Skip to contentSkip to conversation

Oprofile Multiplexing Patch

PATCH In review Robert Richter 2 participants 3 messages Jul 23, 2008 – Jul 29, 2008 x86

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Oprofile Multiplexing Patch

From: Jason Yeh <[email protected]>

This patch introduces multiplexing support for the Oprofile kernel module. It basically adds a new function pointer in oprofile_operator allowing each architecture to supply its callback to switch between different sets of event when the timer expires. Userspace tools can modify the time slice through /dev/oprofile/time_slice.

It also modifies the number of counters exposed to the userspace through /dev/oprofile. For example, the number of counters for AMD CPUs are changed to 32 and multiplexed in the sets of 4.

---
 arch/x86/oprofile/nmi_int.c       |  100 ++++++++++++++++++++++++++++++++++---
 arch/x86/oprofile/op_counter.h    |    3 +-
 arch/x86/oprofile/op_model_amd.c  |   76 +++++++++++++++++-----------
 arch/x86/oprofile/op_model_p4.c   |    4 ++
 arch/x86/oprofile/op_model_ppro.c |    2 +
 arch/x86/oprofile/op_x86_model.h  |    3 +
 drivers/oprofile/oprof.c          |   58 ++++++++++++++++++++-
 drivers/oprofile/oprof.h          |    4 +-
 drivers/oprofile/oprofile_files.c |   39 ++++++++++++++-
 include/linux/oprofile.h          |    3 +
 10 files changed, 247 insertions(+), 45 deletions(-)
arch/x86/oprofile/nmi_int.c +92 -8
index f46d8fc..186ab90 100644
@@ -23,12 +23,18 @@
 #include "op_counter.h"
 #include "op_x86_model.h"
 
+DEFINE_PER_CPU(int, switch_index);
+
 static struct op_x86_model_spec const *model;
 static DEFINE_PER_CPU(struct op_msrs, cpu_msrs);
 static DEFINE_PER_CPU(unsigned long, saved_lvtpc);
 
 static int nmi_start(void);
 static void nmi_stop(void);
+static void nmi_cpu_save_mpx_registers(struct op_msrs *msrs);
+static void nmi_cpu_restore_mpx_registers(struct op_msrs *msrs);
+static void nmi_cpu_stop(void *dummy);
+static void nmi_cpu_start(void *dummy);
 
 /* 0 == registered but off, 1 == registered and on */
 static int nmi_enabled = 0;
@@ -81,6 +87,47 @@ static void exit_sysfs(void)
 #define exit_sysfs() do { } while (0)
 #endif /* CONFIG_PM */
 
+static void nmi_cpu_switch(void *dummy)
+{
+	int cpu = smp_processor_id();
+	int si = per_cpu(switch_index, cpu);
+	struct op_msrs *msrs = &per_cpu(cpu_msrs, cpu);
+
+	nmi_cpu_stop(NULL);
+	nmi_cpu_save_mpx_registers(msrs);
+
+	/* move to next set */
+	si += model->num_hardware_counters;
+	if ((si > model->num_counters) || (counter_config[si].count == 0))
+		per_cpu(switch_index, smp_processor_id()) = 0;
+	else
+		per_cpu(switch_index, smp_processor_id()) = si;
+
+	nmi_cpu_restore_mpx_registers(msrs);
+	model->setup_ctrs(msrs);
+	nmi_cpu_start(NULL);
+}
+
+/*
+ * Quick check to see if multiplexing is necessary.
+ * The check should be sufficient since counters are used
+ * in ordre.
+ */
+static int nmi_multiplex_on(void)
+{
+	return counter_config[model->num_hardware_counters].count ? 0 : -EINVAL;
+}
+
+static int nmi_switch_event(void)
+{
+	if (nmi_multiplex_on() < 0)
+		return -EINVAL;
+
+	on_each_cpu(nmi_cpu_switch, NULL, 0, 1);
+
+	return 0;
+}
+
 static int profile_exceptions_notify(struct notifier_block *self,
 				     unsigned long val, void *data)
 {
@@ -144,11 +191,10 @@ static void free_msrs(void)
 
 static int allocate_msrs(void)
 {
-	int success = 1;
+	int i, success = 1;
 	size_t controls_size = sizeof(struct op_msr) * model->num_controls;
 	size_t counters_size = sizeof(struct op_msr) * model->num_counters;
 
