Re: Crashes in refresh_zone_stat_thresholds when some nodes have no memory

quoted

[    1.717648] Call Trace:
[    1.717687] [c000000ff0707b80] [c000000000270d08] refresh_zone_stat_thresholds+0xb8/0x240 (unreliable)
[    1.717818] [c000000ff0707bd0] [c000000000a1e4d4] init_per_zone_wmark_min+0x94/0xb0
[    1.717932] [c000000ff0707c30] [c00000000000b90c] do_one_initcall+0x6c/0x1d0
[    1.718036] [c000000ff0707cf0] [c000000000d04244] kernel_init_freeable+0x294/0x384
[    1.718150] [c000000ff0707dc0] [c00000000000c1a8] kernel_init+0x28/0x160
[    1.718249] [c000000ff0707e30] [c000000000009968] ret_from_kernel_thread+0x5c/0x74
[    1.718358] Instruction dump:
[    1.718408] 3fc20003 3bde4e34 3b800000 60420000 3860ffff 3fbb0001 4800001c 60420000 
[    1.718575] 3d220003 3929f8e0 7d49502a e93d9c00 <7f8a49ae> 38a30001 38800800 7ca507b4 

It turns out that we can get a pgdat in the online pgdat list where
pgdat->per_cpu_nodestats is NULL.  On my machine the pgdats for nodes
1 and 17 are like this.  All the memory is in nodes 0 and 16.

How is this possible? setup_per_cpu_pageset does

	for_each_online_pgdat(pgdat)
		pgdat->per_cpu_nodestats =
			alloc_percpu(struct per_cpu_nodestat);

so each online node should have the per_cpu_nodestat allocated.
refresh_zone_stat_thresholds then does for_each_online_pgdat and
for_each_populated_zone also shouldn't give any offline pgdat. Is it
possible that this is yet another manifest of 6aa303defb74 ("mm, vmscan:
only allocate and reclaim from zones with pages managed by the buddy
allocator")? I guess the following should be sufficient?

diff --git a/mm/vmstat.c b/mm/vmstat.c
index 73aab319969d..c170932a0101 100644
--- a/mm/vmstat.c
+++ b/mm/vmstat.c

@@ -185,6 +185,9 @@ void refresh_zone_stat_thresholds(void)
 		struct pglist_data *pgdat = zone->zone_pgdat;
 		unsigned long max_drift, tolerate_drift;
 
+		if (!managed_zone(zone))
+			continue;
+
 		threshold = calculate_normal_threshold(zone);
 
 		for_each_online_cpu(cpu) {