+ if (new_thread->sched_pri >= BASEPRI_RTQUEUES) {
+ bit_set(pset->realtime_map, processor->cpu_id);
+ } else {
+ bit_clear(pset->realtime_map, processor->cpu_id);
+ }
+
+ pset_node_t node = pset->node;
+
+ if (bit_count(node->pset_map) == 1) {
+ /* Node has only a single pset, so skip node pset map updates */
+ return;
+ }
+
+ cpumap_t avail_map = pset_available_cpumap(pset);
+
+ if (new_thread->sched_pri >= BASEPRI_RTQUEUES) {
+ if ((avail_map & pset->realtime_map) == avail_map) {
+ /* No more non-RT CPUs in this pset */
+ atomic_bit_clear(&node->pset_non_rt_map, pset->pset_id, memory_order_relaxed);
+ }
+ avail_map &= pset->primary_map;
+ if ((avail_map & pset->realtime_map) == avail_map) {
+ /* No more non-RT primary CPUs in this pset */
+ atomic_bit_clear(&node->pset_non_rt_primary_map, pset->pset_id, memory_order_relaxed);
+ }
+ } else {
+ if ((avail_map & pset->realtime_map) != avail_map) {
+ if (!bit_test(atomic_load(&node->pset_non_rt_map), pset->pset_id)) {
+ atomic_bit_set(&node->pset_non_rt_map, pset->pset_id, memory_order_relaxed);
+ }
+ }
+ avail_map &= pset->primary_map;
+ if ((avail_map & pset->realtime_map) != avail_map) {
+ if (!bit_test(atomic_load(&node->pset_non_rt_primary_map), pset->pset_id)) {
+ atomic_bit_set(&node->pset_non_rt_primary_map, pset->pset_id, memory_order_relaxed);
+ }
+ }
+ }
+}
+
+static processor_t choose_processor_for_realtime_thread(processor_set_t pset, processor_t skip_processor, bool consider_secondaries);
+static bool all_available_primaries_are_running_realtime_threads(processor_set_t pset);
+#if defined(__x86_64__)
+static bool these_processors_are_running_realtime_threads(processor_set_t pset, uint64_t these_map);
+#endif
+static bool sched_ok_to_run_realtime_thread(processor_set_t pset, processor_t processor);
+static bool processor_is_fast_track_candidate_for_realtime_thread(processor_set_t pset, processor_t processor);
+int sched_allow_rt_smt = 1;
+int sched_avoid_cpu0 = 1;
+
+/*
+ * thread_select:
+ *
+ * Select a new thread for the current processor to execute.
+ *
+ * May select the current thread, which must be locked.
+ */
+static thread_t
+thread_select(thread_t thread,
+ processor_t processor,
+ ast_t *reason)
+{
+ processor_set_t pset = processor->processor_set;
+ thread_t new_thread = THREAD_NULL;
+
+ assert(processor == current_processor());
+ assert((thread->state & (TH_RUN | TH_TERMINATE2)) == TH_RUN);
+
+ do {
+ /*
+ * Update the priority.
+ */
+ if (SCHED(can_update_priority)(thread)) {
+ SCHED(update_priority)(thread);
+ }
+
+ pset_lock(pset);
+
+ processor_state_update_from_thread(processor, thread);
+
+restart:
+ /* Acknowledge any pending IPIs here with pset lock held */
+ bit_clear(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ bit_clear(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ bit_clear(pset->pending_deferred_AST_cpu_mask, processor->cpu_id);
+#endif
+
+ bool secondary_can_only_run_realtime_thread = false;
+
+ assert(processor->state != PROCESSOR_OFF_LINE);
+
+ if (!processor->is_recommended) {
+ /*
+ * The performance controller has provided a hint to not dispatch more threads,
+ * unless they are bound to us (and thus we are the only option
+ */
+ if (!SCHED(processor_bound_count)(processor)) {
+ goto idle;
+ }
+ } else if (processor->processor_primary != processor) {
+ /*
+ * Should this secondary SMT processor attempt to find work? For pset runqueue systems,
+ * we should look for work only under the same conditions that choose_processor()
+ * would have assigned work, which is when all primary processors have been assigned work.
+ *
+ * An exception is that bound threads are dispatched to a processor without going through
+ * choose_processor(), so in those cases we should continue trying to dequeue work.
+ */
+ if (!SCHED(processor_bound_count)(processor)) {
+ if ((pset->recommended_bitmask & pset->primary_map & pset->cpu_state_map[PROCESSOR_IDLE]) != 0) {
+ goto idle;
+ }
+
+ /*
+ * TODO: What if a secondary core beat an idle primary to waking up from an IPI?
+ * Should it dequeue immediately, or spin waiting for the primary to wake up?
+ */
+
+ /* There are no idle primaries */
+
+ if (processor->processor_primary->current_pri >= BASEPRI_RTQUEUES) {
+ bool secondary_can_run_realtime_thread = sched_allow_rt_smt && rt_runq_count(pset) && all_available_primaries_are_running_realtime_threads(pset);
+ if (!secondary_can_run_realtime_thread) {
+ goto idle;
+ }
+ secondary_can_only_run_realtime_thread = true;
+ }
+ }
+ }
+
+ /*
+ * Test to see if the current thread should continue
+ * to run on this processor. Must not be attempting to wait, and not
+ * bound to a different processor, nor be in the wrong
+ * processor set, nor be forced to context switch by TH_SUSP.
+ *
+ * Note that there are never any RT threads in the regular runqueue.
+ *
+ * This code is very insanely tricky.
+ */
+
+ /* i.e. not waiting, not TH_SUSP'ed */
+ bool still_running = ((thread->state & (TH_TERMINATE | TH_IDLE | TH_WAIT | TH_RUN | TH_SUSP)) == TH_RUN);
+
+ /*
+ * Threads running on SMT processors are forced to context switch. Don't rebalance realtime threads.
+ * TODO: This should check if it's worth it to rebalance, i.e. 'are there any idle primary processors'
+ * <rdar://problem/47907700>
+ *
+ * A yielding thread shouldn't be forced to context switch.
+ */
+
+ bool is_yielding = (*reason & AST_YIELD) == AST_YIELD;
+
+ bool needs_smt_rebalance = !is_yielding && thread->sched_pri < BASEPRI_RTQUEUES && processor->processor_primary != processor;
+
+ bool affinity_mismatch = thread->affinity_set != AFFINITY_SET_NULL && thread->affinity_set->aset_pset != pset;
+
+ bool bound_elsewhere = thread->bound_processor != PROCESSOR_NULL && thread->bound_processor != processor;
+
+ bool avoid_processor = !is_yielding && SCHED(avoid_processor_enabled) && SCHED(thread_avoid_processor)(processor, thread);
+
+ if (still_running && !needs_smt_rebalance && !affinity_mismatch && !bound_elsewhere && !avoid_processor) {
+ /*
+ * This thread is eligible to keep running on this processor.
+ *
+ * RT threads with un-expired quantum stay on processor,
+ * unless there's a valid RT thread with an earlier deadline.
+ */
+ if (thread->sched_pri >= BASEPRI_RTQUEUES && processor->first_timeslice) {
+ if (rt_runq_count(pset) > 0) {
+ thread_t next_rt = qe_queue_first(&SCHED(rt_runq)(pset)->queue, struct thread, runq_links);
+
+ if (next_rt->realtime.deadline < processor->deadline &&
+ (next_rt->bound_processor == PROCESSOR_NULL ||
+ next_rt->bound_processor == processor)) {
+ /* The next RT thread is better, so pick it off the runqueue. */
+ goto pick_new_rt_thread;
+ }
+ }
+
+ /* This is still the best RT thread to run. */
+ processor->deadline = thread->realtime.deadline;
+
+ sched_update_pset_load_average(pset, 0);
+
+ processor_t next_rt_processor = PROCESSOR_NULL;
+ sched_ipi_type_t next_rt_ipi_type = SCHED_IPI_NONE;
+
+ if (rt_runq_count(pset) - bit_count(pset->pending_AST_URGENT_cpu_mask) > 0) {
+ next_rt_processor = choose_processor_for_realtime_thread(pset, processor, true);
+ if (next_rt_processor) {
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE,
+ (uintptr_t)0, (uintptr_t)-4, next_rt_processor->cpu_id, next_rt_processor->state, 0);
+ if (next_rt_processor->state == PROCESSOR_IDLE) {
+ pset_update_processor_state(pset, next_rt_processor, PROCESSOR_DISPATCHING);
+ }
+ next_rt_ipi_type = sched_ipi_action(next_rt_processor, NULL, false, SCHED_IPI_EVENT_PREEMPT);
+ }
+ }
+ pset_unlock(pset);
+
+ if (next_rt_processor) {
+ sched_ipi_perform(next_rt_processor, next_rt_ipi_type);
+ }
+
+ return thread;
+ }
+
+ if ((rt_runq_count(pset) == 0) &&
+ SCHED(processor_queue_has_priority)(processor, thread->sched_pri, TRUE) == FALSE) {
+ /* This thread is still the highest priority runnable (non-idle) thread */
+ processor->deadline = UINT64_MAX;
+
+ sched_update_pset_load_average(pset, 0);
+ pset_unlock(pset);
+
+ return thread;
+ }
+ } else {
+ /*
+ * This processor must context switch.
+ * If it's due to a rebalance, we should aggressively find this thread a new home.
+ */
+ if (needs_smt_rebalance || affinity_mismatch || bound_elsewhere || avoid_processor) {
+ *reason |= AST_REBALANCE;
+ }
+ }
+
+ bool secondary_forced_idle = ((processor->processor_secondary != PROCESSOR_NULL) &&
+ (thread_no_smt(thread) || (thread->sched_pri >= BASEPRI_RTQUEUES)) &&
+ (processor->processor_secondary->state == PROCESSOR_IDLE));
+
+ /* OK, so we're not going to run the current thread. Look at the RT queue. */
+ bool ok_to_run_realtime_thread = sched_ok_to_run_realtime_thread(pset, processor);
+ if ((rt_runq_count(pset) > 0) && ok_to_run_realtime_thread) {
+ thread_t next_rt = qe_queue_first(&SCHED(rt_runq)(pset)->queue, struct thread, runq_links);
+
+ if (__probable((next_rt->bound_processor == PROCESSOR_NULL ||
+ (next_rt->bound_processor == processor)))) {
+pick_new_rt_thread:
+ new_thread = qe_dequeue_head(&SCHED(rt_runq)(pset)->queue, struct thread, runq_links);
+
+ new_thread->runq = PROCESSOR_NULL;
+ SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset));
+ rt_runq_count_decr(pset);
+
+ processor->deadline = new_thread->realtime.deadline;
+
+ pset_commit_processor_to_new_thread(pset, processor, new_thread);
+
+ sched_update_pset_load_average(pset, 0);
+
+ processor_t ast_processor = PROCESSOR_NULL;
+ processor_t next_rt_processor = PROCESSOR_NULL;
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+ sched_ipi_type_t next_rt_ipi_type = SCHED_IPI_NONE;
+
+ if (processor->processor_secondary != NULL) {
+ processor_t sprocessor = processor->processor_secondary;
+ if ((sprocessor->state == PROCESSOR_RUNNING) || (sprocessor->state == PROCESSOR_DISPATCHING)) {
+ ipi_type = sched_ipi_action(sprocessor, NULL, false, SCHED_IPI_EVENT_SMT_REBAL);
+ ast_processor = sprocessor;
+ }
+ }
+ if (rt_runq_count(pset) - bit_count(pset->pending_AST_URGENT_cpu_mask) > 0) {
+ next_rt_processor = choose_processor_for_realtime_thread(pset, processor, true);
+ if (next_rt_processor) {
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE,
+ (uintptr_t)0, (uintptr_t)-5, next_rt_processor->cpu_id, next_rt_processor->state, 0);
+ if (next_rt_processor->state == PROCESSOR_IDLE) {
+ pset_update_processor_state(pset, next_rt_processor, PROCESSOR_DISPATCHING);
+ }
+ next_rt_ipi_type = sched_ipi_action(next_rt_processor, NULL, false, SCHED_IPI_EVENT_PREEMPT);
+ }
+ }
+ pset_unlock(pset);
+
+ if (ast_processor) {
+ sched_ipi_perform(ast_processor, ipi_type);
+ }
+
+ if (next_rt_processor) {
+ sched_ipi_perform(next_rt_processor, next_rt_ipi_type);
+ }
+
+ return new_thread;
+ }
+ }
+ if (secondary_can_only_run_realtime_thread) {
+ goto idle;
+ }
+
+ processor->deadline = UINT64_MAX;
+
+ /* No RT threads, so let's look at the regular threads. */
+ if ((new_thread = SCHED(choose_thread)(processor, MINPRI, *reason)) != THREAD_NULL) {
+ pset_commit_processor_to_new_thread(pset, processor, new_thread);
+ sched_update_pset_load_average(pset, 0);
+
+ processor_t ast_processor = PROCESSOR_NULL;
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+
+ processor_t sprocessor = processor->processor_secondary;
+ if ((sprocessor != NULL) && (sprocessor->state == PROCESSOR_RUNNING)) {
+ if (thread_no_smt(new_thread)) {
+ ipi_type = sched_ipi_action(sprocessor, NULL, false, SCHED_IPI_EVENT_SMT_REBAL);
+ ast_processor = sprocessor;
+ }
+ } else if (secondary_forced_idle && !thread_no_smt(new_thread) && pset_has_stealable_threads(pset)) {
+ pset_update_processor_state(pset, sprocessor, PROCESSOR_DISPATCHING);
+ ipi_type = sched_ipi_action(sprocessor, NULL, true, SCHED_IPI_EVENT_PREEMPT);
+ ast_processor = sprocessor;
+ }
+ pset_unlock(pset);
+
+ if (ast_processor) {
+ sched_ipi_perform(ast_processor, ipi_type);
+ }
+ return new_thread;
+ }
+
+ if (processor->must_idle) {
+ processor->must_idle = false;
+ goto idle;
+ }
+
+ if (SCHED(steal_thread_enabled)(pset) && (processor->processor_primary == processor)) {
+ /*
+ * No runnable threads, attempt to steal
+ * from other processors. Returns with pset lock dropped.
