tools/perf/util/evlist.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-)
Dummy events are created with an attribute where the period and freq
are zero. evsel__config will then see the uninitialized values and
initialize them in evsel__default_freq_period. As fequency mode is
used by default the dummy event would be set to use frequency
mode. However, this has no effect on the dummy event but does cause
unnecessary timers/interrupts. Avoid this overhead by setting the
period to 1 for dummy events.
evlist__add_aux_dummy calls evlist__add_dummy then sets freq=0 and
period=1. This isn't necessary after this change and so the setting is
removed.
From Stephane:
The dummy event is not counting anything. It is used to collect mmap
records and avoid a race condition during the synthesize mmap phase of
perf record. As such, it should not cause any overhead during active
profiling. Yet, it did. Because of a bug the dummy event was
programmed as a sampling event in frequency mode. Events in that mode
incur more kernel overheads because on timer tick, the kernel has to
look at the number of samples for each event and potentially adjust
the sampling period to achieve the desired frequency. The dummy event
was therefore adding a frequency event to task and ctx contexts we may
otherwise not have any, e.g., perf record -a -e
cpu/event=0x3c,period=10000000/. On each timer tick the
perf_adjust_freq_unthr_context() is invoked and if ctx->nr_freq is
non-zero, then the kernel will loop over ALL the events of the context
looking for frequency mode ones. In doing, so it locks the context,
and enable/disable the PMU of each hw event. If all the events of the
context are in period mode, the kernel will have to traverse the list for
nothing incurring overhead. The overhead is multiplied by a very large
factor when this happens in a guest kernel. There is no need for the
dummy event to be in frequency mode, it does not count anything and
therefore should not cause extra overhead for no reason.
Fixes: 5bae0250237f ("perf evlist: Introduce perf_evlist__new_dummy constructor")
Reported-by: Stephane Eranian <eranian@google.com>
Signed-off-by: Ian Rogers <irogers@google.com>
---
tools/perf/util/evlist.c | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
index 25c3ebe2c2f5..e36da58522ef 100644
--- a/tools/perf/util/evlist.c
+++ b/tools/perf/util/evlist.c
@@ -251,6 +251,9 @@ static struct evsel *evlist__dummy_event(struct evlist *evlist)
.type = PERF_TYPE_SOFTWARE,
.config = PERF_COUNT_SW_DUMMY,
.size = sizeof(attr), /* to capture ABI version */
+ /* Avoid frequency mode for dummy events to avoid associated timers. */
+ .freq = 0,
+ .sample_period = 1,
};
return evsel__new_idx(&attr, evlist->core.nr_entries);
@@ -277,8 +280,6 @@ struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
evsel->core.attr.exclude_kernel = 1;
evsel->core.attr.exclude_guest = 1;
evsel->core.attr.exclude_hv = 1;
- evsel->core.attr.freq = 0;
- evsel->core.attr.sample_period = 1;
evsel->core.system_wide = system_wide;
evsel->no_aux_samples = true;
evsel->name = strdup("dummy:u");
--
2.42.0.459.ge4e396fd5e-googOn Fri, Sep 15, 2023 at 9:10 PM Ian Rogers <irogers@google.com> wrote:
>
> Dummy events are created with an attribute where the period and freq
> are zero. evsel__config will then see the uninitialized values and
> initialize them in evsel__default_freq_period. As fequency mode is
> used by default the dummy event would be set to use frequency
> mode. However, this has no effect on the dummy event but does cause
> unnecessary timers/interrupts. Avoid this overhead by setting the
> period to 1 for dummy events.
>
> evlist__add_aux_dummy calls evlist__add_dummy then sets freq=0 and
> period=1. This isn't necessary after this change and so the setting is
> removed.
>
> From Stephane:
>
> The dummy event is not counting anything. It is used to collect mmap
> records and avoid a race condition during the synthesize mmap phase of
> perf record. As such, it should not cause any overhead during active
> profiling. Yet, it did. Because of a bug the dummy event was
> programmed as a sampling event in frequency mode. Events in that mode
> incur more kernel overheads because on timer tick, the kernel has to
> look at the number of samples for each event and potentially adjust
> the sampling period to achieve the desired frequency. The dummy event
> was therefore adding a frequency event to task and ctx contexts we may
> otherwise not have any, e.g., perf record -a -e
> cpu/event=0x3c,period=10000000/. On each timer tick the
> perf_adjust_freq_unthr_context() is invoked and if ctx->nr_freq is
> non-zero, then the kernel will loop over ALL the events of the context
> looking for frequency mode ones. In doing, so it locks the context,
> and enable/disable the PMU of each hw event. If all the events of the
> context are in period mode, the kernel will have to traverse the list for
> nothing incurring overhead. The overhead is multiplied by a very large
> factor when this happens in a guest kernel. There is no need for the
> dummy event to be in frequency mode, it does not count anything and
> therefore should not cause extra overhead for no reason.
>
> Fixes: 5bae0250237f ("perf evlist: Introduce perf_evlist__new_dummy constructor")
> Reported-by: Stephane Eranian <eranian@google.com>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
> tools/perf/util/evlist.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
> index 25c3ebe2c2f5..e36da58522ef 100644
> --- a/tools/perf/util/evlist.c
> +++ b/tools/perf/util/evlist.c
> @@ -251,6 +251,9 @@ static struct evsel *evlist__dummy_event(struct evlist *evlist)
> .type = PERF_TYPE_SOFTWARE,
> .config = PERF_COUNT_SW_DUMMY,
> .size = sizeof(attr), /* to capture ABI version */
> + /* Avoid frequency mode for dummy events to avoid associated timers. */
> + .freq = 0,
> + .sample_period = 1,
> };
>
> return evsel__new_idx(&attr, evlist->core.nr_entries);
> @@ -277,8 +280,6 @@ struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
> evsel->core.attr.exclude_kernel = 1;
> evsel->core.attr.exclude_guest = 1;
> evsel->core.attr.exclude_hv = 1;
> - evsel->core.attr.freq = 0;
> - evsel->core.attr.sample_period = 1;
> evsel->core.system_wide = system_wide;
> evsel->no_aux_samples = true;
> evsel->name = strdup("dummy:u");
> --
> 2.42.0.459.ge4e396fd5e-goog
>
Hi Greg,
This patch is a critical performance fix for perf and vPMU. Can you
help us dispatch the commit to all stable kernel versions?
