Re: [PATCH] mm/mempolicy: prevent the risk of division by 0

[Gregory Price <gourry@gourry.net>]

On Mon, Sep 08, 2025 at 02:12:29PM -0400, Gregory Price wrote:
> On Sun, Sep 07, 2025 at 09:08:29AM -0700, Chelsy Ratnawat wrote:
> > If no bits are set in the policy's node mask, then nodes will be 0.
> > This patch adds a check if nodes == 0 before dividing.
> > 
> > Signed-off-by: Chelsy Ratnawat <chelsyratnawat2001@gmail.com>
> > ---
> >  mm/mempolicy.c | 2 ++
> >  1 file changed, 2 insertions(+)
> > 
> > diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> > index eb83cff7db8c..faacc604fc16 100644
> > --- a/mm/mempolicy.c
> > +++ b/mm/mempolicy.c
> > @@ -2530,6 +2530,8 @@ static unsigned long alloc_pages_bulk_interleave(gfp_t gfp,
> >  	unsigned long total_allocated = 0;
> >  
> >  	nodes = nodes_weight(pol->nodes);
> > +	if (nodes == 0)
> > +		return 0;
> 
> 3 second look at this code would tell you the only way this can happen
> is a mempolicy where MPOL_INTERLEAVE is set and the nodemask is empty
>

after some more thought... there is actually a race condition here that
is non-obvious

in some places we take a cpusets cookie to prevent pol->nodes from ever
showing up empty while iterating over the policy nodes.

Excerpt from: alloc_pages_bulk_weighted_interleave

```
        /* read the nodes onto the stack, retry if done during rebind */
        do {
                cpuset_mems_cookie = read_mems_allowed_begin();
                nnodes = read_once_policy_nodemask(pol, &nodes);
        } while (read_mems_allowed_retry(cpuset_mems_cookie));

        /* if the nodemask has become invalid, we cannot do anything */
        if (!nnodes)
                return 0;
```

So this actually can happen - otherwise I wouldn't have bothered with
the stack read in this code chunk anyway.

So now i've argued with and against myself, i think this patch
as-written is not sufficient - and in fact there's a very subtle bug in
the normal interleave that isn't present in the weighted interleave
because we take a copy of the nodes.

So mea culpa for the snap read and response.  But could you please
take a look at the weighted interleave code and model the fix off
of that?

~Gregory