On Wed, 18 Dec 2013 16:24:15 -0800 Dave Hansen <dave@xxxxxxxx> wrote: > On 12/17/2013 07:17 AM, Christoph Lameter wrote: > > On Mon, 16 Dec 2013, Dave Hansen wrote: > > > >> I'll do some testing and see if I can coax out any delta from the > >> optimization myself. Christoph went to a lot of trouble to put this > >> together, so I assumed that he had a really good reason, although the > >> changelogs don't really mention any. > > > > The cmpxchg on the struct page avoids disabling interrupts etc and > > therefore simplifies the code significantly. > > > >> I honestly can't imagine that a cmpxchg16 is going to be *THAT* much > >> cheaper than a per-page spinlock. The contended case of the cmpxchg is > >> way more expensive than spinlock contention for sure. > > > > Make sure slub does not set __CMPXCHG_DOUBLE in the kmem_cache flags > > and it will fall back to spinlocks if you want to do a comparison. Most > > non x86 arches will use that fallback code. > > > I did four tests. The first workload allocs a bunch of stuff, then > frees it all with both the cmpxchg-enabled 64-byte struct page and the > 48-byte one that is supposed to use a spinlock. I confirmed the 'struct > page' size in both cases by looking at dmesg. > > Essentially, I see no worthwhile benefit from using the double-cmpxchg > over the spinlock. In fact, the increased cache footprint makes it > *substantially* worse when doing a tight loop. > > Unless somebody can find some holes in this, I think we have no choice > but to unset the HAVE_ALIGNED_STRUCT_PAGE config option and revert using > the cmpxchg, at least for now. > So your scary patch series which shrinks struct page while retaining the cmpxchg_double() might reclaim most of this loss? -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>