On Fri, Mar 16, 2007 at 07:25:53AM +1100, Nick Piggin wrote: > I would just avoid the complexity and setup/teardown costs, and just > use a vmalloc'ed global hash for NUMA.
This patch is not the way to go, but neither are vmalloc()'d global hashtables. When you just happen to hash to the wrong node, you're in for quasi-unreproducible poor performance. The size is never right, at which point RCU resizing is required with all its overhead and memory freeing delays and failure to resize (even if only to contract) under pressure. Better would be to use a different data structure admitting locality of reference and adaptively sizing itself, furthermore localized to the appropriate sharing domain. For file-backed futexes, this would be the struct address_space. For anonymous-backed futexes, this would be the COW sharing group, which an anon_vma could almost be used to represent. Using an object to properly represent the COW sharing group (i.e. Hugh's struct anon) would do the trick, and one might as well move the rmap code over to it while we're at it since the anon_vma scanning tricks are all pointless overhead once the COW sharing group is accurately tracked (the scanning around for nearby vmas with ->anon_vma set is not great anyway, though the overhead is hidden in the noise of large teardown and setup operations; inheriting on fork() is much simpler and faster). In such a manner localization is accomplished while no interface extensions are required. -- wli - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
