Re: Sub-millisecond [autovacuum_]vacuum_cost_delay broken

On Fri, Mar 10, 2023 at 2:45 PM Thomas Munro <thomas.munro@gmail.com> wrote:
> I'll go and see about usec latch waits.  More soon.

Here are some experimental patches along those lines.  Seems good
locally, but I saw a random failure I don't understand on CI so
apparently I need to find a bug; at least this gives an idea of how
this might look.  Unfortunately, the new interface on Linux turned out
to be newer that I first realised: Linux 5.11+ (so RHEL 9, Debian
12/Bookworm, Ubuntu 21.04/Hirsute Hippo), so unless we're OK with it
taking a couple more years to be more widely used, we'll need some
fallback code.  Perhaps something like 0004, which also shows the sort
of thing that we might consider back-patching to 14 and 15 (next
revision I'll move that up the front and put it in back-patchable
form).  It's not exactly beautiful; maybe sharing code with recovery's
lazy PM-exit detection could help.  Of course, the new μs-based wait
API could be used wherever we do timestamp-based waiting, for no
particular reason other than that it is the resolution of our
timestamps, so there is no need to bother rounding; I doubt anyone
would notice or care much about that, but it's vote in favour of μs
rather than the other obvious contender ns, which modern underlying
kernel primitives are using.