Re: safety of *mutex_unlock() (Was: [BUG] signal: sighand unprotected when accessed by /proc)

[Linus Torvalds]

On Mon, Jun 9, 2014 at 11:29 AM, Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
>>
>> And once again, note that the normal mutex is already unsafe (unless I missed
>> something).
>
> Is it unsafe?
>
> This thread was started because of a bug we triggered in -rt, which
> ended up being a change specific to -rt that modified the way slub
> handled SLAB_DESTROY_BY_RCU. What else was wrong with it?

There's a different issue with freeing of mutexes, which is not a bug,
but "by design". Namely that mutexes aren't truly "atomic". They are
complex data structures, and they have issues that a spinlock does not
have.

When unlocking a mutex, the thread doing the unlocking will still
touch the mutex itself _after_ another thread could already have
successfully acquired the mutex. This is not a problem in any normal
use. since all of this is perfectly coherent in general, but it means
that code sequences like:

  mutex_lock(mem->mutex);
  kill_it = !--mem->refcount;
  mutex_unlock(mem->mutex);
  if (kill_it)
     free(mem);

are fundamentally buggy.

Note that if you think of mutexes as truly indivisible atomic
operations, the above is "obviously correct": the last person who got
the mutex marked it for killing. But the fact is, the *next-to-last*
mutex acquirer may still actively be in the *non-indivisible*
mutex_unlock() when the last person frees it, resulting in a
use-after-free. And yes, we've had this bug, and as far as I know it's
possible that the RT code *introduces* this bug when it changes
spinlocks into mutexes. Because we do exactly the above code sequence
with spinlocks. So just replacing spinlocks with mutexes is a very
subtly buggy thing to do in general.

Another example of this kind of situation is using a mutex as a
completion event: that's simply buggy. Again, it's because mutexes are
complex data structures, and you have a very similar use-after-free
issue. It's why the fundamental data structure for a "struct
completion" is a spinlock, not a mutex.

Again, in *theory*, a completion could be just a mutex that starts out
locked, and then the completer completes it by just unlocking it. The
person who waits for a completion does so by just asking for a lock.
Obvious, simple, and trivial, right? Not so, because of the *exact*
same issue above: the completer (who does an "unlock") may still be
accessing the completion data structure when the completion waiter has
successfully gotten the lock. So the standard thing of the completer
freeing the underlying completion memory (which is often on the stack,
so "freeing" is just he act of going out of scope of the liveness of
the completion data structure) would not work if the completion was a
mutex.

This is subtle, and it is basically unavoidable. If a mutex (or
counting semaphore) has a fast-path - and a mutex/semaphore without a
fast-path is shit - then this issue will exist. Exactly because the
fast-path will depend on just one part of the whole big mutex
structure, and the slow-path will have other pieces to it.

There might be reasonable ways to avoid this issue (having the
fastpath locking field share memory with the slow-path locking, for
example), but it's not how our semaphores and mutexes work, and I
suspect it cannot be the case in general (because it limits you too
badly in how to implement the mutex). As a result, this is all "by
design" as opposed to being a bug.

                Linus
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/