/*
- * Copyright (c) 2000-2007 Apple Inc. All rights reserved.
+ * Copyright (c) 2000-2016 Apple Inc. All rights reserved.
*
* @APPLE_OSREFERENCE_LICENSE_HEADER_START@
*
* any improvements or extensions that they make and grant Carnegie Mellon
* the rights to redistribute these changes.
*/
+
+#define ATOMIC_PRIVATE 1
+#define LOCK_PRIVATE 1
+
#include <mach_ldebug.h>
#include <debug.h>
#include <kern/processor.h>
#include <kern/sched_prim.h>
#include <kern/debug.h>
+#include <libkern/section_keywords.h>
+#include <machine/atomic.h>
+#include <machine/machine_cpu.h>
#include <string.h>
-
#include <sys/kdebug.h>
#if CONFIG_DTRACE
#define LCK_MTX_LCK_WAIT_CODE 2
#define LCK_MTX_UNLCK_WAKEUP_CODE 3
+#if MACH_LDEBUG
+#define ALIGN_TEST(p,t) do{if((uintptr_t)p&(sizeof(t)-1)) __builtin_trap();}while(0)
+#else
+#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))
+
+
static queue_head_t lck_grp_queue;
static unsigned int lck_grp_cnt;
decl_lck_mtx_data(static,lck_grp_lock)
static lck_mtx_ext_t lck_grp_lock_ext;
+SECURITY_READ_ONLY_LATE(boolean_t) spinlock_timeout_panic = TRUE;
+
lck_grp_attr_t LockDefaultGroupAttr;
lck_grp_t LockCompatGroup;
lck_attr_t LockDefaultLckAttr;
+#if CONFIG_DTRACE && __SMP__
+#if defined (__x86_64__)
+uint64_t dtrace_spin_threshold = 500; // 500ns
+#elif defined(__arm__) || defined(__arm64__)
+uint64_t dtrace_spin_threshold = LOCK_PANIC_TIMEOUT / 1000000; // 500ns
+#endif
+#endif
+
+uintptr_t
+unslide_for_kdebug(void* object) {
+ if (__improbable(kdebug_enable))
+ return VM_KERNEL_UNSLIDE_OR_PERM(object);
+ else
+ return 0;
+}
+
/*
* Routine: lck_mod_init
*/
if (!PE_parse_boot_argn("lcks", &LcksOpts, sizeof (LcksOpts)))
LcksOpts = 0;
+
+#if (DEVELOPMENT || DEBUG) && defined(__x86_64__)
+ if (!PE_parse_boot_argn("-disable_mtx_chk", &LckDisablePreemptCheck, sizeof (LckDisablePreemptCheck)))
+ LckDisablePreemptCheck = 0;
+#endif /* (DEVELOPMENT || DEBUG) && defined(__x86_64__) */
+
queue_init(&lck_grp_queue);
/*
lck_attr_setdefault(&LockDefaultLckAttr);
lck_mtx_init_ext(&lck_grp_lock, &lck_grp_lock_ext, &LockCompatGroup, &LockDefaultLckAttr);
-
}
/*
void
lck_grp_init(lck_grp_t * grp, const char * grp_name, lck_grp_attr_t * attr)
{
+ /* make sure locking infrastructure has been initialized */
+ assert(lck_grp_cnt > 0);
+
bzero((void *)grp, sizeof(lck_grp_t));
(void)strlcpy(grp->lck_grp_name, grp_name, LCK_GRP_MAX_NAME);
lck_type_t lck_type)
{
unsigned int *lckcnt;
+ int updated;
switch (lck_type) {
case LCK_TYPE_SPIN:
lckcnt = &grp->lck_grp_rwcnt;
break;
default:
- return panic("lck_grp_lckcnt_decr(): invalid lock type: %d\n", lck_type);
+ panic("lck_grp_lckcnt_decr(): invalid lock type: %d\n", lck_type);
+ return;
}
- (void)hw_atomic_sub(lckcnt, 1);
+ updated = (int)hw_atomic_sub(lckcnt, 1);
+ assert(updated >= 0);
}
/*
lck_attr_setdefault(
lck_attr_t *attr)
{
-#if __i386__ || __x86_64__
+#if __arm__ || __arm64__
+ /* <rdar://problem/4404579>: Using LCK_ATTR_DEBUG here causes panic at boot time for arm */
+ attr->lck_attr_val = LCK_ATTR_NONE;
+#elif __i386__ || __x86_64__
#if !DEBUG
if (LcksOpts & enaLkDeb)
attr->lck_attr_val = LCK_ATTR_DEBUG;
kfree(attr, sizeof(lck_attr_t));
}
+/*
+ * Routine: hw_lock_init
+ *
+ * Initialize a hardware lock.
