On Tue, Dec 2, 2025 at 4:03 AM Douglas Raillard
<[email protected]> wrote:
>
> Hi Samuel,
>
> On 01-12-2025 20:24, Samuel Wu wrote:
> > The existing cpu_frequency trace_event can be verbose, emitting a nearly
> > identical trace event for every CPU in the policy even when their
> > frequencies are identical.
> >
> > This patch replaces the cpu_frequency trace event with policy_frequency
> > trace event, a more efficient alternative. From the kernel's
> > perspective, emitting a trace event once per policy instead of once per
> > cpu saves some memory and is less overhead.
>
> I'd be fully behind that as a general guideline.
>
> > From the post-processing
> > perspective, analysis of the trace log is simplified without any loss of
> > information.
>
> Unfortunately I'm not so sure about the "simplified" part (as of today),
> more on that below.
>
> >
> > Signed-off-by: Samuel Wu <[email protected]>
> > ---
> >   drivers/cpufreq/cpufreq.c      | 14 ++------------
> >   drivers/cpufreq/intel_pstate.c |  6 ++++--
> >   include/trace/events/power.h   | 24 +++++++++++++++++++++---
> >   kernel/trace/power-traces.c    |  2 +-
> >   samples/bpf/cpustat_kern.c     |  8 ++++----
> >   samples/bpf/cpustat_user.c     |  6 +++---
> >   tools/perf/builtin-timechart.c | 12 ++++++------
> >   7 files changed, 41 insertions(+), 31 deletions(-)
> >
> > diff --git a/drivers/cpufreq/cpufreq.c b/drivers/cpufreq/cpufreq.c
> > index 4472bb1ec83c..dd3f08f3b958 100644
> > --- a/drivers/cpufreq/cpufreq.c
> > +++ b/drivers/cpufreq/cpufreq.c
> > @@ -309,8 +309,6 @@ static void cpufreq_notify_transition(struct 
> > cpufreq_policy *policy,
> >                                     struct cpufreq_freqs *freqs,
> >                                     unsigned int state)
> >   {
> > -     int cpu;
> > -
> >       BUG_ON(irqs_disabled());
> >
> >       if (cpufreq_disabled())
> > @@ -344,10 +342,7 @@ static void cpufreq_notify_transition(struct 
> > cpufreq_policy *policy,
> >               adjust_jiffies(CPUFREQ_POSTCHANGE, freqs);
> >               pr_debug("FREQ: %u - CPUs: %*pbl\n", freqs->new,
> >                        cpumask_pr_args(policy->cpus));
> > -
> > -             for_each_cpu(cpu, policy->cpus)
> > -                     trace_cpu_frequency(freqs->new, cpu);
> > -
> > +             trace_policy_frequency(freqs->new, policy->cpu, policy->cpus);
> >               srcu_notifier_call_chain(&cpufreq_transition_notifier_list,
> >                                        CPUFREQ_POSTCHANGE, freqs);
> >
> > @@ -2201,7 +2196,6 @@ unsigned int cpufreq_driver_fast_switch(struct 
> > cpufreq_policy *policy,
> >                                       unsigned int target_freq)
> >   {
> >       unsigned int freq;
> > -     int cpu;
> >
> >       target_freq = clamp_val(target_freq, policy->min, policy->max);
> >       freq = cpufreq_driver->fast_switch(policy, target_freq);
> > @@ -2213,11 +2207,7 @@ unsigned int cpufreq_driver_fast_switch(struct 
> > cpufreq_policy *policy,
> >       arch_set_freq_scale(policy->related_cpus, freq,
> >                           arch_scale_freq_ref(policy->cpu));
> >       cpufreq_stats_record_transition(policy, freq);
> > -
> > -     if (trace_cpu_frequency_enabled()) {
> > -             for_each_cpu(cpu, policy->cpus)
> > -                     trace_cpu_frequency(freq, cpu);
> > -     }
> > +     trace_policy_frequency(freq, policy->cpu, policy->cpus);
> >
> >       return freq;
> >   }
> > diff --git a/drivers/cpufreq/intel_pstate.c b/drivers/cpufreq/intel_pstate.c
> > index ec4abe374573..9724b5d19d83 100644
> > --- a/drivers/cpufreq/intel_pstate.c
> > +++ b/drivers/cpufreq/intel_pstate.c
> > @@ -2297,7 +2297,8 @@ static int hwp_get_cpu_scaling(int cpu)
> >
> >   static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
> >   {
> > -     trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
> > +     trace_policy_frequency(pstate * cpu->pstate.scaling, cpu->cpu,
> > +                            cpumask_of(cpu->cpu));
> >       cpu->pstate.current_pstate = pstate;
> >       /*
> >        * Generally, there is no guarantee that this code will always run on
> > @@ -2587,7 +2588,8 @@ static void intel_pstate_adjust_pstate(struct cpudata 
> > *cpu)
> >
> >       target_pstate = get_target_pstate(cpu);
> >       target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
> > -     trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
> > +     trace_policy_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu,
> > +                            cpumask_of(cpu->cpu));
> >       intel_pstate_update_pstate(cpu, target_pstate);
> >
> >       sample = &cpu->sample;
> > diff --git a/include/trace/events/power.h b/include/trace/events/power.h
> > index 370f8df2fdb4..317098ffdd5f 100644
> > --- a/include/trace/events/power.h
> > +++ b/include/trace/events/power.h
> > @@ -182,11 +182,29 @@ TRACE_EVENT(pstate_sample,
> >               { PM_EVENT_RECOVER, "recover" }, \
> >               { PM_EVENT_POWEROFF, "poweroff" })
> >
> > -DEFINE_EVENT(cpu, cpu_frequency,
> > +TRACE_EVENT(policy_frequency,
> >
> > -     TP_PROTO(unsigned int frequency, unsigned int cpu_id),
> > +     TP_PROTO(unsigned int frequency, unsigned int cpu_id,
> > +              const struct cpumask *policy_cpus),
> >
> > -     TP_ARGS(frequency, cpu_id)
> > +     TP_ARGS(frequency, cpu_id, policy_cpus),
> > +
> > +     TP_STRUCT__entry(
> > +             __field(u32, state)
> > +             __field(u32, cpu_id)
> > +             __cpumask(cpumask)
>
> Using a cpumask is the most technically correct option here, but it also 
> carries a big issue.
> Userspace tooling will have a very hard time doing anything with it. A lot of 
> that is down
> to having no appropriate counterpart in "table libraries" in general (e.g. 
> what would you
> map that to in pandas, polars or SQL ?). Some of the lack of support is 
> probably also down to how
> infrequently used it is. For example I don't think Perfetto would be able to 
> handle that
> in the ftrace_event and args table, as the documented supported value types 
> are:

