Re: [RFC PATCH v2] perf evsel: Fallback to task-clock when not system wide

[Athira Rajeev]

> On 05-Dec-2023, at 8:55 PM, Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
> 
> Em Tue, Dec 05, 2023 at 04:51:01PM +0530, Athira Rajeev escreveu:
>> 
>> 
>>> On 21-Nov-2023, at 5:34 AM, Ian Rogers <irogers@xxxxxxxxxx> wrote:
>>> 
>>> When the cycles event isn't available evsel will fallback to the
>>> cpu-clock software event. task-clock is similar to cpu-clock but only
>>> runs when the process is running. Falling back to cpu-clock when not
>>> system wide leads to confusion, by falling back to task-clock it is
>>> hoped the confusion is less.
>>> 
>>> Pass the target to determine if task-clock is more appropriate. Update
>>> a nearby comment and debug string for the change.
>>> 
>>> ---
>>> v2. Use target__has_cpu as suggested by Namhyung.
>>> https://lpc.events/event/17/contributions/1556/
>>> 
>>> Signed-off-by: Ian Rogers <irogers@xxxxxxxxxx>
>>> ---
>>> tools/perf/builtin-record.c |  2 +-
>>> tools/perf/builtin-stat.c   |  2 +-
>>> tools/perf/builtin-top.c    |  2 +-
>>> tools/perf/util/evsel.c     | 18 ++++++++++--------
>>> tools/perf/util/evsel.h     |  3 ++-
>>> 5 files changed, 15 insertions(+), 12 deletions(-)
>>> 
>>> diff --git a/tools/perf/builtin-record.c b/tools/perf/builtin-record.c
>>> index 8ec818568662..d8bb59511fdd 100644
>>> --- a/tools/perf/builtin-record.c
>>> +++ b/tools/perf/builtin-record.c
>>> @@ -1350,7 +1350,7 @@ static int record__open(struct record *rec)
>>> evlist__for_each_entry(evlist, pos) {
>>> try_again:
>>> if (evsel__open(pos, pos->core.cpus, pos->core.threads) < 0) {
>>> - if (evsel__fallback(pos, errno, msg, sizeof(msg))) {
>>> + if (evsel__fallback(pos, &opts->target, errno, msg, sizeof(msg))) {
>> 
>> Hi Ian
>> 
>> Tested this with perf record and I could find the code fallback to using task-clock
>> 
>> ./perf record -v ls
>> Warning:
>> The cycles event is not supported, trying to fall back to task-clock
> 
> Ok, so I'll take that as a Tested-by: you, ok?

Hi Arnaldo,

Please add my Tested-by.

Tested-by: Athira Rajeev <atrajeev@xxxxxxxxxxxxxxxxxx>

> 
> The "perf stat" part can be addressed in a follow up patch, when that
> error handling is researched to remember why we have that ->supported,
> ->errored thing.

Sure, I will take a look at why we have that difference in “perf stat”

