Messages in this thread | | | Date | Thu, 9 Dec 2010 17:40:27 -0600 (CST) | From | Christoph Lameter <> | Subject | Re: [cpuops cmpxchg V1 2/4] x86: this_cpu_cmpxchg and this_cpu_xchg operations |
| |
On Wed, 8 Dec 2010, H. Peter Anvin wrote:
> On 12/08/2010 10:17 AM, Mathieu Desnoyers wrote: > > > > Hi Christoph, > > > > Can you show if this provides savings in terms of: > > > > - instruction cache footprint > > - cycles required to run > > - large-scale impact on the branch prediction buffers > > > > Given that this targets per-cpu data only, the additional impact on cache-line > > exchange traffic of using cmpxchg over xchg (cache-line not grabbed as exclusive > > by the initial read) should not really matter. > > > > I'm CCing Arjan and HPA, because they might have some interesting insight into > > the performance impact of lock-prefixed xchg vs using local cmpxchg in a loop. > > > > XCHG is always locked; it doesn't need the prefix. Unfortunately, > unlike on the 8086 on modern processors locks have a real cost.
So should we use xchg or a loop using prefixless cmpxchg instead when referring to per cpu data and requiring only per cpu atomicness?
| |