source: mainline/kernel/generic/src/proc/scheduler.c@ 23f36a3

/*
 * Copyright (c) 2010 Jakub Jermar
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions
 * are met:
 *
 * - Redistributions of source code must retain the above copyright
 *   notice, this list of conditions and the following disclaimer.
 * - Redistributions in binary form must reproduce the above copyright
 *   notice, this list of conditions and the following disclaimer in the
 *   documentation and/or other materials provided with the distribution.
 * - The name of the author may not be used to endorse or promote products
 *   derived from this software without specific prior written permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

/** @addtogroup kernel_generic_proc
 * @{
 */

/**
 * @file
 * @brief Scheduler and load balancing.
 *
 * This file contains the scheduler and kcpulb kernel thread which
 * performs load-balancing of per-CPU run queues.
 */

#include <assert.h>
#include <atomic.h>
#include <proc/scheduler.h>
#include <proc/thread.h>
#include <proc/task.h>
#include <mm/frame.h>
#include <mm/page.h>
#include <mm/as.h>
#include <time/timeout.h>
#include <time/delay.h>
#include <arch/asm.h>
#include <arch/faddr.h>
#include <arch/cycle.h>
#include <atomic.h>
#include <synch/spinlock.h>
#include <config.h>
#include <context.h>
#include <fpu_context.h>
#include <halt.h>
#include <arch.h>
#include <adt/list.h>
#include <panic.h>
#include <cpu.h>
#include <stdio.h>
#include <log.h>
#include <stacktrace.h>

static void scheduler_separated_stack(void);
static void fpu_restore(void);

atomic_size_t nrdy;  /**< Number of ready threads in the system. */

/** Take actions before new thread runs.
 *
 * Perform actions that need to be
 * taken before the newly selected
 * thread is passed control.
 *
 * THREAD->lock is locked on entry
 *
 */
static void before_thread_runs(void)
{
        before_thread_runs_arch();

        fpu_restore();

#ifdef CONFIG_UDEBUG
        if (THREAD->btrace) {
                istate_t *istate = THREAD->udebug.uspace_state;
                if (istate != NULL) {
                        printf("Thread %" PRIu64 " stack trace:\n", THREAD->tid);
                        stack_trace_istate(istate);
                }

                THREAD->btrace = false;
        }
#endif
}

/** Take actions after THREAD had run.
 *
 * Perform actions that need to be
 * taken after the running thread
 * had been preempted by the scheduler.
 *
 * THREAD->lock is locked on entry
 *
 */
static void after_thread_ran(void)
{
        after_thread_ran_arch();
}

#ifdef CONFIG_FPU_LAZY
void scheduler_fpu_lazy_request(void)
{
        fpu_enable();

        /* We need this lock to ensure synchronization with thread destructor. */
        irq_spinlock_lock(&CPU->fpu_lock, false);

        /* Save old context */
        thread_t *owner = atomic_load_explicit(&CPU->fpu_owner, memory_order_relaxed);
        if (owner != NULL) {
                fpu_context_save(&owner->fpu_context);
                atomic_store_explicit(&CPU->fpu_owner, NULL, memory_order_relaxed);
        }

        irq_spinlock_unlock(&CPU->fpu_lock, false);

        if (THREAD->fpu_context_exists) {
                fpu_context_restore(&THREAD->fpu_context);
        } else {
                fpu_init();
                THREAD->fpu_context_exists = true;
        }

        atomic_store_explicit(&CPU->fpu_owner, THREAD, memory_order_relaxed);
}
#endif /* CONFIG_FPU_LAZY */

/** Initialize scheduler
 *
 * Initialize kernel scheduler.
 *
 */
void scheduler_init(void)
{
}

