Mike,
On Sun, Aug 01 2021 at 17:14, Mike Galbraith wrote:
> On Sun, 2021-08-01 at 05:36 +0200, Mike Galbraith wrote:
>> On Fri, 2021-07-30 at 22:49 +0200, Thomas Gleixner wrote:
>> > >
>> > > First symptom is KDE/Plasma's task manager going comatose. Notice soon
>> >
>> > KDE/Plasma points at the new fangled rtmutex based ww_mutex from
>> > Peter.
>>
>> Seems not. When booting KVM box with nomodeset, there's exactly one
>> early boot ww_mutex lock/unlock, ancient history at the failure point.
>
> As you've probably already surmised given it isn't the ww_mutex bits,
> it's the wake_q bits. Apply the below, 5.14-rt ceases to fail. Take
> perfectly healthy 5.13-rt, apply those bits, and it instantly begins
> failing as 5.14-rt had been.
now staring at it makes it pretty obvious. When I picked up Peter's
patch I thought about it briefly and then ignored my doubts :(
> /* RT mutex specific wake_q wrappers */
> -static __always_inline void rt_mutex_wake_q_add(struct rt_wake_q_head *wqh,
> +static __always_inline void rt_mutex_wake_q_add(struct rt_mutex_wake_q_head *wqh,
> struct rt_mutex_waiter *w)
> {
> if (IS_ENABLED(CONFIG_PREEMPT_RT) && w->wake_state != TASK_NORMAL) {
> - wake_q_add(&wqh->rt_head, w->task);
> + get_task_struct(w->task);
> + wqh->rtlock_task = w->task;
This is the key. With the original asymmetric version the wake_q_add for
wake_state != TASK_NORMAL is storing the task unconditionally in
wqh->rtlock_task.
With that wake_q_add() we end up with the following situation:
Some code, e.g. futex does:
wake_q_add(..., task)
which links task->wake_q->next is !NULL. Ergo the wake_q_add() in the
rtmutex code bails out. Same the other way round if the rtmutex side
queues first then the second - regular wakeup - will not be queued.
There's two ways to fix that:
1) Go back to my original version
2) Add another wake_q head to task_struct
#2 is overkill IMO simply because the rtlock wait is not subject to
multiple wakeups.
Thanks a lot Mike for tracking this down!
Thanks,
tglx
