[RFC PATCH 0/2] tools/perf: fix the event grouping functionality

[RFC PATCH 0/2] tools/perf: fix the event grouping functionality

[Deng-Cheng Zhu]

Two patches relating to the event state and group_fd of grouped events at
creation.

Deng-Cheng Zhu (2):
  tools/perf: Don't set attr->disabled when doing grouping
  tools/perf: Make group_fd static and move its place in
    __perf_evsel__open()

 tools/perf/builtin-record.c |    6 +++---
 tools/perf/builtin-stat.c   |    3 ++-
 tools/perf/util/evsel.c     |    3 +--
 3 files changed, 6 insertions(+), 6 deletions(-)

[RFC PATCH 1/2] tools/perf: Don't set attr->disabled when doing grouping

[Deng-Cheng Zhu]

Events will be created at the state PERF_EVENT_STATE_OFF if attr->disabled
is set. When these events go to arch level validate_group(), the function
won't do anything real because they are filtered out by the state check.

Signed-off-by: Deng-Cheng Zhu <dczhu@mips.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
---
 tools/perf/builtin-record.c |    6 +++---
 tools/perf/builtin-stat.c   |    3 ++-
 2 files changed, 5 insertions(+), 4 deletions(-)

diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index f4c3fbe..67e8e4c 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -161,7 +161,6 @@ static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
 	struct perf_event_attr *attr = &evsel->attr;
 	int track = !evsel->idx; /* only the first counter needs these */
 
-	attr->disabled		= 1;
 	attr->inherit		= !no_inherit;
 	attr->read_format	= PERF_FORMAT_TOTAL_TIME_ENABLED |
 				  PERF_FORMAT_TOTAL_TIME_RUNNING |
@@ -222,10 +221,11 @@ static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
 	attr->mmap		= track;
 	attr->comm		= track;
 
-	if (target_pid == -1 && target_tid == -1 && !system_wide) {
+	if (!group)
 		attr->disabled = 1;
+
+	if (target_pid == -1 && target_tid == -1 && !system_wide)
 		attr->enable_on_exec = 1;
-	}
 }
 
 static bool perf_evlist__equal(struct perf_evlist *evlist,
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index 5deb17d..ca95475 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -284,7 +284,8 @@ static int create_perf_stat_counter(struct perf_evsel *evsel)
 		return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, group);
 
 	if (target_pid == -1 && target_tid == -1) {
-		attr->disabled = 1;
+		if (!group)
+			attr->disabled = 1;
 		attr->enable_on_exec = 1;
 	}
 
-- 
1.7.1

[RFC PATCH 2/2] tools/perf: Make group_fd static and move its place in __perf_evsel__open()

[Deng-Cheng Zhu]

__perf_evsel__open() is called per event, it does not work for all the
grouped events at one time. So, currently group_fd will alway be -1 for
the events in a group. This patch fixes it.

Signed-off-by: Deng-Cheng Zhu <dczhu@mips.com>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
---
 tools/perf/util/evsel.c |    3 +--
 1 files changed, 1 insertions(+), 2 deletions(-)

diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index e389815..7bd0d9d 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -219,9 +219,8 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 	}
 
 	for (cpu = 0; cpu < cpus->nr; cpu++) {
-		int group_fd = -1;
-
 		for (thread = 0; thread < threads->nr; thread++) {
+			static int group_fd = -1;
 
 			if (!evsel->cgrp)
 				pid = threads->map[thread];
-- 
1.7.1

Re: [RFC PATCH 2/2] tools/perf: Make group_fd static and move its place in __perf_evsel__open()

[Arnaldo Carvalho de Melo]

Em Mon, Oct 24, 2011 at 06:57:00PM +0800, Deng-Cheng Zhu escreveu:
> __perf_evsel__open() is called per event, it does not work for all the
> grouped events at one time. So, currently group_fd will alway be -1 for
> the events in a group. This patch fixes it.
> 
> Signed-off-by: Deng-Cheng Zhu <dczhu@mips.com>
> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
> Cc: Paul Mackerras <paulus@samba.org>
> Cc: Ingo Molnar <mingo@elte.hu>
> Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
> ---
>  tools/perf/util/evsel.c |    3 +--
>  1 files changed, 1 insertions(+), 2 deletions(-)
> 
> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> index e389815..7bd0d9d 100644
> --- a/tools/perf/util/evsel.c
> +++ b/tools/perf/util/evsel.c
> @@ -219,9 +219,8 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
>  	}
>  
>  	for (cpu = 0; cpu < cpus->nr; cpu++) {
> -		int group_fd = -1;
> -
>  		for (thread = 0; thread < threads->nr; thread++) {
> +			static int group_fd = -1;
>  
>  			if (!evsel->cgrp)
>  				pid = threads->map[thread];

Lets not do it that way, using statics for this is humm, ugly, IMHO.

Just pass an integer pointer that is a member of perf_evlist, I'll work
on a patch now.

Thanks for reporting the bug tho! 

