Skip to contentSkip to conversation

perf counter patches

PATCH In review Peter Zijlstra 2 participants 18 messages May 23, 2009 – May 25, 2009

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

cover perf counter patches

Not quite tested, since my upgrade to F11 the userspace bits seem to segfault within the dynamic linker :/

#1 lore raw

PATCH 1/7 perf_counter: fix dynamic irq_period logging

We call perf_adjust_freq() from perf_counter_task_tick() which is is called under the rq->lock causing lock recursion. However, it's no longer required to be called under the rq->lock, so remove it from under it.

Also, fix up some related comments.

LKML-Reference: <new-submission>

---
 include/linux/perf_counter.h |    1 +
 kernel/perf_counter.c        |    3 ++-
 kernel/sched.c               |    3 ++-
 3 files changed, 5 insertions(+), 2 deletions(-)
include/linux/perf_counter.h +1 -0
Index: linux-2.6/include/linux/perf_counter.h
===================================================================
@@ -260,6 +260,7 @@ enum perf_event_type {
 	/*
 	 * struct {
 	 * 	struct perf_event_header	header;
+	 * 	u64				time;
 	 * 	u64				irq_period;
 	 * };
 	 */
kernel/perf_counter.c +2 -1
Index: linux-2.6/kernel/perf_counter.c
===================================================================
@@ -2559,7 +2559,8 @@ void perf_counter_munmap(unsigned long a
 }
 
 /*
- *
+ * Log irq_period changes so that analyzing tools can re-normalize the
+ * event flow.
  */
 
 static void perf_log_period(struct perf_counter *counter, u64 period)
kernel/sched.c +2 -1
Index: linux-2.6/kernel/sched.c
===================================================================
@@ -5113,9 +5113,10 @@ void scheduler_tick(void)
 	update_rq_clock(rq);
 	update_cpu_load(rq);
 	curr->sched_class->task_tick(rq, curr, 0);
-	perf_counter_task_tick(curr, cpu);
 	spin_unlock(&rq->lock);
 
+	perf_counter_task_tick(curr, cpu);
+
 #ifdef CONFIG_SMP
 	rq->idle_at_tick = idle_cpu(cpu);
 	trigger_load_balance(rq, cpu);
#2 lore raw

[tip:perfcounters/core] perf_counter: Fix dynamic irq_period logging

Commit-ID:  e220d2dcb944c5c488b6855d15ec66d76900514f
Gitweb:     http://git.kernel.org/tip/e220d2dcb944c5c488b6855d15ec66d76900514f
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Sat, 23 May 2009 18:28:55 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Sat, 23 May 2009 19:37:44 +0200

perf_counter: Fix dynamic irq_period logging

We call perf_adjust_freq() from perf_counter_task_tick() which is is called under the rq->lock causing lock recursion. However, it's no longer required to be called under the rq->lock, so remove it from under it.

Also, fix up some related comments.

Cc Paul Mackerras <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc John Kacur <[email protected]>

LKML-Reference: <[email protected]>

---
 include/linux/perf_counter.h |    1 +
 kernel/perf_counter.c        |    3 ++-
 kernel/sched.c               |    3 ++-
 3 files changed, 5 insertions(+), 2 deletions(-)
include/linux/perf_counter.h +1 -0
index 2eedae8..23ddd29 100644
@@ -260,6 +260,7 @@ enum perf_event_type {
 	/*
 	 * struct {
 	 * 	struct perf_event_header	header;
+	 * 	u64				time;
 	 * 	u64				irq_period;
 	 * };
 	 */
kernel/perf_counter.c +2 -1
index c100554..2f410ea 100644
@@ -2559,7 +2559,8 @@ void perf_counter_munmap(unsigned long addr, unsigned long len,
 }
 
 /*
- *
+ * Log irq_period changes so that analyzing tools can re-normalize the
+ * event flow.
  */
 
 static void perf_log_period(struct perf_counter *counter, u64 period)
kernel/sched.c +2 -1
index 4c0d58b..ad079f0 100644
@@ -4875,9 +4875,10 @@ void scheduler_tick(void)
 	update_rq_clock(rq);
 	update_cpu_load(rq);
 	curr->sched_class->task_tick(rq, curr, 0);
-	perf_counter_task_tick(curr, cpu);
 	spin_unlock(&rq->lock);
 
+	perf_counter_task_tick(curr, cpu);
+
 #ifdef CONFIG_SMP
 	rq->idle_at_tick = idle_cpu(cpu);
 	trigger_load_balance(rq, cpu);
#3 lore raw

Re: [PATCH 1/7] perf_counter: fix dynamic irq_period logging

In reply to Peter Zijlstra
Quoted 2 lines · Peter Zijlstra

We call perf_adjust_freq() from perf_counter_task_tick() which is is called under the rq->lock causing lock recursion.

