Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rric/oprofile

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/rric/oprofile: (31 commits)
  powerpc/oprofile: fix whitespaces in op_model_cell.c
  powerpc/oprofile: IBM CELL: add SPU event profiling support
  powerpc/oprofile: fix cell/pr_util.h
  powerpc/oprofile: IBM CELL: cleanup and restructuring
  oprofile: make new cpu buffer functions part of the api
  oprofile: remove #ifdef CONFIG_OPROFILE_IBS in non-ibs code
  ring_buffer: fix ring_buffer_event_length()
  oprofile: use new data sample format for ibs
  oprofile: add op_cpu_buffer_get_data()
  oprofile: add op_cpu_buffer_add_data()
  oprofile: rework implementation of cpu buffer events
  oprofile: modify op_cpu_buffer_read_entry()
  oprofile: add op_cpu_buffer_write_reserve()
  oprofile: rename variables in add_ibs_begin()
  oprofile: rename add_sample() in cpu_buffer.c
  oprofile: rename variable ibs_allowed to has_ibs in op_model_amd.c
  oprofile: making add_sample_entry() inline
  oprofile: remove backtrace code for ibs
  oprofile: remove unused ibs macro
  oprofile: remove unused components in struct oprofile_cpu_buffer
  ...

arch/powerpc/include/asm/cell-pmu.h

@@ -37,9 +37,11 @@
 #define CBE_PM_STOP_AT_MAX                 0x40000000
 #define CBE_PM_TRACE_MODE_GET(pm_control)  (((pm_control) >> 28) & 0x3)
 #define CBE_PM_TRACE_MODE_SET(mode)        (((mode)  & 0x3) << 28)
+#define CBE_PM_TRACE_BUF_OVFLW(bit)        (((bit) & 0x1) << 17)
 #define CBE_PM_COUNT_MODE_SET(count)       (((count) & 0x3) << 18)
 #define CBE_PM_FREEZE_ALL_CTRS             0x00100000
 #define CBE_PM_ENABLE_EXT_TRACE            0x00008000
+#define CBE_PM_SPU_ADDR_TRACE_SET(msk)     (((msk) & 0x3) << 9)
 
 /* Macros for the trace_address register. */
 #define CBE_PM_TRACE_BUF_FULL              0x00000800

arch/powerpc/include/asm/oprofile_impl.h

@@ -32,6 +32,12 @@ struct op_system_config {
 	unsigned long mmcr0;
 	unsigned long mmcr1;
 	unsigned long mmcra;
+#ifdef CONFIG_OPROFILE_CELL
+	/* Register for oprofile user tool to check cell kernel profiling
+	 * suport.
+	 */
+	unsigned long cell_support;
+#endif
 #endif
 	unsigned long enable_kernel;
 	unsigned long enable_user;

arch/powerpc/oprofile/cell/pr_util.h

@@ -30,6 +30,10 @@
 extern struct delayed_work spu_work;
 extern int spu_prof_running;
 
+#define TRACE_ARRAY_SIZE 1024
+
+extern spinlock_t oprof_spu_smpl_arry_lck;
+
 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 */
@@ -89,10 +93,11 @@ 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);
+int start_spu_profiling_cycles(unsigned int cycles_reset);
-
+void start_spu_profiling_events(void);
-void stop_spu_profiling(void);
 
+void stop_spu_profiling_cycles(void);
+void stop_spu_profiling_events(void);
 
 /* add the necessary profiling hooks */
 int spu_sync_start(void);

arch/powerpc/oprofile/cell/spu_profiler.c

@@ -18,11 +18,21 @@
 #include <asm/cell-pmu.h>
 #include "pr_util.h"
 
-#define TRACE_ARRAY_SIZE 1024
 #define SCALE_SHIFT 14
 
 static u32 *samples;
 
+/* spu_prof_running is a flag used to indicate if spu profiling is enabled
+ * or not.  It is set by the routines start_spu_profiling_cycles() and
+ * start_spu_profiling_events().  The flag is cleared by the routines
+ * stop_spu_profiling_cycles() and stop_spu_profiling_events().  These
+ * routines are called via global_start() and global_stop() which are called in
+ * op_powerpc_start() and op_powerpc_stop().  These routines are called once
+ * per system as a result of the user starting/stopping oprofile.  Hence, only
+ * one CPU per user at a time will be changing  the value of spu_prof_running.
+ * In general, OProfile does not protect against multiple users trying to run
+ * OProfile at a time.
+ */
 int spu_prof_running;
 static unsigned int profiling_interval;
 
@@ -31,8 +41,8 @@ static unsigned int profiling_interval;
 
 #define SPU_PC_MASK	     0xFFFF
 
-static DEFINE_SPINLOCK(sample_array_lock);
+DEFINE_SPINLOCK(oprof_spu_smpl_arry_lck);
-unsigned long sample_array_lock_flags;
+unsigned long oprof_spu_smpl_arry_lck_flags;
 
 void set_spu_profiling_frequency(unsigned int freq_khz, unsigned int cycles_reset)
 {
@@ -145,13 +155,13 @@ static enum hrtimer_restart profile_spus(struct hrtimer *timer)
 		 * 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,
+		spin_lock_irqsave(&oprof_spu_smpl_arry_lck,
-				  sample_array_lock_flags);
+				  oprof_spu_smpl_arry_lck_flags);
 		num_samples = cell_spu_pc_collection(cpu);
 
 		if (num_samples == 0) {
-			spin_unlock_irqrestore(&sample_array_lock,
+			spin_unlock_irqrestore(&oprof_spu_smpl_arry_lck,
-					       sample_array_lock_flags);
+					       oprof_spu_smpl_arry_lck_flags);
 			continue;
 		}
 
@@ -162,8 +172,8 @@ static enum hrtimer_restart profile_spus(struct hrtimer *timer)
 					num_samples);
 		}
 
-		spin_unlock_irqrestore(&sample_array_lock,
+		spin_unlock_irqrestore(&oprof_spu_smpl_arry_lck,
-				       sample_array_lock_flags);
+				       oprof_spu_smpl_arry_lck_flags);
 
 	}
 	smp_wmb();	/* insure spu event buffer updates are written */
@@ -182,13 +192,13 @@ static enum hrtimer_restart profile_spus(struct hrtimer *timer)
 
 static struct hrtimer timer;
 /*
- * Entry point for SPU profiling.
+ * Entry point for SPU cycle 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)
+int start_spu_profiling_cycles(unsigned int cycles_reset)
 {
 	ktime_t kt;
 
@@ -212,10 +222,30 @@ int start_spu_profiling(unsigned int cycles_reset)
 	return 0;
 }
 
-void stop_spu_profiling(void)
+/*
+ * Entry point for SPU event 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.
+ */
+void start_spu_profiling_events(void)
+{
+	spu_prof_running = 1;
+	schedule_delayed_work(&spu_work, DEFAULT_TIMER_EXPIRE);
+
+	return;
+}
+
+void stop_spu_profiling_cycles(void)
 {
 	spu_prof_running = 0;
 	hrtimer_cancel(&timer);
 	kfree(samples);
-	pr_debug("SPU_PROF: stop_spu_profiling issued\n");
+	pr_debug("SPU_PROF: stop_spu_profiling_cycles issued\n");
+}
+
+void stop_spu_profiling_events(void)
+{
+	spu_prof_running = 0;
 }

arch/powerpc/oprofile/common.c

@@ -132,6 +132,28 @@ static int op_powerpc_create_files(struct super_block *sb, struct dentry *root)
 	oprofilefs_create_ulong(sb, root, "mmcr0", &sys.mmcr0);
 	oprofilefs_create_ulong(sb, root, "mmcr1", &sys.mmcr1);
 	oprofilefs_create_ulong(sb, root, "mmcra", &sys.mmcra);
+#ifdef CONFIG_OPROFILE_CELL
+	/* create a file the user tool can check to see what level of profiling
+	 * support exits with this kernel. Initialize bit mask to indicate
+	 * what support the kernel has:
+	 * bit 0      -  Supports SPU event profiling in addition to PPU
+	 *               event and cycles; and SPU cycle profiling
+	 * bits 1-31  -  Currently unused.
+	 *
+	 * If the file does not exist, then the kernel only supports SPU
+	 * cycle profiling, PPU event and cycle profiling.
+	 */
+	oprofilefs_create_ulong(sb, root, "cell_support", &sys.cell_support);
+	sys.cell_support = 0x1; /* Note, the user OProfile tool must check
+				 * that this bit is set before attempting to
+				 * user SPU event profiling.  Older kernels
+				 * will not have this file, hence the user
+				 * tool is not allowed to do SPU event
+				 * profiling on older kernels.  Older kernels
+				 * will accept SPU events but collected data
+				 * is garbage.
+				 */
+#endif
 #endif
 
 	for (i = 0; i < model->num_counters; ++i) {

arch/powerpc/oprofile/op_model_cell.c

@@ -40,14 +40,15 @@
 #include "../platforms/cell/interrupt.h"
 #include "cell/pr_util.h"
 
-static void cell_global_stop_spu(void);
+#define PPU_PROFILING            0
+#define SPU_PROFILING_CYCLES     1
+#define SPU_PROFILING_EVENTS     2
 
-/*
+#define SPU_EVENT_NUM_START      4100
- * spu_cycle_reset is the number of cycles between samples.
+#define SPU_EVENT_NUM_STOP       4399
- * This variable is used for SPU profiling and should ONLY be set
+#define SPU_PROFILE_EVENT_ADDR          4363  /* spu, address trace, decimal */
- * at the beginning of cell_reg_setup; otherwise, it's read-only.
+#define SPU_PROFILE_EVENT_ADDR_MASK_A   0x146 /* sub unit set to zero */
- */
+#define SPU_PROFILE_EVENT_ADDR_MASK_B   0x186 /* sub unit set to zero */
-static unsigned int spu_cycle_reset;
 
 #define NUM_SPUS_PER_NODE    8
 #define SPU_CYCLES_EVENT_NUM 2	/*  event number for SPU_CYCLES */
@@ -66,6 +67,21 @@ static unsigned int spu_cycle_reset;
 
 #define MAX_SPU_COUNT 0xFFFFFF	/* maximum 24 bit LFSR value */
 
+/* Minumum HW interval timer setting to send value to trace buffer is 10 cycle.
+ * To configure counter to send value every N cycles set counter to
+ * 2^32 - 1 - N.
+ */
+#define NUM_INTERVAL_CYC  0xFFFFFFFF - 10
+
+/*
+ * 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;
+static unsigned int profiling_mode;
+static int spu_evnt_phys_spu_indx;
+
 struct pmc_cntrl_data {
 	unsigned long vcntr;
 	unsigned long evnts;
@@ -105,6 +121,8 @@ struct pm_cntrl {
 	u16 trace_mode;
 	u16 freeze;
 	u16 count_mode;
+	u16 spu_addr_trace;
+	u8  trace_buf_ovflw;
 };
 
 static struct {
@@ -122,7 +140,7 @@ static struct {
 #define GET_INPUT_CONTROL(x) ((x & 0x00000004) >> 2)
 
 static DEFINE_PER_CPU(unsigned long[NR_PHYS_CTRS], pmc_values);
-
+static unsigned long spu_pm_cnt[MAX_NUMNODES * NUM_SPUS_PER_NODE];
 static struct pmc_cntrl_data pmc_cntrl[NUM_THREADS][NR_PHYS_CTRS];
 
 /*
@@ -152,6 +170,7 @@ static u32 hdw_thread;
 
 static u32 virt_cntr_inter_mask;
 static struct timer_list timer_virt_cntr;
+static struct timer_list timer_spu_event_swap;
 
