[PATCH 0/3] even more perf counter patches

[PATCH 0/3] even more perf counter patches

[Peter Zijlstra]

-- 

[PATCH 1/3] perf_counter: x86: expose INV and EDGE bits

[Peter Zijlstra]

[-- Attachment #1: perf_counter-x86-masks.patch --]
[-- Type: text/plain, Size: 1493 bytes --]

Expose the INV and EDGE bits of the PMU to raw configs.

LKML-Reference: <new-submission>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 arch/x86/kernel/cpu/perf_counter.c |    8 ++++++++
 1 file changed, 8 insertions(+)

Index: linux-2.6/arch/x86/kernel/cpu/perf_counter.c
===================================================================
--- linux-2.6.orig/arch/x86/kernel/cpu/perf_counter.c
+++ linux-2.6/arch/x86/kernel/cpu/perf_counter.c
@@ -87,11 +87,15 @@ static u64 intel_pmu_raw_event(u64 event
 {
 #define CORE_EVNTSEL_EVENT_MASK		0x000000FFULL
 #define CORE_EVNTSEL_UNIT_MASK		0x0000FF00ULL
+#define CORE_EVNTSEL_EDGE_MASK		0x00040000ULL
+#define CORE_EVNTSEL_INV_MASK		0x00800000ULL
 #define CORE_EVNTSEL_COUNTER_MASK	0xFF000000ULL
 
 #define CORE_EVNTSEL_MASK 		\
 	(CORE_EVNTSEL_EVENT_MASK |	\
 	 CORE_EVNTSEL_UNIT_MASK  |	\
+	 CORE_EVNTSEL_EDGE_MASK  |	\
+	 CORE_EVNTSEL_INV_MASK  |	\
 	 CORE_EVNTSEL_COUNTER_MASK)
 
 	return event & CORE_EVNTSEL_MASK;
@@ -119,11 +123,15 @@ static u64 amd_pmu_raw_event(u64 event)
 {
 #define K7_EVNTSEL_EVENT_MASK	0x7000000FFULL
 #define K7_EVNTSEL_UNIT_MASK	0x00000FF00ULL
+#define K7_EVNTSEL_EDGE_MASK	0x000040000ULL
+#define K7_EVNTSEL_INV_MASK	0x000800000ULL
 #define K7_EVNTSEL_COUNTER_MASK	0x0FF000000ULL
 
 #define K7_EVNTSEL_MASK			\
 	(K7_EVNTSEL_EVENT_MASK |	\
 	 K7_EVNTSEL_UNIT_MASK  |	\
+	 K7_EVNTSEL_EDGE_MASK  |	\
+	 K7_EVNTSEL_INV_MASK   |	\
 	 K7_EVNTSEL_COUNTER_MASK)
 
 	return event & K7_EVNTSEL_MASK;

-- 

[tip:perfcounters/core] perf_counter: x86: Expose INV and EDGE bits

[tip-bot for Peter Zijlstra]

Commit-ID:  ff99be573e02e9f7edc23b472c7f9a5ddba12795
Gitweb:     http://git.kernel.org/tip/ff99be573e02e9f7edc23b472c7f9a5ddba12795
Author:     Peter Zijlstra <a.p.zijlstra@chello.nl>
AuthorDate: Mon, 25 May 2009 17:39:03 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Mon, 25 May 2009 21:41:11 +0200

perf_counter: x86: Expose INV and EDGE bits

Expose the INV and EDGE bits of the PMU to raw configs.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525153931.494709027@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 arch/x86/kernel/cpu/perf_counter.c |    8 ++++++++
 1 files changed, 8 insertions(+), 0 deletions(-)

diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c
index 6cc1660..c14437f 100644
--- a/arch/x86/kernel/cpu/perf_counter.c
+++ b/arch/x86/kernel/cpu/perf_counter.c
@@ -87,11 +87,15 @@ static u64 intel_pmu_raw_event(u64 event)
 {
 #define CORE_EVNTSEL_EVENT_MASK		0x000000FFULL
 #define CORE_EVNTSEL_UNIT_MASK		0x0000FF00ULL
+#define CORE_EVNTSEL_EDGE_MASK		0x00040000ULL
+#define CORE_EVNTSEL_INV_MASK		0x00800000ULL
 #define CORE_EVNTSEL_COUNTER_MASK	0xFF000000ULL
 
 #define CORE_EVNTSEL_MASK 		\
 	(CORE_EVNTSEL_EVENT_MASK |	\
 	 CORE_EVNTSEL_UNIT_MASK  |	\
+	 CORE_EVNTSEL_EDGE_MASK  |	\
+	 CORE_EVNTSEL_INV_MASK  |	\
 	 CORE_EVNTSEL_COUNTER_MASK)
 
 	return event & CORE_EVNTSEL_MASK;
@@ -119,11 +123,15 @@ static u64 amd_pmu_raw_event(u64 event)
 {
 #define K7_EVNTSEL_EVENT_MASK	0x7000000FFULL
 #define K7_EVNTSEL_UNIT_MASK	0x00000FF00ULL
+#define K7_EVNTSEL_EDGE_MASK	0x000040000ULL
+#define K7_EVNTSEL_INV_MASK	0x000800000ULL
 #define K7_EVNTSEL_COUNTER_MASK	0x0FF000000ULL
 