What was the lock recursion? I see perf_adjust_freq taking ctx->lock, but we were careful not to take any rq->lock within a ctx->lock, at least in the past.

Paul.

#4 lore raw

Re: [PATCH 1/7] perf_counter: fix dynamic irq_period logging

Quoted 8 lines · Paul Mackerras
Quoted 2 lines · Peter Zijlstra

We call perf_adjust_freq() from perf_counter_task_tick() which is is called under the rq->lock causing lock recursion.

What was the lock recursion? I see perf_adjust_freq taking ctx->lock, but we were careful not to take any rq->lock within a ctx->lock, at least in the past.

perf_output_end() can end up calling a wake_up().

perf_output_end()
perf_output_unlock()

    perf_output_wakeup()
      perf_counter_wakeup()
        wake_up_all();
#5 lore raw

PATCH 2/7 perf_counter: sanitize counter->mutex

In reply to Peter Zijlstra

s/counter->mutex/counter->child_mutex/ and make sure its only used to protect child_list.

The usage in __perf_counter_exit_task() doesn't appear to be problematic since ctx->mutex also covers anything related to fd tear-down.

LKML-Reference: <new-submission>

---
 include/linux/perf_counter.h |    7 ++----
 kernel/perf_counter.c        |   47 +++++++++++++++++--------------------------
 2 files changed, 22 insertions(+), 32 deletions(-)
include/linux/perf_counter.h +3 -4
Index: linux-2.6/include/linux/perf_counter.h
===================================================================
@@ -452,9 +452,6 @@ struct perf_counter {
 	struct perf_counter_context	*ctx;
 	struct file			*filp;
 
-	struct perf_counter		*parent;
-	struct list_head		child_list;
-
 	/*
 	 * These accumulate total time (in nanoseconds) that children
 	 * counters have been enabled and running, respectively.
@@ -465,7 +462,9 @@ struct perf_counter {
 	/*
 	 * Protect attach/detach and child_list:
 	 */
-	struct mutex			mutex;
+	struct mutex			child_mutex;
+	struct list_head		child_list;
+	struct perf_counter		*parent;
 
 	int				oncpu;
 	int				cpu;
kernel/perf_counter.c +19 -28

137 lines not on this page · raw message

#6 lore raw

[tip:perfcounters/core] perf_counter: Sanitize counter->mutex

Commit-ID:  fccc714b3148ab9741fafc1e90c3876d50df6093
Gitweb:     http://git.kernel.org/tip/fccc714b3148ab9741fafc1e90c3876d50df6093
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Sat, 23 May 2009 18:28:56 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Sat, 23 May 2009 19:37:45 +0200

perf_counter: Sanitize counter->mutex

s/counter->mutex/counter->child_mutex/ and make sure its only used to protect child_list.

The usage in __perf_counter_exit_task() doesn't appear to be problematic since ctx->mutex also covers anything related to fd tear-down.

Cc Paul Mackerras <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc John Kacur <[email protected]>

LKML-Reference: <[email protected]>

---
 include/linux/perf_counter.h |    7 ++---
 kernel/perf_counter.c        |   47 +++++++++++++++++-------------------------
 2 files changed, 22 insertions(+), 32 deletions(-)
include/linux/perf_counter.h +3 -4
index 23ddd29..4ab8050 100644
@@ -452,9 +452,6 @@ struct perf_counter {
 	struct perf_counter_context	*ctx;
 	struct file			*filp;
 
-	struct perf_counter		*parent;
-	struct list_head		child_list;
-
 	/*
 	 * These accumulate total time (in nanoseconds) that children
 	 * counters have been enabled and running, respectively.
@@ -465,7 +462,9 @@ struct perf_counter {
 	/*
 	 * Protect attach/detach and child_list:
 	 */
-	struct mutex			mutex;
+	struct mutex			child_mutex;
+	struct list_head		child_list;
+	struct perf_counter		*parent;
 
 	int				oncpu;
 	int				cpu;
kernel/perf_counter.c +19 -28

137 lines not on this page · raw message

#7 lore raw

PATCH 3/7 perf_counter: sanitize context locking

In reply to Peter Zijlstra

Ensure we're consistent with the context locks.

context->mutex
context->lock

     list_{add,del}_counter();

so that either lock is sufficient to stabilize the context.

