Merge branch 'cell-merge' of git+ssh://master.kernel.org/pub/scm/linux/kernel/git/arnd/cell-2.6

6 files changed, 379 insertions, 217 deletions

diff --git a/arch/powerpc/platforms/cell/pmu.c b/arch/powerpc/platforms/cell/pmu.c
index d04ae1671e6c..66ca4b5a1dbc 100644
--- a/arch/powerpc/platforms/cell/pmu.c
+++ b/arch/powerpc/platforms/cell/pmu.c
@@ -345,18 +345,12 @@ EXPORT_SYMBOL_GPL(cbe_read_trace_buffer);
  * Enabling/disabling interrupts for the entire performance monitoring unit.
  */
 
-u32 cbe_query_pm_interrupts(u32 cpu)
-{
-        return cbe_read_pm(cpu, pm_status);
-}
-EXPORT_SYMBOL_GPL(cbe_query_pm_interrupts);
-
-u32 cbe_clear_pm_interrupts(u32 cpu)
+u32 cbe_get_and_clear_pm_interrupts(u32 cpu)
 {
         /* Reading pm_status clears the interrupt bits. */
-        return cbe_query_pm_interrupts(cpu);
+        return cbe_read_pm(cpu, pm_status);
 }
-EXPORT_SYMBOL_GPL(cbe_clear_pm_interrupts);
+EXPORT_SYMBOL_GPL(cbe_get_and_clear_pm_interrupts);
 
 void cbe_enable_pm_interrupts(u32 cpu, u32 thread, u32 mask)
 {
@@ -371,7 +365,7 @@ EXPORT_SYMBOL_GPL(cbe_enable_pm_interrupts);
 
 void cbe_disable_pm_interrupts(u32 cpu)
 {
-        cbe_clear_pm_interrupts(cpu);
+        cbe_get_and_clear_pm_interrupts(cpu);
         cbe_write_pm(cpu, pm_status, 0);
 }
 EXPORT_SYMBOL_GPL(cbe_disable_pm_interrupts);
diff --git a/arch/powerpc/platforms/cell/spufs/context.c b/arch/powerpc/platforms/cell/spufs/context.c
index 28c718ca3b51..04ad2e364e97 100644
--- a/arch/powerpc/platforms/cell/spufs/context.c
+++ b/arch/powerpc/platforms/cell/spufs/context.c
@@ -42,7 +42,7 @@ struct spu_context *alloc_spu_context(struct spu_gang *gang)
         }
         spin_lock_init(&ctx->mmio_lock);
         kref_init(&ctx->kref);
-        init_rwsem(&ctx->state_sema);
+        mutex_init(&ctx->state_mutex);
         init_MUTEX(&ctx->run_sema);
         init_waitqueue_head(&ctx->ibox_wq);
         init_waitqueue_head(&ctx->wbox_wq);
@@ -53,6 +53,10 @@ struct spu_context *alloc_spu_context(struct spu_gang *gang)
         ctx->owner = get_task_mm(current);
         if (gang)
                 spu_gang_add_ctx(gang, ctx);
+        ctx->rt_priority = current->rt_priority;
+        ctx->policy = current->policy;
+        ctx->prio = current->prio;
+        INIT_DELAYED_WORK(&ctx->sched_work, spu_sched_tick);
         goto out;
 out_free:
         kfree(ctx);
@@ -65,9 +69,9 @@ void destroy_spu_context(struct kref *kref)
 {
         struct spu_context *ctx;
         ctx = container_of(kref, struct spu_context, kref);
-        down_write(&ctx->state_sema);
+        mutex_lock(&ctx->state_mutex);
         spu_deactivate(ctx);
-        up_write(&ctx->state_sema);
+        mutex_unlock(&ctx->state_mutex);
         spu_fini_csa(&ctx->csa);
         if (ctx->gang)
                 spu_gang_remove_ctx(ctx->gang, ctx);
@@ -96,16 +100,6 @@ void spu_forget(struct spu_context *ctx)
         spu_release(ctx);
 }
 
