RE: [virtio-comment] [PATCH v1 1/8] admin: Add theory of operation for device migration

[Parav Pandit <parav@nvidia.com>]

> From: Michael S. Tsirkin <mst@redhat.com>
> Sent: Friday, November 17, 2023 6:00 PM
> To: Parav Pandit <parav@nvidia.com>
> 
> On Fri, Nov 17, 2023 at 12:02:49PM +0000, Parav Pandit wrote:
> >
> >
> > > From: Michael S. Tsirkin <mst@redhat.com>
> > > Sent: Friday, November 17, 2023 5:13 PM
> > >
> > > On Fri, Nov 17, 2023 at 11:20:14AM +0000, Parav Pandit wrote:
> > > >
> > > > > From: Michael S. Tsirkin <mst@redhat.com>
> > > > > Sent: Friday, November 17, 2023 4:41 PM
> > > > >
> > > > > On Fri, Nov 17, 2023 at 10:20:45AM +0000, Parav Pandit wrote:
> > > > > >
> > > > > >
> > > > > > > From: Michael S. Tsirkin <mst@redhat.com>
> > > > > > > Sent: Friday, November 17, 2023 3:38 PM
> > > > > > >
> > > > > > > On Wed, Nov 15, 2023 at 05:39:43PM +0000, Parav Pandit wrote:
> > > > > > > > > >
> > > > > > > > > > Additionally, if hypervisor has put the trap on virtio
> > > > > > > > > > config, and because the memory device already has the
> > > > > > > > > > interface for virtio config,
> > > > > > > > > >
> > > > > > > > > > Hypervisor can directly write/read from the virtual
> > > > > > > > > > config to the member's
> > > > > > > > > config space, without going through the device context, right?
> > > > > > > > >
> > > > > > > > > If it can do it or it can choose to not. I don't see how
> > > > > > > > > it is related to the discussion here.
> > > > > > > > >
> > > > > > > > It is. I donât see a point of hypervisor not using the
> > > > > > > > native interface provided
> > > > > > > by the member device.
> > > > > > >
> > > > > > > So for example, it seems reasonable to a member supporting
> > > > > > > both existing pci register interface for compatibility and
> > > > > > > the future DMA based one for scale. In such a case, it seems
> > > > > > > possible that DMA will expose more features than pci. And
> > > > > > > then a hypervisor might decide to use
> > > > > that in preference to pci registers.
> > > > > >
> > > > > > We donât find it right to involve owner device for mediating
> > > > > > at current scale
> > > > >
> > > > > In this model, device will be its own owner. Should not be a problem.
> > > > >
> > > > I didnât understand above comment.
> > >
> > > We'd add a new group type "self". You can then send admin commands
> > > through VF itself not through PF.
> > >
> > How? The device is owned by the guest. FLR and device reset cannot send the
> admin command reliably.
> 
> It's of the "it hurts when I do this - don't do this then" category.
>
it is donât do medication category, yes due all this weirdness that has been asked.
 
> 
> > >
> > > > > > and to not break TDISP efforts in upcoming time by such design.
> > > > >
> > > > > Look you either stop mentioning TDISP as motivation or actually
> > > > > try to address it. Safe migration with TDISP is really hard.
> > > > But that is not an excuse to say that TDISP migration is not
> > > > present, hence
> > > involve the owner device for config space access.
> > > > This is another hurdle added that further blocks us away from TDISP.
> > > > Hence, we donât want to take the route of involving owner device
> > > > for any
> > > config access.
> > >
> > > This "blocks" is all just wild hunches. hypervisor controls some
> > > aspects of TDISP devices for sure - maybe we actually should use pci
> > > config space as that is generally hypervisor controlled.
> > Even bad to do hypercalls.
> > I showed you last time the role of the PCI config space snippet from the spec.
> 
> Yes I remember. This is just an example though. My point is maybe it is solvable
> maybe it is not.
> 
> > Do you see we are repeating the discussion again?
> 
> One of the reasons is that people bring up irrelevances. TDISP is important but
> has to be addressed or deferred not vaguely referred to.

So lets continue to follow the current TDISP direction of not involving hypervisor for virtio common and device config.

> 
> > >
> > > > > For example, your current patches are clearly broken for TDISP:
> > > > > owner can control queue state at any time making device modify
> > > > > memory in any way it wants.
> > > > >
> > > > When TDISP migration is needed, the admin device can be another
> > > > TVM
> > > outside the HV scope.
> > > > Or an alternative would have device context encrypted not visible to HV at
> all.
> > >
> > > Maybe. Fact remains your patches do conflict with TDISP and you seem
> > > to be fine with it because you have a hunch you can fix it. But we
> > > can't do development based on your hunches.
> > >
> > We have different view.
> > My patches do not conflict with TDISP because TDISP has clear definition of
> not involving hypervisor for transport.
> > And that part is still preserved.
> > Delegating the migration to another TDISP or encrypting is yet to be defined.
> > And current patches will align to both the approaches in future.
> >
> > So you need to re-evaluate your judgment.
> 
> If you like they do not "conflict".  But if used with TDISP they just make it
> insecure and thus completely worthless.  If hypervisor can change ring state to
> make device poke at random guest memory then it is game over and all the
> effort spent was security theater.
Not really, I proposed two options.
1. delegate the task of LM to the TVM. (proposed by two cpu vendors).
In this case all the infra we build here, just works fine.
It also does not require any hypervisor mediation for control plane.