LKML-Reference: <new-submission>

---
 kernel/perf_counter.c |    9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)
kernel/perf_counter.c +7 -2
Index: linux-2.6/kernel/perf_counter.c
===================================================================
@@ -597,6 +597,8 @@ static void add_counter_to_ctx(struct pe
 
 /*
  * Cross CPU call to install and enable a performance counter
+ *
+ * Must be called with ctx->mutex held
  */
 static void __perf_install_in_context(void *info)
 {
@@ -1496,13 +1498,13 @@ static void perf_counter_for_each_siblin
 	struct perf_counter_context *ctx = counter->ctx;
 	struct perf_counter *sibling;
 
-	spin_lock_irq(&ctx->lock);
+	mutex_lock(&ctx->mutex);
 	counter = counter->group_leader;
 
 	func(counter);
 	list_for_each_entry(sibling, &counter->sibling_list, list_entry)
 		func(sibling);
-	spin_unlock_irq(&ctx->lock);
+	mutex_unlock(&ctx->mutex);
 }
 
 static void perf_counter_for_each_child(struct perf_counter *counter,
@@ -3414,7 +3416,10 @@ __perf_counter_exit_task(struct task_str
 	struct perf_counter *parent_counter;
 
 	update_counter_times(child_counter);
+
+	spin_lock_irq(&child_ctx->lock);
 	list_del_counter(child_counter, child_ctx);
+	spin_unlock_irq(&child_ctx->lock);
 
 	parent_counter = child_counter->parent;
 	/*
#8 lore raw

[tip:perfcounters/core] perf_counter: Sanitize context locking

Commit-ID:  682076ae1de0aba9c2da509f7b19dc03e30a6e1f
Gitweb:     http://git.kernel.org/tip/682076ae1de0aba9c2da509f7b19dc03e30a6e1f
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Sat, 23 May 2009 18:28:57 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Sat, 23 May 2009 19:37:46 +0200

perf_counter: Sanitize context locking

Ensure we're consistent with the context locks.

context->mutex
context->lock

     list_{add,del}_counter();

so that either lock is sufficient to stabilize the context.

Cc Paul Mackerras <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc John Kacur <[email protected]>

LKML-Reference: <[email protected]>

---
 kernel/perf_counter.c |    9 +++++++--
 1 files changed, 7 insertions(+), 2 deletions(-)
kernel/perf_counter.c +7 -2
index 679c3b5..d162d2f 100644
@@ -597,6 +597,8 @@ static void add_counter_to_ctx(struct perf_counter *counter,
 
 /*
  * Cross CPU call to install and enable a performance counter
+ *
+ * Must be called with ctx->mutex held
  */
 static void __perf_install_in_context(void *info)
 {
@@ -1496,13 +1498,13 @@ static void perf_counter_for_each_sibling(struct perf_counter *counter,
 	struct perf_counter_context *ctx = counter->ctx;
 	struct perf_counter *sibling;
 
-	spin_lock_irq(&ctx->lock);
+	mutex_lock(&ctx->mutex);
 	counter = counter->group_leader;
 
 	func(counter);
 	list_for_each_entry(sibling, &counter->sibling_list, list_entry)
 		func(sibling);
-	spin_unlock_irq(&ctx->lock);
+	mutex_unlock(&ctx->mutex);
 }
 
 static void perf_counter_for_each_child(struct perf_counter *counter,
@@ -3414,7 +3416,10 @@ __perf_counter_exit_task(struct task_struct *child,
 	struct perf_counter *parent_counter;
 
 	update_counter_times(child_counter);
+
+	spin_lock_irq(&child_ctx->lock);
 	list_del_counter(child_counter, child_ctx);
+	spin_unlock_irq(&child_ctx->lock);
 
 	parent_counter = child_counter->parent;
 	/*
#9 lore raw

PATCH 4/7 perf_counter: fix userspace build

In reply to Peter Zijlstra

recent userspace (F11) seems to already include the linux/unistd.h bits which means we cannot include the version in the kernel sources due to the header guards being the same.

Ensure we include the kernel version first.

LKML-Reference: <new-submission>

---
 Documentation/perf_counter/builtin-record.c |    3 --
 Documentation/perf_counter/builtin-stat.c   |    5 ----
 Documentation/perf_counter/builtin-top.c    |    5 ----
 Documentation/perf_counter/perf.h           |   31 +++++++++++++++++-----------
 4 files changed, 22 insertions(+), 22 deletions(-)
Documentation/perf_counter/builtin-record.c +1 -2
Index: linux-2.6/Documentation/perf_counter/builtin-record.c
===================================================================
@@ -1,5 +1,6 @@
 
 
+#include "perf.h"
 #include "util/util.h"
 
 #include <sys/types.h>
@@ -30,9 +31,7 @@
 #include <linux/unistd.h>
 #include <linux/types.h>
 
-#include "../../include/linux/perf_counter.h"
 
-#include "perf.h"
 
 #define ALIGN(x,a)		__ALIGN_MASK(x,(typeof(x))(a)-1)
 #define __ALIGN_MASK(x,mask)	(((x)+(mask))&~(mask))
Documentation/perf_counter/builtin-stat.c +1 -4
Index: linux-2.6/Documentation/perf_counter/builtin-stat.c
===================================================================
@@ -61,6 +61,7 @@
   * Released under the GPL v2. (and only v2, not any later version)
   */
 
