Re: [PATCH] arm64: Make CONFIG_ARM64_PSEUDO_NMI macro wrap all the pseudo-NMI code

[Mark Rutland]

On Tue, Jan 11, 2022 at 04:52:32PM +0800, He Ying wrote:
> 在 2022/1/10 19:26, Mark Rutland 写道:
> > On Mon, Jan 10, 2022 at 11:00:43AM +0800, He Ying wrote:
> > > 在 2022/1/7 21:19, Mark Rutland 写道:
> > > > On Fri, Jan 07, 2022 at 03:55:36AM -0500, He Ying wrote:
 
> > Due to the large numbers, I suspect this must be due to a specific fast-path,
> > and it's possible that this is due to secondary factors (e.g. alignment of
> > code) rather than the pseudo-NMK code itself.
> > 
> > We need to narrow down *where* time is being spent. Since it appears that this
> > is related to the local IRQ state management, it isn