On Wed, May 14, 2014 at 11:20:51AM -0500, Christoph Lameter wrote: > On Tue, 13 May 2014, Vladimir Davydov wrote: > > > Since the "slow" and the "normal" free's can't coexist at the same time, > > we must assure all conventional free's have finished before switching > > all further free's to the "slow" mode and starting reparenting. To > > achieve that, a percpu refcounter is used. It is taken and held during > > each "normal" free. The refcounter is killed on memcg offline, and the > > cache's pages migration is initiated from the refcounter's release > > function. If we fail to take a ref on kfree, it means all "normal" > > free's have been completed and the cache is being reparented right now, > > so we should free the object using the "slow" mode. > > Argh adding more code to the free path touching more cachelines in the > process. Actually, there is not that much active code added, IMO. In fact, it's only percpu ref get/put for per memcg caches plus a couple of conditionals. The "slow" mode code is meant to be executed very rarely, so we can move it to a separate function under unlikely optimization. I admit that's far not perfect, because kfree is really a hot path, where every byte of code matters, but unfortunately I don't see how we can avoid this in case we want slab re-parenting. Again, I'd like to hear from you if there is any point in moving in this direction, or I should give up and concentrate on some other approach, because you'll never accept it. Thanks. -- 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>