description: destroys a variable previously initialized + * + * Args: + * Arg1: lck_mtx_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_mtx_gate_data. + */ +void +lck_mtx_gate_destroy(lck_mtx_t *lock, gate_t *gate) +{ + (void) lock; + gate_destroy(gate); +} + +/* + * Name: lck_mtx_gate_try_close + * + * Description: Tries to close the gate. + * In case of success the current thread will be set as + * the holder of the gate. + * + * Args: + * Arg1: lck_mtx_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_mtx_gate_data. + * + * Conditions: Lock must be held. Returns with the lock held. + * + * Returns: + * KERN_SUCCESS in case the gate was successfully closed. The current thread is the new holder + * of the gate. + * A matching lck_mtx_gate_open() or lck_mtx_gate_handoff() needs to be called later on + * to wake up possible waiters on the gate before returning to userspace. + * If the intent is to conditionally probe the gate before waiting, the lock must not be dropped + * between the calls to lck_mtx_gate_try_close() and lck_mtx_gate_wait(). + * + * KERN_FAILURE in case the gate was already closed. Will panic if the current thread was already the holder of the gate. + * lck_mtx_gate_wait() should be called instead if the intent is to unconditionally wait on this gate. + * The calls to lck_mtx_gate_try_close() and lck_mtx_gate_wait() should + * be done without dropping the lock that is protecting the gate in between. + */ +int +lck_mtx_gate_try_close(__assert_only lck_mtx_t *lock, gate_t *gate) +{ + LCK_MTX_ASSERT(lock, LCK_MTX_ASSERT_OWNED); + + return gate_try_close(gate); +} + +/* + * Name: lck_mtx_gate_close + * + * Description: Closes the gate. The current thread will be set as + * the holder of the gate. Will panic if the gate is already closed. + * A matching lck_mtx_gate_open() or lck_mtx_gate_handoff() needs to be called later on + * to wake up possible waiters on the gate before returning to userspace. + * + * Args: + * Arg1: lck_mtx_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_mtx_gate_data. + * + * Conditions: Lock must be held. Returns with the lock held. + * The gate must be open. + * + */ void -hw_atomic_or_noret(volatile uint32_t *dest, uint32_t mask) +lck_mtx_gate_close(__assert_only lck_mtx_t *lock, gate_t *gate) { - ALIGN_TEST(dest, uint32_t); - __c11_atomic_fetch_or(ATOMIC_CAST(uint32_t, dest), mask, memory_order_relaxed); + LCK_MTX_ASSERT(lock, LCK_MTX_ASSERT_OWNED); + + return gate_close(gate); +} + +/* + * Name: lck_mtx_gate_open + * + * Description: Opens of the gate and wakes up possible waiters. + * + * Args: + * Arg1: lck_mtx_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_mtx_gate_data. + * + * Conditions: Lock must be held. Returns with the lock held. + * The current thread must be the holder of the gate. + * + */ +void +lck_mtx_gate_open(__assert_only lck_mtx_t *lock, gate_t *gate) +{ + LCK_MTX_ASSERT(lock, LCK_MTX_ASSERT_OWNED); + + gate_open(gate); } -uint32_t -hw_atomic_and(volatile uint32_t *dest, uint32_t mask) +/* + * Name: lck_mtx_gate_handoff + * + * Description: Set the current ownership of the gate. The waiter with highest sched + * priority will be selected as the new holder of the gate, and woken up, + * with the gate remaining in the closed state throughout. + * If no waiters are present, the gate will be kept closed and KERN_NOT_WAITING + * will be returned. + * OPEN_ON_FAILURE flag can be used to specify if the gate should be opened in + * case no waiters were found. + * + * + * Args: + * Arg1: lck_mtx_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_mtx_gate_data. + * Arg3: flags - GATE_NO_FALGS or OPEN_ON_FAILURE + * + * Conditions: Lock must be held. Returns with the lock held. + * The current thread must be the holder of the gate. + * + * Returns: + * KERN_SUCCESS in case one of the waiters became the new holder. + * KERN_NOT_WAITING in case there were no waiters. + * + */ +kern_return_t +lck_mtx_gate_handoff(__assert_only lck_mtx_t *lock, gate_t *gate, int flags) { - ALIGN_TEST(dest, uint32_t); - return __c11_atomic_fetch_and(ATOMIC_CAST(uint32_t, dest), mask, memory_order_relaxed) & mask; + LCK_MTX_ASSERT(lock, LCK_MTX_ASSERT_OWNED); + + return gate_handoff(gate, flags); } +/* + * Name: lck_mtx_gate_steal + * + * Description: Steals the ownership of the gate. It sets the current thread as the + * new holder of the gate. + * A matching lck_mtx_gate_open() or lck_mtx_gate_handoff() needs to be called later on + * to wake up possible waiters on the gate before returning to userspace. + * NOTE: the previous holder should not call lck_mtx_gate_open() or lck_mtx_gate_handoff() + * anymore. + * + * + * Args: + * Arg1: lck_mtx_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_mtx_gate_data. + * + * Conditions: Lock must be held. Returns with the lock held. + * The gate must be closed and the current thread must not already be the holder. + * + */ void -hw_atomic_and_noret(volatile uint32_t *dest, uint32_t mask) +lck_mtx_gate_steal(__assert_only lck_mtx_t *lock, gate_t *gate) +{ + LCK_MTX_ASSERT(lock, LCK_MTX_ASSERT_OWNED); + + gate_steal(gate); +} + +/* + * Name: lck_mtx_gate_wait + * + * Description: Waits for the current thread to become the holder of the gate or for the + * gate to become open. An interruptible mode and deadline can be specified + * to return earlier from the wait. + * + * Args: + * Arg1: lck_mtx_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_mtx_gate_data. + * Arg3: sleep action. LCK_SLEEP_DEFAULT, LCK_SLEEP_UNLOCK, LCK_SLEEP_SPIN, LCK_SLEEP_SPIN_ALWAYS. + * Arg3: interruptible flag for wait. + * Arg4: deadline + * + * Conditions: Lock must be held. Returns with the lock held according to the sleep action specified. + * Lock will be dropped while waiting. + * The gate must be closed. + * + * Returns: Reason why the thread was woken up. + * GATE_HANDOFF - the current thread was handed off the ownership of the gate. + * A matching lck_mtx_gate_open() or lck_mtx_gate_handoff() needs to be called later on + * to wake up possible waiters on the gate before returning to userspace. + * GATE_OPENED - the gate was opened by the holder. + * GATE_TIMED_OUT - the thread was woken up by a timeout. + * GATE_INTERRUPTED - the thread was interrupted while sleeping. + * + */ +gate_wait_result_t +lck_mtx_gate_wait(lck_mtx_t *lock, gate_t *gate, lck_sleep_action_t lck_sleep_action, wait_interrupt_t interruptible, uint64_t deadline) { - ALIGN_TEST(dest, uint32_t); - __c11_atomic_fetch_and(ATOMIC_CAST(uint32_t, dest), mask, memory_order_relaxed); + LCK_MTX_ASSERT(lock, LCK_MTX_ASSERT_OWNED); + + if (lck_sleep_action & LCK_SLEEP_UNLOCK) { + return gate_wait(gate, + interruptible, + deadline, + ^{lck_mtx_unlock(lock);}, + ^{;}); + } else if (lck_sleep_action & LCK_SLEEP_SPIN) { + return gate_wait(gate, + interruptible, + deadline, + ^{lck_mtx_unlock(lock);}, + ^{lck_mtx_lock_spin(lock);}); + } else if (lck_sleep_action & LCK_SLEEP_SPIN_ALWAYS) { + return gate_wait(gate, + interruptible, + deadline, + ^{lck_mtx_unlock(lock);}, + ^{lck_mtx_lock_spin_always(lock);}); + } else { + return gate_wait(gate, + interruptible, + deadline, + ^{lck_mtx_unlock(lock);}, + ^{lck_mtx_lock(lock);}); + } } -uint32_t -hw_compare_and_store(uint32_t oldval, uint32_t newval, volatile uint32_t *dest) +/* + * Name: lck_mtx_gate_assert + * + * Description: asserts that the gate is in the specified state. + * + * Args: + * Arg1: lck_mtx_t lock used to protect the gate. + * Arg2: pointer to the gate data declared with decl_lck_mtx_gate_data. + * Arg3: flags to specified assert type. + * GATE_ASSERT_CLOSED - the gate is currently closed + * GATE_ASSERT_OPEN - the gate is currently opened + * GATE_ASSERT_HELD - the gate is currently closed and the current thread is the holder + */ +void +lck_mtx_gate_assert(__assert_only lck_mtx_t *lock, gate_t *gate, int flags) { - ALIGN_TEST(dest, uint32_t); - return __c11_atomic_compare_exchange_strong(ATOMIC_CAST(uint32_t, dest), &oldval, newval, - memory_order_acq_rel_smp, memory_order_relaxed); + LCK_MTX_ASSERT(lock, LCK_MTX_ASSERT_OWNED); + + gate_assert(gate, flags); }