Hillf Danton <dhillf@xxxxxxxxx> writes: > On Wed, Jul 31, 2013 at 2:37 PM, Joonsoo Kim <iamjoonsoo.kim@xxxxxxx> wrote: >> On Wed, Jul 31, 2013 at 02:21:38PM +0800, Hillf Danton wrote: >>> On Wed, Jul 31, 2013 at 12:41 PM, Joonsoo Kim <iamjoonsoo.kim@xxxxxxx> wrote: >>> > On Wed, Jul 31, 2013 at 10:49:24AM +0800, Hillf Danton wrote: >>> >> On Wed, Jul 31, 2013 at 10:27 AM, Joonsoo Kim <iamjoonsoo.kim@xxxxxxx> wrote: >>> >> > On Mon, Jul 29, 2013 at 03:24:46PM +0800, Hillf Danton wrote: >>> >> >> On Mon, Jul 29, 2013 at 1:31 PM, Joonsoo Kim <iamjoonsoo.kim@xxxxxxx> wrote: >>> >> >> > alloc_huge_page_node() use dequeue_huge_page_node() without >>> >> >> > any validation check, so it can steal reserved page unconditionally. >>> >> >> >>> >> >> Well, why is it illegal to use reserved page here? >>> >> > >>> >> > If we use reserved page here, other processes which are promised to use >>> >> > enough hugepages cannot get enough hugepages and can die. This is >>> >> > unexpected result to them. >>> >> > >>> >> But, how do you determine that a huge page is requested by a process >>> >> that is not allowed to use reserved pages? >>> > >>> > Reserved page is just one for each address or file offset. If we need to >>> > move this page, this means that it already use it's own reserved page, this >>> > page is it. So we should not use other reserved page for moving this page. >>> > >>> Hm, how do you determine "this page" is not buddy? >> >> If this page comes from the buddy, it doesn't matter. It imply that >> this mapping cannot use reserved page pool, because we always allocate >> a page from reserved page pool first. >> > A buddy page also implies, if the mapping can use reserved pages, that no > reserved page was available when requested. Now we can try reserved > page again. I didn't quiet get that. My understanding is, the new page we are allocating here for soft offline should not be allocated from the reserve pool. If we do that we may possibly have allocation failure later for a request that had done page reservation. Now to avoid that we make sure we have enough free pages outside reserve pool so that we can dequeue the huge page. If not we use buddy to allocate the hugepage. -aneesh -- 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>