[PATCH 2/2] Add support to OProfile for profiling Cell BE SPUs

@@ -1455,7 +1455,8 @@ CONFIG_HAS_DMA=y
 # Instrumentation Support
 #
 CONFIG_PROFILING=y
-CONFIG_OPROFILE=y
+CONFIG_OPROFILE=m
+CONFIG_OPROFILE_CELL=y
 # CONFIG_KPROBES is not set
 
 #

@@ -0,0 +1,90 @@
+ /*
+ * 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"
+
+static inline 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;
+}
+
+/* Defines used for sync_start */
+#define SKIP_GENERIC_SYNC 0
+#define SYNC_START_ERROR -1
+#define DO_GENERIC_SYNC 1
+
+struct	spu_overlay_info
+{
+	unsigned int vma;
+	unsigned int size;
+	unsigned int offset;
+	unsigned int buf;
+};
+
+struct vma_to_fileoffset_map
+{
+	struct vma_to_fileoffset_map *next;
+	unsigned int vma;
+	unsigned int size;
+	unsigned int offset;
+	unsigned int guard_ptr;
+	unsigned int guard_val;
+};
+
+/* 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_profiling_frequency(unsigned int freq_khz, unsigned int

@@ -0,0 +1,223 @@
+/*
+ * 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 <linux/time.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;
+
+extern int spu_prof_num_nodes;
+
+
+#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_profiling_frequency(unsigned int freq_khz, unsigned int

@@ -0,0 +1,465 @@
+/*
+ * 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

@@ -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;
+}

@@ -29,6 +29,8 @@ static struct op_powerpc_model *model;
 static struct op_counter_config ctr[OP_MAX_COUNTER];
 static struct op_system_config sys;
 
+static int op_per_cpu_rc;
+
 static void op_handle_interrupt(struct pt_regs *regs)
 {
 	model->handle_interrupt(regs, ctr);

@@ -36,25 +38,41 @@ static void op_handle_interrupt(struct p
 
 static void op_powerpc_cpu_setup(void *dummy)
 {
-	model->cpu_setup(ctr);
+	int ret;
+
+	ret = model->cpu_setup(ctr);
+
+	if (ret != 0)
+		op_per_cpu_rc = ret;
 }
 
 static int op_powerpc_setup(void)
 {
 	int err;
 
+	op_per_cpu_rc = 0;
+
 	/* Grab the hardware */
 	err = reserve_pmc_hardware(op_handle_interrupt);
 	if (err)
 		return err;
 
 	/* Pre-compute the values to stuff in the hardware registers.  */
-	model->reg_setup(ctr, &sys, model->num_counters);
+	op_per_cpu_rc = model->reg_setup(ctr, &sys, model->num_counters);
 
-	/* Configure the registers on all cpus.  */
+	if (op_per_cpu_rc)
+		goto out;
+
+	/* Configure the registers on all cpus.	 If an error occurs on one
+	 * of the cpus, op_per_cpu_rc will be set to the error */
 	on_each_cpu(op_powerpc_cpu_setup, NULL, 0, 1);
 
-	return 0;
+out:	if (op_per_cpu_rc) {
+		/* error on setup release the performance counter hardware */
+		release_pmc_hardware();
+	}
+
+	return op_per_cpu_rc;
 }
 
 static void op_powerpc_shutdown(void)

@@ -64,16 +82,29 @@ static void op_powerpc_shutdown(void)
 
 static void op_powerpc_cpu_start(void *dummy)
 {
-	model->start(ctr);
+	/* If any of the cpus have return an error, set the
+	 * global flag to the error so it can be returned
+	 * to the generic OProfile caller.
+	 */
+	int ret;
+
+	ret = model->start(ctr);
+	if (ret != 0)
+		op_per_cpu_rc = ret;
 }
 
 static int op_powerpc_start(void)
 {
+	op_per_cpu_rc = 0;
+
 	if (model->global_start)
-		model->global_start(ctr);
-	if (model->start)
+		return model->global_start(ctr);
+	if (model->start) {
 		on_each_cpu(op_powerpc_cpu_start, NULL, 0, 1);
-	return 0;
+		return op_per_cpu_rc;
+	}
+	return -EIO; /* No start function is defined for this
+			power architecture */
 }
 
 static inline void op_powerpc_cpu_stop(void *dummy)