+ */
+void
+hw_lock_init(hw_lock_t lock)
+{
+ ordered_store_hw(lock, 0);
+}
+
+/*
+ * Routine: hw_lock_lock_contended
+ *
+ * Spin until lock is acquired or timeout expires.
+ * 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)
+{
+ uint64_t end = 0;
+ uintptr_t holder = lock->lock_data;
+ int i;
+
+ if (timeout == 0)
+ timeout = LOCK_PANIC_TIMEOUT;
+#if CONFIG_DTRACE
+ uint64_t begin;
+ boolean_t dtrace_enabled = lockstat_probemap[LS_LCK_SPIN_LOCK_SPIN] != 0;
+ if (__improbable(dtrace_enabled))
+ begin = mach_absolute_time();
+#endif
+ for ( ; ; ) {
+ for (i = 0; i < LOCK_SNOOP_SPINS; i++) {
+ cpu_pause();
+#if (!__ARM_ENABLE_WFE_) || (LOCK_PRETEST)
+ holder = ordered_load_hw(lock);
+ if (holder != 0)
+ continue;
+#endif
+ if (atomic_compare_exchange(&lock->lock_data, 0, data,
+ memory_order_acquire_smp, TRUE)) {
+#if CONFIG_DTRACE
+ if (__improbable(dtrace_enabled)) {
+ uint64_t spintime = mach_absolute_time() - begin;
+ if (spintime > dtrace_spin_threshold)
+ LOCKSTAT_RECORD2(LS_LCK_SPIN_LOCK_SPIN, lock, spintime, dtrace_spin_threshold);
+ }
+#endif
+ return 1;
+ }
+ }
+ if (end == 0) {
+ end = ml_get_timebase() + timeout;
+ }
+ else if (ml_get_timebase() >= end)
+ break;
+ }
+ if (do_panic) {
+ // Capture the actual time spent blocked, which may be higher than the timeout
+ // if a misbehaving interrupt stole this thread's CPU time.
+ panic("Spinlock timeout after %llu ticks, %p = %lx",
+ (ml_get_timebase() - end + timeout), lock, holder);
+ }
+ return 0;
+}
+#endif // __SMP__
+
+static inline void
+hw_lock_lock_internal(hw_lock_t lock, thread_t thread)
+{
+ uintptr_t state;
+
+ 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)) {
+ goto end;
+ }
+#if LOCK_PRETEST
+contended:
+#endif // LOCK_PRETEST
+ hw_lock_lock_contended(lock, state, 0, spinlock_timeout_panic);
+end:
+#else // __SMP__
+ if (lock->lock_data)
+ panic("Spinlock held %p", lock);
+ lock->lock_data = state;
+#endif // __SMP__
+#if CONFIG_DTRACE
+ LOCKSTAT_RECORD(LS_LCK_SPIN_LOCK_ACQUIRE, lock, 0);
+#endif
+ return;
+}
+
+/*
+ * Routine: hw_lock_lock
+ *
+ * Acquire lock, spinning until it becomes available,
+ * return with preemption disabled.
+ */
+void
+hw_lock_lock(hw_lock_t lock)
+{
+ thread_t thread = current_thread();
+ disable_preemption_for_thread(thread);
+ hw_lock_lock_internal(lock, thread);
+}
+
+/*
+ * Routine: hw_lock_lock_nopreempt
+ *
+ * Acquire lock, spinning until it becomes available.
