Re: Possible race in rt_mutex_adjust_prio_chain

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Jul 06, 2023 at 02:08:20PM +0800, Henry Wu wrote:
> Hi,
> 
> I found that it's not safe to change waiter->prio after waiter
> dequeued from mutex's waiter rbtree because it's still on owner's
> pi_waiters rbtree. From my analysis, waiters on pi_waiters rbtree
> should be protected by pi_lock of task which have pi_waiters and
> corresponding rt_mutex's wait_lock, but pi_waiters is shared by many
> locks owned by this task, so actually we serialize changes on
> pi_waiters only by pi_lock.
> 
> `rt_mutex_adjust_prio_chain' changes key of nodes of pi_waiters rbtree
> without pi_lock and pi_waiters rbtree's invariant is violated. Maybe
> we are enqueuing waiter on other cpu and pi_waiters rbtree will be
> corrupted.

Are you talking about [7];

Where we do waiter_update_prio() while only
holding [L] rtmutex->wait_lock.

VS

rt_mutex_adjust_prio() / task_top_pi_waiter() that accesses ->pi_waiters
while holding [P] task->pi_lock.

?

I'll go stare at that in more detail -- but I wanted to verify that's
what you're talking about.

> I attach a source file which can trigger this violation. I tested it
> on Ubuntu 20.04 LTS with 5.4 kernel.

Well, that's a horribly old kernel :-( Please double check on v6.4 and
consult that code for the discussion above -- I'm really not too
interested in debugging something ancient.