@@ -147,11 +178,13 @@ int __init oprofile_arch_init(struct opr
 
 	switch (cur_cpu_spec->oprofile_type) {
 #ifdef CONFIG_PPC64
-#ifdef CONFIG_PPC_CELL_NATIVE
+#ifdef CONFIG_OPROFILE_CELL
 		case PPC_OPROFILE_CELL:
 			if (firmware_has_feature(FW_FEATURE_LPAR))
 				return -ENODEV;
 			model = &op_model_cell;
+			ops->sync_start = model->sync_start;
+			ops->sync_stop = model->sync_stop;
 			break;
 #endif
 		case PPC_OPROFILE_RS64:

@@ -15,3 +15,10 @@ config OPROFILE
 
 	  If unsure, say N.
 
+config OPROFILE_CELL
+	bool "OProfile for Cell Broadband Engine"
+	depends on (SPU_FS = y && OPROFILE = m) || (SPU_FS = y && OPROFILE =

@@ -11,7 +11,9 @@ DRIVER_OBJS := $(addprefix ../../../driv
 		timer_int.o )
 
 oprofile-y := $(DRIVER_OBJS) common.o backtrace.o
-oprofile-$(CONFIG_PPC_CELL_NATIVE) += op_model_cell.o
+oprofile-$(CONFIG_OPROFILE_CELL) += op_model_cell.o \
+		cell/spu_profiler.o cell/vma_map.o \
+		cell/spu_task_sync.o
 oprofile-$(CONFIG_PPC64) += op_model_rs64.o op_model_power4.o

@@ -5,8 +5,8 @@
  *
  * Author: David Erb (djerb@us.ibm.com)
  * Modifications:
- *         Carl Love <carll@us.ibm.com>
- *         Maynard Johnson <maynardj@us.ibm.com>
+ *	   Carl Love <carll@us.ibm.com>
+ *	   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

@@ -38,12 +38,25 @@
 
 #include "../platforms/cell/interrupt.h"
 #include "../platforms/cell/cbe_regs.h"
+#include "cell/pr_util.h"
+
+static void cell_global_stop_spu(void);
+
+/*
+ * spu_cycle_reset is the number of cycles between samples.
+ * This variable is used for SPU profiling and should ONLY be set
+ * at the beginning of cell_reg_setup; otherwise, it's read-only.
+ */
+static unsigned int spu_cycle_reset;
+
+#define NUM_SPUS_PER_NODE    8
+#define SPU_CYCLES_EVENT_NUM 2	/*  event number for SPU_CYCLES */
 
 #define PPU_CYCLES_EVENT_NUM 1	/*  event number for CYCLES */
-#define PPU_CYCLES_GRP_NUM   1  /* special group number for identifying
-                                 * PPU_CYCLES event
-                                 */
-#define CBE_COUNT_ALL_CYCLES 0x42800000	/* PPU cycle event specifier */
+#define PPU_CYCLES_GRP_NUM   1	/* special group number for identifying
+				 * PPU_CYCLES event
+				 */
+#define CBE_COUNT_ALL_CYCLES 0x42800000 /* PPU cycle event specifier */
 
 #define NUM_THREADS 2         /* number of physical threads in
 			       * physical processor

@@ -51,6 +64,7 @@
 #define NUM_TRACE_BUS_WORDS 4
 #define NUM_INPUT_BUS_WORDS 2
 
+#define MAX_SPU_COUNT 0xFFFFFF	/* maximum 24 bit LFSR value */
 
 struct pmc_cntrl_data {
 	unsigned long vcntr;

@@ -62,11 +76,10 @@ struct pmc_cntrl_data {
 /*
  * ibm,cbe-perftools rtas parameters
  */
-
 struct pm_signal {
 	u16 cpu;		/* Processor to modify */
-	u16 sub_unit;		/* hw subunit this applies to (if applicable) */
-	short int signal_group;	/* Signal Group to Enable/Disable */
+	u16 sub_unit;		/* hw subunit this applies to (if applicable)*/
+	short int signal_group; /* Signal Group to Enable/Disable */
 	u8 bus_word;		/* Enable/Disable on this Trace/Trigger/Event
 				 * Bus Word(s) (bitmask)
 				 */