+ */
+
+ if ((new_thread = SCHED(steal_thread)(pset)) != THREAD_NULL) {
+ /*
+ * Avoid taking the pset_lock unless it is necessary to change state.
+ * It's safe to read processor->state here, as only the current processor can change state
+ * from this point (interrupts are disabled and this processor is committed to run new_thread).
+ */
+ if (processor->state == PROCESSOR_DISPATCHING || processor->state == PROCESSOR_IDLE) {
+ pset_lock(pset);
+ pset_commit_processor_to_new_thread(pset, processor, new_thread);
+ pset_unlock(pset);
+ } else {
+ assert((processor->state == PROCESSOR_RUNNING) || (processor->state == PROCESSOR_SHUTDOWN));
+ processor_state_update_from_thread(processor, new_thread);
+ }
+
+ return new_thread;
+ }
+
+ /*
+ * If other threads have appeared, shortcut
+ * around again.
+ */
+ if (!SCHED(processor_queue_empty)(processor) || (ok_to_run_realtime_thread && (rt_runq_count(pset) > 0))) {
+ continue;
+ }
+
+ pset_lock(pset);
+
+ /* Someone selected this processor while we had dropped the lock */
+ if (bit_test(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id)) {
+ goto restart;
+ }
+ }
+
+idle:
+ /*
+ * Nothing is runnable, so set this processor idle if it
+ * was running.
+ */
+ if ((processor->state == PROCESSOR_RUNNING) || (processor->state == PROCESSOR_DISPATCHING)) {
+ pset_update_processor_state(pset, processor, PROCESSOR_IDLE);
+ processor_state_update_idle(processor);
+ }
+
+ /* Invoked with pset locked, returns with pset unlocked */
+ SCHED(processor_balance)(processor, pset);
+
+ new_thread = processor->idle_thread;
+ } while (new_thread == THREAD_NULL);
+
+ return new_thread;
+}
+
+/*
+ * thread_invoke
+ *
+ * Called at splsched with neither thread locked.
+ *
+ * Perform a context switch and start executing the new thread.
+ *
+ * Returns FALSE when the context switch didn't happen.
+ * The reference to the new thread is still consumed.
+ *
+ * "self" is what is currently running on the processor,
+ * "thread" is the new thread to context switch to
+ * (which may be the same thread in some cases)
+ */
+static boolean_t
+thread_invoke(
+ thread_t self,
+ thread_t thread,
+ ast_t reason)
+{
+ if (__improbable(get_preemption_level() != 0)) {
+ int pl = get_preemption_level();
+ panic("thread_invoke: preemption_level %d, possible cause: %s",
+ pl, (pl < 0 ? "unlocking an unlocked mutex or spinlock" :
+ "blocking while holding a spinlock, or within interrupt context"));
+ }
+
+ thread_continue_t continuation = self->continuation;
+ void *parameter = self->parameter;
+ processor_t processor;
+
+ uint64_t ctime = mach_absolute_time();
+
+#ifdef CONFIG_MACH_APPROXIMATE_TIME
+ commpage_update_mach_approximate_time(ctime);
+#endif
+
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+ if (!((thread->state & TH_IDLE) != 0 ||
+ ((reason & AST_HANDOFF) && self->sched_mode == TH_MODE_REALTIME))) {
+ sched_timeshare_consider_maintenance(ctime);
+ }
+#endif
+
+#if MONOTONIC
+ mt_sched_update(self);
+#endif /* MONOTONIC */
+
+ assert_thread_magic(self);
+ assert(self == current_thread());
+ assert(self->runq == PROCESSOR_NULL);
+ assert((self->state & (TH_RUN | TH_TERMINATE2)) == TH_RUN);
+
+ thread_lock(thread);
+
+ assert_thread_magic(thread);
+ assert((thread->state & (TH_RUN | TH_WAIT | TH_UNINT | TH_TERMINATE | TH_TERMINATE2)) == TH_RUN);
+ assert(thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == current_processor());
+ assert(thread->runq == PROCESSOR_NULL);
+
+ /* Reload precise timing global policy to thread-local policy */
+ thread->precise_user_kernel_time = use_precise_user_kernel_time(thread);
+
+ /* Update SFI class based on other factors */
+ thread->sfi_class = sfi_thread_classify(thread);
+
+ /* Update the same_pri_latency for the thread (used by perfcontrol callouts) */
+ thread->same_pri_latency = ctime - thread->last_basepri_change_time;
+ /*
+ * In case a base_pri update happened between the timestamp and
+ * taking the thread lock
+ */
+ if (ctime <= thread->last_basepri_change_time) {
+ thread->same_pri_latency = ctime - thread->last_made_runnable_time;
+ }
+
+ /* Allow realtime threads to hang onto a stack. */
+ if ((self->sched_mode == TH_MODE_REALTIME) && !self->reserved_stack) {
+ self->reserved_stack = self->kernel_stack;
+ }
+
+ /* Prepare for spin debugging */
+#if INTERRUPT_MASKED_DEBUG
+ ml_spin_debug_clear(thread);
+#endif
+
+ if (continuation != NULL) {
+ if (!thread->kernel_stack) {
+ /*
+ * If we are using a privileged stack,
+ * check to see whether we can exchange it with
+ * that of the other thread.
+ */
+ if (self->kernel_stack == self->reserved_stack && !thread->reserved_stack) {
+ goto need_stack;
+ }
+
+ /*
+ * Context switch by performing a stack handoff.
+ * Requires both threads to be parked in a continuation.
+ */
+ continuation = thread->continuation;
+ parameter = thread->parameter;
+
+ processor = current_processor();
+ processor->active_thread = thread;
+ processor_state_update_from_thread(processor, thread);
+
+ if (thread->last_processor != processor && thread->last_processor != NULL) {
+ if (thread->last_processor->processor_set != processor->processor_set) {
+ thread->ps_switch++;
+ }
+ thread->p_switch++;
+ }
+ thread->last_processor = processor;
+ thread->c_switch++;
+ ast_context(thread);
+
+ thread_unlock(thread);
+
+ self->reason = reason;
+
+ processor->last_dispatch = ctime;
+ self->last_run_time = ctime;
+ processor_timer_switch_thread(ctime, &thread->system_timer);
+ timer_update(&thread->runnable_timer, ctime);
+ processor->kernel_timer = &thread->system_timer;
+
+ /*
+ * Since non-precise user/kernel time doesn't update the state timer
+ * during privilege transitions, synthesize an event now.
+ */
+ if (!thread->precise_user_kernel_time) {
+ timer_update(processor->current_state, ctime);
+ }
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_STACK_HANDOFF) | DBG_FUNC_NONE,
+ self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0);
+
+ if ((thread->chosen_processor != processor) && (thread->chosen_processor != 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);
+
+#if KPERF
+ kperf_off_cpu(self);
+#endif /* KPERF */
+
+ /*
+ * This is where we actually switch thread identity,
+ * and address space if required. However, register
+ * state is not switched - this routine leaves the
+ * stack and register state active on the current CPU.
+ */
+ TLOG(1, "thread_invoke: calling stack_handoff\n");
+ stack_handoff(self, thread);
+
+ /* 'self' is now off core */
+ assert(thread == current_thread_volatile());
+
+ DTRACE_SCHED(on__cpu);
+
+#if KPERF
+ kperf_on_cpu(thread, continuation, NULL);
+#endif /* KPERF */
+
+ thread_dispatch(self, thread);
+
+#if KASAN
+ /* Old thread's stack has been moved to the new thread, so explicitly
+ * unpoison it. */
+ kasan_unpoison_stack(thread->kernel_stack, kernel_stack_size);
+#endif
+
+ thread->continuation = thread->parameter = NULL;
+
+ boolean_t enable_interrupts = TRUE;
+
+ /* idle thread needs to stay interrupts-disabled */
+ if ((thread->state & TH_IDLE)) {
+ enable_interrupts = FALSE;
+ }
+
+ assert(continuation);
+ call_continuation(continuation, parameter,
+ thread->wait_result, enable_interrupts);
+ /*NOTREACHED*/
+ } else if (thread == self) {
+ /* same thread but with continuation */
+ ast_context(self);
+
+ thread_unlock(self);
+
+#if KPERF
+ kperf_on_cpu(thread, continuation, NULL);
+#endif /* KPERF */
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED) | DBG_FUNC_NONE,
+ self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0);
+
+#if KASAN
+ /* stack handoff to self - no thread_dispatch(), so clear the stack
+ * and free the fakestack directly */
+ kasan_fakestack_drop(self);
+ kasan_fakestack_gc(self);
+ kasan_unpoison_stack(self->kernel_stack, kernel_stack_size);
+#endif
+
+ self->continuation = self->parameter = NULL;
+
+ boolean_t enable_interrupts = TRUE;
+
+ /* idle thread needs to stay interrupts-disabled */
+ if ((self->state & TH_IDLE)) {
+ enable_interrupts = FALSE;
+ }
+
+ call_continuation(continuation, parameter,
+ self->wait_result, enable_interrupts);
+ /*NOTREACHED*/
+ }
+ } else {
+ /*
+ * Check that the other thread has a stack
+ */
+ if (!thread->kernel_stack) {
+need_stack:
+ if (!stack_alloc_try(thread)) {
+ thread_unlock(thread);
+ thread_stack_enqueue(thread);
+ return FALSE;
+ }
+ } else if (thread == self) {
+ ast_context(self);
+ thread_unlock(self);
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED) | DBG_FUNC_NONE,
+ self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0);
+
+ return TRUE;
+ }
+ }
+
+ /*
+ * Context switch by full context save.
+ */
+ processor = current_processor();
+ processor->active_thread = thread;
+ processor_state_update_from_thread(processor, thread);
+
+ if (thread->last_processor != processor && thread->last_processor != NULL) {
+ if (thread->last_processor->processor_set != processor->processor_set) {
+ thread->ps_switch++;
+ }
+ thread->p_switch++;
+ }
+ thread->last_processor = processor;
+ thread->c_switch++;
+ ast_context(thread);
+
+ thread_unlock(thread);
+
+ self->reason = reason;
+
+ processor->last_dispatch = ctime;
+ self->last_run_time = ctime;
+ processor_timer_switch_thread(ctime, &thread->system_timer);
+ timer_update(&thread->runnable_timer, ctime);
+ processor->kernel_timer = &thread->system_timer;
+
+ /*
+ * Since non-precise user/kernel time doesn't update the state timer
+ * during privilege transitions, synthesize an event now.
