#include <linux/sched.h>
#include <linux/kernel.h>
#include <linux/mm.h>
+#include <linux/slab.h>
#include <linux/completion.h>
#include <linux/vmalloc.h>
#include <linux/smp.h>
#include <linux/mutex.h>
#include <linux/notifier.h>
#include <linux/kthread.h>
+#include <linux/pid_namespace.h>
+#include <linux/proc_fs.h>
+#include <linux/seq_file.h>
#include <asm/io.h>
#include <asm/mmu_context.h>
#include <asm/spu_csa.h>
#include <asm/spu_priv1.h>
#include "spufs.h"
+#define CREATE_TRACE_POINTS
+#include "sputrace.h"
struct spu_prio_array {
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_waiting;
};
+static unsigned long spu_avenrun[3];
static struct spu_prio_array *spu_prio;
static struct task_struct *spusched_task;
static struct timer_list spusched_timer;
+static struct timer_list spuloadavg_timer;
/*
* Priority of a normal, non-rt, non-niced'd process (aka nice level 0).
ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio);
}
-static inline int node_allowed(int node)
+/*
+ * Update scheduling information from the owning thread.
+ */
+void __spu_update_sched_info(struct spu_context *ctx)
{
- cpumask_t mask;
+ /*
+ * assert that the context is not on the runqueue, so it is safe
+ * to change its scheduling parameters.
+ */
+ BUG_ON(!list_empty(&ctx->rq));
- if (!nr_cpus_node(node))
- return 0;
- mask = node_to_cpumask(node);
- if (!cpus_intersects(mask, current->cpus_allowed))
- return 0;
- return 1;
-}
+ /*
+ * 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;
-/**
- * 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)
-{
- mutex_lock(&spu_prio->active_mutex[spu->node]);
- list_add_tail(&spu->list, &spu_prio->active_list[spu->node]);
- mutex_unlock(&spu_prio->active_mutex[spu->node]);
-}
+ /*
+ * We do our own priority calculations, so we normally want
+ * ->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.
+ */
+ if (rt_prio(current->prio))
+ ctx->prio = current->prio;
+ else
+ ctx->prio = current->static_prio;
+ ctx->policy = current->policy;
-static void __spu_remove_from_active_list(struct spu *spu)
-{
- list_del_init(&spu->list);
+ /*
+ * TO DO: the context may be loaded, so we may need to activate
+ * it again on a different node. But it shouldn't hurt anything
+ * to update its parameters, because we know that the scheduler
+ * is not actively looking at this field, since it is not on the
+ * runqueue. The context will be rescheduled on the proper node
+ * if it is timesliced or preempted.
+ */
+ ctx->cpus_allowed = current->cpus_allowed;
+
+ /* Save the current cpu id for spu interrupt routing. */
+ ctx->last_ran = raw_smp_processor_id();
}
-/**
- * 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)
+void spu_update_sched_info(struct spu_context *ctx)
{
- int node = spu->node;
+ int node;
- mutex_lock(&spu_prio->active_mutex[node]);
- __spu_remove_from_active_list(spu);
- mutex_unlock(&spu_prio->active_mutex[node]);
-}
+ if (ctx->state == SPU_STATE_RUNNABLE) {
+ node = ctx->spu->node;
-static BLOCKING_NOTIFIER_HEAD(spu_switch_notifier);
+ /*
+ * Take list_mutex to sync with find_victim().
