On Wed, 16 Jun 2010, Tejun Heo wrote: > Hello, > > On 06/16/2010 06:33 PM, Christoph Lameter wrote: > > On Wed, 16 Jun 2010, Tejun Heo wrote: > >>> Tejun: Is it somehow possible to reliably use the alloc_percpu() on all > >>> platforms during early boot before the slab allocator is up? > >> > >> Hmmm... first chunk allocation is done using bootmem, so if we give it > >> enough to room (for both chunk itself and alloc map) so that it can > >> serve till slab comes up, it should work fine. I think what's > >> important here is making up our minds and decide on how to order them. > >> If the order is well defined, things can be made to work one way or > >> the other. What happened to the get-rid-of-bootmem effort? Wouldn't > >> that also interact with this? > > > > Ok how do we make sure that the first chunk has enough room? > > It's primarily controlled by PERCPU_DYNAMIC_RESERVE. I don't think > there will be any systematic way to do it other than sizing it > sufficiently. Can you calculate the upper bound? The constant has > been used primarily for optimization so how it's used needs to be > audited if we wanna guarantee free space in the first chunk but I > don't think it would be too difficult. The upper bound is SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu). Thats usually 14 * 104 bytes = 1456 bytes. This may increase to more than 8k given the future plans to add queues into kmem_cache_cpu. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>