sched.c

linux 内核源代码

	}
	spin_unlock(&spu_prio->runq_lock);
	__set_current_state(TASK_RUNNING);
	remove_wait_queue(&ctx->stop_wq, &wait);
}

static struct spu *spu_get_idle(struct spu_context *ctx)
{
	struct spu *spu, *aff_ref_spu;
	int node, n;

	if (ctx->gang) {
		mutex_lock(&ctx->gang->aff_mutex);
		if (has_affinity(ctx)) {
			aff_ref_spu = ctx->gang->aff_ref_spu;
			atomic_inc(&ctx->gang->aff_sched_count);
			mutex_unlock(&ctx->gang->aff_mutex);
			node = aff_ref_spu->node;

			mutex_lock(&cbe_spu_info[node].list_mutex);
			spu = ctx_location(aff_ref_spu, ctx->aff_offset, node);
			if (spu && spu->alloc_state == SPU_FREE)
				goto found;
			mutex_unlock(&cbe_spu_info[node].list_mutex);

			mutex_lock(&ctx->gang->aff_mutex);
			if (atomic_dec_and_test(&ctx->gang->aff_sched_count))
				ctx->gang->aff_ref_spu = NULL;
			mutex_unlock(&ctx->gang->aff_mutex);

			return NULL;
		}
		mutex_unlock(&ctx->gang->aff_mutex);
	}
	node = cpu_to_node(raw_smp_processor_id());
	for (n = 0; n < MAX_NUMNODES; n++, node++) {
		node = (node < MAX_NUMNODES) ? node : 0;
		if (!node_allowed(ctx, node))
			continue;

		mutex_lock(&cbe_spu_info[node].list_mutex);
		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
			if (spu->alloc_state == SPU_FREE)
				goto found;
		}
		mutex_unlock(&cbe_spu_info[node].list_mutex);
	}

	return NULL;

 found:
	spu->alloc_state = SPU_USED;
	mutex_unlock(&cbe_spu_info[node].list_mutex);
	pr_debug("Got SPU %d %d\n", spu->number, spu->node);
	spu_init_channels(spu);
	return spu;
}

/**
 * find_victim - find a lower priority context to preempt
 * @ctx:	canidate context for running
 *
 * Returns the freed physical spu to run the new context on.
 */
static struct spu *find_victim(struct spu_context *ctx)
{
	struct spu_context *victim = NULL;
	struct spu *spu;
	int node, n;

	/*
	 * Look for a possible preemption candidate on the local node first.
	 * If there is no candidate look at the other nodes.  This isn't
	 * exactly fair, but so far the whole spu schedule tries to keep
	 * a strong node affinity.  We might want to fine-tune this in
	 * the future.
	 */
 restart:
	node = cpu_to_node(raw_smp_processor_id());
	for (n = 0; n < MAX_NUMNODES; n++, node++) {
		node = (node < MAX_NUMNODES) ? node : 0;
		if (!node_allowed(ctx, node))
			continue;

		mutex_lock(&cbe_spu_info[node].list_mutex);
		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
			struct spu_context *tmp = spu->ctx;

			if (tmp && tmp->prio > ctx->prio &&
			    (!victim || tmp->prio > victim->prio))
				victim = spu->ctx;
		}
		mutex_unlock(&cbe_spu_info[node].list_mutex);

		if (victim) {
			/*
			 * This nests ctx->state_mutex, but we always lock
			 * higher priority contexts before lower priority
			 * ones, so this is safe until we introduce
			 * priority inheritance schemes.
			 */
			if (!mutex_trylock(&victim->state_mutex)) {
				victim = NULL;
				goto restart;
			}

			spu = victim->spu;
			if (!spu) {
				/*
				 * This race can happen because we've dropped
				 * the active list mutex.  No a problem, just
				 * restart the search.
				 */
				mutex_unlock(&victim->state_mutex);
				victim = NULL;
				goto restart;
			}

			mutex_lock(&cbe_spu_info[node].list_mutex);
			cbe_spu_info[node].nr_active--;
			spu_unbind_context(spu, victim);
			mutex_unlock(&cbe_spu_info[node].list_mutex);

			victim->stats.invol_ctx_switch++;
			spu->stats.invol_ctx_switch++;
			mutex_unlock(&victim->state_mutex);
			/*
			 * We need to break out of the wait loop in spu_run
			 * manually to ensure this context gets put on the
			 * runqueue again ASAP.
			 */
			wake_up(&victim->stop_wq);
			return spu;
		}
	}

	return NULL;
}

/**
 * spu_activate - find a free spu for a context and execute it
 * @ctx:	spu context to schedule
 * @flags:	flags (currently ignored)
 *
 * Tries to find a free spu to run @ctx.  If no free spu is available
 * add the context to the runqueue so it gets woken up once an spu
 * is available.
 */
int spu_activate(struct spu_context *ctx, unsigned long flags)
{
	do {
		struct spu *spu;

