Messages in this thread | | | Date | Fri, 26 Nov 2010 18:05:58 +0100 | From | Tejun Heo <> | Subject | Re: [thiscpuops upgrade 05/10] x86: Use this_cpu_inc_return for nmi counter |
| |
On 11/26/2010 06:02 PM, Christoph Lameter wrote: > On Fri, 26 Nov 2010, Tejun Heo wrote: > >>> - __this_cpu_inc(alert_counter); >>> - if (__this_cpu_read(alert_counter) == 5 * nmi_hz) >>> + if (__this_cpu_inc_return(alert_counter) == 5 * nmi_hz) >> >> Hmmm... one worry I have is that xadd, being not a very popular >> operation, might be slower than add and read. Using it for atomicity >> would probably be beneficial in most cases but have you checked this >> actually is cheaper? > > XADD takes 3 uops. INC 1 and MOV 1 uop. So there is an additiona uop. > > However, a memory fetch from l1 takes a mininum 4 cycles. Doing that twice > already ends up with at least 8 cycles.
Thanks for the explanation. It might be beneficial to note performance characteristics on top of the x86 implementation? Anyways, for this and the following simple conversion patches.
Reviewed-by: Tejun Heo <tj@kernel.org>
-- tejun
| |