Re: Priority Inheritance Test (Real-Time Preemption)

[Ingo Molnar]

* Esben Nielsen <simlo@xxxxxxxxxx> wrote:

> I have an idea about what the error(s) is(are): In rt.c policy ==
> SCHED_NORMAL tasks are threaded specially. A task boosted into the
> real-time realm by mutex_setprio() is _still_ SCHED_NORMAL and do not
> gain all the privileges of a real-time task. I suggest that the tests
> on SCHED_NORMAL are replaced by using the rt_task() test which just
> looks at the current priority and thus also would be true on tasks
> temporarely boosted in the real-time realm. [...]

ah, indeed. I thought i fixed all these places but you are right that
there's one more instance remaining, in pick_new_owner():

+               if (w->task->policy == SCHED_NORMAL)
+                       break;

> [...] Another thing: A SCHED_NORMAL task will not be added to the
> pi_waiters list, but it ought to be when it is later boosted into the
> real-time realm. [...]

ok, this is an important one too. Fixing this bug will complicate
pi_setprio() some more, but should be pretty straightforward: the task
that is boosted can be blocked on a single rt-mutex at most.

i'll fix these bugs in the next release. Your systematic testing of PI
is very useful, it already uncovered _tons_ of PI bugs that would be
quite hard to find via normal testing.

> [...] Also, you ignore all tasks being SCHED_NORMAL in the tail of the
> wait list when you try to find the next owner: It could be that one of
> those is boosted.

correct.

> --- linux-2.6.10-rc2-mm2-V0.7.30-9/drivers/char/blocker.c.orig  2004-11-23 20:18:28.000000000 +0100
> +++ linux-2.6.10-rc2-mm2-V0.7.30-9/drivers/char/blocker.c       2004-11-23 20:41:57.742899751 +0100

thx, i'll add this too. Do you have uploaded the matching pi_test
userspace changes too? (url?)

	Ingo
-
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/