+ if ((thread->chosen_processor != processor) && (thread->chosen_processor != NULL)) {
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_MOVED)|DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), (uintptr_t)thread->chosen_processor->cpu_id, 0, 0, 0);
+ }
+
+ DTRACE_SCHED2(off__cpu, struct thread *, thread, struct proc *, thread->task->bsd_info);
+
+ SCHED_STATS_CSW(processor, self->reason, self->sched_pri, thread->sched_pri);
+
+ /*
+ * This is where we actually switch register context,
+ * and address space if required. We will next run
+ * as a result of a subsequent context switch.
+ *
+ * Once registers are switched and the processor is running "thread",
+ * the stack variables and non-volatile registers will contain whatever
+ * was there the last time that thread blocked. No local variables should
+ * be used after this point, except for the special case of "thread", which
+ * the platform layer returns as the previous thread running on the processor
+ * via the function call ABI as a return register, and "self", which may have
+ * been stored on the stack or a non-volatile register, but a stale idea of
+ * what was on the CPU is newly-accurate because that thread is again
+ * running on the CPU.
+ */
+ assert(continuation == self->continuation);
+ thread = machine_switch_context(self, continuation, thread);
+ assert(self == current_thread());
+ TLOG(1,"thread_invoke: returning machine_switch_context: self %p continuation %p thread %p\n", self, continuation, thread);
+
+ DTRACE_SCHED(on__cpu);
+
+ /*
+ * We have been resumed and are set to run.
+ */
+ thread_dispatch(thread, self);
+
+ if (continuation) {
+ self->continuation = self->parameter = NULL;
+
+ (void) spllo();
+
+ call_continuation(continuation, parameter, self->wait_result);
+ /*NOTREACHED*/
+ }
+
+ return (TRUE);
+}
+
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+/*
+ * pset_cancel_deferred_dispatch:
+ *
+ * Cancels all ASTs that we can cancel for the given processor set
+ * if the current processor is running the last runnable thread in the
+ * system.
+ *
+ * This function assumes the current thread is runnable. This must
+ * be called with the pset unlocked.
+ */
+static void
+pset_cancel_deferred_dispatch(
+ processor_set_t pset,
+ processor_t processor)
+{
+ processor_t active_processor = NULL;
+ uint32_t sampled_sched_run_count;
+
+ pset_lock(pset);
+ sampled_sched_run_count = (volatile uint32_t) sched_run_count;
+
+ /*
+ * If we have emptied the run queue, and our current thread is runnable, we
+ * should tell any processors that are still DISPATCHING that they will
+ * probably not have any work to do. In the event that there are no
+ * pending signals that we can cancel, this is also uninteresting.
+ *
+ * In the unlikely event that another thread becomes runnable while we are
+ * doing this (sched_run_count is atomically updated, not guarded), the
+ * codepath making it runnable SHOULD (a dangerous word) need the pset lock
+ * in order to dispatch it to a processor in our pset. So, the other
+ * codepath will wait while we squash all cancelable ASTs, get the pset
+ * lock, and then dispatch the freshly runnable thread. So this should be
+ * correct (we won't accidentally have a runnable thread that hasn't been
+ * dispatched to an idle processor), if not ideal (we may be restarting the
+ * dispatch process, which could have some overhead).
+ *
+ */
+ if ((sampled_sched_run_count == 1) &&
+ (pset->pending_deferred_AST_cpu_mask)) {
+ qe_foreach_element_safe(active_processor, &pset->active_queue, processor_queue) {
+ /*
+ * If a processor is DISPATCHING, it could be because of
+ * a cancelable signal.
+ *
+ * IF the processor is not our
+ * current processor (the current processor should not
+ * be DISPATCHING, so this is a bit paranoid), AND there
+ * is a cancelable signal pending on the processor, AND
+ * there is no non-cancelable signal pending (as there is
+ * no point trying to backtrack on bringing the processor
+ * up if a signal we cannot cancel is outstanding), THEN
+ * it should make sense to roll back the processor state
+ * to the IDLE state.
+ *
+ * If the racey nature of this approach (as the signal
+ * will be arbitrated by hardware, and can fire as we
+ * roll back state) results in the core responding
+ * despite being pushed back to the IDLE state, it
+ * should be no different than if the core took some
+ * interrupt while IDLE.
+ */
+ if ((active_processor->state == PROCESSOR_DISPATCHING) &&
+ (pset->pending_deferred_AST_cpu_mask & (1ULL << active_processor->cpu_id)) &&
+ (!(pset->pending_AST_cpu_mask & (1ULL << active_processor->cpu_id))) &&
+ (active_processor != processor)) {
+ /*
+ * Squash all of the processor state back to some
+ * reasonable facsimile of PROCESSOR_IDLE.
+ *
+ * TODO: What queue policy do we actually want here?
+ * We want to promote selection of a good processor
+ * to run on. Do we want to enqueue at the head?
+ * The tail? At the (relative) old position in the
+ * queue? Or something else entirely?
+ */
+ re_queue_head(&pset->idle_queue, (queue_entry_t)active_processor);
+
+ assert(active_processor->next_thread == THREAD_NULL);
+
+ active_processor->current_pri = IDLEPRI;
+ active_processor->current_thmode = TH_MODE_FIXED;
+ active_processor->current_sfi_class = SFI_CLASS_KERNEL;
+ active_processor->deadline = UINT64_MAX;
+ active_processor->state = PROCESSOR_IDLE;
+ pset->pending_deferred_AST_cpu_mask &= ~(1U << active_processor->cpu_id);
+ machine_signal_idle_cancel(active_processor);
+ }
+
+ }
+ }
+
+ pset_unlock(pset);
+}
+#else
+/* We don't support deferred ASTs; everything is candycanes and sunshine. */
+#endif
+
+/*
+ * thread_dispatch:
+ *
+ * Handle threads at context switch. Re-dispatch other thread
+ * if still running, otherwise update run state and perform
+ * special actions. Update quantum for other thread and begin
+ * the quantum for ourselves.
