Re: [PATCH v3 03/16] arm64: Allow IPIs to be handled as normal interrupts

[Marc Zyngier]

On Fri, 07 May 2021 08:30:06 +0100,
He Ying <heying24@xxxxxxxxxx> wrote:
> 
> 
> 在 2021/5/6 19:44, Marc Zyngier 写道:
> > On Thu, 06 May 2021 08:50:42 +0100,
> > He Ying <heying24@xxxxxxxxxx> wrote:
> >> Hello Marc,
> >> 
> >> We have faced a performance regression for handling ipis since this
> >> commit. I think it's the same issue reported by Vincent.
> > Can you share more details on what regression you have observed?
> > What's the workload, the system, the performance drop?
> 
> OK. We have just calculated the pmu cycles from the entry of gic_handle_irq
> to the entry of do_handle_ipi. Here is some more information about our test:
> 
> CPU: Hisilicon hip05-d02
> 
> Applying the patch series: 1115 cycles
> Reverting the patch series: 599 cycles

And? How is that meaningful? Interrupts are pretty rare compared to
everything that happens in the system. How does it affect the
behaviour of the system as a whole?

> 
> > 
> >> I found you pointed out the possible two causes:
> >> 
> >> (1) irq_enter/exit on the rescheduling IPI means we reschedule much
> >> more often.
> > It turned out to be a red herring. We don't reschedule more often, but
> > we instead suffer from the overhead of irq_enter()/irq_exit().
> > However, this only matters for silly benchmarks, and no real-life
> > workload showed any significant regression. Have you identified such
> > realistic workload?
> 
> I'm afraid not. We just run some benchmarks and calculated pmu cycle
> counters.  But we have observed running time from the entry of
> gic_handle_irq to the entry of do_handle_ipi almost doubles. Doesn't
> it affect realistic workload?

Then I'm not that interested. Show me an actual regression in a real
workload that affects people, and I'll be a bit more sympathetic to
your complain. But quoting raw numbers do not help.

There is a number of advantages to having IPI as IRQs, as it allows us
to deal with proper allocation (other subsystem want to use IPIs), and
eventually NMIs. There is a trade-off, and if that means wasting a few
cycles, so be it.

> >> (2) irq_domain lookups add some overhead.
> > While this is also a potential source of overhead, it turned out not
> > to be the case.
> OK.
> > 
> >> But I don't see any following patches in mainline. So, are you still
> >> working on this issue?  Looking forward to your reply.
> > See [1]. However, there is probably better things to do than this
> > low-level specialisation of IPIs, and Thomas outlined what needs to be
> > done (see v1 of the patch series).
> 
> OK. I see the patch series. Would it be applied to the mainline
> someday? I notice that more than 5 months have passed since you sent
> the patch series.

I have no plan to merge these patches any time soon, given that nobody
has shown a measurable regression using something other than a trivial
benchmark. If you come up with such an example, I will of course
reconsider this position.

Thanks,

	M.

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