perf tools: unbreak perf record for arm/arm64

Message ID 1507315102-5942-1-git-send-email-mark.rutland@arm.com
State New
Headers show
Series
  • perf tools: unbreak perf record for arm/arm64
Related show

Commit Message

Mark Rutland Oct. 6, 2017, 6:38 p.m.
Currently, perf record is broken on arm/arm64 systems when the PMU is
specified explicitly as part of the event, e.g.

$ ./perf record -e armv8_cortex_a53/cpu_cycles/u true

In such cases, perf record fails to open events unless
perf_event_paranoid is set to -1, even if the PMU in question supports
mode exclusion. Further, even when perf_event_paranoid is toggled, no
samples are recorded.

This is an unintended side effect of commit:

  e3ba76deef23064f ("perf tools: Force uncore events to system wide monitoring)

... which assumes that if a PMU has an associated cpu_map, it is an
uncore PMU, and forces events for such PMUs to be system-wide.

This is not true for arm/arm64 systems, which can have heterogeneous
CPUs. To account for this, multiple CPU PMUs are exposed, each with a
"cpus" field under sysfs, which the perf tool parses into a cpu_map. ARM
PMUs do not have a "cpumask" file, and only have a "cpus" file. For the
gory details as to why, see commit:

 7e3fcffe95544010 ("perf pmu: Support alternative sysfs cpumask")

Given all of this, we can instead identify uncore PMUs by explicitly
checking for a "cpumask" file, and restore arm/arm64 PMU support back to
a working state. This patch does so, adding a new perf_pmu::is_uncore
field, and splitting the existing cpumask parsing so that it can be
reused.

Signed-off-by: Mark Rutland <mark.rutland@arm.com>

Fixes: e3ba76deef23064f ("perf tools: Force uncore events to system wide monitoring)
Cc: stable@vger.kernel.org
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Arnaldo Carvalho de Melo <acme@redhat.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: David Ahern <dsahern@gmail.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Will Deacon <will.deacon@arm.com>
---
 tools/perf/util/parse-events.c |  9 ++++---
 tools/perf/util/pmu.c          | 56 +++++++++++++++++++++++++++++++-----------
 tools/perf/util/pmu.h          |  1 +
 3 files changed, 47 insertions(+), 19 deletions(-)

-- 
1.9.1

Comments

Will Deacon Oct. 9, 2017, 10:22 a.m. | #1
On Fri, Oct 06, 2017 at 07:38:22PM +0100, Mark Rutland wrote:
> Currently, perf record is broken on arm/arm64 systems when the PMU is

> specified explicitly as part of the event, e.g.

> 

> $ ./perf record -e armv8_cortex_a53/cpu_cycles/u true

> 

> In such cases, perf record fails to open events unless

> perf_event_paranoid is set to -1, even if the PMU in question supports

> mode exclusion. Further, even when perf_event_paranoid is toggled, no

> samples are recorded.

> 

> This is an unintended side effect of commit:

> 

>   e3ba76deef23064f ("perf tools: Force uncore events to system wide monitoring)

> 

> ... which assumes that if a PMU has an associated cpu_map, it is an

> uncore PMU, and forces events for such PMUs to be system-wide.

> 

> This is not true for arm/arm64 systems, which can have heterogeneous

> CPUs. To account for this, multiple CPU PMUs are exposed, each with a

> "cpus" field under sysfs, which the perf tool parses into a cpu_map. ARM

> PMUs do not have a "cpumask" file, and only have a "cpus" file. For the

> gory details as to why, see commit:

> 

>  7e3fcffe95544010 ("perf pmu: Support alternative sysfs cpumask")

> 

> Given all of this, we can instead identify uncore PMUs by explicitly

> checking for a "cpumask" file, and restore arm/arm64 PMU support back to

> a working state. This patch does so, adding a new perf_pmu::is_uncore

> field, and splitting the existing cpumask parsing so that it can be

> reused.