+ */
+ if (!thread->precise_user_kernel_time) {
+ timer_update(processor->current_state, ctime);
+ }
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED) | DBG_FUNC_NONE,
+ self->reason, (uintptr_t)thread_tid(thread), self->sched_pri, thread->sched_pri, 0);
+
+ 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);
+
+#if KPERF
+ kperf_off_cpu(self);
+#endif /* KPERF */
+
+ /*
+ * 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.
+ *
+ * If one of the threads is using a continuation, thread_continue
+ * is used to stitch up its context.
+ *
+ * If we are invoking a thread which is resuming from a continuation,
+ * the CPU will invoke thread_continue next.
+ *
+ * If the current thread is parking in a continuation, then its state
+ * won't be saved and the stack will be discarded. When the stack is
+ * re-allocated, it will be configured to resume from thread_continue.
+ */
+ assert(continuation == self->continuation);
+ thread = machine_switch_context(self, continuation, thread);
+ assert(self == current_thread_volatile());
+ TLOG(1, "thread_invoke: returning machine_switch_context: self %p continuation %p thread %p\n", self, continuation, thread);
+
+ assert(continuation == NULL && self->continuation == NULL);
+
+ DTRACE_SCHED(on__cpu);
+
+#if KPERF
+ kperf_on_cpu(self, NULL, __builtin_frame_address(0));
+#endif /* KPERF */
+
+ /* We have been resumed and are set to run. */
+ thread_dispatch(thread, self);
+
+ 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 = os_atomic_load(&sched_run_buckets[TH_BUCKET_RUN], relaxed);
+
+ /*
+ * 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)) {
+ uint64_t dispatching_map = (pset->cpu_state_map[PROCESSOR_DISPATCHING] &
+ pset->pending_deferred_AST_cpu_mask &
+ ~pset->pending_AST_URGENT_cpu_mask);
+ for (int cpuid = lsb_first(dispatching_map); cpuid >= 0; cpuid = lsb_next(dispatching_map, cpuid)) {
+ active_processor = processor_array[cpuid];
+ /*
+ * 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 != processor) {
+ /*
+ * Squash all of the processor state back to some
+ * reasonable facsimile of PROCESSOR_IDLE.
+ */
+
+ processor_state_update_idle(active_processor);
+ active_processor->deadline = UINT64_MAX;
+ pset_update_processor_state(pset, active_processor, PROCESSOR_IDLE);
+ bit_clear(pset->pending_deferred_AST_cpu_mask, 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
+
+static void
+thread_csw_callout(
+ thread_t old,
+ thread_t new,
+ uint64_t timestamp)
+{
+ perfcontrol_event event = (new->state & TH_IDLE) ? IDLE : CONTEXT_SWITCH;
+ uint64_t same_pri_latency = (new->state & TH_IDLE) ? 0 : new->same_pri_latency;
+ machine_switch_perfcontrol_context(event, timestamp, 0,
+ same_pri_latency, old, new);
+}
+
+
+/*
+ * 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;
+ bool was_idle = false;
+
+ assert(processor == current_processor());
+ assert(self == current_thread_volatile());
+ assert(thread != self);
+
+ if (thread != THREAD_NULL) {
+ /*
+ * Do the perfcontrol callout for context switch.
+ * The reason we do this here is:
+ * - thread_dispatch() is called from various places that are not
+ * the direct context switch path for eg. processor shutdown etc.
+ * So adding the callout here covers all those cases.
+ * - We want this callout as early as possible to be close
+ * to the timestamp taken in thread_invoke()
+ * - We want to avoid holding the thread lock while doing the
+ * callout
+ * - We do not want to callout if "thread" is NULL.
+ */
+ thread_csw_callout(thread, self, processor->last_dispatch);
+
+#if KASAN
+ if (thread->continuation != NULL) {
+ /*
+ * Thread has a continuation and the normal stack is going away.
+ * Unpoison the stack and mark all fakestack objects as unused.
+ */
+ kasan_fakestack_drop(thread);
+ if (thread->kernel_stack) {
+ kasan_unpoison_stack(thread->kernel_stack, kernel_stack_size);
+ }
+ }
+
+ /*
+ * Free all unused fakestack objects.
+ */
+ kasan_fakestack_gc(thread);
+#endif
+
+ /*
+ * If blocked at a continuation, discard
+ * the stack.
+ */
+ if (thread->continuation != NULL && thread->kernel_stack != 0) {
+ stack_free(thread);
+ }
+
+ if (thread->state & TH_IDLE) {
+ was_idle = true;
+ 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_buckets[TH_BUCKET_RUN], 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(thread, thread->t_ledger,
+ task_ledgers.cpu_time, consumed);
+ ledger_credit_thread(thread, thread->t_threadledger,
+ thread_ledgers.cpu_time, consumed);
+ if (thread->t_bankledger) {
+ ledger_credit_thread(thread, thread->t_bankledger,
+ bank_ledgers.cpu_time,
+ (consumed - thread->t_deduct_bank_ledger_time));
+ }
+ thread->t_deduct_bank_ledger_time = 0;
+ if (consumed > 0) {
+ /*
+ * This should never be negative, but in traces we are seeing some instances
+ * of consumed being negative.
+ * <rdar://problem/57782596> thread_dispatch() thread CPU consumed calculation sometimes results in negative value
+ */
+ sched_update_pset_avg_execution_time(current_processor()->processor_set, consumed, processor->last_dispatch, thread->th_sched_bucket);
+ }
+ }
+
+ wake_lock(thread);
+ thread_lock(thread);
+
+ /*
+ * Apply a priority floor if the thread holds a kernel resource
+ * Do this before checking starting_pri to avoid overpenalizing
+ * repeated rwlock blockers.
+ */
+ if (__improbable(thread->rwlock_count != 0)) {
+ lck_rw_set_promotion_locked(thread);
+ }
+
+ boolean_t keep_quantum = processor->first_timeslice;
+
+ /*
+ * Treat a thread which has dropped priority since it got on core
+ * as having expired its quantum.
+ */
+ if (processor->starting_pri > thread->sched_pri) {
+ keep_quantum = FALSE;
+ }
+
+ /* Compute remainder of current quantum. */
+ if (keep_quantum &&
+ 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) {
+ KDBG(MACHDBG_CODE(DBG_MACH_SCHED, MACH_CANCEL_RT_DEADLINE) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), thread->realtime.deadline, thread->realtime.computation, 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->state & TH_WAIT)) {
+ /*
+ * Still runnable.
+ */
+ thread->last_made_runnable_time = thread->last_basepri_change_time = processor->last_dispatch;
+
+ machine_thread_going_off_core(thread, FALSE, processor->last_dispatch, TRUE);
+
+ ast_t reason = thread->reason;
+ sched_options_t options = SCHED_NONE;
+
+ if (reason & AST_REBALANCE) {
+ options |= SCHED_REBALANCE;
+ if (reason & AST_QUANTUM) {
+ /*
+ * Having gone to the trouble of forcing this thread off a less preferred core,
+ * we should force the preferable core to reschedule immediately to give this
+ * thread a chance to run instead of just sitting on the run queue where
+ * it may just be stolen back by the idle core we just forced it off.
+ * But only do this at the end of a quantum to prevent cascading effects.
+ */
+ options |= SCHED_PREEMPT;
+ }
+ }
+
+ if (reason & AST_QUANTUM) {
+ options |= SCHED_TAILQ;
+ } else if (reason & AST_PREEMPT) {
+ options |= SCHED_HEADQ;
+ } else {
+ options |= (SCHED_PREEMPT | SCHED_TAILQ);
+ }
+
+ thread_setrun(thread, options);
+
+ 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_buckets[TH_BUCKET_RUN], 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;
+ int thread_state = thread->state;
+
+ /* 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;
+ }
+
+ timer_stop(&thread->runnable_timer, processor->last_dispatch);
+
+ thread_state &= ~TH_RUN;
+ thread->state = thread_state;
+
+ thread->last_made_runnable_time = thread->last_basepri_change_time = THREAD_NOT_RUNNABLE;
+ thread->chosen_processor = PROCESSOR_NULL;
+
+ new_run_count = SCHED(run_count_decr)(thread);
+
+#if CONFIG_SCHED_AUTO_JOIN
+ if ((thread->sched_flags & TH_SFLAG_THREAD_GROUP_AUTO_JOIN) != 0) {
+ work_interval_auto_join_unwind(thread);
+ }
+#endif /* CONFIG_SCHED_AUTO_JOIN */
+
+#if CONFIG_SCHED_SFI
+ if (thread->reason & AST_SFI) {
+ thread->wait_sfi_begin_time = processor->last_dispatch;
+ }
+#endif
+ machine_thread_going_off_core(thread, should_terminate, processor->last_dispatch, FALSE);
+
+ 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);
+
+ if (thread_state & TH_WAIT_REPORT) {
+ (*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);
+ }
+ }
+ }
+ /*
+ * The thread could have been added to the termination queue, so it's
+ * unsafe to use after this point.
+ */
+ thread = THREAD_NULL;
+ }
+
+ int urgency = THREAD_URGENCY_NONE;
+ uint64_t latency = 0;
+
+ /* Update (new) current thread and reprogram running timers */
+ thread_lock(self);
+
+ if (!(self->state & TH_IDLE)) {
+ uint64_t arg1, arg2;
+
+#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
+
+ assertf(processor->last_dispatch >= self->last_made_runnable_time,
+ "Non-monotonic time? dispatch at 0x%llx, runnable at 0x%llx",
+ processor->last_dispatch, self->last_made_runnable_time);
+
+ assert(self->last_made_runnable_time <= self->last_basepri_change_time);
+
+ latency = processor->last_dispatch - self->last_made_runnable_time;
+ assert(latency >= self->same_pri_latency);
+
+ urgency = thread_get_urgency(self, &arg1, &arg2);
+
+ thread_tell_urgency(urgency, arg1, arg2, latency, self);
+
+ /*
+ * 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;
+
+ running_timer_setup(processor, RUNNING_TIMER_QUANTUM, self,
+ processor->quantum_end, processor->last_dispatch);
+ if (was_idle) {
+ /*
+ * kperf's running timer is active whenever the idle thread for a
+ * CPU is not running.
+ */
+ kperf_running_setup(processor, processor->last_dispatch);
+ }
+ running_timers_activate(processor);
+ processor->first_timeslice = TRUE;
+ } else {
+ running_timers_deactivate(processor);
+ processor->first_timeslice = FALSE;
+ thread_tell_urgency(THREAD_URGENCY_NONE, 0, 0, 0, self);
+ }
+
+ assert(self->block_hint == kThreadWaitNone);
+ self->computation_epoch = processor->last_dispatch;
+ self->reason = AST_NONE;
+ processor->starting_pri = self->sched_pri;
+
+ thread_unlock(self);
+
+ machine_thread_going_on_core(self, urgency, latency, self->same_pri_latency,
+ processor->last_dispatch);
+
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ /*
+ * TODO: Can we state that redispatching our old thread is also
+ * uninteresting?
+ */
+ if ((os_atomic_load(&sched_run_buckets[TH_BUCKET_RUN], relaxed) == 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.