 #define K7_EVNTSEL_MASK			\
 	(K7_EVNTSEL_EVENT_MASK |	\
 	 K7_EVNTSEL_UNIT_MASK  |	\
+	 K7_EVNTSEL_EDGE_MASK  |	\
+	 K7_EVNTSEL_INV_MASK   |	\
 	 K7_EVNTSEL_COUNTER_MASK)
 
 	return event & K7_EVNTSEL_MASK;

[PATCH 2/3] perf_counter: x86: remove interrupt throttle

[Peter Zijlstra]

[-- Attachment #1: perf_counter-x86-remove-throttle.patch --]
[-- Type: text/plain, Size: 4127 bytes --]

remove the x86 specific interrupt throttle

LKML-Reference: <new-submission>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 arch/x86/kernel/apic/apic.c        |    2 -
 arch/x86/kernel/cpu/perf_counter.c |   47 +++----------------------------------
 include/linux/perf_counter.h       |    2 -
 3 files changed, 5 insertions(+), 46 deletions(-)

Index: linux-2.6/arch/x86/kernel/apic/apic.c
===================================================================
--- linux-2.6.orig/arch/x86/kernel/apic/apic.c
+++ linux-2.6/arch/x86/kernel/apic/apic.c
@@ -817,8 +817,6 @@ static void local_apic_timer_interrupt(v
 	inc_irq_stat(apic_timer_irqs);
 
 	evt->event_handler(evt);
-
-	perf_counter_unthrottle();
 }
 
 /*
Index: linux-2.6/arch/x86/kernel/cpu/perf_counter.c
===================================================================
--- linux-2.6.orig/arch/x86/kernel/cpu/perf_counter.c
+++ linux-2.6/arch/x86/kernel/cpu/perf_counter.c
@@ -719,11 +719,6 @@ static void intel_pmu_save_and_restart(s
 }
 
 /*
- * Maximum interrupt frequency of 100KHz per CPU
- */
-#define PERFMON_MAX_INTERRUPTS (100000/HZ)
-
-/*
  * This handler is triggered by the local APIC, so the APIC IRQ handling
  * rules apply:
  */
@@ -775,15 +770,14 @@ again:
 	if (status)
 		goto again;
 
-	if (++cpuc->interrupts != PERFMON_MAX_INTERRUPTS)
-		perf_enable();
+	perf_enable();
 
 	return 1;
 }
 
 static int amd_pmu_handle_irq(struct pt_regs *regs, int nmi)
 {
-	int cpu, idx, throttle = 0, handled = 0;
+	int cpu, idx, handled = 0;
 	struct cpu_hw_counters *cpuc;
 	struct perf_counter *counter;
 	struct hw_perf_counter *hwc;
@@ -792,16 +786,7 @@ static int amd_pmu_handle_irq(struct pt_
 	cpu = smp_processor_id();
 	cpuc = &per_cpu(cpu_hw_counters, cpu);
 
-	if (++cpuc->interrupts == PERFMON_MAX_INTERRUPTS) {
-		throttle = 1;
-		__perf_disable();
-		cpuc->enabled = 0;
-		barrier();
-	}
-
 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
-		int disable = 0;
-
 		if (!test_bit(idx, cpuc->active_mask))
 			continue;
 
@@ -809,45 +794,23 @@ static int amd_pmu_handle_irq(struct pt_
 		hwc = &counter->hw;
 
 		if (counter->hw_event.nmi != nmi)
-			goto next;
+			continue;
 
 		val = x86_perf_counter_update(counter, hwc, idx);
 		if (val & (1ULL << (x86_pmu.counter_bits - 1)))
-			goto next;
+			continue;
 
 		/* counter overflow */
 		x86_perf_counter_set_period(counter, hwc, idx);
 		handled = 1;
 		inc_irq_stat(apic_perf_irqs);
-		disable = perf_counter_overflow(counter, nmi, regs, 0);
-
-next:
-		if (disable || throttle)
+		if (perf_counter_overflow(counter, nmi, regs, 0))
 			amd_pmu_disable_counter(hwc, idx);
 	}
 
 	return handled;
 }
 
-void perf_counter_unthrottle(void)
-{
-	struct cpu_hw_counters *cpuc;
-
-	if (!x86_pmu_initialized())
-		return;
-
-	cpuc = &__get_cpu_var(cpu_hw_counters);
-	if (cpuc->interrupts >= PERFMON_MAX_INTERRUPTS) {
-		/*
-		 * Clear them before re-enabling irqs/NMIs again:
-		 */
-		cpuc->interrupts = 0;
-		perf_enable();
-	} else {
-		cpuc->interrupts = 0;
-	}
-}
-
 void smp_perf_counter_interrupt(struct pt_regs *regs)
 {
 	irq_enter();
Index: linux-2.6/include/linux/perf_counter.h
===================================================================
--- linux-2.6.orig/include/linux/perf_counter.h
+++ linux-2.6/include/linux/perf_counter.h
@@ -570,7 +570,6 @@ extern void perf_counter_init_task(struc
 extern void perf_counter_exit_task(struct task_struct *child);
 extern void perf_counter_do_pending(void);
 extern void perf_counter_print_debug(void);
-extern void perf_counter_unthrottle(void);
 extern void __perf_disable(void);
 extern bool __perf_enable(void);
 extern void perf_disable(void);
@@ -635,7 +634,6 @@ static inline void perf_counter_init_tas
 static inline void perf_counter_exit_task(struct task_struct *child)	{ }
 static inline void perf_counter_do_pending(void)			{ }
 static inline void perf_counter_print_debug(void)			{ }
-static inline void perf_counter_unthrottle(void)			{ }
 static inline void perf_disable(void)					{ }
 static inline void perf_enable(void)					{ }
 static inline int perf_counter_task_disable(void)	{ return -EINVAL; }

