Re: [PATCH RFC 00/26] Multifd 🔀 device state transfer support with VFIO consumer

[Maciej S. Szmigiero]

On 24.04.2024 00:35, Peter Xu wrote:
> On Wed, Apr 24, 2024 at 12:25:08AM +0200, Maciej S. Szmigiero wrote:
> > On 24.04.2024 00:20, Peter Xu wrote:
> > > On Tue, Apr 23, 2024 at 06:15:35PM +0200, Maciej S. Szmigiero wrote:
> > > > On 19.04.2024 17:31, Peter Xu wrote:
> > > > > On Fri, Apr 19, 2024 at 11:07:21AM +0100, Daniel P. Berrangé wrote:
> > > > > > On Thu, Apr 18, 2024 at 04:02:49PM -0400, Peter Xu wrote:
> > > > > > > On Thu, Apr 18, 2024 at 08:14:15PM +0200, Maciej S. Szmigiero wrote:
> > > > > > > > I think one of the reasons for these results is that mixed (RAM + device
> > > > > > > > state) multifd channels participate in the RAM sync process
> > > > > > > > (MULTIFD_FLAG_SYNC) whereas device state dedicated channels don't.
> > > > > > >
> > > > > > > Firstly, I'm wondering whether we can have better names for these new
> > > > > > > hooks.  Currently (only comment on the async* stuff):
> > > > > > >
> > > > > > >      - complete_precopy_async
> > > > > > >      - complete_precopy
> > > > > > >      - complete_precopy_async_wait
> > > > > > >
> > > > > > > But perhaps better:
> > > > > > >
> > > > > > >      - complete_precopy_begin
> > > > > > >      - complete_precopy
> > > > > > >      - complete_precopy_end
> > > > > > >
> > > > > > > ?
> > > > > > >
> > > > > > > As I don't see why the device must do something with async in such hook.
> > > > > > > To me it's more like you're splitting one process into multiple, then
> > > > > > > begin/end sounds more generic.
> > > > > > >
> > > > > > > Then, if with that in mind, IIUC we can already split ram_save_complete()
> > > > > > > into >1 phases too. For example, I would be curious whether the performance
> > > > > > > will go back to normal if we offloading multifd_send_sync_main() into the
> > > > > > > complete_precopy_end(), because we really only need one shot of that, and I
> > > > > > > am quite surprised it already greatly affects VFIO dumping its own things.
> > > > > > >
> > > > > > > I would even ask one step further as what Dan was asking: have you thought
> > > > > > > about dumping VFIO states via multifd even during iterations?  Would that
> > > > > > > help even more than this series (which IIUC only helps during the blackout
> > > > > > > phase)?
> > > > > >
> > > > > > To dump during RAM iteration, the VFIO device will need to have
> > > > > > dirty tracking and iterate on its state, because the guest CPUs
> > > > > > will still be running potentially changing VFIO state. That seems
> > > > > > impractical in the general case.
> > > > >
> > > > > We already do such interations in vfio_save_iterate()?
> > > > >
> > > > > My understanding is the recent VFIO work is based on the fact that the VFIO
> > > > > device can track device state changes more or less (besides being able to
> > > > > save/load full states).  E.g. I still remember in our QE tests some old
> > > > > devices report much more dirty pages than expected during the iterations
> > > > > when we were looking into such issue that a huge amount of dirty pages
> > > > > reported.  But newer models seem to have fixed that and report much less.
> > > > >
> > > > > That issue was about GPU not NICs, though, and IIUC a major portion of such
> > > > > tracking used to be for GPU vRAMs.  So maybe I was mixing up these, and
> > > > > maybe they work differently.
> > > >
> > > > The device which this series was developed against (Mellanox ConnectX-7)
> > > > is already transferring its live state before the VM gets stopped (via
> > > > save_live_iterate SaveVMHandler).
> > > >
> > > > It's just that in addition to the live state it has more than 400 MiB
> > > > of state that cannot be transferred while the VM is still running.
> > > > And that fact hurts a lot with respect to the migration downtime.
> > > >
> > > > AFAIK it's a very similar story for (some) GPUs.
> > >
> > > So during iteration phase VFIO cannot yet leverage the multifd channels
> > > when with this series, am I right?
> >
> > That's right.
> >
> > > Is it possible to extend that use case too?
> >
> > I guess so, but since this phase (iteration while the VM is still
> > running) doesn't impact downtime it is much less critical.
>
> But it affects the bandwidth, e.g. even with multifd enabled, the device
> iteration data will still bottleneck at ~15Gbps on a common system setup
> the best case, even if the hosts are 100Gbps direct connected.  Would that
> be a concern in the future too, or it's known problem and it won't be fixed
> anyway?

I think any improvements to the migration performance are good, even if
they don't impact downtime.

It's just that this patch set focuses on the downtime phase as the more
critical thing.

After this gets improved there's no reason why not to look at improving
performance of the VM live phase too if it brings sensible improvements.

> I remember Avihai used to have plan to look into similar issues, I hope
> this is exactly what he is looking for.  Otherwise changing migration
> protocol from time to time is cumbersome; we always need to provide a flag
> to make sure old systems migrates in the old ways, new systems run the new
> ways, and for such a relatively major change I'd want to double check on
> how far away we can support offload VFIO iterations data to multifd.

The device state transfer is indicated by a new flag in the multifd
header (MULTIFD_FLAG_DEVICE_STATE).

If we are to use multifd channels for VM live phase transfers these
could simply re-use the same flag type.

> Thanks,

Thanks,
Maciej