I've performed a larger modification of slab.c, as a preparation for 
NUMA awareness.

Could you please test it? Especially SMP, or non-x86 platforms would be 
interesting.

A kernprof profile, with function counts, for several loads would be 
great, but I don't know if kernprof works with recent 2.5 kernels.

With slab debugging enabled, it's possible to measure the efficiency of 
the headarrays: The block after the last ":" list the hits/misses:

names_cache  [snip] : 262924    118 263042      0

262924 times kmem_cache_alloc(names_cache) found an object in the 
headarray, only 118 times it had to go back into the slab lists.
263042 times kmem_cache_free(names_cache) could store an object directly 
in the per-cpu array, it never had to flush the array. Note that 
additional flushing can happen in a timer, which is not counted right now.

The patch is against 2.5.40. It relies on Ingo's smptimers.

Changes:
- always enable head arrays, right now they are SMP only.
- make the per-cpu arrays strictly LIFO.
- move slabs during kmem_cache_free to the end of the partial
	list, that should reduce the internal fragmentation.
- smp_call_function() doesn't break disable_irq() anymore,
	remove the workarounds that were necessary.
- Add head arrays to all caches, remove the  fallback to no-headarray.
- Add a dynamic limit to the lengh of the free list
- Do not flush headarrays if they were recently accessed.
- Do not flush free slabs, if the free slab list was recently
	accessed.
- Remove the GROWN flag: activity matters, not when the cache
	was grown last.
- flush free list and head array in a per-cpu timer
- remove kmem_ from slab internal functions.
- poison all objects, even objects with constructors
- add further debug checks [wrong interrupt flags, might_sleep]

--
	Manfred