#include <machine/commpage.h> /* for commpage_update_mach_approximate_time */
#endif
+#if MONOTONIC
+#include <kern/monotonic.h>
+#endif /* MONOTONIC */
+
static void sched_update_thread_bucket(thread_t thread);
/*
/*
* We bill CPU time to both the individual thread and its task.
*
- * Because this balance adjustment could potentially attempt to wake this very
- * thread, we must credit the ledger before taking the thread lock. The ledger
- * pointers are only manipulated by the thread itself at the ast boundary.
+ * Because this balance adjustment could potentially attempt to wake this
+ * very thread, we must credit the ledger before taking the thread lock.
+ * The ledger pointers are only manipulated by the thread itself at the ast
+ * boundary.
*
- * TODO: This fails to account for the time between when the timer was armed and when it fired.
- * It should be based on the system_timer and running a thread_timer_event operation here.
+ * TODO: This fails to account for the time between when the timer was
+ * armed and when it fired. It should be based on the system_timer and
+ * running a timer_update operation here.
*/
ledger_credit(thread->t_ledger, task_ledgers.cpu_time, thread->quantum_remaining);
ledger_credit(thread->t_threadledger, thread_ledgers.cpu_time, thread->quantum_remaining);
-#ifdef CONFIG_BANK
if (thread->t_bankledger) {
ledger_credit(thread->t_bankledger, bank_ledgers.cpu_time,
(thread->quantum_remaining - thread->t_deduct_bank_ledger_time));
}
thread->t_deduct_bank_ledger_time = 0;
-#endif
ctime = mach_absolute_time();
commpage_update_mach_approximate_time(ctime);
#endif
+#if MONOTONIC
+ mt_sched_update(thread);
+#endif /* MONOTONIC */
+
thread_lock(thread);
/*
/*
* Check for fail-safe trip.
*/
- if ((thread->sched_mode == TH_MODE_REALTIME || thread->sched_mode == TH_MODE_FIXED) &&
- !(thread->sched_flags & TH_SFLAG_PROMOTED_MASK) &&
- !(thread->options & TH_OPT_SYSTEM_CRITICAL)) {
- uint64_t new_computation;
-
- new_computation = ctime - thread->computation_epoch;
- new_computation += thread->computation_metered;
- if (new_computation > max_unsafe_computation) {
+ if ((thread->sched_mode == TH_MODE_REALTIME || thread->sched_mode == TH_MODE_FIXED) &&
+ !(thread->sched_flags & TH_SFLAG_PROMOTED) &&
+ !(thread->sched_flags & TH_SFLAG_PROMOTE_REASON_MASK) &&
+ !(thread->options & TH_OPT_SYSTEM_CRITICAL)) {
+ uint64_t new_computation;
+
+ new_computation = ctime - thread->computation_epoch;
+ new_computation += thread->computation_metered;
+ if (new_computation > max_unsafe_computation) {
KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_FAILSAFE)|DBG_FUNC_NONE,
(uintptr_t)thread->sched_pri, (uintptr_t)thread->sched_mode, 0, 0, 0);
if (thread->sched_mode != TH_MODE_REALTIME)
SCHED(quantum_expire)(thread);
- processor->current_pri = thread->sched_pri;
- processor->current_thmode = thread->sched_mode;
-
- /* Tell platform layer that we are still running this thread */
- urgency = thread_get_urgency(thread, &ignore1, &ignore2);
- machine_thread_going_on_core(thread, urgency, 0);
+ processor_state_update_from_thread(processor, thread);
/*
* This quantum is up, give this thread another.
* during privilege transitions, synthesize an event now.
*/
if (!thread->precise_user_kernel_time) {
- timer_switch(PROCESSOR_DATA(processor, current_state),
- ctime,
- PROCESSOR_DATA(processor, current_state));
- timer_switch(PROCESSOR_DATA(processor, thread_timer),
- ctime,
- PROCESSOR_DATA(processor, thread_timer));
+ timer_update(PROCESSOR_DATA(processor, current_state), ctime);
+ timer_update(PROCESSOR_DATA(processor, thread_timer), ctime);
+ timer_update(&thread->runnable_timer, ctime);
}
+
processor->quantum_end = ctime + thread->quantum_remaining;
/*
- * Context switch check.