+#include "perf.h"
 #include "util/util.h"
 
 #include <getopt.h>
@@ -83,10 +84,6 @@
 #include <linux/unistd.h>
 #include <linux/types.h>
 
-#include "../../include/linux/perf_counter.h"
-
-#include "perf.h"
-
 #define EVENT_MASK_KERNEL		1
 #define EVENT_MASK_USER			2
 
Documentation/perf_counter/builtin-top.c +1 -4
Index: linux-2.6/Documentation/perf_counter/builtin-top.c
===================================================================
@@ -42,6 +42,7 @@
   * Released under the GPL v2. (and only v2, not any later version)
   */
 
+#include "perf.h"
 #include "util/util.h"
 
 #include <getopt.h>
@@ -64,10 +65,6 @@
 #include <linux/unistd.h>
 #include <linux/types.h>
 
-#include "../../include/linux/perf_counter.h"
-
-#include "perf.h"
-
 static int			system_wide			=  0;
 
 static int			nr_counters			=  0;
Documentation/perf_counter/perf.h +19 -12
Index: linux-2.6/Documentation/perf_counter/perf.h
===================================================================
@@ -1,6 +1,25 @@
 #ifndef _PERF_PERF_H
 #define _PERF_PERF_H
 
+#if defined(__x86_64__) || defined(__i386__)
+#include "../../arch/x86/include/asm/unistd.h"
+#define rmb()		asm volatile("lfence" ::: "memory")
+#define cpu_relax()	asm volatile("rep; nop" ::: "memory");
+#endif
+
+#ifdef __powerpc__
+#include "../../arch/powerpc/include/asm/unistd.h"
+#define rmb()		asm volatile ("sync" ::: "memory")
+#define cpu_relax()	asm volatile ("" ::: "memory");
+#endif
+
+#include <time.h>
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/syscall.h>
+
+#include "../../include/linux/perf_counter.h"
+
 /*
  * prctl(PR_TASK_PERF_COUNTERS_DISABLE) will (cheaply) disable all
  * counters in the current task.
@@ -26,18 +45,6 @@ static inline unsigned long long rdclock
 #define __user
 #define asmlinkage
 
-#if defined(__x86_64__) || defined(__i386__)
-#include "../../arch/x86/include/asm/unistd.h"
-#define rmb()		asm volatile("lfence" ::: "memory")
-#define cpu_relax()	asm volatile("rep; nop" ::: "memory");
-#endif
-
-#ifdef __powerpc__
-#include "../../arch/powerpc/include/asm/unistd.h"
-#define rmb()		asm volatile ("sync" ::: "memory")
-#define cpu_relax()	asm volatile ("" ::: "memory");
-#endif
-
 #define unlikely(x)	__builtin_expect(!!(x), 0)
 #define min(x, y) ({				\
 	typeof(x) _min1 = (x);			\
#10 lore raw

[tip:perfcounters/core] perf_counter: Fix userspace build

Commit-ID:  1a482f38c5aafeb3576079a38a5b21b46619f3d2
Gitweb:     http://git.kernel.org/tip/1a482f38c5aafeb3576079a38a5b21b46619f3d2
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Sat, 23 May 2009 18:28:58 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Sat, 23 May 2009 19:37:46 +0200

perf_counter: Fix userspace build

recent userspace (F11) seems to already include the linux/unistd.h bits which means we cannot include the version in the kernel sources due to the header guards being the same.

Ensure we include the kernel version first.

Cc Paul Mackerras <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc John Kacur <[email protected]>

LKML-Reference: <[email protected]>

---
 Documentation/perf_counter/builtin-record.c |    3 +-
 Documentation/perf_counter/builtin-stat.c   |    5 +---
 Documentation/perf_counter/builtin-top.c    |    5 +---
 Documentation/perf_counter/perf.h           |   31 ++++++++++++++++----------
 4 files changed, 22 insertions(+), 22 deletions(-)
Documentation/perf_counter/builtin-record.c +1 -2
index efb8759..1b19f18 100644
@@ -1,5 +1,6 @@
 
 
+#include "perf.h"
 #include "util/util.h"
 
 #include <sys/types.h>
@@ -30,9 +31,7 @@
 #include <linux/unistd.h>
 #include <linux/types.h>
 
-#include "../../include/linux/perf_counter.h"
 
-#include "perf.h"
 
 #define ALIGN(x,a)		__ALIGN_MASK(x,(typeof(x))(a)-1)
 #define __ALIGN_MASK(x,mask)	(((x)+(mask))&~(mask))
Documentation/perf_counter/builtin-stat.c +1 -4
index 03518d7..8ae01d5 100644
@@ -61,6 +61,7 @@
   * Released under the GPL v2. (and only v2, not any later version)
   */
 
