[PATCH] exec: Check __FMODE_EXEC instead of in_execve for LSMs

[Jann Horn]

On Wed, Jan 24, 2024 at 10:32 PM Kees Cook <keescook at chromium.org> wrote:
>
> On Wed, Jan 24, 2024 at 12:47:34PM -0800, Linus Torvalds wrote:
> > On Wed, 24 Jan 2024 at 12:15, Kees Cook <keescook at chromium.org> wrote:
> > >
> > > Hmpf, and frustratingly Ubuntu (and Debian) still builds with
> > > CONFIG_USELIB, even though it was reported[2] to them almost 4 years ago.
>
> For completeness, Fedora hasn't had CONFIG_USELIB for a while now.
>
> > Well, we could just remove the __FMODE_EXEC from uselib.
> >
> > It's kind of wrong anyway.
>
> Yeah.
>
> > So I think just removing __FMODE_EXEC would just do the
> > RightThing(tm), and changes nothing for any sane situation.
>
> Agreed about these:
>
> - fs/fcntl.c is just doing a bitfield sanity check.
>
> - nfs_open_permission_mask(), as you say, is only checking for
>   unreadable case.
>
> - fsnotify would also see uselib() as a read, but afaict,
>   that's what it would see for an mmap(), so this should
>   be functionally safe.
>
> This one, though, I need some more time to examine:
>
> - AppArmor, TOMOYO, and LandLock will see uselib() as an
>   open-for-read, so that might still be a problem? As you
>   say, it's more of a mmap() call, but that would mean
>   adding something a call like security_mmap_file() into
>   uselib()...
>
> The issue isn't an insane "support uselib() under AppArmor" case, but
> rather "Can uselib() be used to bypass exec/mmap checks?"
>
> This totally untested patch might give appropriate coverage:
>
> diff --git a/fs/exec.c b/fs/exec.c
> index d179abb78a1c..0c9265312c8d 100644
> --- a/fs/exec.c
> +++ b/fs/exec.c
> @@ -143,6 +143,10 @@ SYSCALL_DEFINE1(uselib, const char __user *, library)
>         if (IS_ERR(file))
>                 goto out;
>
> +       error = security_mmap_file(file, PROT_READ | PROT_EXEC, MAP_FIXED | MAP_SHARED);
> +       if (error)
> +               goto exit;

Call path from here is:

sys_uselib -> load_elf_library -> elf_load -> elf_map -> vm_mmap ->
vm_mmap_pgoff

Call path for normal mmap is:

sys_mmap_pgoff -> ksys_mmap_pgoff -> vm_mmap_pgoff

So I think the call paths converge before any real security checks
happen, and the check you're suggesting should be superfluous. (There
is some weird audit call in ksys_mmap_pgoff() but that's just to
record the FD number, so I guess that doesn't matter.)