Re: [PATCH v5 2/9] migration: Implement switchover ack logic

[Alex Williamson]

On Tue, 6 Jun 2023 15:12:13 +0300
Avihai Horon <avihaih@nvidia.com> wrote:
> On 06/06/2023 1:06, Alex Williamson wrote:
> > On Tue, 30 May 2023 17:48:14 +0300
> > Avihai Horon <avihaih@nvidia.com> wrote:
> >> @@ -2700,6 +2713,20 @@ static void 
> >> migration_update_counters(MigrationState *s,
> >>                                 bandwidth, s->threshold_size);
> >>   }
> >>
> >> +static bool migration_can_switchover(MigrationState *s)
> >> +{
> >> +    if (!migrate_switchover_ack()) {
> >> +        return true;
> >> +    }
> >> +
> >> +    /* No reason to wait for switchover ACK if VM is stopped */
> >> +    if (!runstate_is_running()) {
> >> +        return true;
> >> +    }  
> > Is it possible for QEMU to force the migration to continue regardless
> > of receiving an ack from the target and is this the check that would
> > allow that?  
> 
> Yes. If you stop the source VM then migration will not wait for an ACK 
> to do the switchover.
> 
> >
> > It seems that we don't know the downtime allowed for the VM in any of
> > this, nor do we know how much time the target device will require to
> > generate an ack, but we could certainly have conditions where the
> > priority is moving the VM from the source host regardless of the
> > resulting downtime.  
> 
> In such cases you can keep the switchover-ack capability off.

How is that accomplished?

> > Also does the return path requirement preclude offline migration or
> > does the above again take care of that if we pause the VM for an
> > offline migration (ex. save to and restore from file)?  
> 
> I suppose that by offline migration you mean migration where you stop 
> the source VM first and then do migration?

Yes.

> If so, offline migration should work and in that case we don't care 
> about the ACK as downtime is not a concern.
> 
> However, migrating to a file doesn't work with return-path, as you don't 
> have the destination side responding to the source via the return path.
> For this reason, using return-path when migrating to a file doesn't make 
> sense.

So we require return-path for switchover-ack, but switchover-ack is
only required for pre-copy, therefore why do we require return-path for
an offline migration?

If there's a way to turn off switchover-ack capability, is there also a
way to turn off return-path and therefore enable migration to file?
Sorry if I'm simply not familiar with these migration switches.  Thanks,

Alex