+ */
+ mutex_lock(&cbe_spu_info[node].list_mutex);
+ __spu_update_sched_info(ctx);
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
+ } else {
+ __spu_update_sched_info(ctx);
+ }
+}
-static void spu_switch_notify(struct spu *spu, struct spu_context *ctx)
+static int __node_allowed(struct spu_context *ctx, int node)
{
- blocking_notifier_call_chain(&spu_switch_notifier,
- ctx ? ctx->object_id : 0, spu);
+ if (nr_cpus_node(node)) {
+ const struct cpumask *mask = cpumask_of_node(node);
+
+ if (cpumask_intersects(mask, &ctx->cpus_allowed))
+ return 1;
+ }
+
+ return 0;
}
-int spu_switch_event_register(struct notifier_block * n)
+static int node_allowed(struct spu_context *ctx, int node)
{
- return blocking_notifier_chain_register(&spu_switch_notifier, n);
+ int rval;
+
+ spin_lock(&spu_prio->runq_lock);
+ rval = __node_allowed(ctx, node);
+ spin_unlock(&spu_prio->runq_lock);
+
+ return rval;
}
-int spu_switch_event_unregister(struct notifier_block * n)
+void do_notify_spus_active(void)
{
- return blocking_notifier_chain_unregister(&spu_switch_notifier, n);
+ int node;
+
+ /*
+ * 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;
+
+ 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);
+ }
}
/**
*/
static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
{
- pr_debug("%s: pid=%d SPU=%d NODE=%d\n", __FUNCTION__, current->pid,
- spu->number, spu->node);
+ spu_context_trace(spu_bind_context__enter, ctx, spu);
+
+ spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
+
+ if (ctx->flags & SPU_CREATE_NOSCHED)
+ atomic_inc(&cbe_spu_info[spu->node].reserved_spus);
+
+ ctx->stats.slb_flt_base = spu->stats.slb_flt;
+ ctx->stats.class2_intr_base = spu->stats.class2_intr;
+
+ spu_associate_mm(spu, ctx->owner);
+
+ spin_lock_irq(&spu->register_lock);
spu->ctx = ctx;
spu->flags = 0;
ctx->spu = spu;
ctx->ops = &spu_hw_ops;
spu->pid = current->pid;
- spu_associate_mm(spu, ctx->owner);
+ spu->tgid = current->tgid;
spu->ibox_callback = spufs_ibox_callback;
spu->wbox_callback = spufs_wbox_callback;
spu->stop_callback = spufs_stop_callback;
spu->mfc_callback = spufs_mfc_callback;
- spu->dma_callback = spufs_dma_callback;
- mb();
+ spin_unlock_irq(&spu->register_lock);
+
spu_unmap_mappings(ctx);
+
+ spu_switch_log_notify(spu, ctx, SWITCH_LOG_START, 0);
spu_restore(&ctx->csa, spu);
spu->timestamp = jiffies;
- spu_cpu_affinity_set(spu, raw_smp_processor_id());
spu_switch_notify(spu, ctx);
ctx->state = SPU_STATE_RUNNABLE;
+
+ spuctx_switch_state(ctx, SPU_UTIL_USER);
+}
+
+/*
+ * 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));
+}
+
+static void aff_merge_remaining_ctxs(struct spu_gang *gang)
+{
+ struct spu_context *ctx;
+
+ list_for_each_entry(ctx, &gang->aff_list_head, aff_list) {
+ if (list_empty(&ctx->aff_list))
+ list_add(&ctx->aff_list, &gang->aff_list_head);
+ }
+ gang->aff_flags |= AFF_MERGED;
+}
+
+static void aff_set_offsets(struct spu_gang *gang)
+{
+ struct spu_context *ctx;
+ int offset;
+
+ offset = -1;
+ list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
+ aff_list) {
+ if (&ctx->aff_list == &gang->aff_list_head)
+ break;
+ ctx->aff_offset = offset--;
+ }
+
+ offset = 0;
+ list_for_each_entry(ctx, gang->aff_ref_ctx->aff_list.prev, aff_list) {
+ if (&ctx->aff_list == &gang->aff_list_head)
+ break;
+ ctx->aff_offset = offset++;
+ }
+
+ gang->aff_flags |= AFF_OFFSETS_SET;
+}
+
+static struct spu *aff_ref_location(struct spu_context *ctx, int mem_aff,
+ int group_size, int lowest_offset)
+{
+ struct spu *spu;
+ int node, n;
+
+ /*
+ * TODO: A better algorithm could be used to find a good spu to be
+ * used as reference location for the ctxs chain.