/** Get thread to be scheduled
 *
 * Get the optimal thread to be scheduled
 * according to thread accounting and scheduler
 * policy.
 *
 * @return Thread to be scheduled.
 *
 */
static thread_t *try_find_thread(int *rq_index)
{
        assert(interrupts_disabled());
        assert(CPU != NULL);

        if (atomic_load(&CPU->nrdy) == 0)
                return NULL;

        for (int i = 0; i < RQ_COUNT; i++) {
                irq_spinlock_lock(&(CPU->rq[i].lock), false);
                if (CPU->rq[i].n == 0) {
                        /*
                         * If this queue is empty, try a lower-priority queue.
                         */
                        irq_spinlock_unlock(&(CPU->rq[i].lock), false);
                        continue;
                }

                atomic_dec(&CPU->nrdy);
                atomic_dec(&nrdy);
                CPU->rq[i].n--;

                /*
                 * Take the first thread from the queue.
                 */
                thread_t *thread = list_get_instance(
                    list_first(&CPU->rq[i].rq), thread_t, rq_link);
                list_remove(&thread->rq_link);

                irq_spinlock_pass(&(CPU->rq[i].lock), &thread->lock);

                thread->cpu = CPU;
                thread->priority = i;  /* Correct rq index */

                /* Time allocation in microseconds. */
                uint64_t time_to_run = (i + 1) * 10000;

                /* This is safe because interrupts are disabled. */
                CPU_LOCAL->preempt_deadline =
                    CPU_LOCAL->current_clock_tick + us2ticks(time_to_run);

                /*
                 * Clear the stolen flag so that it can be migrated
                 * when load balancing needs emerge.
                 */
                thread->stolen = false;
                irq_spinlock_unlock(&thread->lock, false);

                *rq_index = i;
                return thread;
        }

        return NULL;
}

/** Get thread to be scheduled
 *
 * Get the optimal thread to be scheduled
 * according to thread accounting and scheduler
 * policy.
 *
 * @return Thread to be scheduled.
 *
 */
static thread_t *find_best_thread(int *rq_index)
{
        assert(interrupts_disabled());
        assert(CPU != NULL);

        while (true) {
                thread_t *thread = try_find_thread(rq_index);

                if (thread != NULL)
                        return thread;

                /*
                 * For there was nothing to run, the CPU goes to sleep
                 * until a hardware interrupt or an IPI comes.
                 * This improves energy saving and hyperthreading.
                 */
                CPU_LOCAL->idle = true;

                /*
                 * Go to sleep with interrupts enabled.
                 * Ideally, this should be atomic, but this is not guaranteed on
                 * all platforms yet, so it is possible we will go sleep when
                 * a thread has just become available.
                 */
                cpu_interruptible_sleep();
        }
}

static void switch_task(task_t *task)
{
        /* If the task stays the same, a lot of work is avoided. */
        if (TASK == task)
                return;

        as_t *old_as = AS;
        as_t *new_as = task->as;

        /* It is possible for two tasks to share one address space. */
        if (old_as != new_as)
                as_switch(old_as, new_as);

        if (TASK)
                task_release(TASK);

        TASK = task;

        task_hold(TASK);

        before_task_runs_arch();
}

/** Prevent rq starvation
 *
 * Prevent low priority threads from starving in rq's.
 *
 * When the function decides to relink rq's, it reconnects
 * respective pointers so that in result threads with 'pri'
 * greater or equal start are moved to a higher-priority queue.
 *
 * @param start Threshold priority.
 *
 */
static void relink_rq(int start)
{
        if (CPU_LOCAL->current_clock_tick < CPU_LOCAL->relink_deadline)
                return;

        CPU_LOCAL->relink_deadline = CPU_LOCAL->current_clock_tick + NEEDS_RELINK_MAX;

        /* Temporary cache for lists we are moving. */
        list_t list;
        list_initialize(&list);

        size_t n = 0;

        /* Move every list (except the one with highest priority) one level up. */
        for (int i = RQ_COUNT - 1; i > start; i--) {
                irq_spinlock_lock(&CPU->rq[i].lock, false);

                /* Swap lists. */
                list_swap(&CPU->rq[i].rq, &list);

                /* Swap number of items. */
                size_t tmpn = CPU->rq[i].n;
                CPU->rq[i].n = n;
                n = tmpn;

                irq_spinlock_unlock(&CPU->rq[i].lock, false);
        }

        /* Append the contents of rq[start + 1]  to rq[start]. */
        if (n != 0) {
                irq_spinlock_lock(&CPU->rq[start].lock, false);
                list_concat(&CPU->rq[start].rq, &list);
                CPU->rq[start].n += n;
                irq_spinlock_unlock(&CPU->rq[start].lock, false);
        }
}