-	int i;
 	for_each_possible_cpu(i) {
 		per_cpu(cpu_msrs, i).counters = kmalloc(counters_size,
 								GFP_KERNEL);
@@ -156,8 +202,8 @@ static int allocate_msrs(void)
 			success = 0;
 			break;
 		}
-		per_cpu(cpu_msrs, i).controls = kmalloc(controls_size,
-								GFP_KERNEL);
+		per_cpu(cpu_msrs, i).controls =
+				kmalloc(controls_size, GFP_KERNEL);
 		if (!per_cpu(cpu_msrs, i).controls) {
 			success = 0;
 			break;
@@ -201,7 +247,8 @@ static int nmi_setup(void)
 		return err;
 	}
 
-	/* We need to serialize save and setup for HT because the subset
+	/*
+	 * We need to serialize save and setup for HT because the subset
 	 * of msrs are distinct for save and setup operations
 	 */
 
@@ -217,7 +264,6 @@ static int nmi_setup(void)
 				per_cpu(cpu_msrs, 0).controls,
 				sizeof(struct op_msr) * model->num_controls);
 		}
-
 	}
 	on_each_cpu(nmi_save_registers, NULL, 1);
 	on_each_cpu(nmi_cpu_setup, NULL, 1);
@@ -225,7 +271,41 @@ static int nmi_setup(void)
 	return 0;
 }
 
-static void nmi_restore_registers(struct op_msrs *msrs)
+static void nmi_cpu_save_mpx_registers(struct op_msrs *msrs)
+{
+	unsigned int si = __get_cpu_var(switch_index);
+	unsigned int const nr_ctrs = model->num_hardware_counters;
+	struct op_msr *counters = &msrs->counters[si];
+	unsigned int i;
+
+	for (i = 0; i < nr_ctrs; ++i) {
+		int offset = i + si;
+		if (counters[offset].addr) {
+			rdmsr(counters[offset].addr,
+				counters[offset].multiplex.low,
+				counters[offset].multiplex.high);
+		}
+	}
+}
+
+static void nmi_cpu_restore_mpx_registers(struct op_msrs *msrs)
+{
+	unsigned int si = __get_cpu_var(switch_index);
+	unsigned int const nr_ctrs = model->num_hardware_counters;
+	struct op_msr *counters = &msrs->counters[si];
+	unsigned int i;
+
+	for (i = 0; i < nr_ctrs; ++i) {
+		int offset = i + si;
+		if (counters[offset].addr) {
+			wrmsr(counters[offset].addr,
+				counters[offset].multiplex.low,
+				counters[offset].multiplex.high);
+		}
+	}
+}
+
+static void nmi_cpu_restore_registers(struct op_msrs *msrs)
 {
 	unsigned int const nr_ctrs = model->num_counters;
 	unsigned int const nr_ctrls = model->num_controls;
@@ -265,7 +345,8 @@ static void nmi_cpu_shutdown(void *dummy)
 	apic_write(APIC_LVTERR, v | APIC_LVT_MASKED);
 	apic_write(APIC_LVTPC, per_cpu(saved_lvtpc, cpu));
 	apic_write(APIC_LVTERR, v);
-	nmi_restore_registers(msrs);
+	nmi_cpu_restore_registers(msrs);
+	__get_cpu_var(switch_index) = 0;
 }
 
 static void nmi_shutdown(void)
@@ -328,6 +409,7 @@ static int nmi_create_files(struct super_block *sb, struct dentry *root)
 		oprofilefs_create_ulong(sb, dir, "unit_mask", &counter_config[i].unit_mask);
 		oprofilefs_create_ulong(sb, dir, "kernel", &counter_config[i].kernel);
 		oprofilefs_create_ulong(sb, dir, "user", &counter_config[i].user);
+		counter_config[i].save_count_low = 0;
 	}
 
 	return 0;
@@ -455,12 +537,14 @@ int __init op_nmi_init(struct oprofile_operations *ops)
 	}
 
 	/* default values, can be overwritten by model */
+	__get_cpu_var(switch_index) = 0;
 	ops->create_files = nmi_create_files;
 	ops->setup = nmi_setup;
 	ops->shutdown = nmi_shutdown;
 	ops->start = nmi_start;
 	ops->stop = nmi_stop;
 	ops->cpu_type = cpu_type;
+	ops->switch_events = nmi_switch_event;
 
 	if (model->init)
 		ret = model->init(ops);
arch/x86/oprofile/op_counter.h +2 -1
index 2880b15..786d6e0 100644
@@ -10,13 +10,14 @@
 #ifndef OP_COUNTER_H
 #define OP_COUNTER_H
  
