// Copyright 2017 The Fuchsia Authors
// Copyright (c) 2008-2015 Travis Geiselbrecht
//
// Use of this source code is governed by a MIT-style
// license that can be found in the LICENSE file or at
// https://opensource.org/licenses/MIT
#include <kernel/sched.h>

#include <assert.h>
#include <debug.h>
#include <err.h>
#include <inttypes.h>
#include <kernel/mp.h>
#include <kernel/percpu.h>
#include <kernel/thread.h>
#include <lib/ktrace.h>
#include <list.h>
#include <platform.h>
#include <printf.h>
#include <string.h>
#include <target.h>
#include <trace.h>
#include <vm/vm.h>
#include <zircon/time.h>
#include <zircon/types.h>

// disable priority boosting
#define NO_BOOST 0

#define MAX_PRIORITY_ADJ 4 // +/- priority levels from the base priority

// ktraces just local to this file
#define LOCAL_KTRACE 0

#if LOCAL_KTRACE
#define LOCAL_KTRACE0(probe) ktrace_probe0(probe)
#define LOCAL_KTRACE2(probe, x, y) ktrace_probe2(probe, x, y)
#else
#define LOCAL_KTRACE0(probe)
#define LOCAL_KTRACE2(probe, x, y)
#endif

#define LOCAL_TRACE 0

#define DEBUG_THREAD_CONTEXT_SWITCH 0

#define TRACE_CONTEXT_SWITCH(str, x...)  \
    do {                                 \
        if (DEBUG_THREAD_CONTEXT_SWITCH) \
            printf("CS " str, ##x);      \
    } while (0)

// threads get 10ms to run before they use up their time slice and the scheduler is invoked
#define THREAD_INITIAL_TIME_SLICE ZX_MSEC(10)

static bool local_migrate_if_needed(thread_t* curr_thread);

// compute the effective priority of a thread
static void compute_effec_priority(thread_t* t) {
    int ep = t->base_priority + t->priority_boost;
    if (t->inherited_priority > ep) {
        ep = t->inherited_priority;
    }

    DEBUG_ASSERT(ep >= LOWEST_PRIORITY && ep <= HIGHEST_PRIORITY);

    t->effec_priority = ep;
}

// boost the priority of the thread by +1
static void boost_thread(thread_t* t) {
    if (NO_BOOST) {
        return;
    }

    if (unlikely(thread_is_real_time_or_idle(t))) {
        return;
    }

    if (t->priority_boost < MAX_PRIORITY_ADJ &&
        likely((t->base_priority + t->priority_boost) < HIGHEST_PRIORITY)) {
        t->priority_boost++;
        compute_effec_priority(t);
    }
}

// deboost the priority of the thread by -1.
// If deboosting because the thread is using up all of its time slice,
// then allow the boost to go negative, otherwise only deboost to 0.
static void deboost_thread(thread_t* t, bool quantum_expiration) {
    if (NO_BOOST) {
        return;
    }

    if (unlikely(thread_is_real_time_or_idle(t))) {
        return;
    }

    int boost_floor;
    if (quantum_expiration) {
        // deboost into negative boost
        boost_floor = -MAX_PRIORITY_ADJ;

        // make sure we dont deboost a thread too far
        if (unlikely(t->base_priority + boost_floor < LOWEST_PRIORITY)) {
            boost_floor = t->base_priority - LOWEST_PRIORITY;
        }

    } else {
        // otherwise only deboost to 0
        boost_floor = 0;
    }

    // if we're already bottomed out or below bottomed out, leave it alone
    if (t->priority_boost <= boost_floor) {
        return;
    }

    // drop a level
    t->priority_boost--;
    compute_effec_priority(t);
}

// pick a 'random' cpu out of the passed in mask of cpus
static cpu_mask_t rand_cpu(cpu_mask_t mask) {
    if (unlikely(mask == 0)) {
        return 0;
    }

    // check that the mask passed in has at least one bit set in the active mask
    cpu_mask_t active = mp_get_active_mask();
    mask &= active;
    if (unlikely(mask == 0)) {
        return 0;
    }

    // compute the highest cpu in the mask
    cpu_num_t highest_cpu = highest_cpu_set(mask);

