git commit 316ad248307fba13be40f01e92a22b89457c32bc changed
set_cpu_sibling_map() to a single loop.
On a 2 socket system with 4 nodes, with 8 cores/socket top of tree
now shows
.> cat /proc/cpuinfo | grep cores
cpu cores : 2
cpu cores : 4
cpu cores : 5
cpu cores : 6
cpu cores : 7
cpu cores : 8
cpu cores : 9
cpu cores : 10
cpu cores : 2
cpu cores : 4
cpu cores : 5
cpu cores : 6
cpu cores : 7
cpu cores : 8
cpu cores : 9
cpu cores : 10
cpu cores : 15
cpu cores : 15
cpu cores : 13
cpu cores : 11
cpu cores : 9
cpu cores : 7
cpu cores : 5
cpu cores : 3
cpu cores : 15
cpu cores : 15
cpu cores : 13
cpu cores : 11
cpu cores : 9
cpu cores : 7
cpu cores : 5
cpu cores : 3
This is wrong. The kernel should show
cpu cores : 8
for each entry.
set_cpu_sibling_map() must be two separate loops, the first which
calculates the cpu's sibling_maps and the second which determines
how many cores have been booted. The code does a check on
(cpumask_weight(cpu_sibling_mask(cpu)). The problem is that
the sibling_mask may not be fully populated by the time this
value is checked resulting in incorrect booted_cores values.
I've found this broken and tested the fix on both AMD multi-socket and Intel
multi-socket systems (as well as single socket single node systems).
-----8<-----
The code in set_cpu_sibling_map() checks the value of
cpu_sibling_mask(cpu) before the mask is fully populated. This results in
an incorrect calculation of each cpu's booted_cores. Move back to a two
pass mechanism to properly calculate booted_cores.
Signed-off-by: Prarit Bhargava <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Cc: Ingo Molnar <[email protected]>
Cc: "H. Peter Anvin" <[email protected]>
Cc: [email protected]
---
arch/x86/kernel/smpboot.c | 4 ++++
1 files changed, 4 insertions(+), 0 deletions(-)
diff --git a/arch/x86/kernel/smpboot.c b/arch/x86/kernel/smpboot.c
index f56f96d..019248e 100644
--- a/arch/x86/kernel/smpboot.c
+++ b/arch/x86/kernel/smpboot.c
@@ -381,6 +381,10 @@ void __cpuinit set_cpu_sibling_map(int cpu)
if ((i == cpu) || (has_mc && match_llc(c, o)))
link_mask(llc_shared, cpu, i);
+ }
+
+ for_each_cpu(i, cpu_sibling_setup_mask) {
+ o = &cpu_data(i);
if ((i == cpu) || (has_mc && match_mc(c, o))) {
link_mask(core, cpu, i);
--
1.7.1