On 14/11/16 15:52, Joerg Roedel wrote:

On Mon, Nov 14, 2016 at 12:00:47PM +0000, Robin Murphy wrote:

quoted

If we've already made the decision to move away from bus ops, I don't
see that it makes sense to deliberately introduce new dependencies on
them. Besides, as it stands, this patch literally implements "tell the
iommu-core which hardware-iommus exist in the system and a seperate
iommu_ops ptr for each of them" straight off.

Not sure which code you are looking at, but as I see it we have only
per-device iommu-ops now (with this patch). That is different from
having core-visible hardware-iommu instances where devices could link
to.

The per-device IOMMU ops are already there since 57f98d2f61e1. This
patch generalises the other end, moving the "registering an IOMMU
instance" (i.e. iommu_fwentry) bit into the IOMMU core, from being
OF-specific. I'd be perfectly happy if we rename iommu_fwentry to
iommu_instance, fwnode_iommu_set_ops() to iommu_register_instance(), and
such if that makes the design intent clearer.

If you'd also prefer to replace iommu_fwspec::ops with an opaque
iommu_fwspec::iommu_instance pointer so that things are a bit more
centralised (and users are forced to go through the API rather then call
ops directly), I'd have no major objection either. My main point is that
we've been deliberately putting the relevant building blocks in place -
the of_iommu_{get,set}_ops stuff was designed from the start to
accommodate per-instance ops, via the ops pointer *being* the instance
token; the iommu_fwspec stuff is deliberately intended to provide
per-device ops on top of that. The raw functionality is either there in
iommu.c already, or moving there in patches already written, so if it
doesn't look right all we need to focus on is making it look right.

Also the rest of iommu-core code still makes use of the per-bus ops. The
per-device ops are only used for the of_xlate fn-ptr.

Hence my aforementioned patches intended for 4.10, directly following on
from introducing iommu_fwspec in 4.9:

http://www.mail-archive.com/iommu at lists.linux-foundation.org/msg14576.html

...the purpose being to provide a smooth transition from per-bus ops to
per-device, per-instance ops. Apply those and we're 90% of the way there
for OF-based IOMMU drivers (not that any of those actually need
per-instance ops, admittedly; I did prototype it for the ARM SMMU ages
ago, but it didn't seem worth the bother). Lorenzo's series broadens the
scope to ACPI-based systems and moves the generically-useful parts into
the core where we can easily build on them further if necessary. The
major remaining work is to convert external callers of the current
bus-dependent functions like iommu_domain_alloc(), iommu_present(), etc.
to device-based alternatives.

Robin.