-- 

[tip:perfcounters/core] perf_counter: x86: Remove interrupt throttle

[tip-bot for Peter Zijlstra]

Commit-ID:  48e22d56ecdeddd1ffb42a02fccba5c6ef42b133
Gitweb:     http://git.kernel.org/tip/48e22d56ecdeddd1ffb42a02fccba5c6ef42b133
Author:     Peter Zijlstra <a.p.zijlstra@chello.nl>
AuthorDate: Mon, 25 May 2009 17:39:04 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Mon, 25 May 2009 21:41:12 +0200

perf_counter: x86: Remove interrupt throttle

remove the x86 specific interrupt throttle

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525153931.616671838@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 arch/x86/kernel/apic/apic.c        |    2 -
 arch/x86/kernel/cpu/perf_counter.c |   47 ++++--------------------------------
 include/linux/perf_counter.h       |    2 -
 3 files changed, 5 insertions(+), 46 deletions(-)

diff --git a/arch/x86/kernel/apic/apic.c b/arch/x86/kernel/apic/apic.c
index b4f6440..89b63b5 100644
--- a/arch/x86/kernel/apic/apic.c
+++ b/arch/x86/kernel/apic/apic.c
@@ -763,8 +763,6 @@ static void local_apic_timer_interrupt(void)
 	inc_irq_stat(apic_timer_irqs);
 
 	evt->event_handler(evt);
-
-	perf_counter_unthrottle();
 }
 
 /*
diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c
index c14437f..8c8177f 100644
--- a/arch/x86/kernel/cpu/perf_counter.c
+++ b/arch/x86/kernel/cpu/perf_counter.c
@@ -719,11 +719,6 @@ static void intel_pmu_save_and_restart(struct perf_counter *counter)
 }
 
 /*
- * Maximum interrupt frequency of 100KHz per CPU
- */
-#define PERFMON_MAX_INTERRUPTS (100000/HZ)
-
-/*
  * This handler is triggered by the local APIC, so the APIC IRQ handling
  * rules apply:
  */
@@ -775,15 +770,14 @@ again:
 	if (status)
 		goto again;
 
-	if (++cpuc->interrupts != PERFMON_MAX_INTERRUPTS)
-		perf_enable();
+	perf_enable();
 
 	return 1;
 }
 
 static int amd_pmu_handle_irq(struct pt_regs *regs, int nmi)
 {
-	int cpu, idx, throttle = 0, handled = 0;
+	int cpu, idx, handled = 0;
 	struct cpu_hw_counters *cpuc;
 	struct perf_counter *counter;
 	struct hw_perf_counter *hwc;
@@ -792,16 +786,7 @@ static int amd_pmu_handle_irq(struct pt_regs *regs, int nmi)
 	cpu = smp_processor_id();
 	cpuc = &per_cpu(cpu_hw_counters, cpu);
 
-	if (++cpuc->interrupts == PERFMON_MAX_INTERRUPTS) {
-		throttle = 1;
-		__perf_disable();
-		cpuc->enabled = 0;
-		barrier();
-	}
-
 	for (idx = 0; idx < x86_pmu.num_counters; idx++) {
-		int disable = 0;
-
 		if (!test_bit(idx, cpuc->active_mask))
 			continue;
 
@@ -809,45 +794,23 @@ static int amd_pmu_handle_irq(struct pt_regs *regs, int nmi)
 		hwc = &counter->hw;
 
 		if (counter->hw_event.nmi != nmi)
-			goto next;
+			continue;
 
 		val = x86_perf_counter_update(counter, hwc, idx);
 		if (val & (1ULL << (x86_pmu.counter_bits - 1)))
-			goto next;
+			continue;
 
 		/* counter overflow */
 		x86_perf_counter_set_period(counter, hwc, idx);
 		handled = 1;
 		inc_irq_stat(apic_perf_irqs);
-		disable = perf_counter_overflow(counter, nmi, regs, 0);
-
-next:
-		if (disable || throttle)
+		if (perf_counter_overflow(counter, nmi, regs, 0))
 			amd_pmu_disable_counter(hwc, idx);
 	}
 
