Date:	Tue, 26 Feb 2008 20:14:52 +0100
From:	"Michael Kerrisk" <mtk.manpages@...glemail.com>
To:	"Davide Libenzi" <davidel@...ilserver.org>
Cc:	"Pierre Habouzit" <madcoder@...ian.org>,
	lkml <linux-kernel@...r.kernel.org>,
	"Eric Dumazet" <dada1@...mosbay.com>,
	"Marc Lehmann" <schmorp@...morp.de>,
	"David Schwartz" <davids@...master.com>
Subject: Re: epoll and shared fd's

On Tue, Feb 26, 2008 at 8:04 PM, Davide Libenzi <davidel@...ilserver.org> wrote:
>
> On Tue, 26 Feb 2008, Michael Kerrisk wrote:
>
> > Following up after quite some time:
> >
> > Davide Libenzi wrote:
> > > On Sat, 26 Jan 2008, Michael Kerrisk wrote:
> > >
> > >> On Jan 25, 2008 12:57 AM, Davide Libenzi <davidel@...ilserver.org> wrote:
> > >>> On Thu, 24 Jan 2008, Pierre Habouzit wrote:
> > >>>
> > >>>> On Fri, Jan 18, 2008 at 09:10:18PM +0000, Davide Libenzi wrote:
> > >>>>> On Fri, 18 Jan 2008, Pierre Habouzit wrote:
> > >>>>>
> > >>>>>>   Hi,
> > >>>>>>
> > >>>>>>   I just came across a strange behavior of epoll that seems to
> > >>>>>> contradict the documentation. Here is what happens:
> > >>>>>>
> > >>>>>> * I have two processes P1 and P2, P1 accept()s connections, and send the
> > >>>>>>   resulting file descriptors to P2 through a unix socket.
> > >>>>>>
> > >>>>>> * P2 registers the received socket in his epollfd.
> > >>>>>>
> > >>>>>>   [time passes]
> > >>>>>>
> > >>>>>> * P2 is done with the socket and closes it
> > >>>>>>
> > >>>>>> * P2 gets events for the socket again !
> > >>>>>>
> > >>>>>>
> > >>>>>>   Though the documentation says that if a process closes a file
> > >>>>>> descriptor, it gets unregistered. And yes I'm sure that P2 doens't dup()
> > >>>>>> the file descriptor. Though (because of a bug) it was still open in
> > >>>>>> P1[0], hence the referenced socket still live at the kernel level.
> > >>>>>>
> > >>>>>>   Of course the userland workaround is to force the EPOLL_CTL_DEL before
> > >>>>>> the close, which I now do, but costs me a syscall where I wanted to
> > >>>>>> spare one :|
> > >>>>> For epoll, a close is when the kernel file* is released (that is, when all
> > >>>>> its instances are gone).
> > >>>>> We could put a special handling in filp_close(), but I don't think is a
> > >>>>> good idea, and we're better live with the current behaviour.
> > >>>>   Okay, maybe updating the linux manpages to be more clear about that is
> > >>>> the way to go then. Thanks
> > >>> Sure. I'll send Michael Kerrisk and updated statement for the A6 answer in
> > >>> the epoll man page.
> > >> Thanks Davide -- yes please send me a patch.
> > >> --
> > >> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> > >> the body of a message to majordomo@...r.kernel.org
> > >> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> > >> Please read the FAQ at  http://www.tux.org/lkml/
> > >>
> > >
> > > Something like the one below ...
> > >
> > >
> > > - Davide
> > >
> > >
> > >
> > > --- epoll.4 2008-01-26 12:58:21.000000000 -0800
> > > +++ epoll.4.new     2008-01-26 13:06:36.000000000 -0800
> > > @@ -285,7 +285,19 @@
> > >  sets automatically?
> > >  .TP
> > >  .B A6
> > > -Yes.
> > > +A file descriptor is the userspace counterpart of an internal kernel handle.
> > > +Every time a process calls functions liks
> > > +.BR dup (2),
> > > +.BR dup2 (2)
> > > +or
> > > +.BR fork (2),
> > > +a new file descriptor referring to the same internal kernel handle is
> > > +created. The internal kernel handle remains alive until all the userspace
> > > +file descriptors have been closed.
> > > +The
> > > +.BR epoll (4)
> > > +interface automatically removes the internal kernel handle from the set,
> > > +once all the file descriptor instances have been closed.
> > >  .TP
> > >  .B Q7
> > >  If more than one event occurs between
> >
> > Davide,
> >
> > Two points.
> >
> > a) I did a
> >
> > s/internal kernel handle/open file description/
> >
> > since that is the POSIX term for the internal handle.
> >
> > b) It seems to me that you text doesn't quite make the point explicit
> > enough.  I've tried to rewrite it; could you please check:
> >
> >        A6     Yes, but be aware of the following point.  A  file
> >               descriptor is a reference to an open file descrip-
> >               tion (see  open(2)).   Whenever  a  descriptor  is
> >               duplicated  via dup(2), dup2(2), fcntl(2) F_DUPFD,
> >               or fork(2), a new file descriptor referring to the
> >               same  open  file  description is created.  An open
> >               file description continues to exist until all file
> >               descriptors referring to it have been closed.  The
> >               epoll  interface  automatically  removes  a   file
> >               descriptor  from  an  epoll set only after all the
> >               file descriptors referring to the underlying  open
> >               file  handle  have  been  closed.  This means that
> >               even after a file descriptor that is  part  of  an
> >               epoll  set has been closed, events may be reported
> >               for that file descriptor if other file descriptors
> >               referring  to the same underlying file description
> >               remain open.
> >
> > Does that seem okay?  I plan to include the text in man-pages-2.79.
>
> I agree with Bodo, it is kinda confusing. The name "open file description",
> even though POSIX, looks very similar to "file descriptor".
> I honestly don't know how more easily such concept could be expressed.
> IMHO at least "internal kernel handle" does not play look-alike games with
> "file descriptor".

Okay -- I'll look at it some more.  I am however loathe to drop the
term open file description, because POSIX uses, as well as a number of
other Linux man pages by now.

> > Was there some reason why removing a file descriptor couldn't have been
> > made to do the "expected" thing (i.e., remove notifications for that file
> > descriptor, regardless of whether the underlying file description remains
> > open)?
>
> That'd mean placing an eventpoll custom hook into sys_close(). Looks very
> bad to me, and probably will look even worse to other kernel folks.
> Is not much a performance issue (a check to see if a file* is an eventpoll
> file is as easy as comparing the f_op pointer), but a design/style issue.
> On top of that, the interface is already out by many years, so changing it
> will like going to cause problems.

Oh -- I wasn't suggesting we could make the change now -- it would
break the ABI and all that.  I was just wondering why the decision
wasn't made to do it the other way to begin with.  The existing
semantics are somewhat couterintuitive, and potentially interact
libraries that do private manipulations with file descriptors.

Cheers,

Michael
-- 
Michael Kerrisk
Maintainer of the Linux man-pages project
http://www.kernel.org/doc/man-pages/
Want to report a man-pages bug?  Look here:
http://www.kernel.org/doc/man-pages/reporting_bugs.html
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/

Powered by blists - more mailing lists