[PATCH 0/7] pending sched and perf_counter patches

[PATCH 0/7] pending sched and perf_counter patches

[Peter Zijlstra]

-- 

[PATCH 1/7] sched: rt: document the risk of small values in the bandwidth settings

[Peter Zijlstra]

[-- Attachment #1: sched-rt-docu.patch --]
[-- Type: text/plain, Size: 2062 bytes --]

Thomas noted that we should disallow sysctl_sched_rt_runtime == 0 for (!
RT_GROUP) since the root group always has some RT tasks in it.

Further, update the documentation to inspire clue.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 Documentation/scheduler/sched-rt-group.txt |   18 ++++++++++++++++++
 kernel/sched.c                             |    7 +++++++
 2 files changed, 25 insertions(+)

Index: linux-2.6/Documentation/scheduler/sched-rt-group.txt
===================================================================
--- linux-2.6.orig/Documentation/scheduler/sched-rt-group.txt
+++ linux-2.6/Documentation/scheduler/sched-rt-group.txt
@@ -4,6 +4,7 @@
 CONTENTS
 ========
 
+0. WARNING
 1. Overview
   1.1 The problem
   1.2 The solution
@@ -14,6 +15,23 @@ CONTENTS
 3. Future plans
 
 
+0. WARNING
+==========
+
+ Fiddling with these settings can result in an unstable system, the knobs are
+ root only and assumes root knows what he is doing.
+
+Most notable:
+
+ * very small values in sched_rt_period_us can result in an unstable
+   system when the period is smaller than either the available hrtimer
+   resolution, or the time it takes to handle the budget refresh itself.
+
+ * very small values in sched_rt_runtime_us can result in an unstable
+   system when the runtime is so small the system has difficulty making
+   forward progress (NOTE: the migration thread and kstopmachine both
+   are real-time processes).
+
 1. Overview
 ===========
 
Index: linux-2.6/kernel/sched.c
===================================================================
--- linux-2.6.orig/kernel/sched.c
+++ linux-2.6/kernel/sched.c
@@ -10024,6 +10024,13 @@ static int sched_rt_global_constraints(v
 	if (sysctl_sched_rt_period <= 0)
 		return -EINVAL;
 
+	/*
+	 * There's always some RT tasks in the root group
+	 * -- migration, kstopmachine etc..
+	 */
+	if (sysctl_sched_rt_runtime == 0)
+		return -EBUSY;
+
 	spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
 	for_each_possible_cpu(i) {
 		struct rt_rq *rt_rq = &cpu_rq(i)->rt;

-- 

[tip:sched/core] sched: rt: document the risk of small values in the bandwidth settings

[tip-bot for Peter Zijlstra]

Commit-ID:  60aa605dfce2976e54fa76e805ab0f221372d4d9
Gitweb:     http://git.kernel.org/tip/60aa605dfce2976e54fa76e805ab0f221372d4d9
Author:     Peter Zijlstra <a.p.zijlstra@chello.nl>
AuthorDate: Tue, 5 May 2009 17:50:21 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Tue, 5 May 2009 20:07:57 +0200

sched: rt: document the risk of small values in the bandwidth settings

Thomas noted that we should disallow sysctl_sched_rt_runtime == 0 for
(!RT_GROUP) since the root group always has some RT tasks in it.

Further, update the documentation to inspire clue.

[ Impact: exclude corner-case sysctl_sched_rt_runtime value ]

Reported-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
LKML-Reference: <20090505155436.863098054@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 Documentation/scheduler/sched-rt-group.txt |   18 ++++++++++++++++++
 kernel/sched.c                             |    7 +++++++
 2 files changed, 25 insertions(+), 0 deletions(-)

diff --git a/Documentation/scheduler/sched-rt-group.txt b/Documentation/scheduler/sched-rt-group.txt
index 5ba4d3f..eb74b01 100644
--- a/Documentation/scheduler/sched-rt-group.txt
+++ b/Documentation/scheduler/sched-rt-group.txt
@@ -4,6 +4,7 @@
 CONTENTS
 ========
 
+0. WARNING
 1. Overview
   1.1 The problem
   1.2 The solution
@@ -14,6 +15,23 @@ CONTENTS
 3. Future plans
 
 
+0. WARNING
+==========
+
+ Fiddling with these settings can result in an unstable system, the knobs are
+ root only and assumes root knows what he is doing.
+
+Most notable:
+
+ * very small values in sched_rt_period_us can result in an unstable
+   system when the period is smaller than either the available hrtimer
+   resolution, or the time it takes to handle the budget refresh itself.
+
+ * very small values in sched_rt_runtime_us can result in an unstable
+   system when the runtime is so small the system has difficulty making
+   forward progress (NOTE: the migration thread and kstopmachine both
+   are real-time processes).
+
 1. Overview
 ===========
 
diff --git a/kernel/sched.c b/kernel/sched.c
index 54d67b9..2a43a58 100644
--- a/kernel/sched.c
+++ b/kernel/sched.c
@@ -9917,6 +9917,13 @@ static int sched_rt_global_constraints(void)
 	if (sysctl_sched_rt_period <= 0)
 		return -EINVAL;
 
+	/*
+	 * There's always some RT tasks in the root group
+	 * -- migration, kstopmachine etc..
+	 */
+	if (sysctl_sched_rt_runtime == 0)
+		return -EBUSY;
+
 	spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
 	for_each_possible_cpu(i) {
 		struct rt_rq *rt_rq = &cpu_rq(i)->rt;

[PATCH 2/7] perf_counter: uncouple data_head updates from wakeups

[Peter Zijlstra]

[-- Attachment #1: perf_counter-uncouple-data_head-from-wakeup.patch --]
[-- Type: text/plain, Size: 3159 bytes --]

Keep data_head up-to-date irrespective of notifications, this fixes
the case where you disable a counter and don't get a notification for
the last few pending events, and also allows polling usage.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 include/linux/perf_counter.h |    4 +++-
 kernel/perf_counter.c        |   20 +++++++++-----------
 2 files changed, 12 insertions(+), 12 deletions(-)

Index: linux-2.6/kernel/perf_counter.c
===================================================================
--- linux-2.6.orig/kernel/perf_counter.c
+++ linux-2.6/kernel/perf_counter.c
@@ -1696,7 +1696,6 @@ struct perf_output_handle {
 	struct perf_mmap_data	*data;
 	unsigned int		offset;
 	unsigned int		head;
-	int			wakeup;
 	int			nmi;
 	int			overflow;
 	int			locked;
@@ -1752,8 +1751,7 @@ static void perf_output_unlock(struct pe
 	struct perf_mmap_data *data = handle->data;
 	int head, cpu;
 
-	if (handle->wakeup)
-		data->wakeup_head = data->head;
+	data->done_head = data->head;
 
 	if (!handle->locked)
 		goto out;
@@ -1764,13 +1762,11 @@ again:
 	 * before we publish the new head, matched by a rmb() in userspace when
 	 * reading this position.
 	 */
-	while ((head = atomic_xchg(&data->wakeup_head, 0))) {
+	while ((head = atomic_xchg(&data->done_head, 0)))
 		data->user_page->data_head = head;
-		handle->wakeup = 1;
-	}
 
 	/*
-	 * NMI can happen here, which means we can miss a wakeup_head update.
+	 * NMI can happen here, which means we can miss a done_head update.
 	 */
 
 	cpu = atomic_xchg(&data->lock, 0);
@@ -1779,7 +1775,7 @@ again:
 	/*
 	 * Therefore we have to validate we did not indeed do so.
 	 */
-	if (unlikely(atomic_read(&data->wakeup_head))) {
+	if (unlikely(atomic_read(&data->done_head))) {
 		/*
 		 * Since we had it locked, we can lock it again.
 		 */
@@ -1789,7 +1785,7 @@ again:
 		goto again;
 	}
 
-	if (handle->wakeup)
+	if (atomic_xchg(&data->wakeup, 0))
 		perf_output_wakeup(handle);
 out:
 	local_irq_restore(handle->flags);
@@ -1824,7 +1820,9 @@ static int perf_output_begin(struct perf
 
 	handle->offset	= offset;
 	handle->head	= head;
-	handle->wakeup	= (offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT);
+
+	if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
+		atomic_set(&data->wakeup, 1);
 
 	return 0;
 
@@ -1882,7 +1880,7 @@ static void perf_output_end(struct perf_
 		int events = atomic_inc_return(&data->events);
 		if (events >= wakeup_events) {
 			atomic_sub(wakeup_events, &data->events);
-			handle->wakeup = 1;
+			atomic_set(&data->wakeup, 1);
 		}
 	}
 
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
@@ -362,9 +362,11 @@ struct perf_mmap_data {
 	atomic_t			head;		/* write position    */
 	atomic_t			events;		/* event limit       */
 
-	atomic_t			wakeup_head;	/* completed head    */
+	atomic_t			done_head;	/* completed head    */
 	atomic_t			lock;		/* concurrent writes */
 
+	atomic_t			wakeup;		/* needs a wakeup    */
+
 	struct perf_counter_mmap_page   *user_page;
 	void 				*data_pages[0];
 };

-- 

[tip:perfcounters/core] perf_counter: uncouple data_head updates from wakeups

[tip-bot for Peter Zijlstra]

Commit-ID:  c66de4a5be7913247bd83d79168f8e4420c9cfbc
Gitweb:     http://git.kernel.org/tip/c66de4a5be7913247bd83d79168f8e4420c9cfbc
Author:     Peter Zijlstra <a.p.zijlstra@chello.nl>
AuthorDate: Tue, 5 May 2009 17:50:22 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Tue, 5 May 2009 20:18:30 +0200

perf_counter: uncouple data_head updates from wakeups

Keep data_head up-to-date irrespective of notifications. This fixes
the case where you disable a counter and don't get a notification for
the last few pending events, and it also allows polling usage.