> 

> Signed-off-by: Mark Rutland <mark.rutland@arm.com>

> Fixes: e3ba76deef23064f ("perf tools: Force uncore events to system wide monitoring)


It sucks that we haven't noticed this being broken for so long, but I can
confirm that this fixes the issue:

Acked-by: Will Deacon <will.deacon@arm.com>

Tested-by Will Deacon <will.deacon@arm.com>

Any chance we can get this into 4.14? You'll probably need to do some stable
backports too, since this is a bit spread out.

Will
Mark Rutland Oct. 9, 2017, 10:51 a.m. | #2
On Mon, Oct 09, 2017 at 11:22:00AM +0100, Will Deacon wrote:
> It sucks that we haven't noticed this being broken for so long, but I can

> confirm that this fixes the issue:

> 

> Acked-by: Will Deacon <will.deacon@arm.com>

> Tested-by Will Deacon <will.deacon@arm.com>


Cheers.

> Any chance we can get this into 4.14? You'll probably need to do some

> stable backports too, since this is a bit spread out.


FWIW, I'm happy to do the backport(s) to stable.

As long as the Fixes tag and stable Cc are kept, I should get pinged by
the stable maintainers if/when the patch fails to apply.

Thanks,
Mark.
Arnaldo Carvalho de Melo Oct. 9, 2017, 7 p.m. | #3
Em Mon, Oct 09, 2017 at 11:22:00AM +0100, Will Deacon escreveu:
> On Fri, Oct 06, 2017 at 07:38:22PM +0100, Mark Rutland wrote:

> > Currently, perf record is broken on arm/arm64 systems when the PMU is

> > specified explicitly as part of the event, e.g.

> > 

> > $ ./perf record -e armv8_cortex_a53/cpu_cycles/u true

> > 

> > In such cases, perf record fails to open events unless

> > perf_event_paranoid is set to -1, even if the PMU in question supports

> > mode exclusion. Further, even when perf_event_paranoid is toggled, no

> > samples are recorded.

> > 

> > This is an unintended side effect of commit:

> > 

> >   e3ba76deef23064f ("perf tools: Force uncore events to system wide monitoring)

> > 

> > ... which assumes that if a PMU has an associated cpu_map, it is an

> > uncore PMU, and forces events for such PMUs to be system-wide.

> > 

> > This is not true for arm/arm64 systems, which can have heterogeneous

> > CPUs. To account for this, multiple CPU PMUs are exposed, each with a

> > "cpus" field under sysfs, which the perf tool parses into a cpu_map. ARM

> > PMUs do not have a "cpumask" file, and only have a "cpus" file. For the

> > gory details as to why, see commit:

> > 

> >  7e3fcffe95544010 ("perf pmu: Support alternative sysfs cpumask")

> > 

> > Given all of this, we can instead identify uncore PMUs by explicitly

> > checking for a "cpumask" file, and restore arm/arm64 PMU support back to

> > a working state. This patch does so, adding a new perf_pmu::is_uncore

> > field, and splitting the existing cpumask parsing so that it can be

> > reused.

> > 

> > Signed-off-by: Mark Rutland <mark.rutland@arm.com>

> > Fixes: e3ba76deef23064f ("perf tools: Force uncore events to system wide monitoring)

> 

> It sucks that we haven't noticed this being broken for so long, but I can

> confirm that this fixes the issue:

> 

> Acked-by: Will Deacon <will.deacon@arm.com>

> Tested-by Will Deacon <will.deacon@arm.com>

> 

> Any chance we can get this into 4.14? You'll probably need to do some stable

> backports too, since this is a bit spread out.


Sure, I've added this to my perf/urgent branch, that, together with a
few other fixes is undergoing testing now.

- Arnaldo
Will Deacon Oct. 9, 2017, 7:02 p.m. | #4
On Mon, Oct 09, 2017 at 04:00:00PM -0300, Arnaldo Carvalho de Melo wrote:
> Em Mon, Oct 09, 2017 at 11:22:00AM +0100, Will Deacon escreveu:

> > On Fri, Oct 06, 2017 at 07:38:22PM +0100, Mark Rutland wrote:

> > > Currently, perf record is broken on arm/arm64 systems when the PMU is

> > > specified explicitly as part of the event, e.g.

