Re: [PATCH] hugetlb: allow to free gigantic pages regardless of the configuration

[Mike Kravetz <mike.kravetz@oracle.com>]

On 1/17/19 10:39 AM, Alexandre Ghiti wrote:
> From: Alexandre Ghiti <alex@ghiti.fr>
> 
> On systems without CMA or (MEMORY_ISOLATION && COMPACTION) activated but
> that support gigantic pages, boottime reserved gigantic pages can not be
> freed at all. This patchs simply enables the possibility to hand back
> those pages to memory allocator.
> 
> This commit then renames gigantic_page_supported and
> ARCH_HAS_GIGANTIC_PAGE to make them more accurate. Indeed, those values
> being false does not mean that the system cannot use gigantic pages: it
> just means that runtime allocation of gigantic pages is not supported,
> one can still allocate boottime gigantic pages if the architecture supports
> it.
> 
> Signed-off-by: Alexandre Ghiti <alex@ghiti.fr>

Thank you for doing this!

Reviewed-by: Mike Kravetz <mike.kravetz@oracle.com>

> --- a/include/linux/gfp.h
> +++ b/include/linux/gfp.h
> @@ -589,8 +589,8 @@ static inline bool pm_suspended_storage(void)
>  /* The below functions must be run on a range from a single zone. */
>  extern int alloc_contig_range(unsigned long start, unsigned long end,
>  			      unsigned migratetype, gfp_t gfp_mask);
> -extern void free_contig_range(unsigned long pfn, unsigned nr_pages);
>  #endif
> +extern void free_contig_range(unsigned long pfn, unsigned int nr_pages);

I think nr_pages should be an unsigned long in cma_release() and here
as well, but that is beyond the scope of this patch.  Most callers of
cma_release pass in a truncated unsigned long.  The truncation is unlikely
to cause any issues, just would be nice if types were consistent.  I have
a patch to do that as part of a contiguous allocation series that I will
get back to someday.

> @@ -2350,9 +2355,10 @@ static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count,
>  			break;
>  	}
>  out:
> -	ret = persistent_huge_pages(h);
> +	h->max_huge_pages = persistent_huge_pages(h);
>  	spin_unlock(&hugetlb_lock);
> -	return ret;
> +
> +	return 0;
>  }
>  
>  #define HSTATE_ATTR_RO(_name) \
> @@ -2404,11 +2410,6 @@ static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
>  	int err;
>  	NODEMASK_ALLOC(nodemask_t, nodes_allowed, GFP_KERNEL | __GFP_NORETRY);
>  
> -	if (hstate_is_gigantic(h) && !gigantic_page_supported()) {
> -		err = -EINVAL;
> -		goto out;
> -	}
> -
>  	if (nid == NUMA_NO_NODE) {
>  		/*
>  		 * global hstate attribute
> @@ -2428,7 +2429,9 @@ static ssize_t __nr_hugepages_store_common(bool obey_mempolicy,
>  	} else
>  		nodes_allowed = &node_states[N_MEMORY];
>  
> -	h->max_huge_pages = set_max_huge_pages(h, count, nodes_allowed);
> +	err = set_max_huge_pages(h, count, nodes_allowed);
> +	if (err)
> +		goto out;
>  
>  	if (nodes_allowed != &node_states[N_MEMORY])
>  		NODEMASK_FREE(nodes_allowed);

Yeah!  Those changes causes max_huge_pages to be modified while holding
hugetlb_lock as it should be.
-- 
Mike Kravetz