Merge branch 'for-2.6.23' into merge

13 files changed, 754 insertions, 143 deletions

diff --git a/arch/powerpc/platforms/cell/io-workarounds.c b/arch/powerpc/platforms/cell/io-workarounds.c
index 7fb92f23f38..9d7c2ef940a 100644
--- a/arch/powerpc/platforms/cell/io-workarounds.c
+++ b/arch/powerpc/platforms/cell/io-workarounds.c
@@ -102,7 +102,7 @@ static void spider_io_flush(const volatile void __iomem *addr)
 		vaddr = (unsigned long)PCI_FIX_ADDR(addr);
 
 		/* Check if it's in allowed range for  PIO */
-		if (vaddr < PHBS_IO_BASE || vaddr >= IMALLOC_BASE)
+		if (vaddr < PHB_IO_BASE || vaddr > PHB_IO_END)
 			return;
 
 		/* Try to find a PTE. If not, clear the paddr, we'll do
diff --git a/arch/powerpc/platforms/cell/spu_base.c b/arch/powerpc/platforms/cell/spu_base.c
index a7f5a7653c6..e4d0c9f42ab 100644
--- a/arch/powerpc/platforms/cell/spu_base.c
+++ b/arch/powerpc/platforms/cell/spu_base.c
@@ -183,7 +183,7 @@ static int __spu_trap_data_seg(struct spu *spu, unsigned long ea)
 		spu->slb_replace = 0;
 
 	spu_restart_dma(spu);
-
+	spu->stats.slb_flt++;
 	return 0;
 }
 
@@ -332,6 +332,7 @@ spu_irq_class_2(int irq, void *data)
 	if (stat & 0x10) /* SPU mailbox threshold */
 		spu->wbox_callback(spu);
 
+	spu->stats.class2_intr++;
 	return stat ? IRQ_HANDLED : IRQ_NONE;
 }
 
@@ -462,8 +463,18 @@ void spu_free(struct spu *spu)
 }
 EXPORT_SYMBOL_GPL(spu_free);
 
+static int spu_shutdown(struct sys_device *sysdev)
+{
+	struct spu *spu = container_of(sysdev, struct spu, sysdev);
+
+	spu_free_irqs(spu);
+	spu_destroy_spu(spu);
+	return 0;
+}
+
 struct sysdev_class spu_sysdev_class = {
-	set_kset_name("spu")
+	set_kset_name("spu"),
+	.shutdown = spu_shutdown,
 };
 
 int spu_add_sysdev_attr(struct sysdev_attribute *attr)
@@ -574,6 +585,9 @@ static int __init create_spu(void *data)
 	spin_unlock_irqrestore(&spu_list_lock, flags);
 	mutex_unlock(&spu_mutex);
 
+	spu->stats.utilization_state = SPU_UTIL_IDLE;
+	spu->stats.tstamp = jiffies;
+
 	goto out;
 
 out_free_irqs:
@@ -586,6 +600,45 @@ out:
 	return ret;
 }
 
+static const char *spu_state_names[] = {
+	"user", "system", "iowait", "idle"
+};
+
+static unsigned long long spu_acct_time(struct spu *spu,
+		enum spu_utilization_state state)
+{
+	unsigned long long time = spu->stats.times[state];
+
+	if (spu->stats.utilization_state == state)
+		time += jiffies - spu->stats.tstamp;
+
+	return jiffies_to_msecs(time);
+}
+
+
+static ssize_t spu_stat_show(struct sys_device *sysdev, char *buf)
+{
+	struct spu *spu = container_of(sysdev, struct spu, sysdev);
+
+	return sprintf(buf, "%s %llu %llu %llu %llu "
+		      "%llu %llu %llu %llu %llu %llu %llu %llu\n",
+		spu_state_names[spu->stats.utilization_state],
+		spu_acct_time(spu, SPU_UTIL_USER),
+		spu_acct_time(spu, SPU_UTIL_SYSTEM),
+		spu_acct_time(spu, SPU_UTIL_IOWAIT),
+		spu_acct_time(spu, SPU_UTIL_IDLE),
+		spu->stats.vol_ctx_switch,
+		spu->stats.invol_ctx_switch,
+		spu->stats.slb_flt,
+		spu->stats.hash_flt,
+		spu->stats.min_flt,
+		spu->stats.maj_flt,
+		spu->stats.class2_intr,
+		spu->stats.libassist);
+}
+
+static SYSDEV_ATTR(stat, 0644, spu_stat_show, NULL);
+
 static int __init init_spu_base(void)
 {
 	int i, ret = 0;
@@ -611,6 +664,8 @@ static int __init init_spu_base(void)
 
 	xmon_register_spus(&spu_full_list);
 
+	spu_add_sysdev_attr(&attr_stat);
+
 	return 0;
 
  out_unregister_sysdev_class:
diff --git a/arch/powerpc/platforms/cell/spufs/backing_ops.c b/arch/powerpc/platforms/cell/spufs/backing_ops.c
index d32db9ffc6e..07a0e815abf 100644
--- a/arch/powerpc/platforms/cell/spufs/backing_ops.c
+++ b/arch/powerpc/platforms/cell/spufs/backing_ops.c
@@ -320,6 +320,12 @@ static int spu_backing_set_mfc_query(struct spu_context * ctx, u32 mask,
 	/* FIXME: what are the side-effects of this? */
 	prob->dma_querymask_RW = mask;
 	prob->dma_querytype_RW = mode;
+	/* In the current implementation, the SPU context is always
+	 * acquired in runnable state when new bits are added to the
+	 * mask (tagwait), so it's sufficient just to mask
+	 * dma_tagstatus_R with the 'mask' parameter here.
+	 */
+	ctx->csa.prob.dma_tagstatus_R &= mask;
 out:
 	spin_unlock(&ctx->csa.register_lock);
 
diff --git a/arch/powerpc/platforms/cell/spufs/context.c b/arch/powerpc/platforms/cell/spufs/context.c
index 7c51cb54bca..6d7bd60f538 100644
--- a/arch/powerpc/platforms/cell/spufs/context.c
+++ b/arch/powerpc/platforms/cell/spufs/context.c
@@ -23,10 +23,14 @@
 #include <linux/fs.h>
 #include <linux/mm.h>
 #include <linux/slab.h>
+#include <asm/atomic.h>
 #include <asm/spu.h>
 #include <asm/spu_csa.h>
 #include "spufs.h"
 