[ Impact: increase precision of perfcounter mmap-ed fields ]

Suggested-by: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <20090505155436.925084300@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 include/linux/perf_counter.h |    4 +++-
 kernel/perf_counter.c        |   20 +++++++++-----------
 2 files changed, 12 insertions(+), 12 deletions(-)

diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index a356fa6..17b6310 100644
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -362,9 +362,11 @@ struct perf_mmap_data {
 	atomic_t			head;		/* write position    */
 	atomic_t			events;		/* event limit       */
 
-	atomic_t			wakeup_head;	/* completed head    */
+	atomic_t			done_head;	/* completed head    */
 	atomic_t			lock;		/* concurrent writes */
 
+	atomic_t			wakeup;		/* needs a wakeup    */
+
 	struct perf_counter_mmap_page   *user_page;
 	void 				*data_pages[0];
 };
diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 5f86a11..ba5e921 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -1696,7 +1696,6 @@ struct perf_output_handle {
 	struct perf_mmap_data	*data;
 	unsigned int		offset;
 	unsigned int		head;
-	int			wakeup;
 	int			nmi;
 	int			overflow;
 	int			locked;
@@ -1752,8 +1751,7 @@ static void perf_output_unlock(struct perf_output_handle *handle)
 	struct perf_mmap_data *data = handle->data;
 	int head, cpu;
 
-	if (handle->wakeup)
-		data->wakeup_head = data->head;
+	data->done_head = data->head;
 
 	if (!handle->locked)
 		goto out;
@@ -1764,13 +1762,11 @@ again:
 	 * before we publish the new head, matched by a rmb() in userspace when
 	 * reading this position.
 	 */
-	while ((head = atomic_xchg(&data->wakeup_head, 0))) {
+	while ((head = atomic_xchg(&data->done_head, 0)))
 		data->user_page->data_head = head;
-		handle->wakeup = 1;
-	}
 
 	/*
-	 * NMI can happen here, which means we can miss a wakeup_head update.
+	 * NMI can happen here, which means we can miss a done_head update.
 	 */
 
 	cpu = atomic_xchg(&data->lock, 0);
@@ -1779,7 +1775,7 @@ again:
 	/*
 	 * Therefore we have to validate we did not indeed do so.
 	 */
-	if (unlikely(atomic_read(&data->wakeup_head))) {
+	if (unlikely(atomic_read(&data->done_head))) {
 		/*
 		 * Since we had it locked, we can lock it again.
 		 */
@@ -1789,7 +1785,7 @@ again:
 		goto again;
 	}
 
-	if (handle->wakeup)
+	if (atomic_xchg(&data->wakeup, 0))
 		perf_output_wakeup(handle);
 out:
 	local_irq_restore(handle->flags);
@@ -1824,7 +1820,9 @@ static int perf_output_begin(struct perf_output_handle *handle,
 
 	handle->offset	= offset;
 	handle->head	= head;
-	handle->wakeup	= (offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT);
+
+	if ((offset >> PAGE_SHIFT) != (head >> PAGE_SHIFT))
+		atomic_set(&data->wakeup, 1);
 
 	return 0;
 
@@ -1882,7 +1880,7 @@ static void perf_output_end(struct perf_output_handle *handle)
 		int events = atomic_inc_return(&data->events);
 		if (events >= wakeup_events) {
 			atomic_sub(wakeup_events, &data->events);
-			handle->wakeup = 1;
+			atomic_set(&data->wakeup, 1);
 		}
 	}
 

[PATCH 3/7] perf_counter: ioctl(PERF_COUNTER_IOC_RESET)

[Peter Zijlstra]

[-- Attachment #1: perf_counter-reset.patch --]
[-- Type: text/plain, Size: 1562 bytes --]

Provide a way to reset an existing counter.

Similar to read() it doesn't collapse pending child counters.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 include/linux/perf_counter.h |    1 +
 kernel/perf_counter.c        |    8 ++++++++
 2 files changed, 9 insertions(+)

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
@@ -160,6 +160,7 @@ struct perf_counter_hw_event {
 #define PERF_COUNTER_IOC_ENABLE		_IO ('$', 0)
 #define PERF_COUNTER_IOC_DISABLE	_IO ('$', 1)
 #define PERF_COUNTER_IOC_REFRESH	_IOW('$', 2, u32)
+#define PERF_COUNTER_IOC_RESET		_IO ('$', 3)
 
 /*
  * Structure of the page that can be mapped via mmap
Index: linux-2.6/kernel/perf_counter.c
===================================================================
--- linux-2.6.orig/kernel/perf_counter.c
+++ linux-2.6/kernel/perf_counter.c
@@ -1288,6 +1288,11 @@ static unsigned int perf_poll(struct fil
 	return events;
 }
 
+static void perf_counter_reset(struct perf_counter *counter)
+{
+	atomic_set(&counter->count, 0);
+}
+
 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 {
 	struct perf_counter *counter = file->private_data;
@@ -1303,6 +1308,9 @@ static long perf_ioctl(struct file *file
 	case PERF_COUNTER_IOC_REFRESH:
 		perf_counter_refresh(counter, arg);
 		break;
+	case PERF_COUNTER_IOC_RESET:
+		perf_counter_reset(counter);
+		break;
 	default:
 		err = -ENOTTY;
 	}

-- 

Re: [PATCH 3/7] perf_counter: ioctl(PERF_COUNTER_IOC_RESET)

[Corey Ashford]

Hi Peter,

Peter Zijlstra wrote:

> +static void perf_counter_reset(struct perf_counter *counter)
> +{
> +	atomic_set(&counter->count, 0);
> +}
> +


Thanks for posting a patch for this issue.

As Ingo said, I think the hardware counter needs to be reset as well as 
the value saved in the perf_counter struct.


Regards,

- Corey

Corey Ashford
Software Engineer
IBM Linux Technology Center, Linux Toolchain
cjashfor@us.ibm.com

Re: [PATCH 3/7] perf_counter: ioctl(PERF_COUNTER_IOC_RESET)

[Peter Zijlstra]

On Tue, 2009-05-05 at 11:31 -0700, Corey Ashford wrote:
> Hi Peter,
> 
> Peter Zijlstra wrote:
> 
> > +static void perf_counter_reset(struct perf_counter *counter)
> > +{
> > +	atomic_set(&counter->count, 0);
> > +}
> > +
> 
> 
> Thanks for posting a patch for this issue.
> 
> As Ingo said, I think the hardware counter needs to be reset as well as 
> the value saved in the perf_counter struct.
> 

I don't think that's needed, we calculate a delta between prev_count and
the current read and use that to increment counter->count. Therefore
when we reset counter->count we should not need to touch the hardware
counter.

However, I do think we need the below, first read the hardware counter
to ensure that delta spoken of above is as close to zero as possible
when we reset.

And update the user-page bits.