+ * Context switch check
+ *
+ * non-urgent flags don't affect kernel threads, so upgrade to urgent
+ * to ensure that rebalancing and non-recommendation kick in quickly.
*/
- if ((preempt = csw_check(processor, AST_QUANTUM)) != AST_NONE)
+
+ ast_t check_reason = AST_QUANTUM;
+ if (thread->task == kernel_task)
+ check_reason |= AST_URGENT;
+
+ if ((preempt = csw_check(processor, check_reason)) != AST_NONE)
ast_on(preempt);
+ /*
+ * AST_KEVENT does not send an IPI when setting the AST,
+ * to avoid waiting for the next context switch to propagate the AST,
+ * the AST is propagated here at quantum expiration.
+ */
+ ast_propagate(thread);
+
thread_unlock(thread);
- timer_call_enter1(&processor->quantum_timer, thread,
- processor->quantum_end, TIMER_CALL_SYS_CRITICAL | TIMER_CALL_LOCAL);
+ timer_call_quantum_timer_enter(&processor->quantum_timer, thread,
+ processor->quantum_end, ctime);
+
+ /* Tell platform layer that we are still running this thread */
+ urgency = thread_get_urgency(thread, &ignore1, &ignore2);
+ machine_thread_going_on_core(thread, urgency, 0, 0, ctime);
+ machine_switch_perfcontrol_state_update(QUANTUM_EXPIRY, ctime,
+ 0, thread);
#if defined(CONFIG_SCHED_TIMESHARE_CORE)
sched_timeshare_consider_maintenance(ctime);
#endif /* CONFIG_SCHED_TIMESHARE_CORE */
+#if __arm__ || __arm64__
+ if (thread->sched_mode == TH_MODE_REALTIME)
+ sched_consider_recommended_cores(ctime, thread);
+#endif /* __arm__ || __arm64__ */
KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_QUANTUM_EXPIRED) | DBG_FUNC_END, preempt, 0, 0, 0, 0);
}
sched_set_thread_base_priority(thread_t thread, int priority)
{
assert(priority >= MINPRI);
+ uint64_t ctime = 0;
if (thread->sched_mode == TH_MODE_REALTIME)
assert(priority <= BASEPRI_RTQUEUES);
else
assert(priority < BASEPRI_RTQUEUES);
+ int old_base_pri = thread->base_pri;
thread->base_pri = priority;
+ if ((thread->state & TH_RUN) == TH_RUN) {
+ assert(thread->last_made_runnable_time != THREAD_NOT_RUNNABLE);
+ ctime = mach_approximate_time();
+ thread->last_basepri_change_time = ctime;
+ } else {
+ assert(thread->last_basepri_change_time == THREAD_NOT_RUNNABLE);
+ assert(thread->last_made_runnable_time == THREAD_NOT_RUNNABLE);
+ }
+
+ /*
+ * Currently the perfcontrol_attr depends on the base pri of the
+ * thread. Therefore, we use this function as the hook for the
+ * perfcontrol callout.
+ */
+ if (thread == current_thread() && old_base_pri != priority) {
+ if (!ctime) {
+ ctime = mach_approximate_time();
+ }
+ machine_switch_perfcontrol_state_update(PERFCONTROL_ATTR_UPDATE,
+ ctime, PERFCONTROL_CALLOUT_WAKE_UNSAFE, thread);
+ }
sched_update_thread_bucket(thread);
- thread_recompute_sched_pri(thread, FALSE);
+ thread_recompute_sched_pri(thread, SETPRI_DEFAULT);
}
/*
* according to its base priority if the
* thread has not been promoted or depressed.
*
- * This is the standard way to push base_pri changes into sched_pri,
- * or to recalculate the appropriate sched_pri after clearing
+ * This is the only way to push base_pri changes into sched_pri,
+ * or to recalculate the appropriate sched_pri after changing
* a promotion or depression.
*
* Called at splsched with the thread locked.
