DECLARE_BITMAP(bitmap, MAX_PRIO);
struct list_head runq[MAX_PRIO];
spinlock_t runq_lock;
- struct list_head active_list[MAX_NUMNODES];
- struct mutex active_mutex[MAX_NUMNODES];
- int nr_active[MAX_NUMNODES];
int nr_waiting;
};
void __spu_update_sched_info(struct spu_context *ctx)
{
/*
- * 32-Bit assignment are atomic on powerpc, and we don't care about
- * memory ordering here because retriving the controlling thread is
- * per defintion racy.
+ * 32-Bit assignments are atomic on powerpc, and we don't care about
+ * memory ordering here because retrieving the controlling thread is
+ * per definition racy.
*/
ctx->tid = current->pid;
/*
* We do our own priority calculations, so we normally want
- * ->static_prio to start with. Unfortunately thies field
+ * ->static_prio to start with. Unfortunately this field
* contains junk for threads with a realtime scheduling
* policy so we have to look at ->prio in this case.
*/
ctx->policy = current->policy;
/*
- * A lot of places that don't hold active_mutex poke into
+ * A lot of places that don't hold list_mutex poke into
* cpus_allowed, including grab_runnable_context which
* already holds the runq_lock. So abuse runq_lock
- * to protect this field aswell.
+ * to protect this field as well.
*/
spin_lock(&spu_prio->runq_lock);
ctx->cpus_allowed = current->cpus_allowed;
{
int node = ctx->spu->node;
- mutex_lock(&spu_prio->active_mutex[node]);
+ mutex_lock(&cbe_spu_info[node].list_mutex);
__spu_update_sched_info(ctx);
- mutex_unlock(&spu_prio->active_mutex[node]);
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
}
static int __node_allowed(struct spu_context *ctx, int node)
return rval;
}
-/**
- * spu_add_to_active_list - add spu to active list
- * @spu: spu to add to the active list
- */
-static void spu_add_to_active_list(struct spu *spu)
-{
- int node = spu->node;
-
- mutex_lock(&spu_prio->active_mutex[node]);
- spu_prio->nr_active[node]++;
- list_add_tail(&spu->list, &spu_prio->active_list[node]);
- mutex_unlock(&spu_prio->active_mutex[node]);
-}
+static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
-static void __spu_remove_from_active_list(struct spu *spu)
+void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
{
- list_del_init(&spu->list);
- spu_prio->nr_active[spu->node]--;
+ blocking_notifier_call_chain(&spu_switch_notifier,
+ ctx ? ctx->object_id : 0, spu);
}
-/**
- * spu_remove_from_active_list - remove spu from active list
- * @spu: spu to remove from the active list
- */
-static void spu_remove_from_active_list(struct spu *spu)
+static void notify_spus_active(void)
{
- int node = spu->node;
-
- mutex_lock(&spu_prio->active_mutex[node]);
- __spu_remove_from_active_list(spu);
- mutex_unlock(&spu_prio->active_mutex[node]);
-}
+ int node;
-static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
+ /*
+ * Wake up the active spu_contexts.
+ *
+ * When the awakened processes see their "notify_active" flag is set,
+ * they will call spu_switch_notify().
+ */
+ for_each_online_node(node) {
+ struct spu *spu;
-static void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
-{
- blocking_notifier_call_chain(&spu_switch_notifier,
- ctx ? ctx->object_id : 0, spu);
+ 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) {
+ struct spu_context *ctx = spu->ctx;
+ set_bit(SPU_SCHED_NOTIFY_ACTIVE,
+ &ctx->sched_flags);
+ mb();
+ wake_up_all(&ctx->stop_wq);
+ }
+ }
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
+ }
}
int spu_switch_event_register(struct notifier_block * n)
{
- return blocking_notifier_chain_register(&spu_switch_notifier, n);
+ int ret;
+ ret = blocking_notifier_chain_register(&spu_switch_notifier, n);
+ if (!ret)
+ notify_spus_active();
+ return ret;
}
+EXPORT_SYMBOL_GPL(spu_switch_event_register);
int spu_switch_event_unregister(struct notifier_block * n)
{
return blocking_notifier_chain_unregister(&spu_switch_notifier, n);
}
+EXPORT_SYMBOL_GPL(spu_switch_event_unregister);
/**
* spu_bind_context - bind spu context to physical spu
if (ctx->flags & SPU_CREATE_NOSCHED)
atomic_inc(&cbe_spu_info[spu->node].reserved_spus);
- if (!list_empty(&ctx->aff_list))
- atomic_inc(&ctx->gang->aff_sched_count);
ctx->stats.slb_flt_base = spu->stats.slb_flt;
ctx->stats.class2_intr_base = spu->stats.class2_intr;
ctx->spu = spu;
ctx->ops = &spu_hw_ops;
spu->pid = current->pid;
+ spu->tgid = current->tgid;
spu_associate_mm(spu, ctx->owner);
spu->ibox_callback = spufs_ibox_callback;
spu->wbox_callback = spufs_wbox_callback;
}
/*
- * XXX(hch): needs locking.