 	return handled;
 }
 
-void perf_counter_unthrottle(void)
-{
-	struct cpu_hw_counters *cpuc;
-
-	if (!x86_pmu_initialized())
-		return;
-
-	cpuc = &__get_cpu_var(cpu_hw_counters);
-	if (cpuc->interrupts >= PERFMON_MAX_INTERRUPTS) {
-		/*
-		 * Clear them before re-enabling irqs/NMIs again:
-		 */
-		cpuc->interrupts = 0;
-		perf_enable();
-	} else {
-		cpuc->interrupts = 0;
-	}
-}
-
 void smp_perf_counter_interrupt(struct pt_regs *regs)
 {
 	irq_enter();
diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index d3e85de..0c160be 100644
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -570,7 +570,6 @@ extern int perf_counter_init_task(struct task_struct *child);
 extern void perf_counter_exit_task(struct task_struct *child);
 extern void perf_counter_do_pending(void);
 extern void perf_counter_print_debug(void);
-extern void perf_counter_unthrottle(void);
 extern void __perf_disable(void);
 extern bool __perf_enable(void);
 extern void perf_disable(void);
@@ -635,7 +634,6 @@ static inline int perf_counter_init_task(struct task_struct *child)	{ }
 static inline void perf_counter_exit_task(struct task_struct *child)	{ }
 static inline void perf_counter_do_pending(void)			{ }
 static inline void perf_counter_print_debug(void)			{ }
-static inline void perf_counter_unthrottle(void)			{ }
 static inline void perf_disable(void)					{ }
 static inline void perf_enable(void)					{ }
 static inline int perf_counter_task_disable(void)	{ return -EINVAL; }

[PATCH 3/3] perf_counter: generic per counter interrupt throttle

[Peter Zijlstra]

[-- Attachment #1: perf_counter-generic-throttle.patch --]
[-- Type: text/plain, Size: 6429 bytes --]

Introduce a generic per counter interrupt throttle.

This uses the perf_counter_overflow() quick disable to throttle a specific
counter when its going too fast when a pmu->unthrottle() method is provided
which can undo the quick disable.

Power needs to implement both the quick disable and the unthrottle method.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 arch/x86/kernel/cpu/perf_counter.c |   13 ++++++++
 include/linux/perf_counter.h       |   11 ++++++
 kernel/perf_counter.c              |   59 ++++++++++++++++++++++++++++++++++---
 kernel/sysctl.c                    |    8 +++++
 4 files changed, 87 insertions(+), 4 deletions(-)

Index: linux-2.6/include/linux/perf_counter.h
===================================================================
--- linux-2.6.orig/include/linux/perf_counter.h
+++ linux-2.6/include/linux/perf_counter.h
@@ -267,6 +267,15 @@ enum perf_event_type {
 	PERF_EVENT_PERIOD		= 4,
 
 	/*
+	 * struct {
+	 * 	struct perf_event_header	header;
+	 * 	u64				time;
+	 * };
+	 */
+	PERF_EVENT_THROTTLE		= 5,
+	PERF_EVENT_UNTHROTTLE		= 6,
+
+	/*
 	 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
 	 * will be PERF_RECORD_*
 	 *
@@ -367,6 +376,7 @@ struct pmu {
 	int (*enable)			(struct perf_counter *counter);
 	void (*disable)			(struct perf_counter *counter);
 	void (*read)			(struct perf_counter *counter);
+	void (*unthrottle)		(struct perf_counter *counter);
 };
 
 /**
@@ -613,6 +623,7 @@ extern struct perf_callchain_entry *perf
 
 extern int sysctl_perf_counter_priv;
 extern int sysctl_perf_counter_mlock;
+extern int sysctl_perf_counter_limit;
 
 extern void perf_counter_init(void);
 
Index: linux-2.6/kernel/perf_counter.c
===================================================================
--- linux-2.6.orig/kernel/perf_counter.c
+++ linux-2.6/kernel/perf_counter.c
@@ -46,6 +46,7 @@ static atomic_t nr_comm_tracking __read_
 
 int sysctl_perf_counter_priv __read_mostly; /* do we need to be privileged */
 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
+int sysctl_perf_counter_limit __read_mostly = 100000; /* max NMIs per second */
 
 /*
  * Lock for (sysadmin-configurable) counter reservations:
@@ -1091,12 +1092,15 @@ int perf_counter_task_disable(void)
 	return 0;
 }
 
+#define MAX_INTERRUPTS (~0ULL)
+
+static void perf_log_throttle(struct perf_counter *counter, int enable);
 static void perf_log_period(struct perf_counter *counter, u64 period);
 
 static void perf_adjust_freq(struct perf_counter_context *ctx)
 {
 	struct perf_counter *counter;
-	u64 irq_period;
+	u64 interrupts, irq_period;
 	u64 events, period;
 	s64 delta;
 
@@ -1105,10 +1109,19 @@ static void perf_adjust_freq(struct perf
 		if (counter->state != PERF_COUNTER_STATE_ACTIVE)
 			continue;
 
+		interrupts = counter->hw.interrupts;
+		counter->hw.interrupts = 0;
+
+		if (interrupts == MAX_INTERRUPTS) {
+			perf_log_throttle(counter, 1);
+			counter->pmu->unthrottle(counter);
+			interrupts = 2*sysctl_perf_counter_limit/HZ;
+		}
+
 		if (!counter->hw_event.freq || !counter->hw_event.irq_freq)
 			continue;
 
-		events = HZ * counter->hw.interrupts * counter->hw.irq_period;
+		events = HZ * interrupts * counter->hw.irq_period;
 		period = div64_u64(events, counter->hw_event.irq_freq);
 
 		delta = (s64)(1 + period - counter->hw.irq_period);
@@ -1122,7 +1135,6 @@ static void perf_adjust_freq(struct perf
 		perf_log_period(counter, irq_period);
 
 		counter->hw.irq_period = irq_period;
-		counter->hw.interrupts = 0;
 	}
 	spin_unlock(&ctx->lock);
 }
@@ -2545,6 +2557,35 @@ static void perf_log_period(struct perf_
 }
 