- Arnaldo

Re: [RFC PATCH 1/2] tools/perf: Don't set attr->disabled when doing grouping

[David Ahern]

On 10/24/2011 04:56 AM, Deng-Cheng Zhu wrote:
> Events will be created at the state PERF_EVENT_STATE_OFF if attr->disabled
> is set. When these events go to arch level validate_group(), the function
> won't do anything real because they are filtered out by the state check.

Seems like a bug with the group checks. You should be able to enable and
disable events - including creating them disabled.

David


> 
> Signed-off-by: Deng-Cheng Zhu <dczhu@mips.com>
> Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
> Cc: Paul Mackerras <paulus@samba.org>
> Cc: Ingo Molnar <mingo@elte.hu>
> Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
> ---
>  tools/perf/builtin-record.c |    6 +++---
>  tools/perf/builtin-stat.c   |    3 ++-
>  2 files changed, 5 insertions(+), 4 deletions(-)
> 
> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
> index f4c3fbe..67e8e4c 100644
> --- a/tools/perf/builtin-record.c
> +++ b/tools/perf/builtin-record.c
> @@ -161,7 +161,6 @@ static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
>  	struct perf_event_attr *attr = &evsel->attr;
>  	int track = !evsel->idx; /* only the first counter needs these */
>  
> -	attr->disabled		= 1;
>  	attr->inherit		= !no_inherit;
>  	attr->read_format	= PERF_FORMAT_TOTAL_TIME_ENABLED |
>  				  PERF_FORMAT_TOTAL_TIME_RUNNING |
> @@ -222,10 +221,11 @@ static void config_attr(struct perf_evsel *evsel, struct perf_evlist *evlist)
>  	attr->mmap		= track;
>  	attr->comm		= track;
>  
> -	if (target_pid == -1 && target_tid == -1 && !system_wide) {
> +	if (!group)
>  		attr->disabled = 1;
> +
> +	if (target_pid == -1 && target_tid == -1 && !system_wide)
>  		attr->enable_on_exec = 1;
> -	}
>  }
>  
>  static bool perf_evlist__equal(struct perf_evlist *evlist,
> diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
> index 5deb17d..ca95475 100644
> --- a/tools/perf/builtin-stat.c
> +++ b/tools/perf/builtin-stat.c
> @@ -284,7 +284,8 @@ static int create_perf_stat_counter(struct perf_evsel *evsel)
>  		return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, group);
>  
>  	if (target_pid == -1 && target_tid == -1) {
> -		attr->disabled = 1;
> +		if (!group)
> +			attr->disabled = 1;
>  		attr->enable_on_exec = 1;
>  	}
>  

Re: [RFC PATCH 1/2] tools/perf: Don't set attr->disabled when doing grouping

[Peter Zijlstra]

On Mon, 2011-10-24 at 08:19 -0600, David Ahern wrote:
> 
> Seems like a bug with the group checks. You should be able to enable and
> disable events - including creating them disabled.

Agreed, it makes no sense to allow creating a group you can't use.

Re: [RFC PATCH 2/2] tools/perf: Make group_fd static and move its place in __perf_evsel__open()

[Arnaldo Carvalho de Melo]

Em Mon, Oct 24, 2011 at 12:03:28PM -0200, Arnaldo Carvalho de Melo escreveu:
> Em Mon, Oct 24, 2011 at 06:57:00PM +0800, Deng-Cheng Zhu escreveu:
> > __perf_evsel__open() is called per event, it does not work for all the
> > grouped events at one time. So, currently group_fd will alway be -1 for
> > the events in a group. This patch fixes it.
> > 
> > Signed-off-by: Deng-Cheng Zhu <dczhu@mips.com>
> > Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
> > Cc: Paul Mackerras <paulus@samba.org>
> > Cc: Ingo Molnar <mingo@elte.hu>
> > Cc: Arnaldo Carvalho de Melo <acme@ghostprotocols.net>
> > ---
> >  tools/perf/util/evsel.c |    3 +--
> >  1 files changed, 1 insertions(+), 2 deletions(-)
> > 
> > diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
> > index e389815..7bd0d9d 100644
> > --- a/tools/perf/util/evsel.c
> > +++ b/tools/perf/util/evsel.c
> > @@ -219,9 +219,8 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
> >  	}
> >  
> >  	for (cpu = 0; cpu < cpus->nr; cpu++) {
> > -		int group_fd = -1;
> > -
> >  		for (thread = 0; thread < threads->nr; thread++) {
> > +			static int group_fd = -1;
> >  
> >  			if (!evsel->cgrp)
> >  				pid = threads->map[thread];
> 
> Lets not do it that way, using statics for this is humm, ugly, IMHO.
> 
> Just pass an integer pointer that is a member of perf_evlist, I'll work
> on a patch now.
> 
> Thanks for reporting the bug tho! 

Can you try this patch?

I tested it with:

[root@emilia ~]# perf top -e cycles -e instructions --group

and

[root@emilia ~]# perf record -e cycles -e instructions --group -a sleep 2

Peter,

	Do you see anything wrong with it?

