[Cbe-oss-dev] [PATCH 2/3] spu sched: switch from workqueues to kthread + timer tick
Christoph Hellwig
hch at lst.de
Mon Apr 16 05:31:36 EST 2007
On Thu, Apr 12, 2007 at 08:42:53PM +0200, Christoph Hellwig wrote:
> Get rid of the scheduler workqueues that complicated things a lot to
> a dedicated spu scheduler thread that gets woken by a traditional
> scheduler tick. By default this scheduler tick runs a HZ * 10, aka
> one spu scheduler tick for every 10 cpu ticks.
>
> Currently the tick is not disabled when we have less context than
> available spus, but I will implement this later.
This version of the patch had some issues with timer handling. Updated
version below:
Index: linux-2.6/arch/powerpc/platforms/cell/spufs/sched.c
===================================================================
--- linux-2.6.orig/arch/powerpc/platforms/cell/spufs/sched.c 2007-04-12 19:45:37.000000000 +0200
+++ linux-2.6/arch/powerpc/platforms/cell/spufs/sched.c 2007-04-15 20:29:03.000000000 +0200
@@ -36,6 +36,7 @@
#include <linux/numa.h>
#include <linux/mutex.h>
#include <linux/notifier.h>
+#include <linux/kthread.h>
#include <asm/io.h>
#include <asm/mmu_context.h>
@@ -46,6 +47,8 @@
#define SPU_TIMESLICE (HZ)
+#define SPUSCHED_TICK (HZ / 100)
+
struct spu_prio_array {
DECLARE_BITMAP(bitmap, MAX_PRIO);
struct list_head runq[MAX_PRIO];
@@ -55,7 +58,8 @@ struct spu_prio_array {
};
static struct spu_prio_array *spu_prio;
-static struct workqueue_struct *spu_sched_wq;
+static struct task_struct *spusched_task;
+static struct timer_list spusched_timer;
static inline int node_allowed(int node)
{
@@ -69,31 +73,6 @@ 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);
- }
-}
-
/**
* spu_add_to_active_list - add spu to active list
* @spu: spu to add to the active list
@@ -105,6 +84,11 @@ static void spu_add_to_active_list(struc
mutex_unlock(&spu_prio->active_mutex[spu->node]);
}
+static void __spu_remove_from_active_list(struct spu *spu)
+{
+ list_del_init(&spu->list);
+}
+
/**
* spu_remove_from_active_list - remove spu from active list
* @spu: spu to remove from the active list
@@ -114,7 +98,7 @@ static void spu_remove_from_active_list(
int node = spu->node;
mutex_lock(&spu_prio->active_mutex[node]);
- list_del_init(&spu->list);
+ __spu_remove_from_active_list(spu);
mutex_unlock(&spu_prio->active_mutex[node]);
}
@@ -193,7 +177,6 @@ static void spu_bind_context(struct spu
spu->timestamp = jiffies;
spu_cpu_affinity_set(spu, raw_smp_processor_id());
spu_switch_notify(spu, ctx);
- spu_add_to_active_list(spu);
ctx->state = SPU_STATE_RUNNABLE;
}
@@ -212,7 +195,6 @@ static void spu_unbind_context(struct sp
if (!list_empty(&ctx->aff_list))
if (atomic_dec_and_test(&ctx->gang->aff_sched_count))
ctx->gang->aff_ref_spu = NULL;
- spu_remove_from_active_list(spu);
spu_switch_notify(spu, NULL);
spu_unmap_mappings(ctx);
spu_save(&ctx->csa, spu);
@@ -354,6 +336,7 @@ static struct spu *find_victim(struct sp
victim = NULL;
goto restart;
}
+ spu_remove_from_active_list(spu);
spu_unbind_context(spu, victim);
mutex_unlock(&victim->state_mutex);
/*
@@ -396,6 +379,7 @@ int spu_activate(struct spu_context *ctx
spu = find_victim(ctx);
if (spu) {
spu_bind_context(spu, ctx);
+ spu_add_to_active_list(spu);
return 0;
}
@@ -439,6 +423,7 @@ static int __spu_deactivate(struct spu_c
if (spu) {
new = grab_runnable_context(max_prio);
if (new || force) {
+ spu_remove_from_active_list(spu);
spu_unbind_context(spu, ctx);
spu_free(spu);
if (new)
@@ -477,51 +462,77 @@ void spu_yield(struct spu_context *ctx)
mutex_unlock(&ctx->state_mutex);
}
-void spu_sched_tick(struct work_struct *work)
+static void spusched_tick(struct spu_context *ctx)
{
- struct spu_context *ctx =
- container_of(work, struct spu_context, sched_work.work);
- int preempted;
+ if (ctx->policy != SCHED_RR || --ctx->time_slice)
+ return;
/*
- * 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.
