On Tue, Sep 8, 2009 at 12:42 PM, Michael Burk<burkml AT gmail DOT com> wrote:
> Sorry - here they are.

OK, I don't see anything funny there.  I think it's finally time to
take this to the OpenBSD list and see if they can find anything in
this wilderness.  It might help to provide a pointer to the archive of
this (now *very* long) thread.

I took a look at the kernel source.. EAGAIN only appears in a few
places, and from what I can tell is adequately protected by
O_NONBLOCK/FNONBLOCK in sys_pipe.c.  Looking at sys_fcntl, the F_GETFL
path is very simple and doesn't do any kind of caching that might
change behavior.

At this point, I'm guessing that this may be a race condition, and
that the fcntl() calls are just enough to change the winner of the
race.  File flags are global to the open file -- they are not
duplicated on a fork, and do not track with a particular process.  So,
hypothetically, if amandad opens a file descriptor, hands it off to
sendbackup, and then sets it to nonblocking mode, then we would see
this behavior.  The only wrinkle in this theory is that amandad very
quickly closes its copy of the file, and doesn't ever set it to
nonblocking mode.

At this point, I'm out of options.

Dustin

-- 
Open Source Storage Engineer
http://www.zmanda.com