Re: Recursion bug in -rt

[Steven Rostedt <rostedt@goodmis.org>]

On Tue, 20 Dec 2005, Esben Nielsen wrote:

> >
>
> The same lock taken twice is just a special case of deadlocking. It would
> be very hard to check for the general case in the futex code without
> "fixing" the rt_mutex. Not that the rt_mutex code is broken - it just
> doesn't handle deadlocks very well as it wasn't supposed to. But as the
> futex indirectly exposes the rt_mutex to userspace it becomes a problem.
>
> The only _hack_ I can see is to force all robust futex calls to go through
> one global lock to prevent the futex deadlocks becomming rt_mutex
> deadlocks which again can turn into spin-lock deadlocks.
>
> I instead argue for altering the premisses for the rt_mutex such
> they can handle deadlocks without turning them into spin-lock deadlocks
> blocking the whole system. Then a futex deadlock will become a rt_mutex
> deadlock which can be handled.
>

For the type of deadlock you are talking about is the following:

P1 -- grabs futex A (no system call)
P2 -- grabs futex B (no system call)

P1 -- tries to grab futex B (system call to block and boost P2)
      But holds no other kernel rt_mutex!
P2 -- tries to grab futex A (system call to block and boost P1)
     spinning deadlock here,

So, before P2 blocks on P1, can there be a circular check t see if this is
a deadlock.  You don't need to worry about other kernel rt_mutexes, you
only need to worry about blocked process.

Is this feasible?

-- Steve

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