2. Encrypt the owner device workload to be not seen by hypervisor

Both methods does not affect the current direction.

But if we force trap+emulation, it is 100% broken for TDISP.
And I would not promote that.

> But you know this, don't you? This is why you mentioned encrypting device.
> Maybe that works. It just does not work *as is*.
It works as_is. But current infrastructure does not block the future work.

> 
> 
> > >
> > > > Such encryption is not possible, with the trap+emulation method,
> > > > where HV
> > > will have to decrypt the data coming over MMIO writes.
> > >
> > > I don't how what trap+emulation has to do with it. Do you refer to
> > > the shadow vq thing?
> >
> > The method proposed here does not hinder any TDISP direction.
> 
> direction? No, why would it. we can always add more commands that are safe
> for TDISP. commands you propose here are unsafe for TDISP.
> 
> > Without my proposal, do you have a method that does not involve hypervisor
> intervention for virtio common and device config space, cvq and shadow vq?
> > If so, I would like to hear that as well because that will align with TDISP.
> 	
> I really did not give it much thought.  I suspect for TDISP it just might be cleaner
> to have guest agent migrate device. Certainly removes all the messy questions.
> That, to me impliest there needs to be a way to send migration commands
> through VF itself. Does this "involve hypervisor intervention"? No one should
> care I think.
Too far of the future to envision. May be yes. When such platform is built, for sure whoever migrates need migrate its device side too.
Some knowledge of migration driver is needed.

> 
> 
> > > I am guessing modern platforms with TDISP support are likely to also
> > > support dirty bit in the IOMMU.
> > >
> > It will be some day.
> 
> What does this mean? Which platforms support TDISP and which IOMMUs do
> they use?
I said it will be some day, not right now.

> 
> > >
> > > > > > And for future scale, having new SIOV interface makes more
> > > > > > sense which has
> > > > > its own direct interface to device.
> > > > > >
> > > > > > I finally captured all past discussions in form of a FAQ at [1].
> > > > > >
> > > > > > [1]
> > > > > > https://docs.google.com/document/d/1Iyn-l3Nm0yls3pZaul4lZiVj8x
> > > > > > 1s73
> > > > > > Ed6r
> > > > > > Osmn6LfXc/edit?usp=sharing
> > > > >
> > > > > Yea skimmed that, "Cons: None". Are you 100% sure? Anyway,
> > > > > discussion will take place on the mailing list please.
> > > >
> > > > We cannot keep discussing the register interface every week.
> > > > I remember we have discussed this many times already in following series.
> > > >
> > > > 1. legacy series
> > > > 2. tvq v4 series
> > > > 3. dynamic vq creation series
> > > > 4. again during suspend series under tvq head 5. right now 6. May
> > > > be more that I forgot.
> > > >
> > > > I captured all the direction and options in the doc. One can refer
> > > > when those
> > > questions arise there.
> > > > If we donât work cohesively same reasoning repetition does not help.
> > >
> > > It's still the same too, doc or no doc. You want to build a device
> > > without registers fine but don't force it down everyone's throat.
> > I donât see any compelling reason for inventing new method really.
> > Nor continuing in register mode.
> > Virtio already has VQ.
> > If CVQ is so problematic, one should put everything on registers and not run
> on double standards.
> 
> We should not and neither should we put everything behind a VQ.
>
Why?
 
> 
> > I captured all the reasoning and thoughts. I donât have much to say in support
> of infinite register scale.
> >
> > People who wants to push SIOV does not show single performance reason on
> why SIOV to be done.
> > I have upstreamed SIOVs in Linux as SFs without PASID, and in all our scale
> tests, before the device chocks, the system chocks.
> >
> > So when someone pushes the SIOV series, I will be the first one interested in
> reading the performance numbers to proceed with patches.
> >
> > > And now with 8MBytes
> > > of on-device memory that's needed for migration and that's
> > > apparently fine I am even less interested in saving 256 bytes for config
> space.
> >
> > Again, not the right comparison. When and how to use 256 matters.
> > I havenât come across any device that prefers infinite register scale.
> 
> Why resort to hyperbole? 256 bytes is pretty far from infinite.  But again, if you
> don't want it in registers just add an option to move
> *all* of config space out of registers. 
This does not work in backward compatible way, not brings the predictability.
The proposed method and current direction is just fine. No changes needed.
No one is extending the config space or config registers in virtio anymore anyway.

> cheaper devices will require newer guests.
> Or, 10 years will pass and you will be able to drop compat with old guests. I
> know it's too long a game for you to care but I've been virtio spec editor for
> more than 10 years so to me it seems reasonable to plan like that.

Current proposal of using CVQ is just working fine across most* virtio spec contributors.
I donât see technical reason to change it.