On Tue, Sep 8, 2009 at 10:20 PM, Nathan Stratton
Treadway<nathanst AT ontko DOT com> wrote:
> At that time the discussion was focused on common-src/stream.c, which
> hadn't changed significantly between those versions, but it would be
> interesting to know if there were any changes in the sendbackup code
> path after 2.5.0p1 that seem like they might be related to this bug....

Sigh, that reaches way back into the CVS days and Jean-Louis' memory
-- he's not around right now.  There were *bunches* of changes between
2.5.0 and 2.5.2, not least of which the introduction of the Backup API
(which became the Application API).  If we could reproduce the
behavior in a test program, exploring the history might be a more
fruitful avenue for pursuit, since we'd know what to look for.

One thing we could do is remove all of the other patches and replace
the "fixing" fcntl() calls with something similar -- maybe just
sleep(1)?  If this "fixes" the problem, then that's evidence this was
a race condition.

Alternately, remove all of the other patches and add:

1285     ptr = buffer;
1286     bytes_written = 0;

g_debug("fcntl(%d) => %d", fileno(pipe_fp), fcntl(fileno(pipe_fp),
F_GETFL, 0), O_NONBLOCK);

1287     just_written = full_write(fileno(pipe_fp), ptr, (size_t)bytes_read);
1288     if (just_written < (size_t)bytes_read) {

and try to trigger the same error.  I suspect you'll find that the
fcntl is returning something with bit 2 (O_NONBLOCK) set.

Dustin

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