-void spu_acquire(struct spu_context *ctx)
-{
-        down_read(&ctx->state_sema);
-}
-
-void spu_release(struct spu_context *ctx)
-{
-        up_read(&ctx->state_sema);
-}
-
 void spu_unmap_mappings(struct spu_context *ctx)
 {
         if (ctx->local_store)
@@ -124,83 +118,84 @@ void spu_unmap_mappings(struct spu_context *ctx)
                 unmap_mapping_range(ctx->psmap, 0, 0x20000, 1);
 }
 
+/**
+ * spu_acquire_exclusive - lock spu contex and protect against userspace access
+ * @ctx:        spu contex to lock
+ *
+ * Note:
+ *      Returns 0 and with the context locked on success
+ *      Returns negative error and with the context _unlocked_ on failure.
+ */
 int spu_acquire_exclusive(struct spu_context *ctx)
 {
-        int ret = 0;
+        int ret = -EINVAL;
 
-        down_write(&ctx->state_sema);
-        /* ctx is about to be freed, can't acquire any more */
-        if (!ctx->owner) {
-                ret = -EINVAL;
-                goto out;
-        }
+        spu_acquire(ctx);
+        /*
+         * Context is about to be freed, so we can't acquire it anymore.
+         */
+        if (!ctx->owner)
+                goto out_unlock;
 
         if (ctx->state == SPU_STATE_SAVED) {
                 ret = spu_activate(ctx, 0);
                 if (ret)
-                        goto out;
-                ctx->state = SPU_STATE_RUNNABLE;
+                        goto out_unlock;
         } else {
-                /* We need to exclude userspace access to the context. */
+                /*
+                 * We need to exclude userspace access to the context.
+                 *
+                 * To protect against memory access we invalidate all ptes
+                 * and make sure the pagefault handlers block on the mutex.
+                 */
                 spu_unmap_mappings(ctx);
         }
 
-out:
-        if (ret)
-                up_write(&ctx->state_sema);
+        return 0;
+
+ out_unlock:
+        spu_release(ctx);
         return ret;
 }
 
-int spu_acquire_runnable(struct spu_context *ctx)
+/**
+ * spu_acquire_runnable - lock spu contex and make sure it is in runnable state
+ * @ctx:        spu contex to lock
+ *
+ * Note:
+ *      Returns 0 and with the context locked on success
+ *      Returns negative error and with the context _unlocked_ on failure.
+ */
+int spu_acquire_runnable(struct spu_context *ctx, unsigned long flags)
 {
-        int ret = 0;
-
-        down_read(&ctx->state_sema);
-        if (ctx->state == SPU_STATE_RUNNABLE) {
-                ctx->spu->prio = current->prio;
-                return 0;
-        }
-        up_read(&ctx->state_sema);
-
-        down_write(&ctx->state_sema);
-        /* ctx is about to be freed, can't acquire any more */
-        if (!ctx->owner) {
-                ret = -EINVAL;
-                goto out;
-        }
+        int ret = -EINVAL;
 
+        spu_acquire(ctx);
         if (ctx->state == SPU_STATE_SAVED) {
-                ret = spu_activate(ctx, 0);
+                /*
+                 * Context is about to be freed, so we can't acquire it anymore.
+                 */
+                if (!ctx->owner)
+                        goto out_unlock;
+                ret = spu_activate(ctx, flags);
                 if (ret)
-                        goto out;
-                ctx->state = SPU_STATE_RUNNABLE;
+                        goto out_unlock;
         }
 
-        downgrade_write(&ctx->state_sema);
-        /* On success, we return holding the lock */
-
-        return ret;
-out:
-        /* Release here, to simplify calling code. */
-        up_write(&ctx->state_sema);
+        return 0;
 
+ out_unlock:
+        spu_release(ctx);
         return ret;
 }
 
+/**
+ * spu_acquire_saved - lock spu contex and make sure it is in saved state
+ * @ctx:        spu contex to lock
+ */
 void spu_acquire_saved(struct spu_context *ctx)
 {
-        down_read(&ctx->state_sema);
-
-        if (ctx->state == SPU_STATE_SAVED)
-                return;
-
-        up_read(&ctx->state_sema);
-        down_write(&ctx->state_sema);
-
-        if (ctx->state == SPU_STATE_RUNNABLE) {
+        spu_acquire(ctx);
+        if (ctx->state != SPU_STATE_SAVED)
                 spu_deactivate(ctx);
-                ctx->state = SPU_STATE_SAVED;
-        }
-
-        downgrade_write(&ctx->state_sema);
 }
diff --git a/arch/powerpc/platforms/cell/spufs/file.c b/arch/powerpc/platforms/cell/spufs/file.c
index a528020baa18..b00653d69c01 100644
--- a/arch/powerpc/platforms/cell/spufs/file.c
+++ b/arch/powerpc/platforms/cell/spufs/file.c
@@ -103,6 +103,9 @@ static unsigned long spufs_mem_mmap_nopfn(struct vm_area_struct *vma,
 
         offset += vma->vm_pgoff << PAGE_SHIFT;
 