Appreciate your help. Thanks.
-Mingwei
On Mon, Oct 30, 2023 at 12:04 PM Mingwei Zhang <mizhang@google.com> wrote:
>
> On Fri, Sep 15, 2023 at 9:10 PM Ian Rogers <irogers@google.com> wrote:
> >
> > Dummy events are created with an attribute where the period and freq
> > are zero. evsel__config will then see the uninitialized values and
> > initialize them in evsel__default_freq_period. As fequency mode is
> > used by default the dummy event would be set to use frequency
> > mode. However, this has no effect on the dummy event but does cause
> > unnecessary timers/interrupts. Avoid this overhead by setting the
> > period to 1 for dummy events.
> >
> > evlist__add_aux_dummy calls evlist__add_dummy then sets freq=0 and
> > period=1. This isn't necessary after this change and so the setting is
> > removed.
> >
> > From Stephane:
> >
> > The dummy event is not counting anything. It is used to collect mmap
> > records and avoid a race condition during the synthesize mmap phase of
> > perf record. As such, it should not cause any overhead during active
> > profiling. Yet, it did. Because of a bug the dummy event was
> > programmed as a sampling event in frequency mode. Events in that mode
> > incur more kernel overheads because on timer tick, the kernel has to
> > look at the number of samples for each event and potentially adjust
> > the sampling period to achieve the desired frequency. The dummy event
> > was therefore adding a frequency event to task and ctx contexts we may
> > otherwise not have any, e.g., perf record -a -e
> > cpu/event=0x3c,period=10000000/. On each timer tick the
> > perf_adjust_freq_unthr_context() is invoked and if ctx->nr_freq is
> > non-zero, then the kernel will loop over ALL the events of the context
> > looking for frequency mode ones. In doing, so it locks the context,
> > and enable/disable the PMU of each hw event. If all the events of the
> > context are in period mode, the kernel will have to traverse the list for
> > nothing incurring overhead. The overhead is multiplied by a very large
> > factor when this happens in a guest kernel. There is no need for the
> > dummy event to be in frequency mode, it does not count anything and
> > therefore should not cause extra overhead for no reason.
> >
> > Fixes: 5bae0250237f ("perf evlist: Introduce perf_evlist__new_dummy constructor")
> > Reported-by: Stephane Eranian <eranian@google.com>
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> > tools/perf/util/evlist.c | 5 +++--
> > 1 file changed, 3 insertions(+), 2 deletions(-)
> >
> > diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
> > index 25c3ebe2c2f5..e36da58522ef 100644
> > --- a/tools/perf/util/evlist.c
> > +++ b/tools/perf/util/evlist.c
> > @@ -251,6 +251,9 @@ static struct evsel *evlist__dummy_event(struct evlist *evlist)
> > .type = PERF_TYPE_SOFTWARE,
> > .config = PERF_COUNT_SW_DUMMY,
> > .size = sizeof(attr), /* to capture ABI version */
> > + /* Avoid frequency mode for dummy events to avoid associated timers. */
> > + .freq = 0,
> > + .sample_period = 1,
> > };
> >
> > return evsel__new_idx(&attr, evlist->core.nr_entries);
> > @@ -277,8 +280,6 @@ struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
> > evsel->core.attr.exclude_kernel = 1;
> > evsel->core.attr.exclude_guest = 1;
> > evsel->core.attr.exclude_hv = 1;
> > - evsel->core.attr.freq = 0;
> > - evsel->core.attr.sample_period = 1;
> > evsel->core.system_wide = system_wide;
> > evsel->no_aux_samples = true;
> > evsel->name = strdup("dummy:u");
> > --
> > 2.42.0.459.ge4e396fd5e-goog
> >
>
> Hi Greg,
>
> This patch is a critical performance fix for perf and vPMU. Can you
> help us dispatch the commit to all stable kernel versions?
>
> Appreciate your help. Thanks.
> -Mingwei
Oops... Update target email to: stable@vger.kernel.org
On Mon, Oct 30, 2023 at 01:01:45PM -0700, Mingwei Zhang wrote:
> On Mon, Oct 30, 2023 at 12:04 PM Mingwei Zhang <mizhang@google.com> wrote:
> >
> > On Fri, Sep 15, 2023 at 9:10 PM Ian Rogers <irogers@google.com> wrote:
> > >
> > > Dummy events are created with an attribute where the period and freq
> > > are zero. evsel__config will then see the uninitialized values and
> > > initialize them in evsel__default_freq_period. As fequency mode is
> > > used by default the dummy event would be set to use frequency
> > > mode. However, this has no effect on the dummy event but does cause
> > > unnecessary timers/interrupts. Avoid this overhead by setting the
> > > period to 1 for dummy events.
> > >
> > > evlist__add_aux_dummy calls evlist__add_dummy then sets freq=0 and
> > > period=1. This isn't necessary after this change and so the setting is
> > > removed.
> > >
> > > From Stephane:
> > >
> > > The dummy event is not counting anything. It is used to collect mmap
> > > records and avoid a race condition during the synthesize mmap phase of
> > > perf record. As such, it should not cause any overhead during active
> > > profiling. Yet, it did. Because of a bug the dummy event was
> > > programmed as a sampling event in frequency mode. Events in that mode
> > > incur more kernel overheads because on timer tick, the kernel has to
> > > look at the number of samples for each event and potentially adjust
> > > the sampling period to achieve the desired frequency. The dummy event
> > > was therefore adding a frequency event to task and ctx contexts we may
> > > otherwise not have any, e.g., perf record -a -e
> > > cpu/event=0x3c,period=10000000/. On each timer tick the
> > > perf_adjust_freq_unthr_context() is invoked and if ctx->nr_freq is
> > > non-zero, then the kernel will loop over ALL the events of the context
> > > looking for frequency mode ones. In doing, so it locks the context,
> > > and enable/disable the PMU of each hw event. If all the events of the
> > > context are in period mode, the kernel will have to traverse the list for
> > > nothing incurring overhead. The overhead is multiplied by a very large
> > > factor when this happens in a guest kernel. There is no need for the
> > > dummy event to be in frequency mode, it does not count anything and
> > > therefore should not cause extra overhead for no reason.