+ */
+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;
+
+ 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);
+
+#if PROC_REF_DEBUG
+ if ((continuation != NULL) && (self->task != kernel_task)) {
+ if (uthread_get_proc_refcount(self->uthread) != 0) {
+ panic("thread_block_reason with continuation uthread %p with uu_proc_refcount != 0", self->uthread);
+ }
+ }
+#endif
+
+ 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 reason = AST_NONE;
+
+ if ((self->state & TH_IDLE) == 0) {
+ reason = AST_HANDOFF;
+ }
+
+ /*
+ * If this thread hadn't been setrun'ed, it
+ * might not have a chosen processor, so give it one
+ */
+ if (new_thread->chosen_processor == NULL) {
+ new_thread->chosen_processor = current_processor();
+ }
+
+ self->continuation = continuation;
+ self->parameter = parameter;
+
+ while (!thread_invoke(self, new_thread, reason)) {
+ /* the handoff failed, so we have to fall back to the normal block path */
+ processor_t processor = current_processor();
+
+ reason = AST_NONE;
+
+ thread_lock(self);
+ new_thread = thread_select(self, processor, &reason);
+ thread_unlock(self);
+ }
+
+ return self->wait_result;
+}
+
+/*
+ * thread_continue:
+ *
+ * Called at splsched when a thread first receives
+ * a new stack after a continuation.
+ *
+ * Called with THREAD_NULL as the old thread when
+ * invoked by machine_load_context.
+ */
+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;
+
+ assert(continuation != NULL);
+
+#if KPERF
+ kperf_on_cpu(self, continuation, NULL);
+#endif
+
+ thread_dispatch(thread, self);
+
+ self->continuation = self->parameter = NULL;
+
+#if INTERRUPT_MASKED_DEBUG
+ /* Reset interrupt-masked spin debugging timeout */
+ ml_spin_debug_clear(self);
+#endif
+
+ TLOG(1, "thread_continue: calling call_continuation\n");
+
+ boolean_t enable_interrupts = TRUE;
+
+ /* bootstrap thread, idle thread need to stay interrupts-disabled */
+ if (thread == THREAD_NULL || (self->state & TH_IDLE)) {
+ enable_interrupts = FALSE;
+ }
+
+ call_continuation(continuation, parameter, self->wait_result, enable_interrupts);
+ /*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->th_sched_bucket == TH_BUCKET_SHARE_BG) {
+ return bg_quantum;
+ } else {
+ return std_quantum;
+ }
+}
+
+/*
+ * run_queue_init:
+ *
+ * Initialize a run queue before first use.
+ */
+void
+run_queue_init(
+ run_queue_t rq)
+{
+ rq->highq = NOPRI;
+ for (u_int i = 0; i < BITMAP_LEN(NRQS); i++) {
+ rq->bitmap[i] = 0;
+ }
+ rq->urgency = rq->count = 0;
+ for (int i = 0; i < NRQS; i++) {
+ circle_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,
+ sched_options_t options)
+{
+ thread_t thread;
+ circle_queue_t queue = &rq->queues[rq->highq];
+
+ if (options & SCHED_HEADQ) {
+ thread = cqe_dequeue_head(queue, struct thread, runq_links);
+ } else {
+ thread = cqe_dequeue_tail(queue, struct thread, runq_links);
+ }
+
+ assert(thread != THREAD_NULL);
+ assert_thread_magic(thread);
+
+ 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 (circle_queue_empty(queue)) {
+ bitmap_clear(rq->bitmap, rq->highq);
+ rq->highq = bitmap_first(rq->bitmap, NRQS);
+ }
+
+ 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,
+ sched_options_t options)
+{
+ circle_queue_t queue = &rq->queues[thread->sched_pri];
+ boolean_t result = FALSE;
+
+ assert_thread_magic(thread);
+
+ if (circle_queue_empty(queue)) {
+ circle_enqueue_tail(queue, &thread->runq_links);
+
+ rq_bitmap_set(rq->bitmap, thread->sched_pri);
+ if (thread->sched_pri > rq->highq) {
+ rq->highq = thread->sched_pri;
+ result = TRUE;
+ }
+ } else {
+ if (options & SCHED_TAILQ) {
+ circle_enqueue_tail(queue, &thread->runq_links);
+ } else {
+ circle_enqueue_head(queue, &thread->runq_links);
+ }
+ }
+ 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)
+{
+ circle_queue_t queue = &rq->queues[thread->sched_pri];
+
+ assert(thread->runq != PROCESSOR_NULL);
+ assert_thread_magic(thread);
+
+ circle_dequeue(queue, &thread->runq_links);
+ 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 (circle_queue_empty(queue)) {
+ /* update run queue status */
+ bitmap_clear(rq->bitmap, thread->sched_pri);
+ rq->highq = bitmap_first(rq->bitmap, NRQS);
+ }
+
+ thread->runq = PROCESSOR_NULL;
+}
+
+/*
+ * run_queue_peek
+ *
+ * Peek at the runq and return the highest
+ * priority thread from the runq.
+ *
+ * The run queue must be locked.
+ */
+thread_t
+run_queue_peek(
+ run_queue_t rq)
+{
+ if (rq->count > 0) {
+ circle_queue_t queue = &rq->queues[rq->highq];
+ thread_t thread = cqe_queue_first(queue, struct thread, runq_links);
+ assert_thread_magic(thread);
+ return thread;
+ } else {
+ return THREAD_NULL;
+ }
+}
+
+rt_queue_t
+sched_rtlocal_runq(processor_set_t pset)
+{
+ return &pset->rt_runq;
+}
+
+void
+sched_rtlocal_init(processor_set_t pset)
+{
+ pset_rt_init(pset);
+}
+
+void
+sched_rtlocal_queue_shutdown(processor_t processor)
+{
+ processor_set_t pset = processor->processor_set;
+ thread_t thread;
+ queue_head_t tqueue;
+
+ pset_lock(pset);
+
+ /* We only need to migrate threads if this is the last active or last recommended processor in the pset */
+ if ((pset->online_processor_count > 0) && pset_is_recommended(pset)) {
+ pset_unlock(pset);
+ return;
+ }
+
+ queue_init(&tqueue);
+
+ while (rt_runq_count(pset) > 0) {
+ thread = qe_dequeue_head(&pset->rt_runq.queue, struct thread, runq_links);
+ thread->runq = PROCESSOR_NULL;
+ SCHED_STATS_RUNQ_CHANGE(&pset->rt_runq.runq_stats, rt_runq_count(pset));
+ rt_runq_count_decr(pset);
+ enqueue_tail(&tqueue, &thread->runq_links);
+ }
+ sched_update_pset_load_average(pset, 0);
+ pset_unlock(pset);
+
+ qe_foreach_element_safe(thread, &tqueue, runq_links) {
+ remqueue(&thread->runq_links);
+
+ thread_lock(thread);
+
+ thread_setrun(thread, SCHED_TAILQ);
+
+ thread_unlock(thread);
+ }
+}
+
+/* Assumes RT lock is not held, and acquires splsched/rt_lock itself */
+void
+sched_rtlocal_runq_scan(sched_update_scan_context_t scan_context)
+{
+ thread_t thread;
+
+ pset_node_t node = &pset_node0;
+ processor_set_t pset = node->psets;
+
+ spl_t s = splsched();
+ do {
+ while (pset != NULL) {
+ pset_lock(pset);
+
+ qe_foreach_element_safe(thread, &pset->rt_runq.queue, runq_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;
+ }
+ }
+
+ pset_unlock(pset);
+
+ pset = pset->pset_list;
+ }
+ } while (((node = node->node_list) != NULL) && ((pset = node->psets) != NULL));
+ splx(s);
+}
+
+int64_t
+sched_rtlocal_runq_count_sum(void)
+{
+ pset_node_t node = &pset_node0;
+ processor_set_t pset = node->psets;
+ int64_t count = 0;
+
+ do {
+ while (pset != NULL) {
+ count += pset->rt_runq.runq_stats.count_sum;
+
+ pset = pset->pset_list;
+ }
+ } while (((node = node->node_list) != NULL) && ((pset = node->psets) != NULL));
+
+ return count;
+}
+
+/*
+ * realtime_queue_insert:
+ *
+ * Enqueue a thread for realtime execution.
+ */
+static boolean_t
+realtime_queue_insert(processor_t processor, processor_set_t pset, thread_t thread)
+{
+ queue_t queue = &SCHED(rt_runq)(pset)->queue;
+ uint64_t deadline = thread->realtime.deadline;
+ boolean_t preempt = FALSE;
+
+ pset_assert_locked(pset);
+
+ if (queue_empty(queue)) {
+ enqueue_tail(queue, &thread->runq_links);
+ preempt = TRUE;
+ } else {
+ /* Insert into rt_runq in thread deadline order */
+ queue_entry_t iter;
+ qe_foreach(iter, queue) {
+ thread_t iter_thread = qe_element(iter, struct thread, runq_links);
+ assert_thread_magic(iter_thread);
+
+ if (deadline < iter_thread->realtime.deadline) {
+ if (iter == queue_first(queue)) {
+ preempt = TRUE;
+ }
+ insque(&thread->runq_links, queue_prev(iter));
+ break;
+ } else if (iter == queue_last(queue)) {
+ enqueue_tail(queue, &thread->runq_links);
+ break;
+ }
+ }
+ }
+
+ thread->runq = processor;
+ SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset));
+ rt_runq_count_incr(pset);
+
+ return preempt;
+}
+
+#define MAX_BACKUP_PROCESSORS 7
+#if defined(__x86_64__)
+#define DEFAULT_BACKUP_PROCESSORS 1
+#else
+#define DEFAULT_BACKUP_PROCESSORS 0
+#endif
+
+int sched_rt_n_backup_processors = DEFAULT_BACKUP_PROCESSORS;
+
+int
+sched_get_rt_n_backup_processors(void)
+{
+ return sched_rt_n_backup_processors;
+}
+
+void
+sched_set_rt_n_backup_processors(int n)
+{
+ if (n < 0) {
+ n = 0;
+ } else if (n > MAX_BACKUP_PROCESSORS) {
+ n = MAX_BACKUP_PROCESSORS;
+ }
+
+ sched_rt_n_backup_processors = n;
+}
+
+/*
+ * 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 chosen_processor,
+ thread_t thread)
+{
+ processor_set_t pset = chosen_processor->processor_set;
+ pset_assert_locked(pset);
+ ast_t preempt;
+
+ int n_backup = 0;
+
+ if (thread->realtime.constraint <= rt_constraint_threshold) {
+ n_backup = sched_rt_n_backup_processors;
+ }
+ assert((n_backup >= 0) && (n_backup <= MAX_BACKUP_PROCESSORS));
+
+ sched_ipi_type_t ipi_type[MAX_BACKUP_PROCESSORS + 1] = {};
+ processor_t ipi_processor[MAX_BACKUP_PROCESSORS + 1] = {};
+
+ thread->chosen_processor = chosen_processor;
+
+ /* <rdar://problem/15102234> */
+ assert(thread->bound_processor == PROCESSOR_NULL);
+
+ realtime_queue_insert(chosen_processor, pset, thread);
+
+ processor_t processor = chosen_processor;
+ bool chosen_process_is_secondary = chosen_processor->processor_primary != chosen_processor;
+
+ int count = 0;
+ for (int i = 0; i <= n_backup; i++) {
+ if (i > 0) {
+ processor = choose_processor_for_realtime_thread(pset, chosen_processor, chosen_process_is_secondary);
+ if ((processor == PROCESSOR_NULL) || (sched_avoid_cpu0 && (processor->cpu_id == 0))) {
+ break;
+ }
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), (uintptr_t)-3, processor->cpu_id, processor->state, 0);
+ }
+ ipi_type[i] = SCHED_IPI_NONE;
+ ipi_processor[i] = processor;
+ count++;
+
+ 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;
+ }
+
+ if (preempt != AST_NONE) {
+ if (processor->state == PROCESSOR_IDLE) {
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = thread->realtime.deadline;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+ if (processor == current_processor()) {
+ ast_on(preempt);
+
+ if ((preempt & AST_URGENT) == AST_URGENT) {
+ bit_set(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ }
+
+ if ((preempt & AST_PREEMPT) == AST_PREEMPT) {
+ bit_set(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+ }
+ } else {
+ ipi_type[i] = sched_ipi_action(processor, thread, true, SCHED_IPI_EVENT_PREEMPT);
+ }
+ } else if (processor->state == PROCESSOR_DISPATCHING) {
+ if ((processor->current_pri < thread->sched_pri) || (processor->deadline > thread->realtime.deadline)) {
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = thread->realtime.deadline;
+ }
+ } else {
+ if (processor == current_processor()) {
+ ast_on(preempt);
+
+ if ((preempt & AST_URGENT) == AST_URGENT) {
+ bit_set(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ }
+
+ if ((preempt & AST_PREEMPT) == AST_PREEMPT) {
+ bit_set(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+ }
+ } else {
+ ipi_type[i] = sched_ipi_action(processor, thread, false, SCHED_IPI_EVENT_PREEMPT);
+ }
+ }
+ } else {
+ /* Selected processor was too busy, just keep thread enqueued and let other processors drain it naturally. */
+ }
+ }
+
+ pset_unlock(pset);
+
+ assert((count > 0) && (count <= (n_backup + 1)));
+ for (int i = 0; i < count; i++) {
+ assert(ipi_processor[i] != PROCESSOR_NULL);
+ sched_ipi_perform(ipi_processor[i], ipi_type[i]);
+ }
+}
+
+
+sched_ipi_type_t
+sched_ipi_deferred_policy(processor_set_t pset, processor_t dst,
+ __unused sched_ipi_event_t event)
+{
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ if (!bit_test(pset->pending_deferred_AST_cpu_mask, dst->cpu_id)) {
+ return SCHED_IPI_DEFERRED;
+ }
+#else /* CONFIG_SCHED_DEFERRED_AST */
+ panic("Request for deferred IPI on an unsupported platform; pset: %p CPU: %d", pset, dst->cpu_id);
+#endif /* CONFIG_SCHED_DEFERRED_AST */
+ return SCHED_IPI_NONE;
+}
+
+sched_ipi_type_t
+sched_ipi_action(processor_t dst, thread_t thread, boolean_t dst_idle, sched_ipi_event_t event)
+{
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+ assert(dst != NULL);
+
+ processor_set_t pset = dst->processor_set;
+ if (current_processor() == dst) {
+ return SCHED_IPI_NONE;
+ }
+
+ if (bit_test(pset->pending_AST_URGENT_cpu_mask, dst->cpu_id)) {
+ return SCHED_IPI_NONE;
+ }
+
+ ipi_type = SCHED(ipi_policy)(dst, thread, dst_idle, event);
+ switch (ipi_type) {
+ case SCHED_IPI_NONE:
+ return SCHED_IPI_NONE;
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ case SCHED_IPI_DEFERRED:
+ bit_set(pset->pending_deferred_AST_cpu_mask, dst->cpu_id);
+ break;
+#endif /* CONFIG_SCHED_DEFERRED_AST */
+ default:
+ bit_set(pset->pending_AST_URGENT_cpu_mask, dst->cpu_id);
+ bit_set(pset->pending_AST_PREEMPT_cpu_mask, dst->cpu_id);
+ break;
+ }
+ return ipi_type;
+}
+
+sched_ipi_type_t
+sched_ipi_policy(processor_t dst, thread_t thread, boolean_t dst_idle, sched_ipi_event_t event)
+{
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+ boolean_t deferred_ipi_supported = false;
+ processor_set_t pset = dst->processor_set;
+
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ deferred_ipi_supported = true;
+#endif /* CONFIG_SCHED_DEFERRED_AST */
+
+ switch (event) {
+ case SCHED_IPI_EVENT_SPILL:
+ case SCHED_IPI_EVENT_SMT_REBAL:
+ case SCHED_IPI_EVENT_REBALANCE:
+ case SCHED_IPI_EVENT_BOUND_THR:
+ /*
+ * The spill, SMT rebalance, rebalance and the bound thread
+ * scenarios use immediate IPIs always.
