On Mon, Jun 14, 2021 at 7:39 AM kernel test robot [off-list ref] wrote:

FYI, we noticed a -29.4% regression of vm-scalability.throughput due to commit:
ecf8443e51a8 ("[PATCH v7 02/11] mm/mremap: Fix race between MOVE_PUD mremap and pageout")

Ouch.

I guess it's not a huge surprise, but that's a fairly large regression.

Probably because the pud lock is just one single lock ("No scalability
reason to split PUD locks yet").

What happens if pud_lockptr() were to do the same thing that pmd_lockptr() does?

                Linus