+ *
+ * TODO: Add an 'update urgency' flag to avoid urgency callouts on every rwlock operation
*/
void
-thread_recompute_sched_pri(
- thread_t thread,
- boolean_t override_depress)
+thread_recompute_sched_pri(thread_t thread, set_sched_pri_options_t options)
{
- int priority;
+ uint32_t sched_flags = thread->sched_flags;
+ sched_mode_t sched_mode = thread->sched_mode;
- if (thread->sched_mode == TH_MODE_TIMESHARE)
+ int priority = thread->base_pri;
+
+ if (sched_mode == TH_MODE_TIMESHARE)
priority = SCHED(compute_timeshare_priority)(thread);
- else
- priority = thread->base_pri;
- if ((!(thread->sched_flags & TH_SFLAG_PROMOTED_MASK) || (priority > thread->sched_pri)) &&
- (!(thread->sched_flags & TH_SFLAG_DEPRESSED_MASK) || override_depress)) {
- set_sched_pri(thread, priority);
+ if (sched_flags & TH_SFLAG_DEPRESS) {
+ /* thread_yield_internal overrides kernel mutex promotion */
+ priority = DEPRESSPRI;
+ } else {
+ /* poll-depress is overridden by mutex promotion and promote-reasons */
+ if ((sched_flags & TH_SFLAG_POLLDEPRESS)) {
+ priority = DEPRESSPRI;
+ }
+
+ if (sched_flags & TH_SFLAG_PROMOTED) {
+ priority = MAX(priority, thread->promotion_priority);
+
+ if (sched_mode != TH_MODE_REALTIME)
+ priority = MIN(priority, MAXPRI_PROMOTE);
+ }
+
+ if (sched_flags & TH_SFLAG_PROMOTE_REASON_MASK) {
+ if (sched_flags & TH_SFLAG_RW_PROMOTED)
+ priority = MAX(priority, MINPRI_RWLOCK);
+
+ if (sched_flags & TH_SFLAG_WAITQ_PROMOTED)
+ priority = MAX(priority, MINPRI_WAITQ);
+
+ if (sched_flags & TH_SFLAG_EXEC_PROMOTED)
+ priority = MAX(priority, MINPRI_EXEC);
+ }
}
+
+ set_sched_pri(thread, priority, options);
}
void
priority = sched_compute_timeshare_priority(thread);
- /*
- * Adjust the scheduled priority like thread_recompute_sched_pri,
- * except with the benefit of knowing the thread is on this core.
- */
- if ((!(thread->sched_flags & TH_SFLAG_PROMOTED_MASK) || (priority > thread->sched_pri)) &&
- (!(thread->sched_flags & TH_SFLAG_DEPRESSED_MASK)) &&
- priority != thread->sched_pri) {
-
- thread->sched_pri = priority;
-
- KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHANGE_PRIORITY),
- (uintptr_t)thread_tid(thread),
- thread->base_pri,
- thread->sched_pri,
- 0, /* eventually, 'reason' */
- 0);
- }
+ if (priority != thread->sched_pri)
+ thread_recompute_sched_pri(thread, SETPRI_LAZY);
}
}
*/
extern int sched_pri_decay_band_limit;
+#ifdef CONFIG_EMBEDDED
+
+int
+sched_compute_timeshare_priority(thread_t thread)
+{
+ int decay_amount = (thread->sched_usage >> thread->pri_shift);
+ int decay_limit = sched_pri_decay_band_limit;
+
+ if (thread->base_pri > BASEPRI_FOREGROUND) {
+ decay_limit += (thread->base_pri - BASEPRI_FOREGROUND);
+ }
+
+ if (decay_amount > decay_limit) {
+ decay_amount = decay_limit;
+ }
+
+ /* start with base priority */
+ int priority = thread->base_pri - decay_amount;
+
+ if (priority < MAXPRI_THROTTLE) {
+ if (thread->task->max_priority > MAXPRI_THROTTLE) {
+ priority = MAXPRI_THROTTLE;
+ } else if (priority < MINPRI_USER) {
+ priority = MINPRI_USER;
+ }
+ } else if (priority > MAXPRI_KERNEL) {
+ priority = MAXPRI_KERNEL;
+ }
+
+ return priority;
+}
+
+#else /* CONFIG_EMBEDDED */
int
sched_compute_timeshare_priority(thread_t thread)
return priority;
}
+#endif /* CONFIG_EMBEDDED */
/*
* can_update_priority
thread->sched_stamp += ticks;
- thread->pri_shift = sched_pri_shifts[thread->th_sched_bucket];
-
/* If requested, accelerate aging of sched_usage */
if (sched_decay_usage_age_factor > 1)
ticks *= sched_decay_usage_age_factor;
thread_timer_delta(thread, delta);
if (ticks < SCHED_DECAY_TICKS) {
/*
- * Accumulate timesharing usage only
- * during contention for processor
- * resources.
