Re: [PATCH 04/10] perf record: Filter out POLLHUP'ed file descriptors

[Arnaldo Carvalho de Melo <acme@kernel.org>]

Em Mon, Sep 08, 2014 at 12:38:24PM -0300, Arnaldo Carvalho de Melo escreveu:
> Em Mon, Sep 08, 2014 at 05:10:17PM +0200, Jiri Olsa escreveu:
> > On Mon, Sep 08, 2014 at 11:33:17AM -0300, Arnaldo Carvalho de Melo wrote:
> > > > IMO it's more clear to poll pm all event FDs.. and now with the
> > > > case Adrian described it seems necessary anyway
> 
> > > I would have to check why was that we were polling just the one where
> > > the mmap is done, I don't recall being the one to do it, probably who
> > > did it thought that since the ring buffer is there, it was enough (and
> > > possibly scaled better, dunno) to do the polling in just one of them.
>  
> > for read notification it's ok to poll just for one event, because they
> > all share same ringbuffer and perf_poll checks if there's ANY new data
>  
> > for the hup notification I think we need to poll all of them
> 
> Yeah, I'm convinced of this, I'm working on a patch to make it look at
> all file descriptors at poll time.

Done, its at:

https://git.kernel.org/cgit/linux/kernel/git/acme/linux.git/commit/?h=perf/fdarray.v2&id=46eb798f12d41093deb743b9df10fec23ffae719

this is in a refreshed branch so that it gets in the right time in the
patch kit

Comment:

-----
  perf evlist: We need to poll all event file descriptors
    
  Because we want to notice when they get POLLHUP'ed, so that we can
  figure out when all threads exited in a workload being monitored.
    
  We can't just monitor the fds that were mmaped, we need to notice when
  all the fds that were PERF_EVENT_IOC_SET_OUTPUT'ed too, because the mmap
  stays even after the fd that originally was used to do the mmap call
  went away, its only when all the set-output fds for a mmap are gone that
  the mmap is.
-----

Full branch is available at:

git://git.kernel.org/pub/scm/linux/kernel/git/acme/linux.git perf/fdarray.v2

Web interface:

https://git.kernel.org/cgit/linux/kernel/git/acme/linux.git/log/?h=perf/fdarray.v2
 
> I'm doing it on top of a patch that will close the mmap when it gets a
> POLLHUP, as discussed recently on this thread, in a response I gave to
> Adrian.

> But since multiple fds share an mmap, we can only close that mmap when
> all fds are HUPed, i.e. we need to reference count struct perf_mmap,
> which is what I am doing.
 
> I.e. the fist mmaps and sets perf_mmap.nfds to 1, the next one, just
> after doing that PERF_EVENT_IOC_SET_OUTPUT will bump perf_mmap.nfds,
> unmap gets replaced by mmap_put, that decs and calls munmap when it
> hits zero, yadda, yadda.

Should be ok, now, with the above patch (kinda one liner, just moving
the add_pollfd call) + the patches at the top of the branch that
introduce the mmap refcounting and the fdarray facility to store some
info (what I think you called poll_item in your patchset) associated to
each entry in the file descriptor array).

Please let me know if you can take a look at the branch via the web
interface to see if that solves the problems you reported before I
repost them or if you prefer that I repost so that you can check via
e-mail.

I will repost it one final time, after lunch ;-)

- Arnaldo