Re: [PATCH v2 01/27] migration: Network Failover can't work with a paused guest

[Michael S. Tsirkin]

On Wed, Dec 02, 2020 at 12:01:21PM +0000, Daniel P. Berrangé wrote:
> On Wed, Dec 02, 2020 at 06:37:46AM -0500, Michael S. Tsirkin wrote:
> > On Wed, Dec 02, 2020 at 11:26:39AM +0000, Daniel P. Berrangé wrote:
> > > On Wed, Dec 02, 2020 at 06:19:29AM -0500, Michael S. Tsirkin wrote:
> > > > On Wed, Dec 02, 2020 at 10:55:15AM +0000, Daniel P. Berrangé wrote:
> > > > > On Wed, Dec 02, 2020 at 11:51:05AM +0100, Juan Quintela wrote:
> > > > > > "Michael S. Tsirkin" <mst@redhat.com> wrote:
> > > > > > > On Wed, Dec 02, 2020 at 05:31:53AM -0500, Michael S. Tsirkin 
> > > > > > > wrote:
> > > > > > >> On Wed, Dec 02, 2020 at 10:27:18AM +0000, Daniel P. Berrangé 
> > > > > > >> wrote:
> > > > > > >> > On Wed, Dec 02, 2020 at 05:13:18AM -0500, Michael S. Tsirkin 
> > > > > > >> > wrote:
> > > > > > >> > > On Wed, Nov 18, 2020 at 09:37:22AM +0100, Juan Quintela 
> > > > > > >> > > wrote:
> > > > > > >> > > > If we have a paused guest, it can't unplug the network VF 
> > > > > > >> > > > device, so
> > > > > > >> > > > we wait there forever.  Just change the code to give one 
> > > > > > >> > > > error on that
> > > > > > >> > > > case.
> > > > > > >> > > > 
> > > > > > >> > > > Signed-off-by: Juan Quintela <quintela@redhat.com>
> > > > > > >> > > 
> > > > > > >> > > It's certainly possible but it's management that created
> > > > > > >> > > this situation after all - why do we bother to enforce
> > > > > > >> > > a policy? It is possible that management will unpause 
> > > > > > >> > > immediately
> > > > > > >> > > afterwards and everything will proceed smoothly.
> > > > > > >> > > 
> > > > > > >> > > Yes migration will not happen until guest is
> > > > > > >> > > unpaused but the same it true of e.g. a guest that is stuck
> > > > > > >> > > because of a bug.
> > > > > > >> > 
> > > > > > >> > That's pretty different behaviour from how migration normally 
> > > > > > >> > handles
> > > > > > >> > a paused guest, which is that it is guaranteed to complete the 
> > > > > > >> > migration
> > > > > > >> > in as short a time as network bandwidth allows.
> > > > > > >> > 
> > > > > > >> > Just ignoring the situation I think will lead to surprise apps 
> > > > > > >> > / admins,
> > > > > > >> > because the person/entity invoking the migration is not likely 
> > > > > > >> > to have
> > > > > > >> > checked wether this particular guest uses net failover or not 
> > > > > > >> > before
> > > > > > >> > invoking - they'll just be expecting a paused migration to run 
> > > > > > >> > fast and
> > > > > > >> > be guaranteed to complete.
> > > > > > >> > 
> > > > > > >> > Regards,
> > > > > > >> > Daniel
> > > > > > >> 
> > > > > > >> Okay I guess. But then shouldn't we handle the reverse situation 
> > > > > > >> too:
> > > > > > >> pausing guest after migration started but before device was
> > > > > > >> unplugged?
> > > > > > >> 
> > > > > > >
> > > > > > > Thinking of which, I have no idea how we'd handle it - fail
> > > > > > > pausing guest until migration is cancelled?
> > > > > > >
> > > > > > > All this seems heavy handed to me ...
> > > > > > 
> > > > > > This is the minimal fix that I can think of.
> > > > > > 
> > > > > > Further solution would be:
> > > > > > - Add a new migration parameter: migrate-paused
> > > > > > - change libvirt to use the new parameter if it exist
> > > > > > - in qemu, when we do start migration (but after we wait for the 
> > > > > > unplug
> > > > > >   device) paused the guest before starting migration and resume it 
> > > > > > after
> > > > > >   migration finish.
> > > > > 
> > > > > It would also have to handle issuing of paused after migration has
> > > > > been started - delay the pause request until the nuplug is complete
> > > > > is one answer.
> > > > 
> > > > Hmm my worry would be that pausing is one way to give cpu
> > > > resources back to host. It's problematic if guest can delay
> > > > that indefinitely.
> > > 
> > > hmm, yes, that is awkward.  Perhaps we should just report an explicit
> > > error then.
> > 
> > Report an error in response to which command? Do you mean
> > fail migration?
> 
> If mgt attempt to pause an existing migration that hasn't finished
> the PCI unplug stage, then fail the pause request.

Pause guest not migration ...
Might be tricky ...

Let me ask this, why not just produce a warning
that migration wan't finish until guest actually runs?
User will then know and unpause the guest when he wants
migration to succeed ...


For example, user can restrict the amount of cpu
using cgroups to a level where almost no progress
is made. QEMU can't detect this ....



> > 
> > > In normal cases this won't happen, as unplug will have
> > > easily completed before the mgmt app pauses the running migration.
> > > In broken/malicious guest cases, this at least ives mgmt a heads up
> > > that something is wrong and they might then decide to cancel the
> > > migration.
> 
> Regards,
> Daniel
> -- 
> |: https://berrange.com      -o-    https://www.flickr.com/photos/dberrange :|
> |: https://libvirt.org         -o-            https://fstop138.berrange.com :|
> |: https://entangle-photo.org    -o-    https://www.instagram.com/dberrange :|