Re: recvmmsg() timeout behavior strangeness [RESEND]

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

Em Mon, May 12, 2014 at 12:15:25PM +0200, Michael Kerrisk (man-pages) escreveu:
> Hi Arnaldo,
> 
> Ping!

I acknowledge the problem, the timeout has to be passed to the
underlying ->recvmsg() implementations that should return the time spent
waiting for each packet, so that we can accrue that at recvmmsg level.

We can do either passing an extra timeout parameter to the recvmsg
implementations or using some struct sock member to specify that
timeout.

The first approach is intrusive, touches tons of files, so I'll try
making it all mostly transparent by hooking into sock_rcvtimeo()
somehow.

- Arnaldo
 
> Cheers,
> 
> Michael
> 
> 
> On Wed, Apr 30, 2014 at 3:59 PM, Michael Kerrisk (man-pages)
> <mtk.manpages@gmail.com> wrote:
> > Arnaldo,
> >
> > I raised this issue somewhat more than a year ago, here:
> > http://thread.gmane.org/gmane.linux.man/3477
> > but got no reply from you. (Chris Friesen in that thread agreed
> > that there is a problem though.)
> >
> > Here, a slightly revised version of that mail, since I've just bumper
> > into a related problem in a different context...
> >
> > As part of his attempt to better document the recvmmsg() syscall that
> > you added in commit a2e2725541fad72416326798c2d7fa4dafb7d337, Elie de
> > Brauwer alerted to me to some strangeness in the timeout behavior of
> > the syscall. I suspect there's a bug that needs fixing, as detailed
> > below.
> >
> > AFAICT, the timeout argument was added to this syscall as a result of
> > the discussion here:
> > http://markmail.org/message/m5l2ap4hiiimut6k#query:+page:1+mid:m5l2ap4hiiimut6k+state:results
> > (20-21 May 2009, "[RFC 1/2] net: Introduce recvmmsg...")
> >
> > If I understand correctly, the *intended* purpose of the timeout
> > argument is to set a limit on how long to wait for additional
> > datagrams after the arrival of an initial datagram. However, the
> > syscall behaves in quite a different way. Instead, it potentially
> > blocks forever, regardless of the timeout. The way the timeout seems
> > to work is as follows:
> >
> > 1. The timeout, T, is armed on receipt of first diagram, starting at time X.
> > 2. After each further datagram is received, a check is made if we have
> > reached time X+T. If we have reached that time, then the syscall
> > returns.
> >
> > Since the timeout is only checked after the arrival of each datagram,
> > we can have scenarios like the following:
> >
> > 0. Assume a timeout of 10 seconds, and that vlen is 5.
> > 1. First datagram arrives at time X.
> > 2. Second datagram arrives at time X+2 secs
> > 3. No more datagrams arrive.
> >
> > In this case, the call blocks forever. Is that intended behavior?
> > (Basically, if up to vlen-1 datagrams arrive before X+T, but then no
> > more datagrams arrive, the call will remain blocked forever.) If it's
> > intended behavior, could you elaborate the use case, since it would be
> > good to add that to the man page. If not, a fix seems to be needed,
> > since otherwise, it's hard to see how the recvmmsg() timeout argument
> > can sanely be used.
> >
> > Thanks,
> >
> > Michael
> >
> > --
> > Michael Kerrisk
> > Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
> > Linux/UNIX System Programming Training: http://man7.org/training/
> 
> 
> 
> -- 
> Michael Kerrisk
> Linux man-pages maintainer; http://www.kernel.org/doc/man-pages/
> Linux/UNIX System Programming Training: http://man7.org/training/