On Wed, Aug 03, 2022 at 08:56:44PM +0800, Michal Hocko wrote:
> On Thu 04-08-22 04:43:06, Feng Tang wrote:
> > On Wed, Aug 03, 2022 at 07:28:59PM +0800, Michal Hocko wrote:
> [...]
> > > +struct mempolicy *policy_mbind_nodemask(gfp_t gfp)
> > > +{
> > > +#ifdef CONFIG_MEMPOLICY
> > > + struct mempolicy *mpol = get_task_policy(current);
> > > +
> > > + /*
> > > + * only enforce MBIND which overlaps with cpuset policy (from policy_nodemask)
> > > + * specifically for hugetlb case
> > > + */
> > > + if (mpol->mode == MPOL_BIND &&
> > > + (apply_policy_zone(mpol, gfp_zone(gfp)) &&
> > > + cpuset_nodemask_valid_mems_allowed(&policy->nodes))
> > > + return &mpol->nodes;
> > > +#endif
> > > + return NULL;
> >
> > I saw the logic is not changed, and it confused me that if there is
> > no qualified node, it will still return NULL which effectively equals
> > node_states[N_MEMORY], while I think it should return a all zero
> > nodemasks.
>
> This is a separate thing and I have to admit that the existing code is
> rather non-intuitive or even broken. I guess we do not care all that
> much because MBIND with completely non-overlapping cpusets is just a
> broken configuration. I am not sure this case is interesting or even
> supported.
Fair enough, and moving the policy_mbind_nodemask() into hugetlb.c for
one single caller make it much less severe.
Do we still need the other nodemask API I proposed earlier which has
no parameter of gfp_flag, and used for __nr_hugepages_store_common?
Thanks,
Feng
> --
> Michal Hocko
> SUSE Labs
>
