Re: [PATCH v1 2/9] iommu/arm-smmu-v3: Add alloc_id/free_id functions to arm_smmu_invs

[Jason Gunthorpe <jgg@nvidia.com>]

On Fri, Jan 16, 2026 at 08:58:02AM -0800, Nicolin Chen wrote:
> On Fri, Jan 16, 2026 at 10:41:20AM -0400, Jason Gunthorpe wrote:
> > On Thu, Jan 15, 2026 at 09:13:00PM -0800, Nicolin Chen wrote:
> > > On Fri, Jan 02, 2026 at 11:57:15AM -0400, Jason Gunthorpe wrote:
> > > > On Tue, Dec 30, 2025 at 10:52:53AM -0800, Nicolin Chen wrote:
> > > > > Hi Jason,
> > > > > 
> > > > > On Fri, Dec 19, 2025 at 01:05:51PM -0400, Jason Gunthorpe wrote:
> > > > > > +static int arm_smmu_get_tag(struct arm_smmu_domain *smmu_domain,
> > > > > > +			    struct arm_smmu_master *master,
> > > > > > +			    struct arm_vsmmu *vsmmu,
> > > > > > +			    struct arm_smmu_iotlb_tag *tag, bool no_alloc)
> > > > > [...]
> > > > > > +	case ARM_SMMU_DOMAIN_S2:
> > > > > > +		if (smmu_domain->nest_parent) {
> > > > > > +			/* FIXME we can support attaching a nest_parent without
> > > > > > +			 * a vsmmu, but to do that we need to fix
> > > > > > +			 * arm_smmu_get_id_from_invs() to never return the vmid
> > > > > > +			 * of a vsmmu. Probably by making a
> > > > > > +			 * INV_TYPE_S2_VMID_VSMMU */
> > > > > > +			id = vsmmu->vmid;
> > > > > > +			return 0;
> > > > > > +		}
> > > > > 
> > > > > Would you mind elaborating why arm_smmu_get_id_from_invs() can't
> > > > > return vsmmu->vmid to share with a naked S2 STE?
> > > > 
> > > > A "naked" S2 domain doesn't have a pointer to the vsmmu, so it is
> > > > impossible to get vsmmu->vmid.
> > > 
> > > An S2 parent domain should be per VM. And a vSMMU on top of an S2
> > > should be per SMMU. So, it could have stored a list of vSMMUs and
> > > device attaching to a naked S2 could match its master->smmu with
> > > vSMMU->smmu in the list?
> > 
> > That would cause lifecycle problems if the vSMMU is destroyed
> > while the nake S2 is still attached and trying to use the vSMMU's
> > VMID.
> 
> Well, if vSMMU code does the same get-build-merge in vsmmu_init()
> and build-unref in vsmmu_destroy(), VMID is basically managed by
> the invalidation array. Yes, a naked S2 attachment would still use
> the shared VMID, but I think that's fine for a nesting parent?

The VMID has to be managed by the vsmmu itself, it would be a big
confusing mess any other way. There are multiple invalidation lists,
so you can't rely on that to refcount things.

Jason