Re: [PATCH v2] fuse: prevent possible NULL pointer dereference in fuse_iomap_writeback_{range,submit}()

["Darrick J. Wong" <djwong@kernel.org>]

On Wed, Sep 03, 2025 at 03:32:40PM -0700, Joanne Koong wrote:
> On Wed, Sep 3, 2025 at 1:48 PM Darrick J. Wong <djwong@kernel.org> wrote:
> >
> > On Wed, Sep 03, 2025 at 09:08:12PM +0100, Luis Henriques wrote:
> > > On Wed, Sep 03 2025, Joanne Koong wrote:
> > >
> > > > On Wed, Sep 3, 2025 at 1:35 AM Luis Henriques <luis@igalia.com> wrote:
> > > >>
> > > >> These two functions make use of the WARN_ON_ONCE() macro to help debugging
> > > >> a NULL wpc->wb_ctx.  However, this doesn't prevent the possibility of NULL
> > > >> pointer dereferences in the code.  This patch adds some extra defensive
> > > >> checks to avoid these NULL pointer accesses.
> > > >>
> > > >> Fixes: ef7e7cbb323f ("fuse: use iomap for writeback")
> > > >> Signed-off-by: Luis Henriques <luis@igalia.com>
> > > >> ---
> > > >> Hi!
> > > >>
> > > >> This v2 results from Joanne's inputs -- I now believe that it is better to
> > > >> keep the WARN_ON_ONCE() macros, but it's still good to try to minimise
> > > >> the undesirable effects of a NULL wpc->wb_ctx.
> > > >>
> > > >> I've also added the 'Fixes:' tag to the commit message.
> > > >>
> > > >>  fs/fuse/file.c | 13 +++++++++----
> > > >>  1 file changed, 9 insertions(+), 4 deletions(-)
> > > >>
> > > >> diff --git a/fs/fuse/file.c b/fs/fuse/file.c
> > > >> index 5525a4520b0f..990c287bc3e3 100644
> > > >> --- a/fs/fuse/file.c
> > > >> +++ b/fs/fuse/file.c
> > > >> @@ -2135,14 +2135,18 @@ static ssize_t fuse_iomap_writeback_range(struct iomap_writepage_ctx *wpc,
> > > >>                                           unsigned len, u64 end_pos)
> > > >>  {
> > > >>         struct fuse_fill_wb_data *data = wpc->wb_ctx;
> > > >> -       struct fuse_writepage_args *wpa = data->wpa;
> > > >> -       struct fuse_args_pages *ap = &wpa->ia.ap;
> > > >> +       struct fuse_writepage_args *wpa;
> > > >> +       struct fuse_args_pages *ap;
> > > >>         struct inode *inode = wpc->inode;
> > > >>         struct fuse_inode *fi = get_fuse_inode(inode);
> > > >>         struct fuse_conn *fc = get_fuse_conn(inode);
> > > >>         loff_t offset = offset_in_folio(folio, pos);
> > > >>
> > > >> -       WARN_ON_ONCE(!data);
> > > >> +       if (WARN_ON_ONCE(!data))
> > > >> +               return -EIO;
> > > >
> > > > imo this WARN_ON_ONCE (and the one below) should be left as is instead
> > > > of embedded in the "if" construct. The data pointer passed in is set
> > > > by fuse and as such, we're able to reasonably guarantee that data is a
> > > > valid pointer. Looking at other examples of WARN_ON in the fuse
> > > > codebase, the places where an "if" construct is used are for cases
> > > > where the assumptions that are made are more delicate (eg folio
> > > > mapping state, in fuse_try_move_folio()) and less clearly obvious. I
> > > > think this WARN_ON_ONCE here and below should be left as is.
> > >
> > > OK, thank you for your feedback, Joanne.  So, if Miklos agrees with that,
> > > I guess we can drop this patch.
> 
> I think having the two lines "wpa = data->wpa;" and "ap = &wpa->ia.ap"
> moved to below the "WARN_ON_ONCE(!data);" would still be useful

<shrug> Compilers are magic, they can rearrange the function unless you
explicitly put in barriers or data dependencies to prevent that. 8-)

> >
> > AFAICT, this function can only be called by other iomap-using functions
> > in file.c, and those other functions always set
> > iomap_writepage_ctx::wb_ctx so I /think/ the assertions aren't necessary
> > at all...
> >
> > > Cheers,
> > > --
> > > Luís
> > >
> > > >
> > > >
> > > > Thanks,
> > > > Joanne
> > > >
> > > >> +
> > > >> +       wpa = data->wpa;
> > > >> +       ap = &wpa->ia.ap;
> > > >>
> > > >>         if (!data->ff) {
> >
> > ...because if someone fails to set wpc->wb_ctx, this line will crash the
> > kernel at least as much as the WARN_ON would.  IOWs, the WARN_ONs aren't
> > necessary but I don't think they hurt much.
> >
> 
> Oh, I see. Actually, this explanation makes a lot of sense. When I was
> looking at the other WARN_ON usages in fuse, I noticed they were also
> used even if it's logically proven that the code path can never be
> triggered. But I guess what you're saying is that WARN_ONs in general
> should be used if it's otherwise somehow undetectable / non-obvious
> that the condition is violated? That makes sense to me, and checks out
> with the other fuse WARN_ON uses.
> 
> I'm fine with just removing the WARN_ON(!data) here and below. I think
> I added some more WARN_ONs in my other fuse iomap patchset, so I'll
> remove those as well when I send out a new version.

<nod>

> > You could introduce a CONFIG_FUSE_DEBUG option and hide some assertions
> > and whatnot behind it, ala CONFIG_FUSE_IOMAP_DEBUG*:
> >
> > https://git.kernel.org/pub/scm/linux/kernel/git/djwong/xfs-linux.git/tree/fs/fuse/iomap_priv.h?h=djwong-wtf&id=170269a48ae83ea7ce1e23ea5ff39995600efff0
> >
> 
> In that case, personally I'd much prefer removing the WARN_ONs here
> than having a new config for it.

<nod> I added it to mine because there are a lot of things that iomap
/can/ get cranky about, so it's useful to have a "BAD_DATA" macro that
screams when the fuse server feeds garbage to the kernel.

--D

> Thanks,
> Joanne
> 
> > --D
> >
> > > >>                 data->ff = fuse_write_file_get(fi);
> > > >> @@ -2182,7 +2186,8 @@ static int fuse_iomap_writeback_submit(struct iomap_writepage_ctx *wpc,
> > > >>  {
> > > >>         struct fuse_fill_wb_data *data = wpc->wb_ctx;
> > > >>
> > > >> -       WARN_ON_ONCE(!data);
> > > >> +       if (WARN_ON_ONCE(!data))
> > > >> +               return error ? error : -EIO;
> > > >>
> > > >>         if (data->wpa) {
> > > >>                 WARN_ON(!data->wpa->ia.ap.num_folios);
> > >
> > >