@@ -112,21 +125,42 @@ static DEFINE_PER_CPU(unsigned long[NR_P
 
 static struct pmc_cntrl_data pmc_cntrl[NUM_THREADS][NR_PHYS_CTRS];
 
-/* Interpetation of hdw_thread:
+/*
+ * The CELL profiling code makes rtas calls to setup the debug bus to
+ * route the performance signals.  Additionally, SPU profiling requires
+ * a second rtas call to setup the hardware to capture the SPU PCs.
+ * The EIO error value is returned if the token lookups or the rtas
+ * call fail.  The EIO error number is the best choice of the existing
+ * error numbers.  The probability of rtas related error is very low.

@@ -147,8 +181,8 @@ rtas_ibm_cbe_perftools(int subfunc, int 
 {
 	u64 paddr = __pa(address);
 
-	return rtas_call(pm_rtas_token, 5, 1, NULL, subfunc, passthru,
-			 paddr >> 32, paddr & 0xffffffff, length);
+	return rtas_call(pm_rtas_token, 5, 1, NULL, subfunc,
+			 passthru, paddr >> 32, paddr & 0xffffffff, length);
 }
 
 static void pm_rtas_reset_signals(u32 node)

@@ -156,12 +190,13 @@ static void pm_rtas_reset_signals(u32 no
 	int ret;
 	struct pm_signal pm_signal_local;
 
-	/*  The debug bus is being set to the passthru disable state.
-	 *  However, the FW still expects atleast one legal signal routing
-	 *  entry or it will return an error on the arguments.  If we don't
-	 *  supply a valid entry, we must ignore all return values.  Ignoring
-	 *  all return values means we might miss an error we should be
-	 *  concerned about.
+	/*
+	 * The debug bus is being set to the passthru disable state.
+	 * However, the FW still expects atleast one legal signal routing
+	 * entry or it will return an error on the arguments.	If we don't
+	 * supply a valid entry, we must ignore all return values.  Ignoring
+	 * all return values means we might miss an error we should be
+	 * concerned about.
 	 */
 
 	/*  fw expects physical cpu #. */

@@ -175,18 +210,24 @@ static void pm_rtas_reset_signals(u32 no
 				     &pm_signal_local,
 				     sizeof(struct pm_signal));
 
-	if (ret)
+	if (unlikely(ret))
+		/*
+		 * Not a fatal error. For Oprofile stop, the oprofile
+		 * functions do not support returning an error for
+		 * failure to stop OProfile.
+		 */
 		printk(KERN_WARNING "%s: rtas returned: %d\n",
 		       __FUNCTION__, ret);
 }
 
-static void pm_rtas_activate_signals(u32 node, u32 count)
+static int pm_rtas_activate_signals(u32 node, u32 count)
 {
 	int ret;
 	int i, j;
 	struct pm_signal pm_signal_local[NR_PHYS_CTRS];
 
-	/* There is no debug setup required for the cycles event.
+	/*
+	 * There is no debug setup required for the cycles event.
 	 * Note that only events in the same group can be used.
 	 * Otherwise, there will be conflicts in correctly routing
 	 * the signals on the debug bus.  It is the responsiblity

@@ -213,10 +254,14 @@ static void pm_rtas_activate_signals(u32
 					     pm_signal_local,
 					     i * sizeof(struct pm_signal));
 
-		if (ret)
+		if (unlikely(ret)) {
 			printk(KERN_WARNING "%s: rtas returned: %d\n",
 			       __FUNCTION__, ret);
+			return -EIO;
+		}
 	}
+
+	return 0;
 }
 
 /*

@@ -260,11 +305,12 @@ static void set_pm_event(u32 ctr, int ev
 	pm_regs.pm07_cntrl[ctr] |= PM07_CTR_POLARITY(polarity);
 	pm_regs.pm07_cntrl[ctr] |= PM07_CTR_INPUT_CONTROL(input_control);
 