+ *
+ * "thread" is the old thread that we have switched away from.
+ * "self" is the new current thread that we have context switched to
+ *
+ * Called at splsched.
+ */
+void
+thread_dispatch(
+ thread_t thread,
+ thread_t self)
+{
+ processor_t processor = self->last_processor;
+
+ assert(processor == current_processor());
+ assert(self == current_thread());
+ assert(thread != self);
+
+ if (thread != THREAD_NULL) {
+ /*
+ * If blocked at a continuation, discard
+ * the stack.
+ */
+ if (thread->continuation != NULL && thread->kernel_stack != 0)
+ stack_free(thread);
+
+ if (thread->state & TH_IDLE) {
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED,MACH_DISPATCH) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), 0, thread->state, sched_run_count, 0);
+ } else {
+ int64_t consumed;
+ int64_t remainder = 0;
+
+ if (processor->quantum_end > processor->last_dispatch)
+ remainder = processor->quantum_end -
+ processor->last_dispatch;
+
+ consumed = thread->quantum_remaining - remainder;
+
+ if ((thread->reason & AST_LEDGER) == 0) {
+ /*
+ * Bill CPU time to both the task and
+ * the individual thread.
+ */
+ ledger_credit(thread->t_ledger,
+ task_ledgers.cpu_time, consumed);
+ ledger_credit(thread->t_threadledger,
+ thread_ledgers.cpu_time, consumed);
+#ifdef CONFIG_BANK
+ if (thread->t_bankledger) {
+ ledger_credit(thread->t_bankledger,
+ bank_ledgers.cpu_time,
+ (consumed - thread->t_deduct_bank_ledger_time));
+
+ }
+ thread->t_deduct_bank_ledger_time =0;
+#endif
+ }
+
+ wake_lock(thread);
+ thread_lock(thread);
+
+ /*
+ * Compute remainder of current quantum.
+ */
+ if (processor->first_timeslice &&
+ processor->quantum_end > processor->last_dispatch)
+ thread->quantum_remaining = (uint32_t)remainder;
+ else
+ thread->quantum_remaining = 0;
+
+ if (thread->sched_mode == TH_MODE_REALTIME) {
+ /*
+ * Cancel the deadline if the thread has
+ * consumed the entire quantum.
+ */
+ if (thread->quantum_remaining == 0) {
+ thread->realtime.deadline = UINT64_MAX;
+ }
+ } else {
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+ /*
+ * For non-realtime threads treat a tiny
+ * remaining quantum as an expired quantum
+ * but include what's left next time.
+ */
+ if (thread->quantum_remaining < min_std_quantum) {
+ thread->reason |= AST_QUANTUM;
+ thread->quantum_remaining += SCHED(initial_quantum_size)(thread);
+ }
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
+ }
+
+ /*
+ * If we are doing a direct handoff then
+ * take the remainder of the quantum.
+ */
+ if ((thread->reason & (AST_HANDOFF|AST_QUANTUM)) == AST_HANDOFF) {
+ self->quantum_remaining = thread->quantum_remaining;
+ thread->reason |= AST_QUANTUM;
+ thread->quantum_remaining = 0;
+ } else {
+#if defined(CONFIG_SCHED_MULTIQ)
+ if (SCHED(sched_groups_enabled) &&
+ thread->sched_group == self->sched_group) {
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_QUANTUM_HANDOFF),
+ self->reason, (uintptr_t)thread_tid(thread),
+ self->quantum_remaining, thread->quantum_remaining, 0);
+
+ self->quantum_remaining = thread->quantum_remaining;
+ thread->quantum_remaining = 0;
+ /* Don't set AST_QUANTUM here - old thread might still want to preempt someone else */
+ }
+#endif /* defined(CONFIG_SCHED_MULTIQ) */
+ }
+
+ thread->computation_metered += (processor->last_dispatch - thread->computation_epoch);
+
+ if ((thread->rwlock_count != 0) && !(LcksOpts & disLkRWPrio)) {
+ integer_t priority;
+
+ priority = thread->sched_pri;
+
+ if (priority < thread->base_pri)
+ priority = thread->base_pri;
+ if (priority < BASEPRI_BACKGROUND)
+ priority = BASEPRI_BACKGROUND;
+
+ if ((thread->sched_pri < priority) || !(thread->sched_flags & TH_SFLAG_RW_PROMOTED)) {
+ KERNEL_DEBUG_CONSTANT(
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_RW_PROMOTE) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), thread->sched_pri, thread->base_pri, priority, 0);
+
+ thread->sched_flags |= TH_SFLAG_RW_PROMOTED;
+
+ if (thread->sched_pri < priority)
+ set_sched_pri(thread, priority);
+ }
+ }
+
+ if (!(thread->state & TH_WAIT)) {
+ /*
+ * Still runnable.
+ */
+ thread->last_made_runnable_time = mach_approximate_time();
+
+ machine_thread_going_off_core(thread, FALSE);
+
+ if (thread->reason & AST_QUANTUM)
+ thread_setrun(thread, SCHED_TAILQ);
+ else if (thread->reason & AST_PREEMPT)
+ thread_setrun(thread, SCHED_HEADQ);
+ else
+ thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ);
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED,MACH_DISPATCH) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), thread->reason, thread->state, sched_run_count, 0);
+
+ if (thread->wake_active) {
+ thread->wake_active = FALSE;
+ thread_unlock(thread);
+
+ thread_wakeup(&thread->wake_active);
+ } else {
+ thread_unlock(thread);
+ }
+
+ wake_unlock(thread);
+ } else {
+ /*
+ * Waiting.