 /*
  * pm_signal needs to be global since it is initialized in
@@ -165,7 +184,7 @@ static int spu_rtas_token;   /* token for SPU cycle profiling */
 static u32 reset_value[NR_PHYS_CTRS];
 static int num_counters;
 static int oprofile_running;
-static DEFINE_SPINLOCK(virt_cntr_lock);
+static DEFINE_SPINLOCK(cntr_lock);
 
 static u32 ctr_enabled;
 
@@ -336,13 +355,13 @@ static void set_pm_event(u32 ctr, int event, u32 unit_mask)
 	for (i = 0; i < NUM_DEBUG_BUS_WORDS; i++) {
 		if (bus_word & (1 << i)) {
 			pm_regs.debug_bus_control |=
-			    (bus_type << (30 - (2 * i)));
+				(bus_type << (30 - (2 * i)));
 
 			for (j = 0; j < NUM_INPUT_BUS_WORDS; j++) {
 				if (input_bus[j] == 0xff) {
 					input_bus[j] = i;
 					pm_regs.group_control |=
-					    (i << (30 - (2 * j)));
+						(i << (30 - (2 * j)));
 
 					break;
 				}
@@ -367,12 +386,16 @@ static void write_pm_cntrl(int cpu)
 	if (pm_regs.pm_cntrl.stop_at_max == 1)
 		val |= CBE_PM_STOP_AT_MAX;
 
-	if (pm_regs.pm_cntrl.trace_mode == 1)
+	if (pm_regs.pm_cntrl.trace_mode != 0)
 		val |= CBE_PM_TRACE_MODE_SET(pm_regs.pm_cntrl.trace_mode);
 
+	if (pm_regs.pm_cntrl.trace_buf_ovflw == 1)
+		val |= CBE_PM_TRACE_BUF_OVFLW(pm_regs.pm_cntrl.trace_buf_ovflw);
 	if (pm_regs.pm_cntrl.freeze == 1)
 		val |= CBE_PM_FREEZE_ALL_CTRS;
 
+	val |= CBE_PM_SPU_ADDR_TRACE_SET(pm_regs.pm_cntrl.spu_addr_trace);
+
 	/*
 	 * Routine set_count_mode must be called previously to set
 	 * the count mode based on the user selection of user and kernel.
@@ -441,7 +464,7 @@ static void cell_virtual_cntr(unsigned long data)
 	 * not both playing with the counters on the same node.
 	 */
 
-	spin_lock_irqsave(&virt_cntr_lock, flags);
+	spin_lock_irqsave(&cntr_lock, flags);
 
 	prev_hdw_thread = hdw_thread;
 
@@ -480,7 +503,7 @@ static void cell_virtual_cntr(unsigned long data)
 		cbe_disable_pm_interrupts(cpu);
 		for (i = 0; i < num_counters; i++) {
 			per_cpu(pmc_values, cpu + prev_hdw_thread)[i]
-			    = cbe_read_ctr(cpu, i);
+				= cbe_read_ctr(cpu, i);
 
 			if (per_cpu(pmc_values, cpu + next_hdw_thread)[i]
 			    == 0xFFFFFFFF)
@@ -527,7 +550,7 @@ static void cell_virtual_cntr(unsigned long data)
 		cbe_enable_pm(cpu);
 	}
 
-	spin_unlock_irqrestore(&virt_cntr_lock, flags);
+	spin_unlock_irqrestore(&cntr_lock, flags);
 
 	mod_timer(&timer_virt_cntr, jiffies + HZ / 10);
 }
@@ -541,38 +564,146 @@ static void start_virt_cntrs(void)
 	add_timer(&timer_virt_cntr);
 }
 
-/* This function is called once for all cpus combined */
+static int cell_reg_setup_spu_cycles(struct op_counter_config *ctr,
-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 = ctr[0].count;
-	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",
-			       __func__);
-			return -EIO;
-		}
-	}
-
-	pm_rtas_token = rtas_token("ibm,cbe-perftools");
 
 	/*
-	 * For all events excetp PPU CYCLEs, each node will need to make
+	 * 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",
+		       __func__);
+		return -EIO;
+	}
+	return 0;
+}
+
+/* Unfortunately, the hardware will only support event profiling
+ * on one SPU per node at a time.  Therefore, we must time slice
+ * the profiling across all SPUs in the node.  Note, we do this
+ * in parallel for each node.  The following routine is called
+ * periodically based on kernel timer to switch which SPU is
+ * being monitored in a round robbin fashion.
+ */
+static void spu_evnt_swap(unsigned long data)
+{
+	int node;
+	int cur_phys_spu, nxt_phys_spu, cur_spu_evnt_phys_spu_indx;
+	unsigned long flags;
+	int cpu;
+	int ret;
+	u32 interrupt_mask;
+
+
+	/* enable interrupts on cntr 0 */
+	interrupt_mask = CBE_PM_CTR_OVERFLOW_INTR(0);
+
+	hdw_thread = 0;
+
+	/* Make sure spu event interrupt handler and spu event swap
+	 * don't access the counters simultaneously.
+	 */
+	spin_lock_irqsave(&cntr_lock, flags);
+
+	cur_spu_evnt_phys_spu_indx = spu_evnt_phys_spu_indx;
+
+	if (++(spu_evnt_phys_spu_indx) == NUM_SPUS_PER_NODE)
+		spu_evnt_phys_spu_indx = 0;
+
+	pm_signal[0].sub_unit = spu_evnt_phys_spu_indx;
+	pm_signal[1].sub_unit = spu_evnt_phys_spu_indx;
+	pm_signal[2].sub_unit = spu_evnt_phys_spu_indx;
+
+	/* switch the SPU being profiled on each node */
+	for_each_online_cpu(cpu) {
+		if (cbe_get_hw_thread_id(cpu))
+			continue;
+
+		node = cbe_cpu_to_node(cpu);
+		cur_phys_spu = (node * NUM_SPUS_PER_NODE)
+			+ cur_spu_evnt_phys_spu_indx;
+		nxt_phys_spu = (node * NUM_SPUS_PER_NODE)
+			+ spu_evnt_phys_spu_indx;
+
+		/*
+		 * stop counters, save counter values, restore counts
+		 * for previous physical SPU
+		 */
+		cbe_disable_pm(cpu);
+		cbe_disable_pm_interrupts(cpu);
+
+		spu_pm_cnt[cur_phys_spu]
+			= cbe_read_ctr(cpu, 0);
+
+		/* restore previous count for the next spu to sample */
+		/* NOTE, hardware issue, counter will not start if the
+		 * counter value is at max (0xFFFFFFFF).
+		 */
+		if (spu_pm_cnt[nxt_phys_spu] >= 0xFFFFFFFF)
+			cbe_write_ctr(cpu, 0, 0xFFFFFFF0);
+		 else
+			 cbe_write_ctr(cpu, 0, spu_pm_cnt[nxt_phys_spu]);
+
+		pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
+
+		/* setup the debug bus measure the one event and
+		 * the two events to route the next SPU's PC on
+		 * the debug bus
+		 */
+		ret = pm_rtas_activate_signals(cbe_cpu_to_node(cpu), 3);
+		if (ret)
+			printk(KERN_ERR "%s: pm_rtas_activate_signals failed, "
+			       "SPU event swap\n", __func__);
+
+		/* clear the trace buffer, don't want to take PC for
+		 * previous SPU*/
+		cbe_write_pm(cpu, trace_address, 0);
+
+		enable_ctr(cpu, 0, pm_regs.pm07_cntrl);
+
+		/* Enable interrupts on the CPU thread that is starting */
+		cbe_enable_pm_interrupts(cpu, hdw_thread,
+					 interrupt_mask);
+		cbe_enable_pm(cpu);
+	}
+
+	spin_unlock_irqrestore(&cntr_lock, flags);
+
+	/* swap approximately every 0.1 seconds */
+	mod_timer(&timer_spu_event_swap, jiffies + HZ / 25);
+}
+
+static void start_spu_event_swap(void)
+{
+	init_timer(&timer_spu_event_swap);
+	timer_spu_event_swap.function = spu_evnt_swap;
+	timer_spu_event_swap.data = 0UL;
+	timer_spu_event_swap.expires = jiffies + HZ / 25;
+	add_timer(&timer_spu_event_swap);
+}
+
+static int cell_reg_setup_spu_events(struct op_counter_config *ctr,
+			struct op_system_config *sys, int num_ctrs)
+{
+	int i;
+
+	/* routine is called once for all nodes */
+
+	spu_evnt_phys_spu_indx = 0;
+	/*
+	 * For all events except 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.
 	 */
+	pm_rtas_token = rtas_token("ibm,cbe-perftools");
+
 	if (unlikely(pm_rtas_token == RTAS_UNKNOWN_SERVICE)) {
 		printk(KERN_ERR
 		       "%s: rtas token ibm,cbe-perftools unknown\n",
@@ -580,6 +711,58 @@ static int cell_reg_setup(struct op_counter_config *ctr,
 		return -EIO;
 	}
 
+	/* setup the pm_control register settings,
+	 * settings will be written per node by the
+	 * cell_cpu_setup() function.
+	 */
+	pm_regs.pm_cntrl.trace_buf_ovflw = 1;
+
+	/* Use the occurrence trace mode to have SPU PC saved
+	 * to the trace buffer.  Occurrence data in trace buffer
+	 * is not used.  Bit 2 must be set to store SPU addresses.
+	 */
+	pm_regs.pm_cntrl.trace_mode = 2;
+
+	pm_regs.pm_cntrl.spu_addr_trace = 0x1;  /* using debug bus
+						   event 2 & 3 */
+
+	/* setup the debug bus event array with the SPU PC routing events.
+	*  Note, pm_signal[0] will be filled in by set_pm_event() call below.
+	*/
+	pm_signal[1].signal_group = SPU_PROFILE_EVENT_ADDR / 100;
+	pm_signal[1].bus_word = GET_BUS_WORD(SPU_PROFILE_EVENT_ADDR_MASK_A);
+	pm_signal[1].bit = SPU_PROFILE_EVENT_ADDR % 100;
+	pm_signal[1].sub_unit = spu_evnt_phys_spu_indx;
+
+	pm_signal[2].signal_group = SPU_PROFILE_EVENT_ADDR / 100;
+	pm_signal[2].bus_word = GET_BUS_WORD(SPU_PROFILE_EVENT_ADDR_MASK_B);
+	pm_signal[2].bit = SPU_PROFILE_EVENT_ADDR % 100;
+	pm_signal[2].sub_unit = spu_evnt_phys_spu_indx;
+
+	/* Set the user selected spu event to profile on,
+	 * note, only one SPU profiling event is supported
+	 */
+	num_counters = 1;  /* Only support one SPU event at a time */
+	set_pm_event(0, ctr[0].event, ctr[0].unit_mask);
+
+	reset_value[0] = 0xFFFFFFFF - ctr[0].count;
+
+	/* global, used by cell_cpu_setup */
+	ctr_enabled |= 1;
+
+	/* Initialize the count for each SPU to the reset value */
+	for (i=0; i < MAX_NUMNODES * NUM_SPUS_PER_NODE; i++)
+		spu_pm_cnt[i] = reset_value[0];
+
+	return 0;
+}
+
+static int cell_reg_setup_ppu(struct op_counter_config *ctr,
+			struct op_system_config *sys, int num_ctrs)
+{
+	/* routine is called once for all nodes */
+	int i, j, cpu;
+
 	num_counters = num_ctrs;
 
 	if (unlikely(num_ctrs > NR_PHYS_CTRS)) {
@@ -589,14 +772,6 @@ static int cell_reg_setup(struct op_counter_config *ctr,
 		       __func__);
 		return -EIO;
 	}
-	pm_regs.group_control = 0;
-	pm_regs.debug_bus_control = 0;
-
-	/* setup the pm_control register */
-	memset(&pm_regs.pm_cntrl, 0, sizeof(struct pm_cntrl));
-	pm_regs.pm_cntrl.stop_at_max = 1;
-	pm_regs.pm_cntrl.trace_mode = 0;
-	pm_regs.pm_cntrl.freeze = 1;
 
 	set_count_mode(sys->enable_kernel, sys->enable_user);
 
