Login | Register For Free | Help
Search for: (Advanced)

Mailing List Archive: Linux: Kernel

[PATCH 04/10] perf, x86: Making hardware events translations available in sysfs

 

 

Linux kernel RSS feed   Index | Next | Previous | View Threaded


jolsa at redhat

Jul 3, 2012, 3:00 PM

Post #1 of 10 (104 views)
Permalink
[PATCH 04/10] perf, x86: Making hardware events translations available in sysfs

Making hardware events translations available through the sysfs.
Adding 'events' group attribute under the sysfs x86 PMU record
with attribute/file for each hardware event:

# ls /sys/devices/cpu/events/
branch_instructions
branch_misses
bus_cycles
cache_misses
cache_references
cycles
instructions
ref_cycles
stalled_cycles_backend
stalled_cycles_frontend

The file - hw event ID mappings is:

file hw event ID
---------------------------------------------------------------
cycles PERF_COUNT_HW_CPU_CYCLES
instructions PERF_COUNT_HW_INSTRUCTIONS
cache_references PERF_COUNT_HW_CACHE_REFERENCES
cache_misses PERF_COUNT_HW_CACHE_MISSES
branch_instructions PERF_COUNT_HW_BRANCH_INSTRUCTIONS
branch_misses PERF_COUNT_HW_BRANCH_MISSES
bus_cycles PERF_COUNT_HW_BUS_CYCLES
stalled_cycles_frontend PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
stalled_cycles_backend PERF_COUNT_HW_STALLED_CYCLES_BACKEND
ref_cycles PERF_COUNT_HW_REF_CPU_CYCLES

Each file in 'events' directory contains term translation for the
symbolic hw event for the currently running cpu model.

# cat /sys/devices/cpu/events/instructions
config=0xc0

Signed-off-by: Jiri Olsa <jolsa [at] redhat>
---
arch/x86/kernel/cpu/perf_event.c | 44 ++++++++++++++++++++++++++++++++++++++
1 file changed, 44 insertions(+)

diff --git a/arch/x86/kernel/cpu/perf_event.c b/arch/x86/kernel/cpu/perf_event.c
index 41db515..e7aab56 100644
--- a/arch/x86/kernel/cpu/perf_event.c
+++ b/arch/x86/kernel/cpu/perf_event.c
@@ -1666,9 +1666,53 @@ static struct attribute_group x86_pmu_attr_group = {
.attrs = x86_pmu_attrs,
};

+#define PMU_EVENTS_ATTR_CONFIG(_name, _id) \
+static ssize_t \
+_name##_show(struct device *dev, \
+ struct device_attribute *attr, \
+ char *page) \
+{ \
+ u64 val = x86_pmu.event_map(_id); \
+ BUILD_BUG_ON(_id >= PERF_COUNT_HW_MAX); \
+ return sprintf(page, "config=0x%llx\n", val); \
+} \
+ \
+static struct device_attribute event_attr_##_name = __ATTR_RO(_name)
+
+PMU_EVENTS_ATTR_CONFIG(cycles, PERF_COUNT_HW_CPU_CYCLES);
+PMU_EVENTS_ATTR_CONFIG(instructions, PERF_COUNT_HW_INSTRUCTIONS);
+PMU_EVENTS_ATTR_CONFIG(cache_references, PERF_COUNT_HW_CACHE_REFERENCES);
+PMU_EVENTS_ATTR_CONFIG(cache_misses, PERF_COUNT_HW_CACHE_MISSES);
+PMU_EVENTS_ATTR_CONFIG(branch_instructions, PERF_COUNT_HW_BRANCH_INSTRUCTIONS);
+PMU_EVENTS_ATTR_CONFIG(branch_misses, PERF_COUNT_HW_BRANCH_MISSES);
+PMU_EVENTS_ATTR_CONFIG(bus_cycles, PERF_COUNT_HW_BUS_CYCLES);
+PMU_EVENTS_ATTR_CONFIG(stalled_cycles_frontend, PERF_COUNT_HW_STALLED_CYCLES_FRONTEND);
+PMU_EVENTS_ATTR_CONFIG(stalled_cycles_backend, PERF_COUNT_HW_STALLED_CYCLES_BACKEND);
+PMU_EVENTS_ATTR_CONFIG(ref_cycles, PERF_COUNT_HW_REF_CPU_CYCLES);
+
+static struct attribute *events_attr[] = {
+ &event_attr_cycles.attr,
+ &event_attr_instructions.attr,
+ &event_attr_cache_references.attr,
+ &event_attr_cache_misses.attr,
+ &event_attr_branch_instructions.attr,
+ &event_attr_branch_misses.attr,
+ &event_attr_bus_cycles.attr,
+ &event_attr_stalled_cycles_frontend.attr,
+ &event_attr_stalled_cycles_backend.attr,
+ &event_attr_ref_cycles.attr,
+ NULL,
+};
+
+static struct attribute_group x86_pmu_events_group = {
+ .name = "events",
+ .attrs = events_attr,
+};
+
static const struct attribute_group *x86_pmu_attr_groups[] = {
&x86_pmu_attr_group,
&x86_pmu_format_group,
+ &x86_pmu_events_group,
NULL,
};

