[CELL] oprofile: add support to OProfile for profiling CELL BE SPUs

From: Maynard Johnson <mpjohn@us.ibm.com>

This patch updates the existing arch/powerpc/oprofile/op_model_cell.c
to add in the SPU profiling capabilities.  In addition, a 'cell' subdirectory
was added to arch/powerpc/oprofile to hold Cell-specific SPU profiling code.
Exports spu_set_profile_private_kref and spu_get_profile_private_kref which
are used by OProfile to store private profile information in spufs data
structures.

Also incorporated several fixes from other patches (rrn).  Check pointer
returned from kzalloc.  Eliminated unnecessary cast.  Better error
handling and cleanup in the related area.  64-bit unsigned long parameter
was being demoted to 32-bit unsigned int and eventually promoted back to
unsigned long.

Signed-off-by: Carl Love <carll@us.ibm.com>
Signed-off-by: Maynard Johnson <mpjohn@us.ibm.com>
Signed-off-by: Bob Nelson <rrnelson@us.ibm.com>
Signed-off-by: Arnd Bergmann <arnd.bergmann@de.ibm.com>
Acked-by: Paul Mackerras <paulus@samba.org>

4 files changed, 1089 insertions, 0 deletions

diff --git a/arch/powerpc/oprofile/cell/pr_util.h b/arch/powerpc/oprofile/cell/pr_util.h
new file mode 100644
index 00000000000..e5704f00c8b
--- /dev/null
+++ b/arch/powerpc/oprofile/cell/pr_util.h
@@ -0,0 +1,97 @@
+ /*
+ * Cell Broadband Engine OProfile Support
+ *
+ * (C) Copyright IBM Corporation 2006
+ *
+ * Author: Maynard Johnson <maynardj@us.ibm.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ */
+
+#ifndef PR_UTIL_H
+#define PR_UTIL_H
+
+#include <linux/cpumask.h>
+#include <linux/oprofile.h>
+#include <asm/cell-pmu.h>
+#include <asm/spu.h>
+
+#include "../../platforms/cell/cbe_regs.h"
+
+/* Defines used for sync_start */
+#define SKIP_GENERIC_SYNC 0
+#define SYNC_START_ERROR -1
+#define DO_GENERIC_SYNC 1
+
+struct spu_overlay_info {	/* map of sections within an SPU overlay */
+	unsigned int vma;	/* SPU virtual memory address from elf */
+	unsigned int size;	/* size of section from elf */
+	unsigned int offset;	/* offset of section into elf file */
+	unsigned int buf;
+};
+
+struct vma_to_fileoffset_map {	/* map of sections within an SPU program */
+	struct vma_to_fileoffset_map *next;	/* list pointer */
+	unsigned int vma;	/* SPU virtual memory address from elf */
+	unsigned int size;	/* size of section from elf */
+	unsigned int offset;	/* offset of section into elf file */
+	unsigned int guard_ptr;
+	unsigned int guard_val;
+        /*
+	 * The guard pointer is an entry in the _ovly_buf_table,
+	 * computed using ovly.buf as the index into the table.  Since
+	 * ovly.buf values begin at '1' to reference the first (or 0th)
+	 * entry in the _ovly_buf_table, the computation subtracts 1
+	 * from ovly.buf.
+	 * The guard value is stored in the _ovly_buf_table entry and
+	 * is an index (starting at 1) back to the _ovly_table entry
+	 * that is pointing at this _ovly_buf_table entry.  So, for
+	 * example, for an overlay scenario with one overlay segment
+	 * and two overlay sections:
+	 *      - Section 1 points to the first entry of the
+	 *        _ovly_buf_table, which contains a guard value
+	 *        of '1', referencing the first (index=0) entry of
+	 *        _ovly_table.
+	 *      - Section 2 points to the second entry of the
+	 *        _ovly_buf_table, which contains a guard value
+	 *        of '2', referencing the second (index=1) entry of
+	 *        _ovly_table.
+	 */
+
+};
+
+/* The three functions below are for maintaining and accessing
+ * the vma-to-fileoffset map.
+ */
+struct vma_to_fileoffset_map *create_vma_map(const struct spu *spu,
+					     u64 objectid);
+unsigned int vma_map_lookup(struct vma_to_fileoffset_map *map,
+			    unsigned int vma, const struct spu *aSpu,
+			    int *grd_val);
+void vma_map_free(struct vma_to_fileoffset_map *map);
+
+/*
+ * Entry point for SPU profiling.
+ * cycles_reset is the SPU_CYCLES count value specified by the user.
