It's worse than the initial description protrays - it turns out that if the dnsmasq package is updated (e.g. by "yum update"), all libvirt-initiated dnsmasq processes are summarily killed, and a system-wide dnsmasq process using the "default" dnsmasq configuration (which listens on all interfaces) is started. This prevents libvirt networks from being restarted (their dnsmasq fails to start because it finds the dhcp/dns ports on the interface "busy").

The reason for this is similar to the initial description - the dnsmasq specfile's %post script calls "service dnsmasq condrestart" which is supposed to restart the dnsmasq service *only if it is already running*. But instead of checking for a system-service initiated dnsmasq process, dnsmasq's initscript simply calls "pidof dnsmasq" (which looks for *any* dnsmasq processes), sees that there is at least one, and uses that as an indication to "killproc dnsmasq" then start the system instance of dnsmasq.

What dnsmasq's initscript *should* do is use "pidfileofproc dnsmasq" instead of "pidof dnsmasq" *in all cases*, and terminate/restart dnsmasq only if the system-service instance of dnsmasq is runnning.

I found a report of this same bug against Fedora 11/12 all the way back in December of 2009, with a (partial) patch included: Bug 547605.

I just made and tested a complete patch (and am attaching it to this bug) for the current RHEL6 dnsmasq initscript file which fixes all the problems outlined in this bug:

1) "service dnsmasq stop|restart|condrestart" will only stop dnsmasq processes
   that were originally started by the dnsmasq initscript.

2) "service dnsmasq condrestart" will only start the system dnsmasq if an initscript-initiated dnsmasq was already running

3) "service dnsmasq reload" will only send SIGHUP to the dnsmasq process started
   by the dnsmasq initscript.

To prevent unexpected disruption to libvirt users during an update, this patch should at the very least be included with the next update of dnsmasq, as it puts libvirt's networks in a state that requires undocumented manual intervention (or a host reboot) to escape from.

Created attachment 644338 [details]
fix  dnsmasq initscript to only recognize/act on self-initiated dnsmasq instances

Created attachment 647838 [details]
[PATCH] fix dnsmasq initscript to only recognize/act on self-initiated dnsmasq instances

Corrected version of the initial patch.

Since there will now be a dnsmasq update for RHEL6.4, this patch should be included (see the final paragraph of Comment 2)

Also, I'm not sure what the reason was for your change to the original patch I posted. "pidfileofproc" isn't a binary in /sbin, it is a shell function that is defined in /etc/init.d/functions (sourced right at the beginning of /etc/init.d/dnsmasq), and anyway you only changed one instance of that name in the patch, leaving the other 4 occurences untouched. I believe the original patch is the correct patch to apply.

(In reply to comment #6)
> Also, I'm not sure what the reason was for your change to the original patch
> I posted. "pidfileofproc" isn't a binary in /sbin, it is a shell function
> that is defined in /etc/init.d/functions (sourced right at the beginning of
> /etc/init.d/dnsmasq), and anyway you only changed one instance of that name
> in the patch, leaving the other 4 occurences untouched. I believe the
> original patch is the correct patch to apply.

Yes, you are right about saying that "pidfileofproc" isn't a binary in /sbin.

In the *original patch* you used:
-	    if test "x`/sbin/pidof dnsmasq`" != x; then
+	    if test "x`/sbin/pidfileofproc dnsmasq`" != x; then

Since "pidfileofproc" isn't a binary in /sbin (as I know and as you stated in Comment #6) I corrected it to:
-	    if test "x`/sbin/pidof dnsmasq`" != x; then
+	    if test "x`pidfileofproc dnsmasq`" != x; then

This request was evaluated by Red Hat Product Management for
inclusion in a Red Hat Enterprise Linux release.  Product
Management has requested further review of this request by
Red Hat Engineering, for potential inclusion in a Red Hat
Enterprise Linux release for currently deployed products.
This request is not yet committed for inclusion in a release.

Ah, right you are! I was looking at the patchfiles backwards for some reason. Sorry for the noise :-/

Created attachment 666124 [details]
[PATCH] fix dnsmasq initscript to only recognize/act on self-initiated dnsmasq instances

This is the final patch I used. I changed "kill" back to "killproc" after consulting with colleague that maintains init scripts. killproc does much more than just kill. And also the initscript output was not good with kill. There was no [ OK ] or [ FAILED ].

After inspection the main problem was in the condition in "stop" part of initscript:

-        if test "x`pidof dnsmasq`" != x; then
+        if test "x`pidfileofproc dnsmasq`" != x; then

I tested the final patch and everything worked as should. Calling "service dnsmasq stop" killed just the system instantion with PID in /var/run/dnsmasq.pid

Ah, now I have a better understanding of what was causing libvirt's dnsmasq processes to be erroneously killed - it's not the mere act of using "killproc dnsmasq", but in doing that when there is currently no running system instance of dnsmasq.

If there is a system instance of dnsmasq running, killproc will kill only that process. But if no system instance can be found, "killproc dnsmasq" behaves identically to "killall dnsmasq", so all of libvirt's dnsmasq instances are killed.

I've tested dnsmasq-2.48-12 on a system with libvirt dnsmasq instances running, and all start/stop/restart operations are now behaving properly. Thanks!

Since the problem described in this bug report should be
resolved in a recent advisory, it has been closed with a
resolution of ERRATA.

For information on the advisory, and where to find the updated
files, follow the link below.

If the solution does not work for you, open a new bug report.

http://rhn.redhat.com/errata/RHSA-2013-0277.html