Re: [PATCH-mm v3] mm/list_lru: Optimize memcg_reparent_list_lru_node()

[Roman Gushchin <roman.gushchin@linux.dev>]

On Mon, Mar 28, 2022 at 04:46:39PM -0400, Waiman Long wrote:
> On 3/28/22 15:12, Roman Gushchin wrote:
> > On Sun, Mar 27, 2022 at 08:57:15PM -0400, Waiman Long wrote:
> > > On 3/22/22 22:12, Muchun Song wrote:
> > > > On Wed, Mar 23, 2022 at 9:55 AM Waiman Long <longman@redhat.com> wrote:
> > > > > On 3/22/22 21:06, Muchun Song wrote:
> > > > > > On Wed, Mar 9, 2022 at 10:40 PM Waiman Long <longman@redhat.com> wrote:
> > > > > > > Since commit 2c80cd57c743 ("mm/list_lru.c: fix list_lru_count_node()
> > > > > > > to be race free"), we are tracking the total number of lru
> > > > > > > entries in a list_lru_node in its nr_items field.  In the case of
> > > > > > > memcg_reparent_list_lru_node(), there is nothing to be done if nr_items
> > > > > > > is 0.  We don't even need to take the nlru->lock as no new lru entry
> > > > > > > could be added by a racing list_lru_add() to the draining src_idx memcg
> > > > > > > at this point.
> > > > > > Hi Waiman,
> > > > > > 
> > > > > > Sorry for the late reply.  Quick question: what if there is an inflight
> > > > > > list_lru_add()?  How about the following race?
> > > > > > 
> > > > > > CPU0:                               CPU1:
> > > > > > list_lru_add()
> > > > > >        spin_lock(&nlru->lock)
> > > > > >        l = list_lru_from_kmem(memcg)
> > > > > >                                        memcg_reparent_objcgs(memcg)
> > > > > >                                        memcg_reparent_list_lrus(memcg)
> > > > > >                                            memcg_reparent_list_lru()
> > > > > >                                                memcg_reparent_list_lru_node()
> > > > > >                                                    if (!READ_ONCE(nlru->nr_items))
> > > > > >                                                        // Miss reparenting
> > > > > >                                                        return
> > > > > >        // Assume 0->1
> > > > > >        l->nr_items++
> > > > > >        // Assume 0->1
> > > > > >        nlru->nr_items++
> > > > > > 
> > > > > > IIUC, we use nlru->lock to serialise this scenario.
> > > > > I guess this race is theoretically possible but very unlikely since it
> > > > > means a very long pause between list_lru_from_kmem() and the increment
> > > > > of nr_items.
> > > > It is more possible in a VM.
> > > > 
> > > > > How about the following changes to make sure that this race can't happen?
> > > > > 
> > > > > diff --git a/mm/list_lru.c b/mm/list_lru.c
> > > > > index c669d87001a6..c31a0a8ad4e7 100644
> > > > > --- a/mm/list_lru.c
> > > > > +++ b/mm/list_lru.c
> > > > > @@ -395,9 +395,10 @@ static void memcg_reparent_list_lru_node(struct
> > > > > list_lru *lru, int nid,
> > > > >            struct list_lru_one *src, *dst;
> > > > > 
> > > > >            /*
> > > > > -        * If there is no lru entry in this nlru, we can skip it
> > > > > immediately.
> > > > > +        * If there is no lru entry in this nlru and the nlru->lock is free,
> > > > > +        * we can skip it immediately.
> > > > >             */
> > > > > -       if (!READ_ONCE(nlru->nr_items))
> > > > > +       if (!READ_ONCE(nlru->nr_items) && !spin_is_locked(&nlru->lock))
> > > > I think we also should insert a smp_rmb() between those two loads.
> > > Thinking about this some more, I believe that adding spin_is_locked() check
> > > will be enough for x86. However, that will likely not be enough for arches
> > > with a more relaxed memory semantics. So the safest way to avoid this
> > > possible race is to move the check to within the lock critical section,
> > > though that comes with a slightly higher overhead for the 0 nr_items case. I
> > > will send out a patch to correct that. Thanks for bring this possible race
> > > to my attention.
> > Yes, I think it's not enough:
> > CPU0                                       CPU1
> > READ_ONCE(&nlru->nr_items) -> 0
> >                                             spin_lock(&nlru->lock);
> >                                             nlru->nr_items++;
> >                                             spin_unlock(&nlru->lock);
> > && !spin_is_locked(&nlru->lock) -> 0
> I have actually thought of that. I am even thinking about reading nr_items
> again after spin_is_locked(). Still for arches with relaxed memory
> semantics, when will a memory write by one cpu be propagated to another cpu
> can be highly variable. It is very hard to prove that it is completely safe.
> 
> x86 has a more strict memory semantics and it is the only architecture that
> I have enough confidence that doing the check without taking a lock can be
> safe. Perhaps we could use this optimization just for x86 and do it inside
> locks for the rests.

Hm, is this such a big problem in the real life? Can you describe the setup?
I'm somewhat resistant to an idea of having arch-specific optimizations here
without a HUGE reason.

> 
> > Getting back to the original patch, I wonder if instead we can batch reparenting
> > of lrus so we don't have to grab and release nlru->lock for each reparenting lru.
> 
> nlru is actually a sub-structure within a lru. So if there are m lrus and n
> nodes, we will have m*n nlrus. I don't believe there is anymore batching
> that can be done.

I need to think, I'll share any ideas/patches if there will be any.

Thanks!