+ */
+int start_spu_profiling(unsigned int cycles_reset);
+
+void stop_spu_profiling(void);
+
+
+/* add the necessary profiling hooks */
+int spu_sync_start(void);
+
+/* remove the hooks */
+int spu_sync_stop(void);
+
+/* Record SPU program counter samples to the oprofile event buffer. */
+void spu_sync_buffer(int spu_num, unsigned int *samples,
+		     int num_samples);
+
+void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int cycles_reset);
+
+#endif	  /* PR_UTIL_H */
diff --git a/arch/powerpc/oprofile/cell/spu_profiler.c b/arch/powerpc/oprofile/cell/spu_profiler.c
new file mode 100644
index 00000000000..380d7e21753
--- /dev/null
+++ b/arch/powerpc/oprofile/cell/spu_profiler.c
@@ -0,0 +1,221 @@
+/*
+ * Cell Broadband Engine OProfile Support
+ *
+ * (C) Copyright IBM Corporation 2006
+ *
+ * Authors: Maynard Johnson <maynardj@us.ibm.com>
+ *	    Carl Love <carll@us.ibm.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ */
+
+#include <linux/hrtimer.h>
+#include <linux/smp.h>
+#include <linux/slab.h>
+#include <asm/cell-pmu.h>
+#include "pr_util.h"
+
+#define TRACE_ARRAY_SIZE 1024
+#define SCALE_SHIFT 14
+
+static u32 *samples;
+
+static int spu_prof_running;
+static unsigned int profiling_interval;
+
+#define NUM_SPU_BITS_TRBUF 16
+#define SPUS_PER_TB_ENTRY   4
+#define SPUS_PER_NODE	     8
+
+#define SPU_PC_MASK	     0xFFFF
+
+static DEFINE_SPINLOCK(sample_array_lock);
+unsigned long sample_array_lock_flags;
+
+void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int cycles_reset)
+{
+	unsigned long ns_per_cyc;
+
+	if (!freq_khz)
+		freq_khz = ppc_proc_freq/1000;
+
+	/* To calculate a timeout in nanoseconds, the basic
+	 * formula is ns = cycles_reset * (NSEC_PER_SEC / cpu frequency).
+	 * To avoid floating point math, we use the scale math
+	 * technique as described in linux/jiffies.h.  We use
+	 * a scale factor of SCALE_SHIFT, which provides 4 decimal places
+	 * of precision.  This is close enough for the purpose at hand.
+	 *
+	 * The value of the timeout should be small enough that the hw
+	 * trace buffer will not get more then about 1/3 full for the
+	 * maximum user specified (the LFSR value) hw sampling frequency.
+	 * This is to ensure the trace buffer will never fill even if the
+	 * kernel thread scheduling varies under a heavy system load.
+	 */
+
+	ns_per_cyc = (USEC_PER_SEC << SCALE_SHIFT)/freq_khz;
+	profiling_interval = (ns_per_cyc * cycles_reset) >> SCALE_SHIFT;
+
+}
+
+/*
+ * Extract SPU PC from trace buffer entry
+ */
+static void spu_pc_extract(int cpu, int entry)
+{
+	/* the trace buffer is 128 bits */
+	u64 trace_buffer[2];
+	u64 spu_mask;
+	int spu;
+
+	spu_mask = SPU_PC_MASK;
+
+	/* Each SPU PC is 16 bits; hence, four spus in each of
+	 * the two 64-bit buffer entries that make up the
+	 * 128-bit trace_buffer entry.	Process two 64-bit values
+	 * simultaneously.
+	 * trace[0] SPU PC contents are: 0 1 2 3
+	 * trace[1] SPU PC contents are: 4 5 6 7
+	 */
+
+	cbe_read_trace_buffer(cpu, trace_buffer);
+
+	for (spu = SPUS_PER_TB_ENTRY-1; spu >= 0; spu--) {
+		/* spu PC trace entry is upper 16 bits of the
+		 * 18 bit SPU program counter
+		 */
+		samples[spu * TRACE_ARRAY_SIZE + entry]
+			= (spu_mask & trace_buffer[0]) << 2;
+		samples[(spu + SPUS_PER_TB_ENTRY) * TRACE_ARRAY_SIZE + entry]
+			= (spu_mask & trace_buffer[1]) << 2;
+
+		trace_buffer[0] = trace_buffer[0] >> NUM_SPU_BITS_TRBUF;
+		trace_buffer[1] = trace_buffer[1] >> NUM_SPU_BITS_TRBUF;
+	}
+}
+
+static int cell_spu_pc_collection(int cpu)
+{
+	u32 trace_addr;
+	int entry;
+
+	/* process the collected SPU PC for the node */
+
+	entry = 0;
+
+	trace_addr = cbe_read_pm(cpu, trace_address);
+	while (!(trace_addr & CBE_PM_TRACE_BUF_EMPTY)) {
+		/* there is data in the trace buffer to process */
+		spu_pc_extract(cpu, entry);
+
+		entry++;
+
+		if (entry >= TRACE_ARRAY_SIZE)
+			/* spu_samples is full */
+			break;
+
+		trace_addr = cbe_read_pm(cpu, trace_address);
+	}
+
+	return entry;
+}
+
+
+static enum hrtimer_restart profile_spus(struct hrtimer *timer)
+{
+	ktime_t kt;
+	int cpu, node, k, num_samples, spu_num;
+
+	if (!spu_prof_running)
+		goto stop;
+
+	for_each_online_cpu(cpu) {
+		if (cbe_get_hw_thread_id(cpu))
+			continue;
+
+		node = cbe_cpu_to_node(cpu);
+
+		/* There should only be one kernel thread at a time processing
+		 * the samples.	 In the very unlikely case that the processing
+		 * is taking a very long time and multiple kernel threads are
+		 * started to process the samples.  Make sure only one kernel
+		 * thread is working on the samples array at a time.  The
+		 * sample array must be loaded and then processed for a given
+		 * cpu.	 The sample array is not per cpu.
