Re: [PATCH v3 6/8] admin: Add theory of operation for write recording commands

["Michael S. Tsirkin" <mst@redhat.com>]

On Thu, Nov 09, 2023 at 11:31:27AM +0800, Jason Wang wrote:
> On Wed, Nov 8, 2023 at 4:17âPM Michael S. Tsirkin <mst@redhat.com> wrote:
> >
> > On Wed, Nov 08, 2023 at 12:28:36PM +0800, Jason Wang wrote:
> > > On Tue, Nov 7, 2023 at 3:05âPM Michael S. Tsirkin <mst@redhat.com> wrote:
> > > >
> > > > On Tue, Nov 07, 2023 at 12:04:29PM +0800, Jason Wang wrote:
> > > > > > > > Each virtio and non virtio devices who wants to report their dirty page report,
> > > > > > > will do their way.
> > > > > > > >
> > > > > > > > > 3) inventing it in the virtio layer will be deprecated in the future
> > > > > > > > > for sure, as platform will provide much rich features for logging
> > > > > > > > > e.g it can do it per PASID etc, I don't see any reason virtio need
> > > > > > > > > to compete with the features that will be provided by the platform
> > > > > > > > Can you bring the cpu vendors and committement to virtio tc with timelines
> > > > > > > so that virtio TC can omit?
> > > > > > >
> > > > > > > Why do we need to bring CPU vendors in the virtio TC? Virtio needs to be built
> > > > > > > on top of transport or platform. There's no need to duplicate their job.
> > > > > > > Especially considering that virtio can't do better than them.
> > > > > > >
> > > > > > I wanted to see a strong commitment for the cpu vendors to support dirty page tracking.
> > > > >
> > > > > The RFC of IOMMUFD support can go back to early 2022. Intel, AMD and
> > > > > ARM are all supporting that now.
> > > > >
> > > > > > And the work seems to have started for some platforms.
> > > > >
> > > > > Let me quote from the above link:
> > > > >
> > > > > """
> > > > > Today, AMD Milan (or more recent) supports it while ARM SMMUv3.2
> > > > > alongside VT-D rev3.x also do support.
> > > > > """
> > > > >
> > > > > > Without such platform commitment, virtio also skipping it would not work.
> > > > >
> > > > > Is the above sufficient? I'm a little bit more familiar with vtd, the
> > > > > hw feature has been there for years.
> > > >
> > > >
> > > > Repeating myself - I'm not sure that will work well for all workloads.
> > >
> > > I think this comment applies to this proposal as well.
> >
> > Yes - some systems might be better off with platform tracking.
> > And I think supporting shadow vq better would be nice too.
> 
> For shadow vq, did you mean the work that is done by Eugenio?

Yes.

> >
> > > > Definitely KVM did
> > > > not scan PTEs. It used pagefaults with bit per page and later as VM size
> > > > grew switched to PLM.  This interface is analogous to PLM,
> > >
> > > I think you meant PML actually. And it doesn't work like PML. To
> > > behave like PML it needs to
> > >
> > > 1) log buffers were organized as a queue with indices
> > > 2) device needs to suspend (as a #vmexit in PML) if it runs out of the buffers
> > > 3) device need to send a notification to the driver if it runs out of the buffer
> > >
> > > I don't see any of the above in this proposal. If we do that it would
> > > be less problematic than what is being proposed here.
> >
> > What is common between this and PML is that you get the addresses
> > directly without scanning megabytes of bitmaps or worse -
> > hundreds of megabytes of page tables.
> 
> Yes, it has overhead but this is the method we use for vhost and KVM (earlier).
> 
> To me the  important advantage of PML is that it uses limited
> resources on the host which
> 
> 1) doesn't require resources in the device
> 2) doesn't scale as the guest memory increases. (but this advantage
> doesn't exist in neither this nor bitmap)

it seems 2 exactly exists here.


> >
> > The data structure is different but I don't see why it is critical.
> >
> > I agree that I don't see out of buffers notifications too which implies
> > device has to maintain something like a bitmap internally.  Which I
> > guess could be fine but it is not clear to me how large that bitmap has
> > to be. How does the device know? Needs to be addressed.
> 
> This is the question I asked Parav in another thread. Using host
> memory as a queue with notification (like PML) might be much better.

Well if queue is what you want to do you can just do it internally.
Problem of course is that it might overflow and cause things like
packet drops.


