Re: [PATCH v3 00/14] Generalized Priority Inheritance via Proxy Execution v3

[Peter Zijlstra]

Anyway... tglx and bigeasy just pointed me at a problem with
pi_blocked_on vs sched_submit_work(), where if that function were to
block (specifically use rt-mutex), then we get:

	__rt_mutex_slowlock()
	  task_blocks_on_rt_mutex()
	    current->pi_blocked_on = waiter;
	  rt_mutex_slowlock_block()
	    schedule()
	      sched_submit_work()
	        ...
		  rt_mutex_lock()
		    ...
		      current->pi_blocked_on = waiter; // <--- OOPS

And AFAICT we have the exact same problem with the current DEBUG_MUTEXES
->blocked_on state which in this patch-set is generalized and relied
upon more heavily.