[PATCH v5 5/9] iommu/arm-smmu-v3: Refactor write_ctx_desc

[Michael Shavit]

On Tue, Aug 15, 2023 at 7:38 PM Jason Gunthorpe <jgg at nvidia.com> wrote:
>
> On Tue, Aug 15, 2023 at 01:20:04PM +0800, Michael Shavit wrote:
> > On Thu, Aug 10, 2023 at 11:39 PM Jason Gunthorpe <jgg at nvidia.com> wrote:
> > >
> > > Actually, I don't think this even works as nothing on the PASID path
> > > adds to the list that arm_smmu_write_ctx_desc_devices() iterates over ??
> > >
> > > Then the remaining two calls:
> > >
> > > arm_smmu_share_asid(struct mm_struct *mm, u16 asid)
> > >         arm_smmu_write_ctx_desc_devices(smmu_domain, 0, cd);
> > >
> > >         This is OK only if the sketchy assumption that the CD
> > >         we extracted for a conflicting ASID is not asigned to a PASID.
> > >
> > > static void arm_smmu_mm_release(struct mmu_notifier *mn, struct mm_struct *mm)
> > >         arm_smmu_write_ctx_desc_devices(smmu_domain, mm->pasid, &quiet_cd);
> > >
> > >         This doesn't work because we didn't add the master to the list
> > >         during __arm_smmu_sva_bind and this path is expressly working
> > >         on the PASID binds, not the RID binds.
> >
> > Actually it is working on the RID attached domain (as returned by
> > iommu_get_domain_for_dev() at sva_bind time) not the SVA domain
> > here...
>
> That can't be right, the purpose of that call and arm_smmu_mm_release is to
> disable the PASID that is about the UAF the mm's page table.
>
> Jason

For the sake of this message, let's call "primary domain" whatever RID
domain was attached to a master at the time set_dev_pasid() was called
on that master. That RID domain is locked in while SVA is enabled and
cannot be detached.

The arm-smmu-v3-sva.c implementation creates a mapping between an SVA
domain and this primary domain (through the sva domain's mm). In
arm_smmu_mm_release, the primary domain is looked up and
arm_smmu_write_ctx_desc() is called on all masters that this domain is
attached to.