On 11/11/16 16:27, Joerg Roedel wrote:

On Fri, Nov 11, 2016 at 04:17:37PM +0000, Robin Murphy wrote:

quoted

In the original of_iommu_configure design, the thought was that an ops
structure could be IOMMU-instance-specific (hence the later-removed
"priv" member), so I suppose right now it is mostly a hangover from
that. However, it's also what we initialise a device's fwspec with, so
becomes important again if we're ever going to get past the limitations
of buses-which-are-not-actually-buses[1].

Yeah, I discussed this with a few others at LPC. My current idea is to
tell the iommu-core which hardware-iommus exist in the system and a
seperate iommu_ops ptr for each of them. Then every struct device can
link to the iommu-instance it is translated by.

Er, that sounds very much like a description of what we already have in
4.9-rc. Every struct device now has an iommu_fwspec which encapsulates
both an iommu_ops pointer (which can perfectly well be per-instance if
the IOMMU driver wants) and a place for the IOMMU-private data to
replace the mess of archdata.iommu and driver-internal globals.

We are not there yet, but this will give you the same per-device
iommu-ops as implemented here.

With those two patches I linked to, which make the bulk of the IOMMU
core code per-device-ops-aware off the bat, I'd say we *are* already
pretty much there. It's only iommu_domain_alloc() which needs a
device-based alternative, and the non-of_xlate-based IOMMU drivers to
either call iommu_fwspec_init() for themselves, or perhaps for x86
plumbing in DMAR/IVRS equivalents of the IORT parsing to the
infrastructure provided by this series.

Robin.