@@ -665,6 +840,63 @@ static int cell_reg_setup(struct op_counter_config *ctr,
 }
 
 
+/* This function is called once for all cpus combined */
+static int cell_reg_setup(struct op_counter_config *ctr,
+			struct op_system_config *sys, int num_ctrs)
+{
+	int ret=0;
+	spu_cycle_reset = 0;
+
+	/* initialize the spu_arr_trace value, will be reset if
+	 * doing spu event profiling.
+	 */
+	pm_regs.group_control = 0;
+	pm_regs.debug_bus_control = 0;
+	pm_regs.pm_cntrl.stop_at_max = 1;
+	pm_regs.pm_cntrl.trace_mode = 0;
+	pm_regs.pm_cntrl.freeze = 1;
+	pm_regs.pm_cntrl.trace_buf_ovflw = 0;
+	pm_regs.pm_cntrl.spu_addr_trace = 0;
+
+	/*
+	 * For all events except 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.
+	 */
+	pm_rtas_token = rtas_token("ibm,cbe-perftools");
+
+	if (unlikely(pm_rtas_token == RTAS_UNKNOWN_SERVICE)) {
+		printk(KERN_ERR
+		       "%s: rtas token ibm,cbe-perftools unknown\n",
+		       __func__);
+		return -EIO;
+	}
+
+	if (ctr[0].event == SPU_CYCLES_EVENT_NUM) {
+		profiling_mode = SPU_PROFILING_CYCLES;
+		ret = cell_reg_setup_spu_cycles(ctr, sys, num_ctrs);
+	} else if ((ctr[0].event >= SPU_EVENT_NUM_START) &&
+		   (ctr[0].event <= SPU_EVENT_NUM_STOP)) {
+		profiling_mode = SPU_PROFILING_EVENTS;
+		spu_cycle_reset = ctr[0].count;
+
+		/* for SPU event profiling, need to setup the
+		 * pm_signal array with the events to route the
+		 * SPU PC before making the FW call.  Note, only
+		 * one SPU event for profiling can be specified
+		 * at a time.
+		 */
+		cell_reg_setup_spu_events(ctr, sys, num_ctrs);
+	} else {
+		profiling_mode = PPU_PROFILING;
+		ret = cell_reg_setup_ppu(ctr, sys, num_ctrs);
+	}
+
+	return ret;
+}
+
+
 
 /* This function is called once for each cpu */
 static int cell_cpu_setup(struct op_counter_config *cntr)
@@ -672,8 +904,13 @@ static int cell_cpu_setup(struct op_counter_config *cntr)
 	u32 cpu = smp_processor_id();
 	u32 num_enabled = 0;
 	int i;
+	int ret;
 
-	if (spu_cycle_reset)
+	/* Cycle based SPU profiling does not use the performance
+	 * counters.  The trace array is configured to collect
+	 * the data.
+	 */
+	if (profiling_mode == SPU_PROFILING_CYCLES)
 		return 0;
 
 	/* There is one performance monitor per processor chip (i.e. node),
@@ -686,7 +923,6 @@ static int cell_cpu_setup(struct op_counter_config *cntr)
 	cbe_disable_pm(cpu);
 	cbe_disable_pm_interrupts(cpu);
 
-	cbe_write_pm(cpu, pm_interval, 0);
 	cbe_write_pm(cpu, pm_start_stop, 0);
 	cbe_write_pm(cpu, group_control, pm_regs.group_control);
 	cbe_write_pm(cpu, debug_bus_control, pm_regs.debug_bus_control);
@@ -703,7 +939,20 @@ static int cell_cpu_setup(struct op_counter_config *cntr)
 	 * 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);
+	if (profiling_mode == SPU_PROFILING_EVENTS) {
+		/* For SPU event profiling also need to setup the
+		 * pm interval timer
+		 */
+		ret = pm_rtas_activate_signals(cbe_cpu_to_node(cpu),
+					       num_enabled+2);
+		/* store PC from debug bus to Trace buffer as often
+		 * as possible (every 10 cycles)
+		 */
+		cbe_write_pm(cpu, pm_interval, NUM_INTERVAL_CYC);
+		return ret;
+	} else
+		return pm_rtas_activate_signals(cbe_cpu_to_node(cpu),
+						num_enabled);
 }
 
 #define ENTRIES	 303
@@ -885,7 +1134,122 @@ static struct notifier_block cpu_freq_notifier_block = {
 };
 #endif
 
-static int cell_global_start_spu(struct op_counter_config *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_cycles(void)
+{
+	int subfunc, rtn_value;
+	unsigned int lfsr_value;
+	int cpu;
+
+	oprofile_running = 0;
+	smp_wmb();
+
+#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",
+			       __func__, rtn_value);
+		}
+
+		/* Deactivate the signals */
+		pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
+	}
+
+	stop_spu_profiling_cycles();
+}
+
+static void cell_global_stop_spu_events(void)
+{
+	int cpu;
+	oprofile_running = 0;
+
+	stop_spu_profiling_events();
+	smp_wmb();
+
+	for_each_online_cpu(cpu) {
+		if (cbe_get_hw_thread_id(cpu))
+			continue;
+
+		cbe_sync_irq(cbe_cpu_to_node(cpu));
+		/* Stop the counters */
+		cbe_disable_pm(cpu);
+		cbe_write_pm07_control(cpu, 0, 0);
+
+		/* Deactivate the signals */
+		pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
+
+		/* Deactivate interrupts */
+		cbe_disable_pm_interrupts(cpu);
+	}
+	del_timer_sync(&timer_spu_event_swap);
+}
+
+static void cell_global_stop_ppu(void)
+{
+	int cpu;
+
+	/*
+	 * 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.
+	 */
+	del_timer_sync(&timer_virt_cntr);
+	oprofile_running = 0;
+	smp_wmb();
+
+	for_each_online_cpu(cpu) {
+		if (cbe_get_hw_thread_id(cpu))
+			continue;
+
+		cbe_sync_irq(cbe_cpu_to_node(cpu));
+		/* Stop the counters */
+		cbe_disable_pm(cpu);
+
+		/* Deactivate the signals */
+		pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
+
+		/* Deactivate interrupts */
+		cbe_disable_pm_interrupts(cpu);
+	}
+}
+
+static void cell_global_stop(void)
+{
+	if (profiling_mode == PPU_PROFILING)
+		cell_global_stop_ppu();
+	else if (profiling_mode == SPU_PROFILING_EVENTS)
+		cell_global_stop_spu_events();
+	else
+		cell_global_stop_spu_cycles();
+}
+
+static int cell_global_start_spu_cycles(struct op_counter_config *ctr)
 {
 	int subfunc;
 	unsigned int lfsr_value;
@@ -951,18 +1315,18 @@ static int cell_global_start_spu(struct op_counter_config *ctr)
 
 		/* start profiling */
 		ret = rtas_call(spu_rtas_token, 3, 1, NULL, subfunc,
-		  cbe_cpu_to_node(cpu), lfsr_value);
+				cbe_cpu_to_node(cpu), lfsr_value);
 
 		if (unlikely(ret != 0)) {
 			printk(KERN_ERR
-			       "%s: rtas call ibm,cbe-spu-perftools failed, return = %d\n",
+			       "%s: rtas call ibm,cbe-spu-perftools failed, " \
-			       __func__, ret);
+			       "return = %d\n", __func__, ret);
 			rtas_error = -EIO;
 			goto out;
 		}
 	}
 
-	rtas_error = start_spu_profiling(spu_cycle_reset);
+	rtas_error = start_spu_profiling_cycles(spu_cycle_reset);
 	if (rtas_error)
 		goto out_stop;
 
@@ -970,11 +1334,74 @@ static int cell_global_start_spu(struct op_counter_config *ctr)
 	return 0;
 
 out_stop:
-	cell_global_stop_spu();		/* clean up the PMU/debug bus */
+	cell_global_stop_spu_cycles();	/* clean up the PMU/debug bus */
 out:
 	return rtas_error;
 }
 
+static int cell_global_start_spu_events(struct op_counter_config *ctr)
+{
+	int cpu;
+	u32 interrupt_mask = 0;
+	int rtn = 0;
+
+	hdw_thread = 0;
+
+	/* spu event profiling, uses the performance counters to generate
+	 * an interrupt.  The hardware is setup to store the SPU program
+	 * counter into the trace array.  The occurrence mode is used to
+	 * enable storing data to the trace buffer.  The bits are set
+	 * to send/store the SPU address in the trace buffer.  The debug
+	 * bus must be setup to route the SPU program counter onto the
+	 * debug bus.  The occurrence data in the trace buffer is not used.
+	 */
+
+	/* This routine gets called once for the system.
+	 * There is one performance monitor per node, so we
+	 * only need to perform this function once per node.
+	 */
+
+	for_each_online_cpu(cpu) {
+		if (cbe_get_hw_thread_id(cpu))
+			continue;
+
+		/*
+		 * Setup SPU event-based profiling.
+		 * Set perf_mon_control bit 0 to a zero before
+		 * enabling spu collection hardware.
+		 *
+		 * Only support one SPU event on one SPU per node.
+		 */
+		if (ctr_enabled & 1) {
+			cbe_write_ctr(cpu, 0, reset_value[0]);
+			enable_ctr(cpu, 0, pm_regs.pm07_cntrl);
+			interrupt_mask |=
+				CBE_PM_CTR_OVERFLOW_INTR(0);
+		} else {
+			/* Disable counter */
+			cbe_write_pm07_control(cpu, 0, 0);
+		}
+
+		cbe_get_and_clear_pm_interrupts(cpu);
+		cbe_enable_pm_interrupts(cpu, hdw_thread, interrupt_mask);
+		cbe_enable_pm(cpu);
+
+		/* clear the trace buffer */
+		cbe_write_pm(cpu, trace_address, 0);
+	}
+
+	/* Start the timer to time slice collecting the event profile
+	 * on each of the SPUs.  Note, can collect profile on one SPU
+	 * per node at a time.
+	 */
+	start_spu_event_swap();
+	start_spu_profiling_events();
+	oprofile_running = 1;
+	smp_wmb();
+
+	return rtn;
+}
+
 static int cell_global_start_ppu(struct op_counter_config *ctr)
 {
 	u32 cpu, i;
@@ -994,8 +1421,7 @@ static int cell_global_start_ppu(struct op_counter_config *ctr)
 			if (ctr_enabled & (1 << i)) {
 				cbe_write_ctr(cpu, i, reset_value[i]);
 				enable_ctr(cpu, i, pm_regs.pm07_cntrl);
-				interrupt_mask |=
+				interrupt_mask |= CBE_PM_CTR_OVERFLOW_INTR(i);
-				    CBE_PM_CTR_OVERFLOW_INTR(i);
 			} else {
 				/* Disable counter */
 				cbe_write_pm07_control(cpu, i, 0);
@@ -1024,99 +1450,162 @@ static int cell_global_start_ppu(struct op_counter_config *ctr)
 
 static int cell_global_start(struct op_counter_config *ctr)
 {
-	if (spu_cycle_reset)
+	if (profiling_mode == SPU_PROFILING_CYCLES)
-		return cell_global_start_spu(ctr);
+		return cell_global_start_spu_cycles(ctr);
+	else if (profiling_mode == SPU_PROFILING_EVENTS)
+		return cell_global_start_spu_events(ctr);
 	else
 		return cell_global_start_ppu(ctr);
 }
 
