Virtualization - Feb 2023 - [RFC][PATCH] splice: Prevent gifting of multipage folios

Don't let parts of multipage folios be gifted by (vm)splice into a pipe as
the other end may only be expecting single-page gifts (fuse and virtio
console for example).

replace_page_cache_folio(), for example, will do the wrong thing if it
tries to replace a single paged folio with a multipage folio.

Try to avoid this by making add_to_pipe() remove the gift flag on multipage
folios.

Signed-off-by: David Howells <dhowells at redhat.com>
cc: Matthew Wilcox <willy at infradead.org>
cc: Miklos Szeredi <miklos at szeredi.hu>
cc: Amit Shah <amit at kernel.org>
cc: linux-fsdevel at vger.kernel.org
cc: virtualization at lists.linux-foundation.org
cc: linux-mm at kvack.org
---
 fs/splice.c |    2 ++
 1 file changed, 2 insertions(+)

diff --git a/fs/splice.c b/fs/splice.c
index 2e76dbb81a8f..33caa28a86e4 100644
--- a/fs/splice.c
+++ b/fs/splice.c
@@ -240,6 +240,8 @@ ssize_t add_to_pipe(struct pipe_inode_info *pipe, struct
pipe_buffer *buf)
 	} else if (pipe_full(head, tail, pipe->max_usage)) {
 		ret = -EAGAIN;
 	} else {
+		if (folio_nr_pages(page_folio(buf->page)) > 1)
+			buf->flags &= ~PIPE_BUF_FLAG_GIFT;
 		pipe->bufs[head & mask] = *buf;
 		pipe->head = head + 1;
 		return buf->len;

On Mon, Feb 27, 2023 at 02:23:32PM +0000, David Howells
wrote:
>     
> Don't let parts of multipage folios be gifted by (vm)splice into a pipe
as
> the other end may only be expecting single-page gifts (fuse and virtio
> console for example).
> 
> replace_page_cache_folio(), for example, will do the wrong thing if it
> tries to replace a single paged folio with a multipage folio.
> 
> Try to avoid this by making add_to_pipe() remove the gift flag on multipage
> folios.
> 
> Signed-off-by: David Howells <dhowells at redhat.com>
What should the Fixes: here be?  This was already possible with THPs
(both anon and tmpfs backed) long before I introduced folios.
> cc: Matthew Wilcox <willy at infradead.org>
> cc: Miklos Szeredi <miklos at szeredi.hu>
> cc: Amit Shah <amit at kernel.org>
> cc: linux-fsdevel at vger.kernel.org
> cc: virtualization at lists.linux-foundation.org
> cc: linux-mm at kvack.org
> ---
>  fs/splice.c |    2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/fs/splice.c b/fs/splice.c
> index 2e76dbb81a8f..33caa28a86e4 100644
> --- a/fs/splice.c
> +++ b/fs/splice.c
> @@ -240,6 +240,8 @@ ssize_t add_to_pipe(struct pipe_inode_info *pipe,
struct pipe_buffer *buf)
>  	} else if (pipe_full(head, tail, pipe->max_usage)) {
>  		ret = -EAGAIN;
>  	} else {
> +		if (folio_nr_pages(page_folio(buf->page)) > 1)
> +			buf->flags &= ~PIPE_BUF_FLAG_GIFT;
		if (PageCompound(buf->page))
			buf->flags &= ~PIPE_BUF_FLAG_GIFT;

would be simpler and more backportable.

Matthew Wilcox <willy at infradead.org> wrote:
> What should the Fixes: here be?  This was already possible with THPs
> (both anon and tmpfs backed) long before I introduced folios.
Good question.  How do I tell when THPs were introduced?  It looks like they
might predate git, looking at free_compound_page()'s comment.

David