Re: [PATCH 02/11] iommu-api: Add iommu_map and iommu_unmap functions

[Avi Kivity <avi@redhat.com>]

On 02/07/2010 12:50 PM, Joerg Roedel wrote:
> On Sun, Feb 07, 2010 at 11:38:30AM +0200, Avi Kivity wrote:
>    
>> On 01/28/2010 01:37 PM, Joerg Roedel wrote:
>>      
>>> These two functions provide support for mapping and
>>> unmapping physical addresses to io virtual addresses. The
>>> difference to the iommu_(un)map_range() is that the new
>>> functions take a gfp_order parameter instead of a size. This
>>> allows the IOMMU backend implementations to detect easier if
>>> a given range can be mapped by larger page sizes.
>>> These new functions should replace the old ones in the long
>>> term.
>>>
>>>        
>> These seem to be less flexible in the long term.  Sure, it is easier for
>> the backend to map to multiple page sizes if your iommu supports
>> arbitrary power-of-two page sizes, but we should make the APIs easier
>> for the callers, not the backend.
>>      
> These functions are as flexible as the old ones which just tok a size.
> The benefit of the new interface is that is makes the ability of the
> IOMMU to map the area with a large page (an get the benefit of fewer
> hardware tlb walks) visible to the caller. With the old interface the
> caller is tempted to just map ist whole area using 4kb page sizes.
> It gives a bit more complexity to the caller, thats right. But given
> that the page allocator in Linux always returns pages which are aligned
> to its size takes a lot of that complexity away.
>
>    

You are right - I was thinking of the kvm use case which is range 
oriented, but the ordinary kernel interfaces are gfp_order oriented.

-- 
error compiling committee.c: too many arguments to function