Attachment 874403 Details for Bug 1065304 – sched/numa: fix cpu_power initialization

[patch] sched/numa: fix cpu_power initialization

0001-sched-numa-fix-cpu_power-initialization.patch (text/plain), 2.73 KB, created by Radim Krčmář on 2014-03-14 12:46:26 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Radim Krčmář

Created: 2014-03-14 12:46:26 UTC

Size: 2.73 KB

patch

obsolete

>From a8afb9b3e9b7cb1c88949b7148f958d2f1f2a216 Mon Sep 17 00:00:00 2001
>From: =?UTF-8?q?Radim=20Kr=C4=8Dm=C3=A1=C5=99?= <rkrcmar@redhat.com>
>Date: Wed, 5 Mar 2014 14:05:46 +0100
>Subject: [RHEL6.6 kernel PATCH] sched/numa: fix cpu_power initialization
>MIME-Version: 1.0
>Content-Type: text/plain; charset=UTF-8
>Content-Transfer-Encoding: 8bit
>
>bug:  https://bugzilla.redhat.com/show_bug.cgi?id=1065304
>brew: https://brewweb.devel.redhat.com/brew/taskinfo?taskID=7194671
>Tested by me on hp-dl585-02, ibm-ls22-05 and weird-topology kvm guests.
>A functionally identical patch was verified by the reporter.
>This patch has no upstream equivalent.
>
>Our problem lies in init_numa_sched_groups_power().
>This function expects that every numa domain is a superset of physical,
>so when it is not, some (all) cpus in a domain might not get accounted
>for, which causes division by zero problems on boot.
>(Weird topology example: one cpu with two cores and two numa domains.)
>
>Commit 08d7ef55afc468ed6cb29d892b53063dc382c9fa made weird topologies
>bootable by setting a sane default value of cpu_power.
>(|cpus in domain| * SCHED_POWER_SCALE)
>This was done with a presumption that normal topologies will correct the
>value later.  False presumption.
>Only the first domain does, thus cpu_powers also get imbalanced.
>
>To fix this, we make init_numa_sched_groups_power() aware of the new
>default; by substracting SCHED_POWER_SCALE for every cpu we add from
>physical domains.
>This does not affect normal topologies, where we substract everything we
>added, but keeps SCHED_POWER_SCALE for every cpu missed with weird
>topologies.
>It makes sense to set initial cpu_power in the same for-cycle, (and we
>need nonzero for sd_allnodes) so we do not have to set cpu_power in
>build_numa_sched_groups(). (In the end, I'm keeping it to stay safe.)
>
>Signed-off-by: Radim KrÄmÃ¡Å <rkrcmar@redhat.com>
>---
> kernel/sched.c | 6 ++++++
> 1 file changed, 6 insertions(+)
>
>diff --git a/kernel/sched.c b/kernel/sched.c
>index f8dca86..9f6397c 100644
>--- a/kernel/sched.c
>+++ b/kernel/sched.c
>@@ -9176,6 +9176,9 @@ static void init_numa_sched_groups_power(struct sched_group *group_head)
> 	if (!sg)
> 		return;
> 	do {
>+		/* Estimate the final value to avoid /0 on weird topologies. */
>+		sg->cpu_power = SCHED_LOAD_SCALE * cpumask_weight(
>+		                    sched_group_cpus(sg));
> 		for_each_cpu(j, sched_group_cpus(sg)) {
> 			struct sched_domain *sd;
> 
>@@ -9189,6 +9192,9 @@ static void init_numa_sched_groups_power(struct sched_group *group_head)
> 			}
> 
> 			sg->cpu_power += sd->groups->cpu_power;
>+			/* Discharge initial estimate of updated cpus. */
>+			sg->cpu_power -= SCHED_LOAD_SCALE * cpumask_weight(
>+			                     sched_group_cpus(sd->groups));
> 		}
> 		sg = sg->next;
> 	} while (sg != group_head);
>-- 
>1.8.5.3
>

Actions: View | Diff

Attachments on bug 1065304: 868573 | 870669 | 874403 | 890505