Re: kern/55268: tmpfs is slow

[Andrew Doran <ad%netbsd.org@localhost>]

It's a little better now but there is still a major problem with lock
contention and context switch storms during fault processing because of a
mix of reader/writer holds on the uobject locks.

http://www.netbsd.org/~ad/2020/tmpfs-sucks.svg

I guess it's something to do with tmpfs pages always being marked
UVM_PAGE_STATUS_UNKNOWN unless committed to swap and that affecting the
fault type.

Andrew