+#include "perf.h"
 #include "util/util.h"
 
 #include <getopt.h>
@@ -83,10 +84,6 @@
 #include <linux/unistd.h>
 #include <linux/types.h>
 
-#include "../../include/linux/perf_counter.h"
-
-#include "perf.h"
-
 #define EVENT_MASK_KERNEL		1
 #define EVENT_MASK_USER			2
 
Documentation/perf_counter/builtin-top.c +1 -4
index 814b2e4..a3216a6 100644
@@ -42,6 +42,7 @@
   * Released under the GPL v2. (and only v2, not any later version)
   */
 
+#include "perf.h"
 #include "util/util.h"
 
 #include <getopt.h>
@@ -64,10 +65,6 @@
 #include <linux/unistd.h>
 #include <linux/types.h>
 
-#include "../../include/linux/perf_counter.h"
-
-#include "perf.h"
-
 static int			system_wide			=  0;
 
 static int			nr_counters			=  0;
Documentation/perf_counter/perf.h +19 -12
index 81a7374..a517683 100644
@@ -1,6 +1,25 @@
 #ifndef _PERF_PERF_H
 #define _PERF_PERF_H
 
+#if defined(__x86_64__) || defined(__i386__)
+#include "../../arch/x86/include/asm/unistd.h"
+#define rmb()		asm volatile("lfence" ::: "memory")
+#define cpu_relax()	asm volatile("rep; nop" ::: "memory");
+#endif
+
+#ifdef __powerpc__
+#include "../../arch/powerpc/include/asm/unistd.h"
+#define rmb()		asm volatile ("sync" ::: "memory")
+#define cpu_relax()	asm volatile ("" ::: "memory");
+#endif
+
+#include <time.h>
+#include <unistd.h>
+#include <sys/types.h>
+#include <sys/syscall.h>
+
+#include "../../include/linux/perf_counter.h"
+
 /*
  * prctl(PR_TASK_PERF_COUNTERS_DISABLE) will (cheaply) disable all
  * counters in the current task.
@@ -26,18 +45,6 @@ static inline unsigned long long rdclock(void)
 #define __user
 #define asmlinkage
 
-#if defined(__x86_64__) || defined(__i386__)
-#include "../../arch/x86/include/asm/unistd.h"
-#define rmb()		asm volatile("lfence" ::: "memory")
-#define cpu_relax()	asm volatile("rep; nop" ::: "memory");
-#endif
-
-#ifdef __powerpc__
-#include "../../arch/powerpc/include/asm/unistd.h"
-#define rmb()		asm volatile ("sync" ::: "memory")
-#define cpu_relax()	asm volatile ("" ::: "memory");
-#endif
-
 #define unlikely(x)	__builtin_expect(!!(x), 0)
 #define min(x, y) ({				\
 	typeof(x) _min1 = (x);			\
#11 lore raw

PATCH 5/7 perf_counter: simplify context cleanup

In reply to Peter Zijlstra

Use perf_counter_remove_from_context() to remove counters from the context.

---
 kernel/perf_counter.c |    5 +----
 1 file changed, 1 insertion(+), 4 deletions(-)
kernel/perf_counter.c +1 -4
Index: linux-2.6/kernel/perf_counter.c
===================================================================
@@ -3416,10 +3416,7 @@ __perf_counter_exit_task(struct task_str
 	struct perf_counter *parent_counter;
 
 	update_counter_times(child_counter);
-
-	spin_lock_irq(&child_ctx->lock);
-	list_del_counter(child_counter, child_ctx);
-	spin_unlock_irq(&child_ctx->lock);
+	perf_counter_remove_from_context(child_counter);
 
 	parent_counter = child_counter->parent;
 	/*
#12 lore raw

[tip:perfcounters/core] perf_counter: Simplify context cleanup

Commit-ID:  aa9c67f53d1969cf1db4c9c2db3a78c4ceb96469
Gitweb:     http://git.kernel.org/tip/aa9c67f53d1969cf1db4c9c2db3a78c4ceb96469
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Sat, 23 May 2009 18:28:59 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Sat, 23 May 2009 19:37:47 +0200

perf_counter: Simplify context cleanup

Use perf_counter_remove_from_context() to remove counters from the context.

