Re: [PATCH 2/2] mm/mempolicy: stop copying the nodemask in the interleave paths

[Gregory Price <gourry@gourry.net>]

On Wed, Sep 16, 2026 at 05:57:22PM +0200, David Hildenbrand (Arm) wrote:
> On 8/29/26 03:59, Gregory Price wrote:
> > The interleave node selectors copy pol->nodes onto the stack so the mask
> > cannot change while they walk it.  nodemask_t is 128 bytes at
> > MAX_NUMNODES=1024, and two of the three run per folio fault.
> > 
> > The copy only buys consistency between the node count and the walk.
> > Drop the consistency and just bounds check the walk instead.
> 
> I think you should document here that accessing the node bitnmap is safe (is RCU
> responsible for that? I think yes), but it can get updated concurrently.
> 

Actually, not quite.

RCU stabilizes the weights and makes accessing them safe. RCU doesn't
help us with the node bitmasks themselves (we were using a cpuset cookie
to stabilize them to take a copy - but we're killin that here).

What makes it safe is that we detect the torn read condition (check for
an empty nodemask) and otherwise don't care about weight skews in the
weighted variant.

I think it's reasonable to add a couple comments to spell out.

Are you ok if I just spin an additional commit to tack on rather than
spin a new version just for comments?

> > +	 * The target was calculated in a separate loop, and a concurrent
> > +	 * rebind can change the total number of nodes.  Clamp this loop to
> > +	 * a single pass (nnodes) to keep the walk bounded by node count.
> 
> Might want to explicitly comment here that we are looking at a moving target and
> might race with node bitmap modifications.
> 

ack.

> > -	nnodes = read_once_policy_nodemask(pol, &nodemask);
> > +	nnodes = nodes_weight(pol->nodes);
> >  	if (!nnodes)
> >  		return numa_node_id();
> >  	target = ilx % nnodes;
> > -	nid = first_node(nodemask);
> > -	for (i = 0; i < target; i++)
> > -		nid = next_node(nid, nodemask);
> > +	nid = first_node(pol->nodes);
> 
> Similarly, I wonder whether we should spell out the raciness.
>

ack.

~Gregory