On Wed 30-10-19 13:44:55, Johannes Weiner wrote:
> On Tue, Oct 29, 2019 at 04:47:53PM -0700, Shakeel Butt wrote:
> > Since commit 1ba6fc9af35b ("mm: vmscan: do not share cgroup iteration
> > between reclaimers"), the memcg reclaim does not bail out earlier based
> > on sc->nr_reclaimed and will traverse all the nodes. All the reclaimable
> > pages of the memcg on all the nodes will be scanned relative to the
> > reclaim priority. So, there is no need to maintain state regarding which
> > node to start the memcg reclaim from. Also KCSAN complains data races in
> > the code maintaining the state.
> >
> > This patch effectively reverts the commit 889976dbcb12 ("memcg: reclaim
> > memory from nodes in round-robin order") and the commit 453a9bf347f1
> > ("memcg: fix numa scan information update to be triggered by memory
> > event").
> >
> > Signed-off-by: Shakeel Butt <shakeelb@xxxxxxxxxx>
> > Reported-by: <syzbot+13f93c99c06988391efe@xxxxxxxxxxxxxxxxxxxxxxxxx>
>
> Excellent, thanks Shakeel!
> Acked-by: Johannes Weiner <hannes@xxxxxxxxxxx>
>
> Just a request on this bit:
>
> > @@ -3360,16 +3358,9 @@ unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
> > .may_unmap = 1,
> > .may_swap = may_swap,
> > };
> > + struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
> >
> > set_task_reclaim_state(current, &sc.reclaim_state);
> > - /*
> > - * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
> > - * take care of from where we get pages. So the node where we start the
> > - * scan does not need to be the current node.
> > - */
> > - nid = mem_cgroup_select_victim_node(memcg);
> > -
> > - zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
>
> This works, but it *is* somewhat fragile if we decide to add bail-out
> conditions to reclaim again. And some numa nodes receiving slightly
> less pressure than others could be quite tricky to debug.
>
> Can we add a comment here that points out the assumption that the
> zonelist walk is comprehensive, and that all nodes receive equal
> reclaim pressure?
Makes sense
> Also, I think we should use sc.gfp_mask & ~__GFP_THISNODE, so that
> allocations with a physical node preference still do node-agnostic
> reclaim for the purpose of cgroup accounting.
Do not we exclude that by GFP_RECLAIM_MASK already?
--
Michal Hocko
SUSE Labs
