On Mon, Sep 21, 2015 at 11:42:28AM +1000, David Gibson wrote:

On Sat, Sep 19, 2015 at 04:22:47PM +1000, David Gibson wrote:

quoted

On Fri, Sep 18, 2015 at 09:47:32AM -0600, Alex Williamson wrote:

quoted

On Fri, 2015-09-18 at 16:24 +1000, Gavin Shan wrote:

quoted

This allows to accept IOMMU group (PE) ID from the parameter from userland
when handling EEH operation so that the operation only affects the target
IOMMU group (PE). If the IOMMU group (PE) ID in the parameter from userland
is invalid, all IOMMU groups (PEs) attached to the specified container are
affected as before.

Gavin Shan (2):
  drivers/vfio: Support EEH API revision
  drivers/vfio: Support IOMMU group for EEH operations

 drivers/vfio/vfio_iommu_spapr_tce.c | 50 ++++++++++++++++++++++++++++++++++---
 drivers/vfio/vfio_spapr_eeh.c       | 46 ++++++++++++++++++++++------------
 include/linux/vfio.h                | 13 +++++++---
 include/uapi/linux/vfio.h           |  6 +++++
 4 files changed, 93 insertions(+), 22 deletions(-)

This interface is terrible.  A function named foo_enabled() should
return a bool, yes or no, don't try to overload it to also return a
version.

Sorry, that one's my fault.  I suggested that approach to Gavin
without really thinking it through.

quoted

AFAICT, patch 2/2 breaks current users by changing the offset
of the union in struct vfio_eeh_pe_err.

Yeah, this one's ugly.  We have to preserve the offset, but that means
putting the group in a very awkward place.  Especially since I'm not
sure if there even are any existing users of the single extant union
branch.

Sigh.

Yeah, Perhaps, I should have put "RFC" on the subjects because those
patches are really pre-mature and just intend to bring more discussion
on it.

quoted

quoted

Also, we generally pass group
file descriptors rather than a group ID because we can prove the
ownership of the group through the file descriptor and we don't need to
worry about races with the group because we can hold a reference to it.

Duh.  I finally realised the better, simpler, obvious solution.

Rather than changing the parameter structure, we should move the
ioctl()s so they're on the group fd instead of the container fd.

Obviously we need to keep it on the container fd for backwards compat,
but I think we should just error out if there is more than one group
in the container there.

We will need a new capability too, obviously.  VFIO_EEH_GROUPFD maybe?

Yes, I agree to route EEH ioctl commands to group fd since EEH ioctl
commands operate on granularity of PE (IOMMU group). However, it requires
to extend current code to support that. I'm not sure if it's good idea as
I explained to David through IRC. Waiting for Alex to judge:

- Adding a callback to "struct vfio_group": platform_ioctl();
- When attaching the group to platform, this function is initialized;
- The EEH ioctl commands are routed to platform_ioctl() in vfio_group_fops_unl_ioctl()

Thanks,
Gavin

-- 
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