+ */
+ node = cpu_to_node(raw_smp_processor_id());
+ for (n = 0; n < MAX_NUMNODES; n++, node++) {
+ /*
+ * "available_spus" counts how many spus are not potentially
+ * going to be used by other affinity gangs whose reference
+ * context is already in place. Although this code seeks to
+ * avoid having affinity gangs with a summed amount of
+ * contexts bigger than the amount of spus in the node,
+ * this may happen sporadically. In this case, available_spus
+ * becomes negative, which is harmless.
+ */
+ int available_spus;
+
+ node = (node < MAX_NUMNODES) ? node : 0;
+ if (!node_allowed(ctx, node))
+ continue;
+
+ available_spus = 0;
+ mutex_lock(&cbe_spu_info[node].list_mutex);
+ list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
+ if (spu->ctx && spu->ctx->gang && !spu->ctx->aff_offset
+ && spu->ctx->gang->aff_ref_spu)
+ available_spus -= spu->ctx->gang->contexts;
+ available_spus++;
+ }
+ if (available_spus < ctx->gang->contexts) {
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
+ continue;
+ }
+
+ list_for_each_entry(spu, &cbe_spu_info[node].spus, cbe_list) {
+ if ((!mem_aff || spu->has_mem_affinity) &&
+ sched_spu(spu)) {
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
+ return spu;
+ }
+ }
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
+ }
+ return NULL;
+}
+
+static void aff_set_ref_point_location(struct spu_gang *gang)
+{
+ int mem_aff, gs, lowest_offset;
+ struct spu_context *ctx;
+ struct spu *tmp;
+
+ mem_aff = gang->aff_ref_ctx->flags & SPU_CREATE_AFFINITY_MEM;
+ lowest_offset = 0;
+ gs = 0;
+
+ list_for_each_entry(tmp, &gang->aff_list_head, aff_list)
+ gs++;
+
+ list_for_each_entry_reverse(ctx, &gang->aff_ref_ctx->aff_list,
+ aff_list) {
+ if (&ctx->aff_list == &gang->aff_list_head)
+ break;
+ lowest_offset = ctx->aff_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, 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))
+ offset--;
+ }
+ } 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.
+ */
+static int has_affinity(struct spu_context *ctx)
+{
+ struct spu_gang *gang = ctx->gang;
+
+ if (list_empty(&ctx->aff_list))
+ return 0;
+
+ if (atomic_read(&ctx->gang->aff_sched_count) == 0)
+ ctx->gang->aff_ref_spu = NULL;
+
+ if (!gang->aff_ref_spu) {
+ if (!(gang->aff_flags & AFF_MERGED))
+ aff_merge_remaining_ctxs(gang);
+ if (!(gang->aff_flags & AFF_OFFSETS_SET))
+ aff_set_offsets(gang);
+ aff_set_ref_point_location(gang);
+ }
+
+ return gang->aff_ref_spu != NULL;
}
/**
*/
static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
{
- pr_debug("%s: unbind pid=%d SPU=%d NODE=%d\n", __FUNCTION__,
- spu->pid, spu->number, spu->node);
+ u32 status;
+
+ spu_context_trace(spu_unbind_context__enter, ctx, spu);
+
+ spuctx_switch_state(ctx, SPU_UTIL_SYSTEM);
+
+ if (spu->ctx->flags & SPU_CREATE_NOSCHED)
+ atomic_dec(&cbe_spu_info[spu->node].reserved_spus);
+
+ if (ctx->gang)
+ /*
+ * If ctx->gang->aff_sched_count is positive, SPU affinity is
+ * being considered in this gang. Using atomic_dec_if_positive
+ * allow us to skip an explicit check for affinity in this gang
+ */
+ atomic_dec_if_positive(&ctx->gang->aff_sched_count);
spu_switch_notify(spu, NULL);
spu_unmap_mappings(ctx);
spu_save(&ctx->csa, spu);
+ spu_switch_log_notify(spu, ctx, SWITCH_LOG_STOP, 0);
+
+ spin_lock_irq(&spu->register_lock);
spu->timestamp = jiffies;
ctx->state = SPU_STATE_SAVED;
spu->ibox_callback = NULL;
spu->wbox_callback = NULL;
spu->stop_callback = NULL;
spu->mfc_callback = NULL;