+		 */
+		spin_lock_irqsave(&sample_array_lock,
+				  sample_array_lock_flags);
+		num_samples = cell_spu_pc_collection(cpu);
+
+		if (num_samples == 0) {
+			spin_unlock_irqrestore(&sample_array_lock,
+					       sample_array_lock_flags);
+			continue;
+		}
+
+		for (k = 0; k < SPUS_PER_NODE; k++) {
+			spu_num = k + (node * SPUS_PER_NODE);
+			spu_sync_buffer(spu_num,
+					samples + (k * TRACE_ARRAY_SIZE),
+					num_samples);
+		}
+
+		spin_unlock_irqrestore(&sample_array_lock,
+				       sample_array_lock_flags);
+
+	}
+	smp_wmb();	/* insure spu event buffer updates are written */
+			/* don't want events intermingled... */
+
+	kt = ktime_set(0, profiling_interval);
+	if (!spu_prof_running)
+		goto stop;
+	hrtimer_forward(timer, timer->base->get_time(), kt);
+	return HRTIMER_RESTART;
+
+ stop:
+	printk(KERN_INFO "SPU_PROF: spu-prof timer ending\n");
+	return HRTIMER_NORESTART;
+}
+
+static struct hrtimer timer;
+/*
+ * Entry point for SPU profiling.
+ * NOTE:  SPU profiling is done system-wide, not per-CPU.
+ *
+ * cycles_reset is the count value specified by the user when
+ * setting up OProfile to count SPU_CYCLES.
+ */
+int start_spu_profiling(unsigned int cycles_reset)
+{
+	ktime_t kt;
+
+	pr_debug("timer resolution: %lu\n", TICK_NSEC);
+	kt = ktime_set(0, profiling_interval);
+	hrtimer_init(&timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
+	timer.expires = kt;
+	timer.function = profile_spus;
+
+	/* Allocate arrays for collecting SPU PC samples */
+	samples = kzalloc(SPUS_PER_NODE *
+			  TRACE_ARRAY_SIZE * sizeof(u32), GFP_KERNEL);
+
+	if (!samples)
+		return -ENOMEM;
+
+	spu_prof_running = 1;
+	hrtimer_start(&timer, kt, HRTIMER_MODE_REL);
+
+	return 0;
+}
+
+void stop_spu_profiling(void)
+{
+	spu_prof_running = 0;
+	hrtimer_cancel(&timer);
+	kfree(samples);
+	pr_debug("SPU_PROF: stop_spu_profiling issued\n");
+}
diff --git a/arch/powerpc/oprofile/cell/spu_task_sync.c b/arch/powerpc/oprofile/cell/spu_task_sync.c
new file mode 100644
index 00000000000..133665754a7
--- /dev/null
+++ b/arch/powerpc/oprofile/cell/spu_task_sync.c
@@ -0,0 +1,484 @@
+/*
+ * Cell Broadband Engine OProfile Support
+ *
+ * (C) Copyright IBM Corporation 2006
+ *
+ * Author: Maynard Johnson <maynardj@us.ibm.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ */
+
+/* The purpose of this file is to handle SPU event task switching
+ * and to record SPU context information into the OProfile
+ * event buffer.
+ *
+ * Additionally, the spu_sync_buffer function is provided as a helper
+ * for recoding actual SPU program counter samples to the event buffer.
+ */
+#include <linux/dcookies.h>
+#include <linux/kref.h>
+#include <linux/mm.h>
+#include <linux/module.h>
+#include <linux/notifier.h>
+#include <linux/numa.h>
+#include <linux/oprofile.h>
+#include <linux/spinlock.h>
+#include "pr_util.h"
+
+#define RELEASE_ALL 9999
+
+static DEFINE_SPINLOCK(buffer_lock);
+static DEFINE_SPINLOCK(cache_lock);
+static int num_spu_nodes;
+int spu_prof_num_nodes;
+int last_guard_val[MAX_NUMNODES * 8];
+
+/* Container for caching information about an active SPU task. */
+struct cached_info {
+	struct vma_to_fileoffset_map *map;
+	struct spu *the_spu;	/* needed to access pointer to local_store */
+	struct kref cache_ref;
+};
+
+static struct cached_info *spu_info[MAX_NUMNODES * 8];
+
+static void destroy_cached_info(struct kref *kref)
+{
+	struct cached_info *info;
+
+	info = container_of(kref, struct cached_info, cache_ref);
+	vma_map_free(info->map);
+	kfree(info);
+	module_put(THIS_MODULE);
+}
+
+/* Return the cached_info for the passed SPU number.
+ * ATTENTION:  Callers are responsible for obtaining the
+ *	       cache_lock if needed prior to invoking this function.