+
+atomic_t nr_spu_contexts = ATOMIC_INIT(0);
+
 struct spu_context *alloc_spu_context(struct spu_gang *gang)
 {
 	struct spu_context *ctx;
@@ -53,10 +57,12 @@ struct spu_context *alloc_spu_context(struct spu_gang *gang)
 	INIT_LIST_HEAD(&ctx->rq);
 	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);
+	ctx->cpus_allowed = current->cpus_allowed;
+	spu_set_timeslice(ctx);
+	ctx->stats.execution_state = SPUCTX_UTIL_USER;
+	ctx->stats.tstamp = jiffies;
+
+	atomic_inc(&nr_spu_contexts);
 	goto out;
 out_free:
 	kfree(ctx);
@@ -76,6 +82,7 @@ void destroy_spu_context(struct kref *kref)
 	if (ctx->gang)
 		spu_gang_remove_ctx(ctx->gang, ctx);
 	BUG_ON(!list_empty(&ctx->rq));
+	atomic_dec(&nr_spu_contexts);
 	kfree(ctx);
 }
 
diff --git a/arch/powerpc/platforms/cell/spufs/fault.c b/arch/powerpc/platforms/cell/spufs/fault.c
index 0f75c07e29d..e064d0c0d80 100644
--- a/arch/powerpc/platforms/cell/spufs/fault.c
+++ b/arch/powerpc/platforms/cell/spufs/fault.c
@@ -33,7 +33,8 @@
  * function. Currently, there are a few corner cases that we haven't had
  * to handle fortunately.
  */
-static int spu_handle_mm_fault(struct mm_struct *mm, unsigned long ea, unsigned long dsisr)
+static int spu_handle_mm_fault(struct mm_struct *mm, unsigned long ea,
+		unsigned long dsisr, unsigned *flt)
 {
 	struct vm_area_struct *vma;
 	unsigned long is_write;
@@ -73,7 +74,8 @@ good_area:
 			goto bad_area;
 	}
 	ret = 0;
-	switch (handle_mm_fault(mm, vma, ea, is_write)) {
+	*flt = handle_mm_fault(mm, vma, ea, is_write);
+	switch (*flt) {
 	case VM_FAULT_MINOR:
 		current->min_flt++;
 		break;
@@ -153,6 +155,7 @@ int spufs_handle_class1(struct spu_context *ctx)
 {
 	u64 ea, dsisr, access;
 	unsigned long flags;
+	unsigned flt = 0;
 	int ret;
 
 	/*
@@ -178,9 +181,17 @@ int spufs_handle_class1(struct spu_context *ctx)
 	if (!(dsisr & (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED)))
 		return 0;
 
+	spuctx_switch_state(ctx, SPUCTX_UTIL_IOWAIT);
+
 	pr_debug("ctx %p: ea %016lx, dsisr %016lx state %d\n", ctx, ea,
 		dsisr, ctx->state);
 
+	ctx->stats.hash_flt++;
+	if (ctx->state == SPU_STATE_RUNNABLE) {
+		ctx->spu->stats.hash_flt++;
+		spu_switch_state(ctx->spu, SPU_UTIL_IOWAIT);
+	}
+
 	/* we must not hold the lock when entering spu_handle_mm_fault */
 	spu_release(ctx);
 
@@ -192,7 +203,7 @@ int spufs_handle_class1(struct spu_context *ctx)
 
 	/* hashing failed, so try the actual fault handler */
 	if (ret)
-		ret = spu_handle_mm_fault(current->mm, ea, dsisr);
+		ret = spu_handle_mm_fault(current->mm, ea, dsisr, &flt);
 
 	spu_acquire(ctx);
 	/*
@@ -201,11 +212,23 @@ int spufs_handle_class1(struct spu_context *ctx)
 	 * In case of unhandled error report the problem to user space.
 	 */
 	if (!ret) {
+		if (flt == VM_FAULT_MINOR)
+			ctx->stats.min_flt++;
+		else
+			ctx->stats.maj_flt++;
+		if (ctx->state == SPU_STATE_RUNNABLE) {
+			if (flt == VM_FAULT_MINOR)
+				ctx->spu->stats.min_flt++;
+			else
+				ctx->spu->stats.maj_flt++;
+		}
+
 		if (ctx->spu)
 			ctx->ops->restart_dma(ctx);
 	} else
 		spufs_handle_dma_error(ctx, ea, SPE_EVENT_SPE_DATA_STORAGE);
 
+	spuctx_switch_state(ctx, SPUCTX_UTIL_SYSTEM);
 	return ret;
 }
 EXPORT_SYMBOL_GPL(spufs_handle_class1);
diff --git a/arch/powerpc/platforms/cell/spufs/file.c b/arch/powerpc/platforms/cell/spufs/file.c
index b1e7e2f8a2e..c2814ea96af 100644
--- a/arch/powerpc/platforms/cell/spufs/file.c
+++ b/arch/powerpc/platforms/cell/spufs/file.c
@@ -28,6 +28,7 @@
 #include <linux/pagemap.h>
 #include <linux/poll.h>
 #include <linux/ptrace.h>
+#include <linux/seq_file.h>
 
 #include <asm/io.h>
 #include <asm/semaphore.h>
@@ -39,6 +40,7 @@
 
 #define SPUFS_MMAP_4K (PAGE_SIZE == 0x1000)
 
