Re: [PATCH] sched/deadline/rtmutex: Fix a PI crash for deadline tasks

[Xunlei Pang <xpang@redhat.com>]

On 2016/04/09 at 02:59, Peter Zijlstra wrote:
> On Fri, Apr 08, 2016 at 02:50:55PM -0400, Steven Rostedt wrote:
>> On Fri, 8 Apr 2016 19:38:35 +0200
>> Peter Zijlstra <peterz@infradead.org> wrote:
>>
>>> On Fri, Apr 08, 2016 at 12:25:10PM -0400, Steven Rostedt wrote:
>>>
>>>> So the preempt_disable() is to allow us to set current back to its
>>>> normal priority first before waking up the other task because we don't
>>>> want two tasks at the same priority?  
>>>> What's the point of swapping deboost and the wake up again?  
>>> In the context of this patch, it ensures the new pi_task pointer points
>>> to something that exists -- this is a rather useful property for a
>>> pointer to have.
>> It's not clear to what would make the new pi_task pointer object no
>> longer exist from this patch. I take it that waking up the wake_q, will
>> cause something to change in the code of rt_mutex_adjust_prio(current).
>> If so, it should probably be stated in a comment, because nothing is
>> obvious here.
> Its pretty obvious that a running task can exit :-)
>
> But also, wake_q holds a task ref.
>
>>> It furthermore guarantees that it points to a blocked task, another
>>> useful property.
>> I would think that the slowfn() would have removed anything to do with
>> what's on the wake_q removed from current.
> It cannot.
>
>> What task on what pointer.
>> I'm only looking at this current patch, not anything to do with the
>> original patch of this thread. That is, just the swap of waking up
>> wake_q and calling rt_mutex_adjust_prio().
> This whole patch was in the context of the previous patch, as should be
> clear from the thread.
>
> In any case, I just realized we do not in fact provide this guarantee
> (of pointing to a blocked task) that needs a bit more work.

Current patch calls rt_mutex_adjust_prio() before wake_up_q() the wakee, at that moment
the wakee has been removed by rt_mutex_slowunlock()->mark_wakeup_next_waiter(),
from current's pi_waiters, rt_mutex_adjust_prio() won't see this wakee, so I think this should
not be problem.

Regards,
Xunlei