I've been in touch with the Perfetto team through this process, and it
is an easy update for them to handle. If there are other libraries
using this event, I think their approach would be similar to how
Perfetto's SQL tables handle this new trace event.

> args table:
> value_type      STRING  The type of the value of the arg. Will be one of 
> 'int', 'uint', 'string', 'real', 'pointer', 'bool' or 'json'.
> https://perfetto.dev/docs/analysis/stdlib-docs

In the same URL as above, there is a 'cpu_frequency_counters' SQL
table documentation. The new tracepoint would be parsed by Perfetto's
C++ parser before being inserted into the aforementioned SQL table as
four separate rows, given a policy with 4 CPUs. This effectively
creates the same table with the same data (just slightly different
timestamps) as prior to this patch.

>
> So while I definitely support improving the situation around cpumasks (I 
> lobbied a bit for that),
> I don't think the ecosystem is ready for it yet and having such a core event 
> switched to using it
> is going to cause a lot of pain.
>
> Some alternatives for tooling could be:
> 1. Record the policy cpumasks in a tool-friendly format in the trace header, 
> but no current format I know
>     of provides that, and ftrace does not provide a "JSON blob to be passed 
> through" we could easily append to.
>     Any such addition will therefore require libraries update which will take 
> time.
>
> 3. Doing without the data in the trace. That means collecting and bundling 
> another sidecar file, which
>     is really not convenient and still requires 3rd party tool modifications 
> for end users.
>
> 2. Add policy_frequency event, but not remove cpu_frequency yet. Possibly 
> with a deprecation warning
>     when enabling the event.
>

These alternatives should work, but I feel are probably too complex to
be worth the effort.