+
 static int
 spufs_mem_open(struct inode *inode, struct file *file)
 {
@@ -216,12 +218,12 @@ unsigned long spufs_get_unmapped_area(struct file *file, unsigned long addr,
 #endif /* CONFIG_SPU_FS_64K_LS */
 
 static const struct file_operations spufs_mem_fops = {
-	.open	 		= spufs_mem_open,
-	.release 		= spufs_mem_release,
-	.read   		= spufs_mem_read,
-	.write   		= spufs_mem_write,
-	.llseek  		= generic_file_llseek,
-	.mmap    		= spufs_mem_mmap,
+	.open			= spufs_mem_open,
+	.release		= spufs_mem_release,
+	.read			= spufs_mem_read,
+	.write			= spufs_mem_write,
+	.llseek			= generic_file_llseek,
+	.mmap			= spufs_mem_mmap,
 #ifdef CONFIG_SPU_FS_64K_LS
 	.get_unmapped_area	= spufs_get_unmapped_area,
 #endif
@@ -1497,14 +1499,15 @@ static ssize_t spufs_mfc_write(struct file *file, const char __user *buffer,
 		if (status)
 			ret = status;
 	}
-	spu_release(ctx);
 
 	if (ret)
-		goto out;
+		goto out_unlock;
 
 	ctx->tagwait |= 1 << cmd.tag;
 	ret = size;
 
+out_unlock:
+	spu_release(ctx);
 out:
 	return ret;
 }
@@ -1515,14 +1518,14 @@ static unsigned int spufs_mfc_poll(struct file *file,poll_table *wait)
 	u32 free_elements, tagstatus;
 	unsigned int mask;
 
+	poll_wait(file, &ctx->mfc_wq, wait);
+
 	spu_acquire(ctx);
 	ctx->ops->set_mfc_query(ctx, ctx->tagwait, 2);
 	free_elements = ctx->ops->get_mfc_free_elements(ctx);
 	tagstatus = ctx->ops->read_mfc_tagstatus(ctx);
 	spu_release(ctx);
 
-	poll_wait(file, &ctx->mfc_wq, wait);
-
 	mask = 0;
 	if (free_elements & 0xffff)
 		mask |= POLLOUT | POLLWRNORM;
@@ -1797,6 +1800,29 @@ static int spufs_info_open(struct inode *inode, struct file *file)
 	return 0;
 }
 
+static int spufs_caps_show(struct seq_file *s, void *private)
+{
+	struct spu_context *ctx = s->private;
+
+	if (!(ctx->flags & SPU_CREATE_NOSCHED))
+		seq_puts(s, "sched\n");
+	if (!(ctx->flags & SPU_CREATE_ISOLATE))
+		seq_puts(s, "step\n");
+	return 0;
+}
+
+static int spufs_caps_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, spufs_caps_show, SPUFS_I(inode)->i_ctx);
+}
+
+static const struct file_operations spufs_caps_fops = {
+	.open		= spufs_caps_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
 static ssize_t __spufs_mbox_info_read(struct spu_context *ctx,
 			char __user *buf, size_t len, loff_t *pos)
 {
@@ -2014,7 +2040,105 @@ static const struct file_operations spufs_proxydma_info_fops = {
 	.read = spufs_proxydma_info_read,
 };
 
+static int spufs_show_tid(struct seq_file *s, void *private)
+{
+	struct spu_context *ctx = s->private;
+
+	seq_printf(s, "%d\n", ctx->tid);
+	return 0;
+}
+
+static int spufs_tid_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, spufs_show_tid, SPUFS_I(inode)->i_ctx);
+}
+
+static const struct file_operations spufs_tid_fops = {
+	.open		= spufs_tid_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+static const char *ctx_state_names[] = {
+	"user", "system", "iowait", "loaded"
+};
+
+static unsigned long long spufs_acct_time(struct spu_context *ctx,
+		enum spuctx_execution_state state)
+{
+	unsigned long time = ctx->stats.times[state];
+
+	if (ctx->stats.execution_state == state)
+		time += jiffies - ctx->stats.tstamp;
+
+	return jiffies_to_msecs(time);
+}
+
+static unsigned long long spufs_slb_flts(struct spu_context *ctx)
+{
+	unsigned long long slb_flts = ctx->stats.slb_flt;
+
+	if (ctx->state == SPU_STATE_RUNNABLE) {
+		slb_flts += (ctx->spu->stats.slb_flt -
+			     ctx->stats.slb_flt_base);
+	}
+
+	return slb_flts;
+}
+
+static unsigned long long spufs_class2_intrs(struct spu_context *ctx)
+{
+	unsigned long long class2_intrs = ctx->stats.class2_intr;
+
+	if (ctx->state == SPU_STATE_RUNNABLE) {
+		class2_intrs += (ctx->spu->stats.class2_intr -
+				 ctx->stats.class2_intr_base);
+	}
+
+	return class2_intrs;
+}
+
+
+static int spufs_show_stat(struct seq_file *s, void *private)
+{
+	struct spu_context *ctx = s->private;
+
+	spu_acquire(ctx);
+	seq_printf(s, "%s %llu %llu %llu %llu "
+		      "%llu %llu %llu %llu %llu %llu %llu %llu\n",
+		ctx_state_names[ctx->stats.execution_state],
+		spufs_acct_time(ctx, SPUCTX_UTIL_USER),
+		spufs_acct_time(ctx, SPUCTX_UTIL_SYSTEM),
+		spufs_acct_time(ctx, SPUCTX_UTIL_IOWAIT),
+		spufs_acct_time(ctx, SPUCTX_UTIL_LOADED),
+		ctx->stats.vol_ctx_switch,
+		ctx->stats.invol_ctx_switch,
+		spufs_slb_flts(ctx),
+		ctx->stats.hash_flt,
+		ctx->stats.min_flt,
+		ctx->stats.maj_flt,
+		spufs_class2_intrs(ctx),
+		ctx->stats.libassist);
+	spu_release(ctx);
+	return 0;
+}
+
+static int spufs_stat_open(struct inode *inode, struct file *file)
+{
+	return single_open(file, spufs_show_stat, SPUFS_I(inode)->i_ctx);
+}
+
+static const struct file_operations spufs_stat_fops = {
+	.open		= spufs_stat_open,
+	.read		= seq_read,
+	.llseek		= seq_lseek,
+	.release	= single_release,
+};
+
+
 struct tree_descr spufs_dir_contents[] = {
+	{ "capabilities", &spufs_caps_fops, 0444, },
 	{ "mem",  &spufs_mem_fops,  0666, },
 	{ "regs", &spufs_regs_fops,  0666, },
 	{ "mbox", &spufs_mbox_fops, 0444, },
@@ -2046,10 +2170,13 @@ struct tree_descr spufs_dir_contents[] = {
 	{ "wbox_info", &spufs_wbox_info_fops, 0444, },
 	{ "dma_info", &spufs_dma_info_fops, 0444, },
 	{ "proxydma_info", &spufs_proxydma_info_fops, 0444, },
+	{ "tid", &spufs_tid_fops, 0444, },
+	{ "stat", &spufs_stat_fops, 0444, },
 	{},
 };
 
