Re: [RFC][PATCH RT 0/2] futex priority based wakeup

["Sébastien Dugué" <sebastien.dugue@bull.net>]

On Wed, 2006-05-10 at 19:02 +0200, Thomas Gleixner wrote:
> On Wed, 2006-05-10 at 11:01 -0400, Jakub Jelinek wrote:
> > But, there are 2 other futexes used by condvars - internal condvar's lock
> > and __data.__futex.  If the associated mutex uses PI protocol, then
> > I'm afraid the internal condvar lock needs to follow the same protocol
> > (i.e. use FUTEX_*LOCK_PI), otherwise a low priority task calling
> > pthread_cond_* and acquiring the internal lock, then scheduled away
> > indefinitely because of some middle-priority CPU hog could delay
> > some high priority task calling pthread_cond_* on the same condvar.
> 
> Did not think about __data.__lock
> 
> I said "hack_alert" explicitely as this was just a work around, so we
> could use an PI futex for the outer one, which was used to protect other
> things too. The assmebler code corrupted the lock field of the outer
> mutex with 0/1/2.

  Hm, I wonder what would be the effect of having the external mutex
and __data.__lock use a PI futex and __data.__futex use a regular
futex when the waiters on __data.__futex are requeued on the external
mutex during a broadcast.

> 
> > But, there is a problem here - pthread_cond_{signal,broadcast} don't
> > have any associated mutexes, so you often don't know which type
> > of protocol you want to use for the internal condvar lock.

  Just a wild guess here, but in the broadcast or signal path, couldn't
__data.__mutex be looked up to determine what protocol to use for the
__data.__futex?

> > Now, for the __data.__futex lock POSIX seems to be more clear,
> > all it says is that the wake up should happen according to the scheduling
> > policy (but, on the other side for pthread_mutex_unlock it says the same
> > and we still use FIFO there).
> 
> Sebastians patch might solve this.
> 
> Is there a way to (pre)set the policy of the inner lock or is there any
> other solution in sight ?

  As a corolary to the above couldn't the protocol used by the external
mutex dictates what the __data.__futex should use (PI or not)?

Maybe there is a need for something like a futex_requeue_pi() otherwise
we'll end up mixing PI and non-PI futexes or maybe I'm not understanding
the thing.

  Sébastien.