Messages in this thread | | | From | Barry Song <> | Date | Thu, 25 Nov 2021 01:02:00 +1300 | Subject | Re: [PATCH v2] sched/fair: Remove the cost of a redundant cpumask_next_wrap in select_idle_cpu |
| |
On Thu, Nov 25, 2021 at 12:57 AM Barry Song <21cnbao@gmail.com> wrote: > > On Thu, Nov 25, 2021 at 12:49 AM Barry Song <21cnbao@gmail.com> wrote: > > > > On Thu, Nov 25, 2021 at 12:13 AM Peter Zijlstra <peterz@infradead.org> wrote: > > > > > > On Wed, Nov 24, 2021 at 05:15:46PM +0800, Barry Song wrote: > > > > diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c > > > > index 6e476f6..8cd23f1 100644 > > > > --- a/kernel/sched/fair.c > > > > +++ b/kernel/sched/fair.c > > > > @@ -6278,6 +6278,7 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool > > > > time = cpu_clock(this); > > > > } > > > > > > > > + --nr; > > > > for_each_cpu_wrap(cpu, cpus, target + 1) { > > > > if (has_idle_core) { > > > > i = select_idle_core(p, cpu, cpus, &idle_cpu); > > > > @@ -6285,11 +6286,11 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool > > > > return i; > > > > > > > > } else { > > > > - if (!--nr) > > > > - return -1; > > > > idle_cpu = __select_idle_cpu(cpu, p); > > > > if ((unsigned int)idle_cpu < nr_cpumask_bits) > > > > break; > > > > + if (!--nr) > > > > + return -1; > > > > } > > > > } > > > > > > This way nr can never be 1 for a single iteration -- it current isn't, > > > but that's besides the point. > > > > Yep. nr=1 seems to be a corner case. > > if nr=1, the original code will return -1 directly without scanning > > any cpu. but the new code will scan > > one time because I haven't checked if(!--nr) and return before > > for_each_cpu_wrap(). so this changes > > the behaviour for this corner case. > > > > but if i change "--nr" to "nr--", this new code will scan nr times > > rather than nr-1, this changes the behaviour > > for all cases besides nr!=1. The original code was looping nr times > > but scanning nr-1 times only. > > > > so you prefer here the codes should scan nr times and change the > > scanning amount from nr-1 > > to nr? > > Let me make it clearer. if nr=5, the original code will loop 5 times, > but in the 5th loop, it returns directly, so __select_idle_cpu is > only done 4 times. > > if nr=1, the original code will loop 1 time, but in the 1st loop, > it returns directly, so __select_idle_cpu is done 0 times.
this is also why in the first version of patch, i did this: span_avg = sd->span_weight * avg_idle; if (span_avg > 4*avg_cost) - nr = div_u64(span_avg, avg_cost); + nr = div_u64(span_avg, avg_cost) - 1; else - nr = 4; + nr = 3;
because we are actually scanning 3 times or div_u64(span_avg, avg_cost) - 1 times but not 4 times or div_u64(span_avg, avg_cost) times.
this is not confusing at all. the only thing which is confusing is the original code.
> > if i change the code to if(!nr--), while nr=5, the new code will > do __select_idle_cpu() 5 times rather than 4 times in the > original code. > > but of course the new code changes the __select_idle_cpu > from zero to one time for the corner case nr==1. > > > > Thanks > > Barry
| |