propagate_priority KASSERT

[Stephan Uphoff]

On Tue, 2004-11-23 at 22:21, Mark Gooderum wrote:
> We're currently tripping accross the following KASSERT in 
> subr_turnstile.c (in 5.3.0 RELEASE):
> 
>         /*
>          * XXX: The owner of a turnstile can be stale if it is the
>          * first thread to grab a slock of a sx lock.  In that case
>          * it is possible for us to be at SSLEEP or some other
>          * weird state.  We should probably just return if the state
>          * isn't SRUN or SLOCK.
>          */
>         KASSERT(!TD_IS_SLEEPING(td),
>             ("sleeping thread %p (pid %d) owns non-sleepable lock %p",
>              td, td->td_proc->p_pid, ts))
> 
> This is in propagate_priority().
> 
> However, propagate_priority() is only called in one place, by 
> turnstile_wait(), and turnstile_wait() in turn is only called by 
> _mtx_lock_sleep() which in turn is only used for sleep mutexes.
> 
> TD_IS_SLEEPING() really means the SLEEPING inhibitor is set which in 
> turn seems to mean that thread is in fact on a sleepq - which is used 
> for various wait channels and timeouts.
> 
> So I'm just trying to understand why it's a 100% assertion for any 
> holder of the turnstile for a sleep mutex sleeping.

Priority inheritance is needed to avoid blocking interrupt threads for a
long time. A thread sleeping while holding a mutex would break priority
inheritance. Another reason is that this would add a hidden dependency
to the wakeup thread - this can easily lead to deadlocks.

Can you find out where the owner of the mutex is sleeping?

	Stephan