+ */
+ boolean_t should_terminate = FALSE;
+ uint32_t new_run_count;
+
+ /* Only the first call to thread_dispatch
+ * after explicit termination should add
+ * the thread to the termination queue
+ */
+ if ((thread->state & (TH_TERMINATE|TH_TERMINATE2)) == TH_TERMINATE) {
+ should_terminate = TRUE;
+ thread->state |= TH_TERMINATE2;
+ }
+
+ thread->state &= ~TH_RUN;
+ thread->last_made_runnable_time = ~0ULL;
+ thread->chosen_processor = PROCESSOR_NULL;
+
+ if (thread->sched_mode == TH_MODE_TIMESHARE) {
+ if (thread->sched_flags & TH_SFLAG_THROTTLED)
+ sched_background_decr(thread);
+
+ sched_share_decr(thread);
+ }
+ new_run_count = sched_run_decr(thread);
+
+#if CONFIG_SCHED_SFI
+ if ((thread->state & (TH_WAIT | TH_TERMINATE)) == TH_WAIT) {
+ if (thread->reason & AST_SFI) {
+ thread->wait_sfi_begin_time = processor->last_dispatch;
+ }
+ }
+#endif
+
+ machine_thread_going_off_core(thread, should_terminate);
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED,MACH_DISPATCH) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), thread->reason, thread->state, new_run_count, 0);
+
+ (*thread->sched_call)(SCHED_CALL_BLOCK, thread);
+
+ if (thread->wake_active) {
+ thread->wake_active = FALSE;
+ thread_unlock(thread);
+
+ thread_wakeup(&thread->wake_active);
+ } else {
+ thread_unlock(thread);
+ }
+
+ wake_unlock(thread);
+
+ if (should_terminate)
+ thread_terminate_enqueue(thread);
+ }
+ }
+ }
+
+ /* Update (new) current thread and reprogram quantum timer */
+ thread_lock(self);
+ if (!(self->state & TH_IDLE)) {
+ uint64_t arg1, arg2;
+ int urgency;
+ uint64_t latency;
+
+#if CONFIG_SCHED_SFI
+ ast_t new_ast;
+
+ new_ast = sfi_thread_needs_ast(self, NULL);
+
+ if (new_ast != AST_NONE) {
+ ast_on(new_ast);
+ }
+#endif
+
+ assert(processor->last_dispatch >= self->last_made_runnable_time);
+ latency = processor->last_dispatch - self->last_made_runnable_time;
+
+ urgency = thread_get_urgency(self, &arg1, &arg2);
+
+ thread_tell_urgency(urgency, arg1, arg2, latency, self);
+
+ machine_thread_going_on_core(self, urgency, latency);
+
+ /*
+ * Get a new quantum if none remaining.
+ */
+ if (self->quantum_remaining == 0) {
+ thread_quantum_init(self);
+ }
+
+ /*
+ * Set up quantum timer and timeslice.
+ */
+ processor->quantum_end = processor->last_dispatch + self->quantum_remaining;
+ timer_call_enter1(&processor->quantum_timer, self, processor->quantum_end, TIMER_CALL_SYS_CRITICAL | TIMER_CALL_LOCAL);
+
+ processor->first_timeslice = TRUE;
+ } else {
+ timer_call_cancel(&processor->quantum_timer);
+ processor->first_timeslice = FALSE;
+
+ thread_tell_urgency(THREAD_URGENCY_NONE, 0, 0, 0, self);
+ machine_thread_going_on_core(self, THREAD_URGENCY_NONE, 0);
+ }
+
+ self->computation_epoch = processor->last_dispatch;
+ self->reason = AST_NONE;
+
+ thread_unlock(self);
+
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ /*
+ * TODO: Can we state that redispatching our old thread is also
+ * uninteresting?
+ */
+ if ((((volatile uint32_t)sched_run_count) == 1) &&
+ !(self->state & TH_IDLE)) {
+ pset_cancel_deferred_dispatch(processor->processor_set, processor);
+ }
+#endif
+
+}
+
+/*
+ * thread_block_reason:
+ *
+ * Forces a reschedule, blocking the caller if a wait
+ * has been asserted.
+ *
+ * If a continuation is specified, then thread_invoke will
+ * attempt to discard the thread's kernel stack. When the
+ * thread resumes, it will execute the continuation function
+ * on a new kernel stack.
+ */
+counter(mach_counter_t c_thread_block_calls = 0;)
+
+wait_result_t
+thread_block_reason(
+ thread_continue_t continuation,
+ void *parameter,
+ ast_t reason)
+{
+ thread_t self = current_thread();
+ processor_t processor;
+ thread_t new_thread;
+ spl_t s;
+
+ counter(++c_thread_block_calls);
+
+ s = splsched();
+
+ processor = current_processor();
+
+ /* If we're explicitly yielding, force a subsequent quantum */
+ if (reason & AST_YIELD)
+ processor->first_timeslice = FALSE;
+
+ /* We're handling all scheduling AST's */
+ ast_off(AST_SCHEDULING);
+
+ self->continuation = continuation;
+ self->parameter = parameter;
+
+ if (self->state & ~(TH_RUN | TH_IDLE)) {
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED,MACH_BLOCK),
+ reason, VM_KERNEL_UNSLIDE(continuation), 0, 0, 0);
+ }
+
+ do {
+ thread_lock(self);
+ new_thread = thread_select(self, processor, reason);
+ thread_unlock(self);
+ } while (!thread_invoke(self, new_thread, reason));
+
+ splx(s);
+
+ return (self->wait_result);
+}
+
+/*
+ * thread_block:
+ *
+ * Block the current thread if a wait has been asserted.
