Re: [PATCH 5/5] futex: fix wakeup race by setting TASK_INTERRUPTIBLE before queue_me

[Ingo Molnar <mingo@elte.hu>]

* Darren Hart <dvhltc@us.ibm.com> wrote:

> Eric Dumazet wrote:
>> Darren Hart a ??crit :
>>> PI futexes do not use the same plist_node_empty() test for wakeup. It was
>>> possible for the waiter (in futex_wait_requeue_pi()) to set TASK_INTERRUPTIBLE
>>> after the waker assigned the rtmutex to the waiter. The waiter would then note
>>> the plist was not empty and call schedule(). The task would not be found by any
>>> subsequeuent futex wakeups, resulting in a userspace hang. By moving the
>>> setting of TASK_INTERRUPTIBLE to before the call to queue_me(), the race with
>>> the waker is eliminated. Since we no longer call get_user() from within
>>> queue_me(), there is no need to delay the setting of TASK_INTERRUPTIBLE until
>>> after the call to queue_me().
>>>
>>> The FUTEX_LOCK_PI operation is not affected as futex_lock_pi() relies entirely
>>> on the rtmutex code to handle schedule() and wakeup.  The requeue PI code is
>>> affected because the waiter starts as a non-PI waiter and is woken on a PI
>>> futex.
>>>
>>> Remove the crusty old comment about holding spinlocks() across get_user() as we
>>> no longer do that. Correct the locking statement with a description of why the
>>> test is performed.
>>
>> I am very confused by this ChangeLog...
>>
>>> Signed-off-by: Darren Hart <dvhltc@us.ibm.com>
>>> Cc: Thomas Gleixner <tglx@linutronix.de>
>>> Cc: Peter Zijlstra <peterz@infradead.org>
>>> Cc: Steven Rostedt <rostedt@goodmis.org>
>>> Cc: Ingo Molnar <mingo@elte.hu>
>>> CC: Eric Dumazet <eric.dumazet@gmail.com>
>>> CC: Dinakar Guniguntala <dino@in.ibm.com>
>>> CC: John Stultz <johnstul@us.ibm.com>
>>> ---
>>>
>>>  kernel/futex.c |   15 +++------------
>>>  1 files changed, 3 insertions(+), 12 deletions(-)
>>>
>>> diff --git a/kernel/futex.c b/kernel/futex.c
>>> index f92afbe..463af2e 100644
>>> --- a/kernel/futex.c
>>> +++ b/kernel/futex.c
>>> @@ -1656,17 +1656,8 @@ out:
>>>  static void futex_wait_queue_me(struct futex_hash_bucket *hb, struct futex_q *q,
>>>  				struct hrtimer_sleeper *timeout)
>>>  {
>>> -	queue_me(q, hb);
>>> -
>>> -	/*
>>> -	 * There might have been scheduling since the queue_me(), as we
>>> -	 * cannot hold a spinlock across the get_user() in case it
>>> -	 * faults, and we cannot just set TASK_INTERRUPTIBLE state when
>>> -	 * queueing ourselves into the futex hash. This code thus has to
>>> -	 * rely on the futex_wake() code removing us from hash when it
>>> -	 * wakes us up.
>>> -	 */
>>>  	set_current_state(TASK_INTERRUPTIBLE);
>>
>> Hmm, you missed the smp_mb() properties here...
>>
>> Before :
>>      queue_me()
>>      set_mb(current->state, TASK_INTERRUPTIBLE);
>>      if (timeout) {...}
>>      if (likely(!plist_node_empty(&q->list))) {
>> 	...
>>      }
>>
>> After :
>>      set_mb(current->state, TASK_INTERRUPTIBLE);
>>      queue_me();
>>      if (timeout) {...}
>> // no barrier... why ar we still testing q->list
>> // since it has no synchro between queue_me() and test ?
>
> As Ingo said, the barrier is covered by the spin_unlock() in queue_me()  
> according to memory-barriers.txt:
>
>
>  (2) UNLOCK operation implication:
>
>      Memory operations issued before the UNLOCK will be completed before
>      the UNLOCK operation has completed.

btw., it might make sense to add a comment about this - it's not trivial 
at all and the barrier rules here are tricky ...

	Ingo