[Fwd: 1.5.11-1: sftp performance problem]

Bob Byrnes byrnes@curl.com
Sat Sep 11 17:24:00 GMT 2004

On Sep 11,  3:02pm, Corinna Vinschen wrote:
-- Subject: Re: [Fwd: 1.5.11-1: sftp performance problem]
> http://cygwin.com/acronyms/#PCYMTNQREAIYR

Sorry ... done.

> | One really annoying consequence of socketpairs for sshd is that
> | Windows-native (i.e. non-Cygwin) programs don't know how to write
> | directly to sockets on stdout (or stderr), so if you try to use
> | them via ssh, their output silently disappears.
> Oh lord!  I tried it with `net', `ping' and `nslookup', which are roughly
> the only Windows applications I use on the command line, and all three
> have no problems.  I'm wondering how the above can happen since the
> interactive application side is running in a PTY (which *is* a pipe),
> not on socketpairs.

Try it without a pty:

    ssh cygwin-system-with-sshd-using-sockpairs "win32-native-command"


    ssh cygwin-system-with-sshd-using-sockpairs "win32-native-command | cat"

> | I think OpenSSH uses pipes on most platforms by default, so I'm also
> Only on those platforms which have known problems with (or no implementation
> of) socketpairs.

OK, maybe I am misremembering, or the default changed ages ago.
I see that sshd on Linux uses socketpairs.

> | somewhat concerned that the socketpair-specific code in OpenSSH might be
> | a bit crufty.
> Why?  Socketpairs are in the same shape as any other socket connection
> since they are using the same code.  And I surely do care for socket
> operations.  See the ChangeLogs.  What especially do you think is "crufty"?
-- End of excerpt from Corinna Vinschen

My concern was about the state of the OpenSSH code base, not Cygwin, and
it was just a vague fear, based on the previous (I guess now incorrect)
[mis]understanding that OpenSSH prefers to use pipes.

> | It's not obvious to me why socketpairs would be inherently faster
> | than pipes.

It occurred to me after I wrote this that if no pipes were used at all,
then the select_pipe thread would not be spawned ... the select_socket
thread would handle everything.  This might be win, but we'd need to
measure to see if the effect is significant.

And we'd still have the problem with win32-native apps, which I think
is a show-stopper for socketpairs.


More information about the Cygwin-patches mailing list