+ */
+ ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE;
+ break;
+ case SCHED_IPI_EVENT_PREEMPT:
+ /* In the preemption case, use immediate IPIs for RT threads */
+ if (thread && (thread->sched_pri >= BASEPRI_RTQUEUES)) {
+ ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE;
+ break;
+ }
+
+ /*
+ * For Non-RT threads preemption,
+ * If the core is active, use immediate IPIs.
+ * If the core is idle, use deferred IPIs if supported; otherwise immediate IPI.
+ */
+ if (deferred_ipi_supported && dst_idle) {
+ return sched_ipi_deferred_policy(pset, dst, event);
+ }
+ ipi_type = dst_idle ? SCHED_IPI_IDLE : SCHED_IPI_IMMEDIATE;
+ break;
+ default:
+ panic("Unrecognized scheduler IPI event type %d", event);
+ }
+ assert(ipi_type != SCHED_IPI_NONE);
+ return ipi_type;
+}
+
+void
+sched_ipi_perform(processor_t dst, sched_ipi_type_t ipi)
+{
+ switch (ipi) {
+ case SCHED_IPI_NONE:
+ break;
+ case SCHED_IPI_IDLE:
+ machine_signal_idle(dst);
+ break;
+ case SCHED_IPI_IMMEDIATE:
+ cause_ast_check(dst);
+ break;
+ case SCHED_IPI_DEFERRED:
+ machine_signal_idle_deferred(dst);
+ break;
+ default:
+ panic("Unrecognized scheduler IPI type: %d", ipi);
+ }
+}
+
+#if defined(CONFIG_SCHED_TIMESHARE_CORE)
+
+boolean_t
+priority_is_urgent(int priority)
+{
+ return bitmap_test(sched_preempt_pri, priority) ? 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;
+ pset_assert_locked(pset);
+ ast_t preempt;
+ enum { eExitIdle, eInterruptRunning, eDoNothing } ipi_action = eDoNothing;
+
+ sched_ipi_type_t ipi_type = SCHED_IPI_NONE;
+
+ thread->chosen_processor = processor;
+
+ /*
+ * 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->current_is_eagerpreempt) {
+ 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;
+ }
+
+ if ((options & (SCHED_PREEMPT | SCHED_REBALANCE)) == (SCHED_PREEMPT | SCHED_REBALANCE)) {
+ /*
+ * Having gone to the trouble of forcing this thread off a less preferred core,
+ * we should force the preferable core to reschedule immediately to give this
+ * thread a chance to run instead of just sitting on the run queue where
+ * it may just be stolen back by the idle core we just forced it off.
+ */
+ preempt |= AST_PREEMPT;
+ }
+
+ SCHED(processor_enqueue)(processor, thread, options);
+ sched_update_pset_load_average(pset, 0);
+
+ if (preempt != AST_NONE) {
+ if (processor->state == PROCESSOR_IDLE) {
+ processor_state_update_from_thread(processor, thread);
+ processor->deadline = UINT64_MAX;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+ ipi_action = eExitIdle;
+ } else if (processor->state == PROCESSOR_DISPATCHING) {
+ if (processor->current_pri < thread->sched_pri) {
+ processor_state_update_from_thread(processor, thread);
+ 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_state_update_from_thread(processor, thread);
+ processor->deadline = UINT64_MAX;
+ pset_update_processor_state(pset, processor, PROCESSOR_DISPATCHING);
+
+ ipi_action = eExitIdle;
+ }
+ }
+
+ if (ipi_action != eDoNothing) {
+ if (processor == current_processor()) {
+ if ((preempt = csw_check_locked(processor->active_thread, processor, pset, AST_NONE)) != AST_NONE) {
+ ast_on(preempt);
+ }
+
+ if ((preempt & AST_URGENT) == AST_URGENT) {
+ bit_set(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ } else {
+ bit_clear(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ }
+
+ if ((preempt & AST_PREEMPT) == AST_PREEMPT) {
+ bit_set(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+ } else {
+ bit_clear(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+ }
+ } else {
+ sched_ipi_event_t event = (options & SCHED_REBALANCE) ? SCHED_IPI_EVENT_REBALANCE : SCHED_IPI_EVENT_PREEMPT;
+ ipi_type = sched_ipi_action(processor, thread, (ipi_action == eExitIdle), event);
+ }
+ }
+ pset_unlock(pset);
+ sched_ipi_perform(processor, ipi_type);
+}
+
+/*
+ * 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;
+}
+
+inline static processor_set_t
+change_locked_pset(processor_set_t current_pset, processor_set_t new_pset)
+{
+ if (current_pset != new_pset) {
+ pset_unlock(current_pset);
+ pset_lock(new_pset);
+ }
+
+ return new_pset;
+}
+
+/*
+ * 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 starting_pset,
+ processor_t processor,
+ thread_t thread)
+{
+ processor_set_t pset = starting_pset;
+ processor_set_t nset;
+
+ assert(thread->sched_pri <= BASEPRI_RTQUEUES);
+
+ /*
+ * 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.
+ */
+ if ((thread->sched_pri < BASEPRI_RTQUEUES) || processor_is_fast_track_candidate_for_realtime_thread(pset, processor)) {
+ return processor;
+ }
+ processor = PROCESSOR_NULL;
+ 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_is_fast_track_candidate_for_realtime_thread(pset, processor)) {
+ 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_secondary_priority = MAXPRI + 1;
+ integer_t lowest_unpaired_primary_priority = MAXPRI + 1;
+ integer_t lowest_idle_secondary_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_idle_secondary_processor = PROCESSOR_NULL;
+ processor_t lp_paired_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;
+ }
+
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
+ pset_node_t node = pset->node;
+ int consider_secondaries = (!pset->is_SMT) || (bit_count(node->pset_map) == 1) || (node->pset_non_rt_primary_map == 0);
+ for (; consider_secondaries < 2; consider_secondaries++) {
+ pset = change_locked_pset(pset, starting_pset);
+ do {
+ processor = choose_processor_for_realtime_thread(pset, PROCESSOR_NULL, consider_secondaries);
+ if (processor) {
+ return processor;
+ }
+
+ /* NRG Collect processor stats for furthest deadline etc. here */
+
+ nset = next_pset(pset);
+
+ if (nset != starting_pset) {
+ pset = change_locked_pset(pset, nset);
+ }
+ } while (nset != starting_pset);
+ }
+ /* Or we could just let it change to starting_pset in the loop above */
+ pset = change_locked_pset(pset, starting_pset);
+ }
+
+ do {
+ /*
+ * Choose an idle processor, in pset traversal order
+ */
+
+ uint64_t idle_primary_map = (pset->cpu_state_map[PROCESSOR_IDLE] &
+ pset->primary_map &
+ pset->recommended_bitmask);
+
+ /* there shouldn't be a pending AST if the processor is idle */
+ assert((idle_primary_map & pset->pending_AST_URGENT_cpu_mask) == 0);
+
+ int cpuid = lsb_first(idle_primary_map);
+ if (cpuid >= 0) {
+ processor = processor_array[cpuid];
+ return processor;
+ }
+
+ /*
+ * Otherwise, enumerate active and idle processors to find primary candidates
+ * with lower priority/etc.
+ */
+
+ uint64_t active_map = ((pset->cpu_state_map[PROCESSOR_RUNNING] | pset->cpu_state_map[PROCESSOR_DISPATCHING]) &
+ pset->recommended_bitmask &
+ ~pset->pending_AST_URGENT_cpu_mask);
+
+ if (SCHED(priority_is_urgent)(thread->sched_pri) == FALSE) {
+ active_map &= ~pset->pending_AST_PREEMPT_cpu_mask;
+ }
+
+ active_map = bit_ror64(active_map, (pset->last_chosen + 1));
+ for (int rotid = lsb_first(active_map); rotid >= 0; rotid = lsb_next(active_map, rotid)) {
+ cpuid = ((rotid + pset->last_chosen + 1) & 63);
+ processor = processor_array[cpuid];
+
+ integer_t cpri = processor->current_pri;
+ processor_t primary = processor->processor_primary;
+ if (primary != processor) {
+ /* If primary is running a NO_SMT thread, don't choose its secondary */
+ if (!((primary->state == PROCESSOR_RUNNING) && processor_active_thread_no_smt(primary))) {
+ if (cpri < lowest_secondary_priority) {
+ lowest_secondary_priority = cpri;
+ lp_paired_secondary_processor = processor;
+ }
+ }
+ } else {
+ 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
+ */
+ uint64_t idle_secondary_map = (pset->cpu_state_map[PROCESSOR_IDLE] &
+ ~pset->primary_map &
+ pset->recommended_bitmask);
+
+ /* there shouldn't be a pending AST if the processor is idle */
+ assert((idle_secondary_map & pset->pending_AST_URGENT_cpu_mask) == 0);
+ assert((idle_secondary_map & pset->pending_AST_PREEMPT_cpu_mask) == 0);
+
+ for (cpuid = lsb_first(idle_secondary_map); cpuid >= 0; cpuid = lsb_next(idle_secondary_map, cpuid)) {
+ processor = processor_array[cpuid];
+
+ processor_t cprimary = processor->processor_primary;
+
+ integer_t primary_pri = cprimary->current_pri;
+
+ /*
+ * TODO: This should also make the same decisions
+ * as secondary_can_run_realtime_thread
+ *
+ * TODO: Keep track of the pending preemption priority
+ * of the primary to make this more accurate.