-/*
+
- * Note the generic OProfile stop calls do not support returning
+/* The SPU interrupt handler
- * 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.
+ * SPU event profiling works as follows:
- * Failure to disable the SPU profiling is not an issue.  The FW calls
+ * The pm_signal[0] holds the one SPU event to be measured.  It is routed on
- * to enable the performance counters and debug bus will work even if
+ * the debug bus using word 0 or 1.  The value of pm_signal[1] and
- * the hardware was not cleanly reset.
+ * pm_signal[2] contain the necessary events to route the SPU program
+ * counter for the selected SPU onto the debug bus using words 2 and 3.
+ * The pm_interval register is setup to write the SPU PC value into the
+ * trace buffer at the maximum rate possible.  The trace buffer is configured
+ * to store the PCs, wrapping when it is full.  The performance counter is
+ * intialized to the max hardware count minus the number of events, N, between
+ * samples.  Once the N events have occured, a HW counter overflow occurs
+ * causing the generation of a HW counter interrupt which also stops the
+ * writing of the SPU PC values to the trace buffer.  Hence the last PC
+ * written to the trace buffer is the SPU PC that we want.  Unfortunately,
+ * we have to read from the beginning of the trace buffer to get to the
+ * last value written.  We just hope the PPU has nothing better to do then
+ * service this interrupt. The PC for the specific SPU being profiled is
+ * extracted from the trace buffer processed and stored.  The trace buffer
+ * is cleared, interrupts are cleared, the counter is reset to max - N.
+ * A kernel timer is used to periodically call the routine spu_evnt_swap()
+ * to switch to the next physical SPU in the node to profile in round robbin
+ * order.  This way data is collected for all SPUs on the node. It does mean
+ * that we need to use a relatively small value of N to ensure enough samples
+ * on each SPU are collected each SPU is being profiled 1/8 of the time.
+ * It may also be necessary to use a longer sample collection period.
  */
-static void cell_global_stop_spu(void)
+static void cell_handle_interrupt_spu(struct pt_regs *regs,
+				      struct op_counter_config *ctr)
 {
-	int subfunc, rtn_value;
+	u32 cpu, cpu_tmp;
-	unsigned int lfsr_value;
+	u64 trace_entry;
-	int cpu;
+	u32 interrupt_mask;
+	u64 trace_buffer[2];
+	u64 last_trace_buffer;
+	u32 sample;
+	u32 trace_addr;
+	unsigned long sample_array_lock_flags;
+	int spu_num;
+	unsigned long flags;
 
-	oprofile_running = 0;
+	/* Make sure spu event interrupt handler and spu event swap
+	 * don't access the counters simultaneously.
+	 */
+	cpu = smp_processor_id();
+	spin_lock_irqsave(&cntr_lock, flags);
 
-#ifdef CONFIG_CPU_FREQ
+	cpu_tmp = cpu;
-	cpufreq_unregister_notifier(&cpu_freq_notifier_block,
+	cbe_disable_pm(cpu);
-				    CPUFREQ_TRANSITION_NOTIFIER);
-#endif
 
-	for_each_online_cpu(cpu) {
+	interrupt_mask = cbe_get_and_clear_pm_interrupts(cpu);
-		if (cbe_get_hw_thread_id(cpu))
-			continue;
 
-		subfunc = 3;	/*
+	sample = 0xABCDEF;
-				 * 2 - activate SPU tracing,
+	trace_entry = 0xfedcba;
-				 * 3 - deactivate
+	last_trace_buffer = 0xdeadbeaf;
-				 */
-		lfsr_value = 0x8f100000;
 
-		rtn_value = rtas_call(spu_rtas_token, 3, 1, NULL,
+	if ((oprofile_running == 1) && (interrupt_mask != 0)) {
-				      subfunc, cbe_cpu_to_node(cpu),
+		/* disable writes to trace buff */
-				      lfsr_value);
+		cbe_write_pm(cpu, pm_interval, 0);
 
-		if (unlikely(rtn_value != 0)) {
+		/* only have one perf cntr being used, cntr 0 */
-			printk(KERN_ERR
+		if ((interrupt_mask & CBE_PM_CTR_OVERFLOW_INTR(0))
-			       "%s: rtas call ibm,cbe-spu-perftools failed, return = %d\n",
+		    && ctr[0].enabled)
-			       __func__, rtn_value);
+			/* The SPU PC values will be read
+			 * from the trace buffer, reset counter
+			 */
+
+			cbe_write_ctr(cpu, 0, reset_value[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
+			 * Read the buffer until you get to the last
+			 * entry.  This is the value we want.
+			 */
+
+			cbe_read_trace_buffer(cpu, trace_buffer);
+			trace_addr = cbe_read_pm(cpu, trace_address);
 		}
 
-		/* Deactivate the signals */
+		/* SPU Address 16 bit count format for 128 bit
-		pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
+		 * HW trace buffer is used for the SPU PC storage
+		 *    HDR bits          0:15
+		 *    SPU Addr 0 bits   16:31
+		 *    SPU Addr 1 bits   32:47
+		 *    unused bits       48:127
+		 *
+		 * HDR: bit4 = 1 SPU Address 0 valid
+		 * HDR: bit5 = 1 SPU Address 1 valid
+		 *  - unfortunately, the valid bits don't seem to work
+		 *
+		 * Note trace_buffer[0] holds bits 0:63 of the HW
+		 * trace buffer, trace_buffer[1] holds bits 64:127
+		 */
+
+		trace_entry = trace_buffer[0]
+			& 0x00000000FFFF0000;
+
+		/* only top 16 of the 18 bit SPU PC address
+		 * is stored in trace buffer, hence shift right
+		 * by 16 -2 bits */
+		sample = trace_entry >> 14;
+		last_trace_buffer = trace_buffer[0];
+
+		spu_num = spu_evnt_phys_spu_indx
+			+ (cbe_cpu_to_node(cpu) * NUM_SPUS_PER_NODE);
+
+		/* make sure only one process at a time is calling
+		 * spu_sync_buffer()
+		 */
+		spin_lock_irqsave(&oprof_spu_smpl_arry_lck,
+				  sample_array_lock_flags);
+		spu_sync_buffer(spu_num, &sample, 1);
+		spin_unlock_irqrestore(&oprof_spu_smpl_arry_lck,
+				       sample_array_lock_flags);
+
+		smp_wmb();    /* insure spu event buffer updates are written
+			       * don't want events intermingled... */
+
+		/* The counters were frozen by the interrupt.
+		 * Reenable the interrupt and restart the counters.
+		 */
+		cbe_write_pm(cpu, pm_interval, NUM_INTERVAL_CYC);
+		cbe_enable_pm_interrupts(cpu, hdw_thread,
+					 virt_cntr_inter_mask);
+
+		/* clear the trace buffer, re-enable writes to trace buff */
+		cbe_write_pm(cpu, trace_address, 0);
+		cbe_write_pm(cpu, pm_interval, NUM_INTERVAL_CYC);
+
+		/* 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
+		 * be disabled while writing to the latches.  This is a
+		 * HW design issue.
+		 */
+		write_pm_cntrl(cpu);
+		cbe_enable_pm(cpu);
 	}
-
+	spin_unlock_irqrestore(&cntr_lock, flags);
-	stop_spu_profiling();
 }
 
-static void cell_global_stop_ppu(void)
+static void cell_handle_interrupt_ppu(struct pt_regs *regs,
-{
+				      struct op_counter_config *ctr)
-	int cpu;
-
-	/*
-	 * 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.
-	 */
-	del_timer_sync(&timer_virt_cntr);
-	oprofile_running = 0;
-	smp_wmb();
-
-	for_each_online_cpu(cpu) {
-		if (cbe_get_hw_thread_id(cpu))
-			continue;
-
-		cbe_sync_irq(cbe_cpu_to_node(cpu));
-		/* Stop the counters */
-		cbe_disable_pm(cpu);
-
-		/* Deactivate the signals */
-		pm_rtas_reset_signals(cbe_cpu_to_node(cpu));
-
-		/* Deactivate interrupts */
-		cbe_disable_pm_interrupts(cpu);
-	}
-}
-
-static void cell_global_stop(void)
-{
-	if (spu_cycle_reset)
-		cell_global_stop_spu();
-	else
-		cell_global_stop_ppu();
-}
-
-static void cell_handle_interrupt(struct pt_regs *regs,
-				struct op_counter_config *ctr)
 {
 	u32 cpu;
 	u64 pc;
@@ -1132,7 +1621,7 @@ static void cell_handle_interrupt(struct pt_regs *regs,
 	 * routine are not running at the same time. See the
 	 * cell_virtual_cntr() routine for additional comments.
 	 */
-	spin_lock_irqsave(&virt_cntr_lock, flags);
+	spin_lock_irqsave(&cntr_lock, flags);
 
 	/*
 	 * Need to disable and reenable the performance counters
@@ -1185,7 +1674,16 @@ static void cell_handle_interrupt(struct pt_regs *regs,
 		 */
 		cbe_enable_pm(cpu);
 	}
-	spin_unlock_irqrestore(&virt_cntr_lock, flags);
+	spin_unlock_irqrestore(&cntr_lock, flags);
+}
+
+static void cell_handle_interrupt(struct pt_regs *regs,
+				  struct op_counter_config *ctr)
+{
+	if (profiling_mode == PPU_PROFILING)
+		cell_handle_interrupt_ppu(regs, ctr);
+	else
+		cell_handle_interrupt_spu(regs, ctr);
 }
 
 /*
@@ -1195,7 +1693,8 @@ static void cell_handle_interrupt(struct pt_regs *regs,
  */
 static int cell_sync_start(void)
 {
-	if (spu_cycle_reset)
+	if ((profiling_mode == SPU_PROFILING_CYCLES) ||
+	    (profiling_mode == SPU_PROFILING_EVENTS))
 		return spu_sync_start();
 	else
 		return DO_GENERIC_SYNC;
@@ -1203,7 +1702,8 @@ static int cell_sync_start(void)
 
 static int cell_sync_stop(void)
 {
-	if (spu_cycle_reset)
+	if ((profiling_mode == SPU_PROFILING_CYCLES) ||
+	    (profiling_mode == SPU_PROFILING_EVENTS))
 		return spu_sync_stop();
 	else
 		return 1;

arch/x86/oprofile/op_model_amd.c

@@ -2,7 +2,7 @@
  * @file op_model_amd.c
  * athlon / K7 / K8 / Family 10h model-specific MSR operations
  *
- * @remark Copyright 2002-2008 OProfile authors
+ * @remark Copyright 2002-2009 OProfile authors
  * @remark Read the file COPYING
  *
  * @author John Levon
@@ -10,7 +10,7 @@
  * @author Graydon Hoare
  * @author Robert Richter <robert.richter@amd.com>
  * @author Barry Kasindorf
-*/
+ */
 
 #include <linux/oprofile.h>
 #include <linux/device.h>
@@ -60,53 +60,10 @@ static unsigned long reset_value[NUM_COUNTERS];
 #define IBS_OP_LOW_VALID_BIT		(1ULL<<18)	/* bit 18 */
 #define IBS_OP_LOW_ENABLE		(1ULL<<17)	/* bit 17 */
 
-/* Codes used in cpu_buffer.c */
+#define IBS_FETCH_SIZE	6
-/* This produces duplicate code, need to be fixed */
+#define IBS_OP_SIZE	12
-#define IBS_FETCH_BEGIN 3
-#define IBS_OP_BEGIN    4
 
