[Cbe-oss-dev] PATCH [4/7] decouple spu scheduler from spufs_spu_run [asynchronous scheduling]

Luke Browning lukebr at linux.vnet.ibm.com
Thu Nov 1 11:07:18 EST 2007


This patch changes spufs_spu_run so that the context is queued 
directly to the scheduler and the controlling thread advances
directly to spufs_wait() for spe errors and exceptions.

Isolated (No-sched) contexts are treated the same as before.

Signed-off-by: Luke Browning <lukebr at linux.vnet.ibm.com>

---

Index: linux-2.6.22/arch/powerpc/platforms/cell/spufs/run.c
===================================================================
--- linux-2.6.22.orig/arch/powerpc/platforms/cell/spufs/run.c
+++ linux-2.6.22/arch/powerpc/platforms/cell/spufs/run.c
@@ -39,19 +39,27 @@ void spufs_stop_callback(struct spu *spu
 
 static inline int spu_stopped(struct spu_context *ctx, u32 * stat)
 {
-	struct spu *spu;
-	u64 pte_fault;
+	u64 dsisr;
+	u32 stopped;
 
 	*stat = ctx->ops->status_read(ctx);
 
-	spu = ctx->spu;
-	if (ctx->state != SPU_STATE_RUNNABLE ||
-	    test_bit(SPU_SCHED_NOTIFY_ACTIVE, &ctx->sched_flags))
+	if (test_bit(SPU_SCHED_NOTIFY_ACTIVE, &ctx->sched_flags))
 		return 1;
-	pte_fault = ctx->csa.dsisr &
-	    (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED);
-	return (!(*stat & SPU_STATUS_RUNNING) || pte_fault || ctx->csa.class_0_pending) ?
-		1 : 0;
+
+	stopped = SPU_STATUS_INVALID_INSTR | SPU_STATUS_SINGLE_STEP |
+		SPU_STATUS_STOPPED_BY_HALT | SPU_STATUS_STOPPED_BY_STOP;
+	if (*stat & stopped)
+		return 1;
+
+	dsisr = ctx->csa.dsisr;
+	if (dsisr & (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED))
+		return 1;
+
+	if (ctx->csa.class_0_pending)
+		return 1;
+
+	return 0;
 }
 
 static int spu_setup_isolated(struct spu_context *ctx)
@@ -147,24 +155,27 @@ out:
 
 static int spu_run_init(struct spu_context *ctx, u32 * npc)
 {
-	unsigned long runcntl;
+	unsigned long runcntl = SPU_RUNCNTL_RUNNABLE;
 	int ret;
 
 	spuctx_switch_state(ctx, SPUCTX_UTIL_SYSTEM);
 
-	if (ctx->flags & SPU_CREATE_ISOLATE) {
-		/*
-		 * Force activation of spu.  Isolated state assumes that
-		 * special loader context is loaded and running on spu.
-		 */
+	/*
+	 * NOSCHED is synchronous scheduling with respect to the caller.
+	 * The caller waits for the context to be loaded.
+	 */
+	if (ctx->flags & SPU_CREATE_NOSCHED) {
 		if (ctx->state == SPU_STATE_SAVED) {
-			spu_set_timeslice(ctx);
-
 			ret = spu_activate(ctx, 0);
 			if (ret)
 				return ret;
 		}
+	}
 
+	/*
+	 * Apply special setup as required.
+	 */
+	if (ctx->flags & SPU_CREATE_ISOLATE) {
 		if (!(ctx->ops->status_read(ctx) & SPU_STATUS_ISOLATED_STATE)) {
 			ret = spu_setup_isolated(ctx);
 			if (ret)
@@ -172,17 +183,18 @@ static int spu_run_init(struct spu_conte
 		}
 
 		/*
-		 * If userspace has set the runcntrl register (eg, to
-		 * issue an isolated exit), we need to re-set it here
+		 * If userspace has set the runcntrl register (eg, to issue
+		 * an isolated exit), we need to re-set it here
 		 */
 		runcntl = ctx->ops->runcntl_read(ctx) &
 			(SPU_RUNCNTL_RUNNABLE | SPU_RUNCNTL_ISOLATE);
 		if (runcntl == 0)
 			runcntl = SPU_RUNCNTL_RUNNABLE;
+	}
 
+	if (ctx->flags & SPU_CREATE_NOSCHED) {
 		spuctx_switch_state(ctx, SPUCTX_UTIL_USER);
 		ctx->ops->runcntl_write(ctx, runcntl);
-
 	} else {
 		unsigned long privcntl;
 