---
Index: linux-2.6/kernel/perf_counter.c
===================================================================
--- linux-2.6.orig/kernel/perf_counter.c
+++ linux-2.6/kernel/perf_counter.c
@@ -1299,7 +1299,9 @@ static unsigned int perf_poll(struct fil
 
 static void perf_counter_reset(struct perf_counter *counter)
 {
+	(void)perf_counter_read(counter);
 	atomic_set(&counter->count, 0);
+	perf_counter_update_userpage(counter);
 }
 
 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)

Re: [PATCH 3/7] perf_counter: ioctl(PERF_COUNTER_IOC_RESET)

[Ingo Molnar]

* Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:

> On Tue, 2009-05-05 at 11:31 -0700, Corey Ashford wrote:
> > Hi Peter,
> > 
> > Peter Zijlstra wrote:
> > 
> > > +static void perf_counter_reset(struct perf_counter *counter)
> > > +{
> > > +	atomic_set(&counter->count, 0);
> > > +}
> > > +
> > 
> > 
> > Thanks for posting a patch for this issue.
> > 
> > As Ingo said, I think the hardware counter needs to be reset as well as 
> > the value saved in the perf_counter struct.
> > 
> 
> I don't think that's needed, we calculate a delta between prev_count and
> the current read and use that to increment counter->count. Therefore
> when we reset counter->count we should not need to touch the hardware
> counter.
> 
> However, I do think we need the below, first read the hardware counter
> to ensure that delta spoken of above is as close to zero as possible
> when we reset.
> 
> And update the user-page bits.
> 
> ---
> Index: linux-2.6/kernel/perf_counter.c
> ===================================================================
> --- linux-2.6.orig/kernel/perf_counter.c
> +++ linux-2.6/kernel/perf_counter.c
> @@ -1299,7 +1299,9 @@ static unsigned int perf_poll(struct fil
>  
>  static void perf_counter_reset(struct perf_counter *counter)
>  {
> +	(void)perf_counter_read(counter);
>  	atomic_set(&counter->count, 0);
> +	perf_counter_update_userpage(counter);
>  }

Mind sending a changelogged patch? Thanks,

	Ingo

[tip:perfcounters/core] perf_counter: add ioctl(PERF_COUNTER_IOC_RESET)

[tip-bot for Peter Zijlstra]

Commit-ID:  6de6a7b95705b859b61430fa3afa1403034eb3e6
Gitweb:     http://git.kernel.org/tip/6de6a7b95705b859b61430fa3afa1403034eb3e6
Author:     Peter Zijlstra <a.p.zijlstra@chello.nl>
AuthorDate: Tue, 5 May 2009 17:50:23 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Tue, 5 May 2009 20:18:31 +0200

perf_counter: add ioctl(PERF_COUNTER_IOC_RESET)

Provide a way to reset an existing counter - this eases PAPI
libraries around perfcounters.

Similar to read() it doesn't collapse pending child counters.

[ Impact: new perfcounter fd ioctl method to reset counters ]

Suggested-by: Corey Ashford <cjashfor@linux.vnet.ibm.com>
Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
LKML-Reference: <20090505155437.022272933@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 include/linux/perf_counter.h |    1 +
 kernel/perf_counter.c        |    8 ++++++++
 2 files changed, 9 insertions(+), 0 deletions(-)

diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index 17b6310..0fcbf34 100644
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -160,6 +160,7 @@ struct perf_counter_hw_event {
 #define PERF_COUNTER_IOC_ENABLE		_IO ('$', 0)
 #define PERF_COUNTER_IOC_DISABLE	_IO ('$', 1)
 #define PERF_COUNTER_IOC_REFRESH	_IOW('$', 2, u32)
+#define PERF_COUNTER_IOC_RESET		_IO ('$', 3)
 
 /*
  * Structure of the page that can be mapped via mmap
diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index ba5e921..6e6834e 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -1288,6 +1288,11 @@ static unsigned int perf_poll(struct file *file, poll_table *wait)
 	return events;
 }
 
+static void perf_counter_reset(struct perf_counter *counter)
+{
+	atomic_set(&counter->count, 0);
+}
+
 static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 {
 	struct perf_counter *counter = file->private_data;
@@ -1303,6 +1308,9 @@ static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 	case PERF_COUNTER_IOC_REFRESH:
 		perf_counter_refresh(counter, arg);
 		break;
+	case PERF_COUNTER_IOC_RESET:
+		perf_counter_reset(counter);
+		break;
 	default:
 		err = -ENOTTY;
 	}

[PATCH 4/7] perf_counter: provide an mlock threshold

[Peter Zijlstra]

[-- Attachment #1: perf_counter-mlock-sysctl.patch --]
[-- Type: text/plain, Size: 3498 bytes --]

Provide a threshold to relax the mlock accounting, increasing usability.

Each counter gets perf_counter_mlock_kb for free.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 include/linux/perf_counter.h |    2 ++
 kernel/perf_counter.c        |   15 +++++++++++----
 kernel/sysctl.c              |    8 ++++++++
 3 files changed, 21 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
@@ -358,6 +358,7 @@ struct file;
 struct perf_mmap_data {
 	struct rcu_head			rcu_head;
 	int				nr_pages;	/* nr of data pages  */
+	int				nr_locked;	/* nr pages mlocked  */
 
 	atomic_t			poll;		/* POLL_ for wakeups */
 	atomic_t			head;		/* write position    */
@@ -575,6 +576,7 @@ struct perf_callchain_entry {
 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 
 extern int sysctl_perf_counter_priv;
+extern int sysctl_perf_counter_mlock;
 
 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
@@ -44,6 +44,7 @@ static atomic_t nr_munmap_tracking __rea
 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 = 128; /* 'free' kb per counter */
 
 /*
  * Lock for (sysadmin-configurable) counter reservations:
@@ -1461,7 +1462,7 @@ static void perf_mmap_close(struct vm_ar
 
 	if (atomic_dec_and_mutex_lock(&counter->mmap_count,
 				      &counter->mmap_mutex)) {
-		vma->vm_mm->locked_vm -= counter->data->nr_pages + 1;
+		vma->vm_mm->locked_vm -= counter->data->nr_locked;
 		perf_mmap_data_free(counter);
 		mutex_unlock(&counter->mmap_mutex);
 	}
@@ -1480,6 +1481,7 @@ static int perf_mmap(struct file *file, 
 	unsigned long nr_pages;
 	unsigned long locked, lock_limit;
 	int ret = 0;
+	long extra;
 
 	if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
 		return -EINVAL;
@@ -1507,8 +1509,12 @@ static int perf_mmap(struct file *file, 
 		goto unlock;
 	}
 
-	locked = vma->vm_mm->locked_vm;
-	locked += nr_pages + 1;
+	extra = nr_pages /* + 1 only account the data pages */;
+	extra -= sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
+	if (extra < 0)
+		extra = 0;
+
+	locked = vma->vm_mm->locked_vm + extra;
 
 	lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 	lock_limit >>= PAGE_SHIFT;
@@ -1524,7 +1530,8 @@ static int perf_mmap(struct file *file, 
 		goto unlock;
 
 	atomic_set(&counter->mmap_count, 1);
-	vma->vm_mm->locked_vm += nr_pages + 1;
+	vma->vm_mm->locked_vm += extra;
+	counter->data->nr_locked = extra;
 unlock:
 	mutex_unlock(&counter->mmap_mutex);
 
Index: linux-2.6/kernel/sysctl.c
===================================================================
--- linux-2.6.orig/kernel/sysctl.c
+++ linux-2.6/kernel/sysctl.c
@@ -924,6 +924,14 @@ static struct ctl_table kern_table[] = {
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec,
 	},
+	{
+		.ctl_name	= CTL_UNNUMBERED,
+		.procname	= "perf_counter_mlock_kb",
+		.data		= &sysctl_perf_counter_mlock,
+		.maxlen		= sizeof(sysctl_perf_counter_mlock),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+	},
 #endif
 /*
  * NOTE: do not add new entries to this table unless you have read

-- 

[tip:perfcounters/core] perf_counter: provide an mlock threshold

[tip-bot for Peter Zijlstra]

Commit-ID:  c5078f78b455fbf67ea71442c7e7ca8acf9ff095
Gitweb:     http://git.kernel.org/tip/c5078f78b455fbf67ea71442c7e7ca8acf9ff095
Author:     Peter Zijlstra <a.p.zijlstra@chello.nl>
AuthorDate: Tue, 5 May 2009 17:50:24 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Tue, 5 May 2009 20:18:32 +0200

perf_counter: provide an mlock threshold

Provide a threshold to relax the mlock accounting, increasing usability.

Each counter gets perf_counter_mlock_kb for free.

