On Tue, 2013-01-08 at 18:38 -0800, Michel Lespinasse wrote:

Well no fair, the previous patch (for powerpc as well) has 22
insertions and 93 deletions :)

The benefit is that the new code has lower algorithmic complexity, it
replaces a per-vma loop with O(N) complexity with an outer loop that
finds contiguous slice blocks and passes them to vm_unmapped_area()
which is only O(log N) complexity. So the new code will be faster for
workloads which use lots of vmas.

That said, I do agree that the code that looks for contiguous
available slices looks kinda ugly - just not sure how to make it look
nicer though.

Ok. I think at least you can move that construct:

+               if (addr < SLICE_LOW_TOP) {
+                       slice = GET_LOW_SLICE_INDEX(addr);
+                       addr = (slice + 1) << SLICE_LOW_SHIFT;
+                       if (!(available.low_slices & (1u << slice)))
+                               continue;
+               } else {
+                       slice = GET_HIGH_SLICE_INDEX(addr);
+                       addr = (slice + 1) << SLICE_HIGH_SHIFT;
+                       if (!(available.high_slices & (1u << slice)))
+                               continue;
+               }

Into some kind of helper. It will probably compile to the same thing but
at least it's more readable and it will avoid a fuckup in the future if
somebody changes the algorithm and forgets to update one of the
copies :-)

Cheers,
Ben.