 struct tree_descr spufs_dir_nosched_contents[] = {
+	{ "capabilities", &spufs_caps_fops, 0444, },
 	{ "mem",  &spufs_mem_fops,  0666, },
 	{ "mbox", &spufs_mbox_fops, 0444, },
 	{ "ibox", &spufs_ibox_fops, 0444, },
@@ -2068,6 +2195,8 @@ struct tree_descr spufs_dir_nosched_contents[] = {
 	{ "psmap", &spufs_psmap_fops, 0666, },
 	{ "phys-id", &spufs_id_ops, 0666, },
 	{ "object-id", &spufs_object_id_ops, 0666, },
+	{ "tid", &spufs_tid_fops, 0444, },
+	{ "stat", &spufs_stat_fops, 0444, },
 	{},
 };
 
diff --git a/arch/powerpc/platforms/cell/spufs/inode.c b/arch/powerpc/platforms/cell/spufs/inode.c
index 9807206e021..f37460e5bfd 100644
--- a/arch/powerpc/platforms/cell/spufs/inode.c
+++ b/arch/powerpc/platforms/cell/spufs/inode.c
@@ -232,10 +232,6 @@ static int spufs_dir_close(struct inode *inode, struct file *file)
 	return dcache_dir_close(inode, file);
 }
 
-const struct inode_operations spufs_dir_inode_operations = {
-	.lookup = simple_lookup,
-};
-
 const struct file_operations spufs_context_fops = {
 	.open		= dcache_dir_open,
 	.release	= spufs_dir_close,
@@ -269,7 +265,7 @@ spufs_mkdir(struct inode *dir, struct dentry *dentry, unsigned int flags,
 		goto out_iput;
 
 	ctx->flags = flags;
-	inode->i_op = &spufs_dir_inode_operations;
+	inode->i_op = &simple_dir_inode_operations;
 	inode->i_fop = &simple_dir_operations;
 	if (flags & SPU_CREATE_NOSCHED)
 		ret = spufs_fill_dir(dentry, spufs_dir_nosched_contents,
@@ -386,7 +382,7 @@ spufs_mkgang(struct inode *dir, struct dentry *dentry, int mode)
 	if (!gang)
 		goto out_iput;
 
-	inode->i_op = &spufs_dir_inode_operations;
+	inode->i_op = &simple_dir_inode_operations;
 	inode->i_fop = &simple_dir_operations;
 
 	d_instantiate(dentry, inode);
@@ -593,7 +589,7 @@ spufs_create_root(struct super_block *sb, void *data)
 	if (!inode)
 		goto out;
 
-	inode->i_op = &spufs_dir_inode_operations;
+	inode->i_op = &simple_dir_inode_operations;
 	inode->i_fop = &simple_dir_operations;
 	SPUFS_I(inode)->i_ctx = NULL;
 
diff --git a/arch/powerpc/platforms/cell/spufs/run.c b/arch/powerpc/platforms/cell/spufs/run.c
index 57626600b1a..58ae13b7de8 100644
--- a/arch/powerpc/platforms/cell/spufs/run.c
+++ b/arch/powerpc/platforms/cell/spufs/run.c
@@ -29,7 +29,8 @@ static inline int spu_stopped(struct spu_context *ctx, u32 * stat)
 	spu = ctx->spu;
 	pte_fault = spu->dsisr &
 	    (MFC_DSISR_PTE_NOT_FOUND | MFC_DSISR_ACCESS_DENIED);
-	return (!(*stat & 0x1) || pte_fault || spu->class_0_pending) ? 1 : 0;
+	return (!(*stat & SPU_STATUS_RUNNING) || pte_fault || spu->class_0_pending) ?
+		1 : 0;
 }
 
 static int spu_setup_isolated(struct spu_context *ctx)
@@ -142,8 +143,11 @@ static int spu_run_init(struct spu_context *ctx, u32 * npc)
 			runcntl = SPU_RUNCNTL_RUNNABLE;
 		ctx->ops->runcntl_write(ctx, runcntl);
 	} else {
-		spu_start_tick(ctx);
+		unsigned long mode = SPU_PRIVCNTL_MODE_NORMAL;
 		ctx->ops->npc_write(ctx, *npc);
+		if (test_thread_flag(TIF_SINGLESTEP))
+			mode = SPU_PRIVCNTL_MODE_SINGLE_STEP;
+		out_be64(&ctx->spu->priv2->spu_privcntl_RW, mode);
 		ctx->ops->runcntl_write(ctx, SPU_RUNCNTL_RUNNABLE);
 	}
 
@@ -155,7 +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);
@@ -298,9 +301,22 @@ long spufs_run_spu(struct file *file, struct spu_context *ctx,
 	ctx->ops->master_start(ctx);
 	ctx->event_return = 0;
 
-	ret = spu_acquire_runnable(ctx, 0);
-	if (ret)
-		return ret;
+	spu_acquire(ctx);
+	if (ctx->state == SPU_STATE_SAVED) {
+		__spu_update_sched_info(ctx);
+
+		ret = spu_activate(ctx, 0);
+		if (ret) {
+			spu_release(ctx);
+			goto out;
+		}
+	} else {
+		/*
+		 * We have to update the scheduling priority under active_mutex
+		 * to protect against find_victim().
+		 */
+		spu_update_sched_info(ctx);
+	}
 
 	ret = spu_run_init(ctx, npc);
 	if (ret) {
@@ -325,16 +341,20 @@ 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);
 
