x86, sched: Workaround broken sched domain creation for AMD Magny-Cours

Current sched domain creation code can't handle multi-node processors.
When switching to power_savings scheduling errors show up and
system might hang later on (due to broken sched domain hierarchy):

  # echo 0  >> /sys/devices/system/cpu/sched_mc_power_savings
  CPU0 attaching sched-domain:
   domain 0: span 0-5 level MC
    groups: 0 1 2 3 4 5
    domain 1: span 0-23 level NODE
     groups: 0-5 6-11 18-23 12-17
  ...
  # echo 1  >> /sys/devices/system/cpu/sched_mc_power_savings
  CPU0 attaching sched-domain:
   domain 0: span 0-11 level MC
    groups: 0 1 2 3 4 5 6 7 8 9 10 11
  ERROR: parent span is not a superset of domain->span
    domain 1: span 0-5 level CPU
  ERROR: domain->groups does not contain CPU0
     groups: 6-11 (__cpu_power = 12288)
  ERROR: groups don't span domain->span
     domain 2: span 0-23 level NODE
      groups:
  ERROR: domain->cpu_power not set

  ERROR: groups don't span domain->span
  ...

Fixing all aspects of power-savings scheduling for Magny-Cours needs
some larger changes in the sched domain creation code.

As a short-term and temporary workaround avoid the problems by
extending "the worst possible hack" ;-(
and always use llc_shared_map on AMD Magny-Cours when MC domain span
is calculated.

With this I get:

  # echo 1  >> /sys/devices/system/cpu/sched_mc_power_savings
  CPU0 attaching sched-domain:
   domain 0: span 0-5 level MC
    groups: 0 1 2 3 4 5
    domain 1: span 0-5 level CPU
     groups: 0-5 (__cpu_power = 6144)
     domain 2: span 0-23 level NODE
      groups: 0-5 (__cpu_power = 6144) 6-11 (__cpu_power = 6144) 18-23 (__cpu_power = 6144) 12-17 (__cpu_power = 6144)
  ...

I.e. no errors during sched domain creation, no system hangs, and also
mc_power_savings scheduling works to a certain extend.

Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Signed-off-by: Andreas Herrmann <andreas.herrmann3@amd.com>
Signed-off-by: H. Peter Anvin <hpa@zytor.com>

1 files changed, 2 insertions, 1 deletions

diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
index 2fecda69ee64..c36cc1452cdc 100644
--- a/arch/x86/kernel/smpboot.c
+++ b/arch/x86/kernel/smpboot.c
@@ -434,7 +434,8 @@ const struct cpumask *cpu_coregroup_mask(int cpu)
          * For perf, we return last level cache shared map.
          * And for power savings, we return cpu_core_map
          */
-        if (sched_mc_power_savings || sched_smt_power_savings)
+        if ((sched_mc_power_savings || sched_smt_power_savings) &&
+            !(cpu_has(c, X86_FEATURE_AMD_DCM)))
                 return cpu_core_mask(cpu);
         else
                 return c->llc_shared_map;