+ */
+wait_result_t
+thread_block(
+ thread_continue_t continuation)
+{
+ return thread_block_reason(continuation, NULL, AST_NONE);
+}
+
+wait_result_t
+thread_block_parameter(
+ thread_continue_t continuation,
+ void *parameter)
+{
+ return thread_block_reason(continuation, parameter, AST_NONE);
+}
+
+/*
+ * thread_run:
+ *
+ * Switch directly from the current thread to the
+ * new thread, handing off our quantum if appropriate.
+ *
+ * New thread must be runnable, and not on a run queue.
+ *
+ * Called at splsched.
+ */
+int
+thread_run(
+ thread_t self,
+ thread_continue_t continuation,
+ void *parameter,
+ thread_t new_thread)
+{
+ ast_t handoff = AST_HANDOFF;
+
+ self->continuation = continuation;
+ self->parameter = parameter;
+
+ while (!thread_invoke(self, new_thread, handoff)) {
+ processor_t processor = current_processor();
+
+ thread_lock(self);
+ new_thread = thread_select(self, processor, AST_NONE);
+ thread_unlock(self);
+ handoff = AST_NONE;
+ }
+
+ return (self->wait_result);
+}
+
+/*
+ * thread_continue:
+ *
+ * Called at splsched when a thread first receives
+ * a new stack after a continuation.
+ */
+void
+thread_continue(
+ thread_t thread)
+{
+ thread_t self = current_thread();
+ thread_continue_t continuation;
+ void *parameter;
+
+ DTRACE_SCHED(on__cpu);
+
+ continuation = self->continuation;
+ parameter = self->parameter;
+
+ thread_dispatch(thread, self);
+
+ self->continuation = self->parameter = NULL;
+
+ if (thread != THREAD_NULL)
+ (void)spllo();
+
+ TLOG(1, "thread_continue: calling call_continuation \n");
+ call_continuation(continuation, parameter, self->wait_result);
+ /*NOTREACHED*/
+}
+
+void
+thread_quantum_init(thread_t thread)
+{
+ if (thread->sched_mode == TH_MODE_REALTIME) {
+ thread->quantum_remaining = thread->realtime.computation;
+ } else {
+ thread->quantum_remaining = SCHED(initial_quantum_size)(thread);
+ }
+}
+
+uint32_t
+sched_timeshare_initial_quantum_size(thread_t thread)
+{
+ if ((thread == THREAD_NULL) || !(thread->sched_flags & TH_SFLAG_THROTTLED))
+ return std_quantum;
+ else
+ return bg_quantum;
+}
+
+/*
+ * run_queue_init:
+ *
+ * Initialize a run queue before first use.
+ */
+void
+run_queue_init(
+ run_queue_t rq)
+{
+ int i;
+
+ rq->highq = IDLEPRI;
+ for (i = 0; i < NRQBM; i++)
+ rq->bitmap[i] = 0;
+ setbit(MAXPRI - IDLEPRI, rq->bitmap);
+ rq->urgency = rq->count = 0;
+ for (i = 0; i < NRQS; i++)
+ queue_init(&rq->queues[i]);
+}
+
+/*
+ * run_queue_dequeue:
+ *
+ * Perform a dequeue operation on a run queue,
+ * and return the resulting thread.
+ *
+ * The run queue must be locked (see thread_run_queue_remove()
+ * for more info), and not empty.
+ */
+thread_t
+run_queue_dequeue(
+ run_queue_t rq,
+ integer_t options)
+{
+ thread_t thread;
+ queue_t queue = rq->queues + rq->highq;
+
+ if (options & SCHED_HEADQ) {
+ thread = (thread_t)dequeue_head(queue);
+ }
+ else {
+ thread = (thread_t)dequeue_tail(queue);
+ }
+
+ thread->runq = PROCESSOR_NULL;
+ SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
+ rq->count--;
+ if (SCHED(priority_is_urgent)(rq->highq)) {
+ rq->urgency--; assert(rq->urgency >= 0);
+ }
+ if (queue_empty(queue)) {
+ if (rq->highq != IDLEPRI)
+ clrbit(MAXPRI - rq->highq, rq->bitmap);
+ rq->highq = MAXPRI - ffsbit(rq->bitmap);
+ }
+
+ return (thread);
+}
+
+/*
+ * run_queue_enqueue:
+ *
+ * Perform a enqueue operation on a run queue.
+ *
+ * The run queue must be locked (see thread_run_queue_remove()
+ * for more info).
+ */
+boolean_t
+run_queue_enqueue(
+ run_queue_t rq,
+ thread_t thread,
+ integer_t options)
+{
+ queue_t queue = rq->queues + thread->sched_pri;
+ boolean_t result = FALSE;
+
+ if (queue_empty(queue)) {
+ enqueue_tail(queue, (queue_entry_t)thread);
+
+ setbit(MAXPRI - thread->sched_pri, rq->bitmap);
+ if (thread->sched_pri > rq->highq) {
+ rq->highq = thread->sched_pri;
+ result = TRUE;
+ }
+ } else {
+ if (options & SCHED_TAILQ)
+ enqueue_tail(queue, (queue_entry_t)thread);
+ else
+ enqueue_head(queue, (queue_entry_t)thread);
+ }
+ if (SCHED(priority_is_urgent)(thread->sched_pri))
+ rq->urgency++;
+ SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
+ rq->count++;
+
+ return (result);
+
+}
+
+/*
+ * run_queue_remove:
+ *
+ * Remove a specific thread from a runqueue.
+ *
+ * The run queue must be locked.
