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

[Michal Hocko <mhocko@suse.com>]

Please make sure to CC Mike on hugetlb related changes.

I didn't really get to grasp your proposed solution but it feels goind
sideways. The real issue is that hugetlb uses a dedicated allocation
scheme which is not fully MPOL_PREFERRED_MANY aware AFAICS. I do not
think we should be tricking that by providing some fake nodemasks and
what not.

The good news is that allocation from the pool is MPOL_PREFERRED_MANY
aware because it first tries to allocation from the preffered node mask
and then fall back to the full nodemask (dequeue_huge_page_vma).
If the existing pools cannot really satisfy that allocation then it
tries to allocate a new hugetlb page (alloc_fresh_huge_page) which also
performs 2 stage allocation with the node mask and no node masks. But
both of them might fail.

The bad news is that other allocation functions - including those that
allocate to the pool are not fully MPOL_PREFERRED_MANY aware. E.g.
__nr_hugepages_store_common paths which use the allocating process
policy to fill up the pool so the pool could be under provisioned if
that context is using MPOL_PREFERRED_MANY.

Wrt. allowed_mems_nr (i.e. hugetlb_acct_memory) this is a reservation
code and I have to admit I do not really remember details there. This is
a subtle code and my best guess would be that policy_nodemask_current
should be hugetlb specific and only care about MPOL_BIND.

On Tue 02-08-22 15:39:52, Feng Tang wrote:
> On Tue, Aug 02, 2022 at 02:40:11PM +0800, Muchun Song wrote:
> > On Tue, Aug 02, 2022 at 01:52:05PM +0800, Feng Tang wrote:
> > > On Tue, Aug 02, 2022 at 11:42:52AM +0800, Muchun Song wrote:
> > > > 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.
> > > 
> > > Yep. 
> > > 
> > > I had another thought to add one global all-zero nodemask, for API like
> > > policy_nodemask(), it has 2 types of return value:
> > > * a nodemask with some bits set
> > > * NULL (means all nodes)
> > > 
> > > Here a new type of zero nodemask (a gloabl variable)can be created to
> > > indicate no qualified node.
> > >
> > 
> > I know why you want to introduce a gloable zero nidemask. Since we already
> > have a glable nodemask array, namely node_states, instead of returning NULL
> > for the case of all nodes, how about returing node_states[N_ONLINE] for it?
> > And make it return NULL for the case where no nodes are allowed. Any thought?
> 
> I think return node_states[N_ONLINE] can simplify the code in allowed_mems_nr(),
> the empty zero nodes can simplify further.
> 
> Here is some draft patch (not tested) to show the idea
> 
> Thanks,
> Feng
> 
> ---
>  include/linux/mempolicy.h |  8 ++++++++
>  include/linux/nodemask.h  |  7 +++++++
>  mm/hugetlb.c              |  7 ++++---
>  mm/mempolicy.c            | 17 +++++++++++++++++
>  mm/page_alloc.c           |  3 +++
>  5 files changed, 39 insertions(+), 3 deletions(-)
> 
> diff --git a/include/linux/mempolicy.h b/include/linux/mempolicy.h
> index 668389b4b53d..b5451fef1620 100644
> --- a/include/linux/mempolicy.h
> +++ b/include/linux/mempolicy.h
> @@ -150,6 +150,7 @@ extern bool init_nodemask_of_mempolicy(nodemask_t *mask);
>  extern bool mempolicy_in_oom_domain(struct task_struct *tsk,
>  				const nodemask_t *mask);
>  extern nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy);
> +extern nodemask_t *allowed_policy_nodemask(gfp_t gfp, struct mempolicy *policy);
>  
>  static inline nodemask_t *policy_nodemask_current(gfp_t gfp)
>  {
> @@ -158,6 +159,13 @@ static inline nodemask_t *policy_nodemask_current(gfp_t gfp)
>  	return policy_nodemask(gfp, mpol);
>  }
>  
> +static inline nodemask_t *allowed_policy_nodemask_current(gfp_t gfp)
> +{
> +	struct mempolicy *mpol = get_task_policy(current);
> +
> +	return allowed_policy_nodemask(gfp, mpol);
> +}
> +
>  extern unsigned int mempolicy_slab_node(void);
>  
>  extern enum zone_type policy_zone;
> diff --git a/include/linux/nodemask.h b/include/linux/nodemask.h
> index 0f233b76c9ce..dc5fab38e810 100644
> --- a/include/linux/nodemask.h
> +++ b/include/linux/nodemask.h
> @@ -409,6 +409,13 @@ enum node_states {
>  
>  extern nodemask_t node_states[NR_NODE_STATES];
>  
> +extern nodemask_t zero_nodes;
> +
> +static inline bool is_empty_nodes(nodemask_t *pnodes)
> +{
> +	 return (pnodes == &zero_nodes || __nodes_empty(pnodes, MAX_NUMNODES));
> +}
> +
>  #if MAX_NUMNODES > 1
>  static inline int node_state(int node, enum node_states state)
>  {
> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index a57e1be41401..dc9f4ed32909 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -4340,10 +4340,11 @@ static unsigned int allowed_mems_nr(struct hstate *h)
>  
>  	mpol_allowed = policy_nodemask_current(gfp_mask);
>  
> -	for_each_node_mask(node, cpuset_current_mems_allowed) {
> -		if (!mpol_allowed || node_isset(node, *mpol_allowed))
> +	if (is_empty_nodes(mpol_allowed))
> +		return 0;
> +
> +	for_each_node_mask(node, mpol_allowed)
>  			nr += array[node];
> -	}
>  
>  	return nr;
>  }
> diff --git a/mm/mempolicy.c b/mm/mempolicy.c
> index d39b01fd52fe..3e936b8ca9ea 100644
> --- a/mm/mempolicy.c
> +++ b/mm/mempolicy.c
> @@ -1845,6 +1845,23 @@ nodemask_t *policy_nodemask(gfp_t gfp, struct mempolicy *policy)
>  	return NULL;
>  }
>  
> +/*
> + * Return the allowed nodes mask for a mempolicy and page allocation,
> + * which is a 'stricter' semantic than policy_nodemsk()
> + */
> +nodemask_t *allowed_policy_nodemask(gfp_t gfp, struct mempolicy *policy)
> +{
> +	if (unlikely(policy->mode == MPOL_BIND)) {
> +		if (apply_policy_zone(policy, gfp_zone(gfp)) &&
> +			cpuset_nodemask_valid_mems_allowed(&policy->nodes))
> +			return &policy->nodes;
> +		else
> +			return &zero_nodes;
> +	}
> +
> +	return NULL;
> +}
> +
>  /*
>   * Return the  preferred node id for 'prefer' mempolicy, and return
>   * the given id for all other policies.
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index e008a3df0485..3549ea037588 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -181,6 +181,9 @@ nodemask_t node_states[NR_NODE_STATES] __read_mostly = {
>  };
>  EXPORT_SYMBOL(node_states);
>  
> +nodemask_t zero_nodes = NODE_MASK_NONE;
> +EXPORT_SYMBOL(zero_nodes);
> +
>  atomic_long_t _totalram_pages __read_mostly;
>  EXPORT_SYMBOL(_totalram_pages);
>  unsigned long totalreserve_pages __read_mostly;
> -- 
> 2.27.0
> 

-- 
Michal Hocko
SUSE Labs