/**
 * Do whatever needs to be done with current FPU state before we switch to
 * another thread.
 */
static void fpu_cleanup(void)
{
#if (defined CONFIG_FPU) && (!defined CONFIG_FPU_LAZY)
        fpu_context_save(&THREAD->fpu_context);
#endif
}

/**
 * Set correct FPU state for this thread after switch from another thread.
 */
static void fpu_restore(void)
{
#ifdef CONFIG_FPU_LAZY
        /*
         * The only concurrent modification possible for fpu_owner here is
         * another thread changing it from itself to NULL in its destructor.
         */
        thread_t *owner = atomic_load_explicit(&CPU->fpu_owner,
            memory_order_relaxed);

        if (THREAD == owner)
                fpu_enable();
        else
                fpu_disable();

#elif defined CONFIG_FPU
        fpu_enable();
        if (THREAD->fpu_context_exists)
                fpu_context_restore(&THREAD->fpu_context);
        else {
                fpu_init();
                THREAD->fpu_context_exists = true;
        }
#endif
}

void scheduler(void)
{
        ipl_t ipl = interrupts_disable();

        if (atomic_load(&haltstate))
                halt();

        if (THREAD) {
                irq_spinlock_lock(&THREAD->lock, false);
        }

        scheduler_locked(ipl);
}

/** The scheduler
 *
 * The thread scheduling procedure.
 * Passes control directly to
 * scheduler_separated_stack().
 *
 */
void scheduler_locked(ipl_t ipl)
{
        assert(CPU != NULL);

        if (THREAD) {
                /* Update thread kernel accounting */
                THREAD->kcycles += get_cycle() - THREAD->last_cycle;

                fpu_cleanup();

                if (!context_save(&THREAD->saved_context)) {
                        /*
                         * This is the place where threads leave scheduler();
                         */

                        /* Save current CPU cycle */
                        THREAD->last_cycle = get_cycle();

                        irq_spinlock_unlock(&THREAD->lock, false);
                        interrupts_restore(THREAD->saved_ipl);

                        return;
                }

                /*
                 * Interrupt priority level of preempted thread is recorded
                 * here to facilitate scheduler() invocations from
                 * interrupts_disable()'d code (e.g. waitq_sleep_timeout()).
                 *
                 */
                THREAD->saved_ipl = ipl;
        }

        /*
         * Through the 'CURRENT' structure, we keep track of THREAD, TASK, CPU, AS
         * and preemption counter. At this point CURRENT could be coming either
         * from THREAD's or CPU's stack.
         *
         */
        current_copy(CURRENT, (current_t *) CPU_LOCAL->stack);

        /*
         * We may not keep the old stack.
         * Reason: If we kept the old stack and got blocked, for instance, in
         * find_best_thread(), the old thread could get rescheduled by another
         * CPU and overwrite the part of its own stack that was also used by
         * the scheduler on this CPU.
         *
         * Moreover, we have to bypass the compiler-generated POP sequence
         * which is fooled by SP being set to the very top of the stack.
         * Therefore the scheduler() function continues in
         * scheduler_separated_stack().
         *
         */
        context_t ctx;
        context_save(&ctx);
        context_set(&ctx, FADDR(scheduler_separated_stack),
            (uintptr_t) CPU_LOCAL->stack, STACK_SIZE);
        context_restore(&ctx);

        /* Not reached */
}

/** Scheduler stack switch wrapper
 *
 * Second part of the scheduler() function
 * using new stack. Handling the actual context
 * switch to a new thread.
 *
 */
void scheduler_separated_stack(void)
{
        assert((!THREAD) || (irq_spinlock_locked(&THREAD->lock)));
        assert(CPU != NULL);
        assert(interrupts_disabled());

        if (THREAD) {
                /* Must be run after the switch to scheduler stack */
                after_thread_ran();

                int expected;

                switch (THREAD->state) {
                case Running:
                        irq_spinlock_unlock(&THREAD->lock, false);
                        thread_ready(THREAD);
                        break;

                case Exiting:
                        irq_spinlock_unlock(&THREAD->lock, false);
                        waitq_close(&THREAD->join_wq);