[ Impact: allow more mmap buffering ]

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
LKML-Reference: <20090505155437.112113632@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 include/linux/perf_counter.h |    2 ++
 kernel/perf_counter.c        |   15 +++++++++++----
 kernel/sysctl.c              |    8 ++++++++
 3 files changed, 21 insertions(+), 4 deletions(-)

diff --git a/include/linux/perf_counter.h b/include/linux/perf_counter.h
index 0fcbf34..00081d8 100644
--- a/include/linux/perf_counter.h
+++ b/include/linux/perf_counter.h
@@ -358,6 +358,7 @@ struct file;
 struct perf_mmap_data {
 	struct rcu_head			rcu_head;
 	int				nr_pages;	/* nr of data pages  */
+	int				nr_locked;	/* nr pages mlocked  */
 
 	atomic_t			poll;		/* POLL_ for wakeups */
 	atomic_t			head;		/* write position    */
@@ -575,6 +576,7 @@ struct perf_callchain_entry {
 extern struct perf_callchain_entry *perf_callchain(struct pt_regs *regs);
 
 extern int sysctl_perf_counter_priv;
+extern int sysctl_perf_counter_mlock;
 
 extern void perf_counter_init(void);
 
diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 6e6834e..2d13427 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -44,6 +44,7 @@ static atomic_t nr_munmap_tracking __read_mostly;
 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 = 128; /* 'free' kb per counter */
 
 /*
  * Lock for (sysadmin-configurable) counter reservations:
@@ -1461,7 +1462,7 @@ static void perf_mmap_close(struct vm_area_struct *vma)
 
 	if (atomic_dec_and_mutex_lock(&counter->mmap_count,
 				      &counter->mmap_mutex)) {
-		vma->vm_mm->locked_vm -= counter->data->nr_pages + 1;
+		vma->vm_mm->locked_vm -= counter->data->nr_locked;
 		perf_mmap_data_free(counter);
 		mutex_unlock(&counter->mmap_mutex);
 	}
@@ -1480,6 +1481,7 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma)
 	unsigned long nr_pages;
 	unsigned long locked, lock_limit;
 	int ret = 0;
+	long extra;
 
 	if (!(vma->vm_flags & VM_SHARED) || (vma->vm_flags & VM_WRITE))
 		return -EINVAL;
@@ -1507,8 +1509,12 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma)
 		goto unlock;
 	}
 
-	locked = vma->vm_mm->locked_vm;
-	locked += nr_pages + 1;
+	extra = nr_pages /* + 1 only account the data pages */;
+	extra -= sysctl_perf_counter_mlock >> (PAGE_SHIFT - 10);
+	if (extra < 0)
+		extra = 0;
+
+	locked = vma->vm_mm->locked_vm + extra;
 
 	lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
 	lock_limit >>= PAGE_SHIFT;
@@ -1524,7 +1530,8 @@ static int perf_mmap(struct file *file, struct vm_area_struct *vma)
 		goto unlock;
 
 	atomic_set(&counter->mmap_count, 1);
-	vma->vm_mm->locked_vm += nr_pages + 1;
+	vma->vm_mm->locked_vm += extra;
+	counter->data->nr_locked = extra;
 unlock:
 	mutex_unlock(&counter->mmap_mutex);
 
diff --git a/kernel/sysctl.c b/kernel/sysctl.c
index 8203d70..3b05c2b 100644
--- a/kernel/sysctl.c
+++ b/kernel/sysctl.c
@@ -920,6 +920,14 @@ static struct ctl_table kern_table[] = {
 		.mode		= 0644,
 		.proc_handler	= &proc_dointvec,
 	},
+	{
+		.ctl_name	= CTL_UNNUMBERED,
+		.procname	= "perf_counter_mlock_kb",
+		.data		= &sysctl_perf_counter_mlock,
+		.maxlen		= sizeof(sysctl_perf_counter_mlock),
+		.mode		= 0644,
+		.proc_handler	= &proc_dointvec,
+	},
 #endif
 /*
  * NOTE: do not add new entries to this table unless you have read

[PATCH 5/7] perf_counter: fix the output lock

[Peter Zijlstra]

[-- Attachment #1: perf_counter-fix-output-lock.patch --]
[-- Type: text/plain, Size: 1348 bytes --]

Use -1 instead of 0 as unlocked, since 0 is a valid cpu number.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 kernel/perf_counter.c |   39 ++++++++++++++++++++++++++++++++++-----
 1 file changed, 34 insertions(+), 5 deletions(-)

Index: linux-2.6/kernel/perf_counter.c
===================================================================
--- linux-2.6.orig/kernel/perf_counter.c
+++ linux-2.6/kernel/perf_counter.c
@@ -1409,6 +1417,7 @@ static int perf_mmap_data_alloc(struct p
 	}
 
 	data->nr_pages = nr_pages;
+	atomic_set(&data->lock, -1);
 
 	rcu_assign_pointer(counter->data, data);
 
@@ -1755,7 +1764,7 @@ static void perf_output_lock(struct perf
 	if (in_nmi() && atomic_read(&data->lock) == cpu)
 		return;
 
-	while (atomic_cmpxchg(&data->lock, 0, cpu) != 0)
+	while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
 		cpu_relax();
 
 	handle->locked = 1;
@@ -1784,7 +1793,7 @@ again:
 	 * NMI can happen here, which means we can miss a done_head update.
 	 */
 
-	cpu = atomic_xchg(&data->lock, 0);
+	cpu = atomic_xchg(&data->lock, -1);
 	WARN_ON_ONCE(cpu != smp_processor_id());
 
 	/*
@@ -1794,7 +1803,7 @@ again:
 		/*
 		 * Since we had it locked, we can lock it again.
 		 */
-		while (atomic_cmpxchg(&data->lock, 0, cpu) != 0)
+		while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
 			cpu_relax();
 
 		goto again;

-- 

[tip:perfcounters/core] perf_counter: fix the output lock

[tip-bot for Peter Zijlstra]

Commit-ID:  22c1558e51c210787c6cf75d8905246fc91ec030
Gitweb:     http://git.kernel.org/tip/22c1558e51c210787c6cf75d8905246fc91ec030
Author:     Peter Zijlstra <a.p.zijlstra@chello.nl>
AuthorDate: Tue, 5 May 2009 17:50:25 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Tue, 5 May 2009 20:18:32 +0200

perf_counter: fix the output lock

Use -1 instead of 0 as unlocked, since 0 is a valid cpu number.

( This is not an issue right now but will be once we allow multiple
  counters to output to the same mmap area. )

[ Impact: prepare code for multi-counter profile output ]

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
LKML-Reference: <20090505155437.232686598@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 kernel/perf_counter.c |    7 ++++---
 1 files changed, 4 insertions(+), 3 deletions(-)

diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index 2d13427..c881afe 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -1409,6 +1409,7 @@ static int perf_mmap_data_alloc(struct perf_counter *counter, int nr_pages)
 	}
 
 	data->nr_pages = nr_pages;
+	atomic_set(&data->lock, -1);
 
 	rcu_assign_pointer(counter->data, data);
 
@@ -1755,7 +1756,7 @@ static void perf_output_lock(struct perf_output_handle *handle)
 	if (in_nmi() && atomic_read(&data->lock) == cpu)
 		return;
 
-	while (atomic_cmpxchg(&data->lock, 0, cpu) != 0)
+	while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
 		cpu_relax();
 
 	handle->locked = 1;
@@ -1784,7 +1785,7 @@ again:
 	 * NMI can happen here, which means we can miss a done_head update.
 	 */
 
-	cpu = atomic_xchg(&data->lock, 0);
+	cpu = atomic_xchg(&data->lock, -1);
 	WARN_ON_ONCE(cpu != smp_processor_id());
 
 	/*
@@ -1794,7 +1795,7 @@ again:
 		/*
 		 * Since we had it locked, we can lock it again.
 		 */
-		while (atomic_cmpxchg(&data->lock, 0, cpu) != 0)
+		while (atomic_cmpxchg(&data->lock, -1, cpu) != -1)
 			cpu_relax();
 
 		goto again;

[PATCH 6/7] perf_counter: inheritable sample counters

[Peter Zijlstra]

[-- Attachment #1: perf_counter-inherit.patch --]
[-- Type: text/plain, Size: 2544 bytes --]

Redirect the output to the parent counter and put in some sanity checks.

