Re: [RFC][PATCH] netfs, afs, ceph: Use folios

[Matthew Wilcox <willy@infradead.org>]

On Wed, Aug 11, 2021 at 02:07:51PM +0100, David Howells wrote:
>  (*) Can page_endio() be split into two separate functions, one for read
>      and one for write?  If seems a waste of time to conditionally switch
>      between two different branches.

At this point I'm thinking ...

static inline void folio_end_read(struct folio *folio, int err)
{
	if (!err)
		folio_set_uptodate(folio);
	folio_unlock(folio);
}

Clearly the page isn't uptodate at this point, or ->readpage wouldn't've
been called.  So there's no need to clear it.  And PageError is
completely useless.

Part of this exercise was to find gaps in the API.  It looks like there
aren't too many places I missed that AFS hits.

> @@ -78,7 +78,7 @@ int afs_write_begin(struct file *file, struct address_space *mapping,
>  			goto flush_conflicting_write;
>  	}
>  
> -	*_page = page;
> +	*_page = &folio->page;

Can't do anything about this one; the write_begin API needs to be fixed.

> @@ -87,17 +87,17 @@ int afs_write_begin(struct file *file, struct address_space *mapping,
>  	 */
>  flush_conflicting_write:
>  	_debug("flush conflict");
> -	ret = write_one_page(page);
> +	ret = write_one_page(&folio->page);

I have folio_write_one() in my tree:

https://git.infradead.org/users/willy/pagecache.git/commitdiff/82b9f3c7b258de31bf3d3fa4cc587a6d17b5fe40

it's in the "nice to have", rather than "filesystems depend on it" pile,
so I should move it over.

> @@ -174,40 +175,32 @@ static void afs_kill_pages(struct address_space *mapping,
[...]
> +		folio_clear_uptodate(folio);
> +		folio_end_writeback(folio);
> +		folio_lock(folio);
> +		generic_error_remove_page(mapping, &folio->page);
> +		folio_unlock(folio);
> +		folio_put(folio);

This one I'm entirely missing.  It's awkward.  I'll work on it.

> @@ -497,8 +480,8 @@ static void afs_extend_writeback(struct address_space *mapping,
>  			else if (t == psize || new_content)
>  				stop = false;
>  
> -			index += thp_nr_pages(page);
> -			if (!pagevec_add(&pvec, page))
> +			index += folio_nr_pages(folio);
> +			if (!pagevec_add(&pvec, &folio->page))

Pagevecs are also awkward.  I haven't quite figured out how to
transition them to folios.

> @@ -933,29 +919,28 @@ int afs_launder_page(struct page *page)
>  	unsigned int f, t;
>  	int ret = 0;
>  
> -	_enter("{%lx}", page->index);
> +	_enter("{%lx}", folio_index(folio));
>  
> -	priv = page_private(page);
> -	if (clear_page_dirty_for_io(page)) {
> +	priv = (unsigned long)folio_get_private(folio);
> +	if (folio_clear_dirty_for_io(folio)) {
>  		f = 0;
> -		t = thp_size(page);
> -		if (PagePrivate(page)) {
> -			f = afs_page_dirty_from(page, priv);
> -			t = afs_page_dirty_to(page, priv);
> +		t = folio_size(folio);
> +		if (folio_test_private(folio)) {
> +			f = afs_folio_dirty_from(folio, priv);
> +			t = afs_folio_dirty_to(folio, priv);
>  		}
>  
> -		bv[0].bv_page = page;
> +		bv[0].bv_page = &folio->page;
>  		bv[0].bv_offset = f;
>  		bv[0].bv_len = t - f;

I should probably add a wrapper to init a bvec.

>  zero_out:
> -	zero_user_segments(page, 0, offset, offset + len, thp_size(page));
> +	zero_user_segments(&folio->page, 0, offset, offset + len, folio_size(folio));

Yeah, that's ugly.