Thanks,

- Arnaldo

>From fa5e5376207a3e08e0893c3a15321d8972e51fd1 Mon Sep 17 00:00:00 2001
From: Arnaldo Carvalho de Melo <acme@redhat.com>
Date: Tue, 25 Oct 2011 10:42:19 -0200
Subject: [PATCH] perf evlist: Fix grouping of multiple events

The __perf_evsel__open routing was grouping just the threads for that
specific events per cpu when we want to group all threads in all events
to the first fd opened on that cpu.

So pass the xyarray with the first event, where the other events will be
able to get that first per cpu fd.

At some point top and record will switch to using perf_evlist__open that
takes care of this detail and probably will also handle the fallback
from hw to soft counters, etc.

Reported-by: Deng-Cheng Zhu <dczhu@mips.com>
Cc: David Ahern <dsahern@gmail.com>
Cc: Frederic Weisbecker <fweisbec@gmail.com>
Cc: Mike Galbraith <efault@gmx.de>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Stephane Eranian <eranian@google.com>
Link: http://lkml.kernel.org/n/tip-ebm34rh098i9y9v4cytfdp0x@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 tools/perf/builtin-record.c |   11 +++++++++--
 tools/perf/builtin-stat.c   |   20 ++++++++++++++------
 tools/perf/builtin-test.c   |    6 +++---
 tools/perf/builtin-top.c    |   11 +++++++++--
 tools/perf/util/evlist.c    |   30 ++++++++++++++++++++++++++++++
 tools/perf/util/evlist.h    |    2 ++
 tools/perf/util/evsel.c     |   43 +++++++++++++++++++++++++++++++------------
 tools/perf/util/evsel.h     |   10 +++++++---
 tools/perf/util/python.c    |   31 ++++++++++++++++++++++++++++++-
 9 files changed, 135 insertions(+), 29 deletions(-)

diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
index f82480f..40088a5 100644
--- a/tools/perf/builtin-record.c
+++ b/tools/perf/builtin-record.c
@@ -262,13 +262,16 @@ static bool perf_evlist__equal(struct perf_evlist *evlist,
 
 static void open_counters(struct perf_evlist *evlist)
 {
-	struct perf_evsel *pos;
+	struct perf_evsel *pos, *first;
 
 	if (evlist->cpus->map[0] < 0)
 		no_inherit = true;
 
+	first = list_entry(evlist->entries.next, struct perf_evsel, node);
+
 	list_for_each_entry(pos, &evlist->entries, node) {
 		struct perf_event_attr *attr = &pos->attr;
+		struct xyarray *group_fd = NULL;
 		/*
 		 * Check if parse_single_tracepoint_event has already asked for
 		 * PERF_SAMPLE_TIME.
@@ -283,11 +286,15 @@ static void open_counters(struct perf_evlist *evlist)
 		 */
 		bool time_needed = attr->sample_type & PERF_SAMPLE_TIME;
 
+		if (group && pos != first)
+			group_fd = first->fd;
+
 		config_attr(pos, evlist);
 retry_sample_id:
 		attr->sample_id_all = sample_id_all_avail ? 1 : 0;
 try_again:
-		if (perf_evsel__open(pos, evlist->cpus, evlist->threads, group) < 0) {
+		if (perf_evsel__open(pos, evlist->cpus, evlist->threads, group,
+				     group_fd) < 0) {
 			int err = errno;
 
 			if (err == EPERM || err == EACCES) {
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index 7ce65f5..7d98676 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -278,9 +278,14 @@ struct stats			runtime_itlb_cache_stats[MAX_NR_CPUS];
 struct stats			runtime_dtlb_cache_stats[MAX_NR_CPUS];
 struct stats			walltime_nsecs_stats;
 
-static int create_perf_stat_counter(struct perf_evsel *evsel)
+static int create_perf_stat_counter(struct perf_evsel *evsel,
+				    struct perf_evsel *first)
 {
 	struct perf_event_attr *attr = &evsel->attr;
+	struct xyarray *group_fd = NULL;
+
+	if (group && evsel != first)
+		group_fd = first->fd;
 
 	if (scale)
 		attr->read_format = PERF_FORMAT_TOTAL_TIME_ENABLED |
@@ -289,14 +294,15 @@ static int create_perf_stat_counter(struct perf_evsel *evsel)
 	attr->inherit = !no_inherit;
 
 	if (system_wide)
-		return perf_evsel__open_per_cpu(evsel, evsel_list->cpus, group);
-
+		return perf_evsel__open_per_cpu(evsel, evsel_list->cpus,
+						group, group_fd);
 	if (target_pid == -1 && target_tid == -1) {
 		attr->disabled = 1;
 		attr->enable_on_exec = 1;
 	}
 
-	return perf_evsel__open_per_thread(evsel, evsel_list->threads, group);
+	return perf_evsel__open_per_thread(evsel, evsel_list->threads,
+					   group, group_fd);
 }
 