> >
> >
> > > Even if we manage to do that, it doesn't mean we won't have issues.
> > >
> > > 1) For many reasons it can neither see nor log via GPA, so this
> > > requires a traversal of the vIOMMU mapping tables by the hypervisor
> > > afterwards, it would be expensive and need synchronization with the
> > > guest modification of the IO page table which looks very hard.
> >
> > vIOMMU is fast enough to be used on data path but not fast enough for
> > dirty tracking?
> 
> We set up SPTEs or using nesting offloading where the PTEs could be
> iterated by hardware directly which is fast.

There's a way to have hardware find dirty PTEs for you quickly?
I don't know how it's done. Do tell.


> This is not the case here where software needs to iterate the IO page
> tables in the guest which could be slow.
> 
> > Hard to believe.  If true and you want to speed up
> > vIOMMU then you implement an efficient datastructure for that.
> 
> Besides the issue of performance, it's also racy, assuming we are logging IOVA.
> 
> 0) device log IOVA
> 1) hypervisor fetches IOVA from log buffer
> 2) guest map IOVA to a new GPA
> 3) hypervisor traverse guest table to get IOVA to new GPA
> 
> Then we lost the old GPA.

Interesting and a good point. And by the way e.g. vhost has the same
issue.  You need to flush dirty tracking info when changing the mappings
somehow.  Parav what's the plan for this? Should be addressed in the
spec too.



> >
> > > 2) There are a lot of special or reserved IOVA ranges (for example the
> > > interrupt areas in x86) that need special care which is architectural
> > > and where it is beyond the scope or knowledge of the virtio device but
> > > the platform IOMMU. Things would be more complicated when SVA is
> > > enabled.
> >
> > SVA being what here?
> 
> For example, IOMMU may treat interrupt ranges differently depending on
> whether SVA is enabled or not. It's very hard and unnecessary to teach
> devices about this.

Oh, shared virtual memory. So what you are saying here? virtio
does not care, it just uses some addresses and if you want it to
it can record writes somewhere.

> >
> > > And there could be other architecte specific knowledge (e.g
> > > PAGE_SIZE) that might be needed. There's no easy way to deal with
> > > those cases.
> >
> > Good point about page size actually - using 4k unconditionally
> > is a waste of resources.
> 
> Actually, they are more than just PAGE_SIZE, for example, PASID and others.

what does pasid have to do with it? anyway, just give driver control
over page size.

> >
> >
> > > We wouldn't need to care about all of them if it is done at platform
> > > IOMMU level.
> >
> > If someone logs at IOMMU level then nothing needs to be done
> > in the spec at all. This is about capability at the device level.
> 
> True, but my question is where or not it can be done at the device level easily.

there's no "easily" about live migration ever.
For example on-device iommus are a thing.

> >
> >
> > > > what Lingshan
> > > > proposed is analogous to bit per page - problem unfortunately is
> > > > you can't easily set a bit by DMA.
> > > >
> > >
> > > I'm not saying bit/bytemap is the best, but it has been used by real
> > > hardware. And we have many other options.
> > >
> > > > So I think this dirty tracking is a good option to have.
> > > >
> > > >
> > > >
> > > > > >
> > > > > > > > i.e. in first year of 2024?
> > > > > > >
> > > > > > > Why does it matter in 2024?
> > > > > > Because users needs to use it now.
> > > > > >
> > > > > > >
> > > > > > > > If not, we are better off to offer this, and when/if platform support is, sure,
> > > > > > > this feature can be disabled/not used/not enabled.
> > > > > > > >
> > > > > > > > > 4) if the platform support is missing, we can use software or
> > > > > > > > > leverage transport for assistance like PRI
> > > > > > > > All of these are in theory.
> > > > > > > > Our experiment shows PRI performance is 21x slower than page fault rate
> > > > > > > done by the cpu.
> > > > > > > > It simply does not even pass a simple 10Gbps test.
> > > > > > >
> > > > > > > If you stick to the wire speed during migration, it can converge.
> > > > > > Do you have perf data for this?
> > > > >
> > > > > No, but it's not hard to imagine the worst case. Wrote a small program
> > > > > that dirty every page by a NIC.
> > > > >
> > > > > > In the internal tests we donât see this happening.
> > > > >
> > > > > downtime = dirty_rates * PAGE_SIZE / migration_speed
> > > > >
> > > > > So if we get very high dirty rates (e.g by a high speed NIC), we can't
> > > > > satisfy the requirement of the downtime. Or if you see the converge,
> > > > > you might get help from the auto converge support by the hypervisors
> > > > > like KVM where it tries to throttle the VCPU then you can't reach the
> > > > > wire speed.
> > > >
> > > > Will only work for some device types.
> > > >
> > >
> > > Yes, that's the point. Parav said he doesn't see the issue, it's
> > > probably because he is testing a virtio-net and so the vCPU is
> > > automatically throttled. It doesn't mean it can work for other virito
> > > devices.
> >
> > Only for TX, and I'm pretty sure they had the foresight to test RX not
> > just TX but let's confirm. Parav did you test both directions?
> 
> RX speed somehow depends on the speed of refill, so throttling helps
> more or less.

