Re: [PATCH v3] mm/mempolicy: fix a race between offset_il_node and mpol_rebind_task

[Muchun Song]

On Wed, Aug 18, 2021 at 11:07 PM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>
> On Wed, Aug 18, 2021 at 10:02:46PM +0800, Muchun Song wrote:
> > On Tue, Aug 17, 2021 at 9:43 AM Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
> > > > > +   unsigned int target, nnodes;
> > > > >     int i;
> > > > >     int nid;
> > > > > +   /*
> > > > > +    * The barrier will stabilize the nodemask in a register or on
> > > > > +    * the stack so that it will stop changing under the code.
> > > > > +    *
> > > > > +    * Between first_node() and next_node(), pol->nodes could be changed
> > > > > +    * by other threads. So we put pol->nodes in a local stack.
> > > > > +    */
> > > > > +   barrier();
> > >
> > > I think this could be an smp_rmb()?
> >
> > Hi Matthew,
> >
> > I have a question. Why is barrier() not enough?
>
> I think barrier() may be more than is necessary.  We don't need a
> barrier on non-SMP systems (or do we?)  And we only need to order reads,
> not writes.

Here barrier() is just a compiler barrier, which is cheaper than
smp_rmb() which usually equals to memory barrier instruction
plus barrier(). So I think barrier() , which will stabilize the
nodemask in a register or on the stack, is more appropriate here.

Thanks.