+ * Must be used with the list_mutex held.
*/
static inline int sched_spu(struct spu *spu)
{
+ BUG_ON(!mutex_is_locked(&cbe_spu_info[spu->node].list_mutex));
+
return (!spu->ctx || !(spu->ctx->flags & SPU_CREATE_NOSCHED));
}
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 ((!mem_aff || spu->has_mem_affinity) &&
- sched_spu(spu))
+ sched_spu(spu)) {
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
return spu;
+ }
}
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
}
return NULL;
}
lowest_offset = ctx->aff_offset;
}
- gang->aff_ref_spu = aff_ref_location(ctx, mem_aff, gs, lowest_offset);
+ gang->aff_ref_spu = aff_ref_location(gang->aff_ref_ctx, mem_aff, gs,
+ lowest_offset);
}
-static struct spu *ctx_location(struct spu *ref, int offset)
+static struct spu *ctx_location(struct spu *ref, int offset, int node)
{
struct spu *spu;
spu = NULL;
if (offset >= 0) {
list_for_each_entry(spu, ref->aff_list.prev, aff_list) {
+ BUG_ON(spu->node != node);
if (offset == 0)
break;
if (sched_spu(spu))
}
} else {
list_for_each_entry_reverse(spu, ref->aff_list.next, aff_list) {
+ BUG_ON(spu->node != node);
if (offset == 0)
break;
if (sched_spu(spu))
offset++;
}
}
+
return spu;
}
* affinity_check is called each time a context is going to be scheduled.
* It returns the spu ptr on which the context must run.
*/
-struct spu *affinity_check(struct spu_context *ctx)
+static int has_affinity(struct spu_context *ctx)
{
- struct spu_gang *gang;
+ struct spu_gang *gang = ctx->gang;
if (list_empty(&ctx->aff_list))
- return NULL;
- gang = ctx->gang;
- mutex_lock(&gang->aff_mutex);
+ return 0;
+
if (!gang->aff_ref_spu) {
if (!(gang->aff_flags & AFF_MERGED))
aff_merge_remaining_ctxs(gang);
aff_set_offsets(gang);
aff_set_ref_point_location(gang);
}
- mutex_unlock(&gang->aff_mutex);
- if (!gang->aff_ref_spu)
- return NULL;
- return ctx_location(gang->aff_ref_spu, ctx->aff_offset);
+
+ return gang->aff_ref_spu != NULL;
}
/**
if (spu->ctx->flags & SPU_CREATE_NOSCHED)
atomic_dec(&cbe_spu_info[spu->node].reserved_spus);
- if (!list_empty(&ctx->aff_list))
- if (atomic_dec_and_test(&ctx->gang->aff_sched_count))
- ctx->gang->aff_ref_spu = NULL;
+
+ if (ctx->gang){
+ mutex_lock(&ctx->gang->aff_mutex);
+ if (has_affinity(ctx)) {
+ if (atomic_dec_and_test(&ctx->gang->aff_sched_count))
+ ctx->gang->aff_ref_spu = NULL;
+ }
+ mutex_unlock(&ctx->gang->aff_mutex);
+ }
+
spu_switch_notify(spu, NULL);
spu_unmap_mappings(ctx);
spu_save(&ctx->csa, spu);
spu->dma_callback = NULL;
spu_associate_mm(spu, NULL);
spu->pid = 0;
+ spu->tgid = 0;
ctx->ops = &spu_backing_ops;
spu->flags = 0;
spu->ctx = NULL;
static struct spu *spu_get_idle(struct spu_context *ctx)
{
- struct spu *spu = NULL;
- int node = cpu_to_node(raw_smp_processor_id());
- int n;
+ 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;
- spu = spu_alloc_node(node);
- if (spu)
- break;
+
+ 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;
}
/*
* 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
+ * exactly fair, but so far the whole spu scheduler tries to keep
* a strong node affinity. We might want to fine-tune this in
* the future.
