Re: [PATCH v3 2/2] vmscan,cgroup: apply mems_effective to reclaim

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

On Mon, Apr 21, 2025 at 06:59:20PM -0400, Gregory Price wrote:
> On Mon, Apr 21, 2025 at 10:39:58AM -0700, Shakeel Butt wrote:
> > On Sat, Apr 19, 2025 at 08:14:29PM -0400, Waiman Long wrote:
> > > 
> > > On 4/19/25 2:48 PM, Shakeel Butt wrote:
> > > > On Sat, Apr 19, 2025 at 01:38:24AM -0400, Gregory Price wrote:
> > > > > +bool cpuset_node_allowed(struct cgroup *cgroup, int nid)
> > > > > +{
> > > > > +	struct cgroup_subsys_state *css;
> > > > > +	unsigned long flags;
> > > > > +	struct cpuset *cs;
> > > > > +	bool allowed;
> > > > > +
> > > > > +	css = cgroup_get_e_css(cgroup, &cpuset_cgrp_subsys);
> > > > > +	if (!css)
> > > > > +		return true;
> > > > > +
> > > > > +	cs = container_of(css, struct cpuset, css);
> > > > > +	spin_lock_irqsave(&callback_lock, flags);
> > > > Do we really need callback_lock here? We are not modifying and I am
> > > > wondering if simple rcu read lock is enough here (similar to
> > > > update_nodemasks_hier() where parent's effective_mems is accessed within
> > > > rcu read lock).
> > > 
> > > The callback_lock is required to ensure the stability of the effective_mems
> > > which may be in the process of being changed if not taken.
> > 
> > Stability in what sense? effective_mems will not get freed under us
> > here or is there a chance for corrupted read here? node_isset() and
> > nodes_empty() seems atomic. What's the worst that can happen without
> > callback_lock?
> 
> Fairly sure nodes_empty is not atomic, it's a bitmap search.

For bitmaps smaller than 64 bits, it seems atomic and MAX_NUMNODES seems
smaller than 64 in all the archs.

Anyways I am hoping that we can avoid taking a global lock in reclaim
path which will become a source of contention for memory pressure
situations.