On Tue, 8 Sep 2015, Ingo Molnar wrote:
>
> * Thomas Gleixner <t...@linutronix.de> wrote:
>
> > 3) sched_yield() makes me shudder
> >
> > CPU0 CPU1
> >
> > taskA
> > lock(x->lock)
> >
> > preemption
> > taskC
> > taskB
> > lock(y->lock);
> > x = y->x;
> > if (!try_lock(x->lock)) {
> > unlock(y->lock);
> > boost(taskA);
> > sched_yield(); <- returns immediately
>
> So I'm still struggling with properly parsing the usecase.
>
> If y->x might become invalid the moment we drop y->lock, what makes
> the 'taskA' use (after we've dropped y->lock) safe? Shouldn't we at
> least also have a task_get(taskA)/task_put(taskA) reference count,
> to make sure the boosted task stays around?
Stevens trylock_and_boost() function makes sure that taskA cannot go
away while doing the boosting. It's a bug in my pseudo code, but that
does not make the issue above going away.
> And if we are into getting reference counts, why not solve it at a
> higher level and get a reference count to 'x' to make sure it's safe
> to use? Then we could do:
>
> lock(y->lock);
> retry:
> x = y->x;
> if (!trylock(x->lock)) {
> get_ref(x->count)
> unlock(y->lock);
> lock(x->lock);
> lock(y->lock);
> put_ref(x->count);
> if (y->x != x) { /* Retry if 'x' got dropped meanwhile */
> unlock(x->lock);
> goto retry;
> }
> }
>
> Or so.
In the case of dcache::dentry_kill() we probably do not have to take
refcounts and it might be actually counterproductive to do so. y->x,
i.e. dentry->parent, cannot vanish under us, if I understand the life
time rules correctly.
Aside of that, yes, I was thinking about a similar scheme for
that. I need some more time to grok all the rules there :)
Thanks,
tglx
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
