On Tue, Nov 13, 2018 at 02:17:51PM +0100, Michal Hocko wrote: >On Thu 08-11-18 09:12:04, Wei Yang wrote: >> for_each_zone_zonelist() iterates the zonelist one by one, which means >> it will iterate on zones on the same node. While get_partial_node() >> checks available slab on node base instead of zone. >> >> This patch skip a node in case get_partial_node() fails to acquire slab >> on that node. > >If this is an optimization then it should be accompanied by some >numbers. Let me try to get some test result. Do you have some suggestion on the test suite? Is kernel build a proper test? > >> @@ -1882,6 +1882,9 @@ static void *get_any_partial(struct kmem_cache *s, gfp_t flags, >> enum zone_type high_zoneidx = gfp_zone(flags); >> void *object; >> unsigned int cpuset_mems_cookie; >> + nodemask_t nmask = node_states[N_MEMORY]; > >This will allocate a large bitmask on the stack and that is no-go for >something that might be called from a potentially deep call stack >already. Also are you sure that the micro-optimization offsets the >copying overhead? > You are right. I didn't pay attention to this. I got one other idea to achieve this effect, like the one in get_page_from_freelist(). In get_page_from_freelist(), we use last_pgdat_dirty_limit to track the last node out of dirty limit. I am willing to borrow this idea in get_any_partial() to skip a node. Well, let me do some tests to see whether this is visible. >-- >Michal Hocko >SUSE Labs -- Wei Yang Help you, Help me