+        if (offset >= LS_SIZE)
+                return NOPFN_SIGBUS;
+
         spu_acquire(ctx);
 
         if (ctx->state == SPU_STATE_SAVED) {
@@ -164,7 +167,7 @@ static unsigned long spufs_ps_nopfn(struct vm_area_struct *vma,
         /* error here usually means a signal.. we might want to test
          * the error code more precisely though
          */
-        ret = spu_acquire_runnable(ctx);
+        ret = spu_acquire_runnable(ctx, 0);
         if (ret)
                 return NOPFN_REFAULT;
 
@@ -1306,7 +1309,7 @@ static ssize_t spufs_mfc_write(struct file *file, const char __user *buffer,
         if (ret)
                 goto out;
 
-        spu_acquire_runnable(ctx);
+        spu_acquire_runnable(ctx, 0);
         if (file->f_flags & O_NONBLOCK) {
                 ret = ctx->ops->send_mfc_command(ctx, &cmd);
         } else {
diff --git a/arch/powerpc/platforms/cell/spufs/run.c b/arch/powerpc/platforms/cell/spufs/run.c
index 1acc2ffef8c8..353a8fa07ab8 100644
--- a/arch/powerpc/platforms/cell/spufs/run.c
+++ b/arch/powerpc/platforms/cell/spufs/run.c
@@ -133,7 +133,7 @@ out_drop_priv:
         spu_mfc_sr1_set(ctx->spu, sr1);
 
 out_unlock:
-        spu_release_exclusive(ctx);
+        spu_release(ctx);
 out:
         return ret;
 }
@@ -143,7 +143,7 @@ static inline int spu_run_init(struct spu_context *ctx, u32 * npc)
         int ret;
         unsigned long runcntl = SPU_RUNCNTL_RUNNABLE;
 
-        ret = spu_acquire_runnable(ctx);
+        ret = spu_acquire_runnable(ctx, SPU_ACTIVATE_NOWAKE);
         if (ret)
                 return ret;
 
@@ -155,7 +155,7 @@ static inline int spu_run_init(struct spu_context *ctx, u32 * npc)
                         spu_release(ctx);
                         ret = spu_setup_isolated(ctx);
                         if (!ret)
-                                ret = spu_acquire_runnable(ctx);
+                                ret = spu_acquire_runnable(ctx, SPU_ACTIVATE_NOWAKE);
                 }
 
                 /* if userspace has set the runcntrl register (eg, to issue an
@@ -164,8 +164,10 @@ static inline int spu_run_init(struct spu_context *ctx, u32 * npc)
                         (SPU_RUNCNTL_RUNNABLE | SPU_RUNCNTL_ISOLATE);
                 if (runcntl == 0)
                         runcntl = SPU_RUNCNTL_RUNNABLE;
-        } else
+        } else {
+                spu_start_tick(ctx);
                 ctx->ops->npc_write(ctx, *npc);
+        }
 
         ctx->ops->runcntl_write(ctx, runcntl);
         return ret;
@@ -176,6 +178,7 @@ static inline 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);
@@ -329,8 +332,10 @@ 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)
+                        if (ret) {
+                                spu_stop_tick(ctx);
                                 goto out2;
+                        }
                         continue;
                 }
                 ret = spu_process_events(ctx);
@@ -361,4 +366,3 @@ out:
         up(&ctx->run_sema);
         return ret;
 }
-
diff --git a/arch/powerpc/platforms/cell/spufs/sched.c b/arch/powerpc/platforms/cell/spufs/sched.c
index bd6fe4b7a84b..2f25e68b4bac 100644
--- a/arch/powerpc/platforms/cell/spufs/sched.c
+++ b/arch/powerpc/platforms/cell/spufs/sched.c
@@ -44,17 +44,18 @@
 #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];
-        wait_queue_head_t waitq[MAX_PRIO];
+        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];
 };
 
 static struct spu_prio_array *spu_prio;
+static struct workqueue_struct *spu_sched_wq;
 
 static inline int node_allowed(int node)
 {
@@ -68,6 +69,64 @@ static inline int node_allowed(int node)
         return 1;
 }
 
