Comment # 13 on bug 1013629 from Michal Hocko

(In reply to Howard Guo from comment #12)
> Created attachment 707817 [details]
> vmstat over period of 30 seconds
> 
> Merry Christmas and happy new year!
> 
> The fault did not happen until earlier today. Attached tgz contains capture
> of vmstat file over a period of 30 seconds.

are you sure that kswapd was hogging the CPU during that time period? I do not
see basically any kswapd activity during that 30s time period
                base        diff
pgscan_kswapd_normal        55          0
pgscan_kswapd_dma32         1316447     0
pgsteal_kswapd_normal       7           0
pgsteal_kswapd_dma32        1156969     0
slabs_scanned           8149055     4720
kswapd_inodesteal           1226481     4517
kswapd_high_wmark_hit_quickly   393623726       53777

So we have reclaimed some inodes, didn't scan a single LRU page.
kswapd_high_wmark_hit_quickly is quite high so the kswapd was woken up quite
often but didn't do anything really. The number of LRU pages is more or less
stable.

One reason for that might be a source of high order allocations. THP come to
mind but there is activity there
                base    diff
thp_collapse_alloc_failed       2165    1

maybe there is other source of high order requests. All in all I really do
not see anything that would cause kswapd to hog the CPU.