> > >
> > > Fixes: 5bae0250237f ("perf evlist: Introduce perf_evlist__new_dummy constructor")
> > > Reported-by: Stephane Eranian <eranian@google.com>
> > > Signed-off-by: Ian Rogers <irogers@google.com>
> > > ---
> > > tools/perf/util/evlist.c | 5 +++--
> > > 1 file changed, 3 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
> > > index 25c3ebe2c2f5..e36da58522ef 100644
> > > --- a/tools/perf/util/evlist.c
> > > +++ b/tools/perf/util/evlist.c
> > > @@ -251,6 +251,9 @@ static struct evsel *evlist__dummy_event(struct evlist *evlist)
> > > .type = PERF_TYPE_SOFTWARE,
> > > .config = PERF_COUNT_SW_DUMMY,
> > > .size = sizeof(attr), /* to capture ABI version */
> > > + /* Avoid frequency mode for dummy events to avoid associated timers. */
> > > + .freq = 0,
> > > + .sample_period = 1,
> > > };
> > >
> > > return evsel__new_idx(&attr, evlist->core.nr_entries);
> > > @@ -277,8 +280,6 @@ struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
> > > evsel->core.attr.exclude_kernel = 1;
> > > evsel->core.attr.exclude_guest = 1;
> > > evsel->core.attr.exclude_hv = 1;
> > > - evsel->core.attr.freq = 0;
> > > - evsel->core.attr.sample_period = 1;
> > > evsel->core.system_wide = system_wide;
> > > evsel->no_aux_samples = true;
> > > evsel->name = strdup("dummy:u");
> > > --
> > > 2.42.0.459.ge4e396fd5e-goog
> > >
> >
> > Hi Greg,
> >
> > This patch is a critical performance fix for perf and vPMU. Can you
> > help us dispatch the commit to all stable kernel versions?
> >
> > Appreciate your help. Thanks.
> > -Mingwei
>
> Oops... Update target email to: stable@vger.kernel.org
<formletter>
This is not the correct way to submit patches for inclusion in the
stable kernel tree. Please read:
https://www.kernel.org/doc/html/latest/process/stable-kernel-rules.html
for how to do this properly.
</formletter>
On 16/09/23 07:09, Ian Rogers wrote:
> Dummy events are created with an attribute where the period and freq
> are zero. evsel__config will then see the uninitialized values and
> initialize them in evsel__default_freq_period. As fequency mode is
> used by default the dummy event would be set to use frequency
> mode. However, this has no effect on the dummy event but does cause
> unnecessary timers/interrupts. Avoid this overhead by setting the
> period to 1 for dummy events.
>
> evlist__add_aux_dummy calls evlist__add_dummy then sets freq=0 and
> period=1. This isn't necessary after this change and so the setting is
> removed.
>
> From Stephane:
>
> The dummy event is not counting anything. It is used to collect mmap
> records and avoid a race condition during the synthesize mmap phase of
> perf record. As such, it should not cause any overhead during active
> profiling. Yet, it did. Because of a bug the dummy event was
> programmed as a sampling event in frequency mode. Events in that mode
> incur more kernel overheads because on timer tick, the kernel has to
> look at the number of samples for each event and potentially adjust
> the sampling period to achieve the desired frequency. The dummy event
> was therefore adding a frequency event to task and ctx contexts we may
> otherwise not have any, e.g., perf record -a -e
> cpu/event=0x3c,period=10000000/. On each timer tick the
> perf_adjust_freq_unthr_context() is invoked and if ctx->nr_freq is
> non-zero, then the kernel will loop over ALL the events of the context
> looking for frequency mode ones. In doing, so it locks the context,
> and enable/disable the PMU of each hw event. If all the events of the
> context are in period mode, the kernel will have to traverse the list for
> nothing incurring overhead. The overhead is multiplied by a very large
> factor when this happens in a guest kernel. There is no need for the
> dummy event to be in frequency mode, it does not count anything and
> therefore should not cause extra overhead for no reason.
>
> Fixes: 5bae0250237f ("perf evlist: Introduce perf_evlist__new_dummy constructor")
> Reported-by: Stephane Eranian <eranian@google.com>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
> tools/perf/util/evlist.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
> index 25c3ebe2c2f5..e36da58522ef 100644
> --- a/tools/perf/util/evlist.c
> +++ b/tools/perf/util/evlist.c
> @@ -251,6 +251,9 @@ static struct evsel *evlist__dummy_event(struct evlist *evlist)
> .type = PERF_TYPE_SOFTWARE,
> .config = PERF_COUNT_SW_DUMMY,
> .size = sizeof(attr), /* to capture ABI version */
> + /* Avoid frequency mode for dummy events to avoid associated timers. */
> + .freq = 0,
> + .sample_period = 1,
> };
>
> return evsel__new_idx(&attr, evlist->core.nr_entries);
> @@ -277,8 +280,6 @@ struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
> evsel->core.attr.exclude_kernel = 1;
> evsel->core.attr.exclude_guest = 1;
> evsel->core.attr.exclude_hv = 1;
> - evsel->core.attr.freq = 0;
> - evsel->core.attr.sample_period = 1;
> evsel->core.system_wide = system_wide;
> evsel->no_aux_samples = true;
> evsel->name = strdup("dummy:u");
Note that evsel__config() will put it back to freq if -F is used.
Nevertheless:
Acked-by: Adrian Hunter <adrian.hunter@intel.com>
On Mon, Sep 18, 2023 at 1:14 AM Adrian Hunter <adrian.hunter@intel.com> wrote:
>
> On 16/09/23 07:09, Ian Rogers wrote:
> > Dummy events are created with an attribute where the period and freq
> > are zero. evsel__config will then see the uninitialized values and
> > initialize them in evsel__default_freq_period. As fequency mode is
> > used by default the dummy event would be set to use frequency
> > mode. However, this has no effect on the dummy event but does cause
> > unnecessary timers/interrupts. Avoid this overhead by setting the
> > period to 1 for dummy events.
> >
> > evlist__add_aux_dummy calls evlist__add_dummy then sets freq=0 and
> > period=1. This isn't necessary after this change and so the setting is
> > removed.