-	/* Some of the islands signal selection is based on 64 bit words.
+	/*
+	 * Some of the islands signal selection is based on 64 bit words.
 	 * The debug bus words are 32 bits, the input words to the performance
 	 * counters are defined as 32 bits.  Need to convert the 64 bit island
 	 * specification to the appropriate 32 input bit and bus word for the
-	 * performance counter event selection.  See the CELL Performance
+	 * performance counter event selection.	 See the CELL Performance
 	 * monitoring signals manual and the Perf cntr hardware descriptions
 	 * for the details.
 	 */

@@ -298,6 +344,7 @@ static void set_pm_event(u32 ctr, int ev
 					input_bus[j] = i;
 					pm_regs.group_control |=
 					    (i << (31 - i));
+
 					break;
 				}
 			}

@@ -309,7 +356,8 @@ out:
 
 static void write_pm_cntrl(int cpu)
 {
-	/* Oprofile will use 32 bit counters, set bits 7:10 to 0
+	/*
+	 * Oprofile will use 32 bit counters, set bits 7:10 to 0
 	 * pmregs.pm_cntrl is a global
 	 */
 

@@ -326,7 +374,8 @@ static void write_pm_cntrl(int cpu)
 	if (pm_regs.pm_cntrl.freeze == 1)
 		val |= CBE_PM_FREEZE_ALL_CTRS;
 
-	/* Routine set_count_mode must be called previously to set
+	/*
+	 * Routine set_count_mode must be called previously to set
 	 * the count mode based on the user selection of user and kernel.
 	 */
 	val |= CBE_PM_COUNT_MODE_SET(pm_regs.pm_cntrl.count_mode);

@@ -336,7 +385,8 @@ static void write_pm_cntrl(int cpu)
 static inline void
 set_count_mode(u32 kernel, u32 user)
 {
-	/* The user must specify user and kernel if they want them. If
+	/*
+	 * The user must specify user and kernel if they want them. If
 	 *  neither is specified, OProfile will count in hypervisor mode.
 	 *  pm_regs.pm_cntrl is a global
 	 */

@@ -364,7 +414,7 @@ static inline void enable_ctr(u32 cpu, u
 
 /*
  * Oprofile is expected to collect data on all CPUs simultaneously.
- * However, there is one set of performance counters per node.  There

@@ -377,19 +427,19 @@ static inline void enable_ctr(u32 cpu, u
  * pair of per-cpu arrays is used for storing the previous and next
  * pmc values for a given node.
  * NOTE: We use the per-cpu variable to improve cache performance.
+ *
+ * This routine will alternate loading the virtual counters for
+ * virtual CPUs
  */
 static void cell_virtual_cntr(unsigned long data)
 {
-	/* This routine will alternate loading the virtual counters for
-	 * virtual CPUs
-	 */
 	int i, prev_hdw_thread, next_hdw_thread;
 	u32 cpu;
 	unsigned long flags;
 
-	/* Make sure that the interrupt_hander and
-	 * the virt counter are not both playing with
-	 * the counters on the same node.
+	/*
+	 * Make sure that the interrupt_hander and the virt counter are
+	 * not both playing with the counters on the same node.
 	 */
 
 	spin_lock_irqsave(&virt_cntr_lock, flags);

@@ -400,22 +450,25 @@ static void cell_virtual_cntr(unsigned l
 	hdw_thread = 1 ^ hdw_thread;
 	next_hdw_thread = hdw_thread;
 
-	for (i = 0; i < num_counters; i++)
-	/* There are some per thread events.  Must do the
+	/*
+	 * There are some per thread events.  Must do the
 	 * set event, for the thread that is being started
 	 */
+	for (i = 0; i < num_counters; i++)
 		set_pm_event(i,
 			pmc_cntrl[next_hdw_thread][i].evnts,
 			pmc_cntrl[next_hdw_thread][i].masks);
 
-	/* The following is done only once per each node, but
+	/*
+	 * The following is done only once per each node, but
 	 * we need cpu #, not node #, to pass to the cbe_xxx functions.
 	 */
 	for_each_online_cpu(cpu) {
 		if (cbe_get_hw_thread_id(cpu))
 			continue;
 
-		/* stop counters, save counter values, restore counts
+		/*
+		 * stop counters, save counter values, restore counts
 		 * for previous thread
 		 */
 		cbe_disable_pm(cpu);

