Re: [PATCH v2 08/15] perf record: write trace data into mmap trace files

[Alexei Budankov]

On 26.10.2020 13:32, Jiri Olsa wrote:
> On Mon, Oct 26, 2020 at 11:52:21AM +0300, Alexei Budankov wrote:
>>
>> On 24.10.2020 18:44, Jiri Olsa wrote:
>>> On Wed, Oct 21, 2020 at 07:02:56PM +0300, Alexey Budankov wrote:
>>>
>>> SNIP
>>>
>>>>  
>>>>  	record__synthesize(rec, true);
>>>> -	/* this will be recalculated during process_buildids() */
>>>> -	rec->samples = 0;
>>>>  
>>>>  	if (!err) {
>>>>  		if (!rec->timestamp_filename) {
>>>> @@ -2680,9 +2709,12 @@ int cmd_record(int argc, const char **argv)
>>>>  
>>>>  	}
>>>>  
>>>> -	if (rec->opts.kcore)
>>>> +	if (rec->opts.kcore || record__threads_enabled(rec))
>>>>  		rec->data.is_dir = true;
>>>>  
>>>> +	if (record__threads_enabled(rec))
>>>> +		rec->opts.affinity = PERF_AFFINITY_CPU;
>>>
>>> so all the threads will pin to cpu and back before reading?
>>
>> No, they will not back. Thread mask compares to mmap mask before
>> read and the thread migrates if masks don't match. This happens
>> once on the first mmap read. So explicit pinning can be avoided.
> 
> hum, is that right? the check in record__adjust_affinity
> is checking global 'rec->affinity_mask', at lest I assume
> it's still global ;-)

Yes, rec->affinity_mask should also be per-thread. Good catch. Thanks!

Alexei

> 
>         if (rec->opts.affinity != PERF_AFFINITY_SYS &&
>             !bitmap_equal(rec->affinity_mask.bits, map->affinity_mask.bits,
>                           rec->affinity_mask.nbits)) {
> 
> I think this can never be equal if you have more than one map
> 
> when I check on sched_setaffinity syscalls:
> 
>   # perf trace -e syscalls:sys_enter_sched_setaffinity
> 
> while running record --threads, I see sched_setaffinity
> calls all the time
> 
> jirka
>