-#define OP_MAX_COUNTER 8
+#define OP_MAX_COUNTER 32
  
 /* Per-perfctr configuration as set via
  * oprofilefs.
  */
 struct op_counter_config {
         unsigned long count;
+		unsigned long save_count_low;
         unsigned long enabled;
         unsigned long event;
         unsigned long kernel;
arch/x86/oprofile/op_model_amd.c +46 -30

169 lines not on this page · raw message

arch/x86/oprofile/op_model_p4.c +4 -0
index 56b4757..e641545 100644
@@ -701,6 +701,8 @@ static void p4_shutdown(struct op_msrs const * const msrs)
 struct op_x86_model_spec const op_p4_ht2_spec = {
 	.num_counters = NUM_COUNTERS_HT2,
 	.num_controls = NUM_CONTROLS_HT2,
+	.num_hardware_counters = NUM_COUNTERS_HT2,
+	.num_hardware_controls = NUM_CONTROLS_HT2,
 	.fill_in_addresses = &p4_fill_in_addresses,
 	.setup_ctrs = &p4_setup_ctrs,
 	.check_ctrs = &p4_check_ctrs,
@@ -713,6 +715,8 @@ struct op_x86_model_spec const op_p4_ht2_spec = {
 struct op_x86_model_spec const op_p4_spec = {
 	.num_counters = NUM_COUNTERS_NON_HT,
 	.num_controls = NUM_CONTROLS_NON_HT,
+	.num_hardware_counters = NUM_COUNTERS_NON_HT,
+	.num_hardware_controls = NUM_CONTROLS_NON_HT,
 	.fill_in_addresses = &p4_fill_in_addresses,
 	.setup_ctrs = &p4_setup_ctrs,
 	.check_ctrs = &p4_check_ctrs,
arch/x86/oprofile/op_model_ppro.c +2 -0
index eff431f..e5811aa 100644
@@ -183,6 +183,8 @@ static void ppro_shutdown(struct op_msrs const * const msrs)
 struct op_x86_model_spec const op_ppro_spec = {
 	.num_counters = NUM_COUNTERS,
 	.num_controls = NUM_CONTROLS,
+	.num_hardware_counters = NUM_COUNTERS,
+	.num_hardware_controls = NUM_CONTROLS,
 	.fill_in_addresses = &ppro_fill_in_addresses,
 	.setup_ctrs = &ppro_setup_ctrs,
 	.check_ctrs = &ppro_check_ctrs,
arch/x86/oprofile/op_x86_model.h +3 -0
index 05a0261..e07ba10 100644
@@ -19,6 +19,7 @@ struct op_saved_msr {
 struct op_msr {
 	unsigned long addr;
 	struct op_saved_msr saved;
+	struct op_saved_msr multiplex;
 };
 
 struct op_msrs {
@@ -34,6 +35,8 @@ struct pt_regs;
 struct op_x86_model_spec {
 	int (*init)(struct oprofile_operations *ops);
 	void (*exit)(void);
+	unsigned int const num_hardware_counters;
+	unsigned int const num_hardware_controls;
 	unsigned int const num_counters;
 	unsigned int const num_controls;
 	void (*fill_in_addresses)(struct op_msrs * const msrs);
drivers/oprofile/oprof.c +55 -3

115 lines not on this page · raw message

drivers/oprofile/oprof.h +3 -1
index 1832365..c4406a7 100644
@@ -27,7 +27,8 @@ extern unsigned long fs_buffer_watershed;
 extern struct oprofile_operations oprofile_ops;
 extern unsigned long oprofile_started;
 extern unsigned long backtrace_depth;
- 
+extern unsigned long timeout_jiffies;
+
 struct super_block;
 struct dentry;
 
@@ -35,5 +36,6 @@ void oprofile_create_files(struct super_block * sb, struct dentry * root);
 void oprofile_timer_init(struct oprofile_operations * ops);
 
 int oprofile_set_backtrace(unsigned long depth);
+int oprofile_set_timeout(unsigned long time);
  
 #endif /* OPROF_H */
drivers/oprofile/oprofile_files.c +37 -2
index ef953ba..cc4f5a1 100644
@@ -9,6 +9,7 @@
 