                        /*
                         * Release the reference CPU has for the thread.
                         * If there are no other references (e.g. threads calling join),
                         * the thread structure is deallocated.
                         */
                        thread_put(THREAD);
                        break;

                case Sleeping:
                        expected = SLEEP_INITIAL;

                        /* Only set SLEEP_ASLEEP in sleep pad if it's still in initial state */
                        if (atomic_compare_exchange_strong_explicit(&THREAD->sleep_state,
                            &expected, SLEEP_ASLEEP,
                            memory_order_acq_rel, memory_order_acquire)) {

                                /* Prefer the thread after it's woken up. */
                                THREAD->priority = -1;
                                irq_spinlock_unlock(&THREAD->lock, false);
                        } else {
                                assert(expected == SLEEP_WOKE);
                                /* The thread has already been woken up, requeue immediately. */
                                irq_spinlock_unlock(&THREAD->lock, false);
                                thread_ready(THREAD);
                        }

                        break;

                default:
                        /*
                         * Entering state is unexpected.
                         */
                        panic("tid%" PRIu64 ": unexpected state %s.",
                            THREAD->tid, thread_states[THREAD->state]);
                        break;
                }

                THREAD = NULL;
        }

        int rq_index;
        THREAD = find_best_thread(&rq_index);

        relink_rq(rq_index);

        switch_task(THREAD->task);

        irq_spinlock_lock(&THREAD->lock, false);
        THREAD->state = Running;

#ifdef SCHEDULER_VERBOSE
        log(LF_OTHER, LVL_DEBUG,
            "cpu%u: tid %" PRIu64 " (priority=%d, ticks=%" PRIu64
            ", nrdy=%zu)", CPU->id, THREAD->tid, THREAD->priority,
            THREAD->ticks, atomic_load(&CPU->nrdy));
#endif

        /*
         * Some architectures provide late kernel PA2KA(identity)
         * mapping in a page fault handler. However, the page fault
         * handler uses the kernel stack of the running thread and
         * therefore cannot be used to map it. The kernel stack, if
         * necessary, is to be mapped in before_thread_runs(). This
         * function must be executed before the switch to the new stack.
         */
        before_thread_runs();

        /*
         * Copy the knowledge of CPU, TASK, THREAD and preemption counter to
         * thread's stack.
         */
        current_copy(CURRENT, (current_t *) THREAD->kstack);

        context_restore(&THREAD->saved_context);

        /* Not reached */
}

#ifdef CONFIG_SMP

static thread_t *steal_thread_from(cpu_t *old_cpu, int i)
{
        runq_t *old_rq = &old_cpu->rq[i];
        runq_t *new_rq = &CPU->rq[i];

        ipl_t ipl = interrupts_disable();

        irq_spinlock_lock(&old_rq->lock, false);

        /*
         * If fpu_owner is any thread in the list, its store is seen here thanks to
         * the runqueue lock.
         */
        thread_t *fpu_owner = atomic_load_explicit(&old_cpu->fpu_owner,
            memory_order_relaxed);

        /* Search rq from the back */
        list_foreach_rev(old_rq->rq, rq_link, thread_t, thread) {

                irq_spinlock_lock(&thread->lock, false);

                /*
                 * Do not steal CPU-wired threads, threads
                 * already stolen, threads for which migration
                 * was temporarily disabled or threads whose
                 * FPU context is still in the CPU.
                 */
                if (thread->stolen || thread->nomigrate ||
                    thread == fpu_owner) {
                        irq_spinlock_unlock(&thread->lock, false);
                        continue;
                }

                thread->stolen = true;
                thread->cpu = CPU;

                irq_spinlock_unlock(&thread->lock, false);