+ */
+void
+hw_lock_lock_nopreempt(hw_lock_t lock)
+{
+ thread_t thread = current_thread();
+ if (__improbable(!preemption_disabled_for_thread(thread)))
+ panic("Attempt to take no-preempt spinlock %p in preemptible context", lock);
+ hw_lock_lock_internal(lock, thread);
+}
+
+/*
+ * Routine: hw_lock_to
+ *
+ * Acquire lock, spinning until it becomes available or timeout.
+ * Timeout is in mach_absolute_time ticks, return with
+ * preemption disabled.
+ */
+unsigned int
+hw_lock_to(hw_lock_t lock, uint64_t timeout)
+{
+ thread_t thread;
+ uintptr_t state;
+ unsigned int success = 0;
+
+ thread = current_thread();
+ 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)) {
+ success = 1;
+ goto end;
+ }
+#if LOCK_PRETEST
+contended:
+#endif // LOCK_PRETEST
+ success = hw_lock_lock_contended(lock, state, timeout, FALSE);
+end:
+#else // __SMP__
+ (void)timeout;
+ if (ordered_load_hw(lock) == 0) {
+ ordered_store_hw(lock, state);
+ success = 1;
+ }
+#endif // __SMP__
+#if CONFIG_DTRACE
+ if (success)
+ LOCKSTAT_RECORD(LS_LCK_SPIN_LOCK_ACQUIRE, lock, 0);
+#endif
+ return success;
+}
+
+/*
+ * Routine: hw_lock_try
+ *
+ * returns with preemption disabled on success.
+ */
+static inline unsigned int
+hw_lock_try_internal(hw_lock_t lock, thread_t thread)
+{
+ int success = 0;
+
+#if __SMP__
+#if LOCK_PRETEST
+ if (ordered_load_hw(lock))
+ 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);
+#else
+ if (lock->lock_data == 0) {
+ lock->lock_data = LCK_MTX_THREAD_TO_STATE(thread) | PLATFORM_LCK_ILOCK;
+ success = 1;
+ }
+#endif // __SMP__
+
+#if LOCK_PRETEST
+failed:
+#endif // LOCK_PRETEST
+#if CONFIG_DTRACE
+ if (success)
+ LOCKSTAT_RECORD(LS_LCK_SPIN_LOCK_ACQUIRE, lock, 0);
+#endif
+ return success;
+}
+
+unsigned int
+hw_lock_try(hw_lock_t lock)
+{
+ thread_t thread = current_thread();
+ disable_preemption_for_thread(thread);
+ unsigned int success = hw_lock_try_internal(lock, thread);
+ if (!success)
+ enable_preemption();
+ return success;
+}
+
+unsigned int
+hw_lock_try_nopreempt(hw_lock_t lock)
+{
+ thread_t thread = current_thread();
+ if (__improbable(!preemption_disabled_for_thread(thread)))
+ panic("Attempt to test no-preempt spinlock %p in preemptible context", lock);
+ return hw_lock_try_internal(lock, thread);
+}
+
+/*
+ * Routine: hw_lock_unlock
+ *
+ * Unconditionally release lock, release preemption level.
+ */
+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);
+#if __arm__ || __arm64__
+ // ARM tests are only for open-source exclusion
+ set_event();
+#endif // __arm__ || __arm64__
+#if CONFIG_DTRACE
+ LOCKSTAT_RECORD(LS_LCK_SPIN_UNLOCK_RELEASE, lock, 0);
+#endif /* CONFIG_DTRACE */
+}
+
+void
+hw_lock_unlock(hw_lock_t lock)
+{
+ hw_lock_unlock_internal(lock);
+ enable_preemption();
+}
+
+void
+hw_lock_unlock_nopreempt(hw_lock_t lock)
+{
+ if (__improbable(!preemption_disabled_for_thread(current_thread())))
+ panic("Attempt to release no-preempt spinlock %p in preemptible context", lock);
+ hw_lock_unlock_internal(lock);
+}
+
+/*
+ * Routine hw_lock_held, doesn't change preemption state.
+ * N.B. Racy, of course.