--
1.7.10.4

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo [at] vger
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/


a.p.zijlstra at chello

Jul 4, 2012, 3:22 AM

Post #2 of 10 (106 views)
Permalink
Re: [PATCH 04/10] perf, x86: Making hardware events translations available in sysfs [In reply to]

On Wed, 2012-07-04 at 00:00 +0200, Jiri Olsa wrote:
> +#define PMU_EVENTS_ATTR_CONFIG(_name, _id) \
> +static ssize_t \
> +_name##_show(struct device *dev, \
> + struct device_attribute *attr, \
> + char *page) \
> +{ \
> + u64 val = x86_pmu.event_map(_id); \
> + BUILD_BUG_ON(_id >= PERF_COUNT_HW_MAX); \
> + return sprintf(page, "config=0x%llx\n", val); \
> +} \

How about something like:

u64 config = x86_pmu.event_map(_id);
u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
(config & AMD64_EVENTSEL_EVENT) >> 24;
u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;

WARN_ON_ONCE(config & (ARCH_PERFMON_EVENTSEL_INV |
ARCH_PERFMON_EVENTSEL_CMASK));

sprintf("event=0x%02llx,umask=0x%02llx\n", event, umask);




--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo [at] vger
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/


a.p.zijlstra at chello

Jul 4, 2012, 3:22 AM

Post #3 of 10 (110 views)
Permalink
Re: [PATCH 04/10] perf, x86: Making hardware events translations available in sysfs [In reply to]

On Wed, 2012-07-04 at 00:00 +0200, Jiri Olsa wrote:
> +static struct device_attribute event_attr_##_name = __ATTR_RO(_name)

Note that it would now be trivial to allow writing to this file to
update the kernel table.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo [at] vger
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/


a.p.zijlstra at chello

Jul 4, 2012, 3:24 AM

Post #4 of 10 (104 views)
Permalink
Re: [PATCH 04/10] perf, x86: Making hardware events translations available in sysfs [In reply to]

On Wed, 2012-07-04 at 00:00 +0200, Jiri Olsa wrote:
> +static struct attribute *events_attr[] = {
> + &event_attr_cycles.attr,
> + &event_attr_instructions.attr,
> + &event_attr_cache_references.attr,
> + &event_attr_cache_misses.attr,
> + &event_attr_branch_instructions.attr,
> + &event_attr_branch_misses.attr,
> + &event_attr_bus_cycles.attr,
> + &event_attr_stalled_cycles_frontend.attr,
> + &event_attr_stalled_cycles_backend.attr,
> + &event_attr_ref_cycles.attr,
> + NULL,
> +};

Hmm, should we do:

if (!config)
return -EINVAL;

or somesuch to clearly indicate an event isn't supported?

