On Mon, May 21, 2018 at 01:16:40PM +0300, Kirill Tkhai wrote:
> >> +static int memcg_expand_one_shrinker_map(struct mem_cgroup *memcg,
> >> + int size, int old_size)
> >
> > Nit: No point in passing old_size here. You can instead use
> > memcg_shrinker_map_size directly.
>
> This is made for the readability. All the actions with global variable
> is made in the same function -- memcg_expand_shrinker_maps(), all
> the actions with local variables are also in the same -- memcg_expand_one_shrinker_map().
> Accessing memcg_shrinker_map_size in memcg_expand_one_shrinker_map()
> looks not intuitive and breaks modularity.
I guess it depends on how you look at it. Anyway, it's nitpicking so I
won't mind if you leave it as is.
> >> +static int memcg_alloc_shrinker_maps(struct mem_cgroup *memcg)
> >> +{
> >> + struct memcg_shrinker_map *map;
> >> + int nid, size, ret = 0;
> >> +
> >> + if (mem_cgroup_is_root(memcg))
> >> + return 0;
> >> +
> >> + mutex_lock(&memcg_shrinker_map_mutex);
> >> + size = memcg_shrinker_map_size;
> >> + for_each_node(nid) {
> >> + map = kvzalloc(sizeof(*map) + size, GFP_KERNEL);
> >> + if (!map) {
> >
> >> + memcg_free_shrinker_maps(memcg);
> >
> > Nit: Please don't call this function under the mutex as it isn't
> > necessary. Set 'ret', break the loop, then check 'ret' after releasing
> > the mutex, and call memcg_free_shrinker_maps() if it's not 0.
>
> No, it must be called under the mutex. See the race with memcg_expand_one_shrinker_map().
> NULL maps are not expanded, and this is the indicator we use to differ memcg, which is
> not completely online. If the allocations in memcg_alloc_shrinker_maps() fail at nid == 1,
> then freeing of nid == 0 can race with expanding.
Ah, I see, you're right.
> >> diff --git a/mm/vmscan.c b/mm/vmscan.c
> >> index 3de12a9bdf85..f09ea20d7270 100644
> >> --- a/mm/vmscan.c
> >> +++ b/mm/vmscan.c
> >> @@ -171,6 +171,7 @@ static DECLARE_RWSEM(shrinker_rwsem);
> >>
> >> #ifdef CONFIG_MEMCG_KMEM
> >> static DEFINE_IDR(shrinker_idr);
> >
> >> +static int memcg_shrinker_nr_max;
> >
> > Nit: Please rename it to shrinker_id_max and make it store max shrinker
> > id, not the max number shrinkers that have ever been allocated. This
> > will make it easier to understand IMO.
> >
> > Also, this variable doesn't belong to this patch as you don't really
> > need it to expaned mem cgroup maps. Let's please move it to patch 3
> > (the one that introduces shrinker_idr).
> >
> >>
> >> static int prealloc_memcg_shrinker(struct shrinker *shrinker)
> >> {
> >> @@ -181,6 +182,15 @@ static int prealloc_memcg_shrinker(struct shrinker *shrinker)
> >> ret = id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL);
> >> if (ret < 0)
> >> goto unlock;
> >
> >> +
> >> + if (id >= memcg_shrinker_nr_max) {
> >> + if (memcg_expand_shrinker_maps(id + 1)) {
> >> + idr_remove(&shrinker_idr, id);
> >> + goto unlock;
> >> + }
> >> + memcg_shrinker_nr_max = id + 1;
> >> + }
> >> +
> >
> > Then we'll have here:
> >
> > if (memcg_expaned_shrinker_maps(id)) {
> > idr_remove(shrinker_idr, id);
> > goto unlock;
> > }
> >
> > and from patch 3:
> >
> > shrinker_id_max = MAX(shrinker_id_max, id);
>
> So, shrinker_id_max contains "the max number shrinkers that have ever been allocated" minus 1.
> The only difference to existing logic is "minus 1", which will be needed to reflect in
> shrink_slab_memcg()->for_each_set_bit()...
>
> To have "minus 1" instead of "not to have minus 1" looks a little subjective.
OK, leave 'nr' then, but please consider my other comments:
- rename memcg_shrinker_nr_max to shrinker_nr_max so that the variable
name is consistent with shrinker_idr
- move shrinker_nr_max to patch 3 as you don't need it for expanding
memcg shrinker maps
- don't use shrinker_nr_max to check whether we need to expand memcg
maps - simply call memcg_expand_shrinker_maps() and let it decide -
this will neatly isolate all the logic related to memcg shrinker map
allocation in memcontrol.c
