On Wed, 23 Nov 2011, Miao Xie wrote: > This is a good idea. But I worry that oom will happen easily, because we do > direct reclamation and compact by mems_allowed. > Memory compaction actually iterates through each zone regardless of whether it's allowed or not in the current context. Recall that the nodemask passed into __alloc_pages_nodemask() is non-NULL only when there is a mempolicy that restricts the allocations by MPOL_BIND. That nodemask is not protected by get_mems_allowed(), so there's no change in compaction's behavior with my patch. Direct reclaim does, however, require mems_allowed staying constant without the risk of early oom as you mentioned. It has its own get_mems_allowed(), though, so it doesn't have the opportunity to change until returning to the page allocator. It's possible that mems_allowed will be different on the next call to get_pages_from_freelist() but we don't know anything about that context: it's entirely possible that the set of new mems has an abundance of free memory or are completely depleted as well. So there's no strict need for consistency between the set of allowed nodes during reclaim and the subsequent allocation attempt. All we care about is that reclaim has a consistent set of allowed nodes to determine whether it's making progress or not. -- 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/ . Fight unfair telecom internet charges in Canada: sign http://stopthemeter.ca/ Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>