@@ -428,7 +481,7 @@ static void cell_virtual_cntr(unsigned l
 			    == 0xFFFFFFFF)
 				/* If the cntr value is 0xffffffff, we must
 				 * reset that to 0xfffffff0 when the current
-				 * thread is restarted.  This will generate a
+				 * thread is restarted.	 This will generate a
 				 * new interrupt and make sure that we never
 				 * restore the counters to the max value.  If
 				 * the counters were restored to the max value,

@@ -444,13 +497,15 @@ static void cell_virtual_cntr(unsigned l
 						      next_hdw_thread)[i]);
 		}
 
-		/* Switch to the other thread. Change the interrupt
+		/*
+		 * Switch to the other thread. Change the interrupt
 		 * and control regs to be scheduled on the CPU
 		 * corresponding to the thread to execute.
 		 */
 		for (i = 0; i < num_counters; i++) {
 			if (pmc_cntrl[next_hdw_thread][i].enabled) {
-				/* There are some per thread events.
+				/*
+				 * There are some per thread events.
 				 * Must do the set event, enable_cntr
 				 * for each cpu.
 				 */

@@ -482,17 +537,42 @@ static void start_virt_cntrs(void)
 }
 
 /* This function is called once for all cpus combined */
-static void
-cell_reg_setup(struct op_counter_config *ctr,
-	       struct op_system_config *sys, int num_ctrs)
+static int cell_reg_setup(struct op_counter_config *ctr,
+			struct op_system_config *sys, int num_ctrs)
 {
 	int i, j, cpu;
+	spu_cycle_reset = 0;
+
+	if (ctr[0].event == SPU_CYCLES_EVENT_NUM) {
+		spu_cycle_reset = ctr[0].count;
+
+		/*
+		 * Each node will need to make the rtas call to start
+		 * and stop SPU profiling.  Get the token once and store it.
+		 */
+		spu_rtas_token = rtas_token("ibm,cbe-spu-perftools");
+
+		if (unlikely(spu_rtas_token == RTAS_UNKNOWN_SERVICE)) {
+			printk(KERN_ERR
+			       "%s: rtas token ibm,cbe-spu-perftools unknown\n",
+			       __FUNCTION__);
+			return -EIO;
+		}
+	}
 
 	pm_rtas_token = rtas_token("ibm,cbe-perftools");
-	if (pm_rtas_token == RTAS_UNKNOWN_SERVICE) {
-		printk(KERN_WARNING "%s: RTAS_UNKNOWN_SERVICE\n",
+
+	/*
+	 * For all events excetp PPU CYCLEs, each node will need to make
+	 * the rtas cbe-perftools call to setup and reset the debug bus.
+	 * Make the token lookup call once and store it in the global
+	 * variable pm_rtas_token.
+	 */
+	if (unlikely(pm_rtas_token == RTAS_UNKNOWN_SERVICE)) {
+		printk(KERN_ERR
+		       "%s: rtas token ibm,cbe-perftools unknown\n",
 		       __FUNCTION__);
-		goto out;
+		return -EIO;
 	}
 
 	num_counters = num_ctrs;

@@ -520,7 +600,8 @@ cell_reg_setup(struct op_counter_config 
 			per_cpu(pmc_values, j)[i] = 0;
 	}
 
-	/* Setup the thread 1 events, map the thread 0 event to the
+	/*
+	 * Setup the thread 1 events, map the thread 0 event to the
 	 * equivalent thread 1 event.
 	 */
 	for (i = 0; i < num_ctrs; ++i) {

@@ -544,9 +625,10 @@ cell_reg_setup(struct op_counter_config 
 	for (i = 0; i < NUM_INPUT_BUS_WORDS; i++)
 		input_bus[i] = 0xff;
 
-	/* Our counters count up, and "count" refers to
+	/*
+	 * Our counters count up, and "count" refers to
 	 * how much before the next interrupt, and we interrupt
-	 * on overflow.  So we calculate the starting value
+	 * on overflow.	 So we calculate the starting value
 	 * which will give us "count" until overflow.
 	 * Then we set the events on the enabled counters.
 	 */

@@ -569,28 +651,27 @@ cell_reg_setup(struct op_counter_config 
 		for (i = 0; i < num_counters; ++i) {
 			per_cpu(pmc_values, cpu)[i] = reset_value[i];
 		}
-out:
-	;
+
+	return 0;
 }
 