 /*
+ * IRQ throttle logging
+ */
+
+static void perf_log_throttle(struct perf_counter *counter, int enable)
+{
+	struct perf_output_handle handle;
+	int ret;
+
+	struct {
+		struct perf_event_header	header;
+		u64				time;
+	} throttle_event = {
+		.header = {
+			.type = PERF_EVENT_THROTTLE + 1,
+			.misc = 0,
+			.size = sizeof(throttle_event),
+		},
+		.time = sched_clock(),
+	};
+
+	ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 0, 0);
+	if (ret)
+		return;
+
+	perf_output_put(&handle, throttle_event);
+	perf_output_end(&handle);
+}
+
+/*
  * Generic counter overflow handling.
  */
 
@@ -2552,9 +2593,19 @@ int perf_counter_overflow(struct perf_co
 			  int nmi, struct pt_regs *regs, u64 addr)
 {
 	int events = atomic_read(&counter->event_limit);
+	int throttle = counter->pmu->unthrottle != NULL;
 	int ret = 0;
 
-	counter->hw.interrupts++;
+	if (!throttle) {
+		counter->hw.interrupts++;
+	} else if (counter->hw.interrupts != MAX_INTERRUPTS) {
+		counter->hw.interrupts++;
+		if (HZ*counter->hw.interrupts > (u64)sysctl_perf_counter_limit) {
+			counter->hw.interrupts = MAX_INTERRUPTS;
+			perf_log_throttle(counter, 0);
+			ret = 1;
+		}
+	}
 
 	/*
 	 * XXX event_limit might not quite work as expected on inherited
Index: linux-2.6/kernel/sysctl.c
===================================================================
--- linux-2.6.orig/kernel/sysctl.c
+++ linux-2.6/kernel/sysctl.c
@@ -939,6 +939,14 @@ static struct ctl_table kern_table[] = {
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec,
 	},
+	{
+		.ctl_name	= CTL_UNNUMBERED,
+		.procname	= "perf_counter_int_limit",
+		.data		= &sysctl_perf_counter_limit,
+		.maxlen		= sizeof(sysctl_perf_counter_limit),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+	},
 #endif
 /*
  * NOTE: do not add new entries to this table unless you have read
Index: linux-2.6/arch/x86/kernel/cpu/perf_counter.c
===================================================================
--- linux-2.6.orig/arch/x86/kernel/cpu/perf_counter.c
+++ linux-2.6/arch/x86/kernel/cpu/perf_counter.c
@@ -623,6 +623,18 @@ try_generic:
 	return 0;
 }
 
+static void x86_pmu_unthrottle(struct perf_counter *counter)
+{
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+	struct hw_perf_counter *hwc = &counter->hw;
+
+	if (WARN_ON_ONCE(hwc->idx >= X86_PMC_IDX_MAX ||
+				cpuc->counters[hwc->idx] != counter))
+		return;
+
+	x86_pmu.enable(hwc, hwc->idx);
+}
+
 void perf_counter_print_debug(void)
 {
 	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
@@ -1038,6 +1050,7 @@ static const struct pmu pmu = {
 	.enable		= x86_pmu_enable,
 	.disable	= x86_pmu_disable,
 	.read		= x86_pmu_read,
+	.unthrottle	= x86_pmu_unthrottle,
 };
 
 const struct pmu *hw_perf_counter_init(struct perf_counter *counter)

-- 

[tip:perfcounters/core] perf_counter: Generic per counter interrupt throttle

[tip-bot for Peter Zijlstra]

Commit-ID:  a78ac3258782f3e64cb40beb5990808e1febcc0c
Gitweb:     http://git.kernel.org/tip/a78ac3258782f3e64cb40beb5990808e1febcc0c
Author:     Peter Zijlstra <a.p.zijlstra@chello.nl>
AuthorDate: Mon, 25 May 2009 17:39:05 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Mon, 25 May 2009 21:41:12 +0200

perf_counter: Generic per counter interrupt throttle

Introduce a generic per counter interrupt throttle.

This uses the perf_counter_overflow() quick disable to throttle a specific
counter when its going too fast when a pmu->unthrottle() method is provided
which can undo the quick disable.

Power needs to implement both the quick disable and the unthrottle method.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525153931.703093461@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 arch/x86/kernel/cpu/perf_counter.c |   13 ++++++++
 include/linux/perf_counter.h       |   11 +++++++
 kernel/perf_counter.c              |   59 +++++++++++++++++++++++++++++++++--
 kernel/sysctl.c                    |    8 +++++
 4 files changed, 87 insertions(+), 4 deletions(-)

diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c
index 8c8177f..c4b543d 100644
--- a/arch/x86/kernel/cpu/perf_counter.c
+++ b/arch/x86/kernel/cpu/perf_counter.c
@@ -623,6 +623,18 @@ try_generic:
 	return 0;
 }
 
