| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Message-ID: <ZIFKf2PkIkiRmUaK@chenyu5-mobl2.ccr.corp.intel.com>
Date: Thu, 8 Jun 2023 11:26:55 +0800
From: Chen Yu <yu.c.chen@...el.com>
To: Yicong Yang <yangyicong@...wei.com>
CC: Peter Zijlstra <peterz@...radead.org>, <yangyicong@...ilicon.com>,
<mingo@...hat.com>, <juri.lelli@...hat.com>,
<vincent.guittot@...aro.org>, <dietmar.eggemann@....com>,
<tim.c.chen@...ux.intel.com>, <gautham.shenoy@....com>,
<mgorman@...e.de>, <vschneid@...hat.com>,
<linux-kernel@...r.kernel.org>,
<linux-arm-kernel@...ts.infradead.org>, <rostedt@...dmis.org>,
<bsegall@...gle.com>, <bristot@...hat.com>,
<prime.zeng@...wei.com>, <jonathan.cameron@...wei.com>,
<ego@...ux.vnet.ibm.com>, <srikar@...ux.vnet.ibm.com>,
<linuxarm@...wei.com>, <21cnbao@...il.com>,
<kprateek.nayak@....com>, <wuyun.abel@...edance.com>,
Barry Song <baohua@...nel.org>,
Tim Chen <tim.c.chen@...el.com>,
Ricardo Neri <ricardo.neri@...el.com>,
Len Brown <len.brown@...el.com>
Subject: Re: [PATCH v8 2/2] sched/fair: Scan cluster before scanning LLC in
wake-up path
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?
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;
+ sg = sg->next;
+ } while (sg != pparent->groups);
+
+ pparent->child = tmp;
+ }
+
destroy_sched_domain(parent);
} else
tmp = tmp->parent;
--
2.25.1
Besides per your description I found that something is not quite right:
diff --git a/kernel/sched/topology.c b/kernel/sched/topology.c
index 6682535e37c8..6198fa135176 100644
--- a/kernel/sched/topology.c
+++ b/kernel/sched/topology.c
@@ -747,6 +747,7 @@ cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
*/
do {
sg->flags = 0;
+ sg = sg->next;
} while (sg != sd->groups);
sd->child = NULL;
--
2.25.1
thanks,
Chenyu
Powered by blists - more mailing lists