+ */
+void
+run_queue_remove(
+ run_queue_t rq,
+ thread_t thread)
+{
+
+ remqueue((queue_entry_t)thread);
+ SCHED_STATS_RUNQ_CHANGE(&rq->runq_stats, rq->count);
+ rq->count--;
+ if (SCHED(priority_is_urgent)(thread->sched_pri)) {
+ rq->urgency--; assert(rq->urgency >= 0);
+ }
+
+ if (queue_empty(rq->queues + thread->sched_pri)) {
+ /* update run queue status */
+ if (thread->sched_pri != IDLEPRI)
+ clrbit(MAXPRI - thread->sched_pri, rq->bitmap);
+ rq->highq = MAXPRI - ffsbit(rq->bitmap);
+ }
+
+ thread->runq = PROCESSOR_NULL;
+}
+
+/* Assumes RT lock is not held, and acquires splsched/rt_lock itself */
+void
+rt_runq_scan(sched_update_scan_context_t scan_context)
+{
+ spl_t s;
+ thread_t thread;
+
+ s = splsched();
+ rt_lock_lock();
+
+ qe_foreach_element_safe(thread, &rt_runq.queue, links) {
+ if (thread->last_made_runnable_time < scan_context->earliest_rt_make_runnable_time) {
+ scan_context->earliest_rt_make_runnable_time = thread->last_made_runnable_time;
+ }
+ }
+
+ rt_lock_unlock();
+ splx(s);
+}
+
+
+/*
+ * realtime_queue_insert:
+ *
+ * Enqueue a thread for realtime execution.
+ */
+static boolean_t
+realtime_queue_insert(
+ thread_t thread)
+{
+ queue_t queue = &rt_runq.queue;
+ uint64_t deadline = thread->realtime.deadline;
+ boolean_t preempt = FALSE;
+
+ rt_lock_lock();
+
+ if (queue_empty(queue)) {
+ enqueue_tail(queue, (queue_entry_t)thread);
+ preempt = TRUE;
+ }
+ else {
+ register thread_t entry = (thread_t)queue_first(queue);
+
+ while (TRUE) {
+ if ( queue_end(queue, (queue_entry_t)entry) ||
+ deadline < entry->realtime.deadline ) {
+ entry = (thread_t)queue_prev((queue_entry_t)entry);
+ break;
+ }
+
+ entry = (thread_t)queue_next((queue_entry_t)entry);
+ }
+
+ if ((queue_entry_t)entry == queue)
+ preempt = TRUE;
+
+ insque((queue_entry_t)thread, (queue_entry_t)entry);
+ }
+
+ thread->runq = THREAD_ON_RT_RUNQ;
+ SCHED_STATS_RUNQ_CHANGE(&rt_runq.runq_stats, rt_runq.count);
+ rt_runq.count++;
+
+ rt_lock_unlock();
+
+ return (preempt);
+}
+
+/*
+ * realtime_setrun:
+ *
+ * Dispatch a thread for realtime execution.
+ *
+ * Thread must be locked. Associated pset must
+ * be locked, and is returned unlocked.
+ */
+static void
+realtime_setrun(
+ processor_t processor,
+ thread_t thread)
+{
+ processor_set_t pset = processor->processor_set;
+ ast_t preempt;
+
+ boolean_t do_signal_idle = FALSE, do_cause_ast = FALSE;
+
+ thread->chosen_processor = processor;
+
+ /* <rdar://problem/15102234> */
+ assert(thread->bound_processor == PROCESSOR_NULL);
+
+ /*
+ * Dispatch directly onto idle processor.
+ */
+ if ( (thread->bound_processor == processor)
+ && processor->state == PROCESSOR_IDLE) {
+ remqueue((queue_entry_t)processor);
+ enqueue_tail(&pset->active_queue, (queue_entry_t)processor);
+
+ processor->next_thread = thread;
+ processor->current_pri = thread->sched_pri;
+ processor->current_thmode = thread->sched_mode;
+ processor->current_sfi_class = thread->sfi_class;
+ processor->deadline = thread->realtime.deadline;
+ processor->state = PROCESSOR_DISPATCHING;
+
+ if (processor != current_processor()) {
+ if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) {
+ /* cleared on exit from main processor_idle() loop */
+ pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id);
+ do_signal_idle = TRUE;
+ }
+ }
+ pset_unlock(pset);
+
+ if (do_signal_idle) {
+ machine_signal_idle(processor);
+ }
+ return;
+ }
+
+ if (processor->current_pri < BASEPRI_RTQUEUES)
+ preempt = (AST_PREEMPT | AST_URGENT);
+ else if (thread->realtime.deadline < processor->deadline)
+ preempt = (AST_PREEMPT | AST_URGENT);
+ else
+ preempt = AST_NONE;
+
+ realtime_queue_insert(thread);
+
+ if (preempt != AST_NONE) {
+ if (processor->state == PROCESSOR_IDLE) {
+ remqueue((queue_entry_t)processor);
+ enqueue_tail(&pset->active_queue, (queue_entry_t)processor);
+ processor->next_thread = THREAD_NULL;
+ processor->current_pri = thread->sched_pri;
+ processor->current_thmode = thread->sched_mode;
+ processor->current_sfi_class = thread->sfi_class;
+ processor->deadline = thread->realtime.deadline;
+ processor->state = PROCESSOR_DISPATCHING;
+ if (processor == current_processor()) {
+ ast_on(preempt);
+ } else {
+ if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) {
+ /* cleared on exit from main processor_idle() loop */
+ pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id);
+ do_signal_idle = TRUE;
+ }
+ }
+ } else if (processor->state == PROCESSOR_DISPATCHING) {
+ if ((processor->next_thread == THREAD_NULL) && ((processor->current_pri < thread->sched_pri) || (processor->deadline > thread->realtime.deadline))) {
+ processor->current_pri = thread->sched_pri;
+ processor->current_thmode = thread->sched_mode;
+ processor->current_sfi_class = thread->sfi_class;
+ processor->deadline = thread->realtime.deadline;
+ }
+ } else {
+ if (processor == current_processor()) {
+ ast_on(preempt);
+ } else {
+ if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) {
+ /* cleared after IPI causes csw_check() to be called */
+ pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id);
+ do_cause_ast = TRUE;
+ }
+ }
+ }
+ } else {
+ /* Selected processor was too busy, just keep thread enqueued and let other processors drain it naturally. */
+ }
+
+ pset_unlock(pset);
+
+ if (do_signal_idle) {
+ machine_signal_idle(processor);
+ } else if (do_cause_ast) {
+ cause_ast_check(processor);
+ }
+}
+
+
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+
+boolean_t
+priority_is_urgent(int priority)
+{
+ return testbit(priority, sched_preempt_pri) ? TRUE : FALSE;
+}
+
+#endif /* CONFIG_SCHED_TIMESHARE_CORE */
+
+/*
+ * processor_setrun:
+ *
+ * Dispatch a thread for execution on a
+ * processor.
