Messages in this thread | | | Date | Thu, 19 Apr 2012 14:34:45 +0800 | From | Yong Zhang <> | Subject | Re: RFC [patch] sched,cgroup_sched: convince RT_GROUP_SCHED throttle to work |
| |
On Wed, Apr 18, 2012 at 10:38:07AM +0200, Mike Galbraith wrote: > On Wed, 2012-04-18 at 15:48 +0800, Yong Zhang wrote: > > On Wed, Apr 18, 2012 at 08:27:50AM +0200, Mike Galbraith wrote: > > > On Wed, 2012-04-18 at 13:20 +0800, Yong Zhang wrote: > > > > > > > > --- a/kernel/sched/core.c > > > > > +++ b/kernel/sched/core.c > > > > > @@ -5875,6 +5875,11 @@ cpu_attach_domain(struct sched_domain *s > > > > > sd->child = NULL; > > > > > } > > > > > > > > > > + if (sd) > > > > > + cpumask_clear_cpu(cpu, cpu_isolated_map); > > > > > + else > > > > > + cpumask_set_cpu(cpu, cpu_isolated_map); > > > > > + > > > > > > > > Do we allow this? > > > > > > Yeah, isolating CPUS 2-3... > > > > Hmm...magic cpuset ;-) > > The _only_ cpuset if you have a low jitter requirement. > > > > [ 3011.444345] CPU0 attaching NULL sched-domain. > > > [ 3011.448719] CPU1 attaching NULL sched-domain. > > > [ 3011.453107] CPU2 attaching NULL sched-domain. > > > [ 3011.457466] CPU3 attaching NULL sched-domain. > > > [ 3011.461892] CPU0 attaching sched-domain: > > > [ 3011.465813] domain 0: span 0-1 level MC > > > [ 3011.469761] groups: 0 1 > > > [ 3011.472415] CPU1 attaching sched-domain: > > > [ 3011.476333] domain 0: span 0-1 level MC > > > [ 3011.480266] groups: 1 0 > > > [ 3011.482988] CPU2 attaching sched-domain: > > > [ 3011.486919] domain 0: span 2-3 level MC > > > [ 3011.490851] groups: 2 3 > > > [ 3011.493502] CPU3 attaching sched-domain: > > > [ 3011.497422] domain 0: span 2-3 level MC > > > [ 3011.501367] groups: 3 2 > > > [ 3011.504214] CPU2 attaching NULL sched-domain. > > > [ 3011.508569] CPU3 attaching NULL sched-domain. > > > > But another scenario comes into head: > > What will happen when a rt_rq is throttled the very CPU is > > attached to NULL domain? > > Yup, _somebody_ will hit the throttle only once :) > > That's what gets fixed by either killing the throttle entirely for > isolated CPUs, or making the throttle work until the guy who needed > isolation turns noise maker off. I prefer reconnect the dots, because > that doesn't touch the fast path.
I like it better too :)
Thanks, Yong
| |