Re: [Qemu-devel] [PATCH RFC 0/4] sPAPR: Support multiple PEs in one PHB

[David Gibson]

On Fri, Sep 18, 2015 at 04:30:12PM +1000, Gavin Shan wrote:
> This patchset bases on David Gibson's git tree: 
> git://github.com/dgibson/qemu.git
> (branch: vfio). And it requires host kernel changes which is being reviewed
> this moment.
> 
> https://patchwork.ozlabs.org/patch/519135/
> https://patchwork.ozlabs.org/patch/519136/
> 
> Currently, EEH works with the assumption that every sPAPRPHBState, which
> is associated with VFIO container in VFIO case, only has one attached
> IOMMU group (PE). The request EEH opertion (like reset) is applied to
> all PEs attached to the specified sPAPRPHBState. It breaks the affected
> boundary of the EEH operation if the sPAPRPHBState supports multiple
> IOMMU groups (PEs).
> 
> The patchset intends to resolve above issue by using the newly exposed
> EEH v2 API interface, which accepts IOMMU group (PE) to specify the
> affected domain of the requested EEH operation: Every PE is identified
> with PE address, which is the (PE's primary bus ID + 1) previously.
> After this patchset, it's changed to (IOMMU group ID + 1). The PE adress
> is passed on every requested EEH operation from guest so that it can
> be passed to host to affect the target PE only.

Sorry Gavin,

I've been working on this problem from the other end - trying to get
qemu to work safely with EEH to the limited extent that's possible
with the existing kernel interface, and also getting rid of the
special VFIO host bridge nonsense at the same time.

My code is at git://github.com/dgibson/qemu.git, branch 'eeh'.  I plan
to post as soon as I've given it at least some minimal testing.

It collides with these patches so they'll need a substantial
reworking.

Fwiw, I'd also prefer if you tackled this by first altering the PE
config_addr allocation in qemu so that we can get it working with the
existing broken kernel interface, as long as there is only one vfio
group per container (put there could be multiple containers on a PHB).

Then, the fixed kernel interface can be added on top of that, to
remove the one-group-per-container restriction.

-- 
David Gibson                    | I'll have my music baroque, and my code
david AT gibson.dropbear.id.au  | minimalist, thank you.  NOT _the_ _other_
                                | _way_ _around_!
http://www.ozlabs.org/~dgibson

pgpr6JWd5vbKC.pgp
Description: PGP signature