Re: [rfc] "fair" rw spinlocks

[Nick Piggin]

On Mon, Nov 30, 2009 at 07:22:13AM -0800, Linus Torvalds wrote:
> 
> 
> On Mon, 30 Nov 2009, Nick Piggin wrote:
> > 
> > We do have quite a large number of rwlocks really.
> 
> Dynamically they tend to be unimportant, except for the tasklist_lock. 
> Many of them are in drivers and/or finegrained, or get called only by 
> fairly unusual things (registering filesystems etc).
> 
> > If they are so important as to be rwlocks,
> 
> Stop making total red-herring arguments.

I just mean, if their usage patterns justify their being an rwlock
(as opposed to a plain spinlock). That would mean they can have
significant parallelism in read-side critical sections. And that means
they could have livelock/DoS problems.


> It's not about "so important as to be rwlocks". Quite the reverse. I'm 
> saying that most rwlocks are totally _unimportant_. Being a rwlock does 
> _not_ make anything more important or less important in itself, so your 
> argument is bogus. You have to base importantness on other issues than 
> whether they are rwlocks or not.

I agree there are probably many of them which should just be spinlocks
by nature of their usage patterns. But I would have thought that at
least *some* small percentage of them other than tasklist_lock
would be valid rwlocks.

Anyway, pointless to care about that point I guess... either they exist
or they don't, either way we can't just make all rwlocks fair of course.
So they would have to be tackled one at a time.

 
> As far as I can tell there is _one_ single important rwlock, and that's 
> tasklist_lock. Everything else could probably trivially and individually 
> be turned into a spinlock if fairness matters for them. But tasklist_lock 
> fundamentally depends on the semantics of rwlocks.
> 
> And that one rwlock requires unfair behavior, and is not going to be happy 
> with some more complicated thing (because it is also called from some 
> pretty critical pathways).
> 
> So my argument is purely:
> 
>  - there is absolutely NOBODY who cares about "fair" rwlocks, because no 
>    other user will ever hit its lock enough for it to matter. And if they
>    really do, most of them tend to be fairly simple and localized and 
>    might be turned into spinlocks.
> 
>  - the _one_ major exception to this - somebody who does hit the lock 
>    enough for fairness to matter - is not likely amenable to any kind of 
>    trivial fairness.
> 
> Now, I'd love to come up with some solution to tasklist_lock, but I just 
> don't see it. At least nothing easy that doesn't have tons of downsides 
> (like turning it into a spinlock, using the irq-safe versions, and having 
> irq's potentially disabled for much longer than I think is good).

Well the simple thing I tried earlier was a per-cpu array of nesting
counter there. It's not _too_ expensive, but it does add another cacheline
access and branch there. It seems to work in solving the livelock though.

--
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/