Re: [PATCH] memcg: drain obj stock on cpu hotplug teardown

[Shakeel Butt]

On Tue, Mar 11, 2025 at 11:30:32AM -0400, Johannes Weiner wrote:
> On Mon, Mar 10, 2025 at 04:09:34PM -0700, Shakeel Butt wrote:
> > Currently on cpu hotplug teardown, only memcg stock is drained but we
> > need to drain the obj stock as well otherwise we will miss the stats
> > accumulated on the target cpu as well as the nr_bytes cached. The stats
> > include MEMCG_KMEM, NR_SLAB_RECLAIMABLE_B & NR_SLAB_UNRECLAIMABLE_B. In
> > addition we are leaking reference to struct obj_cgroup object.
> > 
> > Fixes: bf4f059954dc ("mm: memcg/slab: obj_cgroup API")
> > Signed-off-by: Shakeel Butt <shakeel.butt@xxxxxxxxx>
> > Cc: <stable@xxxxxxxxxxxxxxx>
> 
> Wow, that's old. Good catch.
> 
> > ---
> >  mm/memcontrol.c | 9 +++++++++
> >  1 file changed, 9 insertions(+)
> > 
> > diff --git a/mm/memcontrol.c b/mm/memcontrol.c
> > index 4de6acb9b8ec..59dcaf6a3519 100644
> > --- a/mm/memcontrol.c
> > +++ b/mm/memcontrol.c
> > @@ -1921,9 +1921,18 @@ void drain_all_stock(struct mem_cgroup *root_memcg)
> >  static int memcg_hotplug_cpu_dead(unsigned int cpu)
> >  {
> >  	struct memcg_stock_pcp *stock;
> > +	struct obj_cgroup *old;
> > +	unsigned long flags;
> >  
> >  	stock = &per_cpu(memcg_stock, cpu);
> > +
> > +	/* drain_obj_stock requires stock_lock */
> > +	local_lock_irqsave(&memcg_stock.stock_lock, flags);
> > +	old = drain_obj_stock(stock);
> > +	local_unlock_irqrestore(&memcg_stock.stock_lock, flags);
> > +
> >  	drain_stock(stock);
> > +	obj_cgroup_put(old);
> 
> It might be better to call drain_local_stock() directly instead. That
> would prevent a bug of this type to reoccur in the future.

The issue is drain_local_stock() works on the local cpu stock while here
we are working on a remote cpu cpu which is dead (memcg_hotplug_cpu_dead
is in PREPARE section of hotplug teardown which runs after the cpu is
dead).

We can safely call drain_stock() on remote cpu stock here but
drain_obj_stock() is a bit tricky as it can __refill_stock() to local cpu
stock and can call __mod_objcg_mlstate to flush stats. Both of these
requires irq disable for NON-RT kernels and thus I added the local_lock
here.

Anyways I wanted a simple fix for the backports and in parallel I am
working on cleaning up all the stock functions as I plan to add multi
memcg support.

Thanks for taking a look.