+ */
+
+ /* If the primary is running a no-smt thread, then don't choose its secondary */
+ if (cprimary->state == PROCESSOR_RUNNING &&
+ processor_active_thread_no_smt(cprimary)) {
+ continue;
+ }
+
+ /*
+ * Find the idle secondary processor with the lowest priority primary
+ *
+ * We will choose this processor as a fallback if we find no better
+ * primary to preempt.
+ */
+ if (primary_pri < lowest_idle_secondary_priority) {
+ lp_idle_secondary_processor = processor;
+ lowest_idle_secondary_priority = primary_pri;
+ }
+
+ /* Find the the lowest priority active primary with idle secondary */
+ if (primary_pri < lowest_unpaired_primary_priority) {
+ /* 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) {
+ continue;
+ }
+
+ if (!cprimary->is_recommended) {
+ continue;
+ }
+
+ /* if the primary is pending preemption, don't try to re-preempt it */
+ if (bit_test(pset->pending_AST_URGENT_cpu_mask, cprimary->cpu_id)) {
+ continue;
+ }
+
+ if (SCHED(priority_is_urgent)(thread->sched_pri) == FALSE &&
+ bit_test(pset->pending_AST_PREEMPT_cpu_mask, cprimary->cpu_id)) {
+ continue;
+ }
+
+ lowest_unpaired_primary_priority = primary_pri;
+ lp_unpaired_primary_processor = cprimary;
+ }
+ }
+
+ /*
+ * We prefer preempting a primary processor over waking up its secondary.
+ * The secondary will then be woken up by the preempted thread.
+ */
+ if (thread->sched_pri > lowest_unpaired_primary_priority) {
+ pset->last_chosen = lp_unpaired_primary_processor->cpu_id;
+ return lp_unpaired_primary_processor;
+ }
+
+ /*
+ * We prefer preempting a lower priority active processor over directly
+ * waking up an idle secondary.
+ * The preempted thread will then find the idle secondary.
+ */
+ if (thread->sched_pri > lowest_priority) {
+ pset->last_chosen = lp_processor->cpu_id;
+ return lp_processor;
+ }
+
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
+ /*
+ * For realtime threads, the most important aspect is
+ * scheduling latency, so we will pick an active
+ * secondary processor in this pset, or preempt
+ * another RT thread with a further deadline before
+ * going to the next pset.
+ */
+
+ if (sched_allow_rt_smt && (thread->sched_pri > lowest_secondary_priority)) {
+ pset->last_chosen = lp_paired_secondary_processor->cpu_id;
+ return lp_paired_secondary_processor;
+ }
+
+ if (thread->realtime.deadline < furthest_deadline) {
+ return fd_processor;
+ }
+ }
+
+ /*
+ * lc_processor is used to indicate the best processor set run queue
+ * on which to enqueue a thread when all available CPUs are busy with
+ * higher priority threads, so try to make sure it is initialized.
+ */
+ if (lc_processor == PROCESSOR_NULL) {
+ cpumap_t available_map = ((pset->cpu_state_map[PROCESSOR_IDLE] |
+ pset->cpu_state_map[PROCESSOR_RUNNING] |
+ pset->cpu_state_map[PROCESSOR_DISPATCHING]) &
+ pset->recommended_bitmask);
+ cpuid = lsb_first(available_map);
+ if (cpuid >= 0) {
+ lc_processor = processor_array[cpuid];
+ lowest_count = SCHED(processor_runq_count)(lc_processor);
+ }
+ }
+
+ /*
+ * Move onto the next processor set.
+ *
+ * If all primary processors in this pset are running a higher
+ * priority thread, move on to next pset. Only when we have
+ * exhausted the search for primary processors do we
+ * fall back to secondaries.
+ */
+ nset = next_pset(pset);
+
+ if (nset != starting_pset) {
+ pset = change_locked_pset(pset, nset);
+ }
+ } while (nset != starting_pset);
+
+ /*
+ * Make sure that we pick a running processor,
+ * and that the correct processor set is locked.
+ * Since we may have unlocked the candidate processor's
+ * pset, it may have changed state.
+ *
+ * All primary processors are running a higher priority
+ * thread, so the only options left are enqueuing on
+ * the secondary processor that would perturb the least priority
+ * primary, or the least busy primary.
+ */
+ boolean_t fallback_processor = false;
+ do {
+ /* lowest_priority is evaluated in the main loops above */
+ if (lp_idle_secondary_processor != PROCESSOR_NULL) {
+ processor = lp_idle_secondary_processor;
+ lp_idle_secondary_processor = PROCESSOR_NULL;
+ } else if (lp_paired_secondary_processor != PROCESSOR_NULL) {
+ processor = lp_paired_secondary_processor;
+ lp_paired_secondary_processor = PROCESSOR_NULL;
+ } else if (lc_processor != PROCESSOR_NULL) {
+ processor = lc_processor;
+ lc_processor = PROCESSOR_NULL;
+ } else {
+ /*
+ * All processors are executing higher priority threads, and
+ * the lowest_count candidate was not usable.
+ *
+ * For AMP platforms running the clutch scheduler always
+ * return a processor from the requested pset to allow the
+ * thread to be enqueued in the correct runq. For non-AMP
+ * platforms, simply return the master_processor.
+ */
+ fallback_processor = true;
+#if CONFIG_SCHED_EDGE
+ processor = processor_array[lsb_first(starting_pset->primary_map)];
+#else /* CONFIG_SCHED_EDGE */
+ processor = master_processor;
+#endif /* CONFIG_SCHED_EDGE */
+ }
+
+ /*
+ * Check that the correct processor set is
+ * returned locked.
+ */
+ pset = change_locked_pset(pset, processor->processor_set);
+
+ /*
+ * We must verify that the chosen processor is still available.
+ * The cases where we pick the master_processor or the fallback
+ * processor are execptions, since we may need enqueue a thread
+ * on its runqueue if this is the last remaining processor
+ * during pset shutdown.
+ *
+ * <rdar://problem/47559304> would really help here since it
+ * gets rid of the weird last processor SHUTDOWN case where
+ * the pset is still schedulable.
+ */
+ if (processor != master_processor && (fallback_processor == false) && (processor->state == PROCESSOR_SHUTDOWN || processor->state == PROCESSOR_OFF_LINE)) {
+ processor = PROCESSOR_NULL;
+ }
+ } while (processor == PROCESSOR_NULL);
+
+ pset->last_chosen = processor->cpu_id;
+ return processor;
+}
+
+/*
+ * Default implementation of SCHED(choose_node)()
+ * for single node systems
+ */
+pset_node_t
+sched_choose_node(__unused thread_t thread)
+{
+ return &pset_node0;
+}
+
+/*
+ * choose_starting_pset:
+ *
+ * Choose a starting processor set for the thread.
+ * May return a processor hint within the pset.
+ *
+ * Returns a starting processor set, to be used by
+ * choose_processor.
+ *
+ * The thread must be locked. The resulting pset is unlocked on return,
+ * and is chosen without taking any pset locks.
+ */
+processor_set_t
+choose_starting_pset(pset_node_t node, thread_t thread, processor_t *processor_hint)
+{
+ processor_set_t pset;
+ processor_t processor = PROCESSOR_NULL;
+
+ if (thread->affinity_set != AFFINITY_SET_NULL) {
+ /*
+ * Use affinity set policy hint.
+ */
+ pset = thread->affinity_set->aset_pset;
+ } else if (thread->last_processor != PROCESSOR_NULL) {
+ /*
+ * Simple (last processor) affinity case.
+ */
+ processor = thread->last_processor;
+ pset = processor->processor_set;
+ } else {
+ /*
+ * No Affinity case:
+ *
+ * Utilitize a per task hint to spread threads
+ * among the available processor sets.
+ * NRG this seems like the wrong thing to do.
+ * See also task->pset_hint = pset in thread_setrun()
+ */
+ task_t task = thread->task;
+
+ pset = task->pset_hint;
+ if (pset == PROCESSOR_SET_NULL) {
+ pset = current_processor()->processor_set;
+ }
+
+ pset = choose_next_pset(pset);
+ }
+
+ if (!bit_test(node->pset_map, pset->pset_id)) {
+ /* pset is not from this node so choose one that is */
+ int id = lsb_first(node->pset_map);
+ assert(id >= 0);
+ pset = pset_array[id];
+ }
+
+ if (bit_count(node->pset_map) == 1) {
+ /* Only a single pset in this node */
+ goto out;
+ }
+
+ bool avoid_cpu0 = false;
+
+#if defined(__x86_64__)
+ if ((thread->sched_pri >= BASEPRI_RTQUEUES) && sched_avoid_cpu0) {
+ /* Avoid the pset containing cpu0 */
+ avoid_cpu0 = true;
+ /* Assert that cpu0 is in pset0. I expect this to be true on __x86_64__ */
+ assert(bit_test(pset_array[0]->cpu_bitmask, 0));
+ }
+#endif
+
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
+ pset_map_t rt_target_map = atomic_load(&node->pset_non_rt_primary_map);
+ if ((avoid_cpu0 && pset->pset_id == 0) || !bit_test(rt_target_map, pset->pset_id)) {
+ if (avoid_cpu0) {
+ rt_target_map = bit_ror64(rt_target_map, 1);
+ }
+ int rotid = lsb_first(rt_target_map);
+ if (rotid >= 0) {
+ int id = avoid_cpu0 ? ((rotid + 1) & 63) : rotid;
+ pset = pset_array[id];
+ goto out;
+ }
+ }
+ if (!pset->is_SMT || !sched_allow_rt_smt) {
+ /* All psets are full of RT threads - fall back to choose processor to find the furthest deadline RT thread */
+ goto out;
+ }
+ rt_target_map = atomic_load(&node->pset_non_rt_map);
+ if ((avoid_cpu0 && pset->pset_id == 0) || !bit_test(rt_target_map, pset->pset_id)) {
+ if (avoid_cpu0) {
+ rt_target_map = bit_ror64(rt_target_map, 1);
+ }
+ int rotid = lsb_first(rt_target_map);
+ if (rotid >= 0) {
+ int id = avoid_cpu0 ? ((rotid + 1) & 63) : rotid;
+ pset = pset_array[id];
+ goto out;
+ }
+ }
+ /* All psets are full of RT threads - fall back to choose processor to find the furthest deadline RT thread */
+ } else {
+ pset_map_t idle_map = atomic_load(&node->pset_idle_map);
+ if (!bit_test(idle_map, pset->pset_id)) {
+ int next_idle_pset_id = lsb_first(idle_map);
+ if (next_idle_pset_id >= 0) {
+ pset = pset_array[next_idle_pset_id];
+ }
+ }
+ }
+
+out:
+ if ((processor != PROCESSOR_NULL) && (processor->processor_set != pset)) {
+ processor = PROCESSOR_NULL;
+ }
+ if (processor != PROCESSOR_NULL) {
+ *processor_hint = processor;
+ }
+
+ return pset;
+}
+
+/*
+ * thread_setrun:
+ *
+ * Dispatch thread for execution, onto an idle
+ * processor or run queue, and signal a preemption
+ * as appropriate.
+ *
+ * Thread must be locked.
+ */
+void
+thread_setrun(
+ thread_t thread,
+ sched_options_t options)
+{
+ processor_t processor;
+ processor_set_t pset;
+
+ assert((thread->state & (TH_RUN | TH_WAIT | TH_UNINT | TH_TERMINATE | TH_TERMINATE2)) == TH_RUN);
+ assert(thread->runq == PROCESSOR_NULL);
+
+ /*
+ * Update priority if needed.