Cc Paul Mackerras <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc John Kacur <[email protected]>

LKML-Reference: <[email protected]>

---
 kernel/perf_counter.c |    5 +----
 1 files changed, 1 insertions(+), 4 deletions(-)
kernel/perf_counter.c +1 -4
index d162d2f..0e97f89 100644
@@ -3416,10 +3416,7 @@ __perf_counter_exit_task(struct task_struct *child,
 	struct perf_counter *parent_counter;
 
 	update_counter_times(child_counter);
-
-	spin_lock_irq(&child_ctx->lock);
-	list_del_counter(child_counter, child_ctx);
-	spin_unlock_irq(&child_ctx->lock);
+	perf_counter_remove_from_context(child_counter);
 
 	parent_counter = child_counter->parent;
 	/*
#13 lore raw

PATCH 6/7 perf_counter: change pctrl() behaviour

In reply to Peter Zijlstra

Instead of en/dis-abling all counters acting on a particular task, en/dis- able all counters we created.

LKML-Reference: <new-submission>

---
 include/linux/init_task.h    |   10 +++++
 include/linux/perf_counter.h |    3 +
 include/linux/sched.h        |    2 +
 kernel/perf_counter.c        |   84 ++++++++++---------------------------------
 4 files changed, 36 insertions(+), 63 deletions(-)
include/linux/init_task.h +10 -0
Index: linux-2.6/include/linux/init_task.h
===================================================================
@@ -108,6 +108,15 @@ extern struct group_info init_groups;
 
 extern struct cred init_cred;
 
+#ifdef CONFIG_PERF_COUNTERS
+# define INIT_PERF_COUNTERS(tsk)					\
+	.perf_counter_mutex = 						\
+		 __MUTEX_INITIALIZER(tsk.perf_counter_mutex),		\
+	.perf_counter_list = LIST_HEAD_INIT(tsk.perf_counter_list),
+#else
+# define INIT_PERF_COUNTERS(tsk)
+#endif
+
 /*
  *  INIT_TASK is used to set up the first task table, touch at
  * your own risk!. Base=0, limit=0x1fffff (=2MB)
@@ -171,6 +180,7 @@ extern struct cred init_cred;
 	},								\
 	.dirties = INIT_PROP_LOCAL_SINGLE(dirties),			\
 	INIT_IDS							\
+	INIT_PERF_COUNTERS(tsk)						\
 	INIT_TRACE_IRQFLAGS						\
 	INIT_LOCKDEP							\
 	INIT_FTRACE_GRAPH						\
include/linux/perf_counter.h +3 -0
Index: linux-2.6/include/linux/perf_counter.h
===================================================================
@@ -469,6 +469,9 @@ struct perf_counter {
 	int				oncpu;
 	int				cpu;
 
+	struct list_head		owner_entry;
+	struct task_struct		*owner;
+
 	/* mmap bits */
 	struct mutex			mmap_mutex;
 	atomic_t			mmap_count;
include/linux/sched.h +2 -0
Index: linux-2.6/include/linux/sched.h
===================================================================
@@ -1403,6 +1403,8 @@ struct task_struct {
 #endif
 #ifdef CONFIG_PERF_COUNTERS
 	struct perf_counter_context *perf_counter_ctxp;
+	struct mutex perf_counter_mutex;
+	struct list_head perf_counter_list;
 #endif
 #ifdef CONFIG_NUMA
 	struct mempolicy *mempolicy;
kernel/perf_counter.c +21 -63

112 lines not on this page · raw message

#14 lore raw

Re: [PATCH 6/7] perf_counter: change pctrl() behaviour

Quoted 7 lines · Peter Zijlstra

plain text document attachment (perf_counter-ctrl-all.patch) Instead of en/dis-abling all counters acting on a particular task, en/dis- able all counters we created.

LKML-Reference: <new-submission>

---

Please fold in the below so as to avoid an explosion on first sys_perf_counter_open().

kernel/perf_counter.c +3 -0
Index: linux-2.6/kernel/perf_counter.c
===================================================================
@@ -3446,6 +3446,9 @@ void perf_counter_init_task(struct task_
 
 	child->perf_counter_ctxp = NULL;
 
+	mutex_init(&child->perf_counter_mutex);
+	INIT_LIST_HEAD(&child->perf_counter_list);
+
 	/*
 	 * This is executed from the parent task context, so inherit
 	 * counters that have been marked for cloning.
#15 lore raw

[tip:perfcounters/core] perf_counter: Change pctrl() behaviour

In reply to Peter Zijlstra
Commit-ID:  082ff5a2767a0679ee543f14883adbafb631ffbe
Gitweb:     http://git.kernel.org/tip/082ff5a2767a0679ee543f14883adbafb631ffbe
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Sat, 23 May 2009 18:29:00 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Sun, 24 May 2009 08:24:08 +0200

perf_counter: Change pctrl() behaviour

Instead of en/dis-abling all counters acting on a particular task, en/dis- able all counters we created.

