On Wed, 2009-03-25 at 23:30 +0000, Dave Airlie wrote:
> 
> I've no idea when a fault is likely in the fast case, i.e. will it happen
> usually on the first page etc, because if it happens on the last page and 
> you fallback and restart the whole copy, I would think that would be 
> sub-optimal, granted it could get ugly quick, but this code has already 
> hit a few branches on the tree.

The data in question is presumably 'hot', and so unlikely to be swapped
out. Testing the slow paths doing a complete copy is fairly easy, while
testing partial copies would be considerably more difficult. The
combination of these two seems to encourage a careful and simplistic
slow path.

We can always make it 'faster' in the future, but having it wrong from
the first seems sub-optimal.

-- 
keith.packard@intel.com