+ *
+ * Thread must be locked. Associated pset must
+ * be locked, and is returned unlocked.
+ */
+static void
+processor_setrun(
+ processor_t processor,
+ thread_t thread,
+ integer_t options)
+{
+ processor_set_t pset = processor->processor_set;
+ ast_t preempt;
+ enum { eExitIdle, eInterruptRunning, eDoNothing } ipi_action = eDoNothing;
+ enum { eNoSignal, eDoSignal, eDoDeferredSignal } do_signal_idle = eNoSignal;
+
+ boolean_t do_cause_ast = FALSE;
+
+ thread->chosen_processor = processor;
+
+ /*
+ * Dispatch directly onto idle processor.
+ */
+ if ( (SCHED(direct_dispatch_to_idle_processors) ||
+ thread->bound_processor == processor)
+ && processor->state == PROCESSOR_IDLE) {
+ remqueue((queue_entry_t)processor);
+ enqueue_tail(&pset->active_queue, (queue_entry_t)processor);
+
+ processor->next_thread = thread;
+ processor->current_pri = thread->sched_pri;
+ processor->current_thmode = thread->sched_mode;
+ processor->current_sfi_class = thread->sfi_class;
+ processor->deadline = UINT64_MAX;
+ processor->state = PROCESSOR_DISPATCHING;
+
+ if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) {
+ /* cleared on exit from main processor_idle() loop */
+ pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id);
+ do_signal_idle = eDoSignal;
+ }
+
+ pset_unlock(pset);
+
+ if (do_signal_idle == eDoSignal) {
+ machine_signal_idle(processor);
+ }
+
+ return;
+ }
+
+ /*
+ * Set preemption mode.
+ */
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ /* TODO: Do we need to care about urgency (see rdar://problem/20136239)? */
+#endif
+ if (SCHED(priority_is_urgent)(thread->sched_pri) && thread->sched_pri > processor->current_pri)
+ preempt = (AST_PREEMPT | AST_URGENT);
+ else if(processor->active_thread && thread_eager_preemption(processor->active_thread))
+ preempt = (AST_PREEMPT | AST_URGENT);
+ else if ((thread->sched_mode == TH_MODE_TIMESHARE) && (thread->sched_pri < thread->base_pri)) {
+ if(SCHED(priority_is_urgent)(thread->base_pri) && thread->sched_pri > processor->current_pri) {
+ preempt = (options & SCHED_PREEMPT)? AST_PREEMPT: AST_NONE;
+ } else {
+ preempt = AST_NONE;
+ }
+ } else
+ preempt = (options & SCHED_PREEMPT)? AST_PREEMPT: AST_NONE;
+
+ SCHED(processor_enqueue)(processor, thread, options);
+
+ if (preempt != AST_NONE) {
+ if (processor->state == PROCESSOR_IDLE) {
+ remqueue((queue_entry_t)processor);
+ enqueue_tail(&pset->active_queue, (queue_entry_t)processor);
+ processor->next_thread = THREAD_NULL;
+ processor->current_pri = thread->sched_pri;
+ processor->current_thmode = thread->sched_mode;
+ processor->current_sfi_class = thread->sfi_class;
+ processor->deadline = UINT64_MAX;
+ processor->state = PROCESSOR_DISPATCHING;
+
+ ipi_action = eExitIdle;
+ } else if ( processor->state == PROCESSOR_DISPATCHING) {
+ if ((processor->next_thread == THREAD_NULL) && (processor->current_pri < thread->sched_pri)) {
+ processor->current_pri = thread->sched_pri;
+ processor->current_thmode = thread->sched_mode;
+ processor->current_sfi_class = thread->sfi_class;
+ processor->deadline = UINT64_MAX;
+ }
+ } else if ( (processor->state == PROCESSOR_RUNNING ||
+ processor->state == PROCESSOR_SHUTDOWN) &&
+ (thread->sched_pri >= processor->current_pri)) {
+ ipi_action = eInterruptRunning;
+ }
+ } else {
+ /*
+ * New thread is not important enough to preempt what is running, but
+ * special processor states may need special handling
+ */
+ if (processor->state == PROCESSOR_SHUTDOWN &&
+ thread->sched_pri >= processor->current_pri ) {
+ ipi_action = eInterruptRunning;
+ } else if ( processor->state == PROCESSOR_IDLE &&
+ processor != current_processor() ) {
+ remqueue((queue_entry_t)processor);
+ enqueue_tail(&pset->active_queue, (queue_entry_t)processor);
+ processor->next_thread = THREAD_NULL;
+ processor->current_pri = thread->sched_pri;
+ processor->current_thmode = thread->sched_mode;
+ processor->current_sfi_class = thread->sfi_class;
+ processor->deadline = UINT64_MAX;
+ processor->state = PROCESSOR_DISPATCHING;
+
+ ipi_action = eExitIdle;
+ }
+ }
+
+ switch (ipi_action) {
+ case eDoNothing:
+ break;
+ case eExitIdle:
+ if (processor == current_processor()) {
+ if (csw_check_locked(processor, pset, AST_NONE) != AST_NONE)
+ ast_on(preempt);
+ } else {