[ v2: fix crash on first counter enable ]

Cc Paul Mackerras <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc John Kacur <[email protected]>

LKML-Reference: <[email protected]>

---
 include/linux/init_task.h    |   10 +++++
 include/linux/perf_counter.h |    3 +
 include/linux/sched.h        |    2 +
 kernel/perf_counter.c        |   87 +++++++++++------------------------------
 4 files changed, 39 insertions(+), 63 deletions(-)
include/linux/init_task.h +10 -0
index d87247d..353c0ac 100644
@@ -108,6 +108,15 @@ extern struct group_info init_groups;
 
 extern struct cred init_cred;
 
+#ifdef CONFIG_PERF_COUNTERS
+# define INIT_PERF_COUNTERS(tsk)					\
+	.perf_counter_mutex = 						\
+		 __MUTEX_INITIALIZER(tsk.perf_counter_mutex),		\
+	.perf_counter_list = LIST_HEAD_INIT(tsk.perf_counter_list),
+#else
+# define INIT_PERF_COUNTERS(tsk)
+#endif
+
 /*
  *  INIT_TASK is used to set up the first task table, touch at
  * your own risk!. Base=0, limit=0x1fffff (=2MB)
@@ -171,6 +180,7 @@ extern struct cred init_cred;
 	},								\
 	.dirties = INIT_PROP_LOCAL_SINGLE(dirties),			\
 	INIT_IDS							\
+	INIT_PERF_COUNTERS(tsk)						\
 	INIT_TRACE_IRQFLAGS						\
 	INIT_LOCKDEP							\
 	INIT_FTRACE_GRAPH						\
include/linux/perf_counter.h +3 -0
index 4ab8050..4159ee5 100644
@@ -469,6 +469,9 @@ struct perf_counter {
 	int				oncpu;
 	int				cpu;
 
+	struct list_head		owner_entry;
+	struct task_struct		*owner;
+
 	/* mmap bits */
 	struct mutex			mmap_mutex;
 	atomic_t			mmap_count;
include/linux/sched.h +2 -0
index 9714d45..bc9326d 100644
@@ -1389,6 +1389,8 @@ struct task_struct {
 #endif
 #ifdef CONFIG_PERF_COUNTERS
 	struct perf_counter_context *perf_counter_ctxp;
+	struct mutex perf_counter_mutex;
+	struct list_head perf_counter_list;
 #endif
 #ifdef CONFIG_NUMA
 	struct mempolicy *mempolicy;
kernel/perf_counter.c +24 -63

121 lines not on this page · raw message

#16 lore raw

PATCH 7/7 perf_counter: remove perf_counter_context::nr_enabled

In reply to Peter Zijlstra

now that pctrl() no longer disables other people's counters, remove the PMU cache code that deals with that.

LKML-Reference: <new-submission>

---
 include/linux/perf_counter.h |    1 -
 kernel/perf_counter.c        |   11 +----------
 2 files changed, 1 insertion(+), 11 deletions(-)
include/linux/perf_counter.h +0 -1
Index: linux-2.6/include/linux/perf_counter.h
===================================================================
@@ -516,7 +516,6 @@ struct perf_counter_context {
 	struct list_head	event_list;
 	int			nr_counters;
 	int			nr_active;
-	int			nr_enabled;
 	int			is_active;
 	atomic_t		refcount;
 	struct task_struct	*task;
kernel/perf_counter.c +1 -10
Index: linux-2.6/kernel/perf_counter.c
===================================================================
@@ -134,8 +134,6 @@ list_add_counter(struct perf_counter *co
 
 	list_add_rcu(&counter->event_entry, &ctx->event_list);
 	ctx->nr_counters++;
-	if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
-		ctx->nr_enabled++;
 }
 
 /*
@@ -150,8 +148,6 @@ list_del_counter(struct perf_counter *co
 	if (list_empty(&counter->list_entry))
 		return;
 	ctx->nr_counters--;
-	if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
-		ctx->nr_enabled--;
 
 	list_del_init(&counter->list_entry);
 	list_del_rcu(&counter->event_entry);
@@ -406,7 +402,6 @@ static void __perf_counter_disable(void 
 		else
 			counter_sched_out(counter, cpuctx, ctx);
 		counter->state = PERF_COUNTER_STATE_OFF;
-		ctx->nr_enabled--;
 	}
 
 	spin_unlock_irqrestore(&ctx->lock, flags);
@@ -448,7 +443,6 @@ static void perf_counter_disable(struct 
 	if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
 		update_counter_times(counter);
 		counter->state = PERF_COUNTER_STATE_OFF;
-		ctx->nr_enabled--;
 	}
 
 	spin_unlock_irq(&ctx->lock);
@@ -759,7 +753,6 @@ static void __perf_counter_enable(void *
 		goto unlock;
 	counter->state = PERF_COUNTER_STATE_INACTIVE;
 	counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
-	ctx->nr_enabled++;
 
