Re: [PATCH v3 07/10] migration: add new migration state wait-unplug

[Dr. David Alan Gilbert]

* Jens Freimann (address@hidden) wrote:
> On Fri, Oct 11, 2019 at 06:11:33PM +0100, Dr. David Alan Gilbert wrote:
> > * Jens Freimann (address@hidden) wrote:
> > > This patch adds a new migration state called wait-unplug.  It is entered
> > > after the SETUP state and will transition into ACTIVE once all devices
> > > were succesfully unplugged from the guest.
> > > 
> > > So if a guest doesn't respond or takes long to honor the unplug request
> > > the user will see the migration state 'wait-unplug'.
> > > 
> > > In the migration thread we query failover devices if they're are still
> > > pending the guest unplug. When all are unplugged the migration
> > > continues. We give it a defined number of iterations including small
> > > waiting periods before we proceed.
> > > 
> > > Signed-off-by: Jens Freimann <address@hidden>
> [..]
> > > @@ -3260,6 +3271,27 @@ static void *migration_thread(void *opaque)
> > > 
> > >      qemu_savevm_state_setup(s->to_dst_file);
> > > 
> > > +    migrate_set_state(&s->state, MIGRATION_STATUS_SETUP,
> > > +                      MIGRATION_STATUS_WAIT_UNPLUG);
> > 
> > I think I'd prefer if you only went into this state if you had any
> > devices that were going to need unplugging.
> 
> Sure, that makes sense. I'll change it.
> 
> > > +    while (i < FAILOVER_UNPLUG_RETRIES &&
> > > +           s->state == MIGRATION_STATUS_WAIT_UNPLUG) {
> > > +        i++;
> > > +        qemu_sem_timedwait(&s->wait_unplug_sem, 
> > > FAILOVER_GUEST_UNPLUG_WAIT);
> > > +        all_unplugged = qemu_savevm_state_guest_unplug_pending();
> > > +        if (all_unplugged) {
> > > +            break;
> > > +        }
> > > +    }
> > > +
> > > +    if (all_unplugged) {
> > > +        migrate_set_state(&s->state, MIGRATION_STATUS_WAIT_UNPLUG,
> > > +                MIGRATION_STATUS_ACTIVE);
> > > +    } else {
> > > +        migrate_set_state(&s->state, MIGRATION_STATUS_WAIT_UNPLUG,
> > > +                          MIGRATION_STATUS_CANCELLING);
> > > +    }
> > 
> > I think you can get rid of both the timeout and the count and just make
> > sure that migrate_cancel works at this point.
> 
> I see, I need to add the new state to migration_is_setup_or_active() or
> a cancel won't work.

You probably need to do that anyway given all the other places
is_setup_or_active is called.

> > This pushes the problem up a layer, which I think is fine.
> 
> Seems good to me. To be clear, you're saying I should just poll on
> the device unplugged state? Like
> 
>         while (s->state == MIGRATION_STATUS_WAIT_UNPLUG &&
>                !qemu_savevm_state_guest_unplug_pending()) {
> _            /* This block intentionally left blank */
>         }

I'd keep the qemu_sem_timedwait in there, but with a short time out
(e.g. 250ms say); that way it doesn't eat cpu, but also the cancel still
happens quickly.

Dave

> 
> regards,
> Jens
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK