Re: RT and Cascade interrupts

[john cooper <john.cooper@timesys.com>]

Oleg Nesterov wrote:
> 
> CPU_0						CPU_1
> 
> rpc_release_task(task)
> 						__run_timers:
> 							timer->base = NULL;
> 	rpc_delete_timer()
> 		if (timer->base)
> 			// not taken
> 			del_timer_sync();
> 
> 	mempool_free(task);
> 							timer->function(task):
> 								// task already freed/reused
> 								__rpc_default_timer(task);

Ah ok, I was misreading the above.  Now I see your point.
The race if hit loses the synchronization provided by
del_timer_sync() in my case.  Thanks for being persistent.

The other possibility to fix the original problem within
the scope of the RPC code was to replace the bit state of
RPC_TASK_HAS_TIMER with a count so we don't obscure the
fact the timer was requeued during the preemption window.
This happens as rpc_run_timer() does an unconditional
clear_bit(RPC_TASK_HAS_TIMER,..) before returning.

Another would be to check timer->base/timer_pending()
in rpc_run_timer() and to omit doing a clear_bit()
if the timer is found to have been requeued.

The former has simpler locking requirements but requires
an rpc_task data structure modification.  The latter does
not and is a bit more intuitive but needs additional
locking.  Both appear to provide a solution though I
haven't yet attempted either.

Anyone with more ownership of this code care to comment
on a preferred fix?

-john


-- 
john.cooper@timesys.com