@@ -190,20 +202,17 @@ static int spu_run_init(struct spu_conte
 			privcntl = SPU_PRIVCNTL_MODE_SINGLE_STEP;
 		else
 			privcntl = SPU_PRIVCNTL_MODE_NORMAL;
-		runcntl = SPU_RUNCNTL_RUNNABLE;
 
 		ctx->ops->npc_write(ctx, *npc);
 		ctx->ops->privcntl_write(ctx, privcntl);
+		spuctx_switch_state(ctx, SPUCTX_UTIL_USER);
+		ctx->ops->runcntl_write(ctx, runcntl);
 
 		if (ctx->state == SPU_STATE_SAVED) {
-			spu_set_timeslice(ctx);
 			ret = spu_activate(ctx, 0);
 			if (ret)
 				return ret;
 		}
-
-		spuctx_switch_state(ctx, SPUCTX_UTIL_USER);
-		ctx->ops->runcntl_write(ctx, runcntl);
 	}
 
 	return 0;
@@ -222,9 +231,6 @@ static int spu_run_fini(struct spu_conte
 	spuctx_switch_state(ctx, SPUCTX_UTIL_LOADED);
 	spu_release(ctx);
 
-	if (signal_pending(current))
-		ret = -ERESTARTSYS;
-
 	return ret;
 }
 
@@ -353,9 +359,9 @@ static inline int spu_handle_class0_even
 			ret = spu_handle_class0(ctx);
 		}
 	}
-	if (!ret && signal_pending(current))
-		ret = -ERESTARTSYS;
+
 	ctx->csa.class_0_pending = 0;
+
 	return ret;
 }
 
@@ -383,8 +389,8 @@ long spufs_run_spu(struct file *file, st
 	}
 
 	do {
-		ret = spufs_wait(ctx->stop_wq, spu_stopped(ctx, &status));
 		spu = ctx->spu;
+		ret = spufs_wait(ctx->stop_wq, spu_stopped(ctx, &status));
 		if (unlikely(ret))
 			break;
 
@@ -398,6 +404,7 @@ long spufs_run_spu(struct file *file, st
 				continue;
 			}
 		}
+
 		if ((status & SPU_STATUS_STOPPED_BY_STOP) &&
 		    (status >> SPU_STOP_STATUS_SHIFT == 0x2104)) {
 			ret = spu_process_callback(ctx);
@@ -405,21 +412,31 @@ long spufs_run_spu(struct file *file, st
 				break;
 			status &= ~SPU_STATUS_STOPPED_BY_STOP;
 		}
+
 		ret = spu_handle_class1(ctx);
 		if (ret)
 			break;
 
+		ret = spu_handle_class0_events(ctx);
+		if (ret)
+			break;
+
+		if (status & (SPU_STATUS_STOPPED_BY_STOP |
+			      SPU_STATUS_STOPPED_BY_HALT |
+			      SPU_STATUS_SINGLE_STEP))
+			break;
+
+		if (signal_pending(current)) {
+			ret = -ERESTARTSYS;
+			break;
+		}
+
 		if (unlikely(ctx->state != SPU_STATE_RUNNABLE)) {
 			ret = spu_reacquire_runnable(ctx, npc, &status);
 			if (ret)
 				goto out2;
-			continue;
 		}
-		ret = spu_handle_class0_events(ctx);
-
-	} while (!ret && !(status & (SPU_STATUS_STOPPED_BY_STOP |
-				      SPU_STATUS_STOPPED_BY_HALT |
-				       SPU_STATUS_SINGLE_STEP)));
+	} while (1);
 
 	if ((status & SPU_STATUS_STOPPED_BY_STOP) &&
 	    (((status >> SPU_STOP_STATUS_SHIFT) & 0x3f00) == 0x2100) &&
Index: linux-2.6.22/arch/powerpc/platforms/cell/spufs/sched.c
===================================================================
--- linux-2.6.22.orig/arch/powerpc/platforms/cell/spufs/sched.c
+++ linux-2.6.22/arch/powerpc/platforms/cell/spufs/sched.c
@@ -487,6 +487,13 @@ static void __spu_add_to_rq(struct spu_c
 	}
 }
 