-/*
+static int has_ibs;	/* AMD Family10h and later */
- * The function interface needs to be fixed, something like add
- * data. Should then be added to linux/oprofile.h.
- */
-extern void
-oprofile_add_ibs_sample(struct pt_regs * const regs,
-			unsigned int * const ibs_sample, int ibs_code);
-
-struct ibs_fetch_sample {
-	/* MSRC001_1031 IBS Fetch Linear Address Register */
-	unsigned int ibs_fetch_lin_addr_low;
-	unsigned int ibs_fetch_lin_addr_high;
-	/* MSRC001_1030 IBS Fetch Control Register */
-	unsigned int ibs_fetch_ctl_low;
-	unsigned int ibs_fetch_ctl_high;
-	/* MSRC001_1032 IBS Fetch Physical Address Register */
-	unsigned int ibs_fetch_phys_addr_low;
-	unsigned int ibs_fetch_phys_addr_high;
-};
-
-struct ibs_op_sample {
-	/* MSRC001_1034 IBS Op Logical Address Register (IbsRIP) */
-	unsigned int ibs_op_rip_low;
-	unsigned int ibs_op_rip_high;
-	/* MSRC001_1035 IBS Op Data Register */
-	unsigned int ibs_op_data1_low;
-	unsigned int ibs_op_data1_high;
-	/* MSRC001_1036 IBS Op Data 2 Register */
-	unsigned int ibs_op_data2_low;
-	unsigned int ibs_op_data2_high;
-	/* MSRC001_1037 IBS Op Data 3 Register */
-	unsigned int ibs_op_data3_low;
-	unsigned int ibs_op_data3_high;
-	/* MSRC001_1038 IBS DC Linear Address Register (IbsDcLinAd) */
-	unsigned int ibs_dc_linear_low;
-	unsigned int ibs_dc_linear_high;
-	/* MSRC001_1039 IBS DC Physical Address Register (IbsDcPhysAd) */
-	unsigned int ibs_dc_phys_low;
-	unsigned int ibs_dc_phys_high;
-};
-
-static int ibs_allowed;	/* AMD Family10h and later */
 
 struct op_ibs_config {
 	unsigned long op_enabled;
@@ -197,31 +154,29 @@ static inline int
 op_amd_handle_ibs(struct pt_regs * const regs,
 		  struct op_msrs const * const msrs)
 {
-	unsigned int low, high;
+	u32 low, high;
-	struct ibs_fetch_sample ibs_fetch;
+	u64 msr;
-	struct ibs_op_sample ibs_op;
+	struct op_entry entry;
 
-	if (!ibs_allowed)
+	if (!has_ibs)
 		return 1;
 
 	if (ibs_config.fetch_enabled) {
 		rdmsr(MSR_AMD64_IBSFETCHCTL, low, high);
 		if (high & IBS_FETCH_HIGH_VALID_BIT) {
-			ibs_fetch.ibs_fetch_ctl_high = high;
+			rdmsrl(MSR_AMD64_IBSFETCHLINAD, msr);
-			ibs_fetch.ibs_fetch_ctl_low = low;
+			oprofile_write_reserve(&entry, regs, msr,
-			rdmsr(MSR_AMD64_IBSFETCHLINAD, low, high);
+					       IBS_FETCH_CODE, IBS_FETCH_SIZE);
-			ibs_fetch.ibs_fetch_lin_addr_high = high;
+			oprofile_add_data(&entry, (u32)msr);
-			ibs_fetch.ibs_fetch_lin_addr_low = low;
+			oprofile_add_data(&entry, (u32)(msr >> 32));
-			rdmsr(MSR_AMD64_IBSFETCHPHYSAD, low, high);
+			oprofile_add_data(&entry, low);
-			ibs_fetch.ibs_fetch_phys_addr_high = high;
+			oprofile_add_data(&entry, high);
-			ibs_fetch.ibs_fetch_phys_addr_low = low;
+			rdmsrl(MSR_AMD64_IBSFETCHPHYSAD, msr);
-
+			oprofile_add_data(&entry, (u32)msr);
-			oprofile_add_ibs_sample(regs,
+			oprofile_add_data(&entry, (u32)(msr >> 32));
-						(unsigned int *)&ibs_fetch,
+			oprofile_write_commit(&entry);
-						IBS_FETCH_BEGIN);
 
 			/* reenable the IRQ */
-			rdmsr(MSR_AMD64_IBSFETCHCTL, low, high);
 			high &= ~IBS_FETCH_HIGH_VALID_BIT;
 			high |= IBS_FETCH_HIGH_ENABLE;
 			low &= IBS_FETCH_LOW_MAX_CNT_MASK;
@@ -232,30 +187,29 @@ op_amd_handle_ibs(struct pt_regs * const regs,
 	if (ibs_config.op_enabled) {
 		rdmsr(MSR_AMD64_IBSOPCTL, low, high);
 		if (low & IBS_OP_LOW_VALID_BIT) {
-			rdmsr(MSR_AMD64_IBSOPRIP, low, high);
+			rdmsrl(MSR_AMD64_IBSOPRIP, msr);
-			ibs_op.ibs_op_rip_low = low;
+			oprofile_write_reserve(&entry, regs, msr,
-			ibs_op.ibs_op_rip_high = high;
+					       IBS_OP_CODE, IBS_OP_SIZE);
-			rdmsr(MSR_AMD64_IBSOPDATA, low, high);
+			oprofile_add_data(&entry, (u32)msr);
-			ibs_op.ibs_op_data1_low = low;
+			oprofile_add_data(&entry, (u32)(msr >> 32));
-			ibs_op.ibs_op_data1_high = high;
+			rdmsrl(MSR_AMD64_IBSOPDATA, msr);
-			rdmsr(MSR_AMD64_IBSOPDATA2, low, high);
+			oprofile_add_data(&entry, (u32)msr);
-			ibs_op.ibs_op_data2_low = low;
+			oprofile_add_data(&entry, (u32)(msr >> 32));
-			ibs_op.ibs_op_data2_high = high;
+			rdmsrl(MSR_AMD64_IBSOPDATA2, msr);
-			rdmsr(MSR_AMD64_IBSOPDATA3, low, high);
+			oprofile_add_data(&entry, (u32)msr);
-			ibs_op.ibs_op_data3_low = low;
+			oprofile_add_data(&entry, (u32)(msr >> 32));
-			ibs_op.ibs_op_data3_high = high;
+			rdmsrl(MSR_AMD64_IBSOPDATA3, msr);
-			rdmsr(MSR_AMD64_IBSDCLINAD, low, high);
+			oprofile_add_data(&entry, (u32)msr);
-			ibs_op.ibs_dc_linear_low = low;
+			oprofile_add_data(&entry, (u32)(msr >> 32));
-			ibs_op.ibs_dc_linear_high = high;
+			rdmsrl(MSR_AMD64_IBSDCLINAD, msr);
-			rdmsr(MSR_AMD64_IBSDCPHYSAD, low, high);
+			oprofile_add_data(&entry, (u32)msr);
-			ibs_op.ibs_dc_phys_low = low;
+			oprofile_add_data(&entry, (u32)(msr >> 32));
-			ibs_op.ibs_dc_phys_high = high;
+			rdmsrl(MSR_AMD64_IBSDCPHYSAD, msr);
+			oprofile_add_data(&entry, (u32)msr);
+			oprofile_add_data(&entry, (u32)(msr >> 32));
+			oprofile_write_commit(&entry);
 
 			/* reenable the IRQ */
-			oprofile_add_ibs_sample(regs,
-						(unsigned int *)&ibs_op,
-						IBS_OP_BEGIN);
-			rdmsr(MSR_AMD64_IBSOPCTL, low, high);
 			high = 0;
 			low &= ~IBS_OP_LOW_VALID_BIT;
 			low |= IBS_OP_LOW_ENABLE;
@@ -305,14 +259,14 @@ static void op_amd_start(struct op_msrs const * const msrs)
 	}
 
 #ifdef CONFIG_OPROFILE_IBS
-	if (ibs_allowed && ibs_config.fetch_enabled) {
+	if (has_ibs && ibs_config.fetch_enabled) {
 		low = (ibs_config.max_cnt_fetch >> 4) & 0xFFFF;
 		high = ((ibs_config.rand_en & 0x1) << 25) /* bit 57 */
 			+ IBS_FETCH_HIGH_ENABLE;
 		wrmsr(MSR_AMD64_IBSFETCHCTL, low, high);
 	}
 
-	if (ibs_allowed && ibs_config.op_enabled) {
+	if (has_ibs && ibs_config.op_enabled) {
 		low = ((ibs_config.max_cnt_op >> 4) & 0xFFFF)
 			+ ((ibs_config.dispatched_ops & 0x1) << 19) /* bit 19 */
 			+ IBS_OP_LOW_ENABLE;
@@ -341,14 +295,14 @@ static void op_amd_stop(struct op_msrs const * const msrs)
 	}
 
 #ifdef CONFIG_OPROFILE_IBS
-	if (ibs_allowed && ibs_config.fetch_enabled) {
+	if (has_ibs && ibs_config.fetch_enabled) {
 		/* clear max count and enable */
 		low = 0;
 		high = 0;
 		wrmsr(MSR_AMD64_IBSFETCHCTL, low, high);
 	}
 
-	if (ibs_allowed && ibs_config.op_enabled) {
+	if (has_ibs && ibs_config.op_enabled) {
 		/* clear max count and enable */
 		low = 0;
 		high = 0;
@@ -409,6 +363,7 @@ static int init_ibs_nmi(void)
 				       | IBSCTL_LVTOFFSETVAL);
 		pci_read_config_dword(cpu_cfg, IBSCTL, &value);
 		if (value != (ibs_eilvt_off | IBSCTL_LVTOFFSETVAL)) {
+			pci_dev_put(cpu_cfg);
 			printk(KERN_DEBUG "Failed to setup IBS LVT offset, "
 				"IBSCTL = 0x%08x", value);
 			return 1;
@@ -436,20 +391,20 @@ static int init_ibs_nmi(void)
 /* uninitialize the APIC for the IBS interrupts if needed */
 static void clear_ibs_nmi(void)
 {
-	if (ibs_allowed)
+	if (has_ibs)
 		on_each_cpu(apic_clear_ibs_nmi_per_cpu, NULL, 1);
 }
 
 /* initialize the APIC for the IBS interrupts if available */
 static void ibs_init(void)
 {
-	ibs_allowed = boot_cpu_has(X86_FEATURE_IBS);
+	has_ibs = boot_cpu_has(X86_FEATURE_IBS);
 
-	if (!ibs_allowed)
+	if (!has_ibs)
 		return;
 
 	if (init_ibs_nmi()) {
-		ibs_allowed = 0;
+		has_ibs = 0;
 		return;
 	}
 
@@ -458,7 +413,7 @@ static void ibs_init(void)
 
 static void ibs_exit(void)
 {
-	if (!ibs_allowed)
+	if (!has_ibs)
 		return;
 
 	clear_ibs_nmi();
@@ -478,7 +433,7 @@ static int setup_ibs_files(struct super_block *sb, struct dentry *root)
 	if (ret)
 		return ret;
 
-	if (!ibs_allowed)
+	if (!has_ibs)
 		return ret;
 
 	/* model specific files */

drivers/oprofile/buffer_sync.c

@@ -1,11 +1,12 @@
 /**
  * @file buffer_sync.c
  *
- * @remark Copyright 2002 OProfile authors
+ * @remark Copyright 2002-2009 OProfile authors
  * @remark Read the file COPYING
  *
  * @author John Levon <levon@movementarian.org>
  * @author Barry Kasindorf
+ * @author Robert Richter <robert.richter@amd.com>
  *
  * This is the core of the buffer management. Each
  * CPU buffer is processed and entered into the
@@ -315,88 +316,73 @@ static void add_trace_begin(void)
 	add_event_entry(TRACE_BEGIN_CODE);
 }
 