		/*
		 * If there are multiple threads waiting for a single context
		 * only one actually binds the context while the others will
		 * only be able to acquire the state_mutex once the context
		 * already is in runnable state.
		 */
		if (ctx->spu)
			return 0;

		spu = spu_get_idle(ctx);
		/*
		 * If this is a realtime thread we try to get it running by
		 * preempting a lower priority thread.
		 */
		if (!spu && rt_prio(ctx->prio))
			spu = find_victim(ctx);
		if (spu) {
			int node = spu->node;

			mutex_lock(&cbe_spu_info[node].list_mutex);
			spu_bind_context(spu, ctx);
			cbe_spu_info[node].nr_active++;
			mutex_unlock(&cbe_spu_info[node].list_mutex);
			return 0;
		}

		spu_prio_wait(ctx);
	} while (!signal_pending(current));

	return -ERESTARTSYS;
}

/**
 * grab_runnable_context - try to find a runnable context
 *
 * Remove the highest priority context on the runqueue and return it
 * to the caller.  Returns %NULL if no runnable context was found.
 */
static struct spu_context *grab_runnable_context(int prio, int node)
{
	struct spu_context *ctx;
	int best;

	spin_lock(&spu_prio->runq_lock);
	best = find_first_bit(spu_prio->bitmap, prio);
	while (best < prio) {
		struct list_head *rq = &spu_prio->runq[best];

		list_for_each_entry(ctx, rq, rq) {
			/* XXX(hch): check for affinity here aswell */
			if (__node_allowed(ctx, node)) {
				__spu_del_from_rq(ctx);
				goto found;
			}
		}
		best++;
	}
	ctx = NULL;
 found:
	spin_unlock(&spu_prio->runq_lock);
	return ctx;
}

static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
{
	struct spu *spu = ctx->spu;
	struct spu_context *new = NULL;

	if (spu) {
		new = grab_runnable_context(max_prio, spu->node);
		if (new || force) {
			int node = spu->node;

			mutex_lock(&cbe_spu_info[node].list_mutex);
			spu_unbind_context(spu, ctx);
			spu->alloc_state = SPU_FREE;
			cbe_spu_info[node].nr_active--;
			mutex_unlock(&cbe_spu_info[node].list_mutex);

			ctx->stats.vol_ctx_switch++;
			spu->stats.vol_ctx_switch++;

			if (new)
				wake_up(&new->stop_wq);
		}

	}

	return new != NULL;
}

/**
 * spu_deactivate - unbind a context from it's physical spu
 * @ctx:	spu context to unbind
 *
 * Unbind @ctx from the physical spu it is running on and schedule
 * the highest priority context to run on the freed physical spu.
 */
void spu_deactivate(struct spu_context *ctx)
{
	__spu_deactivate(ctx, 1, MAX_PRIO);
}

/**
 * spu_yield -	yield a physical spu if others are waiting
 * @ctx:	spu context to yield
 *
 * Check if there is a higher priority context waiting and if yes
 * unbind @ctx from the physical spu and schedule the highest
 * priority context to run on the freed physical spu instead.
 */
void spu_yield(struct spu_context *ctx)
{
	if (!(ctx->flags & SPU_CREATE_NOSCHED)) {
		mutex_lock(&ctx->state_mutex);
		__spu_deactivate(ctx, 0, MAX_PRIO);
		mutex_unlock(&ctx->state_mutex);
	}
}

static noinline void spusched_tick(struct spu_context *ctx)
{
	if (ctx->flags & SPU_CREATE_NOSCHED)
		return;
	if (ctx->policy == SCHED_FIFO)
		return;

	if (--ctx->time_slice)
		return;

	/*
	 * Unfortunately list_mutex ranks outside of state_mutex, so
	 * we have to trylock here.  If we fail give the context another
	 * tick and try again.
	 */
	if (mutex_trylock(&ctx->state_mutex)) {
		struct spu *spu = ctx->spu;
		struct spu_context *new;

		new = grab_runnable_context(ctx->prio + 1, spu->node);
		if (new) {
			spu_unbind_context(spu, ctx);
			ctx->stats.invol_ctx_switch++;
			spu->stats.invol_ctx_switch++;
			spu->alloc_state = SPU_FREE;
			cbe_spu_info[spu->node].nr_active--;
			wake_up(&new->stop_wq);
			/*
			 * We need to break out of the wait loop in
			 * spu_run manually to ensure this context
			 * gets put on the runqueue again ASAP.
			 */
			wake_up(&ctx->stop_wq);
		}
		spu_set_timeslice(ctx);
		mutex_unlock(&ctx->state_mutex);
	} else {
		ctx->time_slice++;
	}
}

