Re: [RFC PATCH 8/8] irq: boost/unboost in irq/nmi entry/exit and softirq

From: Vineeth Remanan Pillai
Date: Fri Dec 15 2023 - 13:53:01 EST


[...snip..]
> > +#ifdef CONFIG_PARAVIRT_SCHED
> > + if (pv_sched_enabled() && !in_hardirq() && !local_softirq_pending() &&
> > + !need_resched() && !task_is_realtime(current))
> > + pv_sched_unboost_vcpu();
> > +#endif
>
> Symmetry is overrated. Just pick a spot and slap your hackery in.
>
> Aside of that this whole #ifdeffery is tasteless at best.
>
> > instrumentation_end();
>
> > +#ifdef CONFIG_PARAVIRT_SCHED
> > + if (pv_sched_enabled())
> > + pv_sched_boost_vcpu_lazy();
> > +#endif
>
> But what's worse is that this is just another approach of sprinkling
> random hints all over the place and see what sticks.
>
Understood. WIll have a full re-write of guest side logic for v2.

> Abusing KVM as the man in the middle to communicate between guest and
> host scheduler is creative, but ill defined.
>
> From the host scheduler POV the vCPU is just a user space thread and
> making the guest special is just the wrong approach.
>
> The kernel has no proper general interface to convey information from a
> user space task to the scheduler.
>
> So instead of adding some ad-hoc KVM hackery the right thing is to solve
> this problem from ground up in a generic way and KVM can just utilize
> that instead of having the special snow-flake hackery which is just a
> maintainability nightmare.
>
We had a constructive and productive discussion on the cover letter
thread and reaching a consensus on the kvm side of the design. Will
work towards that and post iterative revisions.

Thanks,
Vineeth