-#ifdef CONFIG_OPROFILE_IBS
+static void add_data(struct op_entry *entry, struct mm_struct *mm)
-
-#define IBS_FETCH_CODE_SIZE	2
-#define IBS_OP_CODE_SIZE	5
-
-/*
- * Add IBS fetch and op entries to event buffer
- */
-static void add_ibs_begin(int cpu, int code, struct mm_struct *mm)
 {
-	unsigned long rip;
+	unsigned long code, pc, val;
-	int i, count;
+	unsigned long cookie;
-	unsigned long ibs_cookie = 0;
 	off_t offset;
-	struct op_sample *sample;
 
-	sample = cpu_buffer_read_entry(cpu);
+	if (!op_cpu_buffer_get_data(entry, &code))
-	if (!sample)
+		return;
-		goto Error;
+	if (!op_cpu_buffer_get_data(entry, &pc))
-	rip = sample->eip;
+		return;
-
+	if (!op_cpu_buffer_get_size(entry))
-#ifdef __LP64__
+		return;
-	rip += sample->event << 32;
-#endif
 
 	if (mm) {
-		ibs_cookie = lookup_dcookie(mm, rip, &offset);
+		cookie = lookup_dcookie(mm, pc, &offset);
 
-		if (ibs_cookie == NO_COOKIE)
+		if (cookie == NO_COOKIE)
-			offset = rip;
+			offset = pc;
-		if (ibs_cookie == INVALID_COOKIE) {
+		if (cookie == INVALID_COOKIE) {
 			atomic_inc(&oprofile_stats.sample_lost_no_mapping);
-			offset = rip;
+			offset = pc;
 		}
-		if (ibs_cookie != last_cookie) {
+		if (cookie != last_cookie) {
-			add_cookie_switch(ibs_cookie);
+			add_cookie_switch(cookie);
-			last_cookie = ibs_cookie;
+			last_cookie = cookie;
 		}
 	} else
-		offset = rip;
+		offset = pc;
 
 	add_event_entry(ESCAPE_CODE);
 	add_event_entry(code);
 	add_event_entry(offset);	/* Offset from Dcookie */
 
-	/* we send the Dcookie offset, but send the raw Linear Add also*/
+	while (op_cpu_buffer_get_data(entry, &val))
-	add_event_entry(sample->eip);
+		add_event_entry(val);
-	add_event_entry(sample->event);
-
-	if (code == IBS_FETCH_CODE)
-		count = IBS_FETCH_CODE_SIZE;	/*IBS FETCH is 2 int64s*/
-	else
-		count = IBS_OP_CODE_SIZE;	/*IBS OP is 5 int64s*/
-
-	for (i = 0; i < count; i++) {
-		sample = cpu_buffer_read_entry(cpu);
-		if (!sample)
-			goto Error;
-		add_event_entry(sample->eip);
-		add_event_entry(sample->event);
-	}
-
-	return;
-
-Error:
-	return;
 }
 
-#endif
+static inline void add_sample_entry(unsigned long offset, unsigned long event)
-
-static void add_sample_entry(unsigned long offset, unsigned long event)
 {
 	add_event_entry(offset);
 	add_event_entry(event);
 }
 
 
-static int add_us_sample(struct mm_struct *mm, struct op_sample *s)
+/*
+ * Add a sample to the global event buffer. If possible the
+ * sample is converted into a persistent dentry/offset pair
+ * for later lookup from userspace. Return 0 on failure.
+ */
+static int
+add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel)
 {
 	unsigned long cookie;
 	off_t offset;
 
+	if (in_kernel) {
+		add_sample_entry(s->eip, s->event);
+		return 1;
+	}
+
+	/* add userspace sample */
+
+	if (!mm) {
+		atomic_inc(&oprofile_stats.sample_lost_no_mm);
+		return 0;
+	}
+
 	cookie = lookup_dcookie(mm, s->eip, &offset);
 
 	if (cookie == INVALID_COOKIE) {
@@ -415,25 +401,6 @@ static int add_us_sample(struct mm_struct *mm, struct op_sample *s)
 }
 
 
-/* Add a sample to the global event buffer. If possible the
- * sample is converted into a persistent dentry/offset pair
- * for later lookup from userspace.
- */
-static int
-add_sample(struct mm_struct *mm, struct op_sample *s, int in_kernel)
-{
-	if (in_kernel) {
-		add_sample_entry(s->eip, s->event);
-		return 1;
-	} else if (mm) {
-		return add_us_sample(mm, s);
-	} else {
-		atomic_inc(&oprofile_stats.sample_lost_no_mm);
-	}
-	return 0;
-}
-
-
 static void release_mm(struct mm_struct *mm)
 {
 	if (!mm)
@@ -526,66 +493,69 @@ void sync_buffer(int cpu)
 {
 	struct mm_struct *mm = NULL;
 	struct mm_struct *oldmm;
+	unsigned long val;
 	struct task_struct *new;
 	unsigned long cookie = 0;
 	int in_kernel = 1;
 	sync_buffer_state state = sb_buffer_start;
 	unsigned int i;
 	unsigned long available;
+	unsigned long flags;
+	struct op_entry entry;
+	struct op_sample *sample;
 
 	mutex_lock(&buffer_mutex);
 
 	add_cpu_switch(cpu);
 
-	cpu_buffer_reset(cpu);
+	op_cpu_buffer_reset(cpu);
-	available = cpu_buffer_entries(cpu);
+	available = op_cpu_buffer_entries(cpu);
 
 	for (i = 0; i < available; ++i) {
-		struct op_sample *s = cpu_buffer_read_entry(cpu);
+		sample = op_cpu_buffer_read_entry(&entry, cpu);
-		if (!s)
+		if (!sample)
 			break;
 
-		if (is_code(s->eip)) {
+		if (is_code(sample->eip)) {
-			switch (s->event) {
+			flags = sample->event;
-			case 0:
+			if (flags & TRACE_BEGIN) {
-			case CPU_IS_KERNEL:
-				/* kernel/userspace switch */
-				in_kernel = s->event;
-				if (state == sb_buffer_start)
-					state = sb_sample_start;
-				add_kernel_ctx_switch(s->event);
-				break;
-			case CPU_TRACE_BEGIN:
 				state = sb_bt_start;
 				add_trace_begin();
-				break;
+			}
-#ifdef CONFIG_OPROFILE_IBS
+			if (flags & KERNEL_CTX_SWITCH) {
-			case IBS_FETCH_BEGIN:
+				/* kernel/userspace switch */
-				state = sb_bt_start;
+				in_kernel = flags & IS_KERNEL;
-				add_ibs_begin(cpu, IBS_FETCH_CODE, mm);
+				if (state == sb_buffer_start)
-				break;
+					state = sb_sample_start;
-			case IBS_OP_BEGIN:
+				add_kernel_ctx_switch(flags & IS_KERNEL);
-				state = sb_bt_start;
+			}
-				add_ibs_begin(cpu, IBS_OP_CODE, mm);
+			if (flags & USER_CTX_SWITCH
-				break;
+			    && op_cpu_buffer_get_data(&entry, &val)) {
-#endif
-			default:
 				/* userspace context switch */
+				new = (struct task_struct *)val;
 				oldmm = mm;
-				new = (struct task_struct *)s->event;
 				release_mm(oldmm);
 				mm = take_tasks_mm(new);
 				if (mm != oldmm)
 					cookie = get_exec_dcookie(mm);
 				add_user_ctx_switch(new, cookie);
-				break;
-			}
-		} else if (state >= sb_bt_start &&
-			   !add_sample(mm, s, in_kernel)) {
-			if (state == sb_bt_start) {
-				state = sb_bt_ignore;
-				atomic_inc(&oprofile_stats.bt_lost_no_mapping);
 			}
+			if (op_cpu_buffer_get_size(&entry))
+				add_data(&entry, mm);
+			continue;
+		}
+
+		if (state < sb_bt_start)
+			/* ignore sample */
+			continue;
+
+		if (add_sample(mm, sample, in_kernel))
+			continue;
+
+		/* ignore backtraces if failed to add a sample */
+		if (state == sb_bt_start) {
+			state = sb_bt_ignore;
+			atomic_inc(&oprofile_stats.bt_lost_no_mapping);
 		}
 	}
 	release_mm(mm);

drivers/oprofile/cpu_buffer.c

@@ -1,11 +1,12 @@
 /**
  * @file cpu_buffer.c
  *
- * @remark Copyright 2002 OProfile authors
+ * @remark Copyright 2002-2009 OProfile authors
  * @remark Read the file COPYING
  *
  * @author John Levon <levon@movementarian.org>
  * @author Barry Kasindorf <barry.kasindorf@amd.com>
+ * @author Robert Richter <robert.richter@amd.com>
  *
  * Each CPU has a local buffer that stores PC value/event
  * pairs. We also log context switches when we notice them.
@@ -45,8 +46,8 @@
  * can be changed to a single buffer solution when the ring buffer
  * access is implemented as non-locking atomic code.
  */
-struct ring_buffer *op_ring_buffer_read;
+static struct ring_buffer *op_ring_buffer_read;
-struct ring_buffer *op_ring_buffer_write;
+static struct ring_buffer *op_ring_buffer_write;
 DEFINE_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
 
 static void wq_sync_buffer(struct work_struct *work);
@@ -54,6 +55,19 @@ static void wq_sync_buffer(struct work_struct *work);
 #define DEFAULT_TIMER_EXPIRE (HZ / 10)
 static int work_enabled;
 
+unsigned long oprofile_get_cpu_buffer_size(void)
+{
+	return oprofile_cpu_buffer_size;
+}
+
+void oprofile_cpu_buffer_inc_smpl_lost(void)
+{
+	struct oprofile_cpu_buffer *cpu_buf
+		= &__get_cpu_var(cpu_buffer);
+
+	cpu_buf->sample_lost_overflow++;
+}
+
 void free_cpu_buffers(void)
 {
 	if (op_ring_buffer_read)
@@ -64,24 +78,11 @@ void free_cpu_buffers(void)
 	op_ring_buffer_write = NULL;
 }
 
-unsigned long oprofile_get_cpu_buffer_size(void)
-{
-	return fs_cpu_buffer_size;
-}
-
-void oprofile_cpu_buffer_inc_smpl_lost(void)
-{
-	struct oprofile_cpu_buffer *cpu_buf
-		= &__get_cpu_var(cpu_buffer);
-
-	cpu_buf->sample_lost_overflow++;
-}
-
 int alloc_cpu_buffers(void)
 {
 	int i;
 
-	unsigned long buffer_size = fs_cpu_buffer_size;
+	unsigned long buffer_size = oprofile_cpu_buffer_size;
 
 	op_ring_buffer_read = ring_buffer_alloc(buffer_size, OP_BUFFER_FLAGS);
 	if (!op_ring_buffer_read)
@@ -97,8 +98,6 @@ int alloc_cpu_buffers(void)
 		b->last_is_kernel = -1;
 		b->tracing = 0;
 		b->buffer_size = buffer_size;
-		b->tail_pos = 0;
-		b->head_pos = 0;
 		b->sample_received = 0;
 		b->sample_lost_overflow = 0;
 		b->backtrace_aborted = 0;
@@ -145,47 +144,156 @@ void end_cpu_work(void)
 	flush_scheduled_work();
 }
 