+void spu_start_tick(struct spu_context *ctx)
+{
+        if (ctx->policy == SCHED_RR)
+                queue_delayed_work(spu_sched_wq, &ctx->sched_work, SPU_TIMESLICE);
+}
+
+void spu_stop_tick(struct spu_context *ctx)
+{
+        if (ctx->policy == SCHED_RR)
+                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 rearm = 1;
+
+        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);
+                        rearm = 0;
+                }
+        }
+        mutex_unlock(&ctx->state_mutex);
+
+        if (rearm)
+                spu_start_tick(ctx);
+}
+
+/**
+ * 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]);
+}
+
+/**
+ * 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)
+{
+        int node = spu->node;
+
+        mutex_lock(&spu_prio->active_mutex[node]);
+        list_del_init(&spu->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;
@@ -94,8 +153,12 @@ int spu_switch_event_unregister(struct notifier_block * n)
         return blocking_notifier_chain_unregister(&spu_switch_notifier, n);
 }
 
-
-static inline void bind_context(struct spu *spu, struct spu_context *ctx)
+/**
+ * spu_bind_context - bind spu context to physical spu
+ * @spu:        physical spu to bind to
+ * @ctx:        context to bind
+ */
+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);
@@ -104,7 +167,6 @@ static inline void bind_context(struct spu *spu, struct spu_context *ctx)
         ctx->spu = spu;
         ctx->ops = &spu_hw_ops;
         spu->pid = current->pid;
-        spu->prio = current->prio;
         spu->mm = ctx->owner;
         mm_needs_global_tlbie(spu->mm);
         spu->ibox_callback = spufs_ibox_callback;
@@ -118,12 +180,21 @@ static inline void 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;
 }
 
-static inline void unbind_context(struct spu *spu, struct spu_context *ctx)
+/**
+ * spu_unbind_context - unbind spu context from physical spu
+ * @spu:        physical spu to unbind from
+ * @ctx:        context to unbind
+ */
+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);
@@ -136,95 +207,98 @@ static inline void unbind_context(struct spu *spu, struct spu_context *ctx)
         spu->dma_callback = NULL;
         spu->mm = NULL;
         spu->pid = 0;
-        spu->prio = MAX_PRIO;
         ctx->ops = &spu_backing_ops;
         ctx->spu = NULL;
         spu->flags = 0;
         spu->ctx = NULL;
 }
 
-static inline void spu_add_wq(wait_queue_head_t * wq, wait_queue_t * wait,
-                              int prio)
+/**
+ * spu_add_to_rq - add a context to the runqueue
+ * @ctx:       context to add
+ */
+static void spu_add_to_rq(struct spu_context *ctx)
 {
-        prepare_to_wait_exclusive(wq, wait, TASK_INTERRUPTIBLE);
-        set_bit(prio, spu_prio->bitmap);
+        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);
 }
 
-static inline void spu_del_wq(wait_queue_head_t * wq, wait_queue_t * wait,
-                              int 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)
 {
-        u64 flags;
-
-        __set_current_state(TASK_RUNNING);
-
-        spin_lock_irqsave(&wq->lock, flags);
+        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);
+}
 
-        remove_wait_queue_locked(wq, wait);
-        if (list_empty(&wq->task_list))
-                clear_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)
+{
+        struct list_head *rq = &spu_prio->runq[prio];
 
-        spin_unlock_irqrestore(&wq->lock, flags);
+        if (list_empty(rq))
+                return NULL;
+        return list_entry(rq->next, struct spu_context, rq);
 }
 
-static void spu_prio_wait(struct spu_context *ctx, u64 flags)
+static void spu_prio_wait(struct spu_context *ctx)
 {
-        int prio = current->prio;
-        wait_queue_head_t *wq = &spu_prio->waitq[prio];
         DEFINE_WAIT(wait);
 