+ */
+ if (SCHED(can_update_priority)(thread)) {
+ SCHED(update_priority)(thread);
+ }
+
+ thread->sfi_class = sfi_thread_classify(thread);
+
+ assert(thread->runq == PROCESSOR_NULL);
+
+ if (thread->bound_processor == PROCESSOR_NULL) {
+ /*
+ * Unbound case.
+ */
+ processor_t processor_hint = PROCESSOR_NULL;
+ pset_node_t node = SCHED(choose_node)(thread);
+ processor_set_t starting_pset = choose_starting_pset(node, thread, &processor_hint);
+
+ pset_lock(starting_pset);
+
+ processor = SCHED(choose_processor)(starting_pset, processor_hint, thread);
+ pset = processor->processor_set;
+ task_t task = thread->task;
+ task->pset_hint = pset; /* NRG this is done without holding the task lock */
+
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), (uintptr_t)-1, processor->cpu_id, processor->state, 0);
+ } else {
+ /*
+ * Bound case:
+ *
+ * Unconditionally dispatch on the processor.
+ */
+ processor = thread->bound_processor;
+ pset = processor->processor_set;
+ pset_lock(pset);
+
+ SCHED_DEBUG_CHOOSE_PROCESSOR_KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHOOSE_PROCESSOR) | DBG_FUNC_NONE,
+ (uintptr_t)thread_tid(thread), (uintptr_t)-2, processor->cpu_id, processor->state, 0);
+ }
+
+ /*
+ * Dispatch the thread on the chosen processor.
+ * TODO: This should be based on sched_mode, not sched_pri
+ */
+ if (thread->sched_pri >= BASEPRI_RTQUEUES) {
+ realtime_setrun(processor, thread);
+ } else {
+ processor_setrun(processor, thread, options);
+ }
+ /* pset is now unlocked */
+ if (thread->bound_processor == PROCESSOR_NULL) {
+ SCHED(check_spill)(pset, thread);
+ }
+}
+
+processor_set_t
+task_choose_pset(
+ task_t task)
+{
+ processor_set_t pset = task->pset_hint;
+
+ if (pset != PROCESSOR_SET_NULL) {
+ pset = choose_next_pset(pset);
+ }
+
+ return pset;
+}
+
+/*
+ * Check for a preemption point in
+ * the current context.
+ *
+ * Called at splsched with thread locked.
+ */
+ast_t
+csw_check(
+ thread_t thread,
+ processor_t processor,
+ ast_t check_reason)
+{
+ processor_set_t pset = processor->processor_set;
+
+ assert(thread == processor->active_thread);
+
+ pset_lock(pset);
+
+ processor_state_update_from_thread(processor, thread);
+
+ ast_t preempt = csw_check_locked(thread, processor, pset, check_reason);
+
+ /* Acknowledge the IPI if we decided not to preempt */
+
+ if ((preempt & AST_URGENT) == 0) {
+ bit_clear(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id);
+ }
+
+ if ((preempt & AST_PREEMPT) == 0) {
+ bit_clear(pset->pending_AST_PREEMPT_cpu_mask, processor->cpu_id);
+ }
+
+ pset_unlock(pset);
+
+ return preempt;
+}
+
+/*
+ * Check for preemption at splsched with
+ * pset and thread locked
+ */
+ast_t
+csw_check_locked(
+ thread_t thread,
+ processor_t processor,
+ processor_set_t pset,
+ ast_t check_reason)
+{
+ ast_t result;
+
+ if (processor->first_timeslice) {
+ if (rt_runq_count(pset) > 0) {
+ return check_reason | AST_PREEMPT | AST_URGENT;
+ }
+ } else {
+ if (rt_runq_count(pset) > 0) {
+ if (BASEPRI_RTQUEUES > processor->current_pri) {
+ return check_reason | AST_PREEMPT | AST_URGENT;
+ } else {
+ return check_reason | AST_PREEMPT;
+ }
+ }
+ }
+
+ /*
+ * If the current thread is running on a processor that is no longer recommended,
+ * urgently preempt it, at which point thread_select() should
+ * try to idle the processor and re-dispatch the thread to a recommended processor.
+ */
+ if (!processor->is_recommended) {
+ return check_reason | AST_PREEMPT | AST_URGENT;
+ }
+
+ result = SCHED(processor_csw_check)(processor);
+ if (result != AST_NONE) {
+ return check_reason | result | (thread_is_eager_preempt(thread) ? AST_URGENT : AST_NONE);
+ }
+
+ /*
+ * Same for avoid-processor
+ *
+ * TODO: Should these set AST_REBALANCE?
+ */
+ if (SCHED(avoid_processor_enabled) && SCHED(thread_avoid_processor)(processor, thread)) {
+ return check_reason | AST_PREEMPT;
+ }
+
+ /*
+ * Even though we could continue executing on this processor, a
+ * secondary SMT core should try to shed load to another primary core.
+ *
+ * TODO: Should this do the same check that thread_select does? i.e.
+ * if no bound threads target this processor, and idle primaries exist, preempt
+ * The case of RT threads existing is already taken care of above
+ */
+
+ if (processor->current_pri < BASEPRI_RTQUEUES &&
+ processor->processor_primary != processor) {
+ return check_reason | AST_PREEMPT;
+ }
+
+ if (thread->state & TH_SUSP) {
+ return check_reason | AST_PREEMPT;
+ }
+
+#if CONFIG_SCHED_SFI
+ /*
+ * Current thread may not need to be preempted, but maybe needs
+ * an SFI wait?
+ */
+ result = sfi_thread_needs_ast(thread, NULL);
+ if (result != AST_NONE) {
+ return check_reason | result;
+ }
+#endif
+
+ return AST_NONE;
+}
+
+/*
+ * Handle preemption IPI or IPI in response to setting an AST flag
+ * Triggered by cause_ast_check
+ * Called at splsched
+ */
+void
+ast_check(processor_t processor)
+{
+ if (processor->state != PROCESSOR_RUNNING &&
+ processor->state != PROCESSOR_SHUTDOWN) {
+ return;
+ }
+
+ thread_t thread = processor->active_thread;
+
+ assert(thread == current_thread());
+
+ thread_lock(thread);
+
+ /*
+ * Propagate thread ast to processor.
+ * (handles IPI in response to setting AST flag)
+ */
+ ast_propagate(thread);
+
+ /*
+ * Stash the old urgency and perfctl values to find out if
+ * csw_check updates them.
+ */
+ thread_urgency_t old_urgency = processor->current_urgency;
+ perfcontrol_class_t old_perfctl_class = processor->current_perfctl_class;
+
+ ast_t preempt;
+
+ if ((preempt = csw_check(thread, processor, AST_NONE)) != AST_NONE) {
+ ast_on(preempt);
+ }
+
+ if (old_urgency != processor->current_urgency) {
+ /*
+ * Urgency updates happen with the thread lock held (ugh).
+ * TODO: This doesn't notice QoS changes...
+ */
+ uint64_t urgency_param1, urgency_param2;
+
+ thread_urgency_t urgency = thread_get_urgency(thread, &urgency_param1, &urgency_param2);
+ thread_tell_urgency(urgency, urgency_param1, urgency_param2, 0, thread);
+ }
+
+ thread_unlock(thread);
+
+ if (old_perfctl_class != processor->current_perfctl_class) {
+ /*
+ * We updated the perfctl class of this thread from another core.
+ * Let CLPC know that the currently running thread has a new
+ * class.
+ */
+
+ machine_switch_perfcontrol_state_update(PERFCONTROL_ATTR_UPDATE,
+ mach_approximate_time(), 0, thread);
+ }
+}
+
+
+/*
+ * set_sched_pri:
+ *
+ * Set the scheduled priority of the specified thread.
+ *
+ * This may cause the thread to change queues.
+ *
+ * Thread must be locked.
+ */
+void
+set_sched_pri(
+ thread_t thread,
+ int16_t new_priority,
+ set_sched_pri_options_t options)
+{
+ bool is_current_thread = (thread == current_thread());
+ bool removed_from_runq = false;
+ bool lazy_update = ((options & SETPRI_LAZY) == SETPRI_LAZY);
+
+ int16_t old_priority = thread->sched_pri;
+
+ /* If we're already at this priority, no need to mess with the runqueue */
+ if (new_priority == old_priority) {
+#if CONFIG_SCHED_CLUTCH
+ /* For the first thread in the system, the priority is correct but
+ * th_sched_bucket is still TH_BUCKET_RUN. Since the clutch
+ * scheduler relies on the bucket being set for all threads, update
+ * its bucket here.
+ */
+ if (thread->th_sched_bucket == TH_BUCKET_RUN) {
+ assert(is_current_thread);
+ SCHED(update_thread_bucket)(thread);
+ }
+#endif /* CONFIG_SCHED_CLUTCH */
+
+ return;
+ }
+
+ if (is_current_thread) {
+ assert(thread->state & TH_RUN);
+ assert(thread->runq == PROCESSOR_NULL);
+ } else {
+ removed_from_runq = thread_run_queue_remove(thread);
+ }
+
+ thread->sched_pri = new_priority;
+
+#if CONFIG_SCHED_CLUTCH
+ /*
+ * Since for the clutch scheduler, the thread's bucket determines its runq
+ * in the hierarchy it is important to update the bucket when the thread
+ * lock is held and the thread has been removed from the runq hierarchy.
+ */
+ SCHED(update_thread_bucket)(thread);
+
+#endif /* CONFIG_SCHED_CLUTCH */
+
+ KERNEL_DEBUG_CONSTANT(MACHDBG_CODE(DBG_MACH_SCHED, MACH_SCHED_CHANGE_PRIORITY),
+ (uintptr_t)thread_tid(thread),
+ thread->base_pri,
+ thread->sched_pri,
+ thread->sched_usage,
+ 0);
+
+ if (removed_from_runq) {
+ thread_run_queue_reinsert(thread, SCHED_PREEMPT | SCHED_TAILQ);
+ } else if (is_current_thread) {
+ processor_t processor = thread->last_processor;
+ assert(processor == current_processor());
+
+ thread_urgency_t old_urgency = processor->current_urgency;
+
+ /*
+ * When dropping in priority, check if the thread no longer belongs on core.
+ * If a thread raises its own priority, don't aggressively rebalance it.
+ * <rdar://problem/31699165>
+ *
+ * csw_check does a processor_state_update_from_thread, but
+ * we should do our own if we're being lazy.
+ */
+ if (!lazy_update && new_priority < old_priority) {
+ ast_t preempt;
+
+ if ((preempt = csw_check(thread, processor, AST_NONE)) != AST_NONE) {
+ ast_on(preempt);
+ }
+ } else {
+ processor_state_update_from_thread(processor, thread);
+ }
+
+ /*
+ * set_sched_pri doesn't alter RT params. We expect direct base priority/QoS
+ * class alterations from user space to occur relatively infrequently, hence
+ * those are lazily handled. QoS classes have distinct priority bands, and QoS
+ * inheritance is expected to involve priority changes.
+ */
+ if (processor->current_urgency != old_urgency) {
+ uint64_t urgency_param1, urgency_param2;
+
+ thread_urgency_t new_urgency = thread_get_urgency(thread,
+ &urgency_param1, &urgency_param2);
+
+ thread_tell_urgency(new_urgency, urgency_param1,
+ urgency_param2, 0, thread);
+ }
+
+ /* TODO: only call this if current_perfctl_class changed */
+ uint64_t ctime = mach_approximate_time();
+ machine_thread_going_on_core(thread, processor->current_urgency, 0, 0, ctime);
+ } else if (thread->state & TH_RUN) {
+ processor_t processor = thread->last_processor;
+
+ if (!lazy_update &&
+ processor != PROCESSOR_NULL &&
+ processor != current_processor() &&
+ processor->active_thread == thread) {
+ cause_ast_check(processor);
+ }
+ }
+}
+
+/*
+ * thread_run_queue_remove_for_handoff
+ *
+ * Pull a thread or its (recursive) push target out of the runqueue
+ * so that it is ready for thread_run()
+ *
+ * Called at splsched
+ *
+ * Returns the thread that was pulled or THREAD_NULL if no thread could be pulled.