 /*
@@ -396,7 +402,7 @@ static int read_counter(struct perf_evsel *counter)
 static int run_perf_stat(int argc __used, const char **argv)
 {
 	unsigned long long t0, t1;
-	struct perf_evsel *counter;
+	struct perf_evsel *counter, *first;
 	int status = 0;
 	int child_ready_pipe[2], go_pipe[2];
 	const bool forks = (argc > 0);
@@ -453,8 +459,10 @@ static int run_perf_stat(int argc __used, const char **argv)
 		close(child_ready_pipe[0]);
 	}
 
+	first = list_entry(evsel_list->entries.next, struct perf_evsel, node);
+
 	list_for_each_entry(counter, &evsel_list->entries, node) {
-		if (create_perf_stat_counter(counter) < 0) {
+		if (create_perf_stat_counter(counter, first) < 0) {
 			if (errno == EINVAL || errno == ENOSYS || errno == ENOENT) {
 				if (verbose)
 					ui__warning("%s event is not supported by the kernel.\n",
diff --git a/tools/perf/builtin-test.c b/tools/perf/builtin-test.c
index efe696f..831d1ba 100644
--- a/tools/perf/builtin-test.c
+++ b/tools/perf/builtin-test.c
@@ -291,7 +291,7 @@ static int test__open_syscall_event(void)
 		goto out_thread_map_delete;
 	}
 
-	if (perf_evsel__open_per_thread(evsel, threads, false) < 0) {
+	if (perf_evsel__open_per_thread(evsel, threads, false, NULL) < 0) {
 		pr_debug("failed to open counter: %s, "
 			 "tweak /proc/sys/kernel/perf_event_paranoid?\n",
 			 strerror(errno));
@@ -366,7 +366,7 @@ static int test__open_syscall_event_on_all_cpus(void)
 		goto out_thread_map_delete;
 	}
 
-	if (perf_evsel__open(evsel, cpus, threads, false) < 0) {
+	if (perf_evsel__open(evsel, cpus, threads, false, NULL) < 0) {
 		pr_debug("failed to open counter: %s, "
 			 "tweak /proc/sys/kernel/perf_event_paranoid?\n",
 			 strerror(errno));
@@ -531,7 +531,7 @@ static int test__basic_mmap(void)
 
 		perf_evlist__add(evlist, evsels[i]);
 
-		if (perf_evsel__open(evsels[i], cpus, threads, false) < 0) {
+		if (perf_evsel__open(evsels[i], cpus, threads, false, NULL) < 0) {
 			pr_debug("failed to open counter: %s, "
 				 "tweak /proc/sys/kernel/perf_event_paranoid?\n",
 				 strerror(errno));
diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c
index 7a87171..d2fc754 100644
--- a/tools/perf/builtin-top.c
+++ b/tools/perf/builtin-top.c
@@ -834,10 +834,16 @@ static void perf_session__mmap_read(struct perf_session *self)
 
 static void start_counters(struct perf_evlist *evlist)
 {
-	struct perf_evsel *counter;
+	struct perf_evsel *counter, *first;
+
+	first = list_entry(evlist->entries.next, struct perf_evsel, node);
 
 	list_for_each_entry(counter, &evlist->entries, node) {
 		struct perf_event_attr *attr = &counter->attr;
+		struct xyarray *group_fd = NULL;
+
+		if (group && counter != first)
+			group_fd = first->fd;
 
 		attr->sample_type = PERF_SAMPLE_IP | PERF_SAMPLE_TID;
 
@@ -860,7 +866,8 @@ static void start_counters(struct perf_evlist *evlist)
 		attr->inherit = inherit;
 try_again:
 		if (perf_evsel__open(counter, top.evlist->cpus,
-				     top.evlist->threads, group) < 0) {
+				     top.evlist->threads, group,
+				     group_fd) < 0) {
 			int err = errno;
 
 			if (err == EPERM || err == EACCES) {
diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
index 2f6bc89..fbb4b4a 100644
--- a/tools/perf/util/evlist.c
+++ b/tools/perf/util/evlist.c
@@ -539,3 +539,33 @@ void perf_evlist__set_selected(struct perf_evlist *evlist,
 {
 	evlist->selected = evsel;
 }
+
+int perf_evlist__open(struct perf_evlist *evlist, bool group)
+{
+	struct perf_evsel *evsel, *first;
+	int err, ncpus, nthreads;
+
+	first = list_entry(evlist->entries.next, struct perf_evsel, node);
+
+	list_for_each_entry(evsel, &evlist->entries, node) {
+		struct xyarray *group_fd = NULL;
+
+		if (group && evsel != first)
+			group_fd = first->fd;
+
+		err = perf_evsel__open(evsel, evlist->cpus, evlist->threads,
+				       group, group_fd);
+		if (err < 0)
+			goto out_err;
+	}
+
+	return 0;
+out_err:
+	ncpus = evlist->cpus ? evlist->cpus->nr : 1;
+	nthreads = evlist->threads ? evlist->threads->nr : 1;
+
+	list_for_each_entry_reverse(evsel, &evlist->entries, node)
+		perf_evsel__close(evsel, ncpus, nthreads);
+
+	return err;
+}
diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h
index 6be71fc..1779ffe 100644
--- a/tools/perf/util/evlist.h
+++ b/tools/perf/util/evlist.h
@@ -50,6 +50,8 @@ struct perf_evsel *perf_evlist__id2evsel(struct perf_evlist *evlist, u64 id);
 
 union perf_event *perf_evlist__mmap_read(struct perf_evlist *self, int idx);
 
