Re: 3.6-rc7 boot crash + bisection

[Alex Williamson <alex.williamson-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

On Wed, 2012-09-26 at 17:10 +0200, Roedel, Joerg wrote:
> On Wed, Sep 26, 2012 at 08:35:59AM -0600, Alex Williamson wrote:
> > Hmm, that throws a kink in iommu groups.  So perhaps we need to make an
> > alias interface to iommu groups.  Seems like this could just be an extra
> > parameter to iommu_group_get and iommu_group_add_device (empty in the
> > typical case).  Then we have the problem of what's the type for an
> > alias?  For AMI-Vi, it's a u16, but we need to be more generic than
> > that.  Maybe iommu groups should just treat it as a void* so iommus can
> > use a pointer to some structure or a fixed value like a u16 bus:slot.
> > Thoughts?
> 
> Good question. The iommu-groups are part of the IOMMU-API, with an
> interface to the IOMMU drivers and one to the users of IOMMU-API. So the
> alias handling itself should be a function of the interface to the IOMMU
> driver. In general the interface should not be bus specific.
> 
> So a void pointer seems the only logical choice then. But I would not
> limit its scope to alias handling. How about making it a bus-private
> pointer where IOMMU driver store bus-specific information. That way we
> make sure that there is one struct per bus-type for this pointer, and
> not one structure per IOMMU driver.

I thought of another approach that may actually be more 3.6 worthy.
What if we just make the iommu driver handle it?  For instance,
amd_iommu can walk the alias table looking for entries that use the same
alias and get the device via pci_get_bus_and_slot.  If it finds a device
with an iommu group, it attaches the new device to the same group,
hiding anything about aliases from the group layer.  It just groups all
devices within the range.  I think the only complication is making sure
we're safe around device hotplug while we're doing this.  Thanks,

Alex