+ */
+unsigned int
+hw_lock_held(hw_lock_t lock)
+{
+ return (ordered_load_hw(lock) != 0);
+}
/*
* Routine: lck_spin_sleep
return res;
}
-
-/*
- * Routine: lck_mtx_clear_promoted
- *
- * Handle clearing of TH_SFLAG_PROMOTED,
- * adjusting thread priority as needed.
- *
- * Called with thread lock held
- */
-static void
-lck_mtx_clear_promoted (
- thread_t thread,
- __kdebug_only uintptr_t trace_lck)
-{
- thread->sched_flags &= ~TH_SFLAG_PROMOTED;
-
- if (thread->sched_flags & TH_SFLAG_RW_PROMOTED) {
- /* Thread still has a RW lock promotion */
- } else if (thread->sched_flags & TH_SFLAG_DEPRESSED_MASK) {
- KERNEL_DEBUG_CONSTANT(
- MACHDBG_CODE(DBG_MACH_SCHED,MACH_DEMOTE) | DBG_FUNC_NONE,
- thread->sched_pri, DEPRESSPRI, 0, trace_lck, 0);
- set_sched_pri(thread, DEPRESSPRI);
- } else {
- if (thread->base_pri < thread->sched_pri) {
- KERNEL_DEBUG_CONSTANT(
- MACHDBG_CODE(DBG_MACH_SCHED,MACH_DEMOTE) | DBG_FUNC_NONE,
- thread->sched_pri, thread->base_pri, 0, trace_lck, 0);
- }
- thread_recompute_sched_pri(thread, FALSE);
- }
-}
-
-
/*
* Routine: lck_mtx_sleep
*/
if (!(lck_sleep_action & LCK_SLEEP_UNLOCK)) {
if ((lck_sleep_action & LCK_SLEEP_SPIN))
lck_mtx_lock_spin(lck);
+ else if ((lck_sleep_action & LCK_SLEEP_SPIN_ALWAYS))
+ lck_mtx_lock_spin_always(lck);
else
lck_mtx_lock(lck);
}
if (lck_sleep_action & LCK_SLEEP_PROMOTED_PRI) {
if ((thread->rwlock_count-- == 1 /* field now 0 */) && (thread->sched_flags & TH_SFLAG_RW_PROMOTED)) {
/* sched_flags checked without lock, but will be rechecked while clearing */
- lck_rw_clear_promotion(thread);
+ lck_rw_clear_promotion(thread, unslide_for_kdebug(event));
}
}
if (lck_sleep_action & LCK_SLEEP_PROMOTED_PRI) {
if ((thread->rwlock_count-- == 1 /* field now 0 */) && (thread->sched_flags & TH_SFLAG_RW_PROMOTED)) {
/* sched_flags checked without lock, but will be rechecked while clearing */
- lck_rw_clear_promotion(thread);
+ lck_rw_clear_promotion(thread, unslide_for_kdebug(event));
}
}
}
/*
- * Routine: lck_mtx_lock_wait
+ * Lock Boosting Invariants:
+ *
+ * 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.
+ */
+
+/*
+ * Routine: lck_mtx_lock_wait
*
* Invoked in order to wait on contention.
*
* Called with the interlock locked and
* returns it unlocked.