Its currently a bit of a mixed bag between 0 and -1.. we might want to
clean that up too.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo [at] vger
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/


peterz at infradead

Jul 4, 2012, 3:28 AM

Post #5 of 10 (104 views)
Permalink
Re: [PATCH 04/10] perf, x86: Making hardware events translations available in sysfs [In reply to]

On Wed, 2012-07-04 at 12:24 +0200, Peter Zijlstra wrote:
> On Wed, 2012-07-04 at 00:00 +0200, Jiri Olsa wrote:
> > +static struct attribute *events_attr[] = {
> > + &event_attr_cycles.attr,
> > + &event_attr_instructions.attr,
> > + &event_attr_cache_references.attr,
> > + &event_attr_cache_misses.attr,
> > + &event_attr_branch_instructions.attr,
> > + &event_attr_branch_misses.attr,
> > + &event_attr_bus_cycles.attr,
> > + &event_attr_stalled_cycles_frontend.attr,
> > + &event_attr_stalled_cycles_backend.attr,
> > + &event_attr_ref_cycles.attr,
> > + NULL,
> > +};
>
> Hmm, should we do:
>
> if (!config)
> return -EINVAL;
>
> or somesuch to clearly indicate an event isn't supported?
>
> Its currently a bit of a mixed bag between 0 and -1.. we might want to
> clean that up too.

Alternatively, we'd do something like:

for (i = 0; events_attr[i]; i++) {
if (x86_pmu.event_map(i))
continue;

for (j = i; events_attr[j]; j++)
events_attr[j] = events_attr[j+1];
}

On init to filter out all unset events so they don't even show up in
sysfs.


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo [at] vger
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/


peterz at infradead

Jul 4, 2012, 3:38 AM

Post #6 of 10 (105 views)
Permalink
Re: [PATCH 04/10] perf, x86: Making hardware events translations available in sysfs [In reply to]

On Wed, 2012-07-04 at 12:22 +0200, Peter Zijlstra wrote:
> How about something like:
>
> u64 config = x86_pmu.event_map(_id);
> u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
> (config & AMD64_EVENTSEL_EVENT) >> 24;
> u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
>
> WARN_ON_ONCE(config & (ARCH_PERFMON_EVENTSEL_INV |
> ARCH_PERFMON_EVENTSEL_CMASK));
>
> sprintf("event=0x%02llx,umask=0x%02llx\n", event, umask);

Oh wait, silly me, we actually need the inv and cmask too.

So that gets to be something like:

u64 config = x86_pmu.event_map(_id);
u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
(config & AMD64_EVENTSEL_EVENT) >> 24;
u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
u64 inv = (config & ARCH_PERFMON_EVENTSEL_INV) >> 23;
u64 cmask = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
ssize_t ret;

ret = sprintf(page, "event=0x%02llx", event);
if (umask)
ret += sprintf(page + ret, ",umask=0x%02llx", umask);
if (inv)
ret += sprintf(page + ret, ",inv");
if (cmask)
ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
sprintf(page + ret, "\n");

return ret;

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo [at] vger
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/


jolsa at redhat

Jul 4, 2012, 5:00 AM

Post #7 of 10 (97 views)
Permalink
Re: [PATCH 04/10] perf, x86: Making hardware events translations available in sysfs [In reply to]

On Wed, Jul 04, 2012 at 12:28:58PM +0200, Peter Zijlstra wrote:
> On Wed, 2012-07-04 at 12:24 +0200, Peter Zijlstra wrote:
> > On Wed, 2012-07-04 at 00:00 +0200, Jiri Olsa wrote:
> > > +static struct attribute *events_attr[] = {
> > > + &event_attr_cycles.attr,
> > > + &event_attr_instructions.attr,
> > > + &event_attr_cache_references.attr,
> > > + &event_attr_cache_misses.attr,
> > > + &event_attr_branch_instructions.attr,
> > > + &event_attr_branch_misses.attr,
> > > + &event_attr_bus_cycles.attr,
> > > + &event_attr_stalled_cycles_frontend.attr,
> > > + &event_attr_stalled_cycles_backend.attr,
> > > + &event_attr_ref_cycles.attr,
> > > + NULL,
> > > +};
> >
> > Hmm, should we do:
> >
> > if (!config)
> > return -EINVAL;
> >
> > or somesuch to clearly indicate an event isn't supported?
> >
> > Its currently a bit of a mixed bag between 0 and -1.. we might want to
> > clean that up too.
>
> Alternatively, we'd do something like:
>
> for (i = 0; events_attr[i]; i++) {
> if (x86_pmu.event_map(i))
> continue;
>
> for (j = i; events_attr[j]; j++)
> events_attr[j] = events_attr[j+1];
> }
>
> On init to filter out all unset events so they don't even show up in
> sysfs.
>
>

ok
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo [at] vger
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/


jolsa at redhat

Jul 4, 2012, 5:01 AM

Post #8 of 10 (100 views)
Permalink
Re: [PATCH 04/10] perf, x86: Making hardware events translations available in sysfs [In reply to]

