Re: [PATCH v2 4/5] iommu: Support mm PASID 1:n with sva domains

[Jason Gunthorpe]

On Thu, Aug 31, 2023 at 03:35:36PM +0800, Baolu Lu wrote:
> On 2023/8/31 14:42, Vasant Hegde wrote:
> > > > Also in this function (mm_pasid_drop()), should we check/free sva domains?
> > > No, the driver must support a SVA domain with an empty mm_struct, eg
> > > by hooking release.
> > Sorry. Looks like confused you. Looking into code I got this.
> > 
> > My question was: when PASID is enabled, is there any chance of mm_pasid_drop()
> > getting called before freeing all SVA domains?
> 
> I remember we have discussed this during sva development. If I remember
> it correctly, in any case, mm_pasid_drop() will be called after the
> device is closed. 

Yes, we guarentee this because the SVA domain should be holding a
mmgrab on the mm_struct while it exists. The mmdrop cannot happen
until the SVA domain is freed which puts back that ref.

But drivers must not make any assumptions about this, the lifecycle of
the PASID is a core concern, so long as the driver is asked to attach
a domain to a PASID it should assume the PASID is valid.

A driver should *never* look at the mm_struct PASID, all the examples
of this in-tree today are wrong.

Jason