[PATCH] perf stat: fix case where guest/host monitoring is not supported by kernel

[PATCH] perf stat: fix case where guest/host monitoring is not supported by kernel

[Stephane Eranian]

By default, perf stat sets exclude_guest = 1. But when you run perf on a kernel
which does not support  host/guest filtering, then you get an error saying
the event in unsupported. This comes from the fact that when the
perf_event_attr struct passed by the user is larger than the one known to 
the kernel there is safety check which ensures that all unknown bits are
zero. But here, exclude_guest is 1 (part of the unknown bits) and thus the
perf_event_open() syscall return EINVAL.

To my surprise, running perf record on the same kernel did not exhibit
the problem. The reason is that perf record handles the problem by
catching the error and retrying with guest/host excludes set to zero.
For some reason, this was not done with perf stat. This patch fixes
this problem.

Signed-off-by: Stephane Eranian <eranian@google.com>
---
diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index dde9e17..6e1c1f9 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -283,6 +283,8 @@ static int create_perf_stat_counter(struct perf_evsel *evsel,
 {
 	struct perf_event_attr *attr = &evsel->attr;
 	struct xyarray *group_fd = NULL;
+	bool exclude_guest_missing = false;
+	int ret;
 
 	if (group && evsel != first)
 		group_fd = first->fd;
@@ -293,16 +295,39 @@ 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,
+retry:
+	if (exclude_guest_missing)
+		evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
+
+	if (system_wide) {
+		ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus,
 						group, group_fd);
+		if (ret)
+			goto check_ret;
+		return 0;
+	}
+
 	if (!target_pid && !target_tid && (!group || evsel == first)) {
 		attr->disabled = 1;
 		attr->enable_on_exec = 1;
 	}
 
-	return perf_evsel__open_per_thread(evsel, evsel_list->threads,
-					   group, group_fd);
+	ret = perf_evsel__open_per_thread(evsel, evsel_list->threads,
+					  group, group_fd);
+	if (!ret)
+		return 0;
+	/* fall through */
+check_ret:
+	if (ret && errno == EINVAL) {
+		if (!exclude_guest_missing &&
+		    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
+			pr_debug("Old kernel, cannot exclude "
+				 "guest or host samples.\n");
+			exclude_guest_missing = true;
+			goto retry;
+		}
+	}
+	return ret;
 }
 
 /*

Re: [PATCH] perf stat: fix case where guest/host monitoring is not supported by kernel

[Arnaldo Carvalho de Melo]

Em Fri, Apr 27, 2012 at 02:45:38PM +0200, Stephane Eranian escreveu:
> 
> By default, perf stat sets exclude_guest = 1. But when you run perf on a kernel
> which does not support  host/guest filtering, then you get an error saying
> the event in unsupported. This comes from the fact that when the
> perf_event_attr struct passed by the user is larger than the one known to 
> the kernel there is safety check which ensures that all unknown bits are
> zero. But here, exclude_guest is 1 (part of the unknown bits) and thus the
> perf_event_open() syscall return EINVAL.
> 
> To my surprise, running perf record on the same kernel did not exhibit
> the problem. The reason is that perf record handles the problem by
> catching the error and retrying with guest/host excludes set to zero.
> For some reason, this was not done with perf stat. This patch fixes
> this problem.

yeah, I have to sit down and finish perf_evlist__open, i.e. get all this
capability queries.

Thanks, applying to my perf/urgent branch.

- Arnaldo
 
> Signed-off-by: Stephane Eranian <eranian@google.com>
> ---
> diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
> index dde9e17..6e1c1f9 100644
> --- a/tools/perf/builtin-stat.c
> +++ b/tools/perf/builtin-stat.c
> @@ -283,6 +283,8 @@ static int create_perf_stat_counter(struct perf_evsel *evsel,
>  {
>  	struct perf_event_attr *attr = &evsel->attr;
>  	struct xyarray *group_fd = NULL;
> +	bool exclude_guest_missing = false;
> +	int ret;
>  
>  	if (group && evsel != first)
>  		group_fd = first->fd;
> @@ -293,16 +295,39 @@ 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,
> +retry:
> +	if (exclude_guest_missing)
> +		evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
> +
> +	if (system_wide) {
> +		ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus,
>  						group, group_fd);
> +		if (ret)
> +			goto check_ret;
> +		return 0;
> +	}
> +
>  	if (!target_pid && !target_tid && (!group || evsel == first)) {
>  		attr->disabled = 1;
>  		attr->enable_on_exec = 1;
>  	}
>  
> -	return perf_evsel__open_per_thread(evsel, evsel_list->threads,
> -					   group, group_fd);
> +	ret = perf_evsel__open_per_thread(evsel, evsel_list->threads,
> +					  group, group_fd);
> +	if (!ret)
> +		return 0;
> +	/* fall through */
> +check_ret:
> +	if (ret && errno == EINVAL) {
> +		if (!exclude_guest_missing &&
> +		    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
> +			pr_debug("Old kernel, cannot exclude "
> +				 "guest or host samples.\n");
> +			exclude_guest_missing = true;
> +			goto retry;
> +		}
> +	}
> +	return ret;
>  }
>  
>  /*