+static void x86_pmu_unthrottle(struct perf_counter *counter)
+{
+	struct cpu_hw_counters *cpuc = &__get_cpu_var(cpu_hw_counters);
+	struct hw_perf_counter *hwc = &counter->hw;
+
+	if (WARN_ON_ONCE(hwc->idx >= X86_PMC_IDX_MAX ||
+				cpuc->counters[hwc->idx] != counter))
+		return;
+
+	x86_pmu.enable(hwc, hwc->idx);
+}
+
 void perf_counter_print_debug(void)
 {
 	u64 ctrl, status, overflow, pmc_ctrl, pmc_count, prev_left, fixed;
@@ -1038,6 +1050,7 @@ static const struct pmu pmu = {
 	.enable		= x86_pmu_enable,
 	.disable	= x86_pmu_disable,
 	.read		= x86_pmu_read,
+	.unthrottle	= x86_pmu_unthrottle,
 };
 
 const struct pmu *hw_perf_counter_init(struct perf_counter *counter)
diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index 0c160be..e3a7585 100644
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -267,6 +267,15 @@ enum perf_event_type {
 	PERF_EVENT_PERIOD		= 4,
 
 	/*
+	 * struct {
+	 * 	struct perf_event_header	header;
+	 * 	u64				time;
+	 * };
+	 */
+	PERF_EVENT_THROTTLE		= 5,
+	PERF_EVENT_UNTHROTTLE		= 6,
+
+	/*
 	 * When header.misc & PERF_EVENT_MISC_OVERFLOW the event_type field
 	 * will be PERF_RECORD_*
 	 *
@@ -367,6 +376,7 @@ struct pmu {
 	int (*enable)			(struct perf_counter *counter);
 	void (*disable)			(struct perf_counter *counter);
 	void (*read)			(struct perf_counter *counter);
+	void (*unthrottle)		(struct perf_counter *counter);
 };
 
 /**
@@ -613,6 +623,7 @@ extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 
 extern int sysctl_perf_counter_priv;
 extern int sysctl_perf_counter_mlock;
+extern int sysctl_perf_counter_limit;
 
 extern void perf_counter_init(void);
 
diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 14b1fe9..ec9c400 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -46,6 +46,7 @@ static atomic_t nr_comm_tracking __read_mostly;
 
 int sysctl_perf_counter_priv __read_mostly; /* do we need to be privileged */
 int sysctl_perf_counter_mlock __read_mostly = 512; /* 'free' kb per user */
+int sysctl_perf_counter_limit __read_mostly = 100000; /* max NMIs per second */
 
 /*
  * Lock for (sysadmin-configurable) counter reservations:
@@ -1066,12 +1067,15 @@ static void perf_counter_cpu_sched_in(struct perf_cpu_context *cpuctx, int cpu)
 	__perf_counter_sched_in(ctx, cpuctx, cpu);
 }
 
+#define MAX_INTERRUPTS (~0ULL)
+
+static void perf_log_throttle(struct perf_counter *counter, int enable);
 static void perf_log_period(struct perf_counter *counter, u64 period);
 
 static void perf_adjust_freq(struct perf_counter_context *ctx)
 {
 	struct perf_counter *counter;
-	u64 irq_period;
+	u64 interrupts, irq_period;
 	u64 events, period;
 	s64 delta;
 
@@ -1080,10 +1084,19 @@ static void perf_adjust_freq(struct perf_counter_context *ctx)
 		if (counter->state != PERF_COUNTER_STATE_ACTIVE)
 			continue;
 
+		interrupts = counter->hw.interrupts;
+		counter->hw.interrupts = 0;
+
+		if (interrupts == MAX_INTERRUPTS) {
+			perf_log_throttle(counter, 1);
+			counter->pmu->unthrottle(counter);
+			interrupts = 2*sysctl_perf_counter_limit/HZ;
+		}
+
 		if (!counter->hw_event.freq || !counter->hw_event.irq_freq)
 			continue;
 
-		events = HZ * counter->hw.interrupts * counter->hw.irq_period;
+		events = HZ * interrupts * counter->hw.irq_period;
 		period = div64_u64(events, counter->hw_event.irq_freq);
 
 		delta = (s64)(1 + period - counter->hw.irq_period);
@@ -1097,7 +1110,6 @@ static void perf_adjust_freq(struct perf_counter_context *ctx)
 		perf_log_period(counter, irq_period);
 
 		counter->hw.irq_period = irq_period;
-		counter->hw.interrupts = 0;
 	}
 	spin_unlock(&ctx->lock);
 }
@@ -2544,6 +2556,35 @@ static void perf_log_period(struct perf_counter *counter, u64 period)
 }
 
 /*
+ * IRQ throttle logging
+ */
+
+static void perf_log_throttle(struct perf_counter *counter, int enable)
+{
+	struct perf_output_handle handle;
+	int ret;
+
+	struct {
+		struct perf_event_header	header;
+		u64				time;
+	} throttle_event = {
+		.header = {
+			.type = PERF_EVENT_THROTTLE + 1,
+			.misc = 0,
+			.size = sizeof(throttle_event),
+		},
+		.time = sched_clock(),
+	};
+
+	ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 0, 0);
+	if (ret)
+		return;
+
+	perf_output_put(&handle, throttle_event);
+	perf_output_end(&handle);
+}
+
+/*
  * Generic counter overflow handling.
  */
 
@@ -2551,9 +2592,19 @@ int perf_counter_overflow(struct perf_counter *counter,
 			  int nmi, struct pt_regs *regs, u64 addr)
 {
 	int events = atomic_read(&counter->event_limit);
+	int throttle = counter->pmu->unthrottle != NULL;
 	int ret = 0;
 