- spu->dma_callback = NULL;
- spu_associate_mm(spu, NULL);
spu->pid = 0;
+ spu->tgid = 0;
ctx->ops = &spu_backing_ops;
- ctx->spu = NULL;
spu->flags = 0;
spu->ctx = NULL;
+ spin_unlock_irq(&spu->register_lock);
+
+ spu_associate_mm(spu, NULL);
+
+ ctx->stats.slb_flt +=
+ (spu->stats.slb_flt - ctx->stats.slb_flt_base);
+ ctx->stats.class2_intr +=
+ (spu->stats.class2_intr - ctx->stats.class2_intr_base);
+
+ /* This maps the underlying spu state to idle */
+ spuctx_switch_state(ctx, SPU_UTIL_IDLE_LOADED);
+ ctx->spu = NULL;
+
+ if (spu_stopped(ctx, &status))
+ wake_up_all(&ctx->stop_wq);
}
/**
*/
static void __spu_add_to_rq(struct spu_context *ctx)
{
- int prio = ctx->prio;
+ /*
+ * Unfortunately this code path can be called from multiple threads
+ * on behalf of a single context due to the way the problem state
+ * mmap support works.
+ *
+ * Fortunately we need to wake up all these threads at the same time
+ * and can simply skip the runqueue addition for every but the first
+ * thread getting into this codepath.
+ *
+ * It's still quite hacky, and long-term we should proxy all other
+ * threads through the owner thread so that spu_run is in control
+ * of all the scheduling activity for a given context.
+ */
+ if (list_empty(&ctx->rq)) {
+ list_add_tail(&ctx->rq, &spu_prio->runq[ctx->prio]);
+ set_bit(ctx->prio, spu_prio->bitmap);
+ if (!spu_prio->nr_waiting++)
+ mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
+ }
+}
- list_add_tail(&ctx->rq, &spu_prio->runq[prio]);
- set_bit(prio, spu_prio->bitmap);
+static void spu_add_to_rq(struct spu_context *ctx)
+{
+ spin_lock(&spu_prio->runq_lock);
+ __spu_add_to_rq(ctx);
+ spin_unlock(&spu_prio->runq_lock);
}
static void __spu_del_from_rq(struct spu_context *ctx)
{
int prio = ctx->prio;
- if (!list_empty(&ctx->rq))
+ if (!list_empty(&ctx->rq)) {
+ if (!--spu_prio->nr_waiting)
+ del_timer(&spusched_timer);
list_del_init(&ctx->rq);
- if (list_empty(&spu_prio->runq[prio]))
- clear_bit(prio, spu_prio->bitmap);
+
+ if (list_empty(&spu_prio->runq[prio]))
+ clear_bit(prio, spu_prio->bitmap);
+ }
+}
+
+void spu_del_from_rq(struct spu_context *ctx)
+{
+ spin_lock(&spu_prio->runq_lock);
+ __spu_del_from_rq(ctx);
+ spin_unlock(&spu_prio->runq_lock);
}
static void spu_prio_wait(struct spu_context *ctx)
{
DEFINE_WAIT(wait);
+ /*
+ * The caller must explicitly wait for a context to be loaded
+ * if the nosched flag is set. If NOSCHED is not set, the caller
+ * queues the context and waits for an spu event or error.
+ */
+ BUG_ON(!(ctx->flags & SPU_CREATE_NOSCHED));
+
spin_lock(&spu_prio->runq_lock);
prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
if (!signal_pending(current)) {
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;
+
+ spu_context_nospu_trace(spu_get_idle__enter, ctx);
+
+ 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);
+
+ atomic_dec(&ctx->gang->aff_sched_count);
+ goto not_found;
+ }
+ 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(node))
+ 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);
}
+
+ not_found:
+ spu_context_nospu_trace(spu_get_idle__not_found, ctx);
+ return NULL;
+
+ found:
+ spu->alloc_state = SPU_USED;
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
+ spu_context_trace(spu_get_idle__found, ctx, spu);
+ spu_init_channels(spu);
return spu;
}
struct spu *spu;
int node, n;
+ spu_context_nospu_trace(spu_find_victim__enter, ctx);
+
/*
* 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.
