On Wed, Mar 12, 2025 at 01:34:02PM +1100, Dave Chinner wrote:

Wrong.

Speculative non-blocking IO like NOWAIT is the wrong optimisation to
make for workloads that are very likely to block in the IO path. It
just adds overhead without adding any improvement in performance.

Note that I suspect that most or at least many loop workloads are
read-heavy.  And at least for reads NOWAIT makes perfect sense.

Getting rid of the serialised IO submission problems that the loop
device current has will benefit *all* workloads that use the loop
device, not just those that are fully allocated. Yes, it won't quite
show the same performance as NOWAIT in that case, but it still
should give 90-95% of native performance for the static file case.
And it should also improve all the other cases, too, because now
they will only serialise when the backing file needs IO operations to
serialise (i.e. during allocation).

And I agree that this should be a first step.

*cough*

The whole ublk-zoned is a bit of a bullshit thing where Ming wrote
up something that barely works to block inclusion of the zloop driver
we really need for zoned xfs testing.  Please don't take it serious.