-        if (ctx->spu)
-                return;
-
-        spu_add_wq(wq, &wait, prio);
-
+        set_bit(SPU_SCHED_WAKE, &ctx->sched_flags);
+        prepare_to_wait_exclusive(&ctx->stop_wq, &wait, TASK_INTERRUPTIBLE);
         if (!signal_pending(current)) {
-                up_write(&ctx->state_sema);
-                pr_debug("%s: pid=%d prio=%d\n", __FUNCTION__,
-                         current->pid, current->prio);
+                mutex_unlock(&ctx->state_mutex);
                 schedule();
-                down_write(&ctx->state_sema);
+                mutex_lock(&ctx->state_mutex);
         }
-
-        spu_del_wq(wq, &wait, prio);
+        __set_current_state(TASK_RUNNING);
+        remove_wait_queue(&ctx->stop_wq, &wait);
+        clear_bit(SPU_SCHED_WAKE, &ctx->sched_flags);
 }
 
-static void spu_prio_wakeup(void)
+/**
+ * spu_reschedule - try to find a runnable context for a spu
+ * @spu:       spu available
+ *
+ * This function is called whenever a spu becomes idle.  It looks for the
+ * most suitable runnable spu context and schedules it for execution.
+ */
+static void spu_reschedule(struct spu *spu)
 {
-        int best = sched_find_first_bit(spu_prio->bitmap);
-        if (best < MAX_PRIO) {
-                wait_queue_head_t *wq = &spu_prio->waitq[best];
-                wake_up_interruptible_nr(wq, 1);
-        }
-}
+        int best;
 
-static int get_active_spu(struct spu *spu)
-{
-        int node = spu->node;
-        struct spu *tmp;
-        int rc = 0;
+        spu_free(spu);
 
-        mutex_lock(&spu_prio->active_mutex[node]);
-        list_for_each_entry(tmp, &spu_prio->active_list[node], list) {
-                if (tmp == spu) {
-                        list_del_init(&spu->list);
-                        rc = 1;
-                        break;
-                }
+        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);
         }
-        mutex_unlock(&spu_prio->active_mutex[node]);
-        return rc;
-}
-
-static void put_active_spu(struct spu *spu)
-{
-        int node = spu->node;
-
-        mutex_lock(&spu_prio->active_mutex[node]);
-        list_add_tail(&spu->list, &spu_prio->active_list[node]);
-        mutex_unlock(&spu_prio->active_mutex[node]);
+        spin_unlock(&spu_prio->runq_lock);
 }
 
-static struct spu *spu_get_idle(struct spu_context *ctx, u64 flags)
+static struct spu *spu_get_idle(struct spu_context *ctx)
 {
         struct spu *spu = NULL;
         int node = cpu_to_node(raw_smp_processor_id());
@@ -241,87 +315,154 @@ static struct spu *spu_get_idle(struct spu_context *ctx, u64 flags)
         return spu;
 }
 
-static inline struct spu *spu_get(struct spu_context *ctx, u64 flags)
+/**
+ * find_victim - find a lower priority context to preempt
+ * @ctx:        canidate context for running
+ *
+ * Returns the freed physical spu to run the new context on.
+ */
+static struct spu *find_victim(struct spu_context *ctx)
 {
-        /* Future: spu_get_idle() if possible,
-         * otherwise try to preempt an active
-         * context.
+        struct spu_context *victim = NULL;
+        struct spu *spu;
+        int node, n;
+
+        /*
+         * 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
+         * a strong node affinity.  We might want to fine-tune this in
+         * the future.
          */
-        return spu_get_idle(ctx, flags);
+ restart:
+        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))
+                        continue;
+
+                mutex_lock(&spu_prio->active_mutex[node]);
+                list_for_each_entry(spu, &spu_prio->active_list[node], list) {
+                        struct spu_context *tmp = spu->ctx;
+
+                        if (tmp->rt_priority < ctx->rt_priority &&
+                            (!victim || tmp->rt_priority < victim->rt_priority))
+                                victim = spu->ctx;
+                }
+                mutex_unlock(&spu_prio->active_mutex[node]);
+
+                if (victim) {
+                        /*
+                         * This nests ctx->state_mutex, but we always lock
+                         * higher priority contexts before lower priority
+                         * ones, so this is safe until we introduce
+                         * priority inheritance schemes.
+                         */
+                        if (!mutex_trylock(&victim->state_mutex)) {
+                                victim = NULL;
+                                goto restart;
+                        }
+
+                        spu = victim->spu;
+                        if (!spu) {
+                                /*
+                                 * This race can happen because we've dropped
+                                 * the active list mutex.  No a problem, just
+                                 * restart the search.
+                                 */
+                                mutex_unlock(&victim->state_mutex);
+                                victim = NULL;
+                                goto restart;
+                        }
+                        spu_unbind_context(spu, victim);
+                        mutex_unlock(&victim->state_mutex);
+                        return spu;
+                }
+        }
+
+        return NULL;
 }
 
