Re: [PATCH v7 1/3] iommufd: Add data structure for Intel VT-d stage-1 cache invalidation

From: Jason Gunthorpe
Date: Mon Jan 08 2024 - 08:52:18 EST


On Mon, Jan 08, 2024 at 04:07:12AM +0000, Tian, Kevin wrote:
> > > In concept w/o vSVA it's still possible to assign sibling vdev's to
> > > a same VM as each vdev is allocated with a unique pasid to mark vRID
> > > so can be differentiated from each other in the fault/error path.
> >
> > I thought the SIOV plan was that each "vdev" ie vpci function would
> > get a slice of the pRID's PASID space statically selected at creation?
> >
> > So SVA/etc doesn't matter, you reliably get a disjoint set of pRID &
> > pPASID into each VM.
> >
> > From that view you can't identify the iommufd dev_id without knowing
> > both the pRID and pPASID which will disambiguate the different SIOV
> > iommufd dev_id instances sharing a rid.
>
> true when assigning those instances to different VMs.
>
> Here I was talking about assigning them to a same VM being a problem.
> with rid sharing plus same ENQCMD pPASID potentially used on both
> instances there'd be ambiguity in vSVA e.g. iopf to identify dev_id.

Oh you imaging sharing the pPASID if things have the same translation?
I guess I can see why, but given where things are overall I'd say just
don't do that.

Indeed we can't do that because it makes the vRID unknowable.

(again I continue to think that vt-d cache design is messed up, using
the PASID for the cache tag is a *terrible* design, and causes exactly
these kinds of problems)

> for errors related to descriptor fetch the driver can tell the command
> by looking at the head pointer of the invalidation queue.
>
> command completion is indirectly detected by inserting a wait descriptor
> as fence. completion timeout error is reported in an error register. but
> this register doesn't record pasid, nor does the command location. if there
> are multiple pending devtlb invalidation commands upon timeout
> error the spec suggests the driver to treat all of them timeout as the
> register can only record one rid.

Makes sense, or at least you have to re-issue them one by one

> this is kind of moot. If the driver submits only one command (plus wait)
> at a time it doesn't need hw's help to identify the timeout command.
> If the driver batches invalidation commands it must treat all timeout if
> an timeout error is reported.

Yes

> from this angle whether to record pasid doesn't really matter.

At least for error handling..

Jason