On Tue, Jul 16, 2019 at 5:12 PM Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx> wrote: > > > > On 7/16/19 4:36 PM, Shakeel Butt wrote: > > Adding related people. > > > > The thread starts at: > > http://lkml.kernel.org/r/1562795006.8510.19.camel@xxxxxx > > > > On Mon, Jul 15, 2019 at 8:01 PM Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx> wrote: > >> > >> > >> On 7/15/19 6:36 PM, Qian Cai wrote: > >>>> On Jul 15, 2019, at 8:22 PM, Yang Shi <yang.shi@xxxxxxxxxxxxxxxxx> wrote: > >>>> > >>>> > >>>> > >>>> On 7/15/19 2:23 PM, Qian Cai wrote: > >>>>> On Fri, 2019-07-12 at 12:12 -0700, Yang Shi wrote: > >>>>>>> Another possible lead is that without reverting the those commits below, > >>>>>>> kdump > >>>>>>> kernel would always also crash in shrink_slab_memcg() at this line, > >>>>>>> > >>>>>>> map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map, true); > >>>>>> This looks a little bit weird. It seems nodeinfo[nid] is NULL? I didn't > >>>>>> think of where nodeinfo was freed but memcg was still online. Maybe a > >>>>>> check is needed: > >>>>> Actually, "memcg" is NULL. > >>>> It sounds weird. shrink_slab() is called in mem_cgroup_iter which does pin the memcg. So, the memcg should not go away. > >>> Well, the commit “mm: shrinker: make shrinker not depend on memcg kmem” changed this line in shrink_slab_memcg(), > >>> > >>> - if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)) > >>> + if (!mem_cgroup_online(memcg)) > >>> return 0; > >>> > >>> Since the kdump kernel has the parameter “cgroup_disable=memory”, shrink_slab_memcg() will no longer be able to handle NULL memcg from mem_cgroup_iter() as, > >>> > >>> if (mem_cgroup_disabled()) > >>> return NULL; > >> Aha, yes. memcg_kmem_enabled() implicitly checks !mem_cgroup_disabled(). > >> Thanks for figuring this out. I think we need add mem_cgroup_dsiabled() > >> check before calling shrink_slab_memcg() as below: > >> > >> diff --git a/mm/vmscan.c b/mm/vmscan.c > >> index a0301ed..2f03c61 100644 > >> --- a/mm/vmscan.c > >> +++ b/mm/vmscan.c > >> @@ -701,7 +701,7 @@ static unsigned long shrink_slab(gfp_t gfp_mask, int > >> nid, > >> unsigned long ret, freed = 0; > >> struct shrinker *shrinker; > >> > >> - if (!mem_cgroup_is_root(memcg)) > >> + if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg)) > >> return shrink_slab_memcg(gfp_mask, nid, memcg, priority); > >> > >> if (!down_read_trylock(&shrinker_rwsem)) > >> > > We were seeing unneeded oom-kills on kernels with > > "cgroup_disabled=memory" and Yang's patch series basically expose the > > bug to crash. I think the commit aeed1d325d42 ("mm/vmscan.c: > > generalize shrink_slab() calls in shrink_node()") missed the case for > > "cgroup_disabled=memory". However I am surprised that root_mem_cgroup > > is allocated even for "cgroup_disabled=memory" and it seems like > > css_alloc() is called even before checking if the corresponding > > controller is disabled. > > I'm surprised too. A quick test with drgn shows root memcg is definitely > allocated: > > >>> prog['root_mem_cgroup'] > *(struct mem_cgroup *)0xffff8902cf058000 = { > [snip] > > But, isn't this a bug? It can be treated as a bug as this is not expected but we can discuss and take care of it later. I think we need your patch urgently as memory reclaim and /proc/sys/vm/drop_caches is broken for "cgroup_disabled=memory" kernel. So, please send your patch asap. thanks, Shakeel