-static inline int
+/*
-add_sample(struct oprofile_cpu_buffer *cpu_buf,
+ * This function prepares the cpu buffer to write a sample.
-	   unsigned long pc, unsigned long event)
+ *
+ * Struct op_entry is used during operations on the ring buffer while
+ * struct op_sample contains the data that is stored in the ring
+ * buffer. Struct entry can be uninitialized. The function reserves a
+ * data array that is specified by size. Use
+ * op_cpu_buffer_write_commit() after preparing the sample. In case of
+ * errors a null pointer is returned, otherwise the pointer to the
+ * sample.
+ *
+ */
+struct op_sample
+*op_cpu_buffer_write_reserve(struct op_entry *entry, unsigned long size)
+{
+	entry->event = ring_buffer_lock_reserve
+		(op_ring_buffer_write, sizeof(struct op_sample) +
+		 size * sizeof(entry->sample->data[0]), &entry->irq_flags);
+	if (entry->event)
+		entry->sample = ring_buffer_event_data(entry->event);
+	else
+		entry->sample = NULL;
+
+	if (!entry->sample)
+		return NULL;
+
+	entry->size = size;
+	entry->data = entry->sample->data;
+
+	return entry->sample;
+}
+
+int op_cpu_buffer_write_commit(struct op_entry *entry)
+{
+	return ring_buffer_unlock_commit(op_ring_buffer_write, entry->event,
+					 entry->irq_flags);
+}
+
+struct op_sample *op_cpu_buffer_read_entry(struct op_entry *entry, int cpu)
+{
+	struct ring_buffer_event *e;
+	e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
+	if (e)
+		goto event;
+	if (ring_buffer_swap_cpu(op_ring_buffer_read,
+				 op_ring_buffer_write,
+				 cpu))
+		return NULL;
+	e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
+	if (e)
+		goto event;
+	return NULL;
+
+event:
+	entry->event = e;
+	entry->sample = ring_buffer_event_data(e);
+	entry->size = (ring_buffer_event_length(e) - sizeof(struct op_sample))
+		/ sizeof(entry->sample->data[0]);
+	entry->data = entry->sample->data;
+	return entry->sample;
+}
+
+unsigned long op_cpu_buffer_entries(int cpu)
+{
+	return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
+		+ ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
+}
+
+static int
+op_add_code(struct oprofile_cpu_buffer *cpu_buf, unsigned long backtrace,
+	    int is_kernel, struct task_struct *task)
 {
 	struct op_entry entry;
-	int ret;
+	struct op_sample *sample;
+	unsigned long flags;
+	int size;
 
-	ret = cpu_buffer_write_entry(&entry);
+	flags = 0;
-	if (ret)
-		return ret;
 
-	entry.sample->eip = pc;
+	if (backtrace)
-	entry.sample->event = event;
+		flags |= TRACE_BEGIN;
 
-	ret = cpu_buffer_write_commit(&entry);
+	/* notice a switch from user->kernel or vice versa */
-	if (ret)
+	is_kernel = !!is_kernel;
-		return ret;
+	if (cpu_buf->last_is_kernel != is_kernel) {
+		cpu_buf->last_is_kernel = is_kernel;
+		flags |= KERNEL_CTX_SWITCH;
+		if (is_kernel)
+			flags |= IS_KERNEL;
+	}
+
+	/* notice a task switch */
+	if (cpu_buf->last_task != task) {
+		cpu_buf->last_task = task;
+		flags |= USER_CTX_SWITCH;
+	}
+
+	if (!flags)
+		/* nothing to do */
+		return 0;
+
+	if (flags & USER_CTX_SWITCH)
+		size = 1;
+	else
+		size = 0;
+
+	sample = op_cpu_buffer_write_reserve(&entry, size);
+	if (!sample)
+		return -ENOMEM;
+
+	sample->eip = ESCAPE_CODE;
+	sample->event = flags;
+
+	if (size)
+		op_cpu_buffer_add_data(&entry, (unsigned long)task);
+
+	op_cpu_buffer_write_commit(&entry);
 
 	return 0;
 }
 
 static inline int
-add_code(struct oprofile_cpu_buffer *buffer, unsigned long value)
+op_add_sample(struct oprofile_cpu_buffer *cpu_buf,
+	      unsigned long pc, unsigned long event)
 {
-	return add_sample(buffer, ESCAPE_CODE, value);
+	struct op_entry entry;
+	struct op_sample *sample;
+
+	sample = op_cpu_buffer_write_reserve(&entry, 0);
+	if (!sample)
+		return -ENOMEM;
+
+	sample->eip = pc;
+	sample->event = event;
+
+	return op_cpu_buffer_write_commit(&entry);
 }
 
-/* This must be safe from any context. It's safe writing here
+/*
- * because of the head/tail separation of the writer and reader
+ * This must be safe from any context.
- * of the CPU buffer.
  *
  * is_kernel is needed because on some architectures you cannot
  * tell if you are in kernel or user space simply by looking at
  * pc. We tag this in the buffer by generating kernel enter/exit
  * events whenever is_kernel changes
  */
-static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
+static int
-		      int is_kernel, unsigned long event)
+log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
+	   unsigned long backtrace, int is_kernel, unsigned long event)
 {
-	struct task_struct *task;
-
 	cpu_buf->sample_received++;
 
 	if (pc == ESCAPE_CODE) {
@@ -193,25 +301,10 @@ static int log_sample(struct oprofile_cpu_buffer *cpu_buf, unsigned long pc,
 		return 0;
 	}
 
-	is_kernel = !!is_kernel;
+	if (op_add_code(cpu_buf, backtrace, is_kernel, current))
+		goto fail;
 
-	task = current;
+	if (op_add_sample(cpu_buf, pc, event))
-
-	/* notice a switch from user->kernel or vice versa */
-	if (cpu_buf->last_is_kernel != is_kernel) {
-		cpu_buf->last_is_kernel = is_kernel;
-		if (add_code(cpu_buf, is_kernel))
-			goto fail;
-	}
-
-	/* notice a task switch */
-	if (cpu_buf->last_task != task) {
-		cpu_buf->last_task = task;
-		if (add_code(cpu_buf, (unsigned long)task))
-			goto fail;
-	}
-
-	if (add_sample(cpu_buf, pc, event))
 		goto fail;
 
 	return 1;
@@ -221,109 +314,102 @@ fail:
 	return 0;
 }
 
-static int oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
+static inline void oprofile_begin_trace(struct oprofile_cpu_buffer *cpu_buf)
 {
-	add_code(cpu_buf, CPU_TRACE_BEGIN);
 	cpu_buf->tracing = 1;
-	return 1;
 }
 
-static void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
+static inline void oprofile_end_trace(struct oprofile_cpu_buffer *cpu_buf)
 {
 	cpu_buf->tracing = 0;
 }
 
-void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
+static inline void
-				unsigned long event, int is_kernel)
+__oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
+			  unsigned long event, int is_kernel)
 {
 	struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
-
+	unsigned long backtrace = oprofile_backtrace_depth;
-	if (!backtrace_depth) {
-		log_sample(cpu_buf, pc, is_kernel, event);
-		return;
-	}
-
-	if (!oprofile_begin_trace(cpu_buf))
-		return;
 
 	/*
 	 * if log_sample() fail we can't backtrace since we lost the
 	 * source of this event
 	 */
-	if (log_sample(cpu_buf, pc, is_kernel, event))
+	if (!log_sample(cpu_buf, pc, backtrace, is_kernel, event))
-		oprofile_ops.backtrace(regs, backtrace_depth);
+		/* failed */
+		return;
+
+	if (!backtrace)
+		return;
+
+	oprofile_begin_trace(cpu_buf);
+	oprofile_ops.backtrace(regs, backtrace);
 	oprofile_end_trace(cpu_buf);
 }
 
+void oprofile_add_ext_sample(unsigned long pc, struct pt_regs * const regs,
+			     unsigned long event, int is_kernel)
+{
+	__oprofile_add_ext_sample(pc, regs, event, is_kernel);
+}
+
 void oprofile_add_sample(struct pt_regs * const regs, unsigned long event)
 {
 	int is_kernel = !user_mode(regs);
 	unsigned long pc = profile_pc(regs);
 
-	oprofile_add_ext_sample(pc, regs, event, is_kernel);
+	__oprofile_add_ext_sample(pc, regs, event, is_kernel);
 }
 
-#ifdef CONFIG_OPROFILE_IBS
+/*
-
+ * Add samples with data to the ring buffer.
-#define MAX_IBS_SAMPLE_SIZE 14
+ *
-
+ * Use oprofile_add_data(&entry, val) to add data and
-void oprofile_add_ibs_sample(struct pt_regs * const regs,
+ * oprofile_write_commit(&entry) to commit the sample.
-			     unsigned int * const ibs_sample, int ibs_code)
+ */
+void
+oprofile_write_reserve(struct op_entry *entry, struct pt_regs * const regs,
+		       unsigned long pc, int code, int size)
 {
+	struct op_sample *sample;
 	int is_kernel = !user_mode(regs);
 	struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
-	struct task_struct *task;
-	int fail = 0;
 
 	cpu_buf->sample_received++;
 
-	/* notice a switch from user->kernel or vice versa */
+	/* no backtraces for samples with data */
-	if (cpu_buf->last_is_kernel != is_kernel) {
+	if (op_add_code(cpu_buf, 0, is_kernel, current))
-		if (add_code(cpu_buf, is_kernel))
-			goto fail;
-		cpu_buf->last_is_kernel = is_kernel;
-	}
-
-	/* notice a task switch */
-	if (!is_kernel) {
-		task = current;
-		if (cpu_buf->last_task != task) {
-			if (add_code(cpu_buf, (unsigned long)task))
-				goto fail;
-			cpu_buf->last_task = task;
-		}
-	}
-
-	fail = fail || add_code(cpu_buf, ibs_code);
-	fail = fail || add_sample(cpu_buf, ibs_sample[0], ibs_sample[1]);
-	fail = fail || add_sample(cpu_buf, ibs_sample[2], ibs_sample[3]);
-	fail = fail || add_sample(cpu_buf, ibs_sample[4], ibs_sample[5]);
-
-	if (ibs_code == IBS_OP_BEGIN) {
-		fail = fail || add_sample(cpu_buf, ibs_sample[6], ibs_sample[7]);
-		fail = fail || add_sample(cpu_buf, ibs_sample[8], ibs_sample[9]);
-		fail = fail || add_sample(cpu_buf, ibs_sample[10], ibs_sample[11]);
-	}
-
-	if (fail)
 		goto fail;
 
-	if (backtrace_depth)
+	sample = op_cpu_buffer_write_reserve(entry, size + 2);
-		oprofile_ops.backtrace(regs, backtrace_depth);
+	if (!sample)
+		goto fail;
+	sample->eip = ESCAPE_CODE;
+	sample->event = 0;		/* no flags */
+
+	op_cpu_buffer_add_data(entry, code);
+	op_cpu_buffer_add_data(entry, pc);
 
 	return;
 
 fail:
 	cpu_buf->sample_lost_overflow++;
-	return;
 }
 
-#endif
+int oprofile_add_data(struct op_entry *entry, unsigned long val)
+{
+	return op_cpu_buffer_add_data(entry, val);
+}
+
+int oprofile_write_commit(struct op_entry *entry)
+{
+	return op_cpu_buffer_write_commit(entry);
+}
 
 void oprofile_add_pc(unsigned long pc, int is_kernel, unsigned long event)
 {
 	struct oprofile_cpu_buffer *cpu_buf = &__get_cpu_var(cpu_buffer);
-	log_sample(cpu_buf, pc, is_kernel, event);
+	log_sample(cpu_buf, pc, 0, is_kernel, event);
 }
 
 void oprofile_add_trace(unsigned long pc)
@@ -340,7 +426,7 @@ void oprofile_add_trace(unsigned long pc)
 	if (pc == ESCAPE_CODE)
 		goto fail;
 
