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

[Jakub Jelinek]

On Wed, May 10, 2006 at 04:32:14PM +0200, Thomas Gleixner wrote:
> On Wed, 2006-05-10 at 15:03 +0200, Sébastien Dugué wrote:
> >   Maybe the pthread_cond_*() function should be made to use the
> > PI-futexes support in glibc.
> 
> <hack_alert>
> 
> There is a simple way to do so. Just remove the assembler version of
> pthread_cond_xx and use the generic c code implementation in glibc. That
> allows you to use a PI mutex for the outer mutex.
> 
> </hack_alert>

I don't see how would that help.  Both asm and sysdeps/pthread/*.c
versions call __pthread_mutex_unlock_usercnt/__pthread_mutex_cond_lock,
which will DTRT for the mutex passed as second argument to
pthread_mutex_*wait (assuming you have Ulrich's/mine PI nptl patch).
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.
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.
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).

	Jakub
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/