-/* The three externally callable interfaces
- * for the scheduler begin here.
+/**
+ * spu_activate - find a free spu for a context and execute it
+ * @ctx:        spu context to schedule
+ * @flags:      flags (currently ignored)
  *
- *      spu_activate    - bind a context to SPU, waiting as needed.
- *      spu_deactivate  - unbind a context from its SPU.
- *      spu_yield       - yield an SPU if others are waiting.
+ * Tries to find a free spu to run @ctx.  If no free spu is availble
+ * add the context to the runqueue so it gets woken up once an spu
+ * is available.
  */
-
-int spu_activate(struct spu_context *ctx, u64 flags)
+int spu_activate(struct spu_context *ctx, unsigned long flags)
 {
-        struct spu *spu;
-        int ret = 0;
 
-        for (;;) {
-                if (ctx->spu)
+        if (ctx->spu)
+                return 0;
+
+        do {
+                struct spu *spu;
+
+                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 && ctx->rt_priority)
+                        spu = find_victim(ctx);
+                if (spu) {
+                        spu_bind_context(spu, ctx);
                         return 0;
-                spu = spu_get(ctx, flags);
-                if (spu != NULL) {
-                        if (ctx->spu != NULL) {
-                                spu_free(spu);
-                                spu_prio_wakeup();
-                                break;
-                        }
-                        bind_context(spu, ctx);
-                        put_active_spu(spu);
-                        break;
                 }
-                spu_prio_wait(ctx, flags);
-                if (signal_pending(current)) {
-                        ret = -ERESTARTSYS;
-                        spu_prio_wakeup();
-                        break;
-                }
-        }
-        return ret;
+
+                spu_add_to_rq(ctx);
+                if (!(flags & SPU_ACTIVATE_NOWAKE))
+                        spu_prio_wait(ctx);
+                spu_del_from_rq(ctx);
+        } while (!signal_pending(current));
+
+        return -ERESTARTSYS;
 }
 
+/**
+ * spu_deactivate - unbind a context from it's physical spu
+ * @ctx:        spu context to unbind
+ *
+ * Unbind @ctx from the physical spu it is running on and schedule
+ * the highest priority context to run on the freed physical spu.
+ */
 void spu_deactivate(struct spu_context *ctx)
 {
-        struct spu *spu;
-        int needs_idle;
+        struct spu *spu = ctx->spu;
 
-        spu = ctx->spu;
-        if (!spu)
-                return;
-        needs_idle = get_active_spu(spu);
-        unbind_context(spu, ctx);
-        if (needs_idle) {
-                spu_free(spu);
-                spu_prio_wakeup();
+        if (spu) {
+                spu_unbind_context(spu, ctx);
+                spu_reschedule(spu);
         }
 }
 
+/**
+ * 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
+ * unbind @ctx from the physical spu and schedule the highest
+ * priority context to run on the freed physical spu instead.
+ */
 void spu_yield(struct spu_context *ctx)
 {
         struct spu *spu;
         int need_yield = 0;
 