 	} while (!ret && !(status & (SPU_STATUS_STOPPED_BY_STOP |
-				      SPU_STATUS_STOPPED_BY_HALT)));
+				      SPU_STATUS_STOPPED_BY_HALT |
+				       SPU_STATUS_SINGLE_STEP)));
+
+	if ((status & SPU_STATUS_STOPPED_BY_STOP) &&
+	    (((status >> SPU_STOP_STATUS_SHIFT) & 0x3f00) == 0x2100) &&
+	    (ctx->state == SPU_STATE_RUNNABLE))
+		ctx->stats.libassist++;
 
 	ctx->ops->master_stop(ctx);
 	ret = spu_run_fini(ctx, npc, &status);
@@ -344,10 +364,15 @@ out2:
 	if ((ret == 0) ||
 	    ((ret == -ERESTARTSYS) &&
 	     ((status & SPU_STATUS_STOPPED_BY_HALT) ||
+	      (status & SPU_STATUS_SINGLE_STEP) ||
 	      ((status & SPU_STATUS_STOPPED_BY_STOP) &&
 	       (status >> SPU_STOP_STATUS_SHIFT != 0x2104)))))
 		ret = status;
 
+	/* Note: we don't need to force_sig SIGTRAP on single-step
+	 * since we have TIF_SINGLESTEP set, thus the kernel will do
+	 * it upon return from the syscall anyawy
+	 */
 	if ((status & SPU_STATUS_STOPPED_BY_STOP)
 	    && (status >> SPU_STOP_STATUS_SHIFT) == 0x3fff) {
 		force_sig(SIGTRAP, current);
diff --git a/arch/powerpc/platforms/cell/spufs/sched.c b/arch/powerpc/platforms/cell/spufs/sched.c
index 3b831e07f1e..e5b4dd1db28 100644
--- a/arch/powerpc/platforms/cell/spufs/sched.c
+++ b/arch/powerpc/platforms/cell/spufs/sched.c
@@ -35,6 +35,10 @@
 #include <linux/numa.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>
@@ -43,54 +47,126 @@
 #include <asm/spu_priv1.h>
 #include "spufs.h"
 
-#define SPU_TIMESLICE	(HZ)
-
 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_active[MAX_NUMNODES];
+	int nr_waiting;
 };
 
+static unsigned long spu_avenrun[3];
 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;
+
+/*
+ * Priority of a normal, non-rt, non-niced'd process (aka nice level 0).
+ */
+#define NORMAL_PRIO		120
+
+/*
+ * Frequency of the spu scheduler tick.  By default we do one SPU scheduler
+ * tick for every 10 CPU scheduler ticks.
+ */
+#define SPUSCHED_TICK		(10)
 
-static inline int node_allowed(int node)
+/*
+ * These are the 'tuning knobs' of the scheduler:
+ *
+ * Minimum timeslice is 5 msecs (or 1 spu scheduler tick, whichever is
+ * larger), default timeslice is 100 msecs, maximum timeslice is 800 msecs.
+ */
+#define MIN_SPU_TIMESLICE	max(5 * HZ / (1000 * SPUSCHED_TICK), 1)
+#define DEF_SPU_TIMESLICE	(100 * HZ / (1000 * SPUSCHED_TICK))
+
+#define MAX_USER_PRIO		(MAX_PRIO - MAX_RT_PRIO)
+#define SCALE_PRIO(x, prio) \
+	max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_SPU_TIMESLICE)
+
+/*
+ * scale user-nice values [ -20 ... 0 ... 19 ] to time slice values:
+ * [800ms ... 100ms ... 5ms]
+ *
+ * The higher a thread's priority, the bigger timeslices
+ * it gets during one round of execution. But even the lowest
+ * priority thread gets MIN_TIMESLICE worth of execution time.
+ */
+void spu_set_timeslice(struct spu_context *ctx)
 {
-	cpumask_t mask;
+	if (ctx->prio < NORMAL_PRIO)
+		ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE * 4, ctx->prio);
+	else
+		ctx->time_slice = SCALE_PRIO(DEF_SPU_TIMESLICE, ctx->prio);
+}
 
-	if (!nr_cpus_node(node))
-		return 0;
-	mask = node_to_cpumask(node);
-	if (!cpus_intersects(mask, current->cpus_allowed))
-		return 0;
-	return 1;
+/*
+ * Update scheduling information from the owning thread.
+ */
+void __spu_update_sched_info(struct spu_context *ctx)
+{
+	/*
+	 * 32-Bit assignment are atomic on powerpc, and we don't care about
+	 * memory ordering here because retriving the controlling thread is
+	 * per defintion racy.
+	 */
+	ctx->tid = current->pid;
+
+	/*
+	 * We do our own priority calculations, so we normally want
+	 * ->static_prio to start with. Unfortunately thies 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;
+
+	/*
+	 * A lot of places that don't hold active_mutex poke into
+	 * cpus_allowed, including grab_runnable_context which
+	 * already holds the runq_lock.  So abuse runq_lock
+	 * to protect this field aswell.
+	 */
+	spin_lock(&spu_prio->runq_lock);
+	ctx->cpus_allowed = current->cpus_allowed;
+	spin_unlock(&spu_prio->runq_lock);
 }
 
-void spu_start_tick(struct spu_context *ctx)
+void spu_update_sched_info(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);
-	}
+	int node = ctx->spu->node;
+
+	mutex_lock(&spu_prio->active_mutex[node]);
+	__spu_update_sched_info(ctx);
+	mutex_unlock(&spu_prio->active_mutex[node]);
 }
 
-void spu_stop_tick(struct spu_context *ctx)
+static int __node_allowed(struct spu_context *ctx, int node)
 {
-	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);
+	if (nr_cpus_node(node)) {
+		cpumask_t mask = node_to_cpumask(node);
+
+		if (cpus_intersects(mask, ctx->cpus_allowed))
+			return 1;
 	}
+
+	return 0;
+}
+
+static int node_allowed(struct spu_context *ctx, int node)
+{
+	int rval;
+
+	spin_lock(&spu_prio->runq_lock);
+	rval = __node_allowed(ctx, node);
+	spin_unlock(&spu_prio->runq_lock);
+
+	return rval;
 }
 
