PI futexes + lock stealing woes

[Julia Cartwright]

Hey Thomas, Peter-

Gratian and I have been debugging into a nasty and difficult race w/
futexes seemingly the culprit.  The original symptom we were seeing
was a seemingly spurious -EDEADLK from a futex(LOCK_PI) operation.

On further analysis, however, it appears the thread which gets the
spurious -EDEADLK has observed a weird futex state: a prior
futex(WAIT_REQUEUE_PI) operation has returned -ETIMEDOUT, but the uaddr2
futex word owner field indicates that it's the owner.

Here's an attempt to boil down this situation into a pseudo trace; I'm
happy to forward along the full traces as well, if that would be
helpful:

   waiter                                  waker                                            stealer (prio > waiter)

   futex(WAIT_REQUEUE_PI, uaddr, uaddr2,
         timeout=[N ms])
      futex_wait_requeue_pi()
         futex_wait_queue_me()
            freezable_schedule()
            <scheduled out>
                                           futex(LOCK_PI, uaddr2)
                                           futex(CMP_REQUEUE_PI, uaddr,
                                                 uaddr2, 1, 0)
                                              /* requeues waiter to uaddr2 */
                                           futex(UNLOCK_PI, uaddr2)
                                                 wake_futex_pi()
                                                    cmp_futex_value_locked(uaddr, waiter)
                                                    wake_up_q()
           <woken by waker>
           <hrtimer_wakeup() fires,
            clears sleeper->task>
                                                                                           futex(LOCK_PI, uaddr2)
                                                                                              __rt_mutex_start_proxy_lock()
                                                                                                 try_to_take_rt_mutex() /* steals lock */
                                                                                                    rt_mutex_set_owner(lock, stealer)
                                                                                              <preempted>
         <scheduled in>
         rt_mutex_wait_proxy_lock()
            __rt_mutex_slowlock()
               try_to_take_rt_mutex() /* fails, lock held by stealer */
               if (timeout && !timeout->task)
                  return -ETIMEDOUT;
            fixup_owner()
               /* lock wasn't acquired, so,
                  fixup_pi_state_owner skipped */
   return -ETIMEDOUT;

   /* At this point, we've returned -ETIMEDOUT to userspace, but the
    * futex word shows waiter to be the owner, and the pi_mutex has
    * stealer as the owner */

   futex_lock(LOCK_PI, uaddr2)
     -> bails with EDEADLK, futex word says we're owner.

At some later point in execution, the stealer gets scheduled back in and
will do fixup_owner() which fixes up the futex word, but at that point
it's too late: the waiter has already observed the wonky state.

fixup_owner() used to have additional seemingly relevant checks in place
that were removed 73d786bd043eb ("futex: Rework inconsistent
rt_mutex/futex_q state").

The actual kernel we've been testing is 4.9.33-rt23, w/ 153fbd1226fb3
("futex: Fix more put_pi_state() vs. exit_pi_state_list() races")
cherry-picked w/ PREEMPT_RT_FULL.  However, it appears that this issue
may affect v4.15-rc1?

Thoughts on how to move forward?

Nasty.

   Julia