+ * 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 (test_bit(SPU_SCHED_EXITING, &ctx->sched_flags))
- return;
+ if (mutex_trylock(&ctx->state_mutex)) {
+ struct spu_context *new = grab_runnable_context(ctx->prio + 1);
+ if (new) {
+ struct spu *spu = ctx->spu;
- mutex_lock(&ctx->state_mutex);
- preempted = __spu_deactivate(ctx, 0, ctx->prio + 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);
+ __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);
+ }
+ ctx->time_slice = SPU_DEF_TIMESLICE;
+ mutex_unlock(&ctx->state_mutex);
} else {
- spu_start_tick(ctx);
+ ctx->time_slice++;
+ }
+}
+
+static void spusched_wake(unsigned long data)
+{
+ mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
+ wake_up_process(spusched_task);
+}
+
+static int spusched_thread(void *unused)
+{
+ struct spu *spu, *next;
+ int node;
+
+ setup_timer(&spusched_timer, spusched_wake, 0);
+ __mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
+
+ while (!kthread_should_stop()) {
+ 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]);
+ }
}
+
+ del_timer_sync(&spusched_timer);
+ return 0;
}
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;
- }
+ if (!spu_prio)
+ return -ENOMEM;
+
for (i = 0; i < MAX_PRIO; i++) {
INIT_LIST_HEAD(&spu_prio->runq[i]);
__clear_bit(i, spu_prio->bitmap);
@@ -532,7 +543,14 @@ int __init spu_sched_init(void)
INIT_LIST_HEAD(&spu_prio->active_list[i]);
}
spin_lock_init(&spu_prio->runq_lock);
+
+ spusched_task = kthread_run(spusched_thread, NULL, "spusched");
+ if (IS_ERR(spusched_task)) {
+ kfree(spu_prio);
+ return PTR_ERR(spusched_task);
+ }
return 0;
+
}
void __exit spu_sched_exit(void)
@@ -540,6 +558,8 @@ void __exit spu_sched_exit(void)
struct spu *spu, *tmp;
int node;
+ 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],
@@ -550,7 +570,6 @@ void __exit spu_sched_exit(void)
mutex_unlock(&spu_prio->active_mutex[node]);
}
kfree(spu_prio);
- destroy_workqueue(spu_sched_wq);
}
static void aff_merge_remaining_ctxs(struct spu_gang *gang)
Index: linux-2.6/arch/powerpc/platforms/cell/spufs/context.c
===================================================================
--- linux-2.6.orig/arch/powerpc/platforms/cell/spufs/context.c 2007-04-12 19:45:23.000000000 +0200
+++ linux-2.6/arch/powerpc/platforms/cell/spufs/context.c 2007-04-15 20:29:03.000000000 +0200
@@ -58,7 +58,7 @@ struct spu_context *alloc_spu_context(st
ctx->rt_priority = current->rt_priority;
ctx->policy = current->policy;
ctx->prio = current->prio;
- INIT_DELAYED_WORK(&ctx->sched_work, spu_sched_tick);
+ ctx->time_slice = SPU_DEF_TIMESLICE;
goto out;
out_free:
kfree(ctx);
@@ -179,5 +179,3 @@ void * spu_get_profile_private_kref(stru
return ctx->prof_priv_kref;
}
EXPORT_SYMBOL_GPL(spu_get_profile_private_kref);
-
-
Index: linux-2.6/arch/powerpc/platforms/cell/spufs/run.c
===================================================================
--- linux-2.6.orig/arch/powerpc/platforms/cell/spufs/run.c 2007-04-12 19:45:24.000000000 +0200
+++ linux-2.6/arch/powerpc/platforms/cell/spufs/run.c 2007-04-15 20:25:30.000000000 +0200
@@ -148,7 +148,6 @@ static inline int spu_run_init(struct sp
if (runcntl == 0)
runcntl = SPU_RUNCNTL_RUNNABLE;
} else {
- spu_start_tick(ctx);
ctx->ops->npc_write(ctx, *npc);
}
@@ -161,7 +160,6 @@ static inline int spu_run_fini(struct sp
{
int ret = 0;
- spu_stop_tick(ctx);
*status = ctx->ops->status_read(ctx);
*npc = ctx->ops->npc_read(ctx);
spu_release(ctx);
@@ -326,10 +324,8 @@ long spufs_run_spu(struct file *file, st
if (unlikely(ctx->state != SPU_STATE_RUNNABLE)) {
ret = spu_reacquire_runnable(ctx, npc, &status);
- if (ret) {
- spu_stop_tick(ctx);
+ if (ret)
goto out2;
- }
continue;
}
ret = spu_process_events(ctx);
Index: linux-2.6/arch/powerpc/platforms/cell/spufs/spufs.h
===================================================================
--- linux-2.6.orig/arch/powerpc/platforms/cell/spufs/spufs.h 2007-04-12 19:45:23.000000000 +0200
+++ linux-2.6/arch/powerpc/platforms/cell/spufs/spufs.h 2007-04-15 20:29:03.000000000 +0200
@@ -31,6 +31,8 @@
#include <asm/spu_csa.h>
#include <asm/spu_info.h>
+#define SPU_DEF_TIMESLICE 100
+
/* The magic number for our file system */
enum {
SPUFS_MAGIC = 0x23c9b64e,
@@ -41,8 +43,7 @@ struct spu_gang;
/* ctx->sched_flags */
enum {
- SPU_SCHED_EXITING = 0,
- SPU_SCHED_NOTIFY_ACTIVE,
+ SPU_SCHED_NOTIFY_ACTIVE = 0,
};
struct spu_context {
@@ -86,7 +87,7 @@ struct spu_context {
/* scheduler fields */
struct list_head rq;
- struct delayed_work sched_work;
+ unsigned int time_slice;
unsigned long sched_flags;
unsigned long rt_priority;
int policy;
@@ -222,9 +223,6 @@ int spu_activate(struct spu_context *ctx
void spu_deactivate(struct spu_context *ctx);
void spu_yield(struct spu_context *ctx);
void spu_switch_notify(struct spu *spu, struct spu_context *ctx);
-void spu_start_tick(struct spu_context *ctx);
-void spu_stop_tick(struct spu_context *ctx);
-void spu_sched_tick(struct work_struct *work);
int __init spu_sched_init(void);
void __exit spu_sched_exit(void);
More information about the cbe-oss-dev
mailing list