> > +     ),
> > +
> > +     TP_fast_assign(
> > +             __entry->state = frequency;
> > +             __entry->cpu_id = cpu_id;
> > +             __assign_cpumask(cpumask, policy_cpus);
>
> ipi_send_cpumask uses cpumask_bits():
>
>                 __assign_cpumask(cpumask, cpumask_bits(cpumask));
>
> It's not clear what is best practice, as struct cpumask contains a single 
> member anyway and
> __assign_cpumask() expands to a memcpy() so they are functionally identical.
>
> > +     ),
> > +
> > +     TP_printk("state=%lu cpu_id=%lu policy_cpus=%*pb",
> > +               (unsigned long)__entry->state,
> > +               (unsigned long)__entry->cpu_id,
> > +               cpumask_pr_args((struct cpumask 
> > *)__get_dynamic_array(cpumask)))
>
> Looking at ipi_send_cpumask, this should be:
>
>    __get_cpumask(cpumask)
>
> The cast and cpumask_pr_args() may look like it's working, but there is only 
> a very slim
> chance any downstream tool will know what to do with this. Looking at 
> libtraceevent
> (which trace-cmd is based on):
>
> ./utest/traceevent-utest.c:116: "print fmt: \"cpumask=%s\", 
> __get_cpumask(cpumask)\n";
> ./src/event-parse.c:3674:       if (strcmp(token, "__get_cpumask") == 0 ||
> ./src/event-parse.c:7568:               printf("__get_cpumask(%s)", 
> args->bitmask.bitmask);
>
> But there is no match for "cpumask_pr_args".

Thanks for pointing out libtraceevent- I didn't know about this tool,
but I'll take a look at compatibility and adjust appropriately. The
macros are a little tricky, but I agree that it seems best to follow
the template set out by the pre-existing ipi_send_cpumask.

>
> Considering the gap between what works when using the in-kernel text 
> rendering and what can be
> reasonably expected to work in any other userspace tool, it's a good idea to 
> try
> as many as possible unfortunately.
>

Overall, I appreciate the thorough and insightful feedback Douglas! If
there are any other tools consuming cpu_frequency, I can help update
them appropriately. AFAICT Perfetto is by far, the most widespread
tool consuming cpu_frequency.

