Re: PI patch against 2.6.16-rt9

[Ingo Molnar <mingo@elte.hu>]

* Esben Nielsen <simlo@phys.au.dk> wrote:

> > well, another possibility is that the task got blocked again, and we'll
> > continue boosting _the wrong chain_. I.e. we'll add extra priority to
> > task(s) that might not deserve it at all (it doesnt own the lock we are
> > interested in anymore).
> 
> This can't happen. We are always looking at the first waiter on 
> task->pi_waiter task->pi_lock held when doing the boosting. If task 
> has released the lock the entry task->pi_waiter is gone and no 
> boosting will take place!

no, the task got blocked _again_, as part of a _new_ blocking chain, and 
there's a _new_ PI waiter! How does the two-lock preemptible boosting 
algorithm ensure that if we are in the middle of boosting a 
blocking-dependency chain:

   T1 -> T2 -> ... -> TI -> TI+1 -> ... TN-1 -> TN

we are at TI, and we [the task doing the boosting] now get preempted.

What prevents TI from being part of a _totally new_ blocking-chain, 
where the only similarity between the two chains is that TI is in the 
middle of it:

   T1' -> T2' -> ... -> TI -> TI+1' -> ... TM-1 -> TM'

the only match between the two chains is 'TI'. Now the algorithm will
happily walk the wrong boosting chain, and will boost the wrong tasks.

	Ingo