-	counter->hw.interrupts++;
+	if (!throttle) {
+		counter->hw.interrupts++;
+	} else if (counter->hw.interrupts != MAX_INTERRUPTS) {
+		counter->hw.interrupts++;
+		if (HZ*counter->hw.interrupts > (u64)sysctl_perf_counter_limit) {
+			counter->hw.interrupts = MAX_INTERRUPTS;
+			perf_log_throttle(counter, 0);
+			ret = 1;
+		}
+	}
 
 	/*
 	 * XXX event_limit might not quite work as expected on inherited
diff --git a/kernel/sysctl.c b/kernel/sysctl.c
index 3cb1849..0c4bf86 100644
--- a/kernel/sysctl.c
+++ b/kernel/sysctl.c
@@ -930,6 +930,14 @@ static struct ctl_table kern_table[] = {
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec,
 	},
+	{
+		.ctl_name	= CTL_UNNUMBERED,
+		.procname	= "perf_counter_int_limit",
+		.data		= &sysctl_perf_counter_limit,
+		.maxlen		= sizeof(sysctl_perf_counter_limit),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+	},
 #endif
 /*
  * NOTE: do not add new entries to this table unless you have read

[tip:perfcounters/core] Revert "perf_counter, x86: speed up the scheduling fast-path"

[tip-bot for Ingo Molnar]

Commit-ID:  53b441a565bf4036ab49c8ea04c5ad06ace7dd6b
Gitweb:     http://git.kernel.org/tip/53b441a565bf4036ab49c8ea04c5ad06ace7dd6b
Author:     Ingo Molnar <mingo@elte.hu>
AuthorDate: Mon, 25 May 2009 21:41:28 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Mon, 25 May 2009 21:41:28 +0200

Revert "perf_counter, x86: speed up the scheduling fast-path"

This reverts commit b68f1d2e7aa21029d73c7d453a8046e95d351740.

It is causing problems (stuck/stuttering profiling) - when mixed
NMI and non-NMI counters are used.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525153931.703093461@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 arch/x86/kernel/cpu/perf_counter.c |    5 +++--
 1 files changed, 3 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kernel/cpu/perf_counter.c b/arch/x86/kernel/cpu/perf_counter.c
index c4b543d..189bf9d 100644
--- a/arch/x86/kernel/cpu/perf_counter.c
+++ b/arch/x86/kernel/cpu/perf_counter.c
@@ -293,7 +293,6 @@ static int __hw_perf_counter_init(struct perf_counter *counter)
 			return -EACCES;
 		hwc->nmi = 1;
 	}
-	perf_counters_lapic_init(hwc->nmi);
 
 	if (!hwc->irq_period)
 		hwc->irq_period = x86_pmu.max_period;
@@ -612,6 +611,8 @@ try_generic:
 		hwc->counter_base = x86_pmu.perfctr;
 	}
 
+	perf_counters_lapic_init(hwc->nmi);
+
 	x86_pmu.disable(hwc, idx);
 
 	cpuc->counters[idx] = counter;
@@ -1037,7 +1038,7 @@ void __init init_hw_perf_counters(void)
 
 	pr_info("... counter mask:    %016Lx\n", perf_counter_mask);
 
-	perf_counters_lapic_init(1);
+	perf_counters_lapic_init(0);
 	register_die_notifier(&perf_counter_nmi_notifier);
 }
 

[tip:perfcounters/core] perf_counter: fix warning & lockup

[tip-bot for Ingo Molnar]

Commit-ID:  0127c3ea082ee9f1034789b978dfc7fd83254617
Gitweb:     http://git.kernel.org/tip/0127c3ea082ee9f1034789b978dfc7fd83254617
Author:     Ingo Molnar <mingo@elte.hu>
AuthorDate: Mon, 25 May 2009 22:03:26 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Mon, 25 May 2009 22:02:23 +0200

perf_counter: fix warning & lockup

 - remove bogus warning
 - fix wakeup from NMI path lockup
 - also fix up whitespace noise in perf_counter.h

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: John Kacur <jkacur@redhat.com>
LKML-Reference: <20090525153931.703093461@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 include/linux/perf_counter.h |   78 +++++++++++++++++++++---------------------
 kernel/perf_counter.c        |    4 +--
 2 files changed, 40 insertions(+), 42 deletions(-)

diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index e3a7585..2b16ed3 100644
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -73,7 +73,7 @@ enum sw_event_ids {
 	PERF_SW_EVENTS_MAX		= 7,
 };
 
-#define __PERF_COUNTER_MASK(name) 			\
+#define __PERF_COUNTER_MASK(name)			\
 	(((1ULL << PERF_COUNTER_##name##_BITS) - 1) <<	\
 	 PERF_COUNTER_##name##_SHIFT)
 
@@ -98,14 +98,14 @@ enum sw_event_ids {
  * in the overflow packets.
  */
 enum perf_counter_record_format {
-	PERF_RECORD_IP		= 1U << 0,
-	PERF_RECORD_TID		= 1U << 1,
-	PERF_RECORD_TIME	= 1U << 2,
-	PERF_RECORD_ADDR	= 1U << 3,
-	PERF_RECORD_GROUP	= 1U << 4,
-	PERF_RECORD_CALLCHAIN	= 1U << 5,
-	PERF_RECORD_CONFIG	= 1U << 6,
-	PERF_RECORD_CPU		= 1U << 7,
+	PERF_RECORD_IP			= 1U << 0,
+	PERF_RECORD_TID			= 1U << 1,
+	PERF_RECORD_TIME		= 1U << 2,
+	PERF_RECORD_ADDR		= 1U << 3,
+	PERF_RECORD_GROUP		= 1U << 4,
+	PERF_RECORD_CALLCHAIN		= 1U << 5,
+	PERF_RECORD_CONFIG		= 1U << 6,
+	PERF_RECORD_CPU			= 1U << 7,
 };
 
