Henry reported that rt_mutex_adjust_prio_check() has an ordering
problem and puts the lie to the comment in [7]. Sharing the sort key
between lock->waiters and owner->pi_waiters *does* create problems,
since unlike what the comment claims, holding [L] is insufficient.
Notably, consider:
A
/ \
M1 M2
| |
B C
That is, task A owns both M1 and M2, B and C block on them. In this
case a concurrent chain walk (B & C) will modify their resp. sort keys
in [7] while holding M1->wait_lock and M2->wait_lock. So holding [L]
is meaningless, they're different Ls.
This then gives rise to a race condition between [7] and [11], where
the requeue of pi_waiters will observe an inconsistent tree order.
B C
(holds M1->wait_lock, (holds M2->wait_lock,
holds B->pi_lock) holds A->pi_lock)
[7]
waiter_update_prio();
...
[8]
raw_spin_unlock(B->pi_lock);
...
[10]
raw_spin_lock(A->pi_lock);
[11]
rt_mutex_enqueue_pi();
// observes inconsistent A->pi_waiters
// tree order
Fixing this means either extending the range of the owner lock from
[10-13] to [6-13], with the immediate problem that this means [6-8]
hold both blocked and owner locks, or duplicating the sort key.
Since the locking in chain walk is horrible enough without having to
consider pi_lock nesting rules, duplicate the sort key instead.
By giving each tree their own sort key, the above race becomes
harmless, if C sees B at the old location, then B will correct things
(if they need correcting) when it walks up the chain and reaches A.
Fixes: fb00aca474 ("rtmutex: Turn the plist into an rb-tree")
Reported-by: Henry Wu <triangletrap12@gmail.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Tested-by: Henry Wu <triangletrap12@gmail.com>
Link: https://lkml.kernel.org/r/20230707161052.GF2883469%40hirez.programming.kicks-ass.net