> >   );
> >
> >   TRACE_EVENT(cpu_frequency_limits,
> > diff --git a/kernel/trace/power-traces.c b/kernel/trace/power-traces.c
> > index f2fe33573e54..a537e68a6878 100644
> > --- a/kernel/trace/power-traces.c
> > +++ b/kernel/trace/power-traces.c
> > @@ -16,5 +16,5 @@
> >
> >   EXPORT_TRACEPOINT_SYMBOL_GPL(suspend_resume);
> >   EXPORT_TRACEPOINT_SYMBOL_GPL(cpu_idle);
> > -EXPORT_TRACEPOINT_SYMBOL_GPL(cpu_frequency);
> > +EXPORT_TRACEPOINT_SYMBOL_GPL(policy_frequency);
> >
> > diff --git a/samples/bpf/cpustat_kern.c b/samples/bpf/cpustat_kern.c
> > index 7ec7143e2757..f485de0f89b2 100644
> > --- a/samples/bpf/cpustat_kern.c
> > +++ b/samples/bpf/cpustat_kern.c
> > @@ -75,9 +75,9 @@ struct {
> >   } pstate_duration SEC(".maps");
> >
> >   /*
> > - * The trace events for cpu_idle and cpu_frequency are taken from:
> > + * The trace events for cpu_idle and policy_frequency are taken from:
> >    * /sys/kernel/tracing/events/power/cpu_idle/format
> > - * /sys/kernel/tracing/events/power/cpu_frequency/format
> > + * /sys/kernel/tracing/events/power/policy_frequency/format
> >    *
> >    * These two events have same format, so define one common structure.
> >    */
> > @@ -162,7 +162,7 @@ int bpf_prog1(struct cpu_args *ctx)
> >        */
> >       if (ctx->state != (u32)-1) {
> >
> > -             /* record pstate after have first cpu_frequency event */
> > +             /* record pstate after have first policy_frequency event */
> >               if (!*pts)
> >                       return 0;
> >
> > @@ -208,7 +208,7 @@ int bpf_prog1(struct cpu_args *ctx)
> >       return 0;
> >   }
> >
> > -SEC("tracepoint/power/cpu_frequency")
> > +SEC("tracepoint/power/policy_frequency")
> >   int bpf_prog2(struct cpu_args *ctx)
> >   {
> >       u64 *pts, *cstate, *pstate, cur_ts, delta;
> > diff --git a/samples/bpf/cpustat_user.c b/samples/bpf/cpustat_user.c
> > index 356f756cba0d..f7e81f702358 100644
> > --- a/samples/bpf/cpustat_user.c
> > +++ b/samples/bpf/cpustat_user.c
> > @@ -143,12 +143,12 @@ static int cpu_stat_inject_cpu_idle_event(void)
> >
> >   /*
> >    * It's possible to have no any frequency change for long time and cannot
> > - * get ftrace event 'trace_cpu_frequency' for long period, this introduces
> > + * get ftrace event 'trace_policy_frequency' for long period, this 
> > introduces
> >    * big deviation for pstate statistics.
> >    *
> >    * To solve this issue, below code forces to set 'scaling_max_freq' to 
> > 208MHz
> > - * for triggering ftrace event 'trace_cpu_frequency' and then recovery 
> > back to
> > - * the maximum frequency value 1.2GHz.
> > + * for triggering ftrace event 'trace_policy_frequency' and then recovery 
> > back
> > + * to the maximum frequency value 1.2GHz.
> >    */
> >   static int cpu_stat_inject_cpu_frequency_event(void)
> >   {
> > diff --git a/tools/perf/builtin-timechart.c b/tools/perf/builtin-timechart.c
> > index 22050c640dfa..3ef1a2fd0493 100644
> > --- a/tools/perf/builtin-timechart.c
> > +++ b/tools/perf/builtin-timechart.c
> > @@ -612,10 +612,10 @@ process_sample_cpu_idle(struct timechart *tchart 
> > __maybe_unused,
> >   }
> >
> >   static int
> > -process_sample_cpu_frequency(struct timechart *tchart,
> > -                          struct evsel *evsel,
> > -                          struct perf_sample *sample,
> > -                          const char *backtrace __maybe_unused)
> > +process_sample_policy_frequency(struct timechart *tchart,
> > +                             struct evsel *evsel,
> > +                             struct perf_sample *sample,
> > +                             const char *backtrace __maybe_unused)
> >   {
> >       u32 state  = evsel__intval(evsel, sample, "state");
> >       u32 cpu_id = evsel__intval(evsel, sample, "cpu_id");
> > @@ -1541,7 +1541,7 @@ static int __cmd_timechart(struct timechart *tchart, 
> > const char *output_name)
> >   {
> >       const struct evsel_str_handler power_tracepoints[] = {
> >               { "power:cpu_idle",             process_sample_cpu_idle },
> > -             { "power:cpu_frequency",        process_sample_cpu_frequency 
> > },
> > +             { "power:policy_frequency",     
> > process_sample_policy_frequency },
> >               { "sched:sched_wakeup",         process_sample_sched_wakeup },
> >               { "sched:sched_switch",         process_sample_sched_switch },
> >   #ifdef SUPPORT_OLD_POWER_EVENTS
> > @@ -1804,7 +1804,7 @@ static int timechart__record(struct timechart 
> > *tchart, int argc, const char **ar
> >       unsigned int backtrace_args_no = ARRAY_SIZE(backtrace_args);
> >
> >       const char * const power_args[] = {
> > -             "-e", "power:cpu_frequency",
> > +             "-e", "power:policy_frequency",
> >               "-e", "power:cpu_idle",
> >       };
> >       unsigned int power_args_nr = ARRAY_SIZE(power_args);
>
> --
>
> Douglas
>

Reply via email to