On Wed, 2010-08-04 at 03:29 +0100, Colin Watson wrote: > On Wed, Aug 04, 2010 at 12:36:00AM +0100, Ben Hutchings wrote: > > On Wed, 2010-08-04 at 00:13 +0100, Colin Watson wrote: > > > On Mon, Jul 26, 2010 at 03:43:47AM +0100, Ben Hutchings wrote: > > > > This seems to be a race condition in debconf, not a bug in the postinst > > > > script (which doesn't do anything very interesting). I was able to > > > > reproduce it only once. The debconf frontend was blocked on read while > > > > the postinst script had exited and was a zombie. > > > > > > > > (I don't really understand this behaviour as read on a pipe should > > > > return as soon as the process at the other end of the pipe exits. So it > > > > could even be a bug in the kernel, though that seems unlikely.) > > > > > > I doubt that this is the problem. What's happening is that > > > pawserv.postinst calls update-inetd, which in some circumstances can > > > restart inetd. Thus inetd is left holding the debconf pipe open and > > > debconf never realises that the postinst has exited. > > > > That makes sense. > > > > > The simple fix should be to put 3>&- at the end of each of the > > > update-inetd invocations, so that update-inetd is always disconnected > > > from debconf. > > > > This seems like a bug in update-inetd which should not have to be > > worked-around in every maintainer script. > > I see your point, although I think I could make a reasonably convincing > argument either way. The opposing position is that it's the postinst > script that's using debconf, not update-inetd, and therefore the > postinst should clean up after itself by not passing the extra open fd > on to update-inetd. > > (DebianNet.pm uses debconf, but it uses the Perl bindings, which don't > open fd 3 - that hack is specific to the shell confmodule.)
But there can only be one front-end, so it is no good closing the only
pipe to it! I first thought that I should move fd 3 back to fd 1, but
in fact the Perl confmodule knows about this redirection:
# A truly gross hack. This is only needed if
# /usr/share/debconf/confmodule is loaded, and then this
# perl module is used. In that case, this module needs to write
# to fd #3, rather than stdout. See changelog 0.3.74.
if (exists $ENV{DEBCONF_REDIR} && $ENV{DEBCONF_REDIR}) {
open(STDOUT,">&3");
}
So pawserv's postinst *must not* do any further direction. And while
the bug is not really in update-inetd, that is probably the only place
it can be fixed now.
Ben.
--
Ben Hutchings
Once a job is fouled up, anything done to improve it makes it worse.
signature.asc
Description: This is a digitally signed message part