+ */
+static struct cached_info *get_cached_info(struct spu *the_spu, int spu_num)
+{
+	struct kref *ref;
+	struct cached_info *ret_info;
+
+	if (spu_num >= num_spu_nodes) {
+		printk(KERN_ERR "SPU_PROF: "
+		       "%s, line %d: Invalid index %d into spu info cache\n",
+		       __FUNCTION__, __LINE__, spu_num);
+		ret_info = NULL;
+		goto out;
+	}
+	if (!spu_info[spu_num] && the_spu) {
+		ref = spu_get_profile_private_kref(the_spu->ctx);
+		if (ref) {
+			spu_info[spu_num] = container_of(ref, struct cached_info, cache_ref);
+			kref_get(&spu_info[spu_num]->cache_ref);
+		}
+	}
+
+	ret_info = spu_info[spu_num];
+ out:
+	return ret_info;
+}
+
+
+/* Looks for cached info for the passed spu.  If not found, the
+ * cached info is created for the passed spu.
+ * Returns 0 for success; otherwise, -1 for error.
+ */
+static int
+prepare_cached_spu_info(struct spu *spu, unsigned long objectId)
+{
+	unsigned long flags;
+	struct vma_to_fileoffset_map *new_map;
+	int retval = 0;
+	struct cached_info *info;
+
+	/* We won't bother getting cache_lock here since
+	 * don't do anything with the cached_info that's returned.
+	 */
+	info = get_cached_info(spu, spu->number);
+
+	if (info) {
+		pr_debug("Found cached SPU info.\n");
+		goto out;
+	}
+
+	/* Create cached_info and set spu_info[spu->number] to point to it.
+	 * spu->number is a system-wide value, not a per-node value.
+	 */
+	info = kzalloc(sizeof(struct cached_info), GFP_KERNEL);
+	if (!info) {
+		printk(KERN_ERR "SPU_PROF: "
+		       "%s, line %d: create vma_map failed\n",
+		       __FUNCTION__, __LINE__);
+		retval = -ENOMEM;
+		goto err_alloc;
+	}
+	new_map = create_vma_map(spu, objectId);
+	if (!new_map) {
+		printk(KERN_ERR "SPU_PROF: "
+		       "%s, line %d: create vma_map failed\n",
+		       __FUNCTION__, __LINE__);
+		retval = -ENOMEM;
+		goto err_alloc;
+	}
+
+	pr_debug("Created vma_map\n");
+	info->map = new_map;
+	info->the_spu = spu;
+	kref_init(&info->cache_ref);
+	spin_lock_irqsave(&cache_lock, flags);
+	spu_info[spu->number] = info;
+	/* Increment count before passing off ref to SPUFS. */
+	kref_get(&info->cache_ref);
+
+	/* We increment the module refcount here since SPUFS is
+	 * responsible for the final destruction of the cached_info,
+	 * and it must be able to access the destroy_cached_info()
+	 * function defined in the OProfile module.  We decrement
+	 * the module refcount in destroy_cached_info.
+	 */
+	try_module_get(THIS_MODULE);
+	spu_set_profile_private_kref(spu->ctx, &info->cache_ref,
+				destroy_cached_info);
+	spin_unlock_irqrestore(&cache_lock, flags);
+	goto out;
+
+err_alloc:
+	kfree(info);
+out:
+	return retval;
+}
+
+/*
+ * NOTE:  The caller is responsible for locking the
+ *	  cache_lock prior to calling this function.
+ */
+static int release_cached_info(int spu_index)
+{
+	int index, end;
+
+	if (spu_index == RELEASE_ALL) {
+		end = num_spu_nodes;
+		index = 0;
+	} else {
+		if (spu_index >= num_spu_nodes) {
+			printk(KERN_ERR "SPU_PROF: "
+				"%s, line %d: "
+				"Invalid index %d into spu info cache\n",
+				__FUNCTION__, __LINE__, spu_index);
+			goto out;
+		}
+		end = spu_index + 1;
+		index = spu_index;
+	}
+	for (; index < end; index++) {
+		if (spu_info[index]) {
+			kref_put(&spu_info[index]->cache_ref,
+				 destroy_cached_info);
+			spu_info[index] = NULL;
+		}
+	}
+
+out:
+	return 0;
+}
+
+/* The source code for fast_get_dcookie was "borrowed"
+ * from drivers/oprofile/buffer_sync.c.
+ */
+
+/* Optimisation. We can manage without taking the dcookie sem
+ * because we cannot reach this code without at least one
+ * dcookie user still being registered (namely, the reader
+ * of the event buffer).
+ */
+static inline unsigned long fast_get_dcookie(struct dentry *dentry,
+					     struct vfsmount *vfsmnt)
+{
+	unsigned long cookie;
+
+	if (dentry->d_cookie)
+		return (unsigned long)dentry;
+	get_dcookie(dentry, vfsmnt, &cookie);
+	return cookie;
+}
+
+/* Look up the dcookie for the task's first VM_EXECUTABLE mapping,
+ * which corresponds loosely to "application name". Also, determine
+ * the offset for the SPU ELF object.  If computed offset is
+ * non-zero, it implies an embedded SPU object; otherwise, it's a
+ * separate SPU binary, in which case we retrieve it's dcookie.
+ * For the embedded case, we must determine if SPU ELF is embedded
+ * in the executable application or another file (i.e., shared lib).