> >
> > From Stephane:
> >
> > The dummy event is not counting anything. It is used to collect mmap
> > records and avoid a race condition during the synthesize mmap phase of
> > perf record. As such, it should not cause any overhead during active
> > profiling. Yet, it did. Because of a bug the dummy event was
> > programmed as a sampling event in frequency mode. Events in that mode
> > incur more kernel overheads because on timer tick, the kernel has to
> > look at the number of samples for each event and potentially adjust
> > the sampling period to achieve the desired frequency. The dummy event
> > was therefore adding a frequency event to task and ctx contexts we may
> > otherwise not have any, e.g., perf record -a -e
> > cpu/event=0x3c,period=10000000/. On each timer tick the
> > perf_adjust_freq_unthr_context() is invoked and if ctx->nr_freq is
> > non-zero, then the kernel will loop over ALL the events of the context
> > looking for frequency mode ones. In doing, so it locks the context,
> > and enable/disable the PMU of each hw event. If all the events of the
> > context are in period mode, the kernel will have to traverse the list for
> > nothing incurring overhead. The overhead is multiplied by a very large
> > factor when this happens in a guest kernel. There is no need for the
> > dummy event to be in frequency mode, it does not count anything and
> > therefore should not cause extra overhead for no reason.
> >
> > Fixes: 5bae0250237f ("perf evlist: Introduce perf_evlist__new_dummy constructor")
> > Reported-by: Stephane Eranian <eranian@google.com>
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> > tools/perf/util/evlist.c | 5 +++--
> > 1 file changed, 3 insertions(+), 2 deletions(-)
> >
> > diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
> > index 25c3ebe2c2f5..e36da58522ef 100644
> > --- a/tools/perf/util/evlist.c
> > +++ b/tools/perf/util/evlist.c
> > @@ -251,6 +251,9 @@ static struct evsel *evlist__dummy_event(struct evlist *evlist)
> > .type = PERF_TYPE_SOFTWARE,
> > .config = PERF_COUNT_SW_DUMMY,
> > .size = sizeof(attr), /* to capture ABI version */
> > + /* Avoid frequency mode for dummy events to avoid associated timers. */
> > + .freq = 0,
> > + .sample_period = 1,
> > };
> >
> > return evsel__new_idx(&attr, evlist->core.nr_entries);
> > @@ -277,8 +280,6 @@ struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
> > evsel->core.attr.exclude_kernel = 1;
> > evsel->core.attr.exclude_guest = 1;
> > evsel->core.attr.exclude_hv = 1;
> > - evsel->core.attr.freq = 0;
> > - evsel->core.attr.sample_period = 1;
> > evsel->core.system_wide = system_wide;
> > evsel->no_aux_samples = true;
> > evsel->name = strdup("dummy:u");
>
> Note that evsel__config() will put it back to freq if -F is used.
Right, I was looking for a minimal fix in part for the sake of back
porting. For the -F we could do:
```
diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
index d5363d23f5d3..806185a39e17 100644
--- a/tools/perf/util/evsel.c
+++ b/tools/perf/util/evsel.c
@@ -1083,11 +1083,15 @@ void __weak arch__post_evsel_config(struct
evsel *evsel __maybe_unused,
static void evsel__set_default_freq_period(struct record_opts *opts,
struct perf_event_attr *attr)
{
- if (opts->freq) {
+ bool is_dummy = attr->type == PERF_TYPE_SOFTWARE &&
+ attr->config == PERF_COUNT_SW_DUMMY;
+
+ if (opts->freq && !is_dummy) {
attr->freq = 1;
attr->sample_freq = opts->freq;
} else {
- attr->sample_period = opts->default_interval;
+ attr->freq = 0;
+ attr->sample_period = is_dummy ? 1 : opts->default_interval;
}
}
```
But this felt like it could potentially have other side-effects.
Software events like page faults have a period, but the
dummy/side-band events appear never to check the period - but I could
be wrong. We can add the patch above but perhaps without a fixes tag.
Thanks,
Ian
> Nevertheless:
>
> Acked-by: Adrian Hunter <adrian.hunter@intel.com>
>
On Fri, Sep 15, 2023 at 9:10 PM Ian Rogers <irogers@google.com> wrote:
>
> Dummy events are created with an attribute where the period and freq
> are zero. evsel__config will then see the uninitialized values and
> initialize them in evsel__default_freq_period. As fequency mode is
> used by default the dummy event would be set to use frequency
> mode. However, this has no effect on the dummy event but does cause
> unnecessary timers/interrupts. Avoid this overhead by setting the
> period to 1 for dummy events.
>
> evlist__add_aux_dummy calls evlist__add_dummy then sets freq=0 and
> period=1. This isn't necessary after this change and so the setting is
> removed.
>
> From Stephane:
>
> The dummy event is not counting anything. It is used to collect mmap
> records and avoid a race condition during the synthesize mmap phase of
> perf record. As such, it should not cause any overhead during active
> profiling. Yet, it did. Because of a bug the dummy event was
> programmed as a sampling event in frequency mode. Events in that mode
> incur more kernel overheads because on timer tick, the kernel has to
> look at the number of samples for each event and potentially adjust
> the sampling period to achieve the desired frequency. The dummy event
> was therefore adding a frequency event to task and ctx contexts we may
> otherwise not have any, e.g., perf record -a -e
> cpu/event=0x3c,period=10000000/. On each timer tick the
> perf_adjust_freq_unthr_context() is invoked and if ctx->nr_freq is
> non-zero, then the kernel will loop over ALL the events of the context
> looking for frequency mode ones. In doing, so it locks the context,
> and enable/disable the PMU of each hw event. If all the events of the
> context are in period mode, the kernel will have to traverse the list for
> nothing incurring overhead. The overhead is multiplied by a very large
> factor when this happens in a guest kernel. There is no need for the
> dummy event to be in frequency mode, it does not count anything and
> therefore should not cause extra overhead for no reason.