    // not very random, round robins a bit through the mask until it gets a hit
    for (;;) {
        // protected by THREAD_LOCK, safe to use non atomically
        static uint rot = 0;

        if (++rot > highest_cpu) {
            rot = 0;
        }

        if ((1u << rot) & mask) {
            return (1u << rot);
        }
    }
}

// find a cpu to wake up
static cpu_mask_t find_cpu_mask(thread_t* t) TA_REQ(thread_lock) {
    // get the last cpu the thread ran on
    cpu_mask_t last_ran_cpu_mask = cpu_num_to_mask(t->last_cpu);

    // the current cpu
    cpu_mask_t curr_cpu_mask = cpu_num_to_mask(arch_curr_cpu_num());

    // the thread's affinity mask
    cpu_mask_t cpu_affinity = t->cpu_affinity;

    LTRACEF_LEVEL(2, "last %#x curr %#x aff %#x name %s\n",
                  last_ran_cpu_mask, curr_cpu_mask, cpu_affinity, t->name);

    // get a list of idle cpus and mask off the ones that aren't in our affinity mask
    cpu_mask_t idle_cpu_mask = mp_get_idle_mask();
    cpu_mask_t active_cpu_mask = mp_get_active_mask();
    idle_cpu_mask &= cpu_affinity;
    if (idle_cpu_mask != 0) {
        if (idle_cpu_mask & curr_cpu_mask) {
            // the current cpu is idle and within our affinity mask, so run it here
            return curr_cpu_mask;
        }

        if (last_ran_cpu_mask & idle_cpu_mask) {
            DEBUG_ASSERT(last_ran_cpu_mask & mp_get_active_mask());
            // the last core it ran on is idle and isn't the current cpu
            return last_ran_cpu_mask;
        }

        // pick an idle_cpu
        DEBUG_ASSERT((idle_cpu_mask & mp_get_active_mask()) == idle_cpu_mask);
        return rand_cpu(idle_cpu_mask);
    }

    // no idle cpus in our affinity mask

    // if the last cpu it ran on is in the affinity mask and not the current cpu, pick that
    if ((last_ran_cpu_mask & cpu_affinity & active_cpu_mask) &&
        last_ran_cpu_mask != curr_cpu_mask) {
        return last_ran_cpu_mask;
    }

    // fall back to picking a cpu out of the affinity mask, preferring something other
    // than the local cpu.
    // the affinity mask hard pins the thread to the cpus in the mask, so it's not possible
    // to pick a cpu outside of that list.
    cpu_mask_t mask = cpu_affinity & ~(curr_cpu_mask);
    if (mask == 0) {
        return curr_cpu_mask; // local cpu is the only choice
    }

    mask = rand_cpu(mask);
    if (mask == 0) {
        return curr_cpu_mask; // local cpu is the only choice
    }
    DEBUG_ASSERT((mask & mp_get_active_mask()) == mask);
    return mask;
}

// run queue manipulation
static void insert_in_run_queue_head(cpu_num_t cpu, thread_t* t) TA_REQ(thread_lock) {
    DEBUG_ASSERT(!list_in_list(&t->queue_node));

    list_add_head(&percpu[cpu].run_queue[t->effec_priority], &t->queue_node);
    percpu[cpu].run_queue_bitmap |= (1u << t->effec_priority);

    // mark the cpu as busy since the run queue now has at least one item in it
    mp_set_cpu_busy(cpu);
}

static void insert_in_run_queue_tail(cpu_num_t cpu, thread_t* t) TA_REQ(thread_lock) {
    DEBUG_ASSERT(!list_in_list(&t->queue_node));

    list_add_tail(&percpu[cpu].run_queue[t->effec_priority], &t->queue_node);
    percpu[cpu].run_queue_bitmap |= (1u << t->effec_priority);

    // mark the cpu as busy since the run queue now has at least one item in it
    mp_set_cpu_busy(cpu);
}

// remove the thread from the run queue it's in
static void remove_from_run_queue(thread_t* t, int prio_queue) TA_REQ(thread_lock) {
    DEBUG_ASSERT(t->state == THREAD_READY);
    DEBUG_ASSERT(is_valid_cpu_num(t->curr_cpu));

    list_delete(&t->queue_node);

