[PATCH 3/3] iommu/core: split mapping to page sizes as supported by the hardware

[Ohad Ben-Cohen]

On Tue, Sep 13, 2011 at 1:44 PM, Roedel, Joerg <Joerg.Roedel at amd.com> wrote:
> Not necessarily. You could implement this side-by-side with the old code
> until all drivers are converted and remove the old code then. This keeps
> bisectability.

Ok.

>> > Intel IOMMU does not support arbitrary page-sizes, afaik.
>>
>> It does; besides the usual 4K it has "super page sizes" support of
>> 2MB, 1GB, 512GB and 1TB.
>
> But the value ~0xfffUL indicates support for 4k, 8k, 16k .. 2^63, no?

Yes, I have done this intentionally, in order to retain the existing
behavior for IOMMU drivers which are already capable of handling
arbitrary page sizes (intel-iommu handles this in software, see
hardware_largepage_caps() and the code that uses it).

Long term, it might make more sense to remove
hardware_largepage_caps() (and the logic around it) and instead just
declare the real page sizes the hardware supports when calling
register_iommu(), but I guess it's up to Intel guys. For now it's just
safer to declare ~0xfffUL which really means: keep calling me with
sizes and alignments that are an order of 4KB, just like you always
did.