On Thu, 2008-07-24 at 16:32 +0300, Pekka Enberg wrote:

On Thu, Jul 24, 2008 at 10:56:08PM +1000, Nick Piggin wrote:

quoted

quoted

quoted

OTOH, skb allocation uses kmalloc don't they? So you could still use
SLOB ksize for that I guess.

On Thursday 24 July 2008 23:04, Herbert Xu wrote:

quoted

quoted

Yes I was referring to the data portion which is kmalloc'ed.
That is also why I'm interested in ksize because a priori we
don't know exactly how big it's going to be.  However, we do
know that statistically 1500 will dominate.

I'm not interested in ksize for kmem_cache at all.  So in fact
we could have something simpler that's based on kmalloc's rounding
algorithm instead.

On Thu, 2008-07-24 at 23:13 +1000, Nick Piggin wrote:

quoted

Yes you could definitely have a function that returns allocated
bytes for a given kmalloc size. Should be about as fast or faster
than extracting the size from the kaddr...

Yup, makes sense.

On the other hand, I can imagine useful allocator changes where this
would not be a constant of requested size. For instance, imagine we had
a classless bucket allocator, but with a heuristic to try a larger
bucket when it wasn't cheap/possible to allocate a right-sized object
(because of memory pressure, etc.) and larger ones were available.
This sort of thing is a pretty small change for SLAB/SLUB.

-- 
Mathematics is the supreme nostalgia of our time.