[PATCH v3 6/8] iommu/arm-smmu-v3: Move CD table to arm_smmu_master

[Jason Gunthorpe]

On Wed, Aug 02, 2023 at 07:19:12PM +0800, Michael Shavit wrote:
> On Wed, Aug 2, 2023 at 7:53 AM Jason Gunthorpe <jgg at nvidia.com> wrote:
> >
> > On Wed, Aug 02, 2023 at 02:35:23AM +0800, Michael Shavit wrote:
> > > @@ -2465,6 +2440,22 @@ static int arm_smmu_attach_dev(struct iommu_domain *domain, struct device *dev)
> > >       if (smmu_domain->stage != ARM_SMMU_DOMAIN_BYPASS)
> > >               master->ats_enabled = arm_smmu_ats_supported(master);
> > >
> > > +     if (smmu_domain->stage == ARM_SMMU_DOMAIN_S1) {
> > > +             if (!master->cd_table.cdtab) {
> > > +                     ret = arm_smmu_alloc_cd_tables(master);
> > > +                     if (ret) {
> >
> > Again, I didn't look very closely at your locking, but what lock is
> > being held to protect the read of master->cd_table.cdtab ?
> 
> The cd_table is only written into (with write_ctx_desc) when something
> attaches or detaches (SVA is a little weird, but it handles locking
> internally, and blocks all non-sva attach/detach calls while enabled).
> The cd_table itself is allocated on first attach, and freed on release.
> 
> Doesn't the iommu framework guarantee that attach_dev (and
> release_device) won't have concurrent calls for a given master through
> the group lock? I can add an internal lock if relying on the iommu
> lock is not OK.

Yes that is right.

So, a comment about that in the struct around those variables would be
helpful (locked by the iommu core using the group mutex)

But the code is fine

Reviewed-by: Jason Gunthorpe <jgg at nvidia.com>

Jason