*/
if (!node_allowed(ctx, node))
continue;
- mutex_lock(&spu_prio->active_mutex[node]);
- list_for_each_entry(spu, &spu_prio->active_list[node], list) {
+ 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->prio > ctx->prio &&
+ if (tmp && tmp->prio > ctx->prio &&
(!victim || tmp->prio > victim->prio))
victim = spu->ctx;
}
- mutex_unlock(&spu_prio->active_mutex[node]);
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
if (victim) {
/*
victim = NULL;
goto restart;
}
- spu_remove_from_active_list(spu);
+
+ 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);
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);
- spu_add_to_active_list(spu);
+ cbe_spu_info[node].nr_active++;
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
return 0;
}
if (spu) {
new = grab_runnable_context(max_prio, spu->node);
if (new || force) {
- spu_remove_from_active_list(spu);
+ 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++;
- spu_free(spu);
+
if (new)
wake_up(&new->stop_wq);
}
}
/**
- * spu_yield - yield a physical spu if others are waiting
+ * 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
}
}
-static void spusched_tick(struct spu_context *ctx)
+static noinline void spusched_tick(struct spu_context *ctx)
{
if (ctx->flags & SPU_CREATE_NOSCHED)
return;
return;
/*
- * Unfortunately active_mutex ranks outside of state_mutex, so
+ * 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.
*/
new = grab_runnable_context(ctx->prio + 1, spu->node);
if (new) {
-
- __spu_remove_from_active_list(spu);
spu_unbind_context(spu, ctx);
ctx->stats.invol_ctx_switch++;
spu->stats.invol_ctx_switch++;
- spu_free(spu);
+ 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
*
* Return the number of tasks currently running or waiting to run.
*
- * Note that we don't take runq_lock / active_mutex here. Reading
+ * 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.
*/
int nr_active = 0, node;
for (node = 0; node < MAX_NUMNODES; node++)
- nr_active += spu_prio->nr_active[node];
+ nr_active += cbe_spu_info[node].nr_active;
nr_active += spu_prio->nr_waiting;
return nr_active;
static int spusched_thread(void *unused)
{
- struct spu *spu, *next;
+ struct spu *spu;
int node;
while (!kthread_should_stop()) {
set_current_state(TASK_INTERRUPTIBLE);
schedule();
for (node = 0; node < MAX_NUMNODES; node++) {
- mutex_lock(&spu_prio->active_mutex[node]);
- list_for_each_entry_safe(spu, next,
- &spu_prio->active_list[node],
- list)
- spusched_tick(spu->ctx);
- mutex_unlock(&spu_prio->active_mutex[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;
}
+void spuctx_switch_state(struct spu_context *ctx,
+ enum spu_utilization_state new_state)
+{
+ unsigned long long curtime;
+ signed long long delta;
+ struct timespec ts;
+ struct spu *spu;
+ enum spu_utilization_state old_state;
+
+ ktime_get_ts(&ts);
+ curtime = timespec_to_ns(&ts);
+ delta = curtime - ctx->stats.tstamp;
+
+ WARN_ON(!mutex_is_locked(&ctx->state_mutex));
+ WARN_ON(delta < 0);
+
+ spu = ctx->spu;
+ old_state = ctx->stats.util_state;
+ ctx->stats.util_state = new_state;
+ ctx->stats.tstamp = curtime;
+
+ /*
+ * Update the physical SPU utilization statistics.
+ */
+ if (spu) {
+ ctx->stats.times[old_state] += delta;
+ spu->stats.times[old_state] += delta;
+ spu->stats.util_state = new_state;
+ spu->stats.tstamp = curtime;
+ }
+}
+
#define LOAD_INT(x) ((x) >> FSHIFT)
#define LOAD_FRAC(x) LOAD_INT(((x) & (FIXED_1-1)) * 100)
/*
* Note that last_pid doesn't really make much sense for the
- * SPU loadavg (it even seems very odd on the CPU side..),
+ * 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",
INIT_LIST_HEAD(&spu_prio->runq[i]);
__clear_bit(i, spu_prio->bitmap);
}
- for (i = 0; i < MAX_NUMNODES; i++) {
- mutex_init(&spu_prio->active_mutex[i]);
- INIT_LIST_HEAD(&spu_prio->active_list[i]);
- }
spin_lock_init(&spu_prio->runq_lock);
setup_timer(&spusched_timer, spusched_wake, 0);
void spu_sched_exit(void)
{
- struct spu *spu, *tmp;
+ struct spu *spu;
int node;
remove_proc_entry("spu_loadavg", NULL);
kthread_stop(spusched_task);
for (node = 0; node < MAX_NUMNODES; node++) {
- mutex_lock(&spu_prio->active_mutex[node]);
- list_for_each_entry_safe(spu, tmp, &spu_prio->active_list[node],
- list) {
- list_del_init(&spu->list);
- spu_free(spu);
- }
- mutex_unlock(&spu_prio->active_mutex[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);
}