-        if (down_write_trylock(&ctx->state_sema)) {
+        if (mutex_trylock(&ctx->state_mutex)) {
                 if ((spu = ctx->spu) != NULL) {
                         int best = sched_find_first_bit(spu_prio->bitmap);
                         if (best < MAX_PRIO) {
                                 pr_debug("%s: yielding SPU %d NODE %d\n",
                                          __FUNCTION__, spu->number, spu->node);
                                 spu_deactivate(ctx);
-                                ctx->state = SPU_STATE_SAVED;
                                 need_yield = 1;
-                        } else {
-                                spu->prio = MAX_PRIO;
                         }
                 }
-                up_write(&ctx->state_sema);
+                mutex_unlock(&ctx->state_mutex);
         }
         if (unlikely(need_yield))
                 yield();
@@ -331,14 +472,19 @@ 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++) {
-                init_waitqueue_head(&spu_prio->waitq[i]);
+                INIT_LIST_HEAD(&spu_prio->runq[i]);
                 __clear_bit(i, spu_prio->bitmap);
         }
         __set_bit(MAX_PRIO, spu_prio->bitmap);
@@ -346,6 +492,7 @@ int __init spu_sched_init(void)
                 mutex_init(&spu_prio->active_mutex[i]);
                 INIT_LIST_HEAD(&spu_prio->active_list[i]);
         }
+        spin_lock_init(&spu_prio->runq_lock);
         return 0;
 }
 
@@ -364,4 +511,5 @@ 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 0941c56df9b5..0c437891dfd5 100644
--- a/arch/powerpc/platforms/cell/spufs/spufs.h
+++ b/arch/powerpc/platforms/cell/spufs/spufs.h
@@ -23,7 +23,7 @@
 #define SPUFS_H
 
 #include <linux/kref.h>
-#include <linux/rwsem.h>
+#include <linux/mutex.h>
 #include <linux/spinlock.h>
 #include <linux/fs.h>
 
@@ -37,11 +37,13 @@ enum {
 };
 
 struct spu_context_ops;
-
-#define SPU_CONTEXT_PREEMPT          0UL
-
 struct spu_gang;
 
+/* ctx->sched_flags */
+enum {
+        SPU_SCHED_WAKE = 0,
+};
+
 struct spu_context {
         struct spu *spu;                  /* pointer to a physical SPU */
         struct spu_state csa;             /* SPU context save area. */
@@ -56,7 +58,7 @@ struct spu_context {
         u64 object_id;             /* user space pointer for oprofile */
 
         enum { SPU_STATE_RUNNABLE, SPU_STATE_SAVED } state;
-        struct rw_semaphore state_sema;
+        struct mutex state_mutex;
         struct semaphore run_sema;
 
         struct mm_struct *owner;
@@ -77,6 +79,14 @@ struct spu_context {
 
         struct list_head gang_list;
         struct spu_gang *gang;
+
+        /* scheduler fields */
+        struct list_head rq;
+        struct delayed_work sched_work;
+        unsigned long sched_flags;
+        unsigned long rt_priority;
+        int policy;
+        int prio;
 };
 
 struct spu_gang {
@@ -161,6 +171,16 @@ void spu_gang_remove_ctx(struct spu_gang *gang, struct spu_context *ctx);
 void spu_gang_add_ctx(struct spu_gang *gang, struct spu_context *ctx);
 
 /* context management */
+static inline void spu_acquire(struct spu_context *ctx)
+{
+        mutex_lock(&ctx->state_mutex);
+}
+
+static inline void spu_release(struct spu_context *ctx)
+{
+        mutex_unlock(&ctx->state_mutex);
+}
+
 struct spu_context * alloc_spu_context(struct spu_gang *gang);
 void destroy_spu_context(struct kref *kref);
 struct spu_context * get_spu_context(struct spu_context *ctx);
@@ -168,20 +188,18 @@ int put_spu_context(struct spu_context *ctx);
 void spu_unmap_mappings(struct spu_context *ctx);
 
 void spu_forget(struct spu_context *ctx);
-void spu_acquire(struct spu_context *ctx);
-void spu_release(struct spu_context *ctx);
-int spu_acquire_runnable(struct spu_context *ctx);
+int spu_acquire_runnable(struct spu_context *ctx, unsigned long flags);
 void spu_acquire_saved(struct spu_context *ctx);
 int spu_acquire_exclusive(struct spu_context *ctx);
-
-static inline void spu_release_exclusive(struct spu_context *ctx)
-{
-        up_write(&ctx->state_sema);
-}
-
-int spu_activate(struct spu_context *ctx, u64 flags);
+enum {
+        SPU_ACTIVATE_NOWAKE = 1,
+};
+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);