Re: [PATCH 0/2] arm64: Allow the rescheduling IPI to bypass irq_enter/exit

[Marc Zyngier]

On Fri, 20 Nov 2020 14:17:12 +0000,
Thomas Gleixner <tglx@xxxxxxxxxxxxx> wrote:

Thomas,

> So with the pre 5.8 scheduler IPI we had scheduler_ipi() doing wonderful
> things

[... tons of interesting and helpful stuff ...]

> Hope that clarifies it.

It does in a way, as it maps some of the low-level x86 things onto
generic concepts. It certainly outlines that we have a lot of work
ahead of us, none of which can be expected to be a quick fix. It
requires restructuring a lot of the low-level arm64 code (Mark has
been toying with it for a couple of years now), and rejig it in funny
ways to match the expectations of the core code.

I haven't tried to think about 32bit ARM just yet, and we may have to
sever the IRQ ties that exist between the two architectures if we want
to make forward progress in a reasonable time frame. In general, it
looks that we'll need a new interface from the generic low-level IRQ
entry code into this new common framework.

> > If arm64 has forever been broken, I'd really like to know and fix it.
> 
> Define broken. It kinda works with all the warts and bolts in tracing,
> context tracking and other places. Is it entirely correct? Probably
> not.
>
> The scheduler IPI is trivial compared to the more interesting ones like
> NMI, MCE, breakpoint, debug exceptions etc. We found quite some
> interesting issues with all of that muck during our 'noinstr' chase.

Yup, point taken. However, given the complexity of the above and the
fact that we currently have a measurable performance regression in
5.10, I'll respin the original series with the cleanups you mentioned,
and the added note that we *really* need to sort this.

Thanks,

	M.

-- 
Without deviation from the norm, progress is not possible.