[ovs-dev] [PATCH] xenserver: Kill bond master's dhclient when bringing up bond slave.

[Ben Pfaff]

On Mon, Jun 28, 2010 at 09:56:20AM +0100, Ian Campbell wrote:
> On Fri, 2010-06-25 at 21:22 +0100, Ben Pfaff wrote:
> > interface-reconfigure is never explicitly called to down a bond master.
> > However, when it is called to up a slave it is implicit that we are
> > destroying the master.  The "bridge" version of interface-reconfigure
> > always "ifdown"s the bond master in such a case, but until now the
> > "vswitch" version has not done so.  Usually, it doesn't matter, because
> > the bond master network device disappears when the slave is brought up,
> > but one case was missed: for a bond master with an IP address obtained
> > via DHCP, the dhclient process needs to be killed, and we were not doing
> > it.  This commit starts doing it (by invoking ifdown on the bond master).
> 
> Looks good.

Thank you for the review.

> > The dhclient process that hangs around doesn't cause problems until the
> > bond master is brought back up, at which point "ifup" fails because it
> > refuses to start another dhclient for the same interface.
> > 
> > I do not know for sure whether the converse change is needed: that is,
> > when a bond PIF is brought up, is interface-reconfigure expected to
> > implicitly take down its slave PIFs?  But I do not think so.  This commit
> > does not change that behavior.
> 
> interface-reconfigure is expected to take down slave PIFs if they are
> up. The bridge bring_down_existing brings down:
>       * vlan masters - (I think) not necessary with vswitch due to
>         different behaviour wrt vlans and tagging ports instead of
>         creating separate bridges etc.
>       * bond masters - handled by this patch
>       * the device itself - this brings down the bond slaves too due to
>         the bridge version of bring_down_interface doing so. In vswitch
>         this seems to be handled by extra_down_ports in the
>         configure_datapath function.
> 
> Unless switching from a slave w/ DHCP to a bond leaves a dhclient lying
> around (which I don't think it does) I think we are currently ok.

For completeness I checked this behavior (on 5.5.0).  It already worked
correctly.  So it seems that I was prematurely worried about that part
anyway.

I pushed this change out, after updating the log message.