#include <asm/spu_priv1.h>
#include "spufs.h"
-#define SPU_MIN_TIMESLICE (100 * HZ / 1000)
+#define SPU_TIMESLICE (HZ)
-#define SPU_BITMAP_SIZE (((MAX_PRIO+BITS_PER_LONG)/BITS_PER_LONG)+1)
struct spu_prio_array {
- unsigned long bitmap[SPU_BITMAP_SIZE];
+ DECLARE_BITMAP(bitmap, MAX_PRIO);
struct list_head runq[MAX_PRIO];
spinlock_t runq_lock;
struct list_head active_list[MAX_NUMNODES];
};
static struct spu_prio_array *spu_prio;
+static struct workqueue_struct *spu_sched_wq;
static inline int node_allowed(int node)
{
return 1;
}
+void spu_start_tick(struct spu_context *ctx)
+{
+ if (ctx->policy == SCHED_RR) {
+ /*
+ * Make sure the exiting bit is cleared.
+ */
+ clear_bit(SPU_SCHED_EXITING, &ctx->sched_flags);
+ mb();
+ queue_delayed_work(spu_sched_wq, &ctx->sched_work, SPU_TIMESLICE);
+ }
+}
+
+void spu_stop_tick(struct spu_context *ctx)
+{
+ if (ctx->policy == SCHED_RR) {
+ /*
+ * While the work can be rearming normally setting this flag
+ * makes sure it does not rearm itself anymore.
+ */
+ set_bit(SPU_SCHED_EXITING, &ctx->sched_flags);
+ mb();
+ cancel_delayed_work(&ctx->sched_work);
+ }
+}
+
+void spu_sched_tick(struct work_struct *work)
+{
+ struct spu_context *ctx =
+ container_of(work, struct spu_context, sched_work.work);
+ struct spu *spu;
+ int preempted = 0;
+
+ /*
+ * If this context is being stopped avoid rescheduling from the
+ * scheduler tick because we would block on the state_mutex.
+ * The caller will yield the spu later on anyway.
+ */
+ if (test_bit(SPU_SCHED_EXITING, &ctx->sched_flags))
+ return;
+
+ mutex_lock(&ctx->state_mutex);
+ spu = ctx->spu;
+ if (spu) {
+ int best = sched_find_first_bit(spu_prio->bitmap);
+ if (best <= ctx->prio) {
+ spu_deactivate(ctx);
+ preempted = 1;
+ }
+ }
+ mutex_unlock(&ctx->state_mutex);
+
+ if (preempted) {
+ /*
+ * 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);
+ } else
+ spu_start_tick(ctx);
+}
+
/**
* spu_add_to_active_list - add spu to active list
* @spu: spu to add to the active list
mutex_unlock(&spu_prio->active_mutex[node]);
}
-static inline void mm_needs_global_tlbie(struct mm_struct *mm)
-{
- int nr = (NR_CPUS > 1) ? NR_CPUS : NR_CPUS + 1;
-
- /* Global TLBIE broadcast required with SPEs. */
- __cpus_setall(&mm->cpu_vm_mask, nr);
-}
-
static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
static void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
ctx->spu = spu;
ctx->ops = &spu_hw_ops;
spu->pid = current->pid;
- spu->mm = ctx->owner;
- mm_needs_global_tlbie(spu->mm);
+ spu_associate_mm(spu, ctx->owner);
spu->ibox_callback = spufs_ibox_callback;
spu->wbox_callback = spufs_wbox_callback;
spu->stop_callback = spufs_stop_callback;
spu->stop_callback = NULL;
spu->mfc_callback = NULL;
spu->dma_callback = NULL;
- spu->mm = NULL;
+ spu_associate_mm(spu, NULL);
spu->pid = 0;
ctx->ops = &spu_backing_ops;
ctx->spu = NULL;
* spu_add_to_rq - add a context to the runqueue
* @ctx: context to add
*/
-static void spu_add_to_rq(struct spu_context *ctx)
+static void __spu_add_to_rq(struct spu_context *ctx)
{
- spin_lock(&spu_prio->runq_lock);
- list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]);
- set_bit(ctx->prio, spu_prio->bitmap);
- spin_unlock(&spu_prio->runq_lock);
-}
+ int prio = ctx->prio;
-/**
- * spu_del_from_rq - remove a context from the runqueue
- * @ctx: context to remove
- */
-static void spu_del_from_rq(struct spu_context *ctx)
-{
- spin_lock(&spu_prio->runq_lock);
- list_del_init(&ctx->rq);
- if (list_empty(&spu_prio->runq[ctx->prio]))
- clear_bit(ctx->prio, spu_prio->bitmap);
- spin_unlock(&spu_prio->runq_lock);
+ list_add_tail(&ctx->rq, &spu_prio->runq[prio]);
+ set_bit(prio, spu_prio->bitmap);
}
-/**
- * spu_grab_context - remove one context from the runqueue
- * @prio: priority of the context to be removed
- *
- * This function removes one context from the runqueue for priority @prio.
