Re: [PATCH v2 1/7] vhost: Add a new modparam to allow userspace select vhost_task

[Jason Wang]

On Wed, Oct 9, 2024 at 4:10 PM Stefano Garzarella <sgarzare@xxxxxxxxxx> wrote:
>
> On Wed, Oct 09, 2024 at 03:28:19PM GMT, Jason Wang wrote:
> >On Mon, Oct 7, 2024 at 9:31 PM Stefano Garzarella <sgarzare@xxxxxxxxxx> wrote:
> >>
> >> On Fri, Oct 04, 2024 at 09:58:15AM GMT, Cindy Lu wrote:
> >> >The vhost is now using vhost_task and working as a child of the owner thread.
> >> >While this makes sense from containerization POV, some old userspace is
> >> >confused, as previously vhost not
> >>
> >> not what?
> >>
> >> > and so was allowed to steal cpu resources
> >> >from outside the container. So we add the kthread API support back
> >>
> >> Sorry, but it's not clear the reason.
> >>
> >> I understand that we want to provide a way to bring back the previous
> >> behavior when we had kthreads, but why do we want that?
> >> Do you have examples where the new mechanism is causing problems?
> >>
> >
> >The main difference is whose (kthreadd or the owner) attributes
> >(namespace, affinities) would vhost thread inherit.
> >
> >The owner's attributes tend to be different from kthreadd, so
> >management might be surprised for example, vhost might be created in
> >different namespaces or having different scheduling affinities.
>
> Okay, so this requires some API to allow the managment to understand how
> the device vhost will be created.
>
> But why do we need to restore the old behavior with a kthread where the
> resource accounting is completely disconnected from userspace?
> For the old managments that don't expect this?

Yes, as such change can be easily noticed by the user space that
breaks existing management layers or tools.

>
> BTW I would suggest adding all this information in this commit, in the
> parameter/IOCTL documentation, and in the cover letter because IMHO it
> is very important.

I've asked this in the past. Cindy, please make sure such information
were included in the next version in the cover letter.

Thanks

>
> Thanks,
> Stefano
>