Re: [PATCH 4/9] vfio/migration: Skip pre-copy if dirty page tracking is not supported

[Avihai Horon]

On 5/18/2022 6:50 PM, Jason Gunthorpe wrote:
> On Wed, May 18, 2022 at 01:39:31PM +0200, Juan Quintela wrote:
>
> > > That does seem like a defect in this patch, any SLA constraints should
> > > still all be checked under the assumption all ram is dirty.
> > And how are we going to:
> > - detect the network link speed
> > - to be sure that we are inside downtime limit
> >
> > I think that it is not possible, so basically we are skiping the precopy
> > stage and praying that the other bits are going to be ok.
> Like I keep saying, this is not a real use case, we expect dirty
> tracking to be available in any real configuration. This is just
> trying to make qemu work in some reasonable way if dirty tracking is
> not available but a VFIO migration device is plugged in.
>
> Just pick something simple that makes sense. Like if any SLA is set
> then just refuse to even start. If no SLA then go directly to
> STOP_COPY.

I tried to follow Jason's suggestion to check if there is any SLA and 
block migration, or no SLA at all and allow migration.

It doesn't seem like there is a way to say "no SLA at all".

Migration param downtime_limit takes values between 0 and 2000 seconds.

What does downtime_limit=0 mean?

If it's 0 then MigrationState->threshold_size = 0, so ram_save_pending() 
will never call migration_bitmap_sync_precopy(). Only upon entering 
stop-copy phase will migration_bitmap_sync_precopy() be called, which 
will collect all pages that were dirtied during pre-copy, which could be 
a lot and impact downtime.

In libvirt it seems that downtime_limit = 0 is not valid and can't be set.

Am I missing something here? Is there a way to allow unlimited downtime?


Thanks.