    // clear the old cpu's queue bitmap if that was the last entry
    struct percpu* c = &percpu[t->curr_cpu];
    if (list_is_empty(&c->run_queue[prio_queue])) {
        c->run_queue_bitmap &= ~(1u << prio_queue);
    }
}

// using the per cpu run queue bitmap, find the highest populated queue
static uint highest_run_queue(const struct percpu* c) TA_REQ(thread_lock) {
    return HIGHEST_PRIORITY - __builtin_clz(c->run_queue_bitmap) -
           (sizeof(c->run_queue_bitmap) * CHAR_BIT - NUM_PRIORITIES);
}

static thread_t* sched_get_top_thread(cpu_num_t cpu) TA_REQ(thread_lock) {
    // pop the head of the highest priority queue with any threads
    // queued up on the passed in cpu.

    struct percpu* c = &percpu[cpu];
    if (likely(c->run_queue_bitmap)) {
        uint highest_queue = highest_run_queue(c);

        thread_t* newthread = list_remove_head_type(&c->run_queue[highest_queue], thread_t, queue_node);

        DEBUG_ASSERT(newthread);
        DEBUG_ASSERT_MSG(newthread->cpu_affinity & cpu_num_to_mask(cpu),
                         "thread %p name %s, aff %#x cpu %u\n", newthread, newthread->name,
                         newthread->cpu_affinity, cpu);
        DEBUG_ASSERT(newthread->curr_cpu == cpu);

        if (list_is_empty(&c->run_queue[highest_queue])) {
            c->run_queue_bitmap &= ~(1u << highest_queue);
        }

        LOCAL_KTRACE2("sched_get_top", newthread->priority_boost, newthread->base_priority);

        return newthread;
    }

    // no threads to run, select the idle thread for this cpu
    return &c->idle_thread;
}

void sched_init_thread(thread_t* t, int priority) {
    t->base_priority = priority;
    t->priority_boost = 0;
    t->inherited_priority = -1;
    compute_effec_priority(t);
}

void sched_block() {
    DEBUG_ASSERT(spin_lock_held(&thread_lock));

    __UNUSED thread_t* current_thread = get_current_thread();

    DEBUG_ASSERT(current_thread->magic == THREAD_MAGIC);
    DEBUG_ASSERT(current_thread->state != THREAD_RUNNING);

    LOCAL_KTRACE0("sched_block");

    // we are blocking on something. the blocking code should have already stuck us on a queue
    sched_resched_internal();
}

// find a cpu to run the thread on, put it in the run queue for that cpu, and accumulate a list
// of cpus we'll need to reschedule, including the local cpu.
static void find_cpu_and_insert(thread_t* t, bool* local_resched,
                                cpu_mask_t* accum_cpu_mask) TA_REQ(thread_lock) {
    // find a core to run it on
    cpu_mask_t cpu = find_cpu_mask(t);
    cpu_num_t cpu_num;

    DEBUG_ASSERT(cpu != 0);

    cpu_num = lowest_cpu_set(cpu);
    if (cpu_num == arch_curr_cpu_num()) {
        *local_resched = true;
    } else {
        *accum_cpu_mask |= cpu_num_to_mask(cpu_num);
    }

    t->curr_cpu = cpu_num;
    if (t->remaining_time_slice > 0) {
        insert_in_run_queue_head(cpu_num, t);
    } else {
        insert_in_run_queue_tail(cpu_num, t);
    }
}

bool sched_unblock(thread_t* t) {
    DEBUG_ASSERT(spin_lock_held(&thread_lock));

    DEBUG_ASSERT(t->magic == THREAD_MAGIC);

    LOCAL_KTRACE0("sched_unblock");

    // thread is being woken up, boost its priority
    boost_thread(t);

    // stuff the new thread in the run queue
    t->state = THREAD_READY;

    bool local_resched = false;
    cpu_mask_t mask = 0;
    find_cpu_and_insert(t, &local_resched, &mask);

    if (mask) {
        mp_reschedule(mask, 0);
    }
    return local_resched;
}

bool sched_unblock_list(struct list_node* list) {
    DEBUG_ASSERT(list);
    DEBUG_ASSERT(spin_lock_held(&thread_lock));

    LOCAL_KTRACE0("sched_unblock_list");

    // pop the list of threads and shove into the scheduler
    bool local_resched = false;
    cpu_mask_t accum_cpu_mask = 0;
    thread_t* t;
    while ((t = list_remove_tail_type(list, thread_t, queue_node))) {
        DEBUG_ASSERT(t->magic == THREAD_MAGIC);
        DEBUG_ASSERT(!thread_is_idle(t));

        // thread is being woken up, boost its priority
        boost_thread(t);

        // stuff the new thread in the run queue
        t->state = THREAD_READY;
        find_cpu_and_insert(t, &local_resched, &accum_cpu_mask);
    }

    if (accum_cpu_mask) {
        mp_reschedule(accum_cpu_mask, 0);
    }

    return local_resched;
}

// handle the special case of resuming a newly created idle thread
void sched_unblock_idle(thread_t* t) {
    DEBUG_ASSERT(spin_lock_held(&thread_lock));

    DEBUG_ASSERT(thread_is_idle(t));
    DEBUG_ASSERT(t->cpu_affinity && (t->cpu_affinity & (t->cpu_affinity - 1)) == 0);

    // idle thread is special case, just jam it into the cpu's run queue in the thread's
    // affinity mask and mark it ready.
    t->state = THREAD_READY;
    cpu_num_t cpu = lowest_cpu_set(t->cpu_affinity);
    t->curr_cpu = cpu;
    insert_in_run_queue_head(cpu, t);
}

// the thread is voluntarily giving up its time slice
void sched_yield() {
    DEBUG_ASSERT(spin_lock_held(&thread_lock));

    thread_t* current_thread = get_current_thread();
    DEBUG_ASSERT(!thread_is_idle(current_thread));

    LOCAL_KTRACE0("sched_yield");

    // consume the rest of the time slice, deboost ourself, and go to the end of a queue
    current_thread->remaining_time_slice = 0;
    deboost_thread(current_thread, false);

    current_thread->state = THREAD_READY;

    if (local_migrate_if_needed(current_thread)) {
        return;
    }

    insert_in_run_queue_tail(arch_curr_cpu_num(), current_thread);
    sched_resched_internal();
}

// the current thread is being preempted from interrupt context
void sched_preempt() {
    DEBUG_ASSERT(spin_lock_held(&thread_lock));

    thread_t* current_thread = get_current_thread();
    uint curr_cpu = arch_curr_cpu_num();

    DEBUG_ASSERT(current_thread->curr_cpu == curr_cpu);
    DEBUG_ASSERT(current_thread->last_cpu == current_thread->curr_cpu);
    LOCAL_KTRACE0("sched_preempt");

    current_thread->state = THREAD_READY;

    // idle thread doesn't go in the run queue
    if (likely(!thread_is_idle(current_thread))) {
        if (current_thread->remaining_time_slice <= 0) {
            // if we're out of quantum, deboost the thread and put it at the tail of a queue
            deboost_thread(current_thread, true);
        }

        if (local_migrate_if_needed(current_thread)) {
            return;
        }

        if (current_thread->remaining_time_slice > 0) {
            insert_in_run_queue_head(curr_cpu, current_thread);
        } else {
            insert_in_run_queue_tail(curr_cpu, current_thread);
        }
    }

    sched_resched_internal();
}

// the current thread is voluntarily reevaluating the scheduler on the current cpu
void sched_reschedule() {
    DEBUG_ASSERT(spin_lock_held(&thread_lock));

    thread_t* current_thread = get_current_thread();
    uint curr_cpu = arch_curr_cpu_num();

    if (current_thread->disable_counts != 0) {
        current_thread->preempt_pending = true;
        return;
    }

    DEBUG_ASSERT(current_thread->curr_cpu == curr_cpu);
    DEBUG_ASSERT(current_thread->last_cpu == current_thread->curr_cpu);
    LOCAL_KTRACE0("sched_reschedule");

    current_thread->state = THREAD_READY;

    // idle thread doesn't go in the run queue
    if (likely(!thread_is_idle(current_thread))) {

