mm/filemap.c | 2 ++ 1 file changed, 2 insertions(+)
Add check on filler to prevent NULL pointer dereference condition in
read_cache_folio[1].
[1] https://syzkaller.appspot.com/bug?extid=4089e577072948ac5531
Reported-by: syzbot+4089e577072948ac5531@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=4089e577072948ac5531
Tested-by: syzbot+4089e577072948ac5531@syzkaller.appspotmail.com
Signed-off-by: Gianfranco Trad <gianf.trad@gmail.com>
---
Notes:
changes in v2:
- refactored check on filler.
mm/filemap.c | 2 ++
1 file changed, 2 insertions(+)
diff --git a/mm/filemap.c b/mm/filemap.c
index 4f3753f0a158..88de8029133c 100644
--- a/mm/filemap.c
+++ b/mm/filemap.c
@@ -2360,6 +2360,8 @@ static int filemap_read_folio(struct file *file, filler_t filler,
/* Start the actual read. The read will unlock the page. */
if (unlikely(workingset))
psi_memstall_enter(&pflags);
+ if (!filler)
+ return -EIO;
error = filler(file, folio);
if (unlikely(workingset))
psi_memstall_leave(&pflags);
--
2.43.0On Mon, Sep 30, 2024 at 11:02:26AM +0200, Gianfranco Trad wrote: > @@ -2360,6 +2360,8 @@ static int filemap_read_folio(struct file *file, filler_t filler, > /* Start the actual read. The read will unlock the page. */ > if (unlikely(workingset)) > psi_memstall_enter(&pflags); > + if (!filler) > + return -EIO; This is definitely wrong because you enter memstall, but do not exit it. As Andrew says, the underlying problem is that the filesystem does not implement ->read_folio. Which filesystem is this? > error = filler(file, folio); > if (unlikely(workingset)) > psi_memstall_leave(&pflags); > -- > 2.43.0 >
On 30/09/24 20:14, Matthew Wilcox wrote: > On Mon, Sep 30, 2024 at 11:02:26AM +0200, Gianfranco Trad wrote: >> @@ -2360,6 +2360,8 @@ static int filemap_read_folio(struct file *file, filler_t filler, >> /* Start the actual read. The read will unlock the page. */ >> if (unlikely(workingset)) >> psi_memstall_enter(&pflags); >> + if (!filler) >> + return -EIO; > > This is definitely wrong because you enter memstall, but do not exit it. Got it, thanks. > > As Andrew says, the underlying problem is that the filesystem does not > implement ->read_folio. Which filesystem is this? Reproducer via procfs accesses a bpf map backed by an anonymous inode (anon_inode_fs_type), with mapping->a_ops pointing to anon_aops, hence, read_folio() undefined. > >> error = filler(file, folio); >> if (unlikely(workingset)) >> psi_memstall_leave(&pflags); >> -- >> 2.43.0 >> I suppose the next step would be to contact the proper maintainers(?) If you have any additional suggestions, I'd be more than glad to listen. Thanks to both of you for your time, --Gian
On 04/10/24 14:07, Gianfranco Trad wrote: > On 30/09/24 20:14, Matthew Wilcox wrote: >> On Mon, Sep 30, 2024 at 11:02:26AM +0200, Gianfranco Trad wrote: >>> @@ -2360,6 +2360,8 @@ static int filemap_read_folio(struct file >>> *file, filler_t filler, >>> /* Start the actual read. The read will unlock the page. */ >>> if (unlikely(workingset)) >>> psi_memstall_enter(&pflags); >>> + if (!filler) >>> + return -EIO; >> >> This is definitely wrong because you enter memstall, but do not exit it. > > Got it, thanks. > >> >> As Andrew says, the underlying problem is that the filesystem does not >> implement ->read_folio. Which filesystem is this? > > Reproducer via procfs accesses a bpf map backed by an anonymous > inode (anon_inode_fs_type), with mapping->a_ops pointing to anon_aops, > hence, read_folio() undefined. > >> >>> error = filler(file, folio); >>> if (unlikely(workingset)) >>> psi_memstall_leave(&pflags); >>> -- >>> 2.43.0 >>> > > I suppose the next step would be to contact the proper maintainers(?) > If you have any additional suggestions, I'd be more than glad to listen. > > Thanks to both of you for your time, > > --Gian > Hello, While studying how to implement read_folio in anon_aops for this specific case (bpf map backed by anon_inode_fs_type) I've come up with an intermediate solution that mitigates the null pointer dereference and avoids the memstall issue (compared to my previous patch) immediately, for all filesystems that do not implement read_folio in their address_space_operations. The patch [1] looks like this: diff --git a/mm/filemap.c b/mm/filemap.c index 4f3753f0a158..680d98086c00 100644 --- a/mm/filemap.c +++ b/mm/filemap.c @@ -3775,6 +3775,8 @@ static struct folio *do_read_cache_folio(struct address_space *mapping, struct folio *folio; int err; + if (!filler && (!mapping->a_ops || !mapping->a_ops->read_folio)) + return ERR_PTR(-ENOSYS); if (!filler) filler = mapping->a_ops->read_folio; repeat: Patch was already tested with syzbot on the same reproducer case. Reproducer did not trigger any issue [2]. Let me know if for now this patch looks good enough, therefore I'll send it to you, or if I should work on it more. Thanks for your time, [1] https://syzkaller.appspot.com/text?tag=Patch&x=142e045f980000 [2] https://syzkaller.appspot.com/x/log.txt?x=1551045f980000 -- Gian
On Mon, 30 Sep 2024 11:02:26 +0200 Gianfranco Trad <gianf.trad@gmail.com> wrote: > Add check on filler to prevent NULL pointer dereference condition in > read_cache_folio[1]. > > [1] https://syzkaller.appspot.com/bug?extid=4089e577072948ac5531 Test case https://syzkaller.appspot.com/x/repro.c?x=10a0d880580000 > diff --git a/mm/filemap.c b/mm/filemap.c > index 4f3753f0a158..88de8029133c 100644 > --- a/mm/filemap.c > +++ b/mm/filemap.c > @@ -2360,6 +2360,8 @@ static int filemap_read_folio(struct file *file, filler_t filler, > /* Start the actual read. The read will unlock the page. */ > if (unlikely(workingset)) > psi_memstall_enter(&pflags); > + if (!filler) > + return -EIO; > error = filler(file, folio); > if (unlikely(workingset)) > psi_memstall_leave(&pflags); do_read_cache_folio() already handles a NULL filler (from freader_get_folio()'s read_cache_folio() call). if (!filler) filler = mapping->a_ops->read_folio; so I'm suspecting that an appropriate fix is to teach the underlying address_space_operations (appears to be from /proc/pid/maps) to implement ->read_folio().
© 2016 - 2024 Red Hat, Inc.