Re: [PATCH 3/5] mm: hugetlb: optimize replace_free_hugepage_folios()

[Oscar Salvador <osalvador@suse.de>]

On Mon, Jan 12, 2026 at 11:09:52PM +0800, Kefeng Wang wrote:
> If no free hugepage folios are available, there is no need to perform
> any replacement operations. Additionally, gigantic folios should not
> be replaced under any circumstances. Therefore, we only check for the
> presence of non-gigantic folios.
> 
> To ensure that gigantic folios are not mistakenly replaced, we utilize
> isolate_or_dissolve_huge_folio().
> 
> Lastly, to optimize performance, we skip unnecessary iterations over pfn
> for compound pages and high-order buddy pages to save processing time.
> 
> A simple test on machine with 114G free memory, allocate 120 * 1G
> HugeTLB folios(104 successfully returned),
> 
>   time echo 120 > /sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages
> 
> Before: 0m0.602s
> After:  0m0.431s
> 
> Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
...
>  int replace_free_hugepage_folios(unsigned long start_pfn, unsigned long end_pfn)
>  {
> -	struct folio *folio;
> -	int ret = 0;
> +	unsigned long nr = 0;
> +	struct page *page;
> +	struct hstate *h;
> +	LIST_HEAD(list);
>  
> -	LIST_HEAD(isolate_list);
> +	/* Avoid pfn iterations if no free non-gigantic huge pages */
> +	for_each_hstate(h) {
> +		if (hstate_is_gigantic(h))
> +			continue;
> +
> +		nr += h->free_huge_pages;
> +		if (nr)
> +			break;
> +	}
> +
> +	if (!nr)
> +		return 0;
>  
>  	while (start_pfn < end_pfn) {
> -		folio = pfn_folio(start_pfn);
> +		page = pfn_to_page(start_pfn);
> +		nr = 1;
>  
> -		/* Not to disrupt normal path by vainly holding hugetlb_lock */
> -		if (folio_test_hugetlb(folio) && !folio_ref_count(folio)) {
> -			ret = alloc_and_dissolve_hugetlb_folio(folio, &isolate_list);
> -			if (ret)
> -				break;
> +		if (PageHuge(page) || PageCompound(page)) {
> +			struct folio *folio = page_folio(page);
> +
> +			nr = folio_nr_pages(folio) - folio_page_idx(folio, page);
> +
> +			/* Not to disrupt normal path by vainly holding hugetlb_lock */
> +			if (folio_test_hugetlb(folio) && !folio_ref_count(folio)) {
> +				if (isolate_or_dissolve_huge_folio(folio, &list))

isolate_or_dissolve_huge_folio() can also return -EBUSY, but you supersed that with
ENOMEM in case it fails. Is that alright?
Maybe it is because down the chain we ignore the type of error as long as it is
an error but just brining it up because the log now differs.

 

-- 
Oscar Salvador
SUSE Labs