On Friday 02 January 2015 17:33:53 Murali Karicheri wrote:

I have no experience with IOMMU and may not offer much help here as I 
originally wrote above. Will Deacon has added this API and probably able 
to offer some help in this discussion.

Will Deacon,

Any comment?

It's complicated :(

Looking at the iommu documentation and of_iommu.c, I get a feeling that 
this API is not really used at present as there are no callers of 
of_iommu_set_ops() and I assume this is a WIP.

Right, but we have patches for some iommu drivers based on this API,
and we should migrate all of them eventually.

I believe the way it is 
expected to work is to have the iommu driver of the master IOMMU devices 
call of_iommu_set_ops(). The device node of this master IOMMU device is 
specified as a phandle in the OF node of the device (various bus devices 
such as platform, PCI etc). This allow to retrieve the iommu ops though 
the of_iommu_configure() API and use it in arch_setup_dma_ops(). So my 
gut feeling is that for PCI devices, as there are no DT node, the root 
bus node may specify iommus phandle to IOMMU master OF nodes.

Yes, but we also need to pass a PCI device specific identifier along
with the root bus node, because some iommu drivers take the PCI
bus/device/function number into account for creating per-function
i/o page tables.

W.r.t your comment "We may want to address the comment in
of_iommu_configure about parent nodes. We should be sure these changes 
work with how we would do searching parent nodes",

I believe, the parent node search itself should work the same way in the 
case of PCI as with platform bus case. PCI's case, we are providing the 
OF node of the root bus host bridge. Why should this be any different in 
terms of search?

I see a potential issue with dma-ranges as described in the notes below.
As noted below the usage of dma-range for iommu is to be determined. For 
keystone, the of_iommu_configure() always return false as we don't use 
the iommu. But don't know if this has any consequences for other 
platforms. Or I got your questions wrong. Any help here from others on 
the list?

========================================================================
One possible extension to the above is to use an "iommus" property along 
with a "dma-ranges" property in a bus device node (such as PCI host 
bridges). This can be useful to describe how children on the bus relate 
to the IOMMU if they are not explicitly listed in the device tree (e.g. 
PCI devices). However, the requirements of that use-case haven't been 
fully determined yet. Implementing this is therefore not recommended 
without further discussion and extension of this binding.
=========================================================================

This probably won't ever apply to PCI devices, so let's ignore it for now.
For the moment (and for PCI), we should assume that we either configure
an iommu directly or we use dma-ranges if no iommu is in use.

	Arnd