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 kernel/perf_counter.c |   32 ++++++++++++++++++++++++++++++--
 1 file changed, 30 insertions(+), 2 deletions(-)

Index: linux-2.6/kernel/perf_counter.c
===================================================================
--- linux-2.6.orig/kernel/perf_counter.c
+++ linux-2.6/kernel/perf_counter.c
@@ -738,10 +738,18 @@ static void perf_counter_enable(struct p
 	spin_unlock_irq(&ctx->lock);
 }
 
-static void perf_counter_refresh(struct perf_counter *counter, int refresh)
+static int perf_counter_refresh(struct perf_counter *counter, int refresh)
 {
+	/*
+	 * not supported on inherited counters
+	 */
+	if (counter->hw_event.inherit)
+		return -EINVAL;
+
 	atomic_add(refresh, &counter->event_limit);
 	perf_counter_enable(counter);
+
+	return 0;
 }
 
 /*
@@ -1307,7 +1315,7 @@ static long perf_ioctl(struct file *file
 		perf_counter_disable_family(counter);
 		break;
 	case PERF_COUNTER_IOC_REFRESH:
-		perf_counter_refresh(counter, arg);
+		err = perf_counter_refresh(counter, arg);
 		break;
 	case PERF_COUNTER_IOC_RESET:
 		perf_counter_reset(counter);
@@ -1814,6 +1822,12 @@ static int perf_output_begin(struct perf
 	struct perf_mmap_data *data;
 	unsigned int offset, head;
 
+	/*
+	 * For inherited counters we send all the output towards the parent.
+	 */
+	if (counter->parent)
+		counter = counter->parent;
+
 	rcu_read_lock();
 	data = rcu_dereference(counter->data);
 	if (!data)
@@ -1995,6 +2009,9 @@ static void perf_counter_output(struct p
 	if (record_type & PERF_RECORD_ADDR)
 		perf_output_put(&handle, addr);
 
+	/*
+	 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
+	 */
 	if (record_type & PERF_RECORD_GROUP) {
 		struct perf_counter *leader, *sub;
 		u64 nr = counter->nr_siblings;
@@ -2281,6 +2298,11 @@ int perf_counter_overflow(struct perf_co
 	int events = atomic_read(&counter->event_limit);
 	int ret = 0;
 
+	/*
+	 * XXX event_limit might not quite work as expected on inherited
+	 * counters
+	 */
+
 	counter->pending_kill = POLL_IN;
 	if (events && atomic_dec_and_test(&counter->event_limit)) {
 		ret = 1;
@@ -2801,6 +2823,12 @@ perf_counter_alloc(struct perf_counter_h
 
 	pmu = NULL;
 
+	/*
+	 * we currently do not support PERF_RECORD_GROUP on inherited counters
+	 */
+	if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
+		goto done;
+
 	if (perf_event_raw(hw_event)) {
 		pmu = hw_perf_counter_init(counter);
 		goto done;

-- 

[tip:perfcounters/core] perf_counter: inheritable sample counters

[tip-bot for Peter Zijlstra]

Commit-ID:  2023b359214bbc5bad31571cf50d7fb83b535c0a
Gitweb:     http://git.kernel.org/tip/2023b359214bbc5bad31571cf50d7fb83b535c0a
Author:     Peter Zijlstra <a.p.zijlstra@chello.nl>
AuthorDate: Tue, 5 May 2009 17:50:26 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Tue, 5 May 2009 20:18:33 +0200

perf_counter: inheritable sample counters

Redirect the output to the parent counter and put in some sanity checks.

[ Impact: new perfcounter feature - inherited sampling counters ]

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
LKML-Reference: <20090505155437.331556171@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 kernel/perf_counter.c |   32 ++++++++++++++++++++++++++++++--
 1 files changed, 30 insertions(+), 2 deletions(-)

diff --git a/kernel/perf_counter.c b/kernel/perf_counter.c
index c881afe..60e55f0 100644
--- a/kernel/perf_counter.c
+++ b/kernel/perf_counter.c
@@ -738,10 +738,18 @@ static void perf_counter_enable(struct perf_counter *counter)
 	spin_unlock_irq(&ctx->lock);
 }
 
-static void perf_counter_refresh(struct perf_counter *counter, int refresh)
+static int perf_counter_refresh(struct perf_counter *counter, int refresh)
 {
+	/*
+	 * not supported on inherited counters
+	 */
+	if (counter->hw_event.inherit)
+		return -EINVAL;
+
 	atomic_add(refresh, &counter->event_limit);
 	perf_counter_enable(counter);
+
+	return 0;
 }
 
 /*
@@ -1307,7 +1315,7 @@ static long perf_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
 		perf_counter_disable_family(counter);
 		break;
 	case PERF_COUNTER_IOC_REFRESH:
-		perf_counter_refresh(counter, arg);
+		err = perf_counter_refresh(counter, arg);
 		break;
 	case PERF_COUNTER_IOC_RESET:
 		perf_counter_reset(counter);
@@ -1814,6 +1822,12 @@ static int perf_output_begin(struct perf_output_handle *handle,
 	struct perf_mmap_data *data;
 	unsigned int offset, head;
 
+	/*
+	 * For inherited counters we send all the output towards the parent.
+	 */
+	if (counter->parent)
+		counter = counter->parent;
+
 	rcu_read_lock();
 	data = rcu_dereference(counter->data);
 	if (!data)
@@ -1995,6 +2009,9 @@ static void perf_counter_output(struct perf_counter *counter,
 	if (record_type & PERF_RECORD_ADDR)
 		perf_output_put(&handle, addr);
 
+	/*
+	 * XXX PERF_RECORD_GROUP vs inherited counters seems difficult.
+	 */
 	if (record_type & PERF_RECORD_GROUP) {
 		struct perf_counter *leader, *sub;
 		u64 nr = counter->nr_siblings;
@@ -2281,6 +2298,11 @@ int perf_counter_overflow(struct perf_counter *counter,
 	int events = atomic_read(&counter->event_limit);
 	int ret = 0;
 
+	/*
+	 * XXX event_limit might not quite work as expected on inherited
+	 * counters
+	 */
+
 	counter->pending_kill = POLL_IN;
 	if (events && atomic_dec_and_test(&counter->event_limit)) {
 		ret = 1;
@@ -2801,6 +2823,12 @@ perf_counter_alloc(struct perf_counter_hw_event *hw_event,
 
 	pmu = NULL;
 
+	/*
+	 * we currently do not support PERF_RECORD_GROUP on inherited counters
+	 */
+	if (hw_event->inherit && (hw_event->record_type & PERF_RECORD_GROUP))
+		goto done;
+
 	if (perf_event_raw(hw_event)) {
 		pmu = hw_perf_counter_init(counter);
 		goto done;

[PATCH 7/7] perf_counter: tools: update the tools to support process and inherited counters

[Peter Zijlstra]

[-- Attachment #1: perf_counter-tools-update.patch --]
[-- Type: text/plain, Size: 11339 bytes --]

record:
 - per task counter
 - inherit switch
 - nmi switch

report:
 - userspace/kernel filter

stat:
 - userspace/kernel filter

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
---
 Documentation/perf_counter/builtin-record.c |  155 +++++++++++++++++-----------
 Documentation/perf_counter/builtin-stat.c   |   24 +++-
 Documentation/perf_counter/perf-report.cc   |   27 +++-
 3 files changed, 140 insertions(+), 66 deletions(-)

Index: linux-2.6/Documentation/perf_counter/builtin-record.c
===================================================================
--- linux-2.6.orig/Documentation/perf_counter/builtin-record.c
+++ linux-2.6/Documentation/perf_counter/builtin-record.c
@@ -45,7 +45,10 @@ static unsigned int		mmap_pages			= 16;
 static int			output;
 static char 			*output_name			= "output.perf";
 static int			group				= 0;
-static unsigned int		realtime_prio			=  0;
+static unsigned int		realtime_prio			= 0;
+static int			system_wide			= 0;
+static int			inherit				= 1;
+static int			nmi				= 1;
 
 const unsigned int default_count[] = {
 	1000000,
@@ -167,7 +170,7 @@ static void display_events_help(void)
 static void display_help(void)
 {
 	printf(
-	"Usage: perf-record [<options>]\n"
+	"Usage: perf-record [<options>] <cmd>\n"
 	"perf-record Options (up to %d event types can be specified at once):\n\n",
 		 MAX_COUNTERS);
 
@@ -178,12 +181,13 @@ static void display_help(void)
 	" -m pages  --mmap_pages=<pages> # number of mmap data pages\n"
 	" -o file   --output=<file>      # output file\n"
 	" -r prio   --realtime=<prio>    # use RT prio\n"
+	" -s        --system             # system wide profiling\n"
 	);
 
 	exit(0);
 }
 
-static void process_options(int argc, char *argv[])
+static void process_options(int argc, const char *argv[])
 {
 	int error = 0, counter;
 
@@ -196,9 +200,12 @@ static void process_options(int argc, ch
 			{"mmap_pages",	required_argument,	NULL, 'm'},
 			{"output",	required_argument,	NULL, 'o'},
 			{"realtime",	required_argument,	NULL, 'r'},
+			{"system",	no_argument,		NULL, 's'},
+			{"inherit",	no_argument,		NULL, 'i'},
+			{"nmi",		no_argument,		NULL, 'n'},
 			{NULL,		0,			NULL,  0 }
 		};
-		int c = getopt_long(argc, argv, "+:c:e:m:o:r:",
+		int c = getopt_long(argc, argv, "+:c:e:m:o:r:sin",
 				    long_options, &option_index);
 		if (c == -1)
 			break;
@@ -209,9 +216,16 @@ static void process_options(int argc, ch
 		case 'm': mmap_pages			=   atoi(optarg); break;
 		case 'o': output_name			= strdup(optarg); break;
 		case 'r': realtime_prio			=   atoi(optarg); break;
+		case 's': system_wide                   ^=             1; break;
+		case 'i': inherit			^=	       1; break;
+		case 'n': nmi				^=	       1; break;
 		default: error = 1; break;
 		}
 	}
+
+	if (argc - optind == 0)
+		error = 1;
+
 	if (error)
 		display_help();
 