*/
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(node))
+ 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 &&
- (!victim || tmp->prio > victim->prio))
+ if (tmp && tmp->prio > ctx->prio &&
+ !(tmp->flags & SPU_CREATE_NOSCHED) &&
+ (!victim || tmp->prio > victim->prio)) {
victim = spu->ctx;
+ }
}
- mutex_unlock(&spu_prio->active_mutex[node]);
+ if (victim)
+ get_spu_context(victim);
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
if (victim) {
/*
* higher priority contexts before lower priority
* ones, so this is safe until we introduce
* priority inheritance schemes.
+ *
+ * XXX if the highest priority context is locked,
+ * this can loop a long time. Might be better to
+ * look at another context or give up after X retries.
*/
if (!mutex_trylock(&victim->state_mutex)) {
+ put_spu_context(victim);
victim = NULL;
goto restart;
}
spu = victim->spu;
- if (!spu) {
+ if (!spu || victim->prio <= ctx->prio) {
/*
* This race can happen because we've dropped
- * the active list mutex. No a problem, just
+ * the active list mutex. Not a problem, just
* restart the search.
*/
mutex_unlock(&victim->state_mutex);
+ put_spu_context(victim);
victim = NULL;
goto restart;
}
- spu_remove_from_active_list(spu);
+
+ spu_context_trace(__spu_deactivate__unload, ctx, 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++;
+ if (test_bit(SPU_SCHED_SPU_RUN, &victim->sched_flags))
+ spu_add_to_rq(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);
+ put_spu_context(victim);
+
return spu;
}
}
return NULL;
}
+static void __spu_schedule(struct spu *spu, struct spu_context *ctx)
+{
+ int node = spu->node;
+ int success = 0;
+
+ spu_set_timeslice(ctx);
+
+ mutex_lock(&cbe_spu_info[node].list_mutex);
+ if (spu->ctx == NULL) {
+ spu_bind_context(spu, ctx);
+ cbe_spu_info[node].nr_active++;
+ spu->alloc_state = SPU_USED;
+ success = 1;
+ }
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
+
+ if (success)
+ wake_up_all(&ctx->run_wq);
+ else
+ spu_add_to_rq(ctx);
+}
+
+static void spu_schedule(struct spu *spu, struct spu_context *ctx)
+{
+ /* not a candidate for interruptible because it's called either
+ from the scheduler thread or from spu_deactivate */
+ mutex_lock(&ctx->state_mutex);
+ if (ctx->state == SPU_STATE_SAVED)
+ __spu_schedule(spu, ctx);
+ spu_release(ctx);
+}
+
+/**
+ * spu_unschedule - remove a context from a spu, and possibly release it.
+ * @spu: The SPU to unschedule from
+ * @ctx: The context currently scheduled on the SPU
+ * @free_spu Whether to free the SPU for other contexts
+ *
+ * Unbinds the context @ctx from the SPU @spu. If @free_spu is non-zero, the
+ * SPU is made available for other contexts (ie, may be returned by
+ * spu_get_idle). If this is zero, the caller is expected to schedule another
+ * context to this spu.
+ *
+ * Should be called with ctx->state_mutex held.