+ *
+ * Always aggressively sets the owning thread to promoted,
+ * even if it's the same or higher priority
+ * This prevents it from lowering its own priority while holding a lock
+ *
+ * TODO: Come up with a more efficient way to handle same-priority promotions
+ * <rdar://problem/30737670> ARM mutex contention logic could avoid taking the thread lock
*/
void
lck_mtx_lock_wait (
{
thread_t self = current_thread();
lck_mtx_t *mutex;
- __kdebug_only uintptr_t trace_lck = VM_KERNEL_UNSLIDE_OR_PERM(lck);
- __kdebug_only uintptr_t trace_holder = VM_KERNEL_UNSLIDE_OR_PERM(holder);
- integer_t priority;
- spl_t s = splsched();
+ __kdebug_only uintptr_t trace_lck = unslide_for_kdebug(lck);
+
#if CONFIG_DTRACE
uint64_t sleep_start = 0;
else
mutex = &lck->lck_mtx_ptr->lck_mtx;
- KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_MTX_LCK_WAIT_CODE) | DBG_FUNC_START, trace_lck, trace_holder, 0, 0, 0);
+ 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);
- priority = self->sched_pri;
- if (priority < self->base_pri)
- priority = self->base_pri;
- if (priority < BASEPRI_DEFAULT)
- priority = BASEPRI_DEFAULT;
+ if ((holder->sched_flags & TH_SFLAG_DEPRESS) == 0)
+ assert(holder->sched_pri >= mutex->lck_mtx_pri);
- /* Do not promote past promotion ceiling */
+ integer_t priority = self->sched_pri;
+ priority = MAX(priority, self->base_pri);
+ priority = MAX(priority, BASEPRI_DEFAULT);
priority = MIN(priority, MAXPRI_PROMOTE);
- thread_lock(holder);
- if (mutex->lck_mtx_pri == 0)
- holder->promotions++;
- holder->sched_flags |= TH_SFLAG_PROMOTED;
- if (mutex->lck_mtx_pri < priority && holder->sched_pri < priority) {
- KERNEL_DEBUG_CONSTANT(
- MACHDBG_CODE(DBG_MACH_SCHED,MACH_PROMOTE) | DBG_FUNC_NONE,
- holder->sched_pri, priority, trace_holder, trace_lck, 0);
- set_sched_pri(holder, priority);
+ 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);
if (mutex->lck_mtx_pri < priority)
mutex->lck_mtx_pri = priority;
- if (self->pending_promoter[self->pending_promoter_index] == NULL) {
- self->pending_promoter[self->pending_promoter_index] = mutex;
- mutex->lck_mtx_waiters++;
- }
- else
- if (self->pending_promoter[self->pending_promoter_index] != mutex) {
- self->pending_promoter[++self->pending_promoter_index] = mutex;
+
+ if (self->waiting_for_mutex == NULL) {
+ self->waiting_for_mutex = mutex;
mutex->lck_mtx_waiters++;
}
- assert_wait(LCK_MTX_EVENT(mutex), THREAD_UNINT);
+ assert(self->waiting_for_mutex == mutex);
+
+ thread_set_pending_block_hint(self, kThreadWaitKernelMutex);
+ assert_wait(LCK_MTX_EVENT(mutex), THREAD_UNINT | THREAD_WAIT_NOREPORT_USER);
lck_mtx_ilk_unlock(mutex);
thread_block(THREAD_CONTINUE_NULL);
+ assert(mutex->lck_mtx_waiters > 0);
+
KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_MTX_LCK_WAIT_CODE) | DBG_FUNC_END, 0, 0, 0, 0, 0);
#if CONFIG_DTRACE
/*
- * Record the Dtrace lockstat probe for blocking, block time
+ * Record the DTrace lockstat probe for blocking, block time
* measured from when we were entered.
*/
if (sleep_start) {
thread_t thread = current_thread();
lck_mtx_t *mutex;
integer_t priority;
- spl_t s;
- __kdebug_only uintptr_t trace_lck = VM_KERNEL_UNSLIDE_OR_PERM(lck);
if (lck->lck_mtx_tag != LCK_MTX_TAG_INDIRECT)
mutex = lck;
else
mutex = &lck->lck_mtx_ptr->lck_mtx;
- if (thread->pending_promoter[thread->pending_promoter_index] == mutex) {
- thread->pending_promoter[thread->pending_promoter_index] = NULL;
- if (thread->pending_promoter_index > 0)
- thread->pending_promoter_index--;
+ /*
+ * 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--;
}
- if (mutex->lck_mtx_waiters)
+ assert(thread->waiting_for_mutex == NULL);
+
+ if (mutex->lck_mtx_waiters > 0) {
priority = mutex->lck_mtx_pri;
- else {
+ } 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) {
- s = splsched();
+ __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 (priority) {
- thread->promotions++;
- thread->sched_flags |= TH_SFLAG_PROMOTED;
- if (thread->sched_pri < priority) {
- KERNEL_DEBUG_CONSTANT(
- MACHDBG_CODE(DBG_MACH_SCHED,MACH_PROMOTE) | DBG_FUNC_NONE,
- thread->sched_pri, priority, 0, trace_lck, 0);
- /* Do not promote past promotion ceiling */
- assert(priority <= MAXPRI_PROMOTE);
- set_sched_pri(thread, 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)
- lck_mtx_clear_promoted(thread, trace_lck);
+ 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);
+
thread_unlock(thread);
splx(s);
}
* Invoked on unlock when there is contention.
