On Tue, Jan 22, 2013 at 03:50:39PM +0900, Joonsoo Kim wrote: > > -static DEFINE_PER_CPU(struct global_cwq, global_cwq); > > +static DEFINE_PER_CPU_ALIGNED(struct worker_pool [NR_STD_WORKER_POOLS], > > + cpu_std_worker_pools); > > static DEFINE_PER_CPU_SHARED_ALIGNED(atomic_t, > > pool_nr_running[NR_STD_WORKER_POOLS]); > > Why worker_pool is defined as DEFINE_PER_CPU_ALIGNED? > > And this makes only worker_pool[0] aligned with cacheline. > worker_pool[1] is not aligned with cacheline. > Now, we have a spin_lock for each instance of worker_pool and > each one is independent instance. > So, IMHO, it is better to align worker_pool[1] with cacheline.
Right, it's just carried over from the time when there was single pool implemented as global_gcwq. We probably should define struct worker_pool as cache aligned. Thanks. -- tejun -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majord...@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
