Re: [PATCH V2 13/23] perf evlist: Add evlist__add_dummy_on_all_cpus()

[Adrian Hunter <adrian.hunter@intel.com>]

On 10/05/22 19:24, Arnaldo Carvalho de Melo wrote:
> Em Tue, May 10, 2022 at 05:55:34PM +0300, Adrian Hunter escreveu:
>> On 6/05/22 18:35, Ian Rogers wrote:
>>> On Fri, May 6, 2022 at 8:08 AM Adrian Hunter <adrian.hunter@intel.com> wrote:
>>>>
>>>> On 6/05/22 16:47, Ian Rogers wrote:
>>>>> On Fri, May 6, 2022 at 5:26 AM Adrian Hunter <adrian.hunter@intel.com> wrote:
>>>>>>
>>>>>> Add evlist__add_dummy_on_all_cpus() to enable creating a system-wide dummy
>>>>>> event that sets up the system-wide maps before map propagation.
>>>>>>
>>>>>> For convenience, add evlist__add_aux_dummy() so that the logic can be used
>>>>>> whether or not the event needs to be system-wide.
>>>>>>
>>>>>> Signed-off-by: Adrian Hunter <adrian.hunter@intel.com>
>>>>>> ---
>>>>>>  tools/perf/util/evlist.c | 40 ++++++++++++++++++++++++++++++++++++++++
>>>>>>  tools/perf/util/evlist.h |  5 +++++
>>>>>>  2 files changed, 45 insertions(+)
>>>>>>
>>>>>> diff --git a/tools/perf/util/evlist.c b/tools/perf/util/evlist.c
>>>>>> index 78c47cbafbc2..c16bd4836314 100644
>>>>>> --- a/tools/perf/util/evlist.c
>>>>>> +++ b/tools/perf/util/evlist.c
>>>>>> @@ -264,6 +264,46 @@ int evlist__add_dummy(struct evlist *evlist)
>>>>>>         return 0;
>>>>>>  }
>>>>>>
>>>>>> +static void evlist__add_on_all_cpus(struct evlist *evlist, struct evsel *evsel)
>>>>>> +{
>>>>>> +       evsel->core.system_wide = true;
>>>>>> +
>>>>>> +       /* All CPUs */
>>>>>> +       perf_cpu_map__put(evsel->core.own_cpus);
>>>>>> +       evsel->core.own_cpus = perf_cpu_map__new(NULL);
>>>>>> +       perf_cpu_map__put(evsel->core.cpus);
>>>>>> +       evsel->core.cpus = perf_cpu_map__get(evsel->core.own_cpus);
>>>>>> +
>>>>>> +       /* No threads */
>>>>>> +       perf_thread_map__put(evsel->core.threads);
>>>>>> +       evsel->core.threads = perf_thread_map__new_dummy();
>>>>>> +
>>>>>> +       evlist__add(evlist, evsel);
>>>>>> +}
>>>>>> +
>>>>>> +struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide)
>>>>>> +{
>>>>>> +       struct evsel *evsel = evlist__dummy_event(evlist);
>>>>>> +
>>>>>> +       if (!evsel)
>>>>>> +               return NULL;
>>>>>> +
>>>>>> +       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->no_aux_samples = true;
>>>>>> +       evsel->name = strdup("dummy:u");
>>>>>> +
>>>>>> +       if (system_wide)
>>>>>> +               evlist__add_on_all_cpus(evlist, evsel);
>>>>>> +       else
>>>>>> +               evlist__add(evlist, evsel);
>>>>>> +
>>>>>> +       return evsel;
>>>>>> +}
>>>>>> +
>>>>>>  static int evlist__add_attrs(struct evlist *evlist, struct perf_event_attr *attrs, size_t nr_attrs)
>>>>>>  {
>>>>>>         struct evsel *evsel, *n;
>>>>>> diff --git a/tools/perf/util/evlist.h b/tools/perf/util/evlist.h
>>>>>> index 4062f5aebfc1..1bde9ccf4e7d 100644
>>>>>> --- a/tools/perf/util/evlist.h
>>>>>> +++ b/tools/perf/util/evlist.h
>>>>>> @@ -114,6 +114,11 @@ int arch_evlist__add_default_attrs(struct evlist *evlist);
>>>>>>  struct evsel *arch_evlist__leader(struct list_head *list);
>>>>>>
>>>>>>  int evlist__add_dummy(struct evlist *evlist);
>>>>>> +struct evsel *evlist__add_aux_dummy(struct evlist *evlist, bool system_wide);
>>>>>> +static inline struct evsel *evlist__add_dummy_on_all_cpus(struct evlist *evlist)
>>>>>
>>>>> Sorry to be a language lawyer. What I hope to clean up with CPU maps is that:
>>>>>
>>>>> empty == dummy == any CPU
>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/lib/perf/cpumap.c?h=perf/core#n279
>>>>>
>>>>> Given every CPU map should be empty or contain any CPU then it seems
>>>>> they all meet the definition of empty - so something is wrong.
>>>>
>>>> Nothing is wrong.  I am not against clarifying things, but stop assuming
>>>> natural language has to mean anything exactly.  That is what computer
>>>> languages are for.
>>>>
>>>> Sometimes more abstract language is used, precisely to stop people
>>>> making assumptions about the details.
>>>>
>>>>>
>>>>> The cpu map here is explicitly opened so that it gets all online CPUs:
>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git/tree/tools/lib/perf/cpumap.c?h=perf/core#n174
>>>>>
>>>>> From:
>>>>> https://github.com/torvalds/linux/blob/master/Documentation/admin-guide/cputopology.rst
>>>>> there are example topologies like:
>>>>> kernel_max: 31
>>>>>    offline: 2,4-31,32-63
>>>>>     online: 0-1,3
>>>>>   possible: 0-31
>>>>>    present: 0-31
>>>>>
>>>>> all_cpus could mean the union of offline and online CPUs, possible
>>>>> CPUs or present CPUs. You are saying that in the perf code all_cpus
>>>>> should be the same as all online cpus as only those CPUs are valid
>>>>> with perf_event_open. That's true but offline CPUs can be made online.
>>>>> If that happens here then the dummy events will have a CPU map that
>>>>> rather than being for all CPUs will be for all online CPUs at the
>>>>> point it was opened. Having online in the function name I think
>>>>> captures the time dependent nature of this - but if you think that's
>>>>> too much could we add a comment?
>>>>
>>>> If you ask me it does the exact opposite.  The function of the code
>>>> is to put the event on all CPUS without having to know the details
>>>> of: well actually perf doesn't automagically retain or restore events
>>>> across enabling or disabling CPUs so in fact we deal only in online
>>>> CPUs.
>>>
>>> But 'any CPU' (-1) could map to an offline CPU brought online. Calling
>>> this function twice could also result in this behavior. Via the
>>> topology documentation we have language to describe exactly the
>>> scenario that's happening and I'd prefer not to muddy that by making
>>> all and online synonyms.
>>
>> In this case the caller wants all CPUs, not online CPUs.  The detail
>> that we can't trace offline CPUs that become online is not relevant
>> to the caller.  Why would the caller call a function limited to online
>> CPUs when that is not what the caller wants.
> 
> Agreed, the intention is for all CPUs to be traced, so all that can be
> traced should be traced, at the time of the call to this function.
> 
> An improvement, that would change the workings of this function, but
> that would still do what was asked, would be to have functionality that
> remembers such requestgs for tracing all CPUs and when a CPU that was
> offline is brought online, gets that CPU added to whoever asked for all
> CPUs to be traced.
> 
> If this is something critical to the caller, then perhaps it should
> provide a callback for when CPUs are made online (or offline).
> 
> But then we can add that functionality when the need arises?