 /**
@@ -99,9 +175,18 @@ void spu_stop_tick(struct spu_context *ctx)
  */
 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]);
+	int node = spu->node;
+
+	mutex_lock(&spu_prio->active_mutex[node]);
+	spu_prio->nr_active[node]++;
+	list_add_tail(&spu->list, &spu_prio->active_list[node]);
+	mutex_unlock(&spu_prio->active_mutex[node]);
+}
+
+static void __spu_remove_from_active_list(struct spu *spu)
+{
+	list_del_init(&spu->list);
+	spu_prio->nr_active[spu->node]--;
 }
 
 /**
@@ -113,7 +198,7 @@ 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]);
 }
 
@@ -144,6 +229,10 @@ 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);
+
+	ctx->stats.slb_flt_base = spu->stats.slb_flt;
+	ctx->stats.class2_intr_base = spu->stats.class2_intr;
+
 	spu->ctx = ctx;
 	spu->flags = 0;
 	ctx->spu = spu;
@@ -161,8 +250,8 @@ 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;
+	spu_switch_state(spu, SPU_UTIL_SYSTEM);
 }
 
 /**
@@ -175,7 +264,8 @@ 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_state(spu, SPU_UTIL_IDLE);
+
 	spu_switch_notify(spu, NULL);
 	spu_unmap_mappings(ctx);
 	spu_save(&ctx->csa, spu);
@@ -192,6 +282,11 @@ static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
 	ctx->spu = NULL;
 	spu->flags = 0;
 	spu->ctx = 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);
 }
 
 /**
@@ -200,20 +295,39 @@ static void spu_unbind_context(struct spu *spu, struct spu_context *ctx)
  */
 static void __spu_add_to_rq(struct spu_context *ctx)
 {
-	int prio = ctx->prio;
-
-	list_add_tail(&ctx->rq, &spu_prio->runq[prio]);
-	set_bit(prio, spu_prio->bitmap);
+	/*
+	 * 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);
+	}
 }
 
 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);
+	}
 }
 
 static void spu_prio_wait(struct spu_context *ctx)
@@ -244,7 +358,7 @@ static struct spu *spu_get_idle(struct spu_context *ctx)
 
 	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)
@@ -276,15 +390,15 @@ static struct spu *find_victim(struct spu_context *ctx)
 	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) {
 			struct spu_context *tmp = spu->ctx;
 
-			if (tmp->rt_priority < ctx->rt_priority &&
-			    (!victim || tmp->rt_priority < victim->rt_priority))
+			if (tmp->prio > ctx->prio &&
+			    (!victim || tmp->prio > victim->prio))
 				victim = spu->ctx;
 		}
 		mutex_unlock(&spu_prio->active_mutex[node]);
@@ -312,7 +426,10 @@ static struct spu *find_victim(struct spu_context *ctx)
 				victim = NULL;
 				goto restart;
 			}
+			spu_remove_from_active_list(spu);
 			spu_unbind_context(spu, victim);
+			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
@@ -338,22 +455,30 @@ static struct spu *find_victim(struct spu_context *ctx)
  */
 int spu_activate(struct spu_context *ctx, unsigned long flags)
 {
-
-	if (ctx->spu)
-		return 0;
+	spuctx_switch_state(ctx, SPUCTX_UTIL_SYSTEM);
 
 	do {
 		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 && ctx->rt_priority)
+		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;
 		}
 
@@ -369,23 +494,28 @@ int spu_activate(struct spu_context *ctx, unsigned long flags)
  * 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) {
+	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;
 }
 
@@ -395,9 +525,12 @@ static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
 	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);
+			ctx->stats.vol_ctx_switch++;
+			spu->stats.vol_ctx_switch++;
 			spu_free(spu);
 			if (new)
 				wake_up(&new->stop_wq);
@@ -417,7 +550,17 @@ static int __spu_deactivate(struct spu_context *ctx, int force, int max_prio)
  */
 void spu_deactivate(struct spu_context *ctx)
 {
+	/*
+	 * We must never reach this for a nosched context,
+	 * but handle the case gracefull instead of panicing.
+	 */
+	if (ctx->flags & SPU_CREATE_NOSCHED) {
+		WARN_ON(1);
+		return;
+	}
+
 	__spu_deactivate(ctx, 1, MAX_PRIO);
+	spuctx_switch_state(ctx, SPUCTX_UTIL_USER);
 }
 
 /**
@@ -432,56 +575,178 @@ void spu_yield(struct spu_context *ctx)
 {
 	if (!(ctx->flags & SPU_CREATE_NOSCHED)) {
 		mutex_lock(&ctx->state_mutex);
-		__spu_deactivate(ctx, 0, MAX_PRIO);
+		if (__spu_deactivate(ctx, 0, MAX_PRIO))
+			spuctx_switch_state(ctx, SPUCTX_UTIL_USER);
+		else {
+			spuctx_switch_state(ctx, SPUCTX_UTIL_LOADED);
+			spu_switch_state(ctx->spu, SPU_UTIL_USER);
+		}
 		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->flags & SPU_CREATE_NOSCHED)
+		return;
+	if (ctx->policy == SCHED_FIFO)
+		return;
+
+	if (--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 *spu = ctx->spu;
+		struct spu_context *new;
 
-	mutex_lock(&ctx->state_mutex);
-	preempted = __spu_deactivate(ctx, 0, ctx->prio + 1);
-	mutex_unlock(&ctx->state_mutex);
+		new = grab_runnable_context(ctx->prio + 1, spu->node);
+		if (new) {
 
-	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);
+			ctx->stats.invol_ctx_switch++;
+			spu->stats.invol_ctx_switch++;
+			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);
 	} else {
-		spu_start_tick(ctx);
+		ctx->time_slice++;
 	}
 }
 
-int __init spu_sched_init(void)
+/**
+ * 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 / active_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 i;
+	int nr_active = 0, node;
 
