Re: [Qemu-devel] [PATCH 0/3] Migration cancel with dead network

[Dr. David Alan Gilbert]

* Daniel P. Berrange (address@hidden) wrote:
> On Thu, Jan 08, 2015 at 11:11:29AM +0000, Dr. David Alan Gilbert (git) wrote:
> > From: "Dr. David Alan Gilbert" <address@hidden>
> > 
> > If the remote host, or networking dies during a migration, the socket can be
> > waiting for a long timeout, and migration_cancel can't complete the cancel
> > for a long time (and you can't start a new one to somewhere else).
> > (Where 'long' is the TCP timeout, that's ~15 mins)
> > 
> > This patch set uses the shutdown(2) syscall to unblock any write/sends that
> > are in progress to let the migrate_cancel happen quickly.
> > 
> >   1/3: socket shutdown  - An updated patch from my postcopy world to
> >                           add a shut_down method on QEMUFile - only
> >                           for 'socket' (where the syscall is supported).
> > 
> >   2/3: Handle bi-directional communication for fd migration
> >                         - A patch from Cristian Klein to use the socket
> >                           QEMUFile for FDs that are passed in, if the FDs
> >                           are sockets; this is needed so that libvirt
> >                           migrations can take advantage of the other 
> > patches.
> >                           Again this patch (and its naming) come from the
> >                           postcopy world.
> > 
> >   3/3: migration_cancel: shutdown migration socket
> >                         - A new patch that uses the shutdown in 
> > migrate_fd_cancel
> > 
> > 
> > Note this does not fix the timeout if you try to migrate to an already dead 
> > host;
> > the connect timeout is typically a much shorter 2 minutes anyway.
> 
> In any libvirt managed setup, you'd need to address the connect timeout
> issue in libvirt instead, since libvirt always uses fd based migration.
> ie libvirt estabishes the connection & then passes the TCP Socket to
> QEMU.
> 
> It should be possible for libvirt to use a non-blocking connect()
> call and catch use of its virDomainMigrateCancel APi to stop the
> connection attempt.

Yes, although as I say I've not fixed the (must shorter) connect side timeout
on the QEMU side either.

How does libvirt behave if you cancel a tunnelled migration with the network
dieing in the middle?

Dave

> 
> Regards,
> Daniel
> -- 
> |: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
> |: http://libvirt.org              -o-             http://virt-manager.org :|
> |: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
> |: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK