On Friday 20 July 2007, Andrew Morton wrote:
> + for_each_online_node(node) {
> struct spu *spu;
> mutex_lock(&spu_prio->active_mutex[node]);
> list_for_each_entry(spu, &spu_prio->active_list[node], list) {
> struct spu_context *ctx = spu->ctx;
> set_bit(SPU_SCHED_NOTIFY_ACTIVE, &ctx->sched_flags);
> - mb();
> + mb(); /* make sure any tasks woken up below */
> + /* can see the bit(s) set above */
> wake_up_all(&ctx->stop_wq);
> }
> mutex_unlock(&spu_prio->active_mutex[node]);
> _
>
> I still wonder about that barrier. At the least it should be smp_mb().
> But aren't our set_bit() semantics _alone_ sufficient to make this barrier
> unneeded?
I believe they are sufficient on powerpc, and should be everywhere, so
the mb() can probably go away entirely. set_bit() does a ldarx/stdcx,
which is guaranteed to be atomic across CPUs.
Arnd <><
_______________________________________________
Linuxppc-dev mailing list
[email protected]
https://ozlabs.org/mailman/listinfo/linuxppc-dev