*
* 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.
*/
void
lck_mtx_unlock_wakeup (
{
thread_t thread = current_thread();
lck_mtx_t *mutex;
- __kdebug_only uintptr_t trace_lck = VM_KERNEL_UNSLIDE_OR_PERM(lck);
+ __kdebug_only uintptr_t trace_lck = unslide_for_kdebug(lck);
if (lck->lck_mtx_tag != LCK_MTX_TAG_INDIRECT)
mutex = lck;
if (thread != holder)
panic("lck_mtx_unlock_wakeup: mutex %p holder %p\n", mutex, holder);
- KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_MTX_UNLCK_WAKEUP_CODE) | DBG_FUNC_START, trace_lck, VM_KERNEL_UNSLIDE_OR_PERM(holder), 0, 0, 0);
+ KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_MTX_UNLCK_WAKEUP_CODE) | DBG_FUNC_START,
+ 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
+ */
if (mutex->lck_mtx_waiters > 1)
thread_wakeup_one_with_pri(LCK_MTX_EVENT(lck), lck->lck_mtx_pri);
else
thread_wakeup_one(LCK_MTX_EVENT(lck));
- if (thread->promotions > 0) {
- spl_t s = splsched();
-
+ /* 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);
- if (--thread->promotions == 0 && (thread->sched_flags & TH_SFLAG_PROMOTED))
- lck_mtx_clear_promoted(thread, trace_lck);
+
+ assert(thread->promotions > 0);
+
+ assert_promotions_invariant(thread);
+
+ if (--thread->promotions == 0)
+ sched_thread_unpromote(thread, trace_lck);
+
+ assert_promotions_invariant(thread);
+
thread_unlock(thread);
splx(s);
}
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_unlockspin_wakeup (
- lck_mtx_t *lck)
+lck_mtx_wakeup_adjust_pri(thread_t thread, integer_t priority)
{
- assert(lck->lck_mtx_waiters > 0);
- thread_wakeup_one(LCK_MTX_EVENT(lck));
+ assert(priority <= MAXPRI_PROMOTE);
+ assert(thread->waiting_for_mutex != NULL);
- KERNEL_DEBUG(MACHDBG_CODE(DBG_MACH_LOCKS, LCK_MTX_UNLCK_WAKEUP_CODE) | DBG_FUNC_NONE, VM_KERNEL_UNSLIDE_OR_PERM(lck), 0, 0, 1, 0);
-#if CONFIG_DTRACE
- /*
- * When there are waiters, we skip the hot-patch spot in the
- * fastpath, so we record it here.