+ */
+static void spu_unschedule(struct spu *spu, struct spu_context *ctx,
+ int free_spu)
+{
+ int node = spu->node;
+
+ mutex_lock(&cbe_spu_info[node].list_mutex);
+ cbe_spu_info[node].nr_active--;
+ if (free_spu)
+ spu->alloc_state = SPU_FREE;
+ spu_unbind_context(spu, ctx);
+ ctx->stats.invol_ctx_switch++;
+ spu->stats.invol_ctx_switch++;
+ mutex_unlock(&cbe_spu_info[node].list_mutex);
+}
+
/**
* spu_activate - find a free spu for a context and execute it
* @ctx: spu context to schedule
*/
int spu_activate(struct spu_context *ctx, unsigned long flags)
{
+ 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;
- do {
- struct spu *spu;
+spu_activate_top:
+ if (signal_pending(current))
+ return -ERESTARTSYS;
- 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) {
- spu_bind_context(spu, ctx);
- spu_add_to_active_list(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) {
+ unsigned long runcntl;
+
+ runcntl = ctx->ops->runcntl_read(ctx);
+ __spu_schedule(spu, ctx);
+ if (runcntl & SPU_RUNCNTL_RUNNABLE)
+ spuctx_switch_state(ctx, SPU_UTIL_USER);
+
+ return 0;
+ }
+ if (ctx->flags & SPU_CREATE_NOSCHED) {
spu_prio_wait(ctx);
- } while (!signal_pending(current));
+ goto spu_activate_top;
+ }
+
+ spu_add_to_rq(ctx);
- return -ERESTARTSYS;
+ return 0;
}
/**
* 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)
+static struct spu_context *grab_runnable_context(int prio, int node)
{
- struct spu_context *ctx = NULL;
+ struct spu_context *ctx;
int best;
spin_lock(&spu_prio->runq_lock);
- best = sched_find_first_bit(spu_prio->bitmap);
- if (best < prio) {
+ best = find_first_bit(spu_prio->bitmap, prio);
+ while (best < prio) {
struct list_head *rq = &spu_prio->runq[best];
- BUG_ON(list_empty(rq));
-
- ctx = list_entry(rq->next, struct spu_context, rq);
- __spu_del_from_rq(ctx);
+ 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;
}
struct spu_context *new = NULL;
if (spu) {
- new = grab_runnable_context(max_prio);
+ new = grab_runnable_context(max_prio, spu->node);
if (new || force) {
- spu_remove_from_active_list(spu);
- spu_unbind_context(spu, ctx);
- spu_free(spu);
- if (new)
- wake_up(&new->stop_wq);
+ spu_unschedule(spu, ctx, new == NULL);
+ if (new) {
+ if (new->flags & SPU_CREATE_NOSCHED)
+ wake_up(&new->stop_wq);
+ else {
+ spu_release(ctx);
+ spu_schedule(spu, new);
+ /* this one can't easily be made
+ interruptible */
+ mutex_lock(&ctx->state_mutex);
+ }
+ }
}
-
}
return new != NULL;
*/
void spu_deactivate(struct spu_context *ctx)
{
+ spu_context_nospu_trace(spu_deactivate__enter, ctx);
__spu_deactivate(ctx, 1, MAX_PRIO);
}
/**
- * 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
*/
void spu_yield(struct spu_context *ctx)
{
+ spu_context_nospu_trace(spu_yield__enter, ctx);
if (!(ctx->flags & SPU_CREATE_NOSCHED)) {
mutex_lock(&ctx->state_mutex);
__spu_deactivate(ctx, 0, MAX_PRIO);
}
}
-static void spusched_tick(struct spu_context *ctx)
+static noinline void spusched_tick(struct spu_context *ctx)
{
- if (ctx->policy == SCHED_FIFO || --ctx->time_slice)
- return;
+ struct spu_context *new = NULL;
+ struct spu *spu = NULL;
- /*
- * Unfortunately active_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_context *new = grab_runnable_context(ctx->prio + 1);
- if (new) {
- struct spu *spu = ctx->spu;
-
- __spu_remove_from_active_list(spu);
- spu_unbind_context(spu, ctx);
- spu_free(spu);
- 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);
+ if (spu_acquire(ctx))
+ BUG(); /* a kernel thread never has signals pending */
+
+ if (ctx->state != SPU_STATE_RUNNABLE)
+ goto out;
+ if (ctx->flags & SPU_CREATE_NOSCHED)
+ goto out;
+ if (ctx->policy == SCHED_FIFO)
+ goto out;
+
+ if (--ctx->time_slice && test_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags))
+ goto out;
+
+ spu = ctx->spu;
+
+ spu_context_trace(spusched_tick__preempt, ctx, spu);
+
+ new = grab_runnable_context(ctx->prio + 1, spu->node);
+ if (new) {
+ spu_unschedule(spu, ctx, 0);
+ if (test_bit(SPU_SCHED_SPU_RUN, &ctx->sched_flags))
+ spu_add_to_rq(ctx);
} else {
- ctx->time_slice++;
+ spu_context_nospu_trace(spusched_tick__newslice, ctx);
+ if (!ctx->time_slice)
+ ctx->time_slice++;
}
+out:
+ spu_release(ctx);
+
+ if (new)
+ spu_schedule(spu, new);
+}
+
+/**
+ * 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 - update the avenrun load estimates.
