On Tue, Feb 11, 2025 at 11:41 AM Jeff Layton <jlayton@xxxxxxxxxx> wrote:
>
> On Tue, 2025-02-11 at 11:23 -0800, Joanne Koong wrote:
> > On Tue, Feb 11, 2025 at 6:01 AM Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> > >
> > > On Mon, 2025-02-10 at 17:38 -0500, Jeff Layton wrote:
> > > > On Mon, 2025-02-10 at 20:36 +0000, Matthew Wilcox wrote:
> > > > > On Mon, Feb 10, 2025 at 02:12:35PM -0500, Josef Bacik wrote:
> > > > > > From: Josef Bacik <josef@xxxxxxxxxxxxxx>
> > > > > > Date: Mon, 10 Feb 2025 14:06:40 -0500
> > > > > > Subject: [PATCH] fuse: drop extra put of folio when using pipe splice
> > > > > >
> > > > > > In 3eab9d7bc2f4 ("fuse: convert readahead to use folios"), I converted
> > > > > > us to using the new folio readahead code, which drops the reference on
> > > > > > the folio once it is locked, using an inferred reference on the folio.
> > > > > > Previously we held a reference on the folio for the entire duration of
> > > > > > the readpages call.
> > > > > >
> > > > > > This is fine, however I failed to catch the case for splice pipe
> > > > > > responses where we will remove the old folio and splice in the new
> > > > > > folio. Here we assumed that there is a reference held on the folio for
> > > > > > ap->folios, which is no longer the case.
> > > > > >
> > > > > > To fix this, simply drop the extra put to keep us consistent with the
> > > > > > non-splice variation. This will fix the UAF bug that was reported.
> > > > > >
> > > > > > Link: https://lore.kernel.org/linux-fsdevel/2f681f48-00f5-4e09-8431-2b3dbfaa881e@xxxxxxxxx/
> > > > > > Fixes: 3eab9d7bc2f4 ("fuse: convert readahead to use folios")
> > > > > > Signed-off-by: Josef Bacik <josef@xxxxxxxxxxxxxx>
> > > > > > ---
> > > > > > fs/fuse/dev.c | 2 --
> > > > > > 1 file changed, 2 deletions(-)
> > > > > >
> > > > > > diff --git a/fs/fuse/dev.c b/fs/fuse/dev.c
> > > > > > index 5b5f789b37eb..5bd6e2e184c0 100644
> > > > > > --- a/fs/fuse/dev.c
> > > > > > +++ b/fs/fuse/dev.c
> > > > > > @@ -918,8 +918,6 @@ static int fuse_try_move_page(struct fuse_copy_state *cs, struct page **pagep)
> > > > > > }
> > > > > >
> > > > > > folio_unlock(oldfolio);
> > > > > > - /* Drop ref for ap->pages[] array */
> > > > > > - folio_put(oldfolio);
> > > > > > cs->len = 0;
> > > > >
> > > > > But aren't we now leaking a reference to newfolio? ie shouldn't
> > > > > we also:
> > > > >
> > > > > - folio_get(newfolio);
> > > > >
> > > > > a few lines earlier?
> > > > >
> > > >
> > > >
> > > > I think that ref was leaking without Josef's patch, but your proposed
> > > > fix seems correct to me. There is:
> > > >
> > > > - 1 reference stolen from the pipe_buffer
> > > > - 1 reference taken for the pagecache in replace_page_cache_folio()
> > > > - the folio_get(newfolio) just after that
> > > >
> > > > The pagecache ref doesn't count here, and we only need the reference
> > > > that was stolen from the pipe_buffer to replace the one in pagep.
> > >
> > > Actually, no. I'm wrong here. A little after the folio_get(newfolio)
> > > call, we do:
> > >
> > > /*
> > > * Release while we have extra ref on stolen page. Otherwise
> > > * anon_pipe_buf_release() might think the page can be reused.
> > > */
> > > pipe_buf_release(cs->pipe, buf);
> > >
> > > ...so that accounts for the extra reference. I think the newfolio
> > > refcounting is correct as-is.
> >
> > I think we do need to remove the folio_get(newfolio); here or we are
> > leaking the reference.
> >
> > new_folio = page_folio(buf->page) # ref is 1
> > replace_page_cache_folio() # ref is 2
> > folio_get() # ref is 3
> > pipe_buf_release() # ref is 2
> >
> > One ref belongs to the page cache and will get dropped by that, but
> > the other ref is unaccounted for (since the original patch removed
> > "folio_put()" from fuse_readpages_end()).