- */
- LOCKSTAT_RECORD(LS_LCK_MTX_UNLOCK_RELEASE, lck, 0);
-#endif
+ __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);
}
/* Only if the caller wanted the lck_rw_t returned unlocked should we drop to 0 */
assert(lck_sleep_action & LCK_SLEEP_UNLOCK);
- lck_rw_clear_promotion(thread);
+ lck_rw_clear_promotion(thread, unslide_for_kdebug(event));
}
}
/* Only if the caller wanted the lck_rw_t returned unlocked should we drop to 0 */
assert(lck_sleep_action & LCK_SLEEP_UNLOCK);
- lck_rw_clear_promotion(thread);
+ lck_rw_clear_promotion(thread, unslide_for_kdebug(event));
}
}
*
* We support a limited form of reader-writer
* lock promotion whose effects are:
- *
+ *
* * Qualifying threads have decay disabled
* * Scheduler priority is reset to a floor of
* of their statically assigned priority
- * or BASEPRI_BACKGROUND
+ * or MINPRI_RWLOCK
*
* The rationale is that lck_rw_ts do not have
* a single owner, so we cannot apply a directed
* lck_rw_clear_promotion: Undo priority promotions when the last RW
* lock is released by a thread (if a promotion was active)
*/
-void lck_rw_clear_promotion(thread_t thread)
+void lck_rw_clear_promotion(thread_t thread, uintptr_t trace_obj)
{
assert(thread->rwlock_count == 0);
/* Cancel any promotions if the thread had actually blocked while holding a RW lock */
spl_t s = splsched();
-
thread_lock(thread);
- if (thread->sched_flags & TH_SFLAG_RW_PROMOTED) {
- thread->sched_flags &= ~TH_SFLAG_RW_PROMOTED;
+ if (thread->sched_flags & TH_SFLAG_RW_PROMOTED)
+ sched_thread_unpromote_reason(thread, TH_SFLAG_RW_PROMOTED, trace_obj);
- if (thread->sched_flags & TH_SFLAG_PROMOTED) {
- /* Thread still has a mutex promotion */
- } else if (thread->sched_flags & TH_SFLAG_DEPRESSED_MASK) {
- KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_RW_DEMOTE) | DBG_FUNC_NONE,
- (uintptr_t)thread_tid(thread), thread->sched_pri, DEPRESSPRI, 0, 0);
+ thread_unlock(thread);
+ splx(s);
+}
- set_sched_pri(thread, DEPRESSPRI);
- } else {
- KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_RW_DEMOTE) | DBG_FUNC_NONE,
- (uintptr_t)thread_tid(thread), thread->sched_pri, thread->base_pri, 0, 0);
+/*
+ * Callout from context switch if the thread goes
+ * off core with a positive rwlock_count
+ *
+ * Called at splsched with the thread locked
+ */
+void
+lck_rw_set_promotion_locked(thread_t thread)
+{
+ if (LcksOpts & disLkRWPrio)
+ return;
- thread_recompute_sched_pri(thread, FALSE);
- }
- }
+ assert(thread->rwlock_count > 0);
- thread_unlock(thread);
- splx(s);
+ if (!(thread->sched_flags & TH_SFLAG_RW_PROMOTED))
+ sched_thread_promote_reason(thread, TH_SFLAG_RW_PROMOTED, 0);
}
kern_return_t
return(KERN_SUCCESS);
}
+/*
+ * Atomic primitives, prototyped in kern/simple_lock.h
+ * Noret versions are more efficient on some architectures
+ */
+
+uint32_t
+hw_atomic_add(volatile uint32_t *dest, uint32_t delt)
+{
+ ALIGN_TEST(dest,uint32_t);
+ return __c11_atomic_fetch_add(ATOMIC_CAST(uint32_t,dest), delt, memory_order_relaxed) + delt;
+}
+
+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;
+}
+
+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;
+}
+
+void
+hw_atomic_or_noret(volatile uint32_t *dest, uint32_t mask)
+{
+ ALIGN_TEST(dest,uint32_t);
+ __c11_atomic_fetch_or(ATOMIC_CAST(uint32_t,dest), mask, memory_order_relaxed);
+}
+
+uint32_t
+hw_atomic_and(volatile uint32_t *dest, uint32_t mask)
+{
+ ALIGN_TEST(dest,uint32_t);
+ return __c11_atomic_fetch_and(ATOMIC_CAST(uint32_t,dest), mask, memory_order_relaxed) & mask;
+}
+
+void
+hw_atomic_and_noret(volatile uint32_t *dest, uint32_t mask)
+{
+ ALIGN_TEST(dest,uint32_t);
+ __c11_atomic_fetch_and(ATOMIC_CAST(uint32_t,dest), mask, memory_order_relaxed);
+}
+
+uint32_t
+hw_compare_and_store(uint32_t oldval, uint32_t newval, volatile uint32_t *dest)
+{
+ 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);
+}
+
projects / apple / xnu.git / blobdiff