On Wed, 2011-09-14 at 20:48 +0200, Manfred Spraul wrote:

On 09/14/2011 11:57 AM, Peter Zijlstra wrote:

quoted

Subject: ipc/sem: Rework semaphore wakeups
From: Peter Zijlstra<redacted>
Date: Tue Sep 13 15:09:40 CEST 2011

Current sysv sems have a weird ass wakeup scheme that involves keeping
preemption disabled over a potential O(n^2) loop and busy waiting on
that on other CPUs.

Have you checked that the patch improves the latency?
Note that  the busy wait only happens if there is a simultaneous timeout 
of a semtimedop() and a true wakeup.

The code does:

     spin_lock()
     preempt_disable();
     usually_very_simple_but_worstcase_O_2
     spin_unlock()
     usually_very_simple_but_worstcase_O_1
     preempt_enable();

with your change, it becomes:

     spin_lock()
     usually_very_simple_but_worstcase_O_2
     usually_very_simple_but_worstcase_O_1
     spin_unlock()

The complex ops remain unchanged, they are still under a lock.

preemptible lock (aka pi-mutex) on -rt, so no weird latencies.

What about removing the preempt_disable?
It's only there to cover a rare race on uniprocessor preempt systems.
(a task is woken up simultaneously due to timeout of semtimedop() and a 
true wakeup)

Then fix the that race - something like the attached patch [obviously 
buggy - see the fixme]

sched_yield() is always a bug, as is it here. Its an life-lock if the
woken task is of higher priority than the waking task. A higher prio
FIFO task calling sched_yield() in a loop is just that, a loop, starving
the lower prio waker.

If you've got enough medium prio tasks around to occupy all other cpus,
you're got indefinite priority inversion, so even on smp its a problem.

But yeah its not the prettiest of solutions but it works.. see that
other patch with the wake-list stuff for something that ought to work
for both rt and mainline (except of course it doesn't actually work).