[tip:perf/urgent] perf stat: Fix case where guest/ host monitoring is not supported by kernel

[tip-bot for Stephane Eranian]

Commit-ID:  5622c07b4741e0afd7607bce6e850b76eeb23210
Gitweb:     http://git.kernel.org/tip/5622c07b4741e0afd7607bce6e850b76eeb23210
Author:     Stephane Eranian <eranian@google.com>
AuthorDate: Fri, 27 Apr 2012 14:45:38 +0200
Committer:  Arnaldo Carvalho de Melo <acme@redhat.com>
CommitDate: Tue, 1 May 2012 14:20:00 -0300

perf stat: Fix case where guest/host monitoring is not supported by kernel

By default, perf stat sets exclude_guest = 1. But when you run perf on a
kernel which does not support  host/guest filtering, then you get an
error saying the event in unsupported. This comes from the fact that
when the perf_event_attr struct passed by the user is larger than the
one known to the kernel there is safety check which ensures that all
unknown bits are zero. But here, exclude_guest is 1 (part of the unknown
bits) and thus the perf_event_open() syscall return EINVAL.

To my surprise, running perf record on the same kernel did not exhibit
the problem. The reason is that perf record handles the problem by
catching the error and retrying with guest/host excludes set to zero.
For some reason, this was not done with perf stat. This patch fixes this
problem.

Signed-off-by: Stephane Eranian <eranian@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Joerg Roedel <joro@8bytes.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Robert Richter <robert.richter@amd.com>
Link: http://lkml.kernel.org/r/20120427124538.GA7230@quad
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
---
 tools/perf/builtin-stat.c |   33 +++++++++++++++++++++++++++++----
 1 files changed, 29 insertions(+), 4 deletions(-)

diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
index c941bb6..4532a78 100644
--- a/tools/perf/builtin-stat.c
+++ b/tools/perf/builtin-stat.c
@@ -283,6 +283,8 @@ static int create_perf_stat_counter(struct perf_evsel *evsel,
 {
 	struct perf_event_attr *attr = &evsel->attr;
 	struct xyarray *group_fd = NULL;
+	bool exclude_guest_missing = false;
+	int ret;
 
 	if (group && evsel != first)
 		group_fd = first->fd;
@@ -293,16 +295,39 @@ 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,
+retry:
+	if (exclude_guest_missing)
+		evsel->attr.exclude_guest = evsel->attr.exclude_host = 0;
+
+	if (system_wide) {
+		ret = perf_evsel__open_per_cpu(evsel, evsel_list->cpus,
 						group, group_fd);
+		if (ret)
+			goto check_ret;
+		return 0;
+	}
+
 	if (!target_pid && !target_tid && (!group || evsel == first)) {
 		attr->disabled = 1;
 		attr->enable_on_exec = 1;
 	}
 
-	return perf_evsel__open_per_thread(evsel, evsel_list->threads,
-					   group, group_fd);
+	ret = perf_evsel__open_per_thread(evsel, evsel_list->threads,
+					  group, group_fd);
+	if (!ret)
+		return 0;
+	/* fall through */
+check_ret:
+	if (ret && errno == EINVAL) {
+		if (!exclude_guest_missing &&
+		    (evsel->attr.exclude_guest || evsel->attr.exclude_host)) {
+			pr_debug("Old kernel, cannot exclude "
+				 "guest or host samples.\n");
+			exclude_guest_missing = true;
+			goto retry;
+		}
+	}
+	return ret;
 }
 
 /*