Wietse Venema:
> Postfix VSTREAMs automatically flush output on the next read
> operation; a lot of things depend on this, including the SMTP client
> and SMTP server protocol implementations. This is how Postfix avoids
> sending silly little network packets.
>
> In the case of skipping Milter replies, the idea is that queued
> SMFIC_RCPT messages will be eventually be flushed when Postfix
> reads a response from the Milter application; normally that
> would be at or before the end-of-body reply.
>
> I am surprised that the VSTREAM doesn't flush SMFIC_RCPT as it
> should; normally one has to jump ugly hoops to lose output like
> that.
My hypothesis is as follows: the automatic flush-before-read did
not happen, because the write and the read operations were done
in different processes (and therefore, on different VSTREAMs).
Specifically, the VSTREAM in the smtpd process did not automatically
flush RCPT requests to the Milter socket, because the next read
request on that Milter socket was done on a different VSTREAM in
a cleanup process.
Please see if the patch below solves the problem.
Wietse
*** ./milter8.c- Sat Jul 11 20:27:15 2009
--- ./milter8.c Fri Sep 18 16:38:11 2009
***************
*** 2584,2589 ****
--- 2584,2596 ----
if (msg_verbose)
msg_info("%s: milter %s", myname, milter->m.name);
+ /*
+ * The next read on this Milter socket happens in a different process. It
+ * will not automatically flush the output buffer in this process.
+ */
+ if (milter->fp)
+ vstream_fflush(milter->fp);
+
if (attr_print(stream, ATTR_FLAG_MORE,
ATTR_TYPE_STR, MAIL_ATTR_MILT_NAME, milter->m.name,
ATTR_TYPE_INT, MAIL_ATTR_MILT_VERS, milter->version,