>
> Fixes: 5bae0250237f ("perf evlist: Introduce perf_evlist__new_dummy constructor")
> Reported-by: Stephane Eranian <eranian@google.com>
> Signed-off-by: Ian Rogers <irogers@google.com>
> ---
> tools/perf/util/evlist.c | 5 +++--
> 1 file changed, 3 insertions(+), 2 deletions(-)
>
> diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
> index 25c3ebe2c2f5..e36da58522ef 100644
> --- a/tools/perf/util/evlist.c
> +++ b/tools/perf/util/evlist.c
> @@ -251,6 +251,9 @@ static struct evsel *evlist__dummy_event(struct evlist *evlist)
> .type = PERF_TYPE_SOFTWARE,
> .config = PERF_COUNT_SW_DUMMY,
> .size = sizeof(attr), /* to capture ABI version */
> + /* Avoid frequency mode for dummy events to avoid associated timers. */
> + .freq = 0,
> + .sample_period = 1,
> };
>
> return evsel__new_idx(&attr, evlist->core.nr_entries);
> @@ -277,8 +280,6 @@ struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
> evsel->core.attr.exclude_kernel = 1;
> evsel->core.attr.exclude_guest = 1;
> evsel->core.attr.exclude_hv = 1;
> - evsel->core.attr.freq = 0;
> - evsel->core.attr.sample_period = 1;
> evsel->core.system_wide = system_wide;
> evsel->no_aux_samples = true;
> evsel->name = strdup("dummy:u");
> --
> 2.42.0.459.ge4e396fd5e-goog
>
Thank you very much for the change. I have one quick question about
the PMU unthrottling logic. When I am looking into the function
perf_adjust_freq_unthr_context(), I see the loop with PMU stop and
start in each iteration. Is there a good way to avoid this PMU reset
operation while quickly figuring out the event in frequency mode?
Thanks.
-Mingwei
On Sat, Sep 16, 2023 at 5:46 PM Mingwei Zhang <mizhang@google.com> wrote:
>
> On Fri, Sep 15, 2023 at 9:10 PM Ian Rogers <irogers@google.com> wrote:
> >
> > Dummy events are created with an attribute where the period and freq
> > are zero. evsel__config will then see the uninitialized values and
> > initialize them in evsel__default_freq_period. As fequency mode is
> > used by default the dummy event would be set to use frequency
> > mode. However, this has no effect on the dummy event but does cause
> > unnecessary timers/interrupts. Avoid this overhead by setting the
> > period to 1 for dummy events.
> >
> > evlist__add_aux_dummy calls evlist__add_dummy then sets freq=0 and
> > period=1. This isn't necessary after this change and so the setting is
> > removed.
> >
> > From Stephane:
> >
> > The dummy event is not counting anything. It is used to collect mmap
> > records and avoid a race condition during the synthesize mmap phase of
> > perf record. As such, it should not cause any overhead during active
> > profiling. Yet, it did. Because of a bug the dummy event was
> > programmed as a sampling event in frequency mode. Events in that mode
> > incur more kernel overheads because on timer tick, the kernel has to
> > look at the number of samples for each event and potentially adjust
> > the sampling period to achieve the desired frequency. The dummy event
> > was therefore adding a frequency event to task and ctx contexts we may
> > otherwise not have any, e.g., perf record -a -e
> > cpu/event=0x3c,period=10000000/. On each timer tick the
> > perf_adjust_freq_unthr_context() is invoked and if ctx->nr_freq is
> > non-zero, then the kernel will loop over ALL the events of the context
> > looking for frequency mode ones. In doing, so it locks the context,
> > and enable/disable the PMU of each hw event. If all the events of the
> > context are in period mode, the kernel will have to traverse the list for
> > nothing incurring overhead. The overhead is multiplied by a very large
> > factor when this happens in a guest kernel. There is no need for the
> > dummy event to be in frequency mode, it does not count anything and
> > therefore should not cause extra overhead for no reason.
> >
> > Fixes: 5bae0250237f ("perf evlist: Introduce perf_evlist__new_dummy constructor")
> > Reported-by: Stephane Eranian <eranian@google.com>
> > Signed-off-by: Ian Rogers <irogers@google.com>
> > ---
> > tools/perf/util/evlist.c | 5 +++--
> > 1 file changed, 3 insertions(+), 2 deletions(-)
> >
> > diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
> > index 25c3ebe2c2f5..e36da58522ef 100644
> > --- a/tools/perf/util/evlist.c
> > +++ b/tools/perf/util/evlist.c
> > @@ -251,6 +251,9 @@ static struct evsel *evlist__dummy_event(struct evlist *evlist)
> > .type = PERF_TYPE_SOFTWARE,
> > .config = PERF_COUNT_SW_DUMMY,
> > .size = sizeof(attr), /* to capture ABI version */
> > + /* Avoid frequency mode for dummy events to avoid associated timers. */
> > + .freq = 0,
> > + .sample_period = 1,
> > };
> >
> > return evsel__new_idx(&attr, evlist->core.nr_entries);
> > @@ -277,8 +280,6 @@ struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
> > evsel->core.attr.exclude_kernel = 1;
> > evsel->core.attr.exclude_guest = 1;
> > evsel->core.attr.exclude_hv = 1;
> > - evsel->core.attr.freq = 0;
> > - evsel->core.attr.sample_period = 1;
> > evsel->core.system_wide = system_wide;
> > evsel->no_aux_samples = true;
> > evsel->name = strdup("dummy:u");
> > --
> > 2.42.0.459.ge4e396fd5e-goog
> >
>
> Thank you very much for the change. I have one quick question about
> the PMU unthrottling logic. When I am looking into the function
> perf_adjust_freq_unthr_context(), I see the loop with PMU stop and
> start in each iteration. Is there a good way to avoid this PMU reset
> operation while quickly figuring out the event in frequency mode?
Agreed. I think before the pmu_disable could be avoided for this condition:
```
if (event->hw.interrupts != MAX_INTERRUPTS &&
(!event->attr.freq || !event->attr.sample_freq))
continue;
```
Fixing up the event stop/start looks harder.
Thanks,
Ian
> Thanks.
