Re: [BUG] long freezes on thinkpad t60

[Nick Piggin <nickpiggin@yahoo.com.au>]

Linus Torvalds wrote:
> 
> On Tue, 26 Jun 2007, Nick Piggin wrote:
> 
>>Hmm, not that I have a strong opinion one way or the other, but I
>>don't know that they would encourage bad code. They are not going to
>>reduce latency under a locked section, but will improve determinism
>>in the contended case.
> 
> 
> xadd really generally *is* slower than an add. One is often microcoded, 
> the other is not.

Oh. I found xadd to be not hugely slower on my P4, but it was a little
bit.


> But the real problem is that your "unlock" sequence is now about two 
> orders of magnitude slower than it used to be. So it used to be that a 
> spinlocked sequence only had a single synchronization point, now it has 
> two. *That* is really bad, and I guarantee that it makes your spinlocks 
> effectively twice as slow for the non-contended parts.

I don't know why my unlock sequence should be that much slower? Unlocked
mov vs unlocked add? Definitely in dumb micro-benchmark testing it wasn't
twice as slow (IIRC).


> But your xadd thing might be worth looking at, just to see how expensive 
> it is. As an _alternative_ to spinlocks, it's certainly viable.
> 
> (Side note: why make it a word? Word operations are slower on many x86 
> implementations, because they add yet another prefix. You only need a 
> byte)

No real reason I guess. I'll change it.

-- 
SUSE Labs, Novell Inc.