 	/*
 	 * If the counter is in a group and isn't the group leader,
@@ -850,7 +843,6 @@ static void perf_counter_enable(struct p
 		counter->state = PERF_COUNTER_STATE_INACTIVE;
 		counter->tstamp_enabled =
 			ctx->time - counter->total_time_enabled;
-		ctx->nr_enabled++;
 	}
  out:
 	spin_unlock_irq(&ctx->lock);
@@ -910,8 +902,7 @@ static int context_equiv(struct perf_cou
 			 struct perf_counter_context *ctx2)
 {
 	return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
-		&& ctx1->parent_gen == ctx2->parent_gen
-		&& ctx1->nr_enabled == ctx2->nr_enabled;
+		&& ctx1->parent_gen == ctx2->parent_gen;
 }
 
 /*
#17 lore raw

[tip:perfcounters/core] perf_counter: Remove perf_counter_context::nr_enabled

Commit-ID:  475c55797323b67435083f6e2eb8ee670f6410ec
Gitweb:     http://git.kernel.org/tip/475c55797323b67435083f6e2eb8ee670f6410ec
Author:     Peter Zijlstra <[email protected]>
AuthorDate: Sat, 23 May 2009 18:29:01 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Sun, 24 May 2009 08:24:30 +0200

perf_counter: Remove perf_counter_context::nr_enabled

now that pctrl() no longer disables other people's counters, remove the PMU cache code that deals with that.

Cc Paul Mackerras <[email protected]>
Cc Corey Ashford <[email protected]>
Cc Arnaldo Carvalho de Melo <[email protected]>
Cc John Kacur <[email protected]>

LKML-Reference: <[email protected]>

---
 include/linux/perf_counter.h |    1 -
 kernel/perf_counter.c        |   11 +----------
 2 files changed, 1 insertions(+), 11 deletions(-)
include/linux/perf_counter.h +0 -1
index 4159ee5..2ddf5e3 100644
@@ -516,7 +516,6 @@ struct perf_counter_context {
 	struct list_head	event_list;
 	int			nr_counters;
 	int			nr_active;
-	int			nr_enabled;
 	int			is_active;
 	atomic_t		refcount;
 	struct task_struct	*task;
kernel/perf_counter.c +1 -10
index 4c86a63..cb40625 100644
@@ -134,8 +134,6 @@ list_add_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
 
 	list_add_rcu(&counter->event_entry, &ctx->event_list);
 	ctx->nr_counters++;
-	if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
-		ctx->nr_enabled++;
 }
 
 /*
@@ -150,8 +148,6 @@ list_del_counter(struct perf_counter *counter, struct perf_counter_context *ctx)
 	if (list_empty(&counter->list_entry))
 		return;
 	ctx->nr_counters--;
-	if (counter->state >= PERF_COUNTER_STATE_INACTIVE)
-		ctx->nr_enabled--;
 
 	list_del_init(&counter->list_entry);
 	list_del_rcu(&counter->event_entry);
@@ -406,7 +402,6 @@ static void __perf_counter_disable(void *info)
 		else
 			counter_sched_out(counter, cpuctx, ctx);
 		counter->state = PERF_COUNTER_STATE_OFF;
-		ctx->nr_enabled--;
 	}
 
 	spin_unlock_irqrestore(&ctx->lock, flags);
@@ -448,7 +443,6 @@ static void perf_counter_disable(struct perf_counter *counter)
 	if (counter->state == PERF_COUNTER_STATE_INACTIVE) {
 		update_counter_times(counter);
 		counter->state = PERF_COUNTER_STATE_OFF;
-		ctx->nr_enabled--;
 	}
 
 	spin_unlock_irq(&ctx->lock);
@@ -759,7 +753,6 @@ static void __perf_counter_enable(void *info)
 		goto unlock;
 	counter->state = PERF_COUNTER_STATE_INACTIVE;
 	counter->tstamp_enabled = ctx->time - counter->total_time_enabled;
-	ctx->nr_enabled++;
 
 	/*
 	 * If the counter is in a group and isn't the group leader,
@@ -850,7 +843,6 @@ static void perf_counter_enable(struct perf_counter *counter)
 		counter->state = PERF_COUNTER_STATE_INACTIVE;
 		counter->tstamp_enabled =
 			ctx->time - counter->total_time_enabled;
-		ctx->nr_enabled++;
 	}
  out:
 	spin_unlock_irq(&ctx->lock);
@@ -910,8 +902,7 @@ static int context_equiv(struct perf_counter_context *ctx1,
 			 struct perf_counter_context *ctx2)
 {
 	return ctx1->parent_ctx && ctx1->parent_ctx == ctx2->parent_ctx
-		&& ctx1->parent_gen == ctx2->parent_gen
-		&& ctx1->nr_enabled == ctx2->nr_enabled;
+		&& ctx1->parent_gen == ctx2->parent_gen;
 }
 
 /*