On Wed, 14 Dec 2011 11:46:58 +0100 Michal Hocko <mhocko@xxxxxxx> wrote: > On Tue 13-12-11 10:43:16, Ying Han wrote: > > On Tue, Dec 13, 2011 at 8:21 AM, Michal Hocko <mhocko@xxxxxxx> wrote: > > > On Mon 12-12-11 18:16:27, Ying Han wrote: > > >> In __mem_cgroup_try_charge() function, the parameter "oom" is passed from the > > >> caller indicating whether or not the charge should enter memcg oom kill. In > > >> fact, we should be able to eliminate that by using the existing gfp_mask and > > >> __GFP_NORETRY flag. > > >> > > >> This patch removed the "oom" parameter, and add the __GFP_NORETRY flag into > > >> gfp_mask for those doesn't want to enter memcg oom. There is no functional > > >> change for those setting false to "oom" like mem_cgroup_move_parent(), but > > >> __GFP_NORETRY now is checked for those even setting true to "oom". > > >> > > >> The __GFP_NORETRY is used in page allocator to bypass retry and oom kill. I > > >> believe there is a reason for callers to use that flag, and in memcg charge > > >> we need to respect it as well. > > > > > > What is the reason for this change? > > > To be honest it makes the oom condition more obscure. __GFP_NORETRY > > > documentation doesn't say anything about OOM and one would have to know > > > details about allocator internals to follow this. > > > So I am not saying the patch is bad but I would need some strong reason > > > to like it ;) > > > > Thank you for looking into this :) > > > > This patch was made as part of the effort solving the livelock issue. > > Then it becomes a separate question by itself. > > > > I don't quite understand the mismatch on gfp_mask = __GFP_NORETRY && > > oom_check == true. > > __GFP_NORETRY is a global thingy (because page allocator is global) > while oom_check is internal memcg and it says that we do not want to go > into oom because we cannot charge, consider THP for example. We do not > want to OOM because we would go over hard limit and we rather want to > fallback into a single page allocation. > The first reason that 'oom' was added as argument was to handle 'decreasing limit'. Thanks, -Kame -- 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>