It should always be possible to figure out which evsels to add a newly
online CPU to.  Roughly:

per-thread evsels CPU map = {-1} => never change since the kernel
dynamically schedules per-task contexts.

If the CPU is a user-requested CPU, add it to all per-cpu evsels.

Otherwise add it to evsels marked as system_wide.

Actually adding it would be a lot more challenging.

> 
> - Arnaldo
>  
>>>>> too much could we add a comment? I'm trying to avoid a situation, like
>>>>> with the CPU map code, where all and online are interchangeable
>>>>> leading to the code being unnecessarily confusing unless you read
>>>>> every line.
>>>>
>>>> It is normal to have to read the details of code, and, in my
>>>> experience at least, normal for the code not to work exactly the
>>>> way I'd imagined.
>>>
>>> :-) The problem is that we all need to work with abstractions at some
>>> point, abstraction is pretty much the whole point of computer science.
>>> We need to fix CPU maps empty function, it is just a fundamental level
>>> of contradiction. As with the CPU map index being often mistaken for
>>> the CPU leading to bugs and crashes, I suspect remedying empty will
>>> fix existing and future bugs. With function naming the point is to be
>>> short and succinct, but also to be intention revealing for the sake of
>>> abstraction. Yes you need to read the code, but as with CPU map empty
>>> even that isn't enough and trying to infer behavior from usage can be
>>> a long and painful process.
>>
>> You seem to be insisting that the function be named for its
>> implementation (i.e. offline CPUs are not supported) not its
>> purpose (trace system wide).
>>
>> I can only suggest we go back to the original name, because the
>> function has *nothing* to do with whether or not perf supports
>> tracing per-cpu contexts on offline CPUs that become online.
>>
>>>
>>> Thanks,
>>> Ian
>>>
>>>>>
>>>>> Thanks,
>>>>> Ian
>>>>>
>>>>>> +{
>>>>>> +       return evlist__add_aux_dummy(evlist, true);
>>>>>> +}
>>>>>>
>>>>>>  int evlist__add_sb_event(struct evlist *evlist, struct perf_event_attr *attr,
>>>>>>                          evsel__sb_cb_t cb, void *data);
>>>>>> --
>>>>>> 2.25.1
>>>>>>
>>>>
>