Re: [RFC v2 1/1] memory: Delete assertion in memory_region_unregister_iommu_notifier

[Peter Xu <peterx@redhat.com>]

On Mon, Jul 13, 2020 at 12:04:16PM +0800, Jason Wang wrote:
> 
> On 2020/7/10 下午9:30, Peter Xu wrote:
> > On Fri, Jul 10, 2020 at 02:34:11PM +0800, Jason Wang wrote:
> > > On 2020/7/9 下午10:10, Peter Xu wrote:
> > > > On Thu, Jul 09, 2020 at 01:58:33PM +0800, Jason Wang wrote:
> > > > > > > - If we care the performance, it's better to implement the MAP event for
> > > > > > > vhost, otherwise it could be a lot of IOTLB miss
> > > > > > I feel like these are two things.
> > > > > > 
> > > > > > So far what we are talking about is whether vt-d should have knowledge about
> > > > > > what kind of events one iommu notifier is interested in.  I still think we
> > > > > > should keep this as answered in question 1.
> > > > > > 
> > > > > > The other question is whether we want to switch vhost from UNMAP to MAP/UNMAP
> > > > > > events even without vDMA, so that vhost can establish the mapping even before
> > > > > > IO starts.  IMHO it's doable, but only if the guest runs DPDK workloads.  When
> > > > > > the guest is using dynamic iommu page mappings, I feel like that can be even
> > > > > > slower, because then the worst case is for each IO we'll need to vmexit twice:
> > > > > > 
> > > > > >      - The first vmexit caused by an invalidation to MAP the page tables, so vhost
> > > > > >        will setup the page table before IO starts
> > > > > > 
> > > > > >      - IO/DMA triggers and completes
> > > > > > 
> > > > > >      - The second vmexit caused by another invalidation to UNMAP the page tables
> > > > > > 
> > > > > > So it seems to be worse than when vhost only uses UNMAP like right now.  At
> > > > > > least we only have one vmexit (when UNMAP).  We'll have a vhost translate()
> > > > > > request from kernel to userspace, but IMHO that's cheaper than the vmexit.
> > > > > Right but then I would still prefer to have another notifier.
> > > > > 
> > > > > Since vtd_page_walk has nothing to do with device IOTLB. IOMMU have a
> > > > > dedicated command for flushing device IOTLB. But the check for
> > > > > vtd_as_has_map_notifier is used to skip the device which can do demand
> > > > > paging via ATS or device specific way. If we have two different notifiers,
> > > > > vhost will be on the device iotlb notifier so we don't need it at all?
> > > > But we can still have iommu notifier that only registers to UNMAP even after we
> > > > introduce dev-iotlb notifier?  We don't want to do page walk for them as well.
> > > > TCG should be the only one so far, but I don't know.. maybe there can still be
> > > > new ones?
> > > 
> > > I think you're right. But looking at the codes, it looks like the check of
> > > vtd_as_has_map_notifier() was only used in:
> > > 
> > > 1) vtd_iommu_replay()
> > > 2) vtd_iotlb_page_invalidate_notify() (PSI)
> > > 
> > > For the replay, it's expensive anyhow. For PSI, I think it's just about one
> > > or few mappings, not sure it will have obvious performance impact.
> > > 
> > > And I had two questions:
> > > 
> > > 1) The codes doesn't check map for DSI or GI, does this match what spec
> > > said? (It looks to me the spec is unclear in this part)
> > Both DSI/GI should cover maps too?  E.g. vtd_sync_shadow_page_table() in
> > vtd_iotlb_domain_invalidate().
> 
> 
> I meant the code doesn't check whether there's an MAP notifier :)

It's actually checked, because it loops over vtd_as_with_notifiers, and only
MAP notifiers register to that. :)

But I agree with you that it should be cleaner to introduce the dev-iotlb
notifier type.

> 
> 
> > 
> > > 2) for the replay() I don't see other implementations (either spapr or
> > > generic one) that did unmap (actually they skip unmap explicitly), any
> > > reason for doing this in intel IOMMU?
> > I could be wrong, but I'd guess it's because vt-d implemented the caching mode
> > by leveraging the same invalidation strucuture, so it's harder to make all
> > things right (IOW, we can't clearly identify MAP with UNMAP when we receive an
> > invalidation request, because MAP/UNMAP requests look the same).
> > 
> > I didn't check others, but I believe spapr is doing it differently by using
> > some hypercalls to deliver IOMMU map/unmap requests, which seems a bit close to
> > what virtio-iommu is doing.  Anyway, the point is if we have explicit MAP/UNMAP
> > from the guest, logically the replay indeed does not need to do any unmap
> > because we don't need to call replay() on an already existing device but only
> > for e.g. hot plug.
> 
> 
> But this looks conflict with what memory_region_iommu_replay( ) did, for
> IOMMU that doesn't have a replay method, it skips UNMAP request:
> 
>     for (addr = 0; addr < memory_region_size(mr); addr += granularity) {
>         iotlb = imrc->translate(iommu_mr, addr, IOMMU_NONE, n->iommu_idx);
>         if (iotlb.perm != IOMMU_NONE) {
>             n->notify(n, &iotlb);
>         }
> 
> I guess there's no knowledge of whether guest have an explicit MAP/UMAP for
> this generic code. Or replay implies that guest doesn't have explicit
> MAP/UNMAP?

I think it matches exactly with a hot plug case?  Note that when IOMMU_NONE
could also mean the translation does not exist.  So it's actually trying to map
everything that can be translated and then notify().

> 
> (btw, the code shortcut the memory_region_notify_one(), not sure the reason)

I think it's simply because memory_region_notify_one() came later. :)

> 
> 
> >   VT-d does not have that clear interface, so VT-d needs to
> > maintain its own mapping structures, and also vt-d is using the same replay &
> > page_walk operations to sync all these structures, which complicated the vt-d
> > replay a bit.  With that, we assume replay() can be called anytime on a device,
> > and we won't notify duplicated MAPs to lower layer like vfio if it is mapped
> > before.  At the meantime, since we'll compare the latest mapping with the one
> > we cached in the iova tree, UNMAP becomes possible too.
> 
> 
> AFAIK vtd_iommu_replay() did a completely UNMAP:
> 
>     /*
>      * The replay can be triggered by either a invalidation or a newly
>      * created entry. No matter what, we release existing mappings
>      * (it means flushing caches for UNMAP-only registers).
>      */
>     vtd_address_space_unmap(vtd_as, n);
> 
> Since it doesn't do any comparison with iova tree. Will this cause
> unnecessary UNMAP to be sent to VFIO?

I feel like that can be removed now, but needs some testings...

Thanks,

-- 
Peter Xu