+int perf_evlist__open(struct perf_evlist *evlist, bool group);
+
 int perf_evlist__alloc_mmap(struct perf_evlist *evlist);
 int perf_evlist__mmap(struct perf_evlist *evlist, int pages, bool overwrite);
 void perf_evlist__munmap(struct perf_evlist *evlist);
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index b46f6e4..e426264 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -16,6 +16,7 @@
 #include "thread_map.h"
 
 #define FD(e, x, y) (*(int *)xyarray__entry(e->fd, x, y))
+#define GROUP_FD(group_fd, cpu) (*(int *)xyarray__entry(group_fd, cpu, 0))
 
 int __perf_evsel__sample_size(u64 sample_type)
 {
@@ -204,15 +205,16 @@ int __perf_evsel__read(struct perf_evsel *evsel,
 }
 
 static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
-			      struct thread_map *threads, bool group)
+			      struct thread_map *threads, bool group,
+			      struct xyarray *group_fds)
 {
 	int cpu, thread;
 	unsigned long flags = 0;
-	int pid = -1;
+	int pid = -1, err;
 
 	if (evsel->fd == NULL &&
 	    perf_evsel__alloc_fd(evsel, cpus->nr, threads->nr) < 0)
-		return -1;
+		return -ENOMEM;
 
 	if (evsel->cgrp) {
 		flags = PERF_FLAG_PID_CGROUP;
@@ -220,7 +222,7 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 	}
 
 	for (cpu = 0; cpu < cpus->nr; cpu++) {
-		int group_fd = -1;
+		int group_fd = group_fds ? GROUP_FD(group_fds, cpu) : -1;
 
 		for (thread = 0; thread < threads->nr; thread++) {
 
@@ -231,8 +233,10 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 								     pid,
 								     cpus->map[cpu],
 								     group_fd, flags);
-			if (FD(evsel, cpu, thread) < 0)
+			if (FD(evsel, cpu, thread) < 0) {
+				err = -errno;
 				goto out_close;
+			}
 
 			if (group && group_fd == -1)
 				group_fd = FD(evsel, cpu, thread);
@@ -249,7 +253,17 @@ out_close:
 		}
 		thread = threads->nr;
 	} while (--cpu >= 0);
-	return -1;
+	return err;
+}
+
+void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads)
+{
+	if (evsel->fd == NULL)
+		return;
+
+	perf_evsel__close_fd(evsel, ncpus, nthreads);
+	perf_evsel__free_fd(evsel);
+	evsel->fd = NULL;
 }
 
 static struct {
@@ -269,7 +283,8 @@ static struct {
 };
 
 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
-		     struct thread_map *threads, bool group)
+		     struct thread_map *threads, bool group,
+		     struct xyarray *group_fd)
 {
 	if (cpus == NULL) {
 		/* Work around old compiler warnings about strict aliasing */
@@ -279,19 +294,23 @@ int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
 	if (threads == NULL)
 		threads = &empty_thread_map.map;
 
-	return __perf_evsel__open(evsel, cpus, threads, group);
+	return __perf_evsel__open(evsel, cpus, threads, group, group_fd);
 }
 
 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
-			     struct cpu_map *cpus, bool group)
+			     struct cpu_map *cpus, bool group,
+			     struct xyarray *group_fd)
 {
-	return __perf_evsel__open(evsel, cpus, &empty_thread_map.map, group);
+	return __perf_evsel__open(evsel, cpus, &empty_thread_map.map, group,
+				  group_fd);
 }
 
 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
-				struct thread_map *threads, bool group)
+				struct thread_map *threads, bool group,
+				struct xyarray *group_fd)
 {
-	return __perf_evsel__open(evsel, &empty_cpu_map.map, threads, group);
+	return __perf_evsel__open(evsel, &empty_cpu_map.map, threads, group,
+				  group_fd);
 }
 
 static int perf_event__parse_id_sample(const union perf_event *event, u64 type,
diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
index e9a3155..b1d15e6 100644
--- a/tools/perf/util/evsel.h
+++ b/tools/perf/util/evsel.h
@@ -82,11 +82,15 @@ void perf_evsel__free_id(struct perf_evsel *evsel);
 void perf_evsel__close_fd(struct perf_evsel *evsel, int ncpus, int nthreads);
 
 int perf_evsel__open_per_cpu(struct perf_evsel *evsel,
-			     struct cpu_map *cpus, bool group);
+			     struct cpu_map *cpus, bool group,
+			     struct xyarray *group_fds);
 int perf_evsel__open_per_thread(struct perf_evsel *evsel,
-				struct thread_map *threads, bool group);
+				struct thread_map *threads, bool group,
+				struct xyarray *group_fds);
 int perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
-		     struct thread_map *threads, bool group);
+		     struct thread_map *threads, bool group,
+		     struct xyarray *group_fds);
+void perf_evsel__close(struct perf_evsel *evsel, int ncpus, int nthreads);
 
