Hi Joerg,

On Mon, Oct 10, 2011 at 5:36 PM, Roedel, Joerg [off-list ref] wrote:

The master branch is best to base your patches on for generic work.

Oh, great. thanks.

It can happen when there is a bug somewhere :)

Hmm, bug ? ;)

Ok, I'll add a BUG_ON :)

Yes, somthing like that. Probably the iommu_ops->unmap function should
be turned into a unmap_page function call which only takes an iova and
no size parameter. The iommu-driver unmaps the page pointing to that
iova and returns the size of the page unmapped. This still allows the
simple implementation for the unmap-call.

Yes, exactly. It will take some time to migrate all drivers (today we
have 4 drivers, each of which is implementing a slightly different
->unmap() semantics), but at least let's not accept any new driver
that doesn't adhere to this, otherwise it's going to be even harder
for the API to evolve.

This change is no requirement for this patch-set, but if we agree on it
this patch-set should keep that direction in mind.

Definitely, thanks.

We don't need to really
enforce the calls to be symetric. But we can define that we only give
the guarantee about what will be unmapped when the calls are symetric.

Sounds good to me. I'll add this to the kernel doc patch (which I'll
submit after this patch set materializes), and when/if we move to
symmetric only, we will update it.

The alternative is that we implement page-splitting in the iommu_unmap
function. But that introduces complexity I am not sure we really need.

Yeah, me neither.

Yes, this get_order thing should be changes to size long-term.

Good. That should be a simple change, I'll do it after this patch set.

Thanks,
Ohad.