> >
> > I still think acquiring an explicit reference on the folio before we
> > add it to ap->folio and then dropping it when we're completely done
> > with it in fuse_readpages_end() is the best solution, as that imo
> > makes the refcounting / lifetimes the most explicit / clear. For
> > example, in try_move_pages(), if we get rid of that "folio_get()"
> > call, the page cache is the holder of the remaining reference on it,
> > and we rely on the earlier "folio_clear_uptodate(newfolio);" line in
> > try_move_pages() to guarantee that the newfolio isn't freed out from
> > under us if memory gets tight and it's evicted from the page cache.
> >
> > imo, a patch like this makes the refcounting the most clear:
> >
> > From 923fa98b97cf6dfba3bb486833179c349d566d64 Mon Sep 17 00:00:00 2001
> > From: Joanne Koong <joannelkoong@xxxxxxxxx>
> > Date: Tue, 11 Feb 2025 10:59:40 -0800
> > Subject: [PATCH] fuse: acquire explicit folio refcount for readahead
> >
> > In 3eab9d7bc2f4 ("fuse: convert readahead to use folios"), the logic
> > was converted to using the new folio readahead code, which drops the
> > reference on the folio once it is locked, using an inferred reference
> > on the folio. Previously we held a reference on the folio for the
> > entire duration of the readpages call.
> >
> > This is fine, however for the case for splice pipe responses where we
> > will remove the old folio and splice in the new folio (see
> > fuse_try_move_page()), we assume that there is a reference held on the
> > folio for ap->folios, which is no longer the case.
> >
> > To fix this and make the refcounting explicit, acquire a refcount on the
> > folio before we add it to ap->folios[] and drop it when we are done with
> > the folio in fuse_readpages_end(). This will fix the UAF bug that was
> > reported.
> >
> > Link: https://lore.kernel.org/linux-fsdevel/2f681f48-00f5-4e09-8431-2b3dbfaa881e@xxxxxxxxx/
> > Fixes: 3eab9d7bc2f4 ("fuse: convert readahead to use folios")
> > Signed-off-by: Joanne Koong <joannelkoong@xxxxxxxxx>
> > ---
> > fs/fuse/file.c | 10 +++++++++-
> > 1 file changed, 9 insertions(+), 1 deletion(-)
> >
> > diff --git a/fs/fuse/file.c b/fs/fuse/file.c
> > index 7d92a5479998..6fa535c73d93 100644
> > --- a/fs/fuse/file.c
> > +++ b/fs/fuse/file.c
> > @@ -955,8 +955,10 @@ static void fuse_readpages_end(struct fuse_mount
> > *fm, struct fuse_args *args,
> > fuse_invalidate_atime(inode);
> > }
> >
> > - for (i = 0; i < ap->num_folios; i++)
> > + for (i = 0; i < ap->num_folios; i++) {
> > folio_end_read(ap->folios[i], !err);
> > + folio_put(ap->folios[i]);
> > + }
> > if (ia->ff)
> > fuse_file_put(ia->ff, false);
> >
> > @@ -1049,6 +1051,12 @@ static void fuse_readahead(struct readahead_control *rac)
> >
> > while (ap->num_folios < cur_pages) {
> > folio = readahead_folio(rac);
> > + /*
> > + * Acquire an explicit reference on the folio in case
> > + * it's replaced in the page cache in the splice case
> > + * (see fuse_try_move_page()).
> > + */
> > + folio_get(folio);
> > ap->folios[ap->num_folios] = folio;
> > ap->descs[ap->num_folios].length = folio_size(folio);
> > ap->num_folios++;
>
> That makes sense. My mistake was assuming the pointer in passed in via
> pagep would hold a reference, and that the replacement folio would
> carry one. I like the above better than assuming we have implicit
> reference due to readpages. It's slightly more expensive due to the
> refcounting, but it seems less brittle.
>
> We should couple this with a comment over fuse_try_move_page().
> Something like this maybe?
>
> /*
> * Attempt to steal a page from the splice() pipe and move it into the
> * pagecache. If successful, the pointer in @pagep will be updated. The
> * folio that was originally in @pagep will lose a reference and the new
> * folio returned in @pagep will carry a reference.
> */
Great idea, I'll add this in.
>
> ...
>
> In any case, for this patch:
>
> Reviewed-by: Jeff Layton <jlayton@xxxxxxxxxx>