 #define perf_evsel__match(evsel, t, c)		\
 	(evsel->attr.type == PERF_TYPE_##t &&	\
diff --git a/tools/perf/util/python.c b/tools/perf/util/python.c
index 7624324..9dd47a4 100644
--- a/tools/perf/util/python.c
+++ b/tools/perf/util/python.c
@@ -623,7 +623,11 @@ static PyObject *pyrf_evsel__open(struct pyrf_evsel *pevsel,
 		cpus = ((struct pyrf_cpu_map *)pcpus)->cpus;
 
 	evsel->attr.inherit = inherit;
-	if (perf_evsel__open(evsel, cpus, threads, group) < 0) {
+	/*
+	 * This will group just the fds for this single evsel, to group
+	 * multiple events, use evlist.open().
+	 */
+	if (perf_evsel__open(evsel, cpus, threads, group, NULL) < 0) {
 		PyErr_SetFromErrno(PyExc_OSError);
 		return NULL;
 	}
@@ -814,6 +818,25 @@ static PyObject *pyrf_evlist__read_on_cpu(struct pyrf_evlist *pevlist,
 	return Py_None;
 }
 
+static PyObject *pyrf_evlist__open(struct pyrf_evlist *pevlist,
+				   PyObject *args, PyObject *kwargs)
+{
+	struct perf_evlist *evlist = &pevlist->evlist;
+	int group = 0;
+	static char *kwlist[] = { "group", NULL };
+
+	if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|OOii", kwlist, &group))
+		return NULL;
+
+	if (perf_evlist__open(evlist, group) < 0) {
+		PyErr_SetFromErrno(PyExc_OSError);
+		return NULL;
+	}
+
+	Py_INCREF(Py_None);
+	return Py_None;
+}
+
 static PyMethodDef pyrf_evlist__methods[] = {
 	{
 		.ml_name  = "mmap",
@@ -822,6 +845,12 @@ static PyMethodDef pyrf_evlist__methods[] = {
 		.ml_doc	  = PyDoc_STR("mmap the file descriptor table.")
 	},
 	{
+		.ml_name  = "open",
+		.ml_meth  = (PyCFunction)pyrf_evlist__open,
+		.ml_flags = METH_VARARGS | METH_KEYWORDS,
+		.ml_doc	  = PyDoc_STR("open the file descriptors.")
+	},
+	{
 		.ml_name  = "poll",
 		.ml_meth  = (PyCFunction)pyrf_evlist__poll,
 		.ml_flags = METH_VARARGS | METH_KEYWORDS,
-- 
1.7.1

Re: [RFC PATCH 2/2] tools/perf: Make group_fd static and move its place in __perf_evsel__open()

[Deng-Cheng Zhu]

On 10/25/2011 11:23 PM, Arnaldo Carvalho de Melo wrote:
> Em Mon, Oct 24, 2011 at 12:03:28PM -0200, Arnaldo Carvalho de Melo escreveu:
>> Em Mon, Oct 24, 2011 at 06:57:00PM +0800, Deng-Cheng Zhu escreveu:
>>> __perf_evsel__open() is called per event, it does not work for all the
>>> grouped events at one time. So, currently group_fd will alway be -1 for
>>> the events in a group. This patch fixes it.
>>>
>>> Signed-off-by: Deng-Cheng Zhu<dczhu@mips.com>
>>> Cc: Peter Zijlstra<a.p.zijlstra@chello.nl>
>>> Cc: Paul Mackerras<paulus@samba.org>
>>> Cc: Ingo Molnar<mingo@elte.hu>
>>> Cc: Arnaldo Carvalho de Melo<acme@ghostprotocols.net>
>>> ---
>>>   tools/perf/util/evsel.c |    3 +--
>>>   1 files changed, 1 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
>>> index e389815..7bd0d9d 100644
>>> --- a/tools/perf/util/evsel.c
>>> +++ b/tools/perf/util/evsel.c
>>> @@ -219,9 +219,8 @@ static int __perf_evsel__open(struct perf_evsel *evsel, struct cpu_map *cpus,
>>>   	}
>>>
>>>   	for (cpu = 0; cpu<  cpus->nr; cpu++) {
>>> -		int group_fd = -1;
>>> -
>>>   		for (thread = 0; thread<  threads->nr; thread++) {
>>> +			static int group_fd = -1;
>>>
>>>   			if (!evsel->cgrp)
>>>   				pid = threads->map[thread];
>>
>> Lets not do it that way, using statics for this is humm, ugly, IMHO.
>>
>> Just pass an integer pointer that is a member of perf_evlist, I'll work
>> on a patch now.
>>
>> Thanks for reporting the bug tho!
>
> Can you try this patch?
>
> I tested it with:
>
> [root@emilia ~]# perf top -e cycles -e instructions --group
>
> and
>
> [root@emilia ~]# perf record -e cycles -e instructions --group -a sleep 2
>

