Re: WARNING: refcount bug in find_key_to_update

From: David Howells
Date: Fri Oct 18 2019 - 12:38:15 EST

Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:

> The backtrace looks simple enough, though:
> RIP: 0010:refcount_inc_checked+0x2b/0x30 lib/refcount.c:156
> __key_get include/linux/key.h:281 [inline]
> find_key_to_update+0x67/0x80 security/keys/keyring.c:1127
> key_create_or_update+0x4e5/0xb20 security/keys/key.c:905
> __do_sys_add_key security/keys/keyctl.c:132 [inline]
> __se_sys_add_key security/keys/keyctl.c:72 [inline]
> __x64_sys_add_key+0x219/0x3f0 security/keys/keyctl.c:72
> do_syscall_64+0xd0/0x540 arch/x86/entry/common.c:296
> entry_SYSCALL_64_after_hwframe+0x49/0xbe
> which to me implies that there's some locking bug, and somebody
> released the key without holding a lock.
> That code looks a bit confused to me. Releasing a key without holding
> a lock looks permitted, but if that's the case then __key_get() is
> complete garbage. It would need to use 'refcount_inc_not_zero()' and
> failure would require failing the caller.

find_key_to_update() must be called with the keyring-to-be-searched locked, as
stated in the comment on that function.

If a key-to-be-updated can be found in that keyring, then the keyring must be
holding a ref on that key already, so it's refcount must be > 0, so it
shouldn't be necessary to use refcount_inc_not_zero().

There shouldn't be a race with key_link(), key_unlink(), key_move(),
keyring_clear() or keyring_gc() (garbage collection) as all of those take a
write-lock on the keyring.

> But I haven't followed the key locking rules, so who knows. That "put
> without lock" scenario would explain the crash, though.

That shouldn't explain it. When key_put() reduces the refcount to 0, it just
schedules the garbage collector. It doesn't touch the key again directly.

I would guess that something incorrectly put a ref when it shouldn't have. Do
we know which type of key is involved? Looking at the syzkaller reproducer,
it's adding an encrypted key and a user key to the process keyring -
presumably repeating the procedure within the same process, hence how it finds
something to update.