Re: [RFC -v3 PATCH 2/3] sched: add yield_to function

[Peter Zijlstra]

On Wed, 2011-01-05 at 11:39 +0900, KOSAKI Motohiro wrote:

> After calling pthread_cond_signal(), T1 which cond_signal caller and T2
> which waked start to GIL grab race. But usually T1 is always win because
> lock variable is in T1's cpu cache. Why kernel and userland have so much
> different result? One of a reason is glibc doesn't have any ticket lock scheme. 

The problem is that making locks strictly fair is that that sucks for
performance, iirc most futex ops are fifo-ordered when they his the
block path, but we do allow for lock-stealing.

Lock-stealing greatly improves performance since it avoids lots of
block/wakeup cycles, but it does make things unfair.

I'm not sure we have a futex option to disable lock-stealing, nor do I
think you really want to, performance suffers really badly.

[This btw is the reason why people reported a performance improvement
when they wrapped all mmap() calls in a pthread_mutex, the
rwsem_down_write() thing doesn't allow for lock-stealing since it needs
to be strict fifo-fair]
--
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/