Re: [PATCH] mm: mempolicy: fix policy_nodemask() for MPOL_PREFERRED_MANY case

[Muchun Song <songmuchun@bytedance.com>]

On Mon, Aug 01, 2022 at 05:26:23PM +0800, Feng Tang wrote:
> On Mon, Aug 01, 2022 at 05:06:14PM +0800, Michal Hocko wrote:
> > On Mon 01-08-22 16:42:07, Muchun Song wrote:
> > > policy_nodemask() is supposed to be returned a nodemask representing a mempolicy
> > > for filtering nodes for page allocation, which is a hard restriction (see the user
> > > of allowed_mems_nr() in hugetlb.c).  However, MPOL_PREFERRED_MANY is a preferred
> > > mode not a hard restriction.  Now it breaks the user of HugeTLB.  Remove it from
> > > policy_nodemask() to fix it, which will not affect current users of policy_nodemask()
> > > since all of the users already have handled the case of MPOL_PREFERRED_MANY before
> > > calling it.  BTW, it is found by code inspection.
> > 
> > I am not sure this is the right fix. It is quite true that
> > policy_nodemask is a tricky function to use. It pretends to have a
> > higher level logic but all existing users are expected to be policy
> > aware and they special case allocation for each policy. That would mean
> > that hugetlb should do the same.
> 
> Yes, when I worked on the MPOL_PREFERRED_MANY patches, I was also
> confused about policy_nodemask(), as it is never a 'strict' one as
> the old code is:
> 
> 	if (unlikely(mode == MPOL_BIND) &&
> 		apply_policy_zone(policy, gfp_zone(gfp)) &&
> 		cpuset_nodemask_valid_mems_allowed(&policy->nodes))
> 		return &policy->nodes;
> 
> 	return NULL
> 
> Even when the MPOL_BIND's nodes is not allowed by cpuset, it will 
> still return NULL (equals all nodes).
>

Well, I agree policy_nodemask() is really confusing because of the
shortage of comments and the weird logic.

> From the semantics of allowed_mems_nr(), I think it does get changed
> a little by b27abaccf8e8. And to enforce the 'strict' semantic for
> 'allowed', we may need a more strict nodemask API for it.
>

Maybe this is a good idea to fix this, e.g. introducing a new helper
to return the strict allowed nodemask.

> > I haven't checked the actual behavior implications for hugetlb here. Is
> > MPOL_PREFERRED_MANY even supported for hugetlb? Does this change make it
> > work? From a quick look this just ignores MPOL_PREFERRED_MANY
> > completely.
> 
> IIRC, the hugetlb will hornor MPOL_PREFERRED_MANY. And I can double
> check and report back if otherwise.
>
> > > Fixes: b27abaccf8e8 ("mm/mempolicy: add MPOL_PREFERRED_MANY for multiple preferred nodes")
> > > Signed-off-by: Muchun Song <songmuchun@bytedance.com>
> > > ---
> > >  mm/mempolicy.c | 3 ---
> > >  1 file changed, 3 deletions(-)
> > > 
> > > diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> > > index 6c27acb6cd63..4deec7e598c6 100644
> > > --- a/mm/mempolicy.c
> > > +++ b/mm/mempolicy.c
> > > @@ -1845,9 +1845,6 @@ nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
> > >  		cpuset_nodemask_valid_mems_allowed(&policy->nodes))
> > >  		return &policy->nodes;
> > >  
> > > -	if (mode == MPOL_PREFERRED_MANY)
> > > -		return &policy->nodes;
> 
> I think it will make MPOL_PREFERRED_MANY not usable.
>

Sorry, I didn't got what you mean here. Could you explain more details
about why it is not usable?

Thanks.

> Thanks,
> Feng
> 
> > > -
> > >  	return NULL;
> > >  }
> > >  
> > > -- 
> > > 2.11.0
> > 
> > -- 
> > Michal Hocko
> > SUSE Labs
>