Your patch does fix the group fd issue. But to get event grouping
workable, the event state fix is still needed. Please see the discussion
here: http://www.spinics.net/lists/mips/msg42190.html

With _only_ your patch applied, I tested with the following commands on
MIPS 74K (4 counters available in total):

perf stat -g -e 
L1-dcache-load-misses,cycles,LLC-load-misses,iTLB-loads,instructions 
find / >/dev/null

I tried to group up to 5 events in the hope of seeing NOSPC error. But
the command didn't fail and output:

  Performance counter stats for 'find /':

            9300823 L1-dcache-load-misses
      <not counted> cycles
      <not counted> LLC-load-misses
      <not counted> iTLB-loads
      <not counted> instructions

        8.463207591 seconds time elapsed

This is due to the event state check in validate_group() filtering out
the grouped events in OFF state. They are in OFF state because we are
running the command with the perf tool as opposed to attaching to an
existing task:

builtin-stat.c:create_perf_stat_counter():

         if (target_pid == -1 && target_tid == -1) {
                 attr->disabled = 1;
                 attr->enable_on_exec = 1;
         }

I suppose X86 has this issue too -- collect_events() in validate_group()
won't do real work in the bottom half of the function.


Deng-Cheng

Re: [RFC PATCH 2/2] tools/perf: Make group_fd static and move its place in __perf_evsel__open()

[Arnaldo Carvalho de Melo]

Em Wed, Oct 26, 2011 at 02:26:03PM +0800, Deng-Cheng Zhu escreveu:
> On 10/25/2011 11:23 PM, Arnaldo Carvalho de Melo wrote:
> >Em Mon, Oct 24, 2011 at 12:03:28PM -0200, Arnaldo Carvalho de Melo escreveu:
> >>Em Mon, Oct 24, 2011 at 06:57:00PM +0800, Deng-Cheng Zhu escreveu:
> >>>__perf_evsel__open() is called per event, it does not work for all the
> >>>grouped events at one time. So, currently group_fd will alway be -1 for

<SNIP>

> >>>+			static int group_fd = -1;
> >>>  			if (!evsel->cgrp)
> >>>  				pid = threads->map[thread];

> >>Lets not do it that way, using statics for this is humm, ugly, IMHO.

> >Can you try this patch?

> >I tested it with:

> >[root@emilia ~]# perf top -e cycles -e instructions --group

> Your patch does fix the group fd issue. But to get event grouping
> workable, the event state fix is still needed. Please see the discussion

Can I have your "Reviewed-by:" or "Tested-by:" tag for this patch then?

> here: http://www.spinics.net/lists/mips/msg42190.html
 
> With _only_ your patch applied, I tested with the following commands on
> MIPS 74K (4 counters available in total):
> 
> perf stat -g -e
> L1-dcache-load-misses,cycles,LLC-load-misses,iTLB-loads,instructions
> find / >/dev/null
> 
> I tried to group up to 5 events in the hope of seeing NOSPC error. But
> the command didn't fail and output:
> 
>  Performance counter stats for 'find /':
> 
>            9300823 L1-dcache-load-misses
>      <not counted> cycles
>      <not counted> LLC-load-misses
>      <not counted> iTLB-loads
>      <not counted> instructions
> 
>        8.463207591 seconds time elapsed
> 
> This is due to the event state check in validate_group() filtering out
> the grouped events in OFF state. They are in OFF state because we are
> running the command with the perf tool as opposed to attaching to an
> existing task:
> 
> builtin-stat.c:create_perf_stat_counter():
> 
>         if (target_pid == -1 && target_tid == -1) {
>                 attr->disabled = 1;
>                 attr->enable_on_exec = 1;
>         }

But they should be in OFF state only till the target program gets
exec'ed, right?
 
> I suppose X86 has this issue too -- collect_events() in validate_group()
> won't do real work in the bottom half of the function.

I'm testing that now.

- Arnaldo

Re: [RFC PATCH 2/2] tools/perf: Make group_fd static and move its place in __perf_evsel__open()

[Arnaldo Carvalho de Melo]

Em Wed, Oct 26, 2011 at 09:18:28AM -0200, Arnaldo Carvalho de Melo escreveu:
> Em Wed, Oct 26, 2011 at 02:26:03PM +0800, Deng-Cheng Zhu escreveu:
> > With _only_ your patch applied, I tested with the following commands on
> > MIPS 74K (4 counters available in total):

> > perf stat -g -e
> > L1-dcache-load-misses,cycles,LLC-load-misses,iTLB-loads,instructions
> > find / >/dev/null

> > I tried to group up to 5 events in the hope of seeing NOSPC error. But
> > the command didn't fail and output:

> >  Performance counter stats for 'find /':

> >            9300823 L1-dcache-load-misses
> >      <not counted> cycles
> >      <not counted> LLC-load-misses
> >      <not counted> iTLB-loads
> >      <not counted> instructions

> >        8.463207591 seconds time elapsed

> > This is due to the event state check in validate_group() filtering out
> > the grouped events in OFF state. They are in OFF state because we are
> > running the command with the perf tool as opposed to attaching to an
> > existing task:

> > builtin-stat.c:create_perf_stat_counter():

> >         if (target_pid == -1 && target_tid == -1) {
> >                 attr->disabled = 1;
> >                 attr->enable_on_exec = 1;
> >         }
> 
> But they should be in OFF state only till the target program gets
> exec'ed, right?
>  
> > I suppose X86 has this issue too -- collect_events() in validate_group()
> > won't do real work in the bottom half of the function.
> 
> I'm testing that now.