+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;
@@ -505,6 +512,13 @@ static void spu_prio_wait(struct spu_con
 {
 	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)) {
@@ -604,6 +618,7 @@ static struct spu *find_victim(struct sp
 			struct spu_context *tmp = spu->ctx;
 
 			if (tmp && tmp->prio > ctx->prio &&
+			    !(tmp->flags & SPU_CREATE_NOSCHED) &&
 			    (!victim || tmp->prio > victim->prio))
 				victim = spu->ctx;
 		}
@@ -645,12 +660,8 @@ static struct spu *find_victim(struct sp
 			victim->stats.invol_ctx_switch++;
 			spu->stats.invol_ctx_switch++;
 			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);
+
+			spu_add_to_rq(victim);
 			return spu;
 		}
 	}
@@ -658,6 +669,48 @@ static struct spu *find_victim(struct sp
 	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)
+{
+	spu_acquire(ctx);
+	__spu_schedule(spu, ctx);
+	spu_release(ctx);
+}
+
+static void spu_unschedule(struct spu *spu, struct spu_context *ctx)
+{
+	int node = spu->node;
+
+	mutex_lock(&cbe_spu_info[node].list_mutex);
+	cbe_spu_info[node].nr_active--;
+	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
@@ -669,40 +722,42 @@ static struct spu *find_victim(struct sp
  */
 int spu_activate(struct spu_context *ctx, unsigned long flags)
 {
-	do {
-		struct spu *spu;
+	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;
-
-		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) {
-			int node = spu->node;
+	/*
+	 * 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;
 
-			mutex_lock(&cbe_spu_info[node].list_mutex);
-			spu_bind_context(spu, ctx);
-			cbe_spu_info[node].nr_active++;
-			mutex_unlock(&cbe_spu_info[node].list_mutex);
-			wake_up_all(&ctx->run_wq);
-			return 0;
-		}
+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_schedule(spu, ctx);
+		return 0;
+	}
+
+	if (ctx->flags & SPU_CREATE_NOSCHED) {
 		spu_prio_wait(ctx);
-	} while (!signal_pending(current));
+		goto spu_activate_top;
+	}
 
-	return -ERESTARTSYS;
+	spu_add_to_rq(ctx);
+
+	return 0;
 }
 
 /**
@@ -744,21 +799,17 @@ static int __spu_deactivate(struct spu_c
 	if (spu) {
 		new = grab_runnable_context(max_prio, spu->node);
 		if (new || force) {
-			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++;
-
-			if (new)
-				wake_up(&new->stop_wq);
+			spu_unschedule(spu, ctx);
+			if (new) {
+				if (new->flags & SPU_CREATE_NOSCHED)
+					wake_up(&new->stop_wq);
+				else {
+					spu_release(ctx);
+					spu_schedule(spu, new);
+					spu_acquire(ctx);
+				}
+			}
 		}
-
 	}
 
 	return new != NULL;
@@ -814,21 +865,14 @@ static noinline void spusched_tick(struc
 
 		new = grab_runnable_context(ctx->prio + 1, spu->node);
 		if (new) {
-			spu_unbind_context(spu, ctx);
-			ctx->stats.invol_ctx_switch++;
-			spu->stats.invol_ctx_switch++;
-			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
-			 * spu_run manually to ensure this context
-			 * gets put on the runqueue again ASAP.
-			 */
-			wake_up(&ctx->stop_wq);
+			spu_unschedule(spu, ctx);
+			spu_add_to_rq(ctx);
+		} else {
+			spu_set_timeslice(ctx);
 		}
-		spu_set_timeslice(ctx);
 		mutex_unlock(&ctx->state_mutex);
+		if (new)
+			spu_schedule(spu, new);
 	} else {
 		ctx->time_slice++;
 	}
Index: linux-2.6.22/arch/powerpc/platforms/cell/spufs/file.c
===================================================================
--- linux-2.6.22.orig/arch/powerpc/platforms/cell/spufs/file.c
+++ linux-2.6.22/arch/powerpc/platforms/cell/spufs/file.c
@@ -572,6 +572,9 @@ void spufs_ibox_callback(struct spu *spu
 {
 	struct spu_context *ctx = spu->ctx;
 
+	if (!ctx)
+		return;
+
 	wake_up_all(&ctx->ibox_wq);
 	kill_fasync(&ctx->ibox_fasync, SIGIO, POLLIN);
 }
@@ -708,6 +711,9 @@ void spufs_wbox_callback(struct spu *spu
 {
 	struct spu_context *ctx = spu->ctx;
 
+	if (!ctx)
+		return;
+
 	wake_up_all(&ctx->wbox_wq);
 	kill_fasync(&ctx->wbox_fasync, SIGIO, POLLOUT);
 }
@@ -1320,6 +1326,9 @@ void spufs_mfc_callback(struct spu *spu)
 {
 	struct spu_context *ctx = spu->ctx;
 
+	if (!ctx)
+		return;
+
 	wake_up_all(&ctx->mfc_wq);
 
 	pr_debug("%s %s\n", __FUNCTION__, spu->name);





More information about the cbe-oss-dev mailing list