On Wed, Jul 04, 2012 at 12:38:31PM +0200, Peter Zijlstra wrote:
> On Wed, 2012-07-04 at 12:22 +0200, Peter Zijlstra wrote:
> > How about something like:
> >
> > u64 config = x86_pmu.event_map(_id);
> > u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
> > (config & AMD64_EVENTSEL_EVENT) >> 24;
> > u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
> >
> > WARN_ON_ONCE(config & (ARCH_PERFMON_EVENTSEL_INV |
> > ARCH_PERFMON_EVENTSEL_CMASK));
> >
> > sprintf("event=0x%02llx,umask=0x%02llx\n", event, umask);
>
> Oh wait, silly me, we actually need the inv and cmask too.
>
> So that gets to be something like:
>
> u64 config = x86_pmu.event_map(_id);
> u64 event = (config & ARCH_PERFMON_EVENTSEL_EVENT) |
> (config & AMD64_EVENTSEL_EVENT) >> 24;
> u64 umask = (config & ARCH_PERFMON_EVENTSEL_UMASK) >> 8;
> u64 inv = (config & ARCH_PERFMON_EVENTSEL_INV) >> 23;
> u64 cmask = (config & ARCH_PERFMON_EVENTSEL_CMASK) >> 24;
> ssize_t ret;
>
> ret = sprintf(page, "event=0x%02llx", event);
> if (umask)
> ret += sprintf(page + ret, ",umask=0x%02llx", umask);
> if (inv)
> ret += sprintf(page + ret, ",inv");
> if (cmask)
> ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
> sprintf(page + ret, "\n");
>
> return ret;
>

ok
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo [at] vger
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/


peterz at infradead

Jul 4, 2012, 5:14 AM

Post #9 of 10 (97 views)
Permalink
Re: [PATCH 04/10] perf, x86: Making hardware events translations available in sysfs [In reply to]

On Wed, 2012-07-04 at 14:01 +0200, Jiri Olsa wrote:
> > ret = sprintf(page, "event=0x%02llx", event);
> > if (umask)
> > ret += sprintf(page + ret, ",umask=0x%02llx", umask);
> > if (inv)
> > ret += sprintf(page + ret, ",inv");
> > if (cmask)
> > ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
> > sprintf(page + ret, "\n");

There's a ret += missing there.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo [at] vger
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/


acme at redhat

Jul 4, 2012, 9:35 AM

Post #10 of 10 (98 views)
Permalink
Re: [PATCH 04/10] perf, x86: Making hardware events translations available in sysfs [In reply to]

Em Wed, Jul 04, 2012 at 02:14:48PM +0200, Peter Zijlstra escreveu:
> On Wed, 2012-07-04 at 14:01 +0200, Jiri Olsa wrote:
> > > ret = sprintf(page, "event=0x%02llx", event);
> > > if (umask)
> > > ret += sprintf(page + ret, ",umask=0x%02llx", umask);
> > > if (inv)
> > > ret += sprintf(page + ret, ",inv");
> > > if (cmask)
> > > ret += sprintf(page + ret, ",cmask=0x%02llx", cmask);
> > > sprintf(page + ret, "\n");
>
> There's a ret += missing there.

And also don't use sprintf or snprintf, use scnprintf, the former have
confusing semantics for the return value, see the man pages.

- Arnaldo
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo [at] vger
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Linux kernel RSS feed   Index | Next | Previous | View Threaded
 
 


Interested in having your list archived? Contact Gossamer Threads
 
  Web Applications & Managed Hosting Powered by Gossamer Threads Inc.