+ * If embedded in a shared lib, we must get the dcookie and return
+ * that to the caller.
+ */
+static unsigned long
+get_exec_dcookie_and_offset(struct spu *spu, unsigned int *offsetp,
+			    unsigned long *spu_bin_dcookie,
+			    unsigned long spu_ref)
+{
+	unsigned long app_cookie = 0;
+	unsigned int my_offset = 0;
+	struct file *app = NULL;
+	struct vm_area_struct *vma;
+	struct mm_struct *mm = spu->mm;
+
+	if (!mm)
+		goto out;
+
+	down_read(&mm->mmap_sem);
+
+	for (vma = mm->mmap; vma; vma = vma->vm_next) {
+		if (!vma->vm_file)
+			continue;
+		if (!(vma->vm_flags & VM_EXECUTABLE))
+			continue;
+		app_cookie = fast_get_dcookie(vma->vm_file->f_dentry,
+					  vma->vm_file->f_vfsmnt);
+		pr_debug("got dcookie for %s\n",
+			 vma->vm_file->f_dentry->d_name.name);
+		app = vma->vm_file;
+		break;
+	}
+
+	for (vma = mm->mmap; vma; vma = vma->vm_next) {
+		if (vma->vm_start > spu_ref || vma->vm_end <= spu_ref)
+			continue;
+		my_offset = spu_ref - vma->vm_start;
+		if (!vma->vm_file)
+			goto fail_no_image_cookie;
+
+		pr_debug("Found spu ELF at %X(object-id:%lx) for file %s\n",
+			 my_offset, spu_ref,
+			 vma->vm_file->f_dentry->d_name.name);
+		*offsetp = my_offset;
+		break;
+	}
+
+	*spu_bin_dcookie = fast_get_dcookie(vma->vm_file->f_dentry,
+						 vma->vm_file->f_vfsmnt);
+	pr_debug("got dcookie for %s\n", vma->vm_file->f_dentry->d_name.name);
+
+	up_read(&mm->mmap_sem);
+
+out:
+	return app_cookie;
+
+fail_no_image_cookie:
+	up_read(&mm->mmap_sem);
+
+	printk(KERN_ERR "SPU_PROF: "
+		"%s, line %d: Cannot find dcookie for SPU binary\n",
+		__FUNCTION__, __LINE__);
+	goto out;
+}
+
+
+
+/* This function finds or creates cached context information for the
+ * passed SPU and records SPU context information into the OProfile
+ * event buffer.
+ */
+static int process_context_switch(struct spu *spu, unsigned long objectId)
+{
+	unsigned long flags;
+	int retval;
+	unsigned int offset = 0;
+	unsigned long spu_cookie = 0, app_dcookie;
+
+	retval = prepare_cached_spu_info(spu, objectId);
+	if (retval)
+		goto out;
+
+	/* Get dcookie first because a mutex_lock is taken in that
+	 * code path, so interrupts must not be disabled.
+	 */
+	app_dcookie = get_exec_dcookie_and_offset(spu, &offset, &spu_cookie, objectId);
+	if (!app_dcookie || !spu_cookie) {
+		retval  = -ENOENT;
+		goto out;
+	}
+
+	/* Record context info in event buffer */
+	spin_lock_irqsave(&buffer_lock, flags);
+	add_event_entry(ESCAPE_CODE);
+	add_event_entry(SPU_CTX_SWITCH_CODE);
+	add_event_entry(spu->number);
+	add_event_entry(spu->pid);
+	add_event_entry(spu->tgid);
+	add_event_entry(app_dcookie);
+	add_event_entry(spu_cookie);
+	add_event_entry(offset);
+	spin_unlock_irqrestore(&buffer_lock, flags);
+	smp_wmb();	/* insure spu event buffer updates are written */
+			/* don't want entries intermingled... */
+out:
+	return retval;
+}
+
+/*
+ * This function is invoked on either a bind_context or unbind_context.
+ * If called for an unbind_context, the val arg is 0; otherwise,
+ * it is the object-id value for the spu context.
+ * The data arg is of type 'struct spu *'.
+ */
+static int spu_active_notify(struct notifier_block *self, unsigned long val,
+				void *data)
+{
+	int retval;
+	unsigned long flags;
+	struct spu *the_spu = data;
+
+	pr_debug("SPU event notification arrived\n");
+	if (!val) {
+		spin_lock_irqsave(&cache_lock, flags);
+		retval = release_cached_info(the_spu->number);
+		spin_unlock_irqrestore(&cache_lock, flags);
+	} else {
+		retval = process_context_switch(the_spu, val);
+	}
+	return retval;
+}
+
+static struct notifier_block spu_active = {
+	.notifier_call = spu_active_notify,
+};
+
+static int number_of_online_nodes(void)
+{
+        u32 cpu; u32 tmp;
+        int nodes = 0;
+        for_each_online_cpu(cpu) {
+                tmp = cbe_cpu_to_node(cpu) + 1;
+                if (tmp > nodes)
+                        nodes++;
+        }
+        return nodes;
+}
+
+/* The main purpose of this function is to synchronize
+ * OProfile with SPUFS by registering to be notified of
+ * SPU task switches.