> > > 

> > > $ ./perf record -e armv8_cortex_a53/cpu_cycles/u true

> > > 

> > > In such cases, perf record fails to open events unless

> > > perf_event_paranoid is set to -1, even if the PMU in question supports

> > > mode exclusion. Further, even when perf_event_paranoid is toggled, no

> > > samples are recorded.

> > > 

> > > This is an unintended side effect of commit:

> > > 

> > >   e3ba76deef23064f ("perf tools: Force uncore events to system wide monitoring)

> > > 

> > > ... which assumes that if a PMU has an associated cpu_map, it is an

> > > uncore PMU, and forces events for such PMUs to be system-wide.

> > > 

> > > This is not true for arm/arm64 systems, which can have heterogeneous

> > > CPUs. To account for this, multiple CPU PMUs are exposed, each with a

> > > "cpus" field under sysfs, which the perf tool parses into a cpu_map. ARM

> > > PMUs do not have a "cpumask" file, and only have a "cpus" file. For the

> > > gory details as to why, see commit:

> > > 

> > >  7e3fcffe95544010 ("perf pmu: Support alternative sysfs cpumask")

> > > 

> > > Given all of this, we can instead identify uncore PMUs by explicitly

> > > checking for a "cpumask" file, and restore arm/arm64 PMU support back to

> > > a working state. This patch does so, adding a new perf_pmu::is_uncore

> > > field, and splitting the existing cpumask parsing so that it can be

> > > reused.

> > > 

> > > Signed-off-by: Mark Rutland <mark.rutland@arm.com>

> > > Fixes: e3ba76deef23064f ("perf tools: Force uncore events to system wide monitoring)

> > 

> > It sucks that we haven't noticed this being broken for so long, but I can

> > confirm that this fixes the issue:

> > 

> > Acked-by: Will Deacon <will.deacon@arm.com>

> > Tested-by Will Deacon <will.deacon@arm.com>

> > 

> > Any chance we can get this into 4.14? You'll probably need to do some stable

> > backports too, since this is a bit spread out.

> 

> Sure, I've added this to my perf/urgent branch, that, together with a

> few other fixes is undergoing testing now.


Thanks, Arnaldo.

Will

Patch

diff --git a/tools/perf/util/parse-events.c b/tools/perf/util/parse-events.c
index f6257fb..39b1596 100644
--- a/tools/perf/util/parse-events.c
+++ b/tools/perf/util/parse-events.c
@@ -309,10 +309,11 @@  static char *get_config_name(struct list_head *head_terms)
 static struct perf_evsel *
 __add_event(struct list_head *list, int *idx,
 	    struct perf_event_attr *attr,
-	    char *name, struct cpu_map *cpus,
+	    char *name, struct perf_pmu *pmu,
 	    struct list_head *config_terms, bool auto_merge_stats)
 {
 	struct perf_evsel *evsel;
+	struct cpu_map *cpus = pmu ? pmu->cpus : NULL;
 
 	event_attr_init(attr);
 
@@ -323,7 +324,7 @@  static char *get_config_name(struct list_head *head_terms)
 	(*idx)++;
 	evsel->cpus        = cpu_map__get(cpus);
 	evsel->own_cpus    = cpu_map__get(cpus);
-	evsel->system_wide = !!cpus;
+	evsel->system_wide = pmu ? pmu->is_uncore : false;
 	evsel->auto_merge_stats = auto_merge_stats;
 
 	if (name)
@@ -1233,7 +1234,7 @@  static int __parse_events_add_pmu(struct parse_events_state *parse_state,
 
 	if (!head_config) {
 		attr.type = pmu->type;
-		evsel = __add_event(list, &parse_state->idx, &attr, NULL, pmu->cpus, NULL, auto_merge_stats);
+		evsel = __add_event(list, &parse_state->idx, &attr, NULL, pmu, NULL, auto_merge_stats);
 		return evsel ? 0 : -ENOMEM;
 	}
 