 /*
@@ -235,13 +235,13 @@ enum perf_event_type {
 	 * correlate userspace IPs to code. They have the following structure:
 	 *
 	 * struct {
-	 * 	struct perf_event_header	header;
+	 *	struct perf_event_header	header;
 	 *
-	 * 	u32				pid, tid;
-	 * 	u64				addr;
-	 * 	u64				len;
-	 * 	u64				pgoff;
-	 * 	char				filename[];
+	 *	u32				pid, tid;
+	 *	u64				addr;
+	 *	u64				len;
+	 *	u64				pgoff;
+	 *	char				filename[];
 	 * };
 	 */
 	PERF_EVENT_MMAP			= 1,
@@ -249,27 +249,27 @@ enum perf_event_type {
 
 	/*
 	 * struct {
-	 * 	struct perf_event_header	header;
+	 *	struct perf_event_header	header;
 	 *
-	 * 	u32				pid, tid;
-	 * 	char				comm[];
+	 *	u32				pid, tid;
+	 *	char				comm[];
 	 * };
 	 */
 	PERF_EVENT_COMM			= 3,
 
 	/*
 	 * struct {
-	 * 	struct perf_event_header	header;
-	 * 	u64				time;
-	 * 	u64				irq_period;
+	 *	struct perf_event_header	header;
+	 *	u64				time;
+	 *	u64				irq_period;
 	 * };
 	 */
 	PERF_EVENT_PERIOD		= 4,
 
 	/*
 	 * struct {
-	 * 	struct perf_event_header	header;
-	 * 	u64				time;
+	 *	struct perf_event_header	header;
+	 *	u64				time;
 	 * };
 	 */
 	PERF_EVENT_THROTTLE		= 5,
@@ -280,23 +280,23 @@ enum perf_event_type {
 	 * will be PERF_RECORD_*
 	 *
 	 * struct {
-	 * 	struct perf_event_header	header;
+	 *	struct perf_event_header	header;
 	 *
-	 * 	{ u64			ip;	  } && PERF_RECORD_IP
-	 * 	{ u32			pid, tid; } && PERF_RECORD_TID
-	 * 	{ u64			time;     } && PERF_RECORD_TIME
-	 * 	{ u64			addr;     } && PERF_RECORD_ADDR
-	 * 	{ u64			config;   } && PERF_RECORD_CONFIG
-	 * 	{ u32			cpu, res; } && PERF_RECORD_CPU
+	 *	{ u64			ip;	  } && PERF_RECORD_IP
+	 *	{ u32			pid, tid; } && PERF_RECORD_TID
+	 *	{ u64			time;     } && PERF_RECORD_TIME
+	 *	{ u64			addr;     } && PERF_RECORD_ADDR
+	 *	{ u64			config;   } && PERF_RECORD_CONFIG
+	 *	{ u32			cpu, res; } && PERF_RECORD_CPU
 	 *
-	 * 	{ u64			nr;
-	 * 	  { u64 event, val; } 	cnt[nr];  } && PERF_RECORD_GROUP
+	 *	{ u64			nr;
+	 *	  { u64 event, val; }	cnt[nr];  } && PERF_RECORD_GROUP
 	 *
-	 * 	{ u16			nr,
-	 * 				hv,
-	 * 				kernel,
-	 * 				user;
-	 * 	  u64			ips[nr];  } && PERF_RECORD_CALLCHAIN
+	 *	{ u16			nr,
+	 *				hv,
+	 *				kernel,
+	 *				user;
+	 *	  u64			ips[nr];  } && PERF_RECORD_CALLCHAIN
 	 * };
 	 */
 };
@@ -406,7 +406,7 @@ struct perf_mmap_data {
 	atomic_t			wakeup;		/* needs a wakeup    */
 
 	struct perf_counter_mmap_page   *user_page;
-	void 				*data_pages[0];
+	void				*data_pages[0];
 };
 
 struct perf_pending_entry {
@@ -422,7 +422,7 @@ struct perf_counter {
 	struct list_head		list_entry;
 	struct list_head		event_entry;
 	struct list_head		sibling_list;
-	int 				nr_siblings;
+	int				nr_siblings;
 	struct perf_counter		*group_leader;
 	const struct pmu		*pmu;
 
diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index ec9c400..070f92d 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -2576,7 +2576,7 @@ static void perf_log_throttle(struct perf_counter *counter, int enable)
 		.time = sched_clock(),
 	};
 
-	ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 0, 0);
+	ret = perf_output_begin(&handle, counter, sizeof(throttle_event), 1, 0);
 	if (ret)
 		return;
 
@@ -3449,8 +3449,6 @@ void perf_counter_exit_task(struct task_struct *child)
 	struct perf_counter_context *child_ctx;
 	unsigned long flags;
 
-	WARN_ON_ONCE(child != current);
-
 	child_ctx = child->perf_counter_ctxp;
 
 	if (likely(!child_ctx))