        // deboost the current thread
        deboost_thread(current_thread, false);

        if (local_migrate_if_needed(current_thread)) {
            return;
        }

        if (current_thread->remaining_time_slice > 0) {
            insert_in_run_queue_head(curr_cpu, current_thread);
        } else {
            insert_in_run_queue_tail(curr_cpu, current_thread);
        }
    }

    sched_resched_internal();
}

// migrate the current thread to a new cpu and locally reschedule to seal the deal
static void migrate_current_thread(thread_t* current_thread) TA_REQ(thread_lock) {
    bool local_resched = false;
    cpu_mask_t accum_cpu_mask = 0;

    // current thread, so just shove ourself into another cpu's queue and reschedule locally
    current_thread->state = THREAD_READY;
    find_cpu_and_insert(current_thread, &local_resched, &accum_cpu_mask);
    if (accum_cpu_mask) {
        mp_reschedule(accum_cpu_mask, 0);
    }
    sched_resched_internal();
}

// migrate all non-pinned threads assigned to |old_cpu| to other queues
//
// must be called on |old_cpu|
void sched_transition_off_cpu(cpu_num_t old_cpu) {
    DEBUG_ASSERT(spin_lock_held(&thread_lock));
    DEBUG_ASSERT(old_cpu == arch_curr_cpu_num());

    // Ensure we do not get scheduled on anymore.
    mp_set_curr_cpu_active(false);

    thread_t* t;
    bool local_resched = false;
    cpu_mask_t accum_cpu_mask = 0;
    cpu_mask_t pinned_mask = cpu_num_to_mask(old_cpu);
    list_node_t pinned_threads = LIST_INITIAL_VALUE(pinned_threads);
    while (!thread_is_idle(t = sched_get_top_thread(old_cpu))) {
        // Threads pinned to old_cpu can't run anywhere else, so put them
        // into a temporary list and deal with them later.
        if (t->cpu_affinity != pinned_mask) {
            find_cpu_and_insert(t, &local_resched, &accum_cpu_mask);
            DEBUG_ASSERT(!local_resched);
        } else {
            DEBUG_ASSERT(!list_in_list(&t->queue_node));
            list_add_head(&pinned_threads, &t->queue_node);
        }
    }

    // Put pinned threads back on old_cpu's queue.
    while ((t = list_remove_head_type(&pinned_threads, thread_t, queue_node)) != NULL) {
        insert_in_run_queue_head(old_cpu, t);
    }

    if (accum_cpu_mask) {
        mp_reschedule(accum_cpu_mask, 0);
    }
}

// check to see if the current thread needs to migrate to a new core
// the passed argument must be the current thread and must already be pushed into the READY state
static bool local_migrate_if_needed(thread_t* curr_thread) TA_REQ(thread_lock) {
    DEBUG_ASSERT(curr_thread == get_current_thread());
    DEBUG_ASSERT(curr_thread->state == THREAD_READY);

    // if the affinity mask does not include the current cpu, migrate us right now
    if (unlikely((curr_thread->cpu_affinity & cpu_num_to_mask(curr_thread->curr_cpu)) == 0)) {
        migrate_current_thread(curr_thread);
        return true;
    }
    return false;
}

// potentially migrate a thread to a new core based on the affinity mask on the thread. If it's
// running or in a scheduler queue, handle it.
void sched_migrate(thread_t* t) {
    DEBUG_ASSERT(spin_lock_held(&thread_lock));

    bool local_resched = false;
    cpu_mask_t accum_cpu_mask = 0;
    switch (t->state) {
    case THREAD_RUNNING:
        // see if we need to migrate
        if (t->cpu_affinity & cpu_num_to_mask(t->curr_cpu)) {
            // it's running and the new mask contains the core it's already running on, nothing to do.
            //TRACEF("t %p nomigrate\n", t);
            return;
        }

        // we need to migrate
        if (t == get_current_thread()) {
            // current thread, so just shove ourself into another cpu's queue and reschedule locally
            migrate_current_thread(t);
            return;
        } else {
            // running on another cpu, interrupt and let sched_preempt() sort it out
            accum_cpu_mask = cpu_num_to_mask(t->curr_cpu);
        }
        break;
    case THREAD_READY:
        if (t->cpu_affinity & cpu_num_to_mask(t->curr_cpu)) {
            // it's ready and the new mask contains the core it's already waiting on, nothing to do.
            //TRACEF("t %p nomigrate\n", t);
            return;
        }

