Re: [PATCH] SLUB: revert direct page allocator pass through

[Matt Mackall <mpm@selenic.com>]

On Fri, 2009-01-23 at 10:03 -0500, Christoph Lameter wrote:
> On Fri, 23 Jan 2009, Nick Piggin wrote:
> 
> > Hmm, it lists quite a number of advantages that I guess are being
> > reverted too? What was the test case(s) that prompted this commit
> > in the first place? Better ensure it doesn't slow down...
> 
> The advantage was mainly memory savings and abilty to redefined kmallocs
> to go directly to the page allocator. Totally avoids slab allocator overhead.
> 
> I thought higher order allocations were not supposed to be used in
> performance critical paths? Didnt you want to do everything with order-0
> allocs?
> 
> It seems that we currently need the slab allocators to compensate for the
> performance problems in the page allocator for these higher order allocs.
> I'd rather have the page allocator fixed but things are as they are.

I still think we should experiment with changing the hierarchy:

Rename all the core get_free_page* functions to buddy_*

Make SL*B call into buddy_* with a default order of N (>=0)

Replace the old get_free_page* functions with simple wrappers that
call into SL*B for order <= N or buddy_* for order >= N

This tackles several problems at once:

- fragmentation of SL*B due to small pages
- poor performance of get_free_pages moderate orders
- poor cache-locality for get_free_pages

-- 
http://selenic.com : development and support for Mercurial and Linux