Victor Duchovni:
> On Mon, Dec 01, 2008 at 11:58:42AM -0500, Charles Marcus wrote:
>
> > On 12/1/2008 11:54 AM, Victor Duchovni wrote:
> > > There is nothing wrong with lost connections after QUIT. Newer versions
> > > of Postfix only log "lost connection" in the SMTP server during data
> > > transfer or when sending the "." response. The client is free to
> > > disconnect without "QUIT" at all other SMTP protocol stages.
> > >
> > > Sufficiently new Postfix releases will not log this condition.
> >
> > Hmmm...
> >
> > I'm running 2.5.5, and get this almost every time (maybe every time)
> > when people send through the webmail interface...
> >
>
> Sorry, Postfix won't log clients disconnecting without sending QUIT,
> but it will log failure to send "221 ...".
>
> The reason is that with PIPELINEd ESMTP, the "250 ..." response to
> "." and "221" response to QUIT are often sent in the same I/O operation,
> so it is appropriate to report I/O errors when sending QUIT, at least
> when there are previous responses in the output buffer. Postfix 2.3+
> complains about problems flushing QUIT unconditionally.
>
> quit_cmd(SMTPD_STATE *state, int unused_argc, SMTPD_TOKEN *unused_argv)
> {
>
> /*
> * Don't bother checking the syntax.
> */
> smtpd_chat_reply(state, "221 2.0.0 Bye");
>
> /*
> * When the "." and quit replies are pipelined, make sure they are
> * flushed now, to avoid repeated mail deliveries in case of a crash in
> * the "clean up before disconnect" code.
> *
> * XXX When this was added in Postfix 2.1 we used vstream_fflush(). As
> * of Postfix 2.3 we use smtp_flush() for better error reporting.
> */
> smtp_flush(state->client);
> return (0);
> }
>
> perhaps the flush should be suppressed if there was no pending unwritten
> data in the client vstream buffer prior to the "221 2.0.0 Bye" reply.
Postfix has a vstream_peek() function to count the amount of buffered
input, but there is as of yet no API to count the amount of buffered
output.
I am not sure it is safe to overload vstream_peek() for this purpose,
because that would break with full-duplex VSTREAMs when the last
operation on the VSTREAM was a write.
Wietse