Thanks
Athira
> 
> - Arnaldo
> 
>> But in case of “perf stat”, in my environment, found that the code path won't invoke “evsel__fallback”.
>> 
>> Snippet for builtin-stat.c
>>            if (errno == EINVAL || errno == ENOSYS ||
>>            errno == ENOENT || errno == EOPNOTSUPP ||
>>            errno == ENXIO) {
>>                if (verbose > 0)
>>                        ui__warning("%s event is not supported by the kernel.\n",
>>                                    evsel__name(counter));
>>                counter->supported = false;
>>                /*
>>                 * errored is a sticky flag that means one of the counter's
>>                 * cpu event had a problem and needs to be reexamined.
>>                 */
>>                counter->errored = true;
>> 
>>                if ((evsel__leader(counter) != counter) ||
>>                    !(counter->core.leader->nr_members > 1))
>>                        return COUNTER_SKIP;
>>        } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
>>                if (verbose > 0)
>>                        ui__warning("%s\n", msg);
>>                return COUNTER_RETRY;
>> 
>> So if the perf_event_open returns ENOENT, we won’t do a fallback in builtin-stat.c
>> Should we address cycles differently here ? Any comments ?
>> 
>> Thanks
>> Athira
>>> 
>>> if (verbose > 0)
>>> ui__warning("%s\n", msg);
>>> goto try_again;
>>> diff --git a/tools/perf/builtin-stat.c b/tools/perf/builtin-stat.c
>>> index a3af805a1d57..d8e5d6f7a87a 100644
>>> --- a/tools/perf/builtin-stat.c
>>> +++ b/tools/perf/builtin-stat.c
>>> @@ -653,7 +653,7 @@ static enum counter_recovery stat_handle_error(struct evsel *counter)
>>> if ((evsel__leader(counter) != counter) ||
>>>   !(counter->core.leader->nr_members > 1))
>>> return COUNTER_SKIP;
>>> - } else if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
>>> + } else if (evsel__fallback(counter, &target, errno, msg, sizeof(msg))) {
>>> if (verbose > 0)
>>> ui__warning("%s\n", msg);
>>> return COUNTER_RETRY;
>>> diff --git a/tools/perf/builtin-top.c b/tools/perf/builtin-top.c
>>> index ea8c7eca5eee..1e42bd1c7d5a 100644
>>> --- a/tools/perf/builtin-top.c
>>> +++ b/tools/perf/builtin-top.c
>>> @@ -1044,7 +1044,7 @@ static int perf_top__start_counters(struct perf_top *top)
>>>   perf_top_overwrite_fallback(top, counter))
>>> goto try_again;
>>> 
>>> - if (evsel__fallback(counter, errno, msg, sizeof(msg))) {
>>> + if (evsel__fallback(counter, &opts->target, errno, msg, sizeof(msg))) {
>>> if (verbose > 0)
>>> ui__warning("%s\n", msg);
>>> goto try_again;
>>> diff --git a/tools/perf/util/evsel.c b/tools/perf/util/evsel.c
>>> index a5da74e3a517..532f34d9fcb5 100644
>>> --- a/tools/perf/util/evsel.c
>>> +++ b/tools/perf/util/evsel.c
>>> @@ -2853,7 +2853,8 @@ u64 evsel__intval_common(struct evsel *evsel, struct perf_sample *sample, const
>>> 
>>> #endif
>>> 
>>> -bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
>>> +bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
>>> +     char *msg, size_t msgsize)
>>> {
>>> int paranoid;
>>> 
>>> @@ -2861,18 +2862,19 @@ bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize)
>>>   evsel->core.attr.type   == PERF_TYPE_HARDWARE &&
>>>   evsel->core.attr.config == PERF_COUNT_HW_CPU_CYCLES) {
>>> /*
>>> - * If it's cycles then fall back to hrtimer based
>>> - * cpu-clock-tick sw counter, which is always available even if
>>> - * no PMU support.
>>> + * If it's cycles then fall back to hrtimer based cpu-clock sw
>>> + * counter, which is always available even if no PMU support.
>>> *
>>> * PPC returns ENXIO until 2.6.37 (behavior changed with commit
>>> * b0a873e).
>>> */
>>> - scnprintf(msg, msgsize, "%s",
>>> -"The cycles event is not supported, trying to fall back to cpu-clock-ticks");
>>> -
>>> evsel->core.attr.type   = PERF_TYPE_SOFTWARE;
>>> - evsel->core.attr.config = PERF_COUNT_SW_CPU_CLOCK;
>>> + evsel->core.attr.config = target__has_cpu(target)
>>> + ? PERF_COUNT_SW_CPU_CLOCK
>>> + : PERF_COUNT_SW_TASK_CLOCK;
>>> + scnprintf(msg, msgsize,
>>> + "The cycles event is not supported, trying to fall back to %s",
>>> + target__has_cpu(target) ? "cpu-clock" : "task-clock");
>>> 
>>> zfree(&evsel->name);
>>> return true;
>>> diff --git a/tools/perf/util/evsel.h b/tools/perf/util/evsel.h
>>> index f19ac9f027ef..efbb6e848287 100644
>>> --- a/tools/perf/util/evsel.h
>>> +++ b/tools/perf/util/evsel.h
>>> @@ -460,7 +460,8 @@ static inline bool evsel__is_clock(const struct evsel *evsel)
>>>      evsel__match(evsel, SOFTWARE, SW_TASK_CLOCK);
>>> }
>>> 
>>> -bool evsel__fallback(struct evsel *evsel, int err, char *msg, size_t msgsize);
>>> +bool evsel__fallback(struct evsel *evsel, struct target *target, int err,
>>> +     char *msg, size_t msgsize);
>>> int evsel__open_strerror(struct evsel *evsel, struct target *target,
>>> int err, char *msg, size_t size);
>>> 
>>> -- 
>>> 2.43.0.rc1.413.gea7ed67945-goog
>>> 
>> 
> 
> -- 
> 
> - Arnaldo