-	spu_sched_wq = create_singlethread_workqueue("spusched");
-	if (!spu_sched_wq)
-		return 1;
+	for (node = 0; node < MAX_NUMNODES; node++)
+		nr_active += spu_prio->nr_active[node];
+	nr_active += spu_prio->nr_waiting;
 
-	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;
+	return nr_active;
+}
+
+/**
+ * spu_calc_load - given tick count, update the avenrun load estimates.
+ * @tick:	tick count
+ *
+ * No locking against reading these values from userspace, as for
+ * the CPU loadavg code.
+ */
+static void spu_calc_load(unsigned long ticks)
+{
+	unsigned long active_tasks; /* fixed-point */
+	static int count = LOAD_FREQ;
+
+	count -= ticks;
+
+	if (unlikely(count < 0)) {
+		active_tasks = count_active_contexts() * FIXED_1;
+		do {
+			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);
+			count += LOAD_FREQ;
+		} while (count < 0);
 	}
+}
+
+static void spusched_wake(unsigned long data)
+{
+	mod_timer(&spusched_timer, jiffies + SPUSCHED_TICK);
+	wake_up_process(spusched_task);
+	spu_calc_load(SPUSCHED_TICK);
+}
+
+static int spusched_thread(void *unused)
+{
+	struct spu *spu, *next;
+	int node;
+
+	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]);
+		}
+	}
+
+	return 0;
+}
+
+#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)
+{
+	struct proc_dir_entry *entry;
+	int err = -ENOMEM, i;
+
+	spu_prio = kzalloc(sizeof(struct spu_prio_array), GFP_KERNEL);
+	if (!spu_prio)
+		goto out;
+
 	for (i = 0; i < MAX_PRIO; i++) {
 		INIT_LIST_HEAD(&spu_prio->runq[i]);
 		__clear_bit(i, spu_prio->bitmap);
@@ -492,7 +757,30 @@ int __init spu_sched_init(void)
 		INIT_LIST_HEAD(&spu_prio->active_list[i]);
 	}
 	spin_lock_init(&spu_prio->runq_lock);
+
+	setup_timer(&spusched_timer, spusched_wake, 0);
+
+	spusched_task = kthread_run(spusched_thread, NULL, "spusched");
+	if (IS_ERR(spusched_task)) {
+		err = PTR_ERR(spusched_task);
+		goto out_free_spu_prio;
+	}
+
+	entry = create_proc_entry("spu_loadavg", 0, NULL);
+	if (!entry)
+		goto out_stop_kthread;
+	entry->proc_fops = &spu_loadavg_fops;
+
+	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)
@@ -500,6 +788,11 @@ void __exit spu_sched_exit(void)
 	struct spu *spu, *tmp;
 	int node;
 
+	remove_proc_entry("spu_loadavg", NULL);
+
+	del_timer_sync(&spusched_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],
@@ -510,5 +803,4 @@ 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/spu_restore.c b/arch/powerpc/platforms/cell/spufs/spu_restore.c
index 0bf723dcd67..4e19ed7a075 100644
--- a/arch/powerpc/platforms/cell/spufs/spu_restore.c
+++ b/arch/powerpc/platforms/cell/spufs/spu_restore.c
@@ -296,7 +296,7 @@ static inline void restore_complete(void)
  * This code deviates from the documented sequence in the
  * following aspects:
  *
- * 	1. The EA for LSCSA is passed from PPE in the
+ *	1. The EA for LSCSA is passed from PPE in the
  *	   signal notification channels.
  *	2. The register spill area is pulled by SPU
  *	   into LS, rather than pushed by PPE.
diff --git a/arch/powerpc/platforms/cell/spufs/spu_save.c b/arch/powerpc/platforms/cell/spufs/spu_save.c
index 196033b8a57..ae95cc1701e 100644
--- a/arch/powerpc/platforms/cell/spufs/spu_save.c
+++ b/arch/powerpc/platforms/cell/spufs/spu_save.c
@@ -44,7 +44,7 @@ static inline void save_event_mask(void)
 	 *    Read the SPU_RdEventMsk channel and save to the LSCSA.
 	 */
 	offset = LSCSA_QW_OFFSET(event_mask);
-	regs_spill[offset].slot[0] = spu_readch(SPU_RdEventStatMask);
+	regs_spill[offset].slot[0] = spu_readch(SPU_RdEventMask);
 }
 
 static inline void save_tag_mask(void)
diff --git a/arch/powerpc/platforms/cell/spufs/spufs.h b/arch/powerpc/platforms/cell/spufs/spufs.h
index 47617e8014a..08b3530288a 100644
--- a/arch/powerpc/platforms/cell/spufs/spufs.h
+++ b/arch/powerpc/platforms/cell/spufs/spufs.h
@@ -26,6 +26,7 @@
 #include <linux/mutex.h>
 #include <linux/spinlock.h>
 #include <linux/fs.h>
+#include <linux/cpumask.h>
 
 #include <asm/spu.h>
 #include <asm/spu_csa.h>
@@ -39,9 +40,17 @@ enum {
 struct spu_context_ops;
 struct spu_gang;
 
-/* ctx->sched_flags */
-enum {
-	SPU_SCHED_EXITING = 0,
+/*
+ * This is the state for spu utilization reporting to userspace.
+ * Because this state is visible to userspace it must never change and needs
+ * to be kept strictly separate from any internal state kept by the kernel.
+ */
+enum spuctx_execution_state {
+	SPUCTX_UTIL_USER = 0,
+	SPUCTX_UTIL_SYSTEM,
+	SPUCTX_UTIL_IOWAIT,
+	SPUCTX_UTIL_LOADED,
+	SPUCTX_UTIL_MAX
 };
 
 struct spu_context {
@@ -81,13 +90,34 @@ struct spu_context {
 	struct list_head gang_list;
 	struct spu_gang *gang;
 