        // it's sitting in a run queue somewhere, so pull it out of that one and find a new home
        DEBUG_ASSERT_MSG(list_in_list(&t->queue_node), "thread %p name %s curr_cpu %u\n", t, t->name, t->curr_cpu);
        remove_from_run_queue(t, t->effec_priority);

        find_cpu_and_insert(t, &local_resched, &accum_cpu_mask);
        break;
    default:
        // the other states do not matter, exit
        return;
    }

    // send some ipis based on the previous code
    if (accum_cpu_mask) {
        mp_reschedule(accum_cpu_mask, 0);
    }
    if (local_resched) {
        sched_reschedule();
    }
}

// the effective priority of a thread has changed, do what is necessary to move the thread
// from different queues and inform us if we need to reschedule
static void sched_priority_changed(thread_t* t, int old_prio,
                                   bool* local_resched,
                                   cpu_mask_t* accum_cpu_mask) TA_REQ(thread_lock) {
    switch (t->state) {
    case THREAD_RUNNING:
        if (t->effec_priority < old_prio) {
            // we're currently running and dropped our effective priority, might want to resched
            if (t == get_current_thread()) {
                *local_resched = true;
            } else {
                *accum_cpu_mask |= cpu_num_to_mask(t->curr_cpu);
            }
        }
        break;
    case THREAD_READY:
        // it's sitting in a run queue somewhere, remove and add back to the proper queue on that cpu
        DEBUG_ASSERT_MSG(list_in_list(&t->queue_node), "thread %p name %s curr_cpu %u\n", t, t->name, t->curr_cpu);
        remove_from_run_queue(t, old_prio);

        // insert ourself into the new queue
        if (t->effec_priority > old_prio) {
            insert_in_run_queue_head(t->curr_cpu, t);

            // we may now be higher priority than the current thread on this cpu, reschedule
            if (t->curr_cpu == arch_curr_cpu_num()) {
                *local_resched = true;
            } else {
                *accum_cpu_mask |= cpu_num_to_mask(t->curr_cpu);
            }
        } else {
            insert_in_run_queue_tail(t->curr_cpu, t);
        }

        break;
    case THREAD_BLOCKED:
        // it's blocked on something, sitting in a wait queue, so we may need to move it around
        // within the wait queue.
        // note it's possible to be blocked but not in a wait queue if the thread is in transition
        // from blocked to running
        if (t->blocking_wait_queue) {
            wait_queue_priority_changed(t, old_prio);
        }
        break;
    default:
        // the other states do not matter, exit
        return;
    }
}

// set the priority to the higher value of what it was before and the newly inherited value
// pri < 0 disables priority inheritance and goes back to the naturally computed values
void sched_inherit_priority(thread_t* t, int pri, bool* local_resched) {
    DEBUG_ASSERT(spin_lock_held(&thread_lock));

    if (pri > HIGHEST_PRIORITY) {
        pri = HIGHEST_PRIORITY;
    }

    // if we're setting it to something real and it's less than the current, skip
    if (pri >= 0 && pri <= t->inherited_priority) {
        return;
    }

    // adjust the priority and remember the old value
    t->inherited_priority = pri;
    int old_ep = t->effec_priority;
    compute_effec_priority(t);
    if (old_ep == t->effec_priority) {
        // same effective priority, nothing to do
        return;
    }

    // see if we need to do something based on the state of the thread
    cpu_mask_t accum_cpu_mask = 0;
    sched_priority_changed(t, old_ep, local_resched, &accum_cpu_mask);

    // send some ipis based on the previous code
    if (accum_cpu_mask) {
        mp_reschedule(accum_cpu_mask, 0);
    }
}

// changes the thread's base priority and if the re-computed effective priority changed
//  then the thread is moved to the proper queue on the same processor and a re-schedule
//  might be issued.
void sched_change_priority(thread_t* t, int pri) {
    DEBUG_ASSERT(spin_lock_held(&thread_lock));

    if (unlikely(t->state == THREAD_DEATH)) {
        return;
    }

    if (pri > HIGHEST_PRIORITY) {
        pri = HIGHEST_PRIORITY;
    }

    int old_ep = t->effec_priority;
    t->base_priority = pri;
    t->priority_boost = 0;

    compute_effec_priority(t);
    if (old_ep == t->effec_priority) {
        // No effective change so we exit. The boost has reset but that's ok.
        return;
    }

    cpu_mask_t accum_cpu_mask = 0;
    bool local_resched = false;