> -Mingwei
On Mon, Sep 18, 2023 at 3:43 PM Ian Rogers <irogers@google.com> wrote:
>
> On Sat, Sep 16, 2023 at 5:46 PM Mingwei Zhang <mizhang@google.com> wrote:
> >
> > On Fri, Sep 15, 2023 at 9:10 PM Ian Rogers <irogers@google.com> wrote:
> > >
> > > Dummy events are created with an attribute where the period and freq
> > > are zero. evsel__config will then see the uninitialized values and
> > > initialize them in evsel__default_freq_period. As fequency mode is
> > > used by default the dummy event would be set to use frequency
> > > mode. However, this has no effect on the dummy event but does cause
> > > unnecessary timers/interrupts. Avoid this overhead by setting the
> > > period to 1 for dummy events.
> > >
> > > evlist__add_aux_dummy calls evlist__add_dummy then sets freq=0 and
> > > period=1. This isn't necessary after this change and so the setting is
> > > removed.
> > >
> > > From Stephane:
> > >
> > > The dummy event is not counting anything. It is used to collect mmap
> > > records and avoid a race condition during the synthesize mmap phase of
> > > perf record. As such, it should not cause any overhead during active
> > > profiling. Yet, it did. Because of a bug the dummy event was
> > > programmed as a sampling event in frequency mode. Events in that mode
> > > incur more kernel overheads because on timer tick, the kernel has to
> > > look at the number of samples for each event and potentially adjust
> > > the sampling period to achieve the desired frequency. The dummy event
> > > was therefore adding a frequency event to task and ctx contexts we may
> > > otherwise not have any, e.g., perf record -a -e
> > > cpu/event=0x3c,period=10000000/. On each timer tick the
> > > perf_adjust_freq_unthr_context() is invoked and if ctx->nr_freq is
> > > non-zero, then the kernel will loop over ALL the events of the context
> > > looking for frequency mode ones. In doing, so it locks the context,
> > > and enable/disable the PMU of each hw event. If all the events of the
> > > context are in period mode, the kernel will have to traverse the list for
> > > nothing incurring overhead. The overhead is multiplied by a very large
> > > factor when this happens in a guest kernel. There is no need for the
> > > dummy event to be in frequency mode, it does not count anything and
> > > therefore should not cause extra overhead for no reason.
> > >
> > > Fixes: 5bae0250237f ("perf evlist: Introduce perf_evlist__new_dummy constructor")
> > > Reported-by: Stephane Eranian <eranian@google.com>
> > > Signed-off-by: Ian Rogers <irogers@google.com>
> > > ---
> > > tools/perf/util/evlist.c | 5 +++--
> > > 1 file changed, 3 insertions(+), 2 deletions(-)
> > >
> > > diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
> > > index 25c3ebe2c2f5..e36da58522ef 100644
> > > --- a/tools/perf/util/evlist.c
> > > +++ b/tools/perf/util/evlist.c
> > > @@ -251,6 +251,9 @@ static struct evsel *evlist__dummy_event(struct evlist *evlist)
> > > .type = PERF_TYPE_SOFTWARE,
> > > .config = PERF_COUNT_SW_DUMMY,
> > > .size = sizeof(attr), /* to capture ABI version */
> > > + /* Avoid frequency mode for dummy events to avoid associated timers. */
> > > + .freq = 0,
> > > + .sample_period = 1,
> > > };
> > >
> > > return evsel__new_idx(&attr, evlist->core.nr_entries);
> > > @@ -277,8 +280,6 @@ struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
> > > evsel->core.attr.exclude_kernel = 1;
> > > evsel->core.attr.exclude_guest = 1;
> > > evsel->core.attr.exclude_hv = 1;
> > > - evsel->core.attr.freq = 0;
> > > - evsel->core.attr.sample_period = 1;
> > > evsel->core.system_wide = system_wide;
> > > evsel->no_aux_samples = true;
> > > evsel->name = strdup("dummy:u");
> > > --
> > > 2.42.0.459.ge4e396fd5e-goog
> > >
> >
> > Thank you very much for the change. I have one quick question about
> > the PMU unthrottling logic. When I am looking into the function
> > perf_adjust_freq_unthr_context(), I see the loop with PMU stop and
> > start in each iteration. Is there a good way to avoid this PMU reset
> > operation while quickly figuring out the event in frequency mode?
>
> Agreed. I think before the pmu_disable could be avoided for this condition:
> ```
> if (event->hw.interrupts != MAX_INTERRUPTS &&
> (!event->attr.freq || !event->attr.sample_freq))
> continue;
> ```
> Fixing up the event stop/start looks harder.
>
Right, I think putting the check early before pmu_disable() is already
a great optimization. The only concern I initially had was whether
event->hw.interrupts can be accessed before we disable the pmu. But
after checking this field in other locations, I don't see any problem
at all.
Thanks.
-Mingwei
Hello,
On Wed, Sep 20, 2023 at 10:05 PM Mingwei Zhang <mizhang@google.com> wrote:
>
> On Mon, Sep 18, 2023 at 3:43 PM Ian Rogers <irogers@google.com> wrote:
> >
> > On Sat, Sep 16, 2023 at 5:46 PM Mingwei Zhang <mizhang@google.com> wrote:
> > > Thank you very much for the change. I have one quick question about
> > > the PMU unthrottling logic. When I am looking into the function
> > > perf_adjust_freq_unthr_context(), I see the loop with PMU stop and
> > > start in each iteration. Is there a good way to avoid this PMU reset
> > > operation while quickly figuring out the event in frequency mode?
> >
> > Agreed. I think before the pmu_disable could be avoided for this condition:
> > ```
> > if (event->hw.interrupts != MAX_INTERRUPTS &&
> > (!event->attr.freq || !event->attr.sample_freq))
> > continue;
> > ```
> > Fixing up the event stop/start looks harder.
> >
>
> Right, I think putting the check early before pmu_disable() is already
> a great optimization. The only concern I initially had was whether
> event->hw.interrupts can be accessed before we disable the pmu. But
> after checking this field in other locations, I don't see any problem
> at all.
The event->hw.interrupts would be increased in the NMI handler
so there is a race between the check and the NMI. That's why
I think it checks that after disabling the PMU.
But I think we can skip non-sampling events for sure. Then it
would be better to set attr.sample_period = 0 rather than attr.freq.
if (!is_sampling_event(event))
continue;
perf_pmu_disable(event->pmu);
...
