On Wed, Mar 13, 2019 at 7:35 AM Jan Kara <jack@xxxxxxx> wrote: > On Tue 12-03-19 23:26:22, Kees Cook wrote: > > On Mon, Mar 11, 2019 at 1:42 PM syzbot > > <syzbot+2c49971e251e36216d1f@xxxxxxxxxxxxxxxxxxxxxxxxx> wrote: > > > C reproducer: https://syzkaller.appspot.com/x/repro.c?x=17ee410b200000 > > > [...] > > > ------------[ cut here ]------------ > > > Bad or missing usercopy whitelist? Kernel memory exposure attempt detected > > > from SLAB object 'fanotify_event' (offset 40, size 8)! > > > [...] > > > copy_to_user include/linux/uaccess.h:151 [inline] > > > copy_fid_to_user fs/notify/fanotify/fanotify_user.c:236 [inline] > > > copy_event_to_user fs/notify/fanotify/fanotify_user.c:294 [inline] > > > > Looks like this is the fh/ext_fh union in struct fanotify_fid, field > > "fid" in struct fanotify_event. Given that "fid" is itself in a union > > against a struct path, I think instead of a whitelist using > > KMEM_CACHE_USERCOPY(), this should just use a bounce buffer to avoid > > leaving a whitelist open for path or ext_fh exposure. > > Do you mean to protect it from a situation when some other code (i.e. not > copy_fid_to_user()) would be tricked into copying ext_fh containing slab > pointer to userspace? Yes. That's the design around the usercopy hardening. The "whitelisting" is either via code (with a bounce buffer, so only the specific "expected" code path can copy it), with a kmem_create_usercopy() range marking (generally best for areas that are not unions or when bounce buffers would be too big/slow), or with implicit whitelisting (via a constant copy size that cannot change at run-time, like: copy_to_user(dst, src, 6)). In this case, since there are multiple unions in place and FANOTIFY_INLINE_FH_LEN is small, it seemed best to go with a bounce buffer. -- Kees Cook