    // see if we need to do something based on the state of the thread.
    sched_priority_changed(t, old_ep, &local_resched, &accum_cpu_mask);

    // send some ipis based on the previous code
    if (accum_cpu_mask) {
        mp_reschedule(accum_cpu_mask, 0);
    }
    if (local_resched) {
        sched_reschedule();
    }
}

// preemption timer that is set whenever a thread is scheduled
void sched_preempt_timer_tick(zx_time_t now) {
    // if the preemption timer went off on the idle or a real time thread, ignore it
    thread_t* current_thread = get_current_thread();
    if (unlikely(thread_is_real_time_or_idle(current_thread))) {
        return;
    }

    LOCAL_KTRACE2("sched_preempt_timer_tick", (uint32_t)current_thread->user_tid,
                  current_thread->remaining_time_slice);

    // did this tick complete the time slice?
    DEBUG_ASSERT(now > current_thread->last_started_running);
    zx_duration_t delta = zx_time_sub_time(now, current_thread->last_started_running);
    if (delta >= current_thread->remaining_time_slice) {
        // we completed the time slice, do not restart it and let the scheduler run
        current_thread->remaining_time_slice = 0;

        // set a timer to go off on the time slice interval from now
        timer_preempt_reset(zx_time_add_duration(now, THREAD_INITIAL_TIME_SLICE));

        // Mark a reschedule as pending.  The irq handler will call back
        // into us with sched_preempt().
        thread_preempt_set_pending();
    } else {
        // the timer tick must have fired early, reschedule and continue
        zx_time_t deadline = zx_time_add_duration(current_thread->last_started_running,
                                                  current_thread->remaining_time_slice);
        timer_preempt_reset(deadline);
    }
}

// On ARM64 with safe-stack, it's no longer possible to use the unsafe-sp
// after set_current_thread (we'd now see newthread's unsafe-sp instead!).
// Hence this function and everything it calls between this point and the
// the low-level context switch must be marked with __NO_SAFESTACK.
__NO_SAFESTACK static void final_context_switch(thread_t* oldthread,
                                                thread_t* newthread) {
    set_current_thread(newthread);
    arch_context_switch(oldthread, newthread);
}

// Internal reschedule routine. The current thread needs to already be in whatever
// state and queues it needs to be in. This routine simply picks the next thread and
// switches to it.
void sched_resched_internal() {
    thread_t* current_thread = get_current_thread();
    uint cpu = arch_curr_cpu_num();

    DEBUG_ASSERT(arch_ints_disabled());
    DEBUG_ASSERT(spin_lock_held(&thread_lock));
    DEBUG_ASSERT_MSG(current_thread->state != THREAD_RUNNING, "state %d\n", current_thread->state);
    DEBUG_ASSERT(!arch_blocking_disallowed());

    CPU_STATS_INC(reschedules);

    // pick a new thread to run
    thread_t* newthread = sched_get_top_thread(cpu);

    DEBUG_ASSERT(newthread);

    newthread->state = THREAD_RUNNING;

    thread_t* oldthread = current_thread;
    oldthread->preempt_pending = false;

    LOCAL_KTRACE2("resched old pri", (uint32_t)oldthread->user_tid, effec_priority(oldthread));
    LOCAL_KTRACE2("resched new pri", (uint32_t)newthread->user_tid, effec_priority(newthread));

    // call this even if we're not changing threads, to handle the case where another
    // core rescheduled us but the work disappeared before we got to run.
    mp_prepare_current_cpu_idle_state(thread_is_idle(newthread));

    // if it's the same thread as we're already running, exit
    if (newthread == oldthread) {
        return;
    }

    zx_time_t now = current_time();