Thanks,
Namhyung
On Tue, Oct 3, 2023 at 1:08 PM Namhyung Kim <namhyung@kernel.org> wrote: > > Hello, > > On Wed, Sep 20, 2023 at 10:05 PM Mingwei Zhang <mizhang@google.com> wrote: > > > > On Mon, Sep 18, 2023 at 3:43 PM Ian Rogers <irogers@google.com> wrote: > > > > > > On Sat, Sep 16, 2023 at 5:46 PM Mingwei Zhang <mizhang@google.com> wrote: > > > > Thank you very much for the change. I have one quick question about > > > > the PMU unthrottling logic. When I am looking into the function > > > > perf_adjust_freq_unthr_context(), I see the loop with PMU stop and > > > > start in each iteration. Is there a good way to avoid this PMU reset > > > > operation while quickly figuring out the event in frequency mode? > > > > > > Agreed. I think before the pmu_disable could be avoided for this condition: > > > ``` > > > if (event->hw.interrupts != MAX_INTERRUPTS && > > > (!event->attr.freq || !event->attr.sample_freq)) > > > continue; > > > ``` > > > Fixing up the event stop/start looks harder. > > > > > > > Right, I think putting the check early before pmu_disable() is already > > a great optimization. The only concern I initially had was whether > > event->hw.interrupts can be accessed before we disable the pmu. But > > after checking this field in other locations, I don't see any problem > > at all. > > The event->hw.interrupts would be increased in the NMI handler > so there is a race between the check and the NMI. That's why > I think it checks that after disabling the PMU. > > But I think we can skip non-sampling events for sure. Then it > would be better to set attr.sample_period = 0 rather than attr.freq. > > if (!is_sampling_event(event)) > continue; > > perf_pmu_disable(event->pmu); > ... > > Thanks, > Namhyung With the PMU disabled, isn't there still a risk of an interrupt still being in flight? In other words the disable doesn't prevent a race and we'll catch this on the next timer call to perf_adjust_freq_unthr_context. I think we can also improve the code by just disabling a PMU once, we can take advantage of the perf_event_pmu_context and disable that PMU, iterate its events and then re-enable the PMU - i.e. no need for an enable and disable per event. I'll put a patch together. Thanks, Ian
On Tue, Oct 3, 2023 at 3:36 PM Ian Rogers <irogers@google.com> wrote: > > On Tue, Oct 3, 2023 at 1:08 PM Namhyung Kim <namhyung@kernel.org> wrote: > > > > Hello, > > > > On Wed, Sep 20, 2023 at 10:05 PM Mingwei Zhang <mizhang@google.com> wrote: > > > > > > On Mon, Sep 18, 2023 at 3:43 PM Ian Rogers <irogers@google.com> wrote: > > > > > > > > On Sat, Sep 16, 2023 at 5:46 PM Mingwei Zhang <mizhang@google.com> wrote: > > > > > Thank you very much for the change. I have one quick question about > > > > > the PMU unthrottling logic. When I am looking into the function > > > > > perf_adjust_freq_unthr_context(), I see the loop with PMU stop and > > > > > start in each iteration. Is there a good way to avoid this PMU reset > > > > > operation while quickly figuring out the event in frequency mode? > > > > > > > > Agreed. I think before the pmu_disable could be avoided for this condition: > > > > ``` > > > > if (event->hw.interrupts != MAX_INTERRUPTS && > > > > (!event->attr.freq || !event->attr.sample_freq)) > > > > continue; > > > > ``` > > > > Fixing up the event stop/start looks harder. > > > > > > > > > > Right, I think putting the check early before pmu_disable() is already > > > a great optimization. The only concern I initially had was whether > > > event->hw.interrupts can be accessed before we disable the pmu. But > > > after checking this field in other locations, I don't see any problem > > > at all. > > > > The event->hw.interrupts would be increased in the NMI handler > > so there is a race between the check and the NMI. That's why > > I think it checks that after disabling the PMU. > > > > But I think we can skip non-sampling events for sure. Then it > > would be better to set attr.sample_period = 0 rather than attr.freq. > > > > if (!is_sampling_event(event)) > > continue; > > > > perf_pmu_disable(event->pmu); > > ... > > > > Thanks, > > Namhyung > > With the PMU disabled, isn't there still a risk of an interrupt still > being in flight? In other words the disable doesn't prevent a race and > we'll catch this on the next timer call to > perf_adjust_freq_unthr_context. I think we can also improve the code > by just disabling a PMU once, we can take advantage of the > perf_event_pmu_context and disable that PMU, iterate its events and > then re-enable the PMU - i.e. no need for an enable and disable per > event. I'll put a patch together. > > Thanks, > Ian +Jim Mattson I initially thought this idea was just an alternative, or a more professional fix in the perf subsystem. I was wrong... This would be way better than just skipping frequency events in the loop. Since if we just skip by event, we may still suffer from huge overhead if the event list contains many sampling events in frequency mode. Unfortunately, that is the general case when we do perf record -e 'eventlist' (IIUC all events in eventlist are in frequency mode if we don't specify period=). So the problem actually remains whenever we do perf sampling unless we use something like Intel vtune. On the other hand, since all of the events are presumably CPU core events, with the fix we pay only once for the PMU reset per hrtimer regardless of how many events are in frequency mode. Looking forward to the patch! Please keep us posted if possible. Thanks. -Mingwei
