On Mon, Sep 15, 2014 at 10:10:18AM +0900, Joonsoo Kim wrote: > On Fri, Sep 05, 2014 at 10:14:16AM -0400, Theodore Ts'o wrote: > > On Fri, Sep 05, 2014 at 04:32:48PM +0900, Joonsoo Kim wrote: > > > I also test another approach, such as allocate freepage in CMA > > > reserved region as late as possible, which is also similar to your > > > suggestion and this doesn't works well. When reclaim is started, > > > too many pages reclaim at once, because lru list has successive pages > > > in CMA region and these doesn't help kswapd's reclaim. kswapd stop > > > reclaiming when freepage count is recovered. But, CMA pages isn't > > > counted for freepage for kswapd because they can't be usable for > > > unmovable, reclaimable allocation. So kswap reclaim too many pages > > > at once unnecessarilly. > > > > Have you considered putting the pages in a CMA region in a separate > > zone? After all, that's what we originally did with brain-damaged > > hardware that could only DMA into the low 16M of memory. We just > > reserved a separate zone for that? That way, we could do > > zone-directed reclaim and free pages in that zone, if that was what > > was actually needed. > > Sorry for long delay. It was long holidays. > > No, I haven't consider it. It sounds good idea to place the pages in a > CMA region into a separate zone. Perhaps we can remove one of > migratetype, MIGRATE_CMA, with this way and it would be a good long-term > architecture for CMA. IIRC, Mel suggested two options, ZONE_MOVABLE zone and MIGRATE_ISOLATE. Absolutely, movable zone option is better solution if we consider interacting with reclaim but one problem was CMA had a specific requirement for memory in the middle of an existing zone. And his concern comes true. Look at https://lkml.org/lkml/2014/5/28/64. It starts to add more stuff in allocator's fast path to overcome the problem. :( Let's rethink. We already have a logic to handle overlapping nodes/zones in compaction.c so isn't it possible to have discrete address ranges in a movable zone? If so, movable zone can include specific ranges horrible devices want and it could make allocation/reclaim logic simple than now and add overheads to slow path(ie, linear pfn scanning logic of zone like compaction). > > I don't know exact history and reason why CMA is implemented in current > form. Ccing some experts in this area. > > Thanks. > -- > To unsubscribe from this list: send the line "unsubscribe linux-kernel" in > the body of a message to majordomo@xxxxxxxxxxxxxxx > More majordomo info at http://vger.kernel.org/majordomo-info.html > Please read the FAQ at http://www.tux.org/lkml/ -- Kind regards, Minchan Kim -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