+ * This may be different than the thread that was passed in.
+ */
+thread_t
+thread_run_queue_remove_for_handoff(thread_t thread)
+{
+ thread_t pulled_thread = THREAD_NULL;
+
+ thread_lock(thread);
+
+ /*
+ * Check that the thread is not bound to a different processor,
+ * NO_SMT flag is not set on the thread, cluster type of
+ * processor matches with thread if the thread is pinned to a
+ * particular cluster and that realtime is not involved.
+ *
+ * Next, pull it off its run queue. If it doesn't come, it's not eligible.
+ */
+ processor_t processor = current_processor();
+ if ((thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == processor)
+ && (!thread_no_smt(thread))
+ && (processor->current_pri < BASEPRI_RTQUEUES)
+ && (thread->sched_pri < BASEPRI_RTQUEUES)
+#if __AMP__
+ && ((!(thread->sched_flags & TH_SFLAG_PCORE_ONLY)) ||
+ processor->processor_set->pset_cluster_type == PSET_AMP_P)
+ && ((!(thread->sched_flags & TH_SFLAG_ECORE_ONLY)) ||
+ processor->processor_set->pset_cluster_type == PSET_AMP_E)
+#endif /* __AMP__ */
+ ) {
+ if (thread_run_queue_remove(thread)) {
+ pulled_thread = thread;
+ }
+ }
+
+ thread_unlock(thread);
+
+ return pulled_thread;
+}
+
+/*
+ * thread_prepare_for_handoff
+ *
+ * Make the thread ready for handoff.
+ * If the thread was runnable then pull it off the runq, if the thread could
+ * not be pulled, return NULL.
+ *
+ * If the thread was woken up from wait for handoff, make sure it is not bound to
+ * different processor.
+ *
+ * Called at splsched
+ *
+ * Returns the thread that was pulled or THREAD_NULL if no thread could be pulled.
+ * This may be different than the thread that was passed in.
+ */
+thread_t
+thread_prepare_for_handoff(thread_t thread, thread_handoff_option_t option)
+{
+ thread_t pulled_thread = THREAD_NULL;
+
+ if (option & THREAD_HANDOFF_SETRUN_NEEDED) {
+ processor_t processor = current_processor();
+ thread_lock(thread);
+
+ /*
+ * Check that the thread is not bound to a different processor,
+ * NO_SMT flag is not set on the thread and cluster type of
+ * processor matches with thread if the thread is pinned to a
+ * particular cluster. Call setrun instead if above conditions
+ * are not satisfied.
+ */
+ if ((thread->bound_processor == PROCESSOR_NULL || thread->bound_processor == processor)
+ && (!thread_no_smt(thread))
+#if __AMP__
+ && ((!(thread->sched_flags & TH_SFLAG_PCORE_ONLY)) ||
+ processor->processor_set->pset_cluster_type == PSET_AMP_P)
+ && ((!(thread->sched_flags & TH_SFLAG_ECORE_ONLY)) ||
+ processor->processor_set->pset_cluster_type == PSET_AMP_E)
+#endif /* __AMP__ */
+ ) {
+ pulled_thread = thread;
+ } else {
+ thread_setrun(thread, SCHED_PREEMPT | SCHED_TAILQ);
+ }
+ thread_unlock(thread);
+ } else {
+ pulled_thread = thread_run_queue_remove_for_handoff(thread);
+ }
+
+ return pulled_thread;
+}
+
+/*
+ * thread_run_queue_remove:
+ *
+ * Remove a thread from its current run queue and
+ * return TRUE if successful.
+ *
+ * Thread must be locked.
+ *
+ * If thread->runq is PROCESSOR_NULL, the thread will not re-enter the
+ * run queues because the caller locked the thread. Otherwise
+ * the thread is on a run queue, but could be chosen for dispatch
+ * and removed by another processor under a different lock, which
+ * will set thread->runq to PROCESSOR_NULL.
+ *
+ * Hence the thread select path must not rely on anything that could
+ * be changed under the thread lock after calling this function,
+ * most importantly thread->sched_pri.
+ */
+boolean_t
+thread_run_queue_remove(
+ thread_t thread)
+{
+ boolean_t removed = FALSE;
+ processor_t processor = thread->runq;
+
+ if ((thread->state & (TH_RUN | TH_WAIT)) == TH_WAIT) {
+ /* Thread isn't runnable */
+ assert(thread->runq == PROCESSOR_NULL);
+ return FALSE;
+ }
+
+ if (processor == PROCESSOR_NULL) {
+ /*
+ * The thread is either not on the runq,
+ * or is in the midst of being removed from the runq.
+ *
+ * runq is set to NULL under the pset lock, not the thread
+ * lock, so the thread may still be in the process of being dequeued
+ * from the runq. It will wait in invoke for the thread lock to be
+ * dropped.
+ */
+
+ return FALSE;
+ }
+
+ if (thread->sched_pri < BASEPRI_RTQUEUES) {
+ return SCHED(processor_queue_remove)(processor, thread);
+ }
+
+ processor_set_t pset = processor->processor_set;
+
+ pset_lock(pset);
+
+ if (thread->runq != PROCESSOR_NULL) {
+ /*
+ * Thread is on the RT run queue and we have a lock on
+ * that run queue.
+ */
+
+ remqueue(&thread->runq_links);
+ SCHED_STATS_RUNQ_CHANGE(&SCHED(rt_runq)(pset)->runq_stats, rt_runq_count(pset));
+ rt_runq_count_decr(pset);
+
+ thread->runq = PROCESSOR_NULL;
+
+ removed = TRUE;
+ }
+
+ pset_unlock(pset);
+
+ return removed;
+}
+
+/*
+ * Put the thread back where it goes after a thread_run_queue_remove
+ *
+ * Thread must have been removed under the same thread lock hold
+ *
+ * thread locked, at splsched
+ */
+void
+thread_run_queue_reinsert(thread_t thread, sched_options_t options)
+{
+ assert(thread->runq == PROCESSOR_NULL);
+ assert(thread->state & (TH_RUN));
+
+ thread_setrun(thread, options);
+}
+
+void
+sys_override_cpu_throttle(boolean_t enable_override)
+{
+ if (enable_override) {
+ cpu_throttle_enabled = 0;
+ } else {
+ cpu_throttle_enabled = 1;
+ }
+}
+
+thread_urgency_t
+thread_get_urgency(thread_t thread, uint64_t *arg1, uint64_t *arg2)
+{
+ uint64_t urgency_param1 = 0, urgency_param2 = 0;
+
+ thread_urgency_t urgency;
+
+ if (thread == NULL || (thread->state & TH_IDLE)) {
+ urgency_param1 = 0;
+ urgency_param2 = 0;
+
+ urgency = THREAD_URGENCY_NONE;
+ } else if (thread->sched_mode == TH_MODE_REALTIME) {
+ urgency_param1 = thread->realtime.period;
+ urgency_param2 = thread->realtime.deadline;
+
+ urgency = THREAD_URGENCY_REAL_TIME;
+ } else if (cpu_throttle_enabled &&
+ (thread->sched_pri <= MAXPRI_THROTTLE) &&
+ (thread->base_pri <= MAXPRI_THROTTLE)) {
+ /*
+ * Threads that are running at low priority but are not
+ * tagged with a specific QoS are separated out from
+ * the "background" urgency. Performance management
+ * subsystem can decide to either treat these threads
+ * as normal threads or look at other signals like thermal
+ * levels for optimal power/perf tradeoffs for a platform.
+ */
+ boolean_t thread_lacks_qos = (proc_get_effective_thread_policy(thread, TASK_POLICY_QOS) == THREAD_QOS_UNSPECIFIED); //thread_has_qos_policy(thread);
+ boolean_t task_is_suppressed = (proc_get_effective_task_policy(thread->task, TASK_POLICY_SUP_ACTIVE) == 0x1);
+
+ /*
+ * Background urgency applied when thread priority is
+ * MAXPRI_THROTTLE or lower and thread is not promoted
+ * and thread has a QoS specified
+ */
+ urgency_param1 = thread->sched_pri;
+ urgency_param2 = thread->base_pri;
+
+ if (thread_lacks_qos && !task_is_suppressed) {
+ urgency = THREAD_URGENCY_LOWPRI;
+ } else {
+ urgency = THREAD_URGENCY_BACKGROUND;
+ }
+ } else {
+ /* For otherwise unclassified threads, report throughput QoS parameters */
+ urgency_param1 = proc_get_effective_thread_policy(thread, TASK_POLICY_THROUGH_QOS);
+ urgency_param2 = proc_get_effective_task_policy(thread->task, TASK_POLICY_THROUGH_QOS);
+ urgency = THREAD_URGENCY_NORMAL;
+ }
+
+ if (arg1 != NULL) {
+ *arg1 = urgency_param1;
+ }
+ if (arg2 != NULL) {
+ *arg2 = urgency_param2;
+ }
+
+ return urgency;
+}
+
+perfcontrol_class_t
+thread_get_perfcontrol_class(thread_t thread)
+{
+ /* Special case handling */
+ if (thread->state & TH_IDLE) {
+ return PERFCONTROL_CLASS_IDLE;
+ }
+ if (thread->task == kernel_task) {
+ return PERFCONTROL_CLASS_KERNEL;
+ }
+ if (thread->sched_mode == TH_MODE_REALTIME) {
+ return PERFCONTROL_CLASS_REALTIME;
+ }
+
+ /* perfcontrol_class based on base_pri */
+ if (thread->base_pri <= MAXPRI_THROTTLE) {
+ return PERFCONTROL_CLASS_BACKGROUND;
+ } else if (thread->base_pri <= BASEPRI_UTILITY) {
+ return PERFCONTROL_CLASS_UTILITY;
+ } else if (thread->base_pri <= BASEPRI_DEFAULT) {
+ return PERFCONTROL_CLASS_NONUI;
+ } else if (thread->base_pri <= BASEPRI_FOREGROUND) {
+ return PERFCONTROL_CLASS_UI;
+ } else {
+ return PERFCONTROL_CLASS_ABOVEUI;
+ }
+}
+
+/*
+ * This is the processor idle loop, which just looks for other threads
+ * to execute. Processor idle threads invoke this without supplying a
+ * current thread to idle without an asserted wait state.
+ *
+ * Returns a the next thread to execute if dispatched directly.
+ */
+
+#if 0
+#define IDLE_KERNEL_DEBUG_CONSTANT(...) KERNEL_DEBUG_CONSTANT(__VA_ARGS__)
+#else
+#define IDLE_KERNEL_DEBUG_CONSTANT(...) do { } while(0)
+#endif
+
+thread_t
+processor_idle(
+ thread_t thread,
+ processor_t processor)
+{
+ processor_set_t pset = processor->processor_set;
+
+ (void)splsched();
+
+ KERNEL_DEBUG_CONSTANT_IST(KDEBUG_TRACE,
+ MACHDBG_CODE(DBG_MACH_SCHED, MACH_IDLE) | DBG_FUNC_START,
+ (uintptr_t)thread_tid(thread), 0, 0, 0, 0);
+
+ SCHED_STATS_INC(idle_transitions);
+ assert(processor->running_timers_active == false);
+
+ uint64_t ctime = mach_absolute_time();
+
+ timer_switch(&processor->system_state, ctime, &processor->idle_state);
+ processor->current_state = &processor->idle_state;
+
+ cpu_quiescent_counter_leave(ctime);
+
+ while (1) {
+ /*
+ * Ensure that updates to my processor and pset state,
+ * made by the IPI source processor before sending the IPI,
+ * are visible on this processor now (even though we don't
+ * take the pset lock yet).
+ */
+ atomic_thread_fence(memory_order_acquire);
+
+ if (processor->state != PROCESSOR_IDLE) {
+ break;
+ }
+ if (bit_test(pset->pending_AST_URGENT_cpu_mask, processor->cpu_id)) {
+ break;
+ }
+#if defined(CONFIG_SCHED_DEFERRED_AST)
+ if (bit_test(pset->pending_deferred_AST_cpu_mask, processor->cpu_id)) {
+ break;
projects / apple / xnu.git / blobdiff