Hi Ian, On Tue, Oct 3, 2023 at 4:20 PM Mingwei Zhang <mizhang@google.com> wrote: > > On Tue, Oct 3, 2023 at 3:36 PM Ian Rogers <irogers@google.com> wrote: > > > > On Tue, Oct 3, 2023 at 1:08 PM Namhyung Kim <namhyung@kernel.org> wrote: > > > > > > Hello, > > > > > > On Wed, Sep 20, 2023 at 10:05 PM Mingwei Zhang <mizhang@google.com> wrote: > > > > > > > > On Mon, Sep 18, 2023 at 3:43 PM Ian Rogers <irogers@google.com> wrote: > > > > > > > > > > On Sat, Sep 16, 2023 at 5:46 PM Mingwei Zhang <mizhang@google.com> wrote: > > > > > > Thank you very much for the change. I have one quick question about > > > > > > the PMU unthrottling logic. When I am looking into the function > > > > > > perf_adjust_freq_unthr_context(), I see the loop with PMU stop and > > > > > > start in each iteration. Is there a good way to avoid this PMU reset > > > > > > operation while quickly figuring out the event in frequency mode? > > > > > > > > > > Agreed. I think before the pmu_disable could be avoided for this condition: > > > > > ``` > > > > > if (event->hw.interrupts != MAX_INTERRUPTS && > > > > > (!event->attr.freq || !event->attr.sample_freq)) > > > > > continue; > > > > > ``` > > > > > Fixing up the event stop/start looks harder. > > > > > > > > > > > > > Right, I think putting the check early before pmu_disable() is already > > > > a great optimization. The only concern I initially had was whether > > > > event->hw.interrupts can be accessed before we disable the pmu. But > > > > after checking this field in other locations, I don't see any problem > > > > at all. > > > > > > The event->hw.interrupts would be increased in the NMI handler > > > so there is a race between the check and the NMI. That's why > > > I think it checks that after disabling the PMU. > > > > > > But I think we can skip non-sampling events for sure. Then it > > > would be better to set attr.sample_period = 0 rather than attr.freq. > > > > > > if (!is_sampling_event(event)) > > > continue; > > > > > > perf_pmu_disable(event->pmu); > > > ... > > > > > > Thanks, > > > Namhyung > > > > With the PMU disabled, isn't there still a risk of an interrupt still > > being in flight? In other words the disable doesn't prevent a race and > > we'll catch this on the next timer call to > > perf_adjust_freq_unthr_context. I think we can also improve the code > > by just disabling a PMU once, we can take advantage of the > > perf_event_pmu_context and disable that PMU, iterate its events and > > then re-enable the PMU - i.e. no need for an enable and disable per > > event. I'll put a patch together. > > > > Thanks, > > Ian > > +Jim Mattson > > I initially thought this idea was just an alternative, or a more > professional fix in the perf subsystem. I was wrong... > > This would be way better than just skipping frequency events in the > loop. Since if we just skip by event, we may still suffer from huge > overhead if the event list contains many sampling events in frequency > mode. Unfortunately, that is the general case when we do perf record > -e 'eventlist' (IIUC all events in eventlist are in frequency mode if > we don't specify period=). So the problem actually remains whenever we > do perf sampling unless we use something like Intel vtune. > > On the other hand, since all of the events are presumably CPU core > events, with the fix we pay only once for the PMU reset per hrtimer > regardless of how many events are in frequency mode. > > Looking forward to the patch! Please keep us posted if possible. Any updates? Thanks, Namhyung
On Tue, Oct 3, 2023 at 3:36 PM Ian Rogers <irogers@google.com> wrote: > > On Tue, Oct 3, 2023 at 1:08 PM Namhyung Kim <namhyung@kernel.org> wrote: > > > > Hello, > > > > On Wed, Sep 20, 2023 at 10:05 PM Mingwei Zhang <mizhang@google.com> wrote: > > > > > > On Mon, Sep 18, 2023 at 3:43 PM Ian Rogers <irogers@google.com> wrote: > > > > > > > > On Sat, Sep 16, 2023 at 5:46 PM Mingwei Zhang <mizhang@google.com> wrote: > > > > > Thank you very much for the change. I have one quick question about > > > > > the PMU unthrottling logic. When I am looking into the function > > > > > perf_adjust_freq_unthr_context(), I see the loop with PMU stop and > > > > > start in each iteration. Is there a good way to avoid this PMU reset > > > > > operation while quickly figuring out the event in frequency mode? > > > > > > > > Agreed. I think before the pmu_disable could be avoided for this condition: > > > > ``` > > > > if (event->hw.interrupts != MAX_INTERRUPTS && > > > > (!event->attr.freq || !event->attr.sample_freq)) > > > > continue; > > > > ``` > > > > Fixing up the event stop/start looks harder. > > > > > > > > > > Right, I think putting the check early before pmu_disable() is already > > > a great optimization. The only concern I initially had was whether > > > event->hw.interrupts can be accessed before we disable the pmu. But > > > after checking this field in other locations, I don't see any problem > > > at all. > > > > The event->hw.interrupts would be increased in the NMI handler > > so there is a race between the check and the NMI. That's why > > I think it checks that after disabling the PMU. > > > > But I think we can skip non-sampling events for sure. Then it > > would be better to set attr.sample_period = 0 rather than attr.freq. > > > > if (!is_sampling_event(event)) > > continue; > > > > perf_pmu_disable(event->pmu); > > ... > > > > Thanks, > > Namhyung > > With the PMU disabled, isn't there still a risk of an interrupt still > being in flight? In other words the disable doesn't prevent a race and > we'll catch this on the next timer call to > perf_adjust_freq_unthr_context. I think we can also improve the code > by just disabling a PMU once, we can take advantage of the > perf_event_pmu_context and disable that PMU, iterate its events and > then re-enable the PMU - i.e. no need for an enable and disable per > event. I'll put a patch together. Thanks, I was thinking about that too. It's also a side effect of the context rewrite. Maybe we could iterate pmu_ctx's active lists and skip pmus with PERF_PMU_CAP_NO_INTERRUPT and individual non-sampling events. Thanks, Namhyung
On Tue, Oct 3, 2023 at 4:02 PM Namhyung Kim <namhyung@kernel.org> wrote: > > On Tue, Oct 3, 2023 at 3:36 PM Ian Rogers <irogers@google.com> wrote: > > With the PMU disabled, isn't there still a risk of an interrupt still > > being in flight? In other words the disable doesn't prevent a race and > > we'll catch this on the next timer call to > > perf_adjust_freq_unthr_context. I think we can also improve the code > > by just disabling a PMU once, we can take advantage of the > > perf_event_pmu_context and disable that PMU, iterate its events and > > then re-enable the PMU - i.e. no need for an enable and disable per > > event. I'll put a patch together. > > Thanks, I was thinking about that too. It's also a side effect of > the context rewrite. Maybe we could iterate pmu_ctx's active lists > and skip pmus with PERF_PMU_CAP_NO_INTERRUPT and > individual non-sampling events. Or we can add pmu_ctx->nr_freq and check it before accessing pmu MSRs. Thanks, Namhyung
© 2016 - 2026 Red Hat, Inc.