@@ -325,18 +339,82 @@ static void mmap_read(struct mmap_data *
 
 static volatile int done = 0;
 
-static void sigchld_handler(int sig)
+static void sig_handler(int sig)
 {
-	if (sig == SIGCHLD)
-		done = 1;
+	done = 1;
 }
 
-int cmd_record(int argc, char **argv)
+static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
+static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
+
+static int nr_poll;
+static int nr_cpu;
+
+static void open_counters(int cpu)
 {
-	struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
-	struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
 	struct perf_counter_hw_event hw_event;
-	int i, counter, group_fd, nr_poll = 0;
+	int counter, group_fd;
+	int track = 1;
+	pid_t pid = -1;
+
+	if (cpu < 0)
+		pid = 0;
+
+	group_fd = -1;
+	for (counter = 0; counter < nr_counters; counter++) {
+
+		memset(&hw_event, 0, sizeof(hw_event));
+		hw_event.config		= event_id[counter];
+		hw_event.irq_period	= event_count[counter];
+		hw_event.record_type	= PERF_RECORD_IP | PERF_RECORD_TID;
+		hw_event.nmi		= nmi;
+		hw_event.mmap		= track;
+		hw_event.comm		= track;
+		hw_event.inherit	= (cpu < 0) && inherit;
+
+		track = 0; // only the first counter needs these
+
+		fd[nr_cpu][counter] =
+			sys_perf_counter_open(&hw_event, pid, cpu, group_fd, 0);
+
+		if (fd[nr_cpu][counter] < 0) {
+			int err = errno;
+			printf("kerneltop error: syscall returned with %d (%s)\n",
+					fd[nr_cpu][counter], strerror(err));
+			if (err == EPERM)
+				printf("Are you root?\n");
+			exit(-1);
+		}
+		assert(fd[nr_cpu][counter] >= 0);
+		fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
+
+		/*
+		 * First counter acts as the group leader:
+		 */
+		if (group && group_fd == -1)
+			group_fd = fd[nr_cpu][counter];
+
+		event_array[nr_poll].fd = fd[nr_cpu][counter];
+		event_array[nr_poll].events = POLLIN;
+		nr_poll++;
+
+		mmap_array[nr_cpu][counter].counter = counter;
+		mmap_array[nr_cpu][counter].prev = 0;
+		mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
+		mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
+				PROT_READ, MAP_SHARED, fd[nr_cpu][counter], 0);
+		if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
+			printf("kerneltop error: failed to mmap with %d (%s)\n",
+					errno, strerror(errno));
+			exit(-1);
+		}
+	}
+	nr_cpu++;
+}
+
+int cmd_record(int argc, const char **argv)
+{
+	int i, counter;
 	pid_t pid;
 	int ret;
 
@@ -357,54 +435,13 @@ int cmd_record(int argc, char **argv)
 	argc -= optind;
 	argv += optind;
 
-	for (i = 0; i < nr_cpus; i++) {
-		group_fd = -1;
-		for (counter = 0; counter < nr_counters; counter++) {
-
-			memset(&hw_event, 0, sizeof(hw_event));
-			hw_event.config		= event_id[counter];
-			hw_event.irq_period	= event_count[counter];
-			hw_event.record_type	= PERF_RECORD_IP | PERF_RECORD_TID;
-			hw_event.nmi		= 1;
-			hw_event.mmap		= 1;
-			hw_event.comm		= 1;
-
-			fd[i][counter] = sys_perf_counter_open(&hw_event, -1, i, group_fd, 0);
-			if (fd[i][counter] < 0) {
-				int err = errno;
-				printf("kerneltop error: syscall returned with %d (%s)\n",
-					fd[i][counter], strerror(err));
-				if (err == EPERM)
-					printf("Are you root?\n");
-				exit(-1);
-			}
-			assert(fd[i][counter] >= 0);
-			fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
-
-			/*
-			 * First counter acts as the group leader:
-			 */
-			if (group && group_fd == -1)
-				group_fd = fd[i][counter];
-
-			event_array[nr_poll].fd = fd[i][counter];
-			event_array[nr_poll].events = POLLIN;
-			nr_poll++;
-
-			mmap_array[i][counter].counter = counter;
-			mmap_array[i][counter].prev = 0;
-			mmap_array[i][counter].mask = mmap_pages*page_size - 1;
-			mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
-					PROT_READ, MAP_SHARED, fd[i][counter], 0);
-			if (mmap_array[i][counter].base == MAP_FAILED) {
-				printf("kerneltop error: failed to mmap with %d (%s)\n",
-						errno, strerror(errno));
-				exit(-1);
-			}
-		}
-	}
+	if (!system_wide)
+		open_counters(-1);
+	else for (i = 0; i < nr_cpus; i++)
+		open_counters(i);
 
-	signal(SIGCHLD, sigchld_handler);
+	signal(SIGCHLD, sig_handler);
+	signal(SIGINT, sig_handler);
 
 	pid = fork();
 	if (pid < 0)
@@ -434,7 +471,7 @@ int cmd_record(int argc, char **argv)
 	while (!done) {
 		int hits = events;
 
-		for (i = 0; i < nr_cpus; i++) {
+		for (i = 0; i < nr_cpu; i++) {
 			for (counter = 0; counter < nr_counters; counter++)
 				mmap_read(&mmap_array[i][counter]);
 		}
Index: linux-2.6/Documentation/perf_counter/builtin-stat.c
===================================================================
--- linux-2.6.orig/Documentation/perf_counter/builtin-stat.c
+++ linux-2.6/Documentation/perf_counter/builtin-stat.c
@@ -87,6 +87,9 @@
 
 #include "perf.h"
 
+#define EVENT_MASK_KERNEL		1
+#define EVENT_MASK_USER			2
+
 static int			system_wide			=  0;
 
 static int			nr_counters			=  0;
@@ -104,6 +107,7 @@ static __u64			event_id[MAX_COUNTERS]		=
 static int			default_interval = 100000;
 static int			event_count[MAX_COUNTERS];
 static int			fd[MAX_NR_CPUS][MAX_COUNTERS];
+static int			event_mask[MAX_COUNTERS];
 
 static int			tid				= -1;
 static int			profile_cpu			= -1;
@@ -258,12 +262,23 @@ static __u64 match_event_symbols(char *s
 	__u64 config, id;
 	int type;
 	unsigned int i;
+	char mask_str[4];
 
 	if (sscanf(str, "r%llx", &config) == 1)
 		return config | PERF_COUNTER_RAW_MASK;
 
-	if (sscanf(str, "%d:%llu", &type, &id) == 2)
-		return EID(type, id);
+	switch (sscanf(str, "%d:%llu:%2s", &type, &id, mask_str)) {
+		case 3:
+			if (strchr(mask_str, 'u'))
+				event_mask[nr_counters] |= EVENT_MASK_USER;
+			if (strchr(mask_str, 'k'))
+				event_mask[nr_counters] |= EVENT_MASK_KERNEL;
+		case 2:
+			return EID(type, id);
+
+		default:
+			break;
+	}
 
 	for (i = 0; i < ARRAY_SIZE(event_symbols); i++) {
 		if (!strncmp(str, event_symbols[i].symbol,
@@ -313,6 +328,11 @@ static void create_perfstat_counter(int 
 	hw_event.config		= event_id[counter];
 	hw_event.record_type	= 0;
 	hw_event.nmi		= 0;
+	hw_event.exclude_kernel = event_mask[counter] & EVENT_MASK_KERNEL;
+	hw_event.exclude_user   = event_mask[counter] & EVENT_MASK_USER;
+
+printf("exclude: %d\n", event_mask[counter]);
+
 	if (scale)
 		hw_event.read_format	= PERF_FORMAT_TOTAL_TIME_ENABLED |
 					  PERF_FORMAT_TOTAL_TIME_RUNNING;
Index: linux-2.6/Documentation/perf_counter/perf-report.cc
===================================================================
--- linux-2.6.orig/Documentation/perf_counter/perf-report.cc
+++ linux-2.6/Documentation/perf_counter/perf-report.cc
@@ -33,8 +33,13 @@
 #include <string>
 
