futex: Handle early deadlock return correctly
authorThomas Gleixner <tglx@linutronix.de>
Tue, 29 Jan 2019 22:15:12 +0000 (23:15 +0100)
committerThomas Gleixner <tglx@linutronix.de>
Fri, 8 Feb 2019 12:00:36 +0000 (13:00 +0100)
commit1a1fb985f2e2b85ec0d3dc2e519ee48389ec2434
treedb20b498bbd047982a926c49bc419096cf268955
parent6f568ebe2afefdc33a6fb06ef20a94f8b96455f1
futex: Handle early deadlock return correctly

commit 56222b212e8e ("futex: Drop hb->lock before enqueueing on the
rtmutex") changed the locking rules in the futex code so that the hash
bucket lock is not longer held while the waiter is enqueued into the
rtmutex wait list. This made the lock and the unlock path symmetric, but
unfortunately the possible early exit from __rt_mutex_proxy_start() due to
a detected deadlock was not updated accordingly. That allows a concurrent
unlocker to observe inconsitent state which triggers the warning in the
unlock path.

futex_lock_pi()                         futex_unlock_pi()
  lock(hb->lock)
  queue(hb_waiter) lock(hb->lock)
  lock(rtmutex->wait_lock)
  unlock(hb->lock)
                                        // acquired hb->lock
                                        hb_waiter = futex_top_waiter()
                                        lock(rtmutex->wait_lock)
  __rt_mutex_proxy_start()
     ---> fail
          remove(rtmutex_waiter);
     ---> returns -EDEADLOCK
  unlock(rtmutex->wait_lock)
                                        // acquired wait_lock
                                        wake_futex_pi()
                                        rt_mutex_next_owner()
  --> returns NULL
                                          --> WARN

  lock(hb->lock)
  unqueue(hb_waiter)

The problem is caused by the remove(rtmutex_waiter) in the failure case of
__rt_mutex_proxy_start() as this lets the unlocker observe a waiter in the
hash bucket but no waiter on the rtmutex, i.e. inconsistent state.

The original commit handles this correctly for the other early return cases
(timeout, signal) by delaying the removal of the rtmutex waiter until the
returning task reacquired the hash bucket lock.

Treat the failure case of __rt_mutex_proxy_start() in the same way and let
the existing cleanup code handle the eventual handover of the rtmutex
gracefully. The regular rt_mutex_proxy_start() gains the rtmutex waiter
removal for the failure case, so that the other callsites are still
operating correctly.

Add proper comments to the code so all these details are fully documented.

Thanks to Peter for helping with the analysis and writing the really
valuable code comments.

Fixes: 56222b212e8e ("futex: Drop hb->lock before enqueueing on the rtmutex")
Reported-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Co-developed-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: linux-s390@vger.kernel.org
Cc: Stefan Liebler <stli@linux.ibm.com>
Cc: Sebastian Sewior <bigeasy@linutronix.de>
Cc: stable@vger.kernel.org
Link: https://lkml.kernel.org/r/alpine.DEB.2.21.1901292311410.1950@nanos.tec.linutronix.de
kernel/futex.c
kernel/locking/rtmutex.c