[openib-general] [PATCH] mthca: memfree completion with error workaround

[Fabian Tillier]

Hi Roland,

On 6/12/06, Roland Dreier <rdreier at cisco.com> wrote:
>    Michael> It might not be too bad - there's a single additional
>    Michael> test on fastpath, and I am guessing both wqe_index and
>    Michael> rq.max should be in registers already.  Once wqe_index is
>    Michael> out of rq.max range we are on slow path.
>
> But it bloats the function and adds to i-cache footprint.  I'm sure it
> benchmarks fine but it adds to general cache usage that pushes useful
> code out of cache.
>
> Unfortunately I don't see a clean way to move it out of line.

Why not just have multiple implemenations of the function, and setup
the function pointer in the verbs according to what firmware and
device is in use?  That way devices not affected could continue to use
the optimized version...

Just a thought - it does have the drawback of having multiple similar functions.

- Fab