Re: [PATCH 1/1] iommu: Set owner token to sva and nested domains

[Baolu Lu]

On 12/7/23 9:36 PM, Jason Gunthorpe wrote:
> On Thu, Dec 07, 2023 at 09:56:10AM +0000, Robin Murphy wrote:
> > On 2023-12-07 2:19 am, Lu Baolu wrote:
> > > Commit a9c362db3920 ("iommu: Validate that devices match domains") added
> > > an owner token to an iommu_domain. This token is checked during domain
> > > attachment to RID or PASID through the generic iommu interfaces.
> > >
> > > The sva and nested domains are attached to device or PASID through the
> > > generic iommu interfaces. Therefore, they require the owner token to be
> > > set during allocation. Otherwise, they fail to attach.
> > Oops, I missed that iommu_sva_domain_alloc() is a thing - when did we get
> > such a confusing proliferation of domain allocation paths? Sigh...
> We have alot of different kinds of domains now, APIs that are giant
> multiplexers are not good.
>
> What I've been wanting to do for a while is to have the drivers call a
> helper to allocate their domain struct and the helper would initialize
> the common iommu_domain instead of doing this after the op
> returns. This is more typical kernel pattern and avoids some of the
> confusion about when struct members are valid or not (notice some of
> driver code needs iommu_domain stuff set earlier and we confusingly
> initialize things twice :()
>
> > I think we should set the owner generically there, since presumably it's
> > being missed for SMMUv3/AMD/etc. SVA domains as well. Nested domains are
> > supposed to be OK since both ->domain_alloc_user callsites are covered, or
> > is there some other sneaky path I've also missed?
> Indeed, I also think the first hunk is not needed, the second hunk was
> missed.

Oh, yes! I overlooked that iommufd has already done that for nested
domain. I will update it with a v2.

Best regards,
baolu