On Do, 2019-04-25 at 18:05 +0200, Johan Hovold wrote:
> Fix two long-standing bugs which could potentially lead to memory
> corruption or leave the port throttled until it is reopened (on weakly
> ordered systems), respectively, when read-URB completion races with
> unthrottle().
>
> First, the URB must not be marked as free before processing is complete
> to prevent it from being submitted by unthrottle() on another CPU.
>
> CPU 1 CPU 2
> ================ ================
> complete() unthrottle()
> process_urb();
> smp_mb__before_atomic();
> set_bit(i, free); if (test_and_clear_bit(i, free))
> submit_urb();
>
> Second, the URB must be marked as free before checking the throttled
> flag to prevent unthrottle() on another CPU from failing to observe that
> the URB needs to be submitted if complete() sees that the throttled flag
> is set.
>
> CPU 1 CPU 2
> ================ ================
> complete() unthrottle()
> set_bit(i, free); throttled = 0;
> smp_mb__after_atomic(); smp_mb();
> if (throttled) if (test_and_clear_bit(i, free))
> return; submit_urb();
>
> Note that test_and_clear_bit() only implies barriers when the test is
> successful. To handle the case where the URB is still in use an explicit
> barrier needs to be added to unthrottle() for the second race condition.
>
> Also note that the first race was fixed by 36e59e0d70d6 ("cdc-acm: fix
> race between callback and unthrottle") back in 2015, but the bug was
> reintroduced a year later.
>
> Fixes: 1aba579f3cf5 ("cdc-acm: handle read pipe errors")
> Fixes: 088c64f81284 ("USB: cdc-acm: re-write read processing")
> Signed-off-by: Johan Hovold <jo...@kernel.org>
Acked-by: Oliver Neukum <oneu...@suse.com>