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ if (!(pset->pending_deferred_AST_cpu_mask & (1ULL << processor->cpu_id)) &&
+ !(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) {
+ /* cleared on exit from main processor_idle() loop */
+ pset->pending_deferred_AST_cpu_mask |= (1ULL << processor->cpu_id);
+ do_signal_idle = eDoDeferredSignal;
+ }
+#else
+ if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) {
+ /* cleared on exit from main processor_idle() loop */
+ pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id);
+ do_signal_idle = eDoSignal;
+ }
+#endif
+ }
+ break;
+ case eInterruptRunning:
+ if (processor == current_processor()) {
+ if (csw_check_locked(processor, pset, AST_NONE) != AST_NONE)
+ ast_on(preempt);
+ } else {
+ if (!(pset->pending_AST_cpu_mask & (1ULL << processor->cpu_id))) {
+ /* cleared after IPI causes csw_check() to be called */
+ pset->pending_AST_cpu_mask |= (1ULL << processor->cpu_id);
+ do_cause_ast = TRUE;
+ }
+ }
+ break;
+ }
+
+ pset_unlock(pset);
+
+ if (do_signal_idle == eDoSignal) {
+ machine_signal_idle(processor);
+ }
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ else if (do_signal_idle == eDoDeferredSignal) {
+ /*
+ * TODO: The ability to cancel this signal could make
+ * sending it outside of the pset lock an issue. Do
+ * we need to address this? Or would the only fallout
+ * be that the core takes a signal? As long as we do
+ * not run the risk of having a core marked as signal
+ * outstanding, with no real signal outstanding, the
+ * only result should be that we fail to cancel some
+ * signals.
+ */
+ machine_signal_idle_deferred(processor);
+ }
+#endif
+ else if (do_cause_ast) {
+ cause_ast_check(processor);
+ }
+}
+
+/*
+ * choose_next_pset:
+ *
+ * Return the next sibling pset containing
+ * available processors.
+ *
+ * Returns the original pset if none other is
+ * suitable.
+ */
+static processor_set_t
+choose_next_pset(
+ processor_set_t pset)
+{
+ processor_set_t nset = pset;
+
+ do {
+ nset = next_pset(nset);
+ } while (nset->online_processor_count < 1 && nset != pset);
+
+ return (nset);
+}
+
+/*
+ * choose_processor:
+ *
+ * Choose a processor for the thread, beginning at
+ * the pset. Accepts an optional processor hint in
+ * the pset.
+ *
+ * Returns a processor, possibly from a different pset.
+ *
+ * The thread must be locked. The pset must be locked,
+ * and the resulting pset is locked on return.
+ */
+processor_t
+choose_processor(
+ processor_set_t pset,
+ processor_t processor,
+ thread_t thread)
+{
+ processor_set_t nset, cset = pset;
+
+ /*
+ * Prefer the hinted processor, when appropriate.
+ */
+
+ /* Fold last processor hint from secondary processor to its primary */
+ if (processor != PROCESSOR_NULL) {
+ processor = processor->processor_primary;
+ }
+
+ /*
+ * Only consult platform layer if pset is active, which
+ * it may not be in some cases when a multi-set system
+ * is going to sleep.
+ */
+ if (pset->online_processor_count) {
+ if ((processor == PROCESSOR_NULL) || (processor->processor_set == pset && processor->state == PROCESSOR_IDLE)) {
+ processor_t mc_processor = machine_choose_processor(pset, processor);
+ if (mc_processor != PROCESSOR_NULL)
+ processor = mc_processor->processor_primary;
+ }
+ }
+
+ /*
+ * At this point, we may have a processor hint, and we may have
+ * an initial starting pset. If the hint is not in the pset, or
+ * if the hint is for a processor in an invalid state, discard
+ * the hint.
+ */
+ if (processor != PROCESSOR_NULL) {
+ if (processor->processor_set != pset) {
+ processor = PROCESSOR_NULL;
+ } else if (!processor->is_recommended) {
+ processor = PROCESSOR_NULL;
+ } else {
+ switch (processor->state) {
+ case PROCESSOR_START:
+ case PROCESSOR_SHUTDOWN:
+ case PROCESSOR_OFF_LINE:
+ /*
+ * Hint is for a processor that cannot support running new threads.
+ */
+ processor = PROCESSOR_NULL;
+ break;
+ case PROCESSOR_IDLE:
+ /*
+ * Hint is for an idle processor. Assume it is no worse than any other
+ * idle processor. The platform layer had an opportunity to provide
+ * the "least cost idle" processor above.
+ */
+ return (processor);
+ break;
+ case PROCESSOR_RUNNING:
+ case PROCESSOR_DISPATCHING:
+ /*
+ * Hint is for an active CPU. This fast-path allows
+ * realtime threads to preempt non-realtime threads
+ * to regain their previous executing processor.