+ * Accumulate timesharing usage only during contention for processor
+ * resources. Use the pri_shift from the previous tick window to
+ * determine if the system was in a contended state.
*/
if (thread->pri_shift < INT8_MAX)
thread->sched_usage += delta;
}
/*
- * Recompute scheduled priority if appropriate.
+ * Now that the thread's CPU usage has been accumulated and aged
+ * based on contention of the previous tick window, update the
+ * pri_shift of the thread to match the current global load/shift
+ * values. The updated pri_shift would be used to calculate the
+ * new priority of the thread.
*/
- if (thread->sched_mode == TH_MODE_TIMESHARE) {
- int priority = sched_compute_timeshare_priority(thread);
-
- /*
- * Adjust the scheduled priority like thread_recompute_sched_pri,
- * except without setting an AST.
- */
- if ((!(thread->sched_flags & TH_SFLAG_PROMOTED_MASK) || (priority > thread->sched_pri)) &&
- (!(thread->sched_flags & TH_SFLAG_DEPRESSED_MASK)) &&
- priority != thread->sched_pri) {
-
- boolean_t removed = thread_run_queue_remove(thread);
-
- thread->sched_pri = priority;
-
- KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHANGE_PRIORITY),
- (uintptr_t)thread_tid(thread),
- thread->base_pri,
- thread->sched_pri,
- 0, /* eventually, 'reason' */
- 0);
-
- if (removed)
- thread_run_queue_reinsert(thread, SCHED_TAILQ);
- }
- }
+ thread->pri_shift = sched_pri_shifts[thread->th_sched_bucket];
- return;
+ /* Recompute scheduled priority if appropriate. */
+ if (thread->sched_mode == TH_MODE_TIMESHARE)
+ thread_recompute_sched_pri(thread, SETPRI_LAZY);
}
#endif /* CONFIG_SCHED_TIMESHARE_CORE */
break;
case TH_MODE_TIMESHARE:
- if (thread->base_pri > BASEPRI_UTILITY)
+ if (thread->base_pri > BASEPRI_DEFAULT)
new_bucket = TH_BUCKET_SHARE_FG;
+ else if (thread->base_pri > BASEPRI_UTILITY)
+ new_bucket = TH_BUCKET_SHARE_DF;
else if (thread->base_pri > MAXPRI_THROTTLE)
new_bucket = TH_BUCKET_SHARE_UT;
else
thread_run_queue_reinsert(thread, SCHED_TAILQ);
}
+/*
+ * Promote thread to a specific priority
+ *
+ * Promotion must not last past syscall boundary
+ * Clients must always pair promote and unpromote 1:1
+ *
+ * Called at splsched with thread locked
+ */
+void
+sched_thread_promote_to_pri(thread_t thread,
+ int priority,
+ __kdebug_only uintptr_t trace_obj /* already unslid */)
+{
+ assert((thread->sched_flags & TH_SFLAG_PROMOTED) != TH_SFLAG_PROMOTED);
+ assert(thread->promotion_priority == 0);
+ assert(priority <= MAXPRI_PROMOTE);
+ assert(priority > 0);
+
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_PROMOTED),
+ thread_tid(thread), trace_obj, priority);
+
+ thread->sched_flags |= TH_SFLAG_PROMOTED;
+ thread->promotion_priority = priority;
+
+ thread_recompute_sched_pri(thread, SETPRI_DEFAULT);
+}
+
+
+/*
+ * Update a pre-existing priority promotion to have a higher priority floor
+ * Priority can only go up from the previous value
+ * Update must occur while a promotion is active
+ *
+ * Called at splsched with thread locked
+ */
+void
+sched_thread_update_promotion_to_pri(thread_t thread,
+ int priority,
+ __kdebug_only uintptr_t trace_obj /* already unslid */)
+{
+ assert(thread->promotions > 0);
+ assert((thread->sched_flags & TH_SFLAG_PROMOTED) == TH_SFLAG_PROMOTED);
+ assert(thread->promotion_priority > 0);
+ assert(priority <= MAXPRI_PROMOTE);
+
+ if (thread->promotion_priority < priority) {
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_PROMOTED_UPDATE),
