The kernel boots up on the BP with an initial memory policy of
MPOL_INTERLEAVE (start_kernel() -> numa_policy_init()). This is
done to avoid all boot data structure allocations to be off the boot
node. Current mainline resets the memory policy to MPOL_DEFAULT at init(),
just before calling userspace init. But, this is too late and leaves
the kernel idle thread with MPOL_INTERLEAVE, causing later allocations
off interrupt/BH context to use MPOL_INTERLEAVE, rather than MPOL_DEFAULT
(if the interrupt occurs on an idle cpu). This can be fixed by changing
the mempolicy to MPOL_DEFAULT just before the 'init' kernel thread is
spawned, OR just before smp_init() (smp_init would spawn idle threads).
The following patch uses the latter approach and moves numa_default_policy()
to just before smp_init().
Signed-off-by: Alok N Kataria <[email protected]>
Signed-off-by: Ravikiran Thirumalai <[email protected]>
Signed-off-by: Shai Fultheim <[email protected]>
Index: linux-2.6.18/init/main.c
===================================================================
--- linux-2.6.18.orig/init/main.c 2006-09-19 20:42:06.000000000 -0700
+++ linux-2.6.18/init/main.c 2006-09-28 15:13:01.000000000 -0700
@@ -703,6 +703,8 @@ static int init(void * unused)
do_pre_smp_initcalls();
+ numa_default_policy();
+
smp_init();
sched_init_smp();
@@ -738,7 +740,6 @@ static int init(void * unused)
unlock_kernel();
mark_rodata_ro();
system_state = SYSTEM_RUNNING;
- numa_default_policy();
if (sys_open((const char __user *) "/dev/console", O_RDWR, 0) < 0)
printk(KERN_WARNING "Warning: unable to open an initial console.\n");