+ *
+ * NOTE: When profiling SPUs, we must ensure that only
+ * spu_sync_start is invoked and not the generic sync_start
+ * in drivers/oprofile/oprof.c.	 A return value of
+ * SKIP_GENERIC_SYNC or SYNC_START_ERROR will
+ * accomplish this.
+ */
+int spu_sync_start(void)
+{
+	int k;
+	int ret = SKIP_GENERIC_SYNC;
+	int register_ret;
+	unsigned long flags = 0;
+
+	spu_prof_num_nodes = number_of_online_nodes();
+	num_spu_nodes = spu_prof_num_nodes * 8;
+
+	spin_lock_irqsave(&buffer_lock, flags);
+	add_event_entry(ESCAPE_CODE);
+	add_event_entry(SPU_PROFILING_CODE);
+	add_event_entry(num_spu_nodes);
+	spin_unlock_irqrestore(&buffer_lock, flags);
+
+	/* Register for SPU events  */
+	register_ret = spu_switch_event_register(&spu_active);
+	if (register_ret) {
+		ret = SYNC_START_ERROR;
+		goto out;
+	}
+
+	for (k = 0; k < (MAX_NUMNODES * 8); k++)
+		last_guard_val[k] = 0;
+	pr_debug("spu_sync_start -- running.\n");
+out:
+	return ret;
+}
+
+/* Record SPU program counter samples to the oprofile event buffer. */
+void spu_sync_buffer(int spu_num, unsigned int *samples,
+		     int num_samples)
+{
+	unsigned long long file_offset;
+	unsigned long flags;
+	int i;
+	struct vma_to_fileoffset_map *map;
+	struct spu *the_spu;
+	unsigned long long spu_num_ll = spu_num;
+	unsigned long long spu_num_shifted = spu_num_ll << 32;
+	struct cached_info *c_info;
+
+	/* We need to obtain the cache_lock here because it's
+	 * possible that after getting the cached_info, the SPU job
+	 * corresponding to this cached_info may end, thus resulting
+	 * in the destruction of the cached_info.
+	 */
+	spin_lock_irqsave(&cache_lock, flags);
+	c_info = get_cached_info(NULL, spu_num);
+	if (!c_info) {
+		/* This legitimately happens when the SPU task ends before all
+		 * samples are recorded.
+		 * No big deal -- so we just drop a few samples.
+		 */
+		pr_debug("SPU_PROF: No cached SPU contex "
+			  "for SPU #%d. Dropping samples.\n", spu_num);
+		goto out;
+	}
+
+	map = c_info->map;
+	the_spu = c_info->the_spu;
+	spin_lock(&buffer_lock);
+	for (i = 0; i < num_samples; i++) {
+		unsigned int sample = *(samples+i);
+		int grd_val = 0;
+		file_offset = 0;
+		if (sample == 0)
+			continue;
+		file_offset = vma_map_lookup( map, sample, the_spu, &grd_val);
+
+		/* If overlays are used by this SPU application, the guard
+		 * value is non-zero, indicating which overlay section is in
+		 * use.	 We need to discard samples taken during the time
+		 * period which an overlay occurs (i.e., guard value changes).
+		 */
+		if (grd_val && grd_val != last_guard_val[spu_num]) {
+			last_guard_val[spu_num] = grd_val;
+			/* Drop the rest of the samples. */
+			break;
+		}
+
+		add_event_entry(file_offset | spu_num_shifted);
+	}
+	spin_unlock(&buffer_lock);
+out:
+	spin_unlock_irqrestore(&cache_lock, flags);
+}
+
+
+int spu_sync_stop(void)
+{
+	unsigned long flags = 0;
+	int ret = spu_switch_event_unregister(&spu_active);
+	if (ret) {
+		printk(KERN_ERR "SPU_PROF: "
+			"%s, line %d: spu_switch_event_unregister returned %d\n",
+			__FUNCTION__, __LINE__, ret);
+		goto out;
+	}
+
+	spin_lock_irqsave(&cache_lock, flags);
+	ret = release_cached_info(RELEASE_ALL);
+	spin_unlock_irqrestore(&cache_lock, flags);
+out:
+	pr_debug("spu_sync_stop -- done.\n");
+	return ret;
+}
+
+
diff --git a/arch/powerpc/oprofile/cell/vma_map.c b/arch/powerpc/oprofile/cell/vma_map.c
new file mode 100644
index 00000000000..76ec1d16aef
--- /dev/null
+++ b/arch/powerpc/oprofile/cell/vma_map.c
@@ -0,0 +1,287 @@
+/*
+ * Cell Broadband Engine OProfile Support
+ *
+ * (C) Copyright IBM Corporation 2006
+ *
+ * Author: Maynard Johnson <maynardj@us.ibm.com>
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ */
+
+/* The code in this source file is responsible for generating
+ * vma-to-fileOffset maps for both overlay and non-overlay SPU
+ * applications.