It doesn't depend on speed of refill you just underrun and drop
packets. then your nice 10usec latency becomes more like 10sec.

> >
> > > >
> > > >
> > > > > >
> > > > > > >
> > > > > > > > There is no requirement for mandating PRI either.
> > > > > > > > So it is unusable.
> > > > > > >
> > > > > > > It's not about mandating, it's about doing things in the correct layer. If PRI is
> > > > > > > slow, PCI can evolve for sure.
> > > > > > You should try.
> > > > >
> > > > > Not my duty, I just want to make sure things are done in the correct
> > > > > layer, and once it needs to be done in the virtio, there's nothing
> > > > > obviously wrong.
> > > >
> > > > Yea but just vague questions don't help to make sure eiter way.
> > >
> > > I don't think it's vague, I have explained, if something in the virito
> > > slows down the PRI, we can try to fix them.
> >
> > I don't believe you are going to make PRI fast. No one managed so far.
> 
> So it's the fault of PRI not virito, but it doesn't mean we need to do
> it in virtio.

I keep saying with this approach we would just say "e1000 emulation is
slow and encumbered this is the fault of e1000" and never get virtio at
all.  Assigning blame only gets you so far.

> >
> > > Missing functions in
> > > platform or transport is not a good excuse to try to workaround it in
> > > the virtio. It's a layer violation and we never had any feature like
> > > this in the past.
> >
> > Yes missing functionality in the platform is exactly why virtio
> > was born in the first place.
> 
> Well the platform can't do device specific logic. But that's not the
> case of dirty page tracking which is device logic agnostic.

Not true platforms have things like NICs on board and have for many
years. It's about performance really. So I'd like Parav to publish some
experiment results and/or some estimates.


> >
> > > >
> > > > > > In the current state, it is mandating.
> > > > > > And if you think PRI is the only way,
> > > > >
> > > > > I don't, it's just an example where virtio can leverage from either
> > > > > transport or platform. Or if it's the fault in virtio that slows down
> > > > > the PRI, then it is something we can do.
> > > > >
> > > > > >  than you should propose that in the dirty page tracking series that you listed above to not do dirty page tracking. Rather depend on PRI, right?
> > > > >
> > > > > No, the point is to not duplicate works especially considering virtio
> > > > > can't do better than platform or transport.
> > > >
> > > > If someone says they tried and platform's migration support does not
> > > > work for them and they want to build a solution in virtio then
> > > > what exactly is the objection?
> > >
> > > The discussion is to make sure whether virtio can do this easily and
> > > correctly, then we can have a conclusion. I've stated some issues
> > > above, and I've asked other questions related to them which are still
> > > not answered.
> > >
> > > I think we had a very hard time in bypassing IOMMU in the past that we
> > > don't want to repeat.
> > >
> > > We've gone through several methods of logging dirty pages in the past
> > > (each with pros/cons), but this proposal never explains why it chooses
> > > one of them but not others. Spec needs to find the best path instead
> > > of just a possible path without any rationale about why.
> >
> > Adding more rationale isn't a bad thing.
> > In particular if platform supplies dirty tracking then how does
> > driver decide which to use platform or device capability?
> > A bit of discussion around this is a good idea.
> >
> >
> > > > virtio is here in the
> > > > first place because emulating devices didn't work well.
> > >
> > > I don't understand here. We have supported emulated devices for years.
> > > I'm pretty sure a lot of issues could be uncovered if this proposal
> > > can be prototyped with an emulated device first.
> > >
> > > Thanks
> >
> > virtio was originally PV as opposed to emulation. That there's now
> > hardware virtio and you call software implementation "an emulation" is
> > very meta.
> 
> Yes but I don't see how it relates to dirty page tracking. When we
> find a way it should work for both software and hardware devices.
> 
> Thanks

It has to work well on a variety of existing platforms. If it does then
sure, why would we roll our own.

-- 
MST