 
+#define SHOW_KERNEL	1
+#define SHOW_USER	2
+#define SHOW_HV		4
+
 static char 		const *input_name = "output.perf";
 static int		input;
+static int		show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
 
 static unsigned long	page_size;
 static unsigned long	mmap_window = 32;
@@ -359,15 +364,21 @@ static void process_options(int argc, ch
 		/** Options for getopt */
 		static struct option long_options[] = {
 			{"input",	required_argument,	NULL, 'i'},
+			{"no-user",	no_argument,		NULL, 'u'},
+			{"no-kernel",	no_argument,		NULL, 'k'},
+			{"no-hv",	no_argument,		NULL, 'h'},
 			{NULL,		0,			NULL,  0 }
 		};
-		int c = getopt_long(argc, argv, "+:i:",
+		int c = getopt_long(argc, argv, "+:i:kuh",
 				    long_options, &option_index);
 		if (c == -1)
 			break;
 
 		switch (c) {
 		case 'i': input_name			= strdup(optarg); break;
+		case 'k': show_mask &= ~SHOW_KERNEL; break;
+		case 'u': show_mask &= ~SHOW_USER; break;
+		case 'h': show_mask &= ~SHOW_HV; break;
 		default: error = 1; break;
 		}
 	}
@@ -443,22 +454,28 @@ more:
 
 	if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
 		std::string comm, sym, level;
+		int show = 0;
 		char output[1024];
 
 		if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
+			show |= SHOW_KERNEL;
 			level = " [k] ";
 			sym = resolve_kernel_symbol(event->ip.ip);
 		} else if (event->header.misc & PERF_EVENT_MISC_USER) {
+			show |= SHOW_USER;
 			level = " [.] ";
 			sym = resolve_user_symbol(event->ip.pid, event->ip.ip);
 		} else {
+			show |= SHOW_HV;
 			level = " [H] ";
 		}
-		comm = resolve_comm(event->ip.pid);
 
-		snprintf(output, sizeof(output), "%16s %s %s",
-				comm.c_str(), level.c_str(), sym.c_str());
-		hist[output]++;
+		if (show & show_mask) {
+			comm = resolve_comm(event->ip.pid);
+			snprintf(output, sizeof(output), "%16s %s %s",
+					comm.c_str(), level.c_str(), sym.c_str());
+			hist[output]++;
+		}
 
 		total++;
 

-- 

[tip:perfcounters/core] perf_counter: tools: update the tools to support process and inherited counters

[tip-bot for Peter Zijlstra]

Commit-ID:  16c8a10932aef971292c9570eb5f60b5d4e83ed2
Gitweb:     http://git.kernel.org/tip/16c8a10932aef971292c9570eb5f60b5d4e83ed2
Author:     Peter Zijlstra <a.p.zijlstra@chello.nl>
AuthorDate: Tue, 5 May 2009 17:50:27 +0200
Committer:  Ingo Molnar <mingo@elte.hu>
CommitDate: Tue, 5 May 2009 20:18:33 +0200

perf_counter: tools: update the tools to support process and inherited counters

"perf record":
 - per task counter
 - inherit switch
 - nmi switch

"perf report":
 - userspace/kernel filter

"perf stat":
 - userspace/kernel filter

Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com>
LKML-Reference: <20090505155437.389163017@chello.nl>
Signed-off-by: Ingo Molnar <mingo@elte.hu>


---
 Documentation/perf_counter/builtin-record.c |  155 +++++++++++++++++----------
 Documentation/perf_counter/builtin-stat.c   |   24 ++++-
 Documentation/perf_counter/perf-report.cc   |   27 ++++-
 3 files changed, 140 insertions(+), 66 deletions(-)

diff --git a/Documentation/perf_counter/builtin-record.c b/Documentation/perf_counter/builtin-record.c
index ddfdcf8..5f5e6df 100644
--- a/Documentation/perf_counter/builtin-record.c
+++ b/Documentation/perf_counter/builtin-record.c
@@ -45,7 +45,10 @@ static unsigned int		mmap_pages			= 16;
 static int			output;
 static char 			*output_name			= "output.perf";
 static int			group				= 0;
-static unsigned int		realtime_prio			=  0;
+static unsigned int		realtime_prio			= 0;
+static int			system_wide			= 0;
+static int			inherit				= 1;
+static int			nmi				= 1;
 
 const unsigned int default_count[] = {
 	1000000,
@@ -167,7 +170,7 @@ static void display_events_help(void)
 static void display_help(void)
 {
 	printf(
-	"Usage: perf-record [<options>]\n"
+	"Usage: perf-record [<options>] <cmd>\n"
 	"perf-record Options (up to %d event types can be specified at once):\n\n",
 		 MAX_COUNTERS);
 
@@ -178,12 +181,13 @@ static void display_help(void)
 	" -m pages  --mmap_pages=<pages> # number of mmap data pages\n"
 	" -o file   --output=<file>      # output file\n"
 	" -r prio   --realtime=<prio>    # use RT prio\n"
+	" -s        --system             # system wide profiling\n"
 	);
 
 	exit(0);
 }
 
-static void process_options(int argc, char *argv[])
+static void process_options(int argc, const char *argv[])
 {
 	int error = 0, counter;
 
@@ -196,9 +200,12 @@ static void process_options(int argc, char *argv[])
 			{"mmap_pages",	required_argument,	NULL, 'm'},
 			{"output",	required_argument,	NULL, 'o'},
 			{"realtime",	required_argument,	NULL, 'r'},
+			{"system",	no_argument,		NULL, 's'},
+			{"inherit",	no_argument,		NULL, 'i'},
+			{"nmi",		no_argument,		NULL, 'n'},
 			{NULL,		0,			NULL,  0 }
 		};
-		int c = getopt_long(argc, argv, "+:c:e:m:o:r:",
+		int c = getopt_long(argc, argv, "+:c:e:m:o:r:sin",
 				    long_options, &option_index);
 		if (c == -1)
 			break;
@@ -209,9 +216,16 @@ static void process_options(int argc, char *argv[])
 		case 'm': mmap_pages			=   atoi(optarg); break;
 		case 'o': output_name			= strdup(optarg); break;
 		case 'r': realtime_prio			=   atoi(optarg); break;
+		case 's': system_wide                   ^=             1; break;
+		case 'i': inherit			^=	       1; break;
+		case 'n': nmi				^=	       1; break;
 		default: error = 1; break;
 		}
 	}
+
+	if (argc - optind == 0)
+		error = 1;
+
 	if (error)
 		display_help();
 
@@ -325,18 +339,82 @@ static void mmap_read(struct mmap_data *md)
 
 static volatile int done = 0;
 
-static void sigchld_handler(int sig)
+static void sig_handler(int sig)
 {
-	if (sig == SIGCHLD)
-		done = 1;
+	done = 1;
 }
 
-int cmd_record(int argc, char **argv)
+static struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
+static struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
+
+static int nr_poll;
+static int nr_cpu;
+
+static void open_counters(int cpu)
 {
-	struct pollfd event_array[MAX_NR_CPUS * MAX_COUNTERS];
-	struct mmap_data mmap_array[MAX_NR_CPUS][MAX_COUNTERS];
 	struct perf_counter_hw_event hw_event;
-	int i, counter, group_fd, nr_poll = 0;
+	int counter, group_fd;
+	int track = 1;
+	pid_t pid = -1;
+
+	if (cpu < 0)
+		pid = 0;
+
+	group_fd = -1;
+	for (counter = 0; counter < nr_counters; counter++) {
+
+		memset(&hw_event, 0, sizeof(hw_event));
+		hw_event.config		= event_id[counter];
+		hw_event.irq_period	= event_count[counter];
+		hw_event.record_type	= PERF_RECORD_IP | PERF_RECORD_TID;
+		hw_event.nmi		= nmi;
+		hw_event.mmap		= track;
+		hw_event.comm		= track;
+		hw_event.inherit	= (cpu < 0) && inherit;
+
+		track = 0; // only the first counter needs these
+
+		fd[nr_cpu][counter] =
+			sys_perf_counter_open(&hw_event, pid, cpu, group_fd, 0);
+
+		if (fd[nr_cpu][counter] < 0) {
+			int err = errno;
+			printf("kerneltop error: syscall returned with %d (%s)\n",
+					fd[nr_cpu][counter], strerror(err));
+			if (err == EPERM)
+				printf("Are you root?\n");
+			exit(-1);
+		}
+		assert(fd[nr_cpu][counter] >= 0);
+		fcntl(fd[nr_cpu][counter], F_SETFL, O_NONBLOCK);
+
+		/*
+		 * First counter acts as the group leader:
+		 */
+		if (group && group_fd == -1)
+			group_fd = fd[nr_cpu][counter];
+
+		event_array[nr_poll].fd = fd[nr_cpu][counter];
+		event_array[nr_poll].events = POLLIN;
+		nr_poll++;
+
+		mmap_array[nr_cpu][counter].counter = counter;
+		mmap_array[nr_cpu][counter].prev = 0;
+		mmap_array[nr_cpu][counter].mask = mmap_pages*page_size - 1;
+		mmap_array[nr_cpu][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
+				PROT_READ, MAP_SHARED, fd[nr_cpu][counter], 0);
+		if (mmap_array[nr_cpu][counter].base == MAP_FAILED) {
+			printf("kerneltop error: failed to mmap with %d (%s)\n",
+					errno, strerror(errno));
+			exit(-1);
+		}
+	}
+	nr_cpu++;
+}
+
+int cmd_record(int argc, const char **argv)
+{
+	int i, counter;
 	pid_t pid;
 	int ret;
 
