On 07/13/2018 03:18 PM, Michael Bringmann wrote:
pmt/numa: Disable arch_update_cpu_topology during post migration
CPU readd updates when evaluating device-tree changes after LPM
to avoid thread deadlocks trying to update node assignments.
System timing between all of the threads and timers restarted in
a migrated system overlapped frequently allowing tasks to start
acquiring resources (get_online_cpus) needed by rebuild_sched_domains.
Defer the operation of that function until after the CPU readd has
completed.
Signed-off-by: Michael Bringmann <m...@linux.vnet.ibm.com>
---
arch/powerpc/platforms/pseries/hotplug-cpu.c | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/arch/powerpc/platforms/pseries/hotplug-cpu.c
b/arch/powerpc/platforms/pseries/hotplug-cpu.c
index 1906ee57..df1791b 100644
--- a/arch/powerpc/platforms/pseries/hotplug-cpu.c
+++ b/arch/powerpc/platforms/pseries/hotplug-cpu.c
@@ -26,6 +26,7 @@
#include <linux/sched.h> /* for idle_task_exit */
#include <linux/sched/hotplug.h>
#include <linux/cpu.h>
+#include <linux/cpuset.h>
#include <linux/of.h>
#include <linux/slab.h>
#include <asm/prom.h>
@@ -684,9 +685,15 @@ static int dlpar_cpu_readd_by_index(u32 drc_index)
pr_info("Attempting to re-add CPU, drc index %x\n", drc_index);
+ arch_update_cpu_topology_suspend();
rc = dlpar_cpu_remove_by_index(drc_index, false);
- if (!rc)
+ arch_update_cpu_topology_resume();
+
+ if (!rc) {
+ arch_update_cpu_topology_suspend();
rc = dlpar_cpu_add(drc_index, false);
+ arch_update_cpu_topology_resume();
+ }
A couple of questions...Why not disable across the entire remove and add
operations instead of disabling for each operation?
Also, what about other CPU add/remove routines, do they need to do
similar disabling?
-Nathan
if (rc)
pr_info("Failed to update cpu at drc_index %lx\n",
