On Mon, Feb 18, 2008 at 07:32:49PM +0000, Thiemo Seufer wrote:
> Fold the SB-1 specific implementation of clear_page/copy_page in the
> generic version, and rewrite that one in tlbex style. The immediate
> benefits:
> - It converts the compile-time workaround for SB-1 pass 1 prefetches
> to a more efficient run-time check.
> - It allows adjustment of loop unfolling, which helps to reduce the
> number of redundant cdex cache ops.
> - It fixes some esoteric cornercases (the cache line length calculations
> can go wrong, and support for 64k pages without prefetch instructions
> will overflow the addiu immediate).
> - Somewhat better guesses of "good" prefetch values.
>
>
> Signed-off-by: Thiemo Seufer <ths@networkno.de>
> ---
>
> Lmbench3 running on a BCM1480 system shows improvements for some
> benchmarks (three runs with the original kernel, then three runs
> with the patched kernel), most markedly (~5%) for open/close and
> exec:
The patch is certainly a good thing but I somehow doubt for many of the
improvments that they are a result of the patch and not the unavoidable
benchmarking noise.
Queued for 2.6.26,
Ralf
|