+
+
 /* This function is called once for each cpu */
-static void cell_cpu_setup(struct op_counter_config *cntr)
+static int cell_cpu_setup(struct op_counter_config *cntr)
 {
 	u32 cpu = smp_processor_id();
 	u32 num_enabled = 0;
 	int i;
 
+	if (spu_cycle_reset)
+		return 0;
+
 	/* There is one performance monitor per processor chip (i.e. node),
 	 * so we only need to perform this function once per node.
 	 */
 	if (cbe_get_hw_thread_id(cpu))
-		goto out;
-
-	if (pm_rtas_token == RTAS_UNKNOWN_SERVICE) {
-		printk(KERN_WARNING "%s: RTAS_UNKNOWN_SERVICE\n",
-		       __FUNCTION__);
-		goto out;
-	}
+		return 0;
 
 	/* Stop all counters */
 	cbe_disable_pm(cpu);

@@ -609,16 +690,286 @@ static void cell_cpu_setup(struct op_cou
 		}
 	}
 
-	pm_rtas_activate_signals(cbe_cpu_to_node(cpu), num_enabled);
+	/*
+	 * The pm_rtas_activate_signals will return -EIO if the FW
+	 * call failed.
+	 */
+	return pm_rtas_activate_signals(cbe_cpu_to_node(cpu), num_enabled);
+}
+
+#define ENTRIES	 303
+#define MAXLFSR	 0xFFFFFF
+
+/* precomputed table of 24 bit LFSR values */
+static int initial_lfsr[] = {
+ 8221349, 12579195, 5379618, 10097839, 7512963, 7519310, 3955098,

@@ -651,19 +1002,80 @@ static void cell_global_start(struct op_
 	oprofile_running = 1;
 	smp_wmb();
 
-	/* NOTE: start_virt_cntrs will result in cell_virtual_cntr() being
-	 * executed which manipulates the PMU.  We start the "virtual counter"
+	/*
+	 * NOTE: start_virt_cntrs will result in cell_virtual_cntr() being
+	 * executed which manipulates the PMU.	We start the "virtual counter"
 	 * here so that we do not need to synchronize access to the PMU in
 	 * the above for-loop.
 	 */
 	start_virt_cntrs();
+
+	return 0;
 }
 
-static void cell_global_stop(void)
+static int cell_global_start(struct op_counter_config *ctr)
+{
+	if (spu_cycle_reset) {
+		return cell_global_start_spu(ctr);
+	} else {
+		return cell_global_start_ppu(ctr);
+	}
+}
+
+/*
+ * Note the generic OProfile stop calls do not support returning
+ * an error on stop.  Hence, will not return an error if the FW
+ * calls fail on stop.	Failure to reset the debug bus is not an issue.
+ * Failure to disable the SPU profiling is not an issue.  The FW calls
+ * to enable the performance counters and debug bus will work even if
+ * the hardware was not cleanly reset.
+ */
+static void cell_global_stop_spu(void)
+{
+	int subfunc, rtn_value;
+	unsigned int lfsr_value;
+	int cpu;
+
+	oprofile_running = 0;
+
+#ifdef CONFIG_CPU_FREQ
+	cpufreq_unregister_notifier(&cpu_freq_notifier_block,
+				    CPUFREQ_TRANSITION_NOTIFIER);
+#endif
+
+	for_each_online_cpu(cpu) {
+		if (cbe_get_hw_thread_id(cpu))
+			continue;
+
+		subfunc = 3;	/*
+				 * 2 - activate SPU tracing,
+				 * 3 - deactivate
+				 */
+		lfsr_value = 0x8f100000;
+
+		rtn_value = rtas_call(spu_rtas_token, 3, 1, NULL,
+				      subfunc, cbe_cpu_to_node(cpu),
+				      lfsr_value);
+
+		if (unlikely(rtn_value != 0)) {
+			printk(KERN_ERR
+			       "%s: rtas call ibm,cbe-spu-perftools failed, return = %d\n",
+			       __FUNCTION__, rtn_value);
+		}
+
+		/* Deactivate the signals */
+		pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
+	}
+
+	stop_spu_profiling();
+}
+
+static void cell_global_stop_ppu(void)
 {
 	int cpu;
 
