Richard Stallman <[EMAIL PROTECTED]> writes:
> However, using --cvs-workaround=e seems to fix it. Again I tried
> three times, and it worked all three times.
That's good news.
> Niels, what precisely does --cvs-workaround=e do that differs from the
> usual case?
The "usual case" is that lsh puts its stderr file descriptor into
non-blocking mode. With --cvs-workaround=e, it doesn't do that.
Instead, it creates a pipe, forks off an extra process that copies
data from the readable end of the pipe to the real stderr. lsh then
puts the writeable end of the pipe into non-blocking mode and uses
that fd instead of the real stderr.
Data written to stderr includes
1. lsh's own diagnostic messages. There's no reason to use
non-blocking mode for this, the writes are actually performed in a
blocking way regardless.
2. Data from the stderr of the remote process. This is written using
lsh's central select/poll-loop, and is properly flow-controlled. If
writing to stderr would block, at most one window of data is
buffered, and the remote end is told not to send any more.
Flow-control is a central feature of the "channel" mechanism in the
ssh protocol.
/Niels