@@ -1254,7 +1255,7 @@  static int __parse_events_add_pmu(struct parse_events_state *parse_state,
 		return -EINVAL;
 
 	evsel = __add_event(list, &parse_state->idx, &attr,
-			    get_config_name(head_config), pmu->cpus,
+			    get_config_name(head_config), pmu,
 			    &config_terms, auto_merge_stats);
 	if (evsel) {
 		evsel->unit = info.unit;
diff --git a/tools/perf/util/pmu.c b/tools/perf/util/pmu.c
index ac16a9d..1c4d7b4 100644
--- a/tools/perf/util/pmu.c
+++ b/tools/perf/util/pmu.c
@@ -470,17 +470,36 @@  static void pmu_read_sysfs(void)
 	closedir(dir);
 }
 
+static struct cpu_map *__pmu_cpumask(const char *path)
+{
+	FILE *file;
+	struct cpu_map *cpus;
+
+	file = fopen(path, "r");
+	if (!file)
+		return NULL;
+
+	cpus = cpu_map__read(file);
+	fclose(file);
+	return cpus;
+}
+
+/*
+ * Uncore PMUs have a "cpumask" file under sysfs. CPU PMUs (e.g. on arm/arm64)
+ * may have a "cpus" file.
+ */
+#define CPUS_TEMPLATE_UNCORE	"%s/bus/event_source/devices/%s/cpumask"
+#define CPUS_TEMPLATE_CPU	"%s/bus/event_source/devices/%s/cpus"
+
 static struct cpu_map *pmu_cpumask(const char *name)
 {
-	struct stat st;
 	char path[PATH_MAX];
-	FILE *file;
 	struct cpu_map *cpus;
 	const char *sysfs = sysfs__mountpoint();
 	const char *templates[] = {
-		 "%s/bus/event_source/devices/%s/cpumask",
-		 "%s/bus/event_source/devices/%s/cpus",
-		 NULL
+		CPUS_TEMPLATE_UNCORE,
+		CPUS_TEMPLATE_CPU,
+		NULL
 	};
 	const char **template;
 
@@ -489,20 +508,25 @@  static struct cpu_map *pmu_cpumask(const char *name)
 
 	for (template = templates; *template; template++) {
 		snprintf(path, PATH_MAX, *template, sysfs, name);
-		if (stat(path, &st) == 0)
-			break;
+		cpus = __pmu_cpumask(path);
+		if (cpus)
+			return cpus;
 	}
 
-	if (!*template)
-		return NULL;
+	return NULL;
+}
 
-	file = fopen(path, "r");
-	if (!file)
-		return NULL;
+static bool pmu_is_uncore(const char *name)
+{
+	char path[PATH_MAX];
+	struct cpu_map *cpus;
+	const char *sysfs = sysfs__mountpoint();
 
-	cpus = cpu_map__read(file);
-	fclose(file);
-	return cpus;
+	snprintf(path, PATH_MAX, CPUS_TEMPLATE_UNCORE, sysfs, name);
+	cpus = __pmu_cpumask(path);
+	cpu_map__put(cpus);
+
+	return !!cpus;
 }
 
 /*
@@ -617,6 +641,8 @@  static struct perf_pmu *pmu_lookup(const char *name)
 
 	pmu->cpus = pmu_cpumask(name);
 
+	pmu->is_uncore = pmu_is_uncore(name);
+
 	INIT_LIST_HEAD(&pmu->format);
 	INIT_LIST_HEAD(&pmu->aliases);
 	list_splice(&format, &pmu->format);
diff --git a/tools/perf/util/pmu.h b/tools/perf/util/pmu.h
index 389e972..fe0de05 100644
--- a/tools/perf/util/pmu.h
+++ b/tools/perf/util/pmu.h
@@ -22,6 +22,7 @@  struct perf_pmu {
 	char *name;
 	__u32 type;
 	bool selectable;
+	bool is_uncore;
 	struct perf_event_attr *default_config;
 	struct cpu_map *cpus;
 	struct list_head format;  /* HEAD struct perf_pmu_format -> list */