[PATCH] ramfs: convert alloc_pages() to folio_alloc() in ramfs_nommu_expand_for_mapping()

[PATCH] ramfs: convert alloc_pages() to folio_alloc() in ramfs_nommu_expand_for_mapping()

[AnishMulay]

Currently, ramfs_nommu_expand_for_mapping() utilizes the deprecated
alloc_pages() API. This patch converts the allocation step to use
the modern folio_alloc() API, removing a legacy caller and helping
pave the way for the eventual removal of alloc_pages().

Because nommu architectures require physically contiguous memory that
often needs to be trimmed to the exact requested size, the allocated
folio is immediately shattered using split_page().

Since split_page() destroys the compound folio metadata, using folio
iteration helpers (like folio_page) becomes unsafe. Therefore, this
patch deliberately drops back to a standard struct page array after
the split. This safely isolates the folio conversion to the allocation
phase while strictly preserving the existing trimming and page cache
insertion behavior.

Signed-off-by: AnishMulay <anishm7030@gmail.com>
---
 fs/ramfs/file-nommu.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/fs/ramfs/file-nommu.c b/fs/ramfs/file-nommu.c
index 77b8ca2757e0d..767de8fc56f0f 100644
--- a/fs/ramfs/file-nommu.c
+++ b/fs/ramfs/file-nommu.c
@@ -62,6 +62,7 @@ const struct inode_operations ramfs_file_inode_operations = {
 int ramfs_nommu_expand_for_mapping(struct inode *inode, size_t newsize)
 {
 	unsigned long npages, xpages, loop;
+	struct folio *folio;
 	struct page *pages;
 	unsigned order;
 	void *data;
@@ -81,10 +82,12 @@ int ramfs_nommu_expand_for_mapping(struct inode *inode, size_t newsize)
 
 	/* allocate enough contiguous pages to be able to satisfy the
 	 * request */
-	pages = alloc_pages(gfp, order);
-	if (!pages)
+	folio = folio_alloc(gfp, order);
+	if (!folio)
 		return -ENOMEM;
 
+	pages = &folio->page;
+
 	/* split the high-order page into an array of single pages */
 	xpages = 1UL << order;
 	npages = (newsize + PAGE_SIZE - 1) >> PAGE_SHIFT;
-- 
2.51.0

Re: [PATCH] ramfs: convert alloc_pages() to folio_alloc() in ramfs_nommu_expand_for_mapping()

[Viacheslav Dubeyko]

CC: linux-fsdevel@vger.kernel.org


On Tue, 2026-02-24 at 15:31 -0500, AnishMulay wrote:
> Currently, ramfs_nommu_expand_for_mapping() utilizes the deprecated
> alloc_pages() API. This patch converts the allocation step to use
> the modern folio_alloc() API, removing a legacy caller and helping
> pave the way for the eventual removal of alloc_pages().
> 
> Because nommu architectures require physically contiguous memory that
> often needs to be trimmed to the exact requested size, the allocated
> folio is immediately shattered using split_page().
> 
> Since split_page() destroys the compound folio metadata, using folio
> iteration helpers (like folio_page) becomes unsafe. Therefore, this
> patch deliberately drops back to a standard struct page array after
> the split. This safely isolates the folio conversion to the allocation
> phase while strictly preserving the existing trimming and page cache
> insertion behavior.
> 
> Signed-off-by: AnishMulay <anishm7030@gmail.com>
> ---
>  fs/ramfs/file-nommu.c | 7 +++++--
>  1 file changed, 5 insertions(+), 2 deletions(-)
> 
> diff --git a/fs/ramfs/file-nommu.c b/fs/ramfs/file-nommu.c
> index 77b8ca2757e0d..767de8fc56f0f 100644
> --- a/fs/ramfs/file-nommu.c
> +++ b/fs/ramfs/file-nommu.c
> @@ -62,6 +62,7 @@ const struct inode_operations ramfs_file_inode_operations = {
>  int ramfs_nommu_expand_for_mapping(struct inode *inode, size_t newsize)

This method uses old page concept everywhere. If you would like to switch this
method for using the folio concept, then I assume this method requires more
rework.

Thanks,
Slava.

>  {
>  	unsigned long npages, xpages, loop;
> +	struct folio *folio;
>  	struct page *pages;
>  	unsigned order;
>  	void *data;
> @@ -81,10 +82,12 @@ int ramfs_nommu_expand_for_mapping(struct inode *inode, size_t newsize)
>  
>  	/* allocate enough contiguous pages to be able to satisfy the
>  	 * request */
> -	pages = alloc_pages(gfp, order);
> -	if (!pages)
> +	folio = folio_alloc(gfp, order);
> +	if (!folio)
>  		return -ENOMEM;
>  
> +	pages = &folio->page;
> +
>  	/* split the high-order page into an array of single pages */
>  	xpages = 1UL << order;
>  	npages = (newsize + PAGE_SIZE - 1) >> PAGE_SHIFT;

Re: [PATCH] ramfs: convert alloc_pages() to folio_alloc() in ramfs_nommu_expand_for_mapping()

[Matthew Wilcox]

On Wed, Feb 25, 2026 at 06:44:28PM +0000, Viacheslav Dubeyko wrote:
> CC: linux-fsdevel@vger.kernel.org

Good plan.

> On Tue, 2026-02-24 at 15:31 -0500, AnishMulay wrote:
> > Currently, ramfs_nommu_expand_for_mapping() utilizes the deprecated
> > alloc_pages() API. This patch converts the allocation step to use

alloc_pages() is not deprecated.  There are no plans for its removal.

> > the modern folio_alloc() API, removing a legacy caller and helping
> > pave the way for the eventual removal of alloc_pages().
> > 
> > Because nommu architectures require physically contiguous memory that
> > often needs to be trimmed to the exact requested size, the allocated
> > folio is immediately shattered using split_page().

Except that split_page() doesn't work on folios.  It says so right there
in the documentation:

 * split_page takes a non-compound higher-order page, and splits it into

Folios are compound pages.  This should have told you that you were
going the wrong thing.

> > Since split_page() destroys the compound folio metadata, using folio
> > iteration helpers (like folio_page) becomes unsafe. Therefore, this
> > patch deliberately drops back to a standard struct page array after
> > the split. This safely isolates the folio conversion to the allocation
> > phase while strictly preserving the existing trimming and page cache
> > insertion behavior.

Please stop trying to help, particularly for code you can't test.