2005-10-25 11:17:49

by Magnus Damm

[permalink] [raw]
Subject: [PATCH] NUMA: broken per cpu pageset counters

The NUMA counters in struct per_cpu_pageset (linux/mmzone.h) are never cleared
today. This works ok for CPU 0 on NUMA machines because boot_pageset[] is
already zero, but for other CPU:s this results in uninitialized counters.

Signed-off-by: Magnus Damm <[email protected]>
---

Tested on dual x86_64 hardware with 2.6.14-rc5-git5.

/proc/zoneinfo:
....
numa_hit: 15064664480448666958
numa_miss: 15061573751478353360
numa_foreign: 15082960697523852788
interleave_hit: 15083101074612482007
local_node: 15084090651888212814
other_node: 5860683432910092144
....

--- linux-2.6.14-rc5-git5/mm/page_alloc.c 2005-10-24 15:37:48.000000000 +0900
+++ linux-2.6.14-rc5-git5-setup_pageset_zero_fix/mm/page_alloc.c 2005-10-25 19:48:06.000000000 +0900
@@ -1750,6 +1750,8 @@ inline void setup_pageset(struct per_cpu
{
struct per_cpu_pages *pcp;

+ memset(p, 0, sizeof(*p));
+
pcp = &p->pcp[0]; /* hot */
pcp->count = 0;
pcp->low = 2 * batch;