Re: [PATCH v8 2/2] sched/fair: Scan cluster before scanning LLC in wake-up path

[Chen Yu <yu.c.chen@intel.com>]

On 2023-06-08 at 14:45:54 +0800, Yicong Yang wrote:
> On 2023/6/8 11:26, Chen Yu wrote:
> > On 2023-05-31 at 16:21:00 +0800, Yicong Yang wrote:
> >> On 2023/5/30 22:39, Yicong Yang wrote:
> >>> On 2023/5/30 19:55, Peter Zijlstra wrote:
> >>>> On Tue, May 30, 2023 at 03:02:53PM +0800, Yicong Yang wrote:
> >>>>
> >>>>> diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
> >>>>> index 373ff5f55884..b8c129ed8b47 100644
> >>>>> --- a/kernel/sched/fair.c
> >>>>> +++ b/kernel/sched/fair.c
> >>>>> @@ -6994,6 +6994,30 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, bool
> >>>>>  		}
> >>>>>  	}
> >>>>>  
> >>>>> +	if (static_branch_unlikely(&sched_cluster_active)) {
> >>>>> +		struct sched_domain *sdc = rcu_dereference(per_cpu(sd_cluster, target));
> >>>>> +
> >>>>> +		if (sdc) {
> >>>>> +			for_each_cpu_wrap(cpu, sched_domain_span(sdc), target + 1) {
> >>>>> +				if (!cpumask_test_cpu(cpu, cpus))
> >>>>> +					continue;
> >>>>> +
> >>>>> +				if (has_idle_core) {
> >>>>> +					i = select_idle_core(p, cpu, cpus, &idle_cpu);
> >>>>> +					if ((unsigned int)i < nr_cpumask_bits)
> >>>>> +						return i;
> >>>>> +				} else {
> >>>>> +					if (--nr <= 0)
> >>>>> +						return -1;
> >>>>> +					idle_cpu = __select_idle_cpu(cpu, p);
> >>>>> +					if ((unsigned int)idle_cpu < nr_cpumask_bits)
> >>>>> +						return idle_cpu;
> >>>>> +				}
> >>>>> +			}
> >>>>> +			cpumask_andnot(cpus, cpus, sched_domain_span(sdc));
> >>>>> +		}
> >>>>> +	}
> >>>>
> >>>> Would not this:
> >>>>
> >>>> --- a/kernel/sched/fair.c
> >>>> +++ b/kernel/sched/fair.c
> >>>> @@ -6994,6 +6994,29 @@ static int select_idle_cpu(struct task_s
> >>>>  		}
> >>>>  	}
> >>>>  
> >>>> +	if (static_branch_unlikely(&sched_cluster_active)) {
> >>>> +		struct sched_group *sg = sd->groups;
> >>>> +		if (sg->flags & SD_CLUSTER) {
> >>>> +			for_each_cpu_wrap(cpu, sched_group_span(sg), target+1) {
> >>>> +				if (!cpumask_test_cpu(cpu, cpus))
> >>>> +					continue;
> >>>> +
> >>>> +				if (has_idle_core) {
> >>>> +					i = select_idle_core(p, cpu, cpus, &idle_cpu);
> >>>> +					if ((unsigned)i < nr_cpumask_bits)
> >>>> +						return 1;
> >>>> +				} else {
> >>>> +					if (--nr <= 0)
> >>>> +						return -1;
> >>>> +					idle_cpu = __select_idle_cpu(cpu, p);
> >>>> +					if ((unsigned)idle_cpu < nr_cpumask_bits)
> >>>> +						return idle_cpu;
> >>>> +				}
> >>>> +			}
> >>>> +			cpumask_andnot(cpus, cpus, sched_group_span(sg));
> >>>> +		}
> >>>> +	}
> >>>> +
> >>>>  	for_each_cpu_wrap(cpu, cpus, target + 1) {
> >>>>  		if (has_idle_core) {
> >>>>  			i = select_idle_core(p, cpu, cpus, &idle_cpu);
> >>>>
> >>>> also work? Then we can avoid the extra sd_cluster per-cpu variable.
> >>>>
> >>>
> >>> I thought it will be fine since sg->flags is derived from the child domain. But practically it doesn't.
> >>> Tested on a 2P Skylake server with no clusters, add some debug messages to see how sg->flags appears:
> >>>
> >>> diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
> >>> index 69968ed9ffb9..5c443b74abf5 100644
> >>> --- a/kernel/sched/topology.c
> >>> +++ b/kernel/sched/topology.c
> >>> @@ -90,8 +90,8 @@ static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
> >>>
> >>>                 cpumask_or(groupmask, groupmask, sched_group_span(group));
> >>>
> >>> -               printk(KERN_CONT " %d:{ span=%*pbl",
> >>> -                               group->sgc->id,
> >>> +               printk(KERN_CONT " %d:{ cluster: %s span=%*pbl",
> >>> +                               group->sgc->id, group->flags & SD_CLUSTER ? "true" : "false",
> >>>                                 cpumask_pr_args(sched_group_span(group)));
> >>>
> >>>                 if ((sd->flags & SD_OVERLAP) &&
> >>>
> >>> Unfortunately the result doesn't match what I expected, the MC domain's sg->flags still marked
> >>> as cluster:
> >>>
> >>> [    8.886099] CPU0 attaching sched-domain(s):
> >>> [    8.889539]  domain-0: span=0,40 level=SMT
> >>> [    8.893538]   groups: 0:{ cluster: false span=0 }, 40:{ cluster: false span=40 }
> >>> [    8.897538]   domain-1: span=0-19,40-59 level=MC
