Hi Dave,
* Dave Voutila wrote:
> While working back and forth with dlg@ on his idea of adding in async
> task queues to vmd devices, he started to introduce a cleaner way to
> handle virtqueues. In short, we can just track the host virtual address
> and read/write to that location.
>
> This has the benefit of cleaning up a good chunk of code, mostly in the
> vioscsi device. If my count is correct, it's close to 100 lines
> removed.
>
> Testing on a variety of systems, I see anywhere from a small (5%) to a
> large (50%) gain in things like vionet performance depending on the host
> hardware and the guest OS. (Linux guests push packets faster, on
> average.)
>
> If/when this lands, it'll be easier to introduce the async taskq diff
> and also clean up more of the virtio codebase. Lots of duplicate logic
> in basic virtio-pci config handling, like updating queue address or
> size.
>
> Testing is appreciated, but also looking for OKs.
I have your patch running on -current with OpenBSD -current, -stable and
Linux (voidlinux) guests. Stable so far, no regressions and Linux "seems"
faster.
Cheers
Matthias