/**
 * count_active_contexts - count nr of active tasks
 *
 * Return the number of tasks currently running or waiting to run.
 *
 * Note that we don't take runq_lock / list_mutex here.  Reading
 * a single 32bit value is atomic on powerpc, and we don't care
 * about memory ordering issues here.
 */
static unsigned long count_active_contexts(void)
{
	int nr_active = 0, node;

	for (node = 0; node < MAX_NUMNODES; node++)
		nr_active += cbe_spu_info[node].nr_active;
	nr_active += spu_prio->nr_waiting;

	return nr_active;
}

/**
 * spu_calc_load - given tick count, update the avenrun load estimates.
 * @tick:	tick count
 *
 * No locking against reading these values from userspace, as for
 * the CPU loadavg code.
 */
static void spu_calc_load(unsigned long ticks)
{
	unsigned long active_tasks; /* fixed-point */
	static int count = LOAD_FREQ;

	count -= ticks;

	if (unlikely(count < 0)) {
		active_tasks = count_active_contexts() * FIXED_1;
		do {
			CALC_LOAD(spu_avenrun[0], EXP_1, active_tasks);
			CALC_LOAD(spu_avenrun[1], EXP_5, active_tasks);
			CALC_LOAD(spu_avenrun[2], EXP_15, active_tasks);
			count += LOAD_FREQ;
		} while (count < 0);
	}
}

static void spusched_wake(unsigned long data)
{
	mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
	wake_up_process(spusched_task);
	spu_calc_load(SPUSCHED_TICK);
}

static int spusched_thread(void *unused)
{
	struct spu *spu;
	int node;

	while (!kthread_should_stop()) {
		set_current_state(TASK_INTERRUPTIBLE);
		schedule();
		for (node = 0; node < MAX_NUMNODES; node++) {
			mutex_lock(&cbe_spu_info[node].list_mutex);
			list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list)
				if (spu->ctx)
					spusched_tick(spu->ctx);
			mutex_unlock(&cbe_spu_info[node].list_mutex);
		}
	}

	return 0;
}

#define LOAD_INT(x) ((x) >> FSHIFT)
#define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)

static int show_spu_loadavg(struct seq_file *s, void *private)
{
	int a, b, c;

	a = spu_avenrun[0] + (FIXED_1/200);
	b = spu_avenrun[1] + (FIXED_1/200);
	c = spu_avenrun[2] + (FIXED_1/200);

	/*
	 * Note that last_pid doesn't really make much sense for the
	 * SPU loadavg (it even seems very odd on the CPU side..),
	 * but we include it here to have a 100% compatible interface.
	 */
	seq_printf(s, "%d.%02d %d.%02d %d.%02d %ld/%d %d\n",
		LOAD_INT(a), LOAD_FRAC(a),
		LOAD_INT(b), LOAD_FRAC(b),
		LOAD_INT(c), LOAD_FRAC(c),
		count_active_contexts(),
		atomic_read(&nr_spu_contexts),
		current->nsproxy->pid_ns->last_pid);
	return 0;
}

static int spu_loadavg_open(struct inode *inode, struct file *file)
{
	return single_open(file, show_spu_loadavg, NULL);
}

static const struct file_operations spu_loadavg_fops = {
	.open		= spu_loadavg_open,
	.read		= seq_read,
	.llseek		= seq_lseek,
	.release	= single_release,
};

int __init spu_sched_init(void)
{
	struct proc_dir_entry *entry;
	int err = -ENOMEM, i;

	spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
	if (!spu_prio)
		goto out;

	for (i = 0; i < MAX_PRIO; i++) {
		INIT_LIST_HEAD(&spu_prio->runq[i]);
		__clear_bit(i, spu_prio->bitmap);
	}
	spin_lock_init(&spu_prio->runq_lock);

	setup_timer(&spusched_timer, spusched_wake, 0);

	spusched_task = kthread_run(spusched_thread, NULL, "spusched");
	if (IS_ERR(spusched_task)) {
		err = PTR_ERR(spusched_task);
		goto out_free_spu_prio;
	}

	entry = create_proc_entry("spu_loadavg", 0, NULL);
	if (!entry)
		goto out_stop_kthread;
	entry->proc_fops = &spu_loadavg_fops;

	pr_debug("spusched: tick: %d, min ticks: %d, default ticks: %d\n",
			SPUSCHED_TICK, MIN_SPU_TIMESLICE, DEF_SPU_TIMESLICE);
	return 0;

 out_stop_kthread:
	kthread_stop(spusched_task);
 out_free_spu_prio:
	kfree(spu_prio);
 out:
	return err;
}

void spu_sched_exit(void)
{
	struct spu *spu;
	int node;

	remove_proc_entry("spu_loadavg", NULL);

	del_timer_sync(&spusched_timer);
	kthread_stop(spusched_task);

	for (node = 0; node < MAX_NUMNODES; node++) {
		mutex_lock(&cbe_spu_info[node].list_mutex);
		list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list)
			if (spu->alloc_state != SPU_FREE)
				spu->alloc_state = SPU_FREE;
		mutex_unlock(&cbe_spu_info[node].list_mutex);
	}
	kfree(spu_prio);
}