Re: [PATCH v7 01/13] tools/libperf: introduce notion of static polled file descriptors

From: Jiri Olsa <jolsa@redhat.com>
To: Alexey Budankov <alexey.budankov@linux.intel.com>
Cc: Arnaldo Carvalho de Melo <acme@kernel.org>,
	Namhyung Kim <namhyung@kernel.org>,
	Alexander Shishkin <alexander.shishkin@linux.intel.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Ingo Molnar <mingo@redhat.com>, Andi Kleen <ak@linux.intel.com>,
	linux-kernel <linux-kernel@vger.kernel.org>
Subject: Re: [PATCH v7 01/13] tools/libperf: introduce notion of static polled file descriptors
Date: Mon, 22 Jun 2020 14:11:20 +0200	[thread overview]
Message-ID: <20200622121120.GA2584593@krava> (raw)
In-Reply-To: <cad3d9a6-da28-c627-de73-17169a7c36a1@linux.intel.com>

On Mon, Jun 22, 2020 at 01:50:03PM +0300, Alexey Budankov wrote:
> 
> On 22.06.2020 13:21, Jiri Olsa wrote:
> > On Mon, Jun 22, 2020 at 12:47:19PM +0300, Alexey Budankov wrote:
> > 
> > SNIP
> > 
> >>>>>>>> fdarray__del(array, fdkey);
> >>>>>>>
> >>>>>>> I think there's solution without having filterable type,
> >>>>>>> I'm not sure why you think this is needed
> >>>>>>>
> >>>>>>> I'm busy with other things this week, but I think I can
> >>>>>>> come up with some patch early next week if needed
> >>>>>>
> >>>>>> Friendly reminder.
> >>>>>
> >>>>> hm? I believe we discussed this in here:
> >>>>>   https://lore.kernel.org/lkml/20200609145611.GI1558310@krava/
> >>>>
> >>>> Do you want it to be implemented like in the patch posted by the link?
> >>>
> >>> no idea.. looking for good solution ;-)
> >>>
> >>> how about switching completely to epoll? I tried and it
> >>> does not look that bad
> >>
> >> Well, epoll() is perhaps possible but why does it want switching to epoll()?
> >> What are the benefits and/or specific task being solved by this switch? 
> > 
> > epoll change fixes the same issues as the patch you took in v8
> > 
> > on top of it it's not a hack and wil make polling more user
> > friendly because of the clear interface
> 
> Clear. The opposite thing is /proc/sys/fs/epoll/max_user_watches limit that
> will affect Perf tool usage additionally to the current process limit on 
> a number of simultaneously open file descriptors (ulimit -n). So move to 
> epoll() will impose one limit what can affect Perf tool scalability.

hum, I dont think this will be a problem:

    Allowing top 4% of low memory (per user) to be allocated in epoll watches,
    we have:

    LOMEM    MAX_WATCHES (per user)
    512MB    ~178000
    1GB      ~356000
    2GB      ~712000

my laptop has 19841945 allowed watches per user

> 
> > 
> >>
> >>>
> >>> there might be some loose ends (interface change), but
> >>> I think this would solve our problems with fdarray
> >>
> >> Your first patch accomodated in v8 actually avoids fds typing
> >> and solves pos (=fdarray__add()) staleness issue with fdarray.
> > 
> > yea, it was a change meant for discussion (which never happened),
> > and I considered it to be more a hack than a solution
> > 
> > I suppose we can live with that for a while, but I'd like to
> > have clean solution for polling as well
> 
> I wouldn't treat it as a hack but more as a fix because returned
> pos is now a part of interface that can be safely used in callers.
> Can we go with this fix for the patch set?

apart from this one I still have a problem with that stat factoring
having 1 complicated function deal with both fork and no fork processing,
which I already commented on, but you ignored ;-)

I'll try to go through that once more, and post some comments

jirka