-	/* This routine will be called once for the system.
+	/*
+	 * This routine will be called once for the system.
 	 * There is one performance monitor per node, so we
 	 * only need to perform this function once per node.
 	 */

@@ -687,8 +1099,17 @@ static void cell_global_stop(void)
 	}
 }
 
-static void
-cell_handle_interrupt(struct pt_regs *regs, struct op_counter_config

@@ -699,13 +1120,15 @@ cell_handle_interrupt(struct pt_regs *re
 
 	cpu = smp_processor_id();
 
-	/* Need to make sure the interrupt handler and the virt counter
+	/*
+	 * Need to make sure the interrupt handler and the virt counter
 	 * routine are not running at the same time. See the
 	 * cell_virtual_cntr() routine for additional comments.
 	 */
 	spin_lock_irqsave(&virt_cntr_lock, flags);
 
-	/* Need to disable and reenable the performance counters
+	/*
+	 * Need to disable and reenable the performance counters
 	 * to get the desired behavior from the hardware.  This
 	 * is hardware specific.
 	 */

@@ -714,7 +1137,8 @@ cell_handle_interrupt(struct pt_regs *re
 
 	interrupt_mask = cbe_get_and_clear_pm_interrupts(cpu);
 
-	/* If the interrupt mask has been cleared, then the virt cntr
+	/*
+	 * If the interrupt mask has been cleared, then the virt cntr
 	 * has cleared the interrupt.  When the thread that generated
 	 * the interrupt is restored, the data count will be restored to
 	 * 0xffffff0 to cause the interrupt to be regenerated.

@@ -732,18 +1156,20 @@ cell_handle_interrupt(struct pt_regs *re
 			}
 		}
 
-		/* The counters were frozen by the interrupt.
+		/*
+		 * The counters were frozen by the interrupt.
 		 * Reenable the interrupt and restart the counters.
 		 * If there was a race between the interrupt handler and
-		 * the virtual counter routine.  The virutal counter
+		 * the virtual counter routine.	 The virutal counter
 		 * routine may have cleared the interrupts.  Hence must
 		 * use the virt_cntr_inter_mask to re-enable the interrupts.
 		 */
 		cbe_enable_pm_interrupts(cpu, hdw_thread,
 					 virt_cntr_inter_mask);
 
-		/* The writes to the various performance counters only writes
-		 * to a latch.  The new values (interrupt setting bits, reset
+		/*
+		 * The writes to the various performance counters only writes
+		 * to a latch.	The new values (interrupt setting bits, reset
 		 * counter value etc.) are not copied to the actual registers
 		 * until the performance monitor is enabled.  In order to get
 		 * this to work as desired, the permormance monitor needs to

@@ -755,10 +1181,33 @@ cell_handle_interrupt(struct pt_regs *re
 	spin_unlock_irqrestore(&virt_cntr_lock, flags);
 }
 
+/*
+ * This function is called from the generic OProfile
+ * driver.  When profiling PPUs, we need to do the
+ * generic sync start; otherwise, do spu_sync_start.
+ */
+static int cell_sync_start(void)
+{
+	if (spu_cycle_reset)
+		return spu_sync_start();
+	else
+		return DO_GENERIC_SYNC;
+}
+
+static int cell_sync_stop(void)
+{
+	if (spu_cycle_reset)
+		return spu_sync_stop();
+	else
+		return 1;
+}
+
 struct op_powerpc_model op_model_cell = {
 	.reg_setup = cell_reg_setup,
 	.cpu_setup = cell_cpu_setup,
 	.global_start = cell_global_start,
 	.global_stop = cell_global_stop,
+	.sync_start = cell_sync_start,
+	.sync_stop = cell_sync_stop,
 	.handle_interrupt = cell_handle_interrupt,
 };

@@ -264,6 +264,7 @@ static void spu_bind_context(struct spu 
 	ctx->spu = spu;
 	ctx->ops = &spu_hw_ops;
 	spu->pid = current->pid;
+	spu->tgid = current->tgid;
 	spu_associate_mm(spu, ctx->owner);
 	spu->ibox_callback = spufs_ibox_callback;
 	spu->wbox_callback = spufs_wbox_callback;