    // account for time used on the old thread
    DEBUG_ASSERT(now >= oldthread->last_started_running);
    zx_duration_t old_runtime = zx_time_sub_time(now, oldthread->last_started_running);
    oldthread->runtime_ns = zx_duration_add_duration(oldthread->runtime_ns, old_runtime);
    oldthread->remaining_time_slice = zx_duration_sub_duration(
        oldthread->remaining_time_slice, MIN(old_runtime, oldthread->remaining_time_slice));

    // set up quantum for the new thread if it was consumed
    if (newthread->remaining_time_slice == 0) {
        newthread->remaining_time_slice = THREAD_INITIAL_TIME_SLICE;
    }

    newthread->last_started_running = now;

    // mark the cpu ownership of the threads
    if (oldthread->state != THREAD_READY) {
        oldthread->curr_cpu = INVALID_CPU;
    }
    newthread->last_cpu = cpu;
    newthread->curr_cpu = cpu;

    // if we selected the idle thread the cpu's run queue must be empty, so mark the
    // cpu as idle
    if (thread_is_idle(newthread)) {
        mp_set_cpu_idle(cpu);
    }

    if (thread_is_realtime(newthread)) {
        mp_set_cpu_realtime(cpu);
    } else {
        mp_set_cpu_non_realtime(cpu);
    }

    CPU_STATS_INC(context_switches);

    if (thread_is_idle(oldthread)) {
        zx_duration_t delta = zx_time_sub_time(now, oldthread->last_started_running);
        percpu[cpu].stats.idle_time = zx_duration_add_duration(percpu[cpu].stats.idle_time, delta);
    }

    LOCAL_KTRACE2("CS timeslice old", (uint32_t)oldthread->user_tid, oldthread->remaining_time_slice);
    LOCAL_KTRACE2("CS timeslice new", (uint32_t)newthread->user_tid, newthread->remaining_time_slice);

    ktrace(TAG_CONTEXT_SWITCH, (uint32_t)newthread->user_tid,
           (cpu | (oldthread->state << 8) |
            (oldthread->effec_priority << 16) | (newthread->effec_priority << 24)),
           (uint32_t)(uintptr_t)oldthread, (uint32_t)(uintptr_t)newthread);

    if (thread_is_real_time_or_idle(newthread)) {
        if (!thread_is_real_time_or_idle(oldthread)) {
            // if we're switching from a non real time to a real time, cancel
            // the preemption timer.
            TRACE_CONTEXT_SWITCH("stop preempt, cpu %u, old %p (%s), new %p (%s)\n",
                                 cpu, oldthread, oldthread->name, newthread, newthread->name);
            timer_preempt_cancel();
        }
    } else {
        // set up a one shot timer to handle the remaining time slice on this thread
        TRACE_CONTEXT_SWITCH("start preempt, cpu %u, old %p (%s), new %p (%s)\n",
                             cpu, oldthread, oldthread->name, newthread, newthread->name);

        // make sure the time slice is reasonable
        DEBUG_ASSERT(newthread->remaining_time_slice > 0 && newthread->remaining_time_slice < ZX_SEC(1));

        timer_preempt_reset(zx_time_add_duration(now, newthread->remaining_time_slice));
    }

    // set some optional target debug leds
    target_set_debug_led(0, !thread_is_idle(newthread));

    TRACE_CONTEXT_SWITCH("cpu %u old %p (%s, pri %d [%d:%d], flags 0x%x) "
                         "new %p (%s, pri %d [%d:%d], flags 0x%x)\n",
                         cpu, oldthread, oldthread->name, oldthread->effec_priority,
                         oldthread->base_priority, oldthread->priority_boost,
                         oldthread->flags, newthread, newthread->name,
                         newthread->effec_priority, newthread->base_priority,
                         newthread->priority_boost, newthread->flags);

    // see if we need to swap mmu context
    if (newthread->aspace != oldthread->aspace) {
        vmm_context_switch(oldthread->aspace, newthread->aspace);
    }

    // do the low level context switch
    final_context_switch(oldthread, newthread);
}

void sched_init_early() {
    // initialize the run queues
    for (unsigned int cpu = 0; cpu < SMP_MAX_CPUS; cpu++)
        for (unsigned int i = 0; i < NUM_PRIORITIES; i++) {
            list_initialize(&percpu[cpu].run_queue[i]);
        }
}