+ *
+ * No locking against reading these values from userspace, as for
+ * the CPU loadavg code.
+ */
+static void spu_calc_load(void)
+{
+ unsigned long active_tasks; /* fixed-point */
+
+ active_tasks = count_active_contexts() * FIXED_1;
+ 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);
}
static void spusched_wake(unsigned long data)
wake_up_process(spusched_task);
}
+static void spuloadavg_wake(unsigned long data)
+{
+ mod_timer(&spuloadavg_timer, jiffies + LOAD_FREQ);
+ spu_calc_load();
+}
+
static int spusched_thread(void *unused)
{
- struct spu *spu, *next;
+ struct spu *spu;
int node;
- setup_timer(&spusched_timer, spusched_wake, 0);
- __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
-
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]);
+ struct mutex *mtx = &cbe_spu_info[node].list_mutex;
+
+ mutex_lock(mtx);
+ list_for_each_entry(spu, &cbe_spu_info[node].spus,
+ cbe_list) {
+ struct spu_context *ctx = spu->ctx;
+
+ if (ctx) {
+ get_spu_context(ctx);
+ mutex_unlock(mtx);
+ spusched_tick(ctx);
+ mutex_lock(mtx);
+ put_spu_context(ctx);
+ }
+ }
+ mutex_unlock(mtx);
}
}
- del_timer_sync(&spusched_timer);
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;
+ int node;
+
+ 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;
+ node = spu->node;
+ if (old_state == SPU_UTIL_USER)
+ atomic_dec(&cbe_spu_info[node].busy_spus);
+ if (new_state == SPU_UTIL_USER)
+ atomic_inc(&cbe_spu_info[node].busy_spus);
+ }
+}
+
+#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)
{
- int i;
+ struct proc_dir_entry *entry;
+ int err = -ENOMEM, i;
spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
if (!spu_prio)
- return -ENOMEM;
+ goto out;
for (i = 0; i < MAX_PRIO; i++) {
INIT_LIST_HEAD(&spu_prio->runq[i]);
__clear_bit(i, spu_prio->bitmap);
}
- __set_bit(MAX_PRIO, 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);
+ setup_timer(&spuloadavg_timer, spuloadavg_wake, 0);
+
spusched_task = kthread_run(spusched_thread, NULL, "spusched");
if (IS_ERR(spusched_task)) {
- kfree(spu_prio);
- return PTR_ERR(spusched_task);
+ err = PTR_ERR(spusched_task);
+ goto out_free_spu_prio;
}
+ mod_timer(&spuloadavg_timer, 0);
+
+ entry = proc_create("spu_loadavg", 0, NULL, &spu_loadavg_fops);
+ if (!entry)
+ goto out_stop_kthread;
+
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 __exit spu_sched_exit(void)
+void spu_sched_exit(void)
{
- struct spu *spu, *tmp;
+ struct spu *spu;
int node;
+ remove_proc_entry("spu_loadavg", NULL);
+
+ del_timer_sync(&spusched_timer);
+ del_timer_sync(&spuloadavg_timer);
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);
}