Re: [PATCH v4 07/13] iommu/arm-smmu-v3: Keep track of attached ssids

[Nicolin Chen <nicolinc@nvidia.com>]

On Fri, Jul 14, 2023 at 05:30:42PM +0800, Michael Shavit wrote:
 
> > And I don't quite get this part. Prior to this change, it issues
> > one ATC_INV command covering all ATC entries per comments inside
> > arm_smmu_atc_inv_to_cmd(). But now we replace that single command
> > with all attached subdomains in the list? Any reason for such a
> > change here?
> 
> Because we don't necessarily want to invalidate all PASID-domains
> attached to a master. If arm_smmu_atc_inv_domain() is called on a
> domain that is only attached with Pasid, we can restrict the
> invalidations to those specific PASID by looping over them.
>
> But yeah,
> you're right that we could potentially optimize this?
> * Skip the per-pasid invalidations if the domain is also attached to
> this master without PASID as we have to invalidate all its pasids in
> that case anyways. It's hard to imagine clients attaching a domain
> both with pasid and without pasid to the same device but could be
> possible.
> * Always invalidate all pasids by issuing atc invalidations on SSID 0.
> This sounds like the wrong trade-off??

Well, firstly it's kinda odd to have this optimization hidden in
a big rework patch. And I am not sure if it alone would work for
all use cases, as it impacts the arm_smmu_atc_inv_domain() that
passes in a zero ssid, in which case the affected function is not
used by a pasid case all the time:
	/*
	 * ATS and PASID:
...
	 * When using STRTAB_STE_1_S1DSS_SSID0 (reserving CD 0 for non-PASID
	 * traffic), translation requests without PASID create ATC entries
	 * without PASID, which must be invalidated with substream_valid clear.
	 * This has the unpleasant side-effect of invalidating all PASID-tagged
	 * ATC entries within the address range.
	 */

Thanks
Nicolin