Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <4763358a-91fa-3607-dd54-749b40f977fc@intel.com>
Date:   Wed, 7 Sep 2022 08:05:27 +0300
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:91.0) Gecko/20100101
 Firefox/91.0 Thunderbird/91.11.0
Subject: Re: [PATCH V2] libperf evlist: Fix per-thread mmaps for
 multi-threaded targets
Content-Language: en-US
To:     Jiri Olsa <olsajiri@gmail.com>
Cc:     Arnaldo Carvalho de Melo <acme@kernel.org>,
        Jiri Olsa <jolsa@redhat.com>,
        Namhyung Kim <namhyung@kernel.org>,
        Ian Rogers <irogers@google.com>, linux-kernel@vger.kernel.org,
        linux-perf-users@vger.kernel.org,
        Peter Zijlstra <peterz@infradead.org>,
        Ingo Molnar <mingo@redhat.com>
References: <20220905114209.8389-1-adrian.hunter@intel.com>
 <YxdEQkmkZjBC6wGb@krava> <60b5c9bf-4ec9-957e-17dd-aa0a50411ff9@intel.com>
 <YxejRK2/LO7QouOZ@krava>
From:   Adrian Hunter <adrian.hunter@intel.com>
Organization: Intel Finland Oy, Registered Address: PL 281, 00181 Helsinki,
 Business Identity Code: 0357606 - 4, Domiciled in Helsinki
In-Reply-To: <YxejRK2/LO7QouOZ@krava>
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 7bit
Precedence: bulk

On 6/09/22 22:45, Jiri Olsa wrote:
> On Tue, Sep 06, 2022 at 05:04:45PM +0300, Adrian Hunter wrote:
>> On 6/09/22 15:59, Jiri Olsa wrote:
>>> On Mon, Sep 05, 2022 at 02:42:09PM +0300, Adrian Hunter wrote:
>>>
>>> SNIP
>>>
>>>> diff --git a/tools/lib/perf/evlist.c b/tools/lib/perf/evlist.c
>>>> index e6c98a6e3908..6b1bafe267a4 100644
>>>> --- a/tools/lib/perf/evlist.c
>>>> +++ b/tools/lib/perf/evlist.c
>>>> @@ -486,6 +486,7 @@ mmap_per_evsel(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
>>>>  			if (ops->idx)
>>>>  				ops->idx(evlist, evsel, mp, idx);
>>>>  
>>>> +			pr_debug("idx %d: mmapping fd %d\n", idx, *output);
>>>>  			if (ops->mmap(map, mp, *output, evlist_cpu) < 0)
>>>>  				return -1;
>>>>  
>>>> @@ -494,6 +495,7 @@ mmap_per_evsel(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
>>>>  			if (!idx)
>>>>  				perf_evlist__set_mmap_first(evlist, map, overwrite);
>>>>  		} else {
>>>> +			pr_debug("idx %d: set output fd %d -> %d\n", idx, fd, *output);
>>>>  			if (ioctl(fd, PERF_EVENT_IOC_SET_OUTPUT, *output) != 0)
>>>>  				return -1;
>>>>  
>>>> @@ -519,6 +521,48 @@ mmap_per_evsel(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
>>>>  	return 0;
>>>>  }
>>>>  
>>>> +static int
>>>> +mmap_per_thread(struct perf_evlist *evlist, struct perf_evlist_mmap_ops *ops,
>>>> +		struct perf_mmap_param *mp)
>>>> +{
>>>> +	int nr_threads = perf_thread_map__nr(evlist->threads);
>>>> +	int nr_cpus    = perf_cpu_map__nr(evlist->all_cpus);
>>>> +	int cpu, thread, idx = 0;
>>>> +	int nr_mmaps = 0;
>>>> +
>>>> +	pr_debug("%s: nr cpu values (may include -1) %d nr threads %d\n",
>>>> +		 __func__, nr_cpus, nr_threads);
>>>
>>> -1 as cpu value is only for 'empty' perf_cpu_map, right?
>>
>> The cpu map is a map of valid 3rd arguments to perf_event_open, so -1
>> means all CPUs which is per-thread by necessity.
>>
>>>
>>>> +
>>>> +	/* per-thread mmaps */
>>>> +	for (thread = 0; thread < nr_threads; thread++, idx++) {
>>>> +		int output = -1;
>>>> +		int output_overwrite = -1;
>>>> +
>>>> +		if (mmap_per_evsel(evlist, ops, idx, mp, 0, thread, &output,
>>>> +				   &output_overwrite, &nr_mmaps))
>>>> +			goto out_unmap;
>>>> +	}
>>>> +
>>>> +	/* system-wide mmaps i.e. per-cpu */
>>>> +	for (cpu = 1; cpu < nr_cpus; cpu++, idx++) {
>>>> +		int output = -1;
>>>> +		int output_overwrite = -1;
>>>> +
>>>> +		if (mmap_per_evsel(evlist, ops, idx, mp, cpu, 0, &output,
>>>> +				   &output_overwrite, &nr_mmaps))
>>>> +			goto out_unmap;
>>>> +	}
>>>
>>> will this loop be executed? we are in here because all_cpus is empty, right?
>>
>> Yes it is executed.  I put back the code that was there before ae4f8ae16a07
>> ("libperf evlist: Allow mixing per-thread and per-cpu mmaps"), which uses
> 
> hm, but commit ae4f8ae16a07 does not have similar cpu loop

It was calling mmap_per_cpu() for that case.

The 2 cases: mmap_per_cpu() and mmap_per_thread() could still be
combined into a single function.

> 
>> perf_cpu_map__empty() which only checks the first entry is -1:
>>
>> bool perf_cpu_map__empty(const struct perf_cpu_map *map)
>> {
>> 	return map ? map->map[0].cpu == -1 : true;
>> }
>>
>> But there can be more CPUs in the map, so perf_cpu_map__empty()
>> returns true for the per-thread case, as desired, even if there
>> are also system-wide CPUs.
> 
> I don't see how, if I'd see -1 together with other cpu values in
> perf_cpu_map I'd think it's a bug, but I might be missing some
> auxtrace usage,

Yes, it is for system-wide collection of events that can affect
every CPU.  Currently text_poke is always system-wide - see the
Intel PT example.

> 
> I thought we use -1 just for empty cpu map, so in per-thread case
> -1 is properly passed to perf_event_open syscall

Yes, but it does not need to be limited to that case.

> 
> jirka
> 
>>
>> I guess perf_cpu_map__empty() needs renaming.
>>
>>>
>>> thanks,
>>> jirka
>>>
>>>> +
>>>> +	if (nr_mmaps != evlist->nr_mmaps)
>>>> +		pr_err("Miscounted nr_mmaps %d vs %d\n", nr_mmaps, evlist->nr_mmaps);
>>>> +
>>>> +	return 0;
>>>> +
>>>> +out_unmap:
>>>> +	perf_evlist__munmap(evlist);
>>>> +	return -1;
>>>> +}
>>>
>>> SNIP
>>