+ */
+
+#include <linux/mm.h>
+#include <linux/string.h>
+#include <linux/uaccess.h>
+#include <linux/elf.h>
+#include "pr_util.h"
+
+
+void vma_map_free(struct vma_to_fileoffset_map *map)
+{
+	while (map) {
+		struct vma_to_fileoffset_map *next = map->next;
+		kfree(map);
+		map = next;
+	}
+}
+
+unsigned int
+vma_map_lookup(struct vma_to_fileoffset_map *map, unsigned int vma,
+	       const struct spu *aSpu, int *grd_val)
+{
+	/*
+	 * Default the offset to the physical address + a flag value.
+	 * Addresses of dynamically generated code can't be found in the vma
+	 * map.  For those addresses the flagged value will be sent on to
+	 * the user space tools so they can be reported rather than just
+	 * thrown away.
+	 */
+	u32 offset = 0x10000000 + vma;
+	u32 ovly_grd;
+
+	for (; map; map = map->next) {
+		if (vma < map->vma || vma >= map->vma + map->size)
+			continue;
+
+		if (map->guard_ptr) {
+			ovly_grd = *(u32 *)(aSpu->local_store + map->guard_ptr);
+			if (ovly_grd != map->guard_val)
+				continue;
+			*grd_val = ovly_grd;
+		}
+		offset = vma - map->vma + map->offset;
+		break;
+	}
+
+	return offset;
+}
+
+static struct vma_to_fileoffset_map *
+vma_map_add(struct vma_to_fileoffset_map *map, unsigned int vma,
+	    unsigned int size, unsigned int offset, unsigned int guard_ptr,
+	    unsigned int guard_val)
+{
+	struct vma_to_fileoffset_map *new =
+		kzalloc(sizeof(struct vma_to_fileoffset_map), GFP_KERNEL);
+	if (!new) {
+		printk(KERN_ERR "SPU_PROF: %s, line %d: malloc failed\n",
+		       __FUNCTION__, __LINE__);
+		vma_map_free(map);
+		return NULL;
+	}
+
+	new->next = map;
+	new->vma = vma;
+	new->size = size;
+	new->offset = offset;
+	new->guard_ptr = guard_ptr;
+	new->guard_val = guard_val;
+
+	return new;
+}
+
+
+/* Parse SPE ELF header and generate a list of vma_maps.
+ * A pointer to the first vma_map in the generated list
+ * of vma_maps is returned.  */
+struct vma_to_fileoffset_map *create_vma_map(const struct spu *aSpu,
+					     unsigned long spu_elf_start)
+{
+	static const unsigned char expected[EI_PAD] = {
+		[EI_MAG0] = ELFMAG0,
+		[EI_MAG1] = ELFMAG1,
+		[EI_MAG2] = ELFMAG2,
+		[EI_MAG3] = ELFMAG3,
+		[EI_CLASS] = ELFCLASS32,
+		[EI_DATA] = ELFDATA2MSB,
+		[EI_VERSION] = EV_CURRENT,
+		[EI_OSABI] = ELFOSABI_NONE
+	};
+
+	int grd_val;
+	struct vma_to_fileoffset_map *map = NULL;
+	struct spu_overlay_info ovly;
+	unsigned int overlay_tbl_offset = -1;
+	unsigned long phdr_start, shdr_start;
+	Elf32_Ehdr ehdr;
+	Elf32_Phdr phdr;
+	Elf32_Shdr shdr, shdr_str;
+	Elf32_Sym sym;
+	int i, j;
+	char name[32];
+
+	unsigned int ovly_table_sym = 0;
+	unsigned int ovly_buf_table_sym = 0;
+	unsigned int ovly_table_end_sym = 0;
+	unsigned int ovly_buf_table_end_sym = 0;
+	unsigned long ovly_table;
+	unsigned int n_ovlys;
+
+	/* Get and validate ELF header.	 */
+
+	if (copy_from_user(&ehdr, (void *) spu_elf_start, sizeof (ehdr)))
+		goto fail;
+
+	if (memcmp(ehdr.e_ident, expected, EI_PAD) != 0) {
+		printk(KERN_ERR "SPU_PROF: "
+		       "%s, line %d: Unexpected e_ident parsing SPU ELF\n",
+		       __FUNCTION__, __LINE__);
+		goto fail;
+	}
+	if (ehdr.e_machine != EM_SPU) {
+		printk(KERN_ERR "SPU_PROF: "
+		       "%s, line %d: Unexpected e_machine parsing SPU ELF\n",
+		       __FUNCTION__,  __LINE__);
+		goto fail;
+	}
+	if (ehdr.e_type != ET_EXEC) {
+		printk(KERN_ERR "SPU_PROF: "
+		       "%s, line %d: Unexpected e_type parsing SPU ELF\n",
+		       __FUNCTION__, __LINE__);
+		goto fail;
+	}
+	phdr_start = spu_elf_start + ehdr.e_phoff;
+	shdr_start = spu_elf_start + ehdr.e_shoff;
+
+	/* Traverse program headers.  */
+	for (i = 0; i < ehdr.e_phnum; i++) {
+		if (copy_from_user(&phdr,
+				   (void *) (phdr_start + i * sizeof(phdr)),
+				   sizeof(phdr)))
+			goto fail;
+
+		if (phdr.p_type != PT_LOAD)
+			continue;
+		if (phdr.p_flags & (1 << 27))
+			continue;
+
+		map = vma_map_add(map, phdr.p_vaddr, phdr.p_memsz,
+				  phdr.p_offset, 0, 0);