+ */
+ if ((thread->sched_pri >= BASEPRI_RTQUEUES) &&
+ (processor->current_pri < BASEPRI_RTQUEUES))
+ return (processor);
+
+ /* Otherwise, use hint as part of search below */
+ break;
+ default:
+ processor = PROCESSOR_NULL;
+ break;
+ }
+ }
+ }
+
+ /*
+ * Iterate through the processor sets to locate
+ * an appropriate processor. Seed results with
+ * a last-processor hint, if available, so that
+ * a search must find something strictly better
+ * to replace it.
+ *
+ * A primary/secondary pair of SMT processors are
+ * "unpaired" if the primary is busy but its
+ * corresponding secondary is idle (so the physical
+ * core has full use of its resources).
+ */
+
+ integer_t lowest_priority = MAXPRI + 1;
+ integer_t lowest_unpaired_primary_priority = MAXPRI + 1;
+ integer_t lowest_count = INT_MAX;
+ uint64_t furthest_deadline = 1;
+ processor_t lp_processor = PROCESSOR_NULL;
+ processor_t lp_unpaired_primary_processor = PROCESSOR_NULL;
+ processor_t lp_unpaired_secondary_processor = PROCESSOR_NULL;
+ processor_t lc_processor = PROCESSOR_NULL;
+ processor_t fd_processor = PROCESSOR_NULL;
+
+ if (processor != PROCESSOR_NULL) {
+ /* All other states should be enumerated above. */
+ assert(processor->state == PROCESSOR_RUNNING || processor->state == PROCESSOR_DISPATCHING);
+
+ lowest_priority = processor->current_pri;
+ lp_processor = processor;
+
+ if (processor->current_pri >= BASEPRI_RTQUEUES) {
+ furthest_deadline = processor->deadline;
+ fd_processor = processor;
+ }
+
+ lowest_count = SCHED(processor_runq_count)(processor);
+ lc_processor = processor;
+ }
+
+ do {
+
+ /*
+ * Choose an idle processor, in pset traversal order
+ */
+ qe_foreach_element(processor, &cset->idle_queue, processor_queue) {
+ if (processor->is_recommended)
+ return processor;
+ }
+
+ /*
+ * Otherwise, enumerate active and idle processors to find candidates
+ * with lower priority/etc.
+ */
+
+ qe_foreach_element(processor, &cset->active_queue, processor_queue) {
+
+ if (!processor->is_recommended) {
+ continue;
+ }
+
+ integer_t cpri = processor->current_pri;
+ if (cpri < lowest_priority) {
+ lowest_priority = cpri;
+ lp_processor = processor;
+ }
+
+ if ((cpri >= BASEPRI_RTQUEUES) && (processor->deadline > furthest_deadline)) {
+ furthest_deadline = processor->deadline;
+ fd_processor = processor;
+ }
+
+ integer_t ccount = SCHED(processor_runq_count)(processor);
+ if (ccount < lowest_count) {
+ lowest_count = ccount;
+ lc_processor = processor;
+ }
+ }
+
+ /*
+ * For SMT configs, these idle secondary processors must have active primary. Otherwise
+ * the idle primary would have short-circuited the loop above
+ */
+ qe_foreach_element(processor, &cset->idle_secondary_queue, processor_queue) {
+
+ if (!processor->is_recommended) {
+ continue;
+ }
+
+ processor_t cprimary = processor->processor_primary;
+
+ /* If the primary processor is offline or starting up, it's not a candidate for this path */
+ if (cprimary->state == PROCESSOR_RUNNING || cprimary->state == PROCESSOR_DISPATCHING) {
+ integer_t primary_pri = cprimary->current_pri;
+
+ if (primary_pri < lowest_unpaired_primary_priority) {
+ lowest_unpaired_primary_priority = primary_pri;
+ lp_unpaired_primary_processor = cprimary;
+ lp_unpaired_secondary_processor = processor;
+ }
+ }
+ }
+
+
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
+
+ /*
+ * For realtime threads, the most important aspect is
+ * scheduling latency, so we attempt to assign threads
+ * to good preemption candidates (assuming an idle primary
+ * processor was not available above).
+ */
+
+ if (thread->sched_pri > lowest_unpaired_primary_priority) {
+ /* Move to end of active queue so that the next thread doesn't also pick it */
+ re_queue_tail(&cset->active_queue, (queue_entry_t)lp_unpaired_primary_processor);
+ return lp_unpaired_primary_processor;
+ }
+ if (thread->sched_pri > lowest_priority) {
+ /* Move to end of active queue so that the next thread doesn't also pick it */
+ re_queue_tail(&cset->active_queue, (queue_entry_t)lp_processor);
+ return lp_processor;
+ }
+ if (thread->realtime.deadline < furthest_deadline)
+ return fd_processor;
+
+ /*
+ * If all primary and secondary CPUs are busy with realtime
+ * threads with deadlines earlier than us, move on to next
+ * pset.
+ */
+ }
+ else {
+
+ if (thread->sched_pri > lowest_unpaired_primary_priority) {
+ /* Move to end of active queue so that the next thread doesn't also pick it */
+ re_queue_tail(&cset->active_queue, (queue_entry_t)lp_unpaired_primary_processor);
+ return lp_unpaired_primary_processor;
+ }
+ if (thread->sched_pri > lowest_priority) {
+ /* Move to end of active queue so that the next thread doesn't also pick it */
+ re_queue_tail(&cset->active_queue, (queue_entry_t)lp_processor);
+ return lp_processor;
+ }
+
+ /*
+ * If all primary processor in this pset are running a higher
+ * priority thread, move on to next pset. Only when we have
+ * exhausted this search do we fall back to other heuristics.
+ */
+ }
+
+ /*
+ * Move onto the next processor set.
+ */
+ nset = next_pset(cset);
projects / apple / xnu.git / blobdiff