-	if (add_sample(cpu_buf, pc, 0))
+	if (op_add_sample(cpu_buf, pc, 0))
 		goto fail;
 
 	return;

drivers/oprofile/cpu_buffer.h

@@ -1,10 +1,11 @@
 /**
  * @file cpu_buffer.h
  *
- * @remark Copyright 2002 OProfile authors
+ * @remark Copyright 2002-2009 OProfile authors
  * @remark Read the file COPYING
  *
  * @author John Levon <levon@movementarian.org>
+ * @author Robert Richter <robert.richter@amd.com>
  */
 
 #ifndef OPROFILE_CPU_BUFFER_H
@@ -31,17 +32,12 @@ void end_cpu_work(void);
 struct op_sample {
 	unsigned long eip;
 	unsigned long event;
+	unsigned long data[0];
 };
 
-struct op_entry {
+struct op_entry;
-	struct ring_buffer_event *event;
-	struct op_sample *sample;
-	unsigned long irq_flags;
-};
 
 struct oprofile_cpu_buffer {
-	volatile unsigned long head_pos;
-	volatile unsigned long tail_pos;
 	unsigned long buffer_size;
 	struct task_struct *last_task;
 	int last_is_kernel;
@@ -54,8 +50,6 @@ struct oprofile_cpu_buffer {
 	struct delayed_work work;
 };
 
-extern struct ring_buffer *op_ring_buffer_read;
-extern struct ring_buffer *op_ring_buffer_write;
 DECLARE_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
 
 /*
@@ -64,7 +58,7 @@ DECLARE_PER_CPU(struct oprofile_cpu_buffer, cpu_buffer);
  * reset these to invalid values; the next sample collected will
  * populate the buffer with proper values to initialize the buffer
  */
-static inline void cpu_buffer_reset(int cpu)
+static inline void op_cpu_buffer_reset(int cpu)
 {
 	struct oprofile_cpu_buffer *cpu_buf = &per_cpu(cpu_buffer, cpu);
 
@@ -72,55 +66,48 @@ static inline void cpu_buffer_reset(int cpu)
 	cpu_buf->last_task = NULL;
 }
 
-static inline int cpu_buffer_write_entry(struct op_entry *entry)
+struct op_sample
+*op_cpu_buffer_write_reserve(struct op_entry *entry, unsigned long size);
+int op_cpu_buffer_write_commit(struct op_entry *entry);
+struct op_sample *op_cpu_buffer_read_entry(struct op_entry *entry, int cpu);
+unsigned long op_cpu_buffer_entries(int cpu);
+
+/* returns the remaining free size of data in the entry */
+static inline
+int op_cpu_buffer_add_data(struct op_entry *entry, unsigned long val)
 {
-	entry->event = ring_buffer_lock_reserve(op_ring_buffer_write,
+	if (!entry->size)
-						sizeof(struct op_sample),
+		return 0;
-						&entry->irq_flags);
+	*entry->data = val;
-	if (entry->event)
+	entry->size--;
-		entry->sample = ring_buffer_event_data(entry->event);
+	entry->data++;
-	else
+	return entry->size;
-		entry->sample = NULL;
-
-	if (!entry->sample)
-		return -ENOMEM;
-
-	return 0;
 }
 
-static inline int cpu_buffer_write_commit(struct op_entry *entry)
+/* returns the size of data in the entry */
+static inline
+int op_cpu_buffer_get_size(struct op_entry *entry)
 {
-	return ring_buffer_unlock_commit(op_ring_buffer_write, entry->event,
+	return entry->size;
-					 entry->irq_flags);
 }
 
-static inline struct op_sample *cpu_buffer_read_entry(int cpu)
+/* returns 0 if empty or the size of data including the current value */
+static inline
+int op_cpu_buffer_get_data(struct op_entry *entry, unsigned long *val)
 {
-	struct ring_buffer_event *e;
+	int size = entry->size;
-	e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
+	if (!size)
-	if (e)
+		return 0;
-		return ring_buffer_event_data(e);
+	*val = *entry->data;
-	if (ring_buffer_swap_cpu(op_ring_buffer_read,
+	entry->size--;
-				 op_ring_buffer_write,
+	entry->data++;
-				 cpu))
+	return size;
-		return NULL;
-	e = ring_buffer_consume(op_ring_buffer_read, cpu, NULL);
-	if (e)
-		return ring_buffer_event_data(e);
-	return NULL;
 }
 
-/* "acquire" as many cpu buffer slots as we can */
+/* extra data flags */
-static inline unsigned long cpu_buffer_entries(int cpu)
+#define KERNEL_CTX_SWITCH	(1UL << 0)
-{
+#define IS_KERNEL		(1UL << 1)
-	return ring_buffer_entries_cpu(op_ring_buffer_read, cpu)
+#define TRACE_BEGIN		(1UL << 2)
-		+ ring_buffer_entries_cpu(op_ring_buffer_write, cpu);
+#define USER_CTX_SWITCH		(1UL << 3)
-}
-
-/* transient events for the CPU buffer -> event buffer */
-#define CPU_IS_KERNEL 1
-#define CPU_TRACE_BEGIN 2
-#define IBS_FETCH_BEGIN 3
-#define IBS_OP_BEGIN    4
 
 #endif /* OPROFILE_CPU_BUFFER_H */

drivers/oprofile/event_buffer.c

@@ -73,8 +73,8 @@ int alloc_event_buffer(void)
 	unsigned long flags;
 
 	spin_lock_irqsave(&oprofilefs_lock, flags);
-	buffer_size = fs_buffer_size;
+	buffer_size = oprofile_buffer_size;
-	buffer_watershed = fs_buffer_watershed;
+	buffer_watershed = oprofile_buffer_watershed;
 	spin_unlock_irqrestore(&oprofilefs_lock, flags);
 
 	if (buffer_watershed >= buffer_size)

drivers/oprofile/oprof.c

@@ -23,7 +23,7 @@
 struct oprofile_operations oprofile_ops;
 
 unsigned long oprofile_started;
-unsigned long backtrace_depth;
+unsigned long oprofile_backtrace_depth;
 static unsigned long is_setup;
 static DEFINE_MUTEX(start_mutex);
 
@@ -172,7 +172,7 @@ int oprofile_set_backtrace(unsigned long val)
 		goto out;
 	}
 
-	backtrace_depth = val;
+	oprofile_backtrace_depth = val;
 
 out:
 	mutex_unlock(&start_mutex);

drivers/oprofile/oprof.h

@@ -21,12 +21,12 @@ void oprofile_stop(void);
 
 struct oprofile_operations;
 
-extern unsigned long fs_buffer_size;
+extern unsigned long oprofile_buffer_size;
-extern unsigned long fs_cpu_buffer_size;
+extern unsigned long oprofile_cpu_buffer_size;
-extern unsigned long fs_buffer_watershed;
+extern unsigned long oprofile_buffer_watershed;
 extern struct oprofile_operations oprofile_ops;
 extern unsigned long oprofile_started;
-extern unsigned long backtrace_depth;
+extern unsigned long oprofile_backtrace_depth;
 
 struct super_block;
 struct dentry;

drivers/oprofile/oprofile_files.c

@@ -14,17 +14,18 @@
 #include "oprofile_stats.h"
 #include "oprof.h"
 
-#define FS_BUFFER_SIZE_DEFAULT		131072
+#define BUFFER_SIZE_DEFAULT		131072
-#define FS_CPU_BUFFER_SIZE_DEFAULT	8192
+#define CPU_BUFFER_SIZE_DEFAULT		8192
-#define FS_BUFFER_WATERSHED_DEFAULT	32768	/* FIXME: tune */
+#define BUFFER_WATERSHED_DEFAULT	32768	/* FIXME: tune */
 
-unsigned long fs_buffer_size;
+unsigned long oprofile_buffer_size;
-unsigned long fs_cpu_buffer_size;
+unsigned long oprofile_cpu_buffer_size;
-unsigned long fs_buffer_watershed;
+unsigned long oprofile_buffer_watershed;
 
 static ssize_t depth_read(struct file *file, char __user *buf, size_t count, loff_t *offset)
 {
-	return oprofilefs_ulong_to_user(backtrace_depth, buf, count, offset);
+	return oprofilefs_ulong_to_user(oprofile_backtrace_depth, buf, count,
+					offset);
 }
 
 
@@ -125,16 +126,16 @@ static const struct file_operations dump_fops = {
 void oprofile_create_files(struct super_block *sb, struct dentry *root)
 {
 	/* reinitialize default values */
-	fs_buffer_size =	FS_BUFFER_SIZE_DEFAULT;
+	oprofile_buffer_size =		BUFFER_SIZE_DEFAULT;
-	fs_cpu_buffer_size =	FS_CPU_BUFFER_SIZE_DEFAULT;
+	oprofile_cpu_buffer_size =	CPU_BUFFER_SIZE_DEFAULT;
-	fs_buffer_watershed =	FS_BUFFER_WATERSHED_DEFAULT;
+	oprofile_buffer_watershed =	BUFFER_WATERSHED_DEFAULT;
 
 	oprofilefs_create_file(sb, root, "enable", &enable_fops);
 	oprofilefs_create_file_perm(sb, root, "dump", &dump_fops, 0666);
 	oprofilefs_create_file(sb, root, "buffer", &event_buffer_fops);
-	oprofilefs_create_ulong(sb, root, "buffer_size", &fs_buffer_size);
+	oprofilefs_create_ulong(sb, root, "buffer_size", &oprofile_buffer_size);
-	oprofilefs_create_ulong(sb, root, "buffer_watershed", &fs_buffer_watershed);
+	oprofilefs_create_ulong(sb, root, "buffer_watershed", &oprofile_buffer_watershed);
-	oprofilefs_create_ulong(sb, root, "cpu_buffer_size", &fs_cpu_buffer_size);
+	oprofilefs_create_ulong(sb, root, "cpu_buffer_size", &oprofile_cpu_buffer_size);
 	oprofilefs_create_file(sb, root, "cpu_type", &cpu_type_fops);
 	oprofilefs_create_file(sb, root, "backtrace_depth", &depth_fops);
 	oprofilefs_create_file(sb, root, "pointer_size", &pointer_size_fops);

include/linux/oprofile.h

@@ -164,4 +164,22 @@ void oprofile_put_buff(unsigned long *buf, unsigned int start,
 unsigned long oprofile_get_cpu_buffer_size(void);
 void oprofile_cpu_buffer_inc_smpl_lost(void);
  
+/* cpu buffer functions */
+
+struct op_sample;
+
+struct op_entry {
+	struct ring_buffer_event *event;
+	struct op_sample *sample;
+	unsigned long irq_flags;
+	unsigned long size;
+	unsigned long *data;
+};
+
+void oprofile_write_reserve(struct op_entry *entry,
+			    struct pt_regs * const regs,
+			    unsigned long pc, int code, int size);
+int oprofile_add_data(struct op_entry *entry, unsigned long val);
+int oprofile_write_commit(struct op_entry *entry);
+
 #endif /* OPROFILE_H */

kernel/trace/ring_buffer.c

@@ -168,7 +168,13 @@ rb_event_length(struct ring_buffer_event *event)
  */
 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
 {
-	return rb_event_length(event);
+	unsigned length = rb_event_length(event);
+	if (event->type != RINGBUF_TYPE_DATA)
+		return length;
+	length -= RB_EVNT_HDR_SIZE;
+	if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
+                length -= sizeof(event->array[0]);
+	return length;
 }
 EXPORT_SYMBOL_GPL(ring_buffer_event_length);