Re: [PATCH v8 01/13] tools/libperf: avoid moving of fds at fdarray__filter() call

[Alexey Budankov]

On 26.06.2020 12:37, Jiri Olsa wrote:
> On Thu, Jun 25, 2020 at 10:32:29PM +0300, Alexey Budankov wrote:
>>
>> On 25.06.2020 20:14, Jiri Olsa wrote:
>>> On Wed, Jun 24, 2020 at 08:19:32PM +0300, Alexey Budankov wrote:
>>>>
>>>> On 17.06.2020 11:35, Alexey Budankov wrote:
>>>>>
>>>>> Skip fds with zeroed revents field from count and avoid fds moving
>>>>> at fdarray__filter() call so fds indices returned by fdarray__add()
>>>>> call stay the same and can be used for direct access and processing
>>>>> of fd revents status field at entries array of struct fdarray object.
>>>>>
>>>>> Signed-off-by: Alexey Budankov <alexey.budankov@xxxxxxxxxxxxxxx>
>>>>> ---
>>>>>  tools/lib/api/fd/array.c   | 11 +++++------
>>>>>  tools/perf/tests/fdarray.c | 20 ++------------------
>>>>>  2 files changed, 7 insertions(+), 24 deletions(-)
>>>>>
>>>>> diff --git a/tools/lib/api/fd/array.c b/tools/lib/api/fd/array.c
>>>>> index 58d44d5eee31..97843a837370 100644
>>>>> --- a/tools/lib/api/fd/array.c
>>>>> +++ b/tools/lib/api/fd/array.c
>>>>> @@ -93,22 +93,21 @@ int fdarray__filter(struct fdarray *fda, short revents,
>>>>>  		return 0;
>>>>>  
>>>>>  	for (fd = 0; fd < fda->nr; ++fd) {
>>>>> +		if (!fda->entries[fd].revents)
>>>>> +			continue;
>>>>> +
>>>>
>>>> So it looks like this condition also filters out non signaling events fds, not only
>>>> control and others fds, and this should be somehow avoided so such event related fds
>>>> would be counted. Several options have been proposed so far:
>>>>
>>>> 1) Explicit typing of fds via API extension and filtering based on the types:
>>>>    a) with separate fdarray__add_stat() call
>>>>    b) with type arg of existing fdarray__add() call
>>>>    c) various memory management design is possible
>>>>
>>>> 2) Playing tricks with fd positions inside entries and assumptions on fdarray API calls ordering
>>>>    - looks more like a hack than a designed solution
>>>>
>>>> 3) Rewrite of fdarray class to allocate separate object for every added fds
>>>>    - can be replaced with nonscrewing of fds by __filter()
>>>>
>>>> 4) Distinct between fds types at fdarray__filter() using .revents == 0 condition
>>>>    - seems to have corner cases and thus not applicable
>>>>
>>>> 5) Extension of fdarray__poll(, *arg_ptr, arg_size) with arg of fds array to atomically poll
>>>>    on fdarray_add()-ed fds and external arg fds and then external arg fds processing
>>>>
>>>> 6) Rewrite of fdarray class on epoll() call basis
>>>>    - introduces new scalability restrictions for Perf tool
>>>
>>> hum, how many fds for polling do you expect in your workloads?
>>
>> Currently it is several hundreds so default of 1K is easily hit and 
>> "Profile a Large Number of PMU Events on Multi-Core Systems" section [1]
>> recommends:
>>
>> soft nofile 65535
>> hard nofile 65535
> 
> I'm confused, are you talking about file descriptors limit now?
> this wont be affected by epoll change.. what do I miss?

Currently there is already uname -n limit on the amount of open file descriptors
and Perf tool process is affected by that limit.

Moving to epoll() will impose one more max_user_watches limit and that can additionally
confine Perf applicability even though default value on some machines currently
is high enough.

~Alexey

> 
> I thought your concern was fs.epoll.max_user_watches, which has
> default value that seems to be enough:
> 
> 	$ cat /proc/sys/fs/epoll/max_user_watches
> 	3169996
> 
> jirka
> 
> 
>>
>> for for /etc/security/limits.conf settings.
>>
>> ~Alexey
>>
>> [1] https://software.intel.com/content/www/us/en/develop/documentation/vtune-cookbook/top/configuration-recipes/profiling-hardware-without-sampling-drivers.html
>>
>>>
>>> jirka
>>>
>>
>