+		if (!map)
+			goto fail;
+	}
+
+	pr_debug("SPU_PROF: Created non-overlay maps\n");
+	/* Traverse section table and search for overlay-related symbols.  */
+	for (i = 0; i < ehdr.e_shnum; i++) {
+		if (copy_from_user(&shdr,
+				   (void *) (shdr_start + i * sizeof(shdr)),
+				   sizeof(shdr)))
+			goto fail;
+
+		if (shdr.sh_type != SHT_SYMTAB)
+			continue;
+		if (shdr.sh_entsize != sizeof (sym))
+			continue;
+
+		if (copy_from_user(&shdr_str,
+				   (void *) (shdr_start + shdr.sh_link *
+					     sizeof(shdr)),
+				   sizeof(shdr)))
+			goto fail;
+
+		if (shdr_str.sh_type != SHT_STRTAB)
+			goto fail;;
+
+		for (j = 0; j < shdr.sh_size / sizeof (sym); j++) {
+			if (copy_from_user(&sym, (void *) (spu_elf_start +
+						       shdr.sh_offset + j *
+							   sizeof (sym)),
+					   sizeof (sym)))
+				goto fail;
+
+			if (copy_from_user(name, (void *)
+					   (spu_elf_start + shdr_str.sh_offset +
+					    sym.st_name),
+					   20))
+				goto fail;
+
+			if (memcmp(name, "_ovly_table", 12) == 0)
+				ovly_table_sym = sym.st_value;
+			if (memcmp(name, "_ovly_buf_table", 16) == 0)
+				ovly_buf_table_sym = sym.st_value;
+			if (memcmp(name, "_ovly_table_end", 16) == 0)
+				ovly_table_end_sym = sym.st_value;
+			if (memcmp(name, "_ovly_buf_table_end", 20) == 0)
+				ovly_buf_table_end_sym = sym.st_value;
+		}
+	}
+
+	/* If we don't have overlays, we're done.  */
+	if (ovly_table_sym == 0 || ovly_buf_table_sym == 0
+	    || ovly_table_end_sym == 0 || ovly_buf_table_end_sym == 0) {
+		pr_debug("SPU_PROF: No overlay table found\n");
+		goto out;
+	} else {
+		pr_debug("SPU_PROF: Overlay table found\n");
+	}
+
+	/* The _ovly_table symbol represents a table with one entry
+	 * per overlay section.	 The _ovly_buf_table symbol represents
+	 * a table with one entry per overlay region.
+	 * The struct spu_overlay_info gives the structure of the _ovly_table
+	 * entries.  The structure of _ovly_table_buf is simply one
+	 * u32 word per entry.
+	 */
+	overlay_tbl_offset = vma_map_lookup(map, ovly_table_sym,
+					    aSpu, &grd_val);
+	if (overlay_tbl_offset < 0) {
+		printk(KERN_ERR "SPU_PROF: "
+		       "%s, line %d: Error finding SPU overlay table\n",
+		       __FUNCTION__, __LINE__);
+		goto fail;
+	}
+	ovly_table = spu_elf_start + overlay_tbl_offset;
+
+	n_ovlys = (ovly_table_end_sym -
+		   ovly_table_sym) / sizeof (ovly);
+
+	/* Traverse overlay table.  */
+	for (i = 0; i < n_ovlys; i++) {
+		if (copy_from_user(&ovly, (void *)
+				   (ovly_table + i * sizeof (ovly)),
+				   sizeof (ovly)))
+			goto fail;
+
+		/* The ovly.vma/size/offset arguments are analogous to the same
+		 * arguments used above for non-overlay maps.  The final two
+		 * args are referred to as the guard pointer and the guard
+		 * value.
+		 * The guard pointer is an entry in the _ovly_buf_table,
+		 * computed using ovly.buf as the index into the table.	 Since
+		 * ovly.buf values begin at '1' to reference the first (or 0th)
+		 * entry in the _ovly_buf_table, the computation subtracts 1
+		 * from ovly.buf.
+		 * The guard value is stored in the _ovly_buf_table entry and
+		 * is an index (starting at 1) back to the _ovly_table entry
+		 * that is pointing at this _ovly_buf_table entry.  So, for
+		 * example, for an overlay scenario with one overlay segment
+		 * and two overlay sections:
+		 *	- Section 1 points to the first entry of the
+		 *	  _ovly_buf_table, which contains a guard value
+		 *	  of '1', referencing the first (index=0) entry of
+		 *	  _ovly_table.
+		 *	- Section 2 points to the second entry of the
+		 *	  _ovly_buf_table, which contains a guard value
+		 *	  of '2', referencing the second (index=1) entry of
+		 *	  _ovly_table.
+		 */
+		map = vma_map_add(map, ovly.vma, ovly.size, ovly.offset,
+				  ovly_buf_table_sym + (ovly.buf-1) * 4, i+1);
+		if (!map)
+			goto fail;
+	}
+	goto out;
+
+ fail:
+	map = NULL;
+ out:
+	return map;
+}