Re: [PATCH v3 4/4] futex: Avoid taking hb lock if nothing to wakeup

[Linus Torvalds]

On Thu, Dec 19, 2013 at 3:53 PM, Linus Torvalds
<torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
>
> So how could we miss this? Explain to me what the separate counter
> does that isn't done by the spinlock head counter.

Hmm. Trying to answer this myself by looking at the code. And I just
realized that I described things wrong ("spin_contended()" rather than
"spin_is_locked()"). So that was a bug in my description.

One difference is that by avoiding the counter, the "do we have
waiters" is two values rather than one ("is spin locked" + "node list
head is empty"). So there are possible memory ordering issues wrt
reading those two fields, but I don't see it mattering: since you'd
need to read both "not locked" _and_ "list empty", any re-ordering
that shows both of those cases should be able to show the "waiters ==
0" case in the explicit separate counter model.

I don't know why I care, though. For some reason that extra counter
just makes me go "I'm convinced it's not needed", even though I can't
really explain why I should care about it.

        Linus
--
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/