- * If there is more than one context with the given priority the first
- * task on the runqueue will be taken.
- *
- * Returns the spu_context it just removed.
- *
- * Must be called with spu_prio->runq_lock held.
- */
-static struct spu_context *spu_grab_context(int prio)
+static void __spu_del_from_rq(struct spu_context *ctx)
{
- struct list_head *rq = &spu_prio->runq[prio];
+ int prio = ctx->prio;
- if (list_empty(rq))
- return NULL;
- return list_entry(rq->next, struct spu_context, rq);
+ if (!list_empty(&ctx->rq))
+ list_del_init(&ctx->rq);
+ if (list_empty(&spu_prio->runq[prio]))
+ clear_bit(prio, spu_prio->bitmap);
}
static void spu_prio_wait(struct spu_context *ctx)
{
DEFINE_WAIT(wait);
- set_bit(SPU_SCHED_WAKE, &ctx->sched_flags);
+ spin_lock(&spu_prio->runq_lock);
prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
if (!signal_pending(current)) {
+ __spu_add_to_rq(ctx);
+ spin_unlock(&spu_prio->runq_lock);
mutex_unlock(&ctx->state_mutex);
schedule();
mutex_lock(&ctx->state_mutex);
+ spin_lock(&spu_prio->runq_lock);
+ __spu_del_from_rq(ctx);
}
+ spin_unlock(&spu_prio->runq_lock);
__set_current_state(TASK_RUNNING);
remove_wait_queue(&ctx->stop_wq, &wait);
- clear_bit(SPU_SCHED_WAKE, &ctx->sched_flags);
}
/**
spin_lock(&spu_prio->runq_lock);
best = sched_find_first_bit(spu_prio->bitmap);
if (best < MAX_PRIO) {
- struct spu_context *ctx = spu_grab_context(best);
- if (ctx && test_bit(SPU_SCHED_WAKE, &ctx->sched_flags))
- wake_up(&ctx->stop_wq);
+ struct list_head *rq = &spu_prio->runq[best];
+ struct spu_context *ctx;
+
+ BUG_ON(list_empty(rq));
+
+ ctx = list_entry(rq->next, struct spu_context, rq);
+ __spu_del_from_rq(ctx);
+ wake_up(&ctx->stop_wq);
}
spin_unlock(&spu_prio->runq_lock);
}
}
spu_unbind_context(spu, victim);
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;
}
}
* @ctx: spu context to schedule
* @flags: flags (currently ignored)
*
- * Tries to find a free spu to run @ctx. If no free spu is availble
+ * 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.
*/
return 0;
}
- spu_add_to_rq(ctx);
- if (!(flags & SPU_ACTIVATE_NOWAKE))
- spu_prio_wait(ctx);
- spu_del_from_rq(ctx);
+ spu_prio_wait(ctx);
} while (!signal_pending(current));
return -ERESTARTSYS;
void spu_yield(struct spu_context *ctx)
{
struct spu *spu;
- int need_yield = 0;
if (mutex_trylock(&ctx->state_mutex)) {
if ((spu = ctx->spu) != NULL) {
pr_debug("%s: yielding SPU %d NODE %d\n",
__FUNCTION__, spu->number, spu->node);
spu_deactivate(ctx);
- need_yield = 1;
}
}
mutex_unlock(&ctx->state_mutex);
}
- if (unlikely(need_yield))
- yield();
}
int __init spu_sched_init(void)
{
int i;
+ spu_sched_wq = create_singlethread_workqueue("spusched");
+ if (!spu_sched_wq)
+ return 1;
+
spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
if (!spu_prio) {
printk(KERN_WARNING "%s: Unable to allocate priority queue.\n",
__FUNCTION__);
+ destroy_workqueue(spu_sched_wq);
return 1;
}
for (i = 0; i < MAX_PRIO; i++) {
mutex_unlock(&spu_prio->active_mutex[node]);
}
kfree(spu_prio);
+ destroy_workqueue(spu_sched_wq);
}