> >>> [    8.901538]    groups: 0:{ cluster: true span=0,40 cap=2048 }, 1:{ cluster: true span=1,41 cap=2048 }, 2:{ cluster: true span=2,42 cap=2048 }, 3:{ cluster: true span=3,43 cap=2048 }, 4:{ cluster: true span=4,44 cap=2048 }, 5:{ cluster: true span=5,45 cap=2048 }, 6:{ cluster: true span=6,46 cap=2048 }, 7:{ cluster: true span=7,47 cap=2048 }, 8:{ cluster: true span=8,48 cap=2048 }, 9:{ cluster: true span=9,49 cap=2048 }, 10:{ cluster: true span=10,50 cap=2048 }, 11:{ cluster: true span=11,51 cap=2048 }, 12:{ cluster: true span=12,52 cap=2048 }, 13:{ cluster: true span=13,53 cap=2048 }, 14:{ cluster: true span=14,54 cap=2048 }, 15:{ cluster: true span=15,55 cap=2048 }, 16:{ cluster: true span=16,56 cap=2048 }, 17:{ cluster: true span=17,57 cap=2048 }, 18:{ cluster: true span=18,58 cap=2048 }, 19:{ cluster: true span=19,59 cap=2048 }
> >>> [    8.905538]    domain-2: span=0-79 level=NUMA
> >>> [    8.909538]     groups: 0:{ cluster: false span=0-19,40-59 cap=40960 }, 20:{ cluster: false span=20-39,60-79 cap=40960 }
> >>>
> >>> I assume we didn't handle the sg->flags correctly on the domain degeneration. Simply checked the code seems
> >>> we've already make sg->flags = 0 on degeneration, maybe I need to check where's wrong.
> >>>
> >>
> >> Currently we only update the groups' flags to 0 for the final lowest domain in [1]. The upper
> >> domains' group won't be updated if degeneration happens. So we cannot use the suggested approach
> >> for cluster scanning and sd_cluster per-cpu variable is still needed.
> >>
> >> [1] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/sched/topology.c?h=v6.4-rc4#n749
> >>
> >> Thanks.
> >>
> >>
> > Is this an issue? Suppose sched domain A's parent domain
> > is B, B's parent sched domain is C. When B degenerates, C's child domain
> > pointer is adjusted to A. However, currently the code does not adjust C's
> > sched groups' flags. Should we adjust C's sched groups flags to be the same
> > as A to keep consistency?
> 
> It depends on whether we're going to use it. currently only asym_smt_can_pull_tasks() uses
> it within the SMT so I think update the lowest domain's group flag works. For correctness
> all the domain group's flag should derives from its real child. I tried to solve this at group
> building but seems hard to do, at that time we don't know whether a domain is going to degenerate
> or not since the groups it not built.
> 
> > 
> > diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
> > index 6198fa135176..fe3fd70f2313 100644
> > --- a/kernel/sched/topology.c
> > +++ b/kernel/sched/topology.c
> > @@ -713,14 +713,13 @@ cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
> >  
> >  	/* Remove the sched domains which do not contribute to scheduling. */
> >  	for (tmp = sd; tmp; ) {
> > -		struct sched_domain *parent = tmp->parent;
> > +		struct sched_domain *parent = tmp->parent, *pparent;
> >  		if (!parent)
> >  			break;
> >  
> >  		if (sd_parent_degenerate(tmp, parent)) {
> > -			tmp->parent = parent->parent;
> > -			if (parent->parent)
> > -				parent->parent->child = tmp;
> > +			pparent = parent->parent;
> > +			tmp->parent = pparent;
> >  			/*
> >  			 * Transfer SD_PREFER_SIBLING down in case of a
> >  			 * degenerate parent; the spans match for this
> > @@ -728,6 +727,18 @@ cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
> >  			 */
> >  			if (parent->flags & SD_PREFER_SIBLING)
> >  				tmp->flags |= SD_PREFER_SIBLING;
> > +
> > +			if (pparent) {
> > +				struct sched_group *sg = pparent->groups;
> > +
> > +				do {
> > +					sg->flags = tmp->flags;
> 
> May need to test on some heterogeous platforms. Does it always stand that child domain of CPU from
> remote group have the same flags with @tmp?
>
Good question, for heterogenous platforms sched groups within the same domain might not always
have the same flags, because if group1 and group2 sit in big/small-core child domain, they could
have different balance flags in theory. Maybe only update the local group's flag is accurate.

I found Tim has proposed a fix for a similar scenario, and it is for SD_SHARE_CPUCAPACITY, and it
should be in tip:
https://lore.kernel.org/lkml/168372654916.404.6677242284447941021.tip-bot2@tip-bot2/
We could adjust it based on his change to remove SD_CLUSTER, or we can
replace groups->flag with tmp->flag directly, in case we have other flags to be
adjusted in the future.

thanks,
Chenyu