@@ -304,6 +305,7 @@ static void spu_unbind_context(struct sp
 	spu->dma_callback = NULL;
 	spu_associate_mm(spu, NULL);
 	spu->pid = 0;
+	spu->tgid = 0;
 	ctx->ops = &spu_backing_ops;
 	ctx->spu = NULL;
 	spu->flags = 0;

@@ -590,7 +592,7 @@ void spu_deactivate(struct spu_context *
 }
 
 /**
- * spu_yield -  yield a physical spu if others are waiting
+ * spu_yield -	yield a physical spu if others are waiting
  * @ctx:	spu context to yield
  *
  * Check if there is a higher priority context waiting and if yes

@@ -26,8 +26,9 @@
 #include <linux/profile.h>
 #include <linux/module.h>
 #include <linux/fs.h>
+#include <linux/oprofile.h>
 #include <linux/sched.h>
- 
+
 #include "oprofile_stats.h"
 #include "event_buffer.h"
 #include "cpu_buffer.h"

@@ -19,28 +19,10 @@ void free_event_buffer(void);
  
 /* wake up the process sleeping on the event file */
 void wake_up_buffer_waiter(void);
- 
-/* Each escaped entry is prefixed by ESCAPE_CODE
- * then one of the following codes, then the
- * relevant data.
- */
-#define ESCAPE_CODE			~0UL
-#define CTX_SWITCH_CODE 		1
-#define CPU_SWITCH_CODE 		2
-#define COOKIE_SWITCH_CODE 		3
-#define KERNEL_ENTER_SWITCH_CODE	4
-#define KERNEL_EXIT_SWITCH_CODE		5
-#define MODULE_LOADED_CODE		6
-#define CTX_TGID_CODE			7
-#define TRACE_BEGIN_CODE		8
-#define TRACE_END_CODE			9
- 
+
 #define INVALID_COOKIE ~0UL
 #define NO_COOKIE 0UL
 
-/* add data to the event buffer */
-void add_event_entry(unsigned long data);
- 
 extern const struct file_operations event_buffer_fops;
  
 /* mutex between sync_cpu_buffers() and the

@@ -53,9 +53,24 @@ int oprofile_setup(void)
 	 * us missing task deaths and eventually oopsing
 	 * when trying to process the event buffer.
 	 */
+	if (oprofile_ops.sync_start) {
+		int sync_ret = oprofile_ops.sync_start();
+		switch (sync_ret) {
+		case 0:
+			goto post_sync;
+		case 1:
+			goto do_generic;
+		case -1:
+			goto out3;
+		default:
+			goto out3;
+		}
+	}
+do_generic:
 	if ((err = sync_start()))
 		goto out3;
 
+post_sync:
 	is_setup = 1;
 	mutex_unlock(&start_mutex);
 	return 0;

@@ -118,7 +133,20 @@ out:
 void oprofile_shutdown(void)
 {
 	mutex_lock(&start_mutex);
+	if (oprofile_ops.sync_stop) {
+		int sync_ret = oprofile_ops.sync_stop();
+		switch (sync_ret) {
+		case 0:
+			goto post_sync;
+		case 1:
+			goto do_generic;
+		default:
+			goto post_sync;
+		}
+	}
+do_generic:
 	sync_stop();
+post_sync:
 	if (oprofile_ops.shutdown)
 		oprofile_ops.shutdown();
 	is_setup = 0;

@@ -39,14 +39,16 @@ struct op_system_config {
 
 /* Per-arch configuration */
 struct op_powerpc_model {
-	void (*reg_setup) (struct op_counter_config *,
+	int (*reg_setup) (struct op_counter_config *,
 			   struct op_system_config *,
 			   int num_counters);
-	void (*cpu_setup) (struct op_counter_config *);
-	void (*start) (struct op_counter_config *);
-        void (*global_start) (struct op_counter_config *);
+	int  (*cpu_setup) (struct op_counter_config *);
+	int  (*start) (struct op_counter_config *);
+	int  (*global_start) (struct op_counter_config *);
 	void (*stop) (void);
 	void (*global_stop) (void);
+	int (*sync_start)(void);
+	int (*sync_stop)(void);
 	void (*handle_interrupt) (struct pt_regs *,
 				  struct op_counter_config *);
 	int num_counters;