Skip to contentSkip to conversation

x86: set cpu_llc_id on AMD CPUs

PATCH In review Andreas Herrmann 1 participant 2 messages Jun 19, 2009 – Jun 21, 2009 x86

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86: set cpu_llc_id on AMD CPUs

This counts when building sched domains in case NUMA information is not available.
(See cpu_coregroup_mask() which uses llc_shared_map which in turn is created based on cpu_llc_id.)

Currently Linux builds domains as follows
(example from a dual socket quad-core system)

CPU0 attaching sched-domain:
domain 0: span 0-7 level CPU
groups: 0 1 2 3 4 5 6 7

...

CPU7 attaching sched-domain:
domain 0: span 0-7 level CPU
groups: 7 0 1 2 3 4 5 6

Ever since that is borked for multi-core AMD CPU systems. This patch fixes that and now we get

CPU0 attaching sched-domain:
domain 0: span 0-3 level MC
groups: 0 1 2 3
domain 1: span 0-7 level CPU

    groups: 0-3 4-7

...

CPU7 attaching sched-domain:
domain 0: span 4-7 level MC
groups: 7 4 5 6
domain 1: span 0-7 level CPU

    groups: 4-7 0-3

This allows scheduler to assign tasks to cores on different sockets (i.e. that don't share last level cache) for performance reasons.

---
 arch/x86/kernel/cpu/amd.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)

Please apply.
Maybe that fix should go to stable-tree(s) as well.

Thanks,
Andreas

arch/x86/kernel/cpu/amd.c +3 -1
index e5b27d8..28e5f59 100644
@@ -258,13 +258,15 @@ static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
 {
 #ifdef CONFIG_X86_HT
 	unsigned bits;
+	int cpu = smp_processor_id();
 
 	bits = c->x86_coreid_bits;
-
 	/* Low order bits define the core id (index of core in socket) */
 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
 	/* Convert the initial APIC ID into the socket ID */
 	c->phys_proc_id = c->initial_apicid >> bits;
+	/* use socket ID also for last level cache */
+	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
 #endif
 }
 
Signature · 1 line
1.6.3.1
#1 lore raw

[tip:x86/urgent] x86: Set cpu_llc_id on AMD CPUs

Commit-ID:  99bd0c0fc4b04da54cb311953ef9489931c19c63
Gitweb:     http://git.kernel.org/tip/99bd0c0fc4b04da54cb311953ef9489931c19c63
Author:     Andreas Herrmann <[email protected]>
AuthorDate: Fri, 19 Jun 2009 10:59:09 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Sun, 21 Jun 2009 10:13:32 +0200

x86: Set cpu_llc_id on AMD CPUs

This counts when building sched domains in case NUMA information is not available.

( See cpu_coregroup_mask() which uses llc_shared_map which in turn is
created based on cpu_llc_id. )

Currently Linux builds domains as follows:
(example from a dual socket quad-core system)

CPU0 attaching sched-domain:
domain 0: span 0-7 level CPU
groups: 0 1 2 3 4 5 6 7

...

CPU7 attaching sched-domain:
domain 0: span 0-7 level CPU
groups: 7 0 1 2 3 4 5 6

Ever since that is borked for multi-core AMD CPU systems. This patch fixes that and now we get a proper:

CPU0 attaching sched-domain:
domain 0: span 0-3 level MC
groups: 0 1 2 3
domain 1: span 0-7 level CPU

    groups: 0-3 4-7

...

CPU7 attaching sched-domain:
domain 0: span 4-7 level MC
groups: 7 4 5 6
domain 1: span 0-7 level CPU

    groups: 4-7 0-3

This allows scheduler to assign tasks to cores on different sockets (i.e. that don't share last level cache) for performance reasons.

LKML-Reference: <[email protected]>

---
 arch/x86/kernel/cpu/amd.c |    4 +++-
 1 files changed, 3 insertions(+), 1 deletions(-)
arch/x86/kernel/cpu/amd.c +3 -1
index e5b27d8..28e5f59 100644
@@ -258,13 +258,15 @@ static void __cpuinit amd_detect_cmp(struct cpuinfo_x86 *c)
 {
 #ifdef CONFIG_X86_HT
 	unsigned bits;
+	int cpu = smp_processor_id();
 
 	bits = c->x86_coreid_bits;
-
 	/* Low order bits define the core id (index of core in socket) */
 	c->cpu_core_id = c->initial_apicid & ((1 << bits)-1);
 	/* Convert the initial APIC ID into the socket ID */
 	c->phys_proc_id = c->initial_apicid >> bits;
+	/* use socket ID also for last level cache */
+	per_cpu(cpu_llc_id, cpu) = c->phys_proc_id;
 #endif
 }