Re: [PATCH 1/6] hugetlb: open-code hugetlb folio lookup index conversion

[Matthew Wilcox]

On Thu, Apr 09, 2026 at 05:41:52PM -0600, Jane Chu wrote:
> This patch removes `filemap_lock_hugetlb_folio()` and open-codes
> the index conversion at each call site, making it explicit when
> hugetlb code is translating a hugepage index into the base-page index
> expected by `filemap_lock_folio()`

I think this is too large a piece to break off in a single patch.
The first thing I did was look at hugetlbfs_read_iter() and wonder
why we're not able to use generic_file_read_iter() here?  It used
to be necessary because we used to index the page cache in units
of hsize, but now we don't, it seems to me that we could use
generic_file_read_iter() instead.

Now, what hugetlbfs_read_iter() does have is support for hwpoison
handling.  I suspect this is something we want in
generic_file_read_iter(), it's just nobody's done it yet.

So perhaps that's patch 1 -- add hwpoison support to
generic_file_read_iter().  Then patch 2 removes hugetlbfs_read_iter() in
favour of using generic_file_read_iter().  

Patch 3 is purely this:
(and you can put my Reviewed-by on it).

> @@ -652,10 +652,10 @@ static void hugetlbfs_zero_partial_page(struct hstate *h,
>  					loff_t start,
>  					loff_t end)
>  {
> -	pgoff_t idx = start >> huge_page_shift(h);
> +	pgoff_t index = start >> PAGE_SHIFT; 
>  	struct folio *folio;
>  
> -	folio = filemap_lock_hugetlb_folio(h, mapping, idx);
> +	folio = filemap_lock_folio(mapping, index);
>  	if (IS_ERR(folio))
>  		return;
>  

Now for patch 4 ...

> diff --git a/mm/hugetlb.c b/mm/hugetlb.c
> index a786034ac95c..38b39eaf46cc 100644
> --- a/mm/hugetlb.c
> +++ b/mm/hugetlb.c
> @@ -5724,7 +5724,7 @@ static vm_fault_t hugetlb_no_page(struct address_space *mapping,
>  	 * before we get page_table_lock.
>  	 */
>  	new_folio = false;
> -	folio = filemap_lock_hugetlb_folio(h, mapping, vmf->pgoff);
> +	folio = filemap_lock_folio(mapping, vmf->pgoff << huge_page_order(h));
>  	if (IS_ERR(folio)) {
>  		size = i_size_read(mapping->host) >> huge_page_shift(h);
>  		if (vmf->pgoff >= size)

This points to a horrible problem.  Everywhere else in the VM has
vmf->pgoff in PAGE_SIZE units, and of course hugetlb works in units of
hpagesize.  So this is an entirely different piece of work where we
convert vmf->pgoff to be in units of PAGE_SIZE.  That'll be fun!

> @@ -6208,7 +6208,7 @@ int hugetlb_mfill_atomic_pte(pte_t *dst_pte,
>  
>  	if (is_continue) {
>  		ret = -EFAULT;
> -		folio = filemap_lock_hugetlb_folio(h, mapping, idx);
> +		folio = filemap_lock_folio(mapping, idx << huge_page_order(h));
>  		if (IS_ERR(folio))
>  			goto out;
>  		folio_in_pagecache = true;

This is a much smaller (more contained) problem.  At least idx is local
to this function, so you can calculate it using linear_page_index()
and modify the whole function.

Finally, you can delete filemap_lock_hugetlb_folio():

> diff --git a/include/linux/hugetlb.h b/include/linux/hugetlb.h
> index 9c098a02a09e..c64c6e5e50f5 100644
> --- a/include/linux/hugetlb.h
> +++ b/include/linux/hugetlb.h
> @@ -829,12 +829,6 @@ static inline unsigned int blocks_per_huge_page(struct hstate *h)
>  	return huge_page_size(h) / 512;
>  }
>  
> -static inline struct folio *filemap_lock_hugetlb_folio(struct hstate *h,
> -				struct address_space *mapping, pgoff_t idx)
> -{
> -	return filemap_lock_folio(mapping, idx << huge_page_order(h));
> -}
> -
>  #include <asm/hugetlb.h>
>  
>  #ifndef is_hugepage_only_range
> @@ -1106,12 +1100,6 @@ static inline struct hugepage_subpool *hugetlb_folio_subpool(struct folio *folio
>  	return NULL;
>  }
>  
> -static inline struct folio *filemap_lock_hugetlb_folio(struct hstate *h,
> -				struct address_space *mapping, pgoff_t idx)
> -{
> -	return NULL;
> -}
> -
>  static inline int isolate_or_dissolve_huge_folio(struct folio *folio,
>  						struct list_head *list)
>  {