On Mon, Oct 17, 2016 at 05:21:54PM +0800, Xishi Qiu wrote: > On 2016/10/13 16:08, js1304@xxxxxxxxx wrote: > > > From: Joonsoo Kim <iamjoonsoo.kim@xxxxxxx> > > > > Currently, freeing page can stay longer in the buddy list if next higher > > order page is in the buddy list in order to help coalescence. However, > > it doesn't work for the simplest sequential free case. For example, think > > about the situation that 8 consecutive pages are freed in sequential > > order. > > > > page 0: attached at the head of order 0 list > > page 1: merged with page 0, attached at the head of order 1 list > > page 2: attached at the tail of order 0 list > > page 3: merged with page 2 and then merged with page 0, attached at > > the head of order 2 list > > page 4: attached at the head of order 0 list > > page 5: merged with page 4, attached at the tail of order 1 list > > page 6: attached at the tail of order 0 list > > page 7: merged with page 6 and then merged with page 4. Lastly, merged > > with page 0 and we get order 3 freepage. > > > > With excluding page 0 case, there are three cases that freeing page is > > attached at the head of buddy list in this example and if just one > > corresponding ordered allocation request comes at that moment, this page > > in being a high order page will be allocated and we would fail to make > > order-3 freepage. > > > > Allocation usually happens in sequential order and free also does. So, it > > would be important to detect such a situation and to give some chance > > to be coalesced. > > > > I think that simple and effective heuristic about this case is just > > attaching freeing page at the tail of the buddy list unconditionally. > > If freeing isn't merged during one rotation, it would be actual > > fragmentation and we don't need to care about it for coalescence. > > > > Hi Joonsoo, > > I find another two places to reduce fragmentation. > > 1) > __rmqueue_fallback > steal_suitable_fallback > move_freepages_block > move_freepages > list_move > If we steal some free pages, we will add these page at the head of start_migratetype list, > this will cause more fixed migratetype, because this pages will be allocated more easily. > So how about use list_move_tail instead of list_move? Yeah... I don't think deeply but, at a glance, it would be helpful. > > 2) > __rmqueue_fallback > expand > list_add > How about use list_add_tail instead of list_add? If add the tail, then the rest of pages > will be hard to be allocated and we can merge them again as soon as the page freed. I guess that it has no effect. When we do __rmqueue_fallback() and expand(), we don't have any freepage on this or more order. So, list_add or list_add_tail will show the same result. 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>