Re: [PATCH v2] memcg: skip cgroup_file_notify if spinning is not allowed

[Shakeel Butt <shakeel.butt@linux.dev>]

On Mon, Sep 22, 2025 at 04:43:28PM -0700, Andrew Morton wrote:
> On Mon, 22 Sep 2025 16:22:57 -0700 Shakeel Butt <shakeel.butt@linux.dev> wrote:
> 
> > > 
> > > > --- a/mm/memcontrol.c
> > > > +++ b/mm/memcontrol.c
> > > > @@ -2307,12 +2307,13 @@ static int try_charge_memcg(struct mem_cgroup *memcg, gfp_t gfp_mask,
> > > >  	bool drained = false;
> > > >  	bool raised_max_event = false;
> > > >  	unsigned long pflags;
> > > > +	bool allow_spinning = gfpflags_allow_spinning(gfp_mask);
> > > >  
> > > 
> > > Does this affect only the problematic call chain which you have
> > > identified, or might other callers be undesirably affected?
> > 
> > It will only affect the call chain which can not spin due to possibly
> > NMI context and at the moment only bpf programs can cause that.
> 
> "possibly" NMI context? 

NMI is one source which can cause recursive context but bpf programs
attached to specific call chains can also cause this recursion. For
example in [1], a bpf program related to sched_ext was called with
scheduler locks held and then that program made a memcg charged
allocation which can potentially call cgroup_file_notify(). The
notification call chain again tries to grab scheduler locks and
potentially causing deadlocks.

[1] https://lore.kernel.org/all/20250905061919.439648-1-yepeilin@google.com/

> Is it possible that a bpf caller which could
> have taken locks will now skip the notifications?  Or do the gfp_flags
> get propagated all the way through?

The bpf programs which might be called/triggerd in a context where
spinning on a lock might not be safe, will skip the notifications. The
gfp_flags are plugged through. Basically we have kmalloc_nolock()
interfaces coming up which will make sure the correct gfp flags are
passed through.