@@ -357,54 +435,13 @@ int cmd_record(int argc, char **argv)
 	argc -= optind;
 	argv += optind;
 
-	for (i = 0; i < nr_cpus; i++) {
-		group_fd = -1;
-		for (counter = 0; counter < nr_counters; counter++) {
-
-			memset(&hw_event, 0, sizeof(hw_event));
-			hw_event.config		= event_id[counter];
-			hw_event.irq_period	= event_count[counter];
-			hw_event.record_type	= PERF_RECORD_IP | PERF_RECORD_TID;
-			hw_event.nmi		= 1;
-			hw_event.mmap		= 1;
-			hw_event.comm		= 1;
-
-			fd[i][counter] = sys_perf_counter_open(&hw_event, -1, i, group_fd, 0);
-			if (fd[i][counter] < 0) {
-				int err = errno;
-				printf("kerneltop error: syscall returned with %d (%s)\n",
-					fd[i][counter], strerror(err));
-				if (err == EPERM)
-					printf("Are you root?\n");
-				exit(-1);
-			}
-			assert(fd[i][counter] >= 0);
-			fcntl(fd[i][counter], F_SETFL, O_NONBLOCK);
-
-			/*
-			 * First counter acts as the group leader:
-			 */
-			if (group && group_fd == -1)
-				group_fd = fd[i][counter];
-
-			event_array[nr_poll].fd = fd[i][counter];
-			event_array[nr_poll].events = POLLIN;
-			nr_poll++;
-
-			mmap_array[i][counter].counter = counter;
-			mmap_array[i][counter].prev = 0;
-			mmap_array[i][counter].mask = mmap_pages*page_size - 1;
-			mmap_array[i][counter].base = mmap(NULL, (mmap_pages+1)*page_size,
-					PROT_READ, MAP_SHARED, fd[i][counter], 0);
-			if (mmap_array[i][counter].base == MAP_FAILED) {
-				printf("kerneltop error: failed to mmap with %d (%s)\n",
-						errno, strerror(errno));
-				exit(-1);
-			}
-		}
-	}
+	if (!system_wide)
+		open_counters(-1);
+	else for (i = 0; i < nr_cpus; i++)
+		open_counters(i);
 
-	signal(SIGCHLD, sigchld_handler);
+	signal(SIGCHLD, sig_handler);
+	signal(SIGINT, sig_handler);
 
 	pid = fork();
 	if (pid < 0)
@@ -434,7 +471,7 @@ int cmd_record(int argc, char **argv)
 	while (!done) {
 		int hits = events;
 
-		for (i = 0; i < nr_cpus; i++) {
+		for (i = 0; i < nr_cpu; i++) {
 			for (counter = 0; counter < nr_counters; counter++)
 				mmap_read(&mmap_array[i][counter]);
 		}
diff --git a/Documentation/perf_counter/builtin-stat.c b/Documentation/perf_counter/builtin-stat.c
index 6de38d2..e2fa117 100644
--- a/Documentation/perf_counter/builtin-stat.c
+++ b/Documentation/perf_counter/builtin-stat.c
@@ -87,6 +87,9 @@
 
 #include "perf.h"
 
+#define EVENT_MASK_KERNEL		1
+#define EVENT_MASK_USER			2
+
 static int			system_wide			=  0;
 
 static int			nr_counters			=  0;
@@ -104,6 +107,7 @@ static __u64			event_id[MAX_COUNTERS]		= {
 static int			default_interval = 100000;
 static int			event_count[MAX_COUNTERS];
 static int			fd[MAX_NR_CPUS][MAX_COUNTERS];
+static int			event_mask[MAX_COUNTERS];
 
 static int			tid				= -1;
 static int			profile_cpu			= -1;
@@ -258,12 +262,23 @@ static __u64 match_event_symbols(char *str)
 	__u64 config, id;
 	int type;
 	unsigned int i;
+	char mask_str[4];
 
 	if (sscanf(str, "r%llx", &config) == 1)
 		return config | PERF_COUNTER_RAW_MASK;
 
-	if (sscanf(str, "%d:%llu", &type, &id) == 2)
-		return EID(type, id);
+	switch (sscanf(str, "%d:%llu:%2s", &type, &id, mask_str)) {
+		case 3:
+			if (strchr(mask_str, 'u'))
+				event_mask[nr_counters] |= EVENT_MASK_USER;
+			if (strchr(mask_str, 'k'))
+				event_mask[nr_counters] |= EVENT_MASK_KERNEL;
+		case 2:
+			return EID(type, id);
+
+		default:
+			break;
+	}
 
 	for (i = 0; i < ARRAY_SIZE(event_symbols); i++) {
 		if (!strncmp(str, event_symbols[i].symbol,
@@ -313,6 +328,11 @@ static void create_perfstat_counter(int counter)
 	hw_event.config		= event_id[counter];
 	hw_event.record_type	= 0;
 	hw_event.nmi		= 0;
+	hw_event.exclude_kernel = event_mask[counter] & EVENT_MASK_KERNEL;
+	hw_event.exclude_user   = event_mask[counter] & EVENT_MASK_USER;
+
+printf("exclude: %d\n", event_mask[counter]);
+
 	if (scale)
 		hw_event.read_format	= PERF_FORMAT_TOTAL_TIME_ENABLED |
 					  PERF_FORMAT_TOTAL_TIME_RUNNING;
diff --git a/Documentation/perf_counter/perf-report.cc b/Documentation/perf_counter/perf-report.cc
index 911d7f3..8855107 100644
--- a/Documentation/perf_counter/perf-report.cc
+++ b/Documentation/perf_counter/perf-report.cc
@@ -33,8 +33,13 @@
 #include <string>
 
 
+#define SHOW_KERNEL	1
+#define SHOW_USER	2
+#define SHOW_HV		4
+
 static char 		const *input_name = "output.perf";
 static int		input;
+static int		show_mask = SHOW_KERNEL | SHOW_USER | SHOW_HV;
 
 static unsigned long	page_size;
 static unsigned long	mmap_window = 32;
@@ -359,15 +364,21 @@ static void process_options(int argc, char *argv[])
 		/** Options for getopt */
 		static struct option long_options[] = {
 			{"input",	required_argument,	NULL, 'i'},
+			{"no-user",	no_argument,		NULL, 'u'},
+			{"no-kernel",	no_argument,		NULL, 'k'},
+			{"no-hv",	no_argument,		NULL, 'h'},
 			{NULL,		0,			NULL,  0 }
 		};
-		int c = getopt_long(argc, argv, "+:i:",
+		int c = getopt_long(argc, argv, "+:i:kuh",
 				    long_options, &option_index);
 		if (c == -1)
 			break;
 
 		switch (c) {
 		case 'i': input_name			= strdup(optarg); break;
+		case 'k': show_mask &= ~SHOW_KERNEL; break;
+		case 'u': show_mask &= ~SHOW_USER; break;
+		case 'h': show_mask &= ~SHOW_HV; break;
 		default: error = 1; break;
 		}
 	}
@@ -443,22 +454,28 @@ more:
 
 	if (event->header.misc & PERF_EVENT_MISC_OVERFLOW) {
 		std::string comm, sym, level;
+		int show = 0;
 		char output[1024];
 
 		if (event->header.misc & PERF_EVENT_MISC_KERNEL) {
+			show |= SHOW_KERNEL;
 			level = " [k] ";
 			sym = resolve_kernel_symbol(event->ip.ip);
 		} else if (event->header.misc & PERF_EVENT_MISC_USER) {
+			show |= SHOW_USER;
 			level = " [.] ";
 			sym = resolve_user_symbol(event->ip.pid, event->ip.ip);
 		} else {
+			show |= SHOW_HV;
 			level = " [H] ";
 		}
-		comm = resolve_comm(event->ip.pid);
 
-		snprintf(output, sizeof(output), "%16s %s %s",
-				comm.c_str(), level.c_str(), sym.c_str());
-		hist[output]++;
+		if (show & show_mask) {
+			comm = resolve_comm(event->ip.pid);
+			snprintf(output, sizeof(output), "%16s %s %s",
+					comm.c_str(), level.c_str(), sym.c_str());
+			hist[output]++;
+		}
 
 		total++;