[POWERPC] spusched: Switch from workqueues to kthread + timer tick

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.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Signed-off-by: Jeremy Kerr <jk@ozlabs.org>
Signed-off-by: Paul Mackerras <paulus@samba.org>

4 files changed, 86 insertions, 76 deletions

diff --git a/arch/powerpc/platforms/cell/spufs/context.c b/arch/powerpc/platforms/cell/spufs/context.c
index 7c51cb54bca1..f084667e4f50 100644
--- a/arch/powerpc/platforms/cell/spufs/context.c
+++ b/arch/powerpc/platforms/cell/spufs/context.c
@@ -56,7 +56,7 @@ struct spu_context *alloc_spu_context(struct spu_gang *gang)
         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);
diff --git a/arch/powerpc/platforms/cell/spufs/run.c b/arch/powerpc/platforms/cell/spufs/run.c
index 3ba30cea764a..89b02b6bfc55 100644
--- a/arch/powerpc/platforms/cell/spufs/run.c
+++ b/arch/powerpc/platforms/cell/spufs/run.c
@@ -144,7 +144,6 @@ static int spu_run_init(struct spu_context *ctx, u32 * npc)
                 ctx->ops->runcntl_write(ctx, runcntl);
         } else {
                 unsigned long mode = SPU_PRIVCNTL_MODE_NORMAL;
-                spu_start_tick(ctx);
                 ctx->ops->npc_write(ctx, *npc);
                 if (test_thread_flag(TIF_SINGLESTEP))
                         mode = SPU_PRIVCNTL_MODE_SINGLE_STEP;
@@ -160,7 +159,6 @@ static int spu_run_fini(struct spu_context *ctx, u32 * npc,
 {
         int ret = 0;
 
-        spu_stop_tick(ctx);
         *status = ctx->ops->status_read(ctx);
         *npc = ctx->ops->npc_read(ctx);
         spu_release(ctx);
@@ -330,10 +328,8 @@ long spufs_run_spu(struct file *file, struct spu_context *ctx,
 
                 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);
diff --git a/arch/powerpc/platforms/cell/spufs/sched.c b/arch/powerpc/platforms/cell/spufs/sched.c
index 3b831e07f1ed..d673353b6d33 100644
--- a/arch/powerpc/platforms/cell/spufs/sched.c
+++ b/arch/powerpc/platforms/cell/spufs/sched.c
@@ -35,6 +35,7 @@
 #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>
@@ -45,6 +46,8 @@
 
 #define SPU_TIMESLICE   (HZ)
 
+#define SPUSCHED_TICK   (HZ / 100)
+
 struct spu_prio_array {
         DECLARE_BITMAP(bitmap, MAX_PRIO);
         struct list_head runq[MAX_PRIO];
@@ -54,7 +57,8 @@ 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)
 {
@@ -68,31 +72,6 @@ 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
@@ -104,6 +83,11 @@ static void spu_add_to_active_list(struct spu *spu)
         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
@@ -113,7 +97,7 @@ static void spu_remove_from_active_list(struct spu *spu)
         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]);
 }
 
@@ -161,7 +145,6 @@ static void spu_bind_context(struct spu *spu, struct spu_context *ctx)
         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;
 }
 
@@ -175,7 +158,6 @@ 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);
 
-        spu_remove_from_active_list(spu);
         spu_switch_notify(spu, NULL);
         spu_unmap_mappings(ctx);
         spu_save(&ctx->csa, spu);
@@ -312,6 +294,7 @@ static struct spu *find_victim(struct spu_context *ctx)
                                 victim = NULL;
                                 goto restart;
                         }
+                        spu_remove_from_active_list(spu);
                         spu_unbind_context(spu, victim);
                         mutex_unlock(&victim->state_mutex);
                         /*
@@ -354,6 +337,7 @@ int spu_activate(struct spu_context *ctx, unsigned long flags)
                         spu = find_victim(ctx);
                 if (spu) {
                         spu_bind_context(spu, ctx);
+                        spu_add_to_active_list(spu);
                         return 0;
                 }
 
@@ -397,6 +381,7 @@ static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
         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)
@@ -437,51 +422,78 @@ void spu_yield(struct spu_context *ctx)
         }
 }
 
-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;
-
-        mutex_lock(&ctx->state_mutex);
-        preempted = __spu_deactivate(ctx, 0, ctx->prio + 1);
-        mutex_unlock(&ctx->state_mutex);
+        if (mutex_trylock(&ctx->state_mutex)) {
+                struct spu_context *new = grab_runnable_context(ctx->prio + 1);
+                if (new) {
+                        struct spu *spu = ctx->spu;
 
-        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()) {
+                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]);
+                }
+        }
+
+        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);
@@ -492,7 +504,14 @@ 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)
@@ -500,6 +519,8 @@ 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],
@@ -510,5 +531,4 @@ void __exit spu_sched_exit(void)
                 mutex_unlock(&spu_prio->active_mutex[node]);
         }
         kfree(spu_prio);
-        destroy_workqueue(spu_sched_wq);
 }
diff --git a/arch/powerpc/platforms/cell/spufs/spufs.h b/arch/powerpc/platforms/cell/spufs/spufs.h
index 47617e8014a5..8068171dfa9c 100644
--- a/arch/powerpc/platforms/cell/spufs/spufs.h
+++ b/arch/powerpc/platforms/cell/spufs/spufs.h
@@ -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,
@@ -39,11 +41,6 @@ enum {
 struct spu_context_ops;
 struct spu_gang;
 
-/* ctx->sched_flags */
-enum {
-        SPU_SCHED_EXITING = 0,
-};
-
 struct spu_context {
         struct spu *spu;                  /* pointer to a physical SPU */
         struct spu_state csa;             /* SPU context save area. */
@@ -83,7 +80,7 @@ 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;
@@ -200,9 +197,6 @@ void spu_acquire_saved(struct spu_context *ctx);
 int spu_activate(struct spu_context *ctx, unsigned long flags);
 void spu_deactivate(struct spu_context *ctx);
 void spu_yield(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);