                /*
                 * Ready thread on local CPU
                 */

#ifdef KCPULB_VERBOSE
                log(LF_OTHER, LVL_DEBUG,
                    "kcpulb%u: TID %" PRIu64 " -> cpu%u, "
                    "nrdy=%ld, avg=%ld", CPU->id, thread->tid,
                    CPU->id, atomic_load(&CPU->nrdy),
                    atomic_load(&nrdy) / config.cpu_active);
#endif

                /* Remove thread from ready queue. */
                old_rq->n--;
                list_remove(&thread->rq_link);
                irq_spinlock_unlock(&old_rq->lock, false);

                /* Append thread to local queue. */
                irq_spinlock_lock(&new_rq->lock, false);
                list_append(&thread->rq_link, &new_rq->rq);
                new_rq->n++;
                irq_spinlock_unlock(&new_rq->lock, false);

                atomic_dec(&old_cpu->nrdy);
                atomic_inc(&CPU->nrdy);
                interrupts_restore(ipl);
                return thread;
        }

        irq_spinlock_unlock(&old_rq->lock, false);
        interrupts_restore(ipl);
        return NULL;
}

/** Load balancing thread
 *
 * SMP load balancing thread, supervising thread supplies
 * for the CPU it's wired to.
 *
 * @param arg Generic thread argument (unused).
 *
 */
void kcpulb(void *arg)
{
        size_t average;
        size_t rdy;

loop:
        /*
         * Work in 1s intervals.
         */
        thread_sleep(1);

not_satisfied:
        /*
         * Calculate the number of threads that will be migrated/stolen from
         * other CPU's. Note that situation can have changed between two
         * passes. Each time get the most up to date counts.
         *
         */
        average = atomic_load(&nrdy) / config.cpu_active + 1;
        rdy = atomic_load(&CPU->nrdy);

        if (average <= rdy)
                goto satisfied;

        size_t count = average - rdy;

        /*
         * Searching least priority queues on all CPU's first and most priority
         * queues on all CPU's last.
         */
        size_t acpu;
        int rq;

        for (rq = RQ_COUNT - 1; rq >= 0; rq--) {
                for (acpu = 0; acpu < config.cpu_active; acpu++) {
                        cpu_t *cpu = &cpus[acpu];

                        /*
                         * Not interested in ourselves.
                         * Doesn't require interrupt disabling for kcpulb has
                         * THREAD_FLAG_WIRED.
                         *
                         */
                        if (CPU == cpu)
                                continue;

                        if (atomic_load(&cpu->nrdy) <= average)
                                continue;

                        if (steal_thread_from(cpu, rq) && --count == 0)
                                goto satisfied;
                }
        }

        if (atomic_load(&CPU->nrdy)) {
                /*
                 * Be a little bit light-weight and let migrated threads run.
                 *
                 */
                scheduler();
        } else {
                /*
                 * We failed to migrate a single thread.
                 * Give up this turn.
                 *
                 */
                goto loop;
        }

        goto not_satisfied;

satisfied:
        goto loop;
}
#endif /* CONFIG_SMP */

/** Print information about threads & scheduler queues
 *
 */
void sched_print_list(void)
{
        size_t cpu;
        for (cpu = 0; cpu < config.cpu_count; cpu++) {
                if (!cpus[cpu].active)
                        continue;

                printf("cpu%u: address=%p, nrdy=%zu\n",
                    cpus[cpu].id, &cpus[cpu], atomic_load(&cpus[cpu].nrdy));

                unsigned int i;
                for (i = 0; i < RQ_COUNT; i++) {
                        irq_spinlock_lock(&(cpus[cpu].rq[i].lock), false);
                        if (cpus[cpu].rq[i].n == 0) {
                                irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
                                continue;
                        }

                        printf("\trq[%u]: ", i);
                        list_foreach(cpus[cpu].rq[i].rq, rq_link, thread_t,
                            thread) {
                                printf("%" PRIu64 "(%s) ", thread->tid,
                                    thread_states[thread->state]);
                        }
                        printf("\n");

                        irq_spinlock_unlock(&(cpus[cpu].rq[i].lock), false);
                }
        }
}

/** @}
 */