+ thread_tid(thread), trace_obj, priority);
+
+ thread->promotion_priority = priority;
+ thread_recompute_sched_pri(thread, SETPRI_DEFAULT);
+ }
+}
+
+/*
+ * End a priority promotion
+ * Demotes a thread back to its expected priority without the promotion in place
+ *
+ * Called at splsched with thread locked
+ */
+void
+sched_thread_unpromote(thread_t thread,
+ __kdebug_only uintptr_t trace_obj /* already unslid */)
+{
+ assert((thread->sched_flags & TH_SFLAG_PROMOTED) == TH_SFLAG_PROMOTED);
+ assert(thread->promotion_priority > 0);
+
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_UNPROMOTED),
+ thread_tid(thread), trace_obj, 0);
+
+ thread->sched_flags &= ~TH_SFLAG_PROMOTED;
+ thread->promotion_priority = 0;
+
+ thread_recompute_sched_pri(thread, SETPRI_DEFAULT);
+}
+
+/* called with thread locked */
+void
+assert_promotions_invariant(thread_t thread)
+{
+ if (thread->promotions > 0)
+ assert((thread->sched_flags & TH_SFLAG_PROMOTED) == TH_SFLAG_PROMOTED);
+
+ if (thread->promotions == 0)
+ assert((thread->sched_flags & TH_SFLAG_PROMOTED) != TH_SFLAG_PROMOTED);
+}
+
+/*
+ * Promote thread to have a sched pri floor for a specific reason
+ *
+ * Promotion must not last past syscall boundary
+ * Clients must always pair promote and demote 1:1,
+ * Handling nesting of the same promote reason is the client's responsibility
+ *
+ * Called at splsched with thread locked
+ */
+void
+sched_thread_promote_reason(thread_t thread,
+ uint32_t reason,
+ __kdebug_only uintptr_t trace_obj /* already unslid */)
+{
+ assert(reason & TH_SFLAG_PROMOTE_REASON_MASK);
+ assert((thread->sched_flags & reason) != reason);
+
+ switch (reason) {
+ case TH_SFLAG_RW_PROMOTED:
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_RW_PROMOTE),
+ thread_tid(thread), thread->sched_pri,
+ thread->base_pri, trace_obj);
+ break;
+ case TH_SFLAG_WAITQ_PROMOTED:
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAITQ_PROMOTE),
+ thread_tid(thread), thread->sched_pri,
+ thread->base_pri, trace_obj);
+ break;
+ case TH_SFLAG_EXEC_PROMOTED:
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_EXEC_PROMOTE),
+ thread_tid(thread), thread->sched_pri,
+ thread->base_pri, trace_obj);
+ break;
+ }
+
+ thread->sched_flags |= reason;
+
+ thread_recompute_sched_pri(thread, SETPRI_DEFAULT);
+}
+
+/*
+ * End a specific promotion reason
+ * Demotes a thread back to its expected priority without the promotion in place
+ *
+ * Called at splsched with thread locked
+ */
+void
+sched_thread_unpromote_reason(thread_t thread,
+ uint32_t reason,
+ __kdebug_only uintptr_t trace_obj /* already unslid */)
+{
+ assert(reason & TH_SFLAG_PROMOTE_REASON_MASK);
+ assert((thread->sched_flags & reason) == reason);
+
+ switch (reason) {
+ case TH_SFLAG_RW_PROMOTED:
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_RW_DEMOTE),
+ thread_tid(thread), thread->sched_pri,
+ thread->base_pri, trace_obj);
+ break;
+ case TH_SFLAG_WAITQ_PROMOTED:
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_WAITQ_DEMOTE),
+ thread_tid(thread), thread->sched_pri,
+ thread->base_pri, trace_obj);
+ break;
+ case TH_SFLAG_EXEC_PROMOTED:
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_EXEC_DEMOTE),
+ thread_tid(thread), thread->sched_pri,
+ thread->base_pri, trace_obj);
+ break;
+ }
+
+ thread->sched_flags &= ~reason;
+
+ thread_recompute_sched_pri(thread, SETPRI_DEFAULT);
+}
+
projects / apple / xnu.git / blobdiff