+	/* owner thread */
+	pid_t tid;
+
 	/* scheduler fields */
- 	struct list_head rq;
-	struct delayed_work sched_work;
+	struct list_head rq;
+	unsigned int time_slice;
 	unsigned long sched_flags;
-	unsigned long rt_priority;
+	cpumask_t cpus_allowed;
 	int policy;
 	int prio;
+
+	/* statistics */
+	struct {
+		/* updates protected by ctx->state_mutex */
+		enum spuctx_execution_state execution_state;
+		unsigned long tstamp;		/* time of last ctx switch */
+		unsigned long times[SPUCTX_UTIL_MAX];
+		unsigned long long vol_ctx_switch;
+		unsigned long long invol_ctx_switch;
+		unsigned long long min_flt;
+		unsigned long long maj_flt;
+		unsigned long long hash_flt;
+		unsigned long long slb_flt;
+		unsigned long long slb_flt_base; /* # at last ctx switch */
+		unsigned long long class2_intr;
+		unsigned long long class2_intr_base; /* # at last ctx switch */
+		unsigned long long libassist;
+	} stats;
 };
 
 struct spu_gang {
@@ -177,6 +207,7 @@ void spu_gang_add_ctx(struct spu_gang *gang, struct spu_context *ctx);
 int spufs_handle_class1(struct spu_context *ctx);
 
 /* context management */
+extern atomic_t nr_spu_contexts;
 static inline void spu_acquire(struct spu_context *ctx)
 {
 	mutex_lock(&ctx->state_mutex);
@@ -200,9 +231,9 @@ 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);
+void spu_set_timeslice(struct spu_context *ctx);
+void spu_update_sched_info(struct spu_context *ctx);
+void __spu_update_sched_info(struct spu_context *ctx);
 int __init spu_sched_init(void);
 void __exit spu_sched_exit(void);
 
@@ -210,7 +241,7 @@ extern char *isolated_loader;
 
 /*
  * spufs_wait
- * 	Same as wait_event_interruptible(), except that here
+ *	Same as wait_event_interruptible(), except that here
  *	we need to call spu_release(ctx) before sleeping, and
  *	then spu_acquire(ctx) when awoken.
  */
@@ -256,4 +287,37 @@ struct spufs_coredump_reader {
 extern struct spufs_coredump_reader spufs_coredump_read[];
 extern int spufs_coredump_num_notes;
 
+/*
+ * This function is a little bit too large for an inline, but
+ * as fault.c is built into the kernel we can't move it out of
+ * line.
+ */
+static inline void spuctx_switch_state(struct spu_context *ctx,
+		enum spuctx_execution_state new_state)
+{
+	WARN_ON(!mutex_is_locked(&ctx->state_mutex));
+
+	if (ctx->stats.execution_state != new_state) {
+		unsigned long curtime = jiffies;
+
+		ctx->stats.times[ctx->stats.execution_state] +=
+				 curtime - ctx->stats.tstamp;
+		ctx->stats.tstamp = curtime;
+		ctx->stats.execution_state = new_state;
+	}
+}
+
+static inline void spu_switch_state(struct spu *spu,
+		enum spuctx_execution_state new_state)
+{
+	if (spu->stats.utilization_state != new_state) {
+		unsigned long curtime = jiffies;
+
+		spu->stats.times[spu->stats.utilization_state] +=
+				 curtime - spu->stats.tstamp;
+		spu->stats.tstamp = curtime;
+		spu->stats.utilization_state = new_state;
+	}
+}
+
 #endif
diff --git a/arch/powerpc/platforms/cell/spufs/switch.c b/arch/powerpc/platforms/cell/spufs/switch.c
index 71a0b41adb8..9c506ba08cd 100644
--- a/arch/powerpc/platforms/cell/spufs/switch.c
+++ b/arch/powerpc/platforms/cell/spufs/switch.c
@@ -70,7 +70,7 @@
   }
 #endif				/* debug */
 
-#define POLL_WHILE_FALSE(_c) 	POLL_WHILE_TRUE(!(_c))
+#define POLL_WHILE_FALSE(_c)	POLL_WHILE_TRUE(!(_c))
 
 static inline void acquire_spu_lock(struct spu *spu)
 {
@@ -387,6 +387,19 @@ static inline void save_ppu_querytype(struct spu_state *csa, struct spu *spu)
 	csa->prob.dma_querytype_RW = in_be32(&prob->dma_querytype_RW);
 }
 
+static inline void save_ppu_tagstatus(struct spu_state *csa, struct spu *spu)
+{
+	struct spu_problem __iomem *prob = spu->problem;
+
+	/* Save the Prxy_TagStatus register in the CSA.
+	 *
+	 * It is unnecessary to restore dma_tagstatus_R, however,
+	 * dma_tagstatus_R in the CSA is accessed via backing_ops, so
+	 * we must save it.
+	 */
+	csa->prob.dma_tagstatus_R = in_be32(&prob->dma_tagstatus_R);
+}
+
 static inline void save_mfc_csr_tsq(struct spu_state *csa, struct spu *spu)
 {
 	struct spu_priv2 __iomem *priv2 = spu->priv2;
@@ -1812,6 +1825,7 @@ static void save_csa(struct spu_state *prev, struct spu *spu)
 	save_mfc_queues(prev, spu);	/* Step 19. */
 	save_ppu_querymask(prev, spu);	/* Step 20. */
 	save_ppu_querytype(prev, spu);	/* Step 21. */
+	save_ppu_tagstatus(prev, spu);  /* NEW.     */
 	save_mfc_csr_tsq(prev, spu);	/* Step 22. */
 	save_mfc_csr_cmd(prev, spu);	/* Step 23. */
 	save_mfc_csr_ato(prev, spu);	/* Step 24. */
@@ -1930,7 +1944,7 @@ static void harvest(struct spu_state *prev, struct spu *spu)
 	reset_spu_privcntl(prev, spu);	        /* Step 16. */
 	reset_spu_lslr(prev, spu);              /* Step 17. */
 	setup_mfc_sr1(prev, spu);	        /* Step 18. */
-	spu_invalidate_slbs(spu);        	/* Step 19. */
+	spu_invalidate_slbs(spu);		/* Step 19. */
 	reset_ch_part1(prev, spu);	        /* Step 20. */
 	reset_ch_part2(prev, spu);	        /* Step 21. */
 	enable_interrupts(prev, spu);	        /* Step 22. */