On Wed 03-08-22 14:41:20, Feng Tang wrote:
> On Tue, Aug 02, 2022 at 05:02:37PM +0800, Michal Hocko wrote:
> > Please make sure to CC Mike on hugetlb related changes.
>
> OK.
>
> > 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.
>
> Thanks for the check!
>
> So you mean if the prferred nodes don't have enough pages, we should
> also fallback to all like dequeue_huge_page_vma() does?
>
> Or we can user a policy API which return nodemask for MPOL_BIND and
> NULL for all other policies, like allowed_mems_nr() needs.
>
> --- a/include/linux/mempolicy.h
> +++ b/include/linux/mempolicy.h
> @@ -158,6 +158,18 @@ static inline nodemask_t *policy_nodemask_current(gfp_t gfp)
> return policy_nodemask(gfp, mpol);
> }
>
> +#ifdef CONFIG_HUGETLB_FS
> +static inline nodemask_t *strict_policy_nodemask_current(void)
> +{
> + struct mempolicy *mpol = get_task_policy(current);
> +
> + if (mpol->mode == MPOL_BIND)
> + return &mpol->nodes;
> +
> + return NULL;
> +}
> +#endif
Yes something like this, except that I would also move this into hugetlb
proper because this doesn't seem generally useful.
> > 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.
>
> The API needed by allowed_mem_nr() is a little different as it has gfp
> flag and cpuset config to consider.
Why would gfp mask matter?
--
Michal Hocko
SUSE Labs
