Postfix 20010228 is OK on FreeBSD 4.2-RELEASE.

I just checked that the system behaves sanely when Postfix does

        connect()
        write()
        close()

before the server has accept()ed the connection.

Reportedly, some later FreeBSD versions introduce a race condition
here. If the server accept()s the connection before the client
close()s it, then the server may receive the data. If the server
accept()s the connection after the client close()s it, the data is
lost. In both cases, the client is told that the data was transmitted
without error.

This change in socket behavior breaks the Postfix flush service,
and probably also breaks other parts of Postfix as well. It all
depends on how quickly a server can call accept(), and thus, it
all depends on how loaded a server is.

This is a bad change.

uname -a output:

    FreeBSD hades.porcupine.org 4.2-RELEASE FreeBSD 4.2-RELEASE
    #5: Sun Mar 4 18:58:52 EST 2001
    wietsehades.porcupine.org:/usr/src/sys/compile/GENERIC i386

        Wietse

Wietse Venema:
> That's it. Postfix connects/writes/disconnects. The changed behavior
> causes loss of data that was already written and acknowledged.
>
> The bad part is that there is no way for the sender to find out.
>
> Wietse
>
> Arjan de Vet:
> > In article <20010306213602.11432BC06Dspike.porcupine.org> you write:
> >
> > >Every 100 seconds means whenever the master 'wakeup' timer goes
> > >off. This means that UNIX-domain socket behavior has changed.
> >
> > I've looked through recent commits to socket code and this one may be
> > the problem (not verified yet, it's time to go to sleep ;-):
> >
> > jlemon 2001/02/13 18:09:12 PST
> >
> > Modified files:
> > sys/kern uipc_socket.c uipc_syscalls.c
> > Log:
> > Return ECONNABORTED from accept if connection is closed while on the
> > listen queue, as well as the current behavior of a zero-length sockaddr.
> >
> > Obtained from: KAME
> > Reviewed by: -net
> >
> > Revision Changes Path
> > 1.88 +3 -6 src/sys/kern/uipc_socket.c
> > 1.85 +15 -2 src/sys/kern/uipc_syscalls.c
> >
> > This change was applied on Feb 24 to FreeBSD 4.2-stable.
> >
> > The change in uipc_socket.c is:
> >
> > -365,11 +365,8
> > so->so_state &= ~SS_NOFDREF;
> > if ((so->so_state & SS_ISDISCONNECTED) == 0)
> > error = (*so->so_proto->pr_usrreqs->pru_accept)(so,
> > nam);
> > - else {
> > - if (nam)
> > - *nam = 0;
> > - error = 0;
> > - }
> > + else
> > + error = ECONNABORTED;
> > splx(s);
> > return (error);
> > }
> >
> > ECONNABORTED is exactly the error I'm seeing now.
> >
> > >At the very least, the kernel could deliver the data that
> > >is send with connect/write/disconnect.
> >
> > Arjan
> >
> > --
> > Arjan de Vet, Eindhoven, The Netherlands <Arjan.deVetadv.iae.nl>
> > URL: http://www.iae.nl/users/devet/ for PGP key: finger devetiae.nl
> >
> >
> >
>
>
>
>

• Messages sorted by: [ date ] [ thread ] [ subject ] [ author ]