Alan Cox wrote:
On Sad, 2005-09-03 at 11:40 +1000, Nick Piggin wrote:
We'll see how things go. I'm fairly sure that for my usage it will
be a win even if it is costly. It is replacing an atomic_inc_return,
and a read_lock/read_unlock pair.
Make sure you bench both AMD and Intel - I'd expect it to be a big loss
on AMD because the AMD stuff will perform atomic locked operations very
efficiently if they are already exclusive on this CPU or a prefetch_w()
on them was done 200+ clocks before.
I will try to get numbers for both.
I would be surprised if it was a big loss... but I'm assuming
a locked cmpxchg isn't outlandishly expensive. Basically:
read_lock_irqsave(cacheline1);
atomic_inc_return(cacheline2);
read_unlock_irqrestore(cacheline1);
Turns into
atomic_cmpxchg();
I'll do some microbenchmarks and get back to you. I'm quite
interested now ;) What sort of AMDs did you have in mind,
Opterons?
Thanks,
Nick
--
SUSE Labs, Novell Inc.
Send instant messages to your online friends http://au.messenger.yahoo.com
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to [EMAIL PROTECTED]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