 #include <linux/fs.h>
 #include <linux/oprofile.h>
+#include <linux/jiffies.h>
 
 #include "event_buffer.h"
 #include "oprofile_stats.h"
@@ -18,6 +19,40 @@ unsigned long fs_buffer_size = 131072;
 unsigned long fs_cpu_buffer_size = 8192;
 unsigned long fs_buffer_watershed = 32768; /* FIXME: tune */
 
+static ssize_t timeout_read(struct file *file, char __user *buf,
+		size_t count, loff_t *offset)
+{
+	return oprofilefs_ulong_to_user(jiffies_to_msecs(timeout_jiffies),
+				buf, count, offset);
+}
+
+
+static ssize_t timeout_write(struct file *file, char const __user *buf,
+		size_t count, loff_t *offset)
+{
+	unsigned long val;
+	int retval;
+
+	if (*offset)
+		return -EINVAL;
+
+	retval = oprofilefs_ulong_from_user(&val, buf, count);
+	if (retval)
+		return retval;
+
+	retval = oprofile_set_timeout(val);
+
+	if (retval)
+		return retval;
+	return count;
+}
+
+static const struct file_operations timeout_fops = {
+	.read		= timeout_read,
+	.write		= timeout_write,
+};
+
+
 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);
@@ -85,11 +120,10 @@ static ssize_t enable_write(struct file * file, char const __user * buf, size_t
 
 	if (*offset)
 		return -EINVAL;
-
 	retval = oprofilefs_ulong_from_user(&val, buf, count);
 	if (retval)
 		return retval;
- 
+
 	if (val)
 		retval = oprofile_start();
 	else
@@ -129,6 +163,7 @@ void oprofile_create_files(struct super_block * sb, struct dentry * root)
 	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);
+	oprofilefs_create_file(sb, root, "timeout_ms", &timeout_fops);
 	oprofile_create_stats_files(sb, root);
 	if (oprofile_ops.create_files)
 		oprofile_ops.create_files(sb, root);
include/linux/oprofile.h +3 -0
index bcb8f72..687f2f4 100644
@@ -67,6 +67,9 @@ struct oprofile_operations {
 
 	/* Initiate a stack backtrace. Optional. */
 	void (*backtrace)(struct pt_regs * const regs, unsigned int depth);
+
+	/* Multiplex between different events. Optional. */
+	int (*switch_events)(void);
 	/* CPU identification string. */
 	char * cpu_type;
 };
Signature · 1 line
1.5.5.4
#1 lore raw

arch/x86/oprofile/op_model_p4.c static void p4_shutdown(struct op_msrs const * const msrs) · line 705

Quoted 13 lines

From: Jason Yeh <[email protected]>

arch/x86/oprofile/op_model_p4.c +2 -0
index 56b4757..e641545 100644
@@ -701,6 +701,8 @@ static void p4_shutdown(struct op_msrs const * const msrs)
 struct op_x86_model_spec const op_p4_ht2_spec = {
 	.num_counters = NUM_COUNTERS_HT2,
 	.num_controls = NUM_CONTROLS_HT2,
+	.num_hardware_counters = NUM_COUNTERS_HT2,
+	.num_hardware_controls = NUM_CONTROLS_HT2,

is there any further change needed for Intel CPUs to benefit from multiplexing too, up to the max of 32 controls? (AFAICS it should work as-is, correct?)

Ingo

#2 lore raw
Quoted 17 lines · Ingo Molnar

arch/x86/oprofile/op_model_p4.c static void p4_shutdown(struct op_msrs const * const msrs) · line 705

Quoted 13 lines

From: Jason Yeh <[email protected]>

arch/x86/oprofile/op_model_p4.c +2 -0
index 56b4757..e641545 100644
@@ -701,6 +701,8 @@ static void p4_shutdown(struct op_msrs const * const msrs)
 struct op_x86_model_spec const op_p4_ht2_spec = {
 	.num_counters = NUM_COUNTERS_HT2,
 	.num_controls = NUM_CONTROLS_HT2,
+	.num_hardware_counters = NUM_COUNTERS_HT2,
+	.num_hardware_controls = NUM_CONTROLS_HT2,

is there any further change needed for Intel CPUs to benefit from multiplexing too, up to the max of 32 controls? (AFAICS it should work as-is, correct?)

I would like to wait with this change until the code base is stable. There could be some follow-up changes that would require more rework then.

-Robert

Quoted 8 lines

Ingo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Signature · 3 lines
Advanced Micro Devices, Inc.
Operating System Research Center
email: [email protected]