Bogus load average and cpu times on x86_64 SMP kernels

I've been seeing bogus values from /proc/loadavg on an x86-64 SMP kernel (but not UP).

$ cat /proc/loadavg
-1012098.26 922203.26 -982431.60 1/112 2688

This is in the current git tree. I'm also seeing strange values in /proc/stat:

cpu 2489 40 920 60530 9398 171 288 1844674407350
cpu0 2509 60 940 60550 9418 191 308 0

The first line is the sum of all cpus (I only have one), so it's picking up up bad data from the non-present cpus. The last value, stolen time, is completely bogus since that value is only ever used on s390.

It looks to me like there is some problem with how the per-cpu structures are being initialized, or are getting corrupted. I have not been able to test i386 SMP yet to see if the problem is x86_64 specific.

--
Brian Gerst

I found the culprit: CPU hotplug. The problem is that prefill_possible_map() is called after setup_per_cpu_areas(). This leaves the per-cpu data sections for the future cpus uninitialized (still pointing to the original per-cpu data, which is initmem). Since the cpus exists in cpu_possible_map, for_each_cpu will iterate over them even though the per-cpu data is invalid.

--
Brian Gerst

Initialize cpu_possible_map properly in the hotplug cpu case. Before this, the map was filled in after the per-cpu areas were set up. This left those cpus pointing to initmem and causing invalid data in /proc/stat and elsewhere.

Subject: [PATCH] Fix hotplug cpu on x86_64

Initialize cpu_possible_map properly in the hotplug cpu case. Before this, the map was filled in after the per-cpu areas were set up. This left those cpus pointing to initmem and causing invalid data in /proc/stat and elsewhere.

---

 arch/x86_64/kernel/smpboot.c |   41 ++++++++++++++++-------------------------
 1 files changed, 16 insertions(+), 25 deletions(-)

12027fcffd85447f0fbf28264c5ee072715c345b

@@ -80,7 +80,23 @@ EXPORT_SYMBOL(cpu_online_map);
 cpumask_t cpu_callin_map;
 cpumask_t cpu_callout_map;
 
+#ifdef CONFIG_HOTPLUG_CPU
+/*
+ * cpu_possible_map should be static, it cannot change as cpu's
+ * are onlined, or offlined. The reason is per-cpu data-structures
+ * are allocated by some modules at init time, and dont expect to
+ * do this dynamically on cpu arrival/departure.
+ * cpu_present_map on the other hand can change dynamically.
+ * In case when cpu_hotplug is not compiled, then we resort to current
+ * behaviour, which is cpu_possible == cpu_present.
+ * If cpu-hotplug is supported, then we need to preallocate for all
+ * those NR_CPUS, hence cpu_possible_map represents entire NR_CPUS range.
+ * - Ashok Raj
+ */
+cpumask_t cpu_possible_map = CPU_MASK_ALL;
+#else
 cpumask_t cpu_possible_map;
+#endif
 EXPORT_SYMBOL(cpu_possible_map);
 
 /* Per CPU bogomips and other parameters */

@@ -879,27 +895,6 @@ static __init void disable_smp(void)
 	cpu_set(0, cpu_core_map[0]);
 }
 
-#ifdef CONFIG_HOTPLUG_CPU
-/*
- * cpu_possible_map should be static, it cannot change as cpu's
- * are onlined, or offlined. The reason is per-cpu data-structures
- * are allocated by some modules at init time, and dont expect to
- * do this dynamically on cpu arrival/departure.
- * cpu_present_map on the other hand can change dynamically.
- * In case when cpu_hotplug is not compiled, then we resort to current
- * behaviour, which is cpu_possible == cpu_present.
- * If cpu-hotplug is supported, then we need to preallocate for all
- * those NR_CPUS, hence cpu_possible_map represents entire NR_CPUS range.
- * - Ashok Raj
- */
-static void prefill_possible_map(void)
-{
-	int i;
-	for (i = 0; i < NR_CPUS; i++)
-		cpu_set(i, cpu_possible_map);
-}
-#endif
-
 /*
  * Various sanity checks.
  */

@@ -967,10 +962,6 @@ void __init smp_prepare_cpus(unsigned in
 	current_cpu_data = boot_cpu_data;
 	current_thread_info()->cpu = 0;  /* needed? */
 
-#ifdef CONFIG_HOTPLUG_CPU
-	prefill_possible_map();
-#endif
-
 	if (smp_sanity_check(max_cpus) < 0) {
 		printk(KERN_INFO "SMP disabled\n");
 		disable_smp();

A similar patch is already queued with Linus.

I also have a followon patch to avoid the extreme memory wastage currently caused by hotplug CPUs (e.g. with NR_CPUS==128 you currently lose 4MB of memory just for preallocated per CPU data). But that is something for post 2.6.14.

See ftp://ftp.firstfloor.org/pub/ak/x86_64/quilt-current/patches/* for the current queue.

-Andi

Hi,

Im interested in doing that on ppc64 too. Are you currently only creating per cpu data areas for possible cpus? The generic code does NR_CPUS worth, we should change that in 2.6.15.

Anton

Yes. In fact that caused some of the problems that lead to the investigation of this.

You need to audit all architecture code then that fills up possible map. (at least for the architectures that support CPU hotplug)

With that change x86-64 could move back to the generic code BTW - the numa allocation in the architecture specific code doesn't work anymore because we usually find out about the cpu<->node mapping now only afterwards :/

-Andi