[root@emilia ~]# perf stat -g -e L1-dcache-load-misses,cycles,LLC-load-misses,iTLB-loads,instructions find / >/dev/null

 Performance counter stats for 'find /':

       158,807,949 L1-dcache-load-misses                                       
     <not counted> cycles                  
     <not counted> LLC-load-misses         
     <not counted> iTLB-loads              
     <not counted> instructions            

     100.676645958 seconds time elapsed

[root@emilia ~]# 

Yeah, reading more code...

- Arnaldo

RE: [RFC PATCH 2/2] tools/perf: Make group_fd static and move its place in __perf_evsel__open()

[Zhu, DengCheng]

> ________________________________________
> From: Arnaldo Carvalho de Melo [arnaldo.melo@gmail.com] on behalf of Arnaldo Carvalho de Melo [acme@ghostprotocols.net]
> Sent: Wednesday, October 26, 2011 7:18 PM
> To: Zhu, DengCheng
> Cc: linux-kernel@vger.kernel.org; Peter Zijlstra; Paul Mackerras; Ingo Molnar
> Subject: Re: [RFC PATCH 2/2] tools/perf: Make group_fd static and move its place in __perf_evsel__open()
>
> Em Wed, Oct 26, 2011 at 02:26:03PM +0800, Deng-Cheng Zhu escreveu:
>> On 10/25/2011 11:23 PM, Arnaldo Carvalho de Melo wrote:
>> >Em Mon, Oct 24, 2011 at 12:03:28PM -0200, Arnaldo Carvalho de Melo escreveu:
>> >>Em Mon, Oct 24, 2011 at 06:57:00PM +0800, Deng-Cheng Zhu escreveu:
>> >>>__perf_evsel__open() is called per event, it does not work for all the
>> >>>grouped events at one time. So, currently group_fd will alway be -1 for
>
> <SNIP>
>
>> >>>+                  static int group_fd = -1;
>> >>>                   if (!evsel->cgrp)
>> >>>                           pid = threads->map[thread];
>
>> >>Lets not do it that way, using statics for this is humm, ugly, IMHO.
>
>> >Can you try this patch?
>
>> >I tested it with:
>
>> >[root@emilia ~]# perf top -e cycles -e instructions --group
>
>> Your patch does fix the group fd issue. But to get event grouping
>> workable, the event state fix is still needed. Please see the discussion
>
> Can I have your "Reviewed-by:" or "Tested-by:" tag for this patch then?

Yes, you may add my:

Tested-by: Deng-Cheng Zhu <dczhu@mips.com>

>> here: http://www.spinics.net/lists/mips/msg42190.html
>
>> With _only_ your patch applied, I tested with the following commands on
>> MIPS 74K (4 counters available in total):
>>
>> perf stat -g -e
>> L1-dcache-load-misses,cycles,LLC-load-misses,iTLB-loads,instructions
>> find / >/dev/null
>>
>> I tried to group up to 5 events in the hope of seeing NOSPC error. But
>> the command didn't fail and output:
>>
>>  Performance counter stats for 'find /':
>>
>>            9300823 L1-dcache-load-misses
>>      <not counted> cycles
>>      <not counted> LLC-load-misses
>>      <not counted> iTLB-loads
>>      <not counted> instructions
>>
>>        8.463207591 seconds time elapsed
>>
>> This is due to the event state check in validate_group() filtering out
>> the grouped events in OFF state. They are in OFF state because we are
>> running the command with the perf tool as opposed to attaching to an
>> existing task:
>>
>> builtin-stat.c:create_perf_stat_counter():
>>
>>         if (target_pid == -1 && target_tid == -1) {
>>                 attr->disabled = 1;
>>                 attr->enable_on_exec = 1;
>>         }
>
> But they should be in OFF state only till the target program gets
> exec'ed, right?
>

Yes, or else event timestamps won't be updated on exec. That's why I came
to a new idea in one of my MIPS perf-events patches (link provided above). 

>> I suppose X86 has this issue too -- collect_events() in validate_group()
>> won't do real work in the bottom half of the function.
>
> I'm testing that now.

As you can see from the test results in your another post, I think we see
"not counted" because collect_events() in validate_group() does not do real
work on events in OFF state. On x86 a patch may be needed in event init,
but on MIPS (or ARM) I suppose the pmu and event state checks in
validate_event() perhaps could be deleted...


Deng-Cheng