Comment 64 for bug 1674838

So, I think the issue occurs because of the pte_lockptr lock being used on a PMD and is incontention with the lock on the same PMD. There are several possible points where this can happen, for example, pages being migrated on NUMA systems (unlikely in most of these bug reports) or even mremap() being used remap a mapped region to a new location. It may be because a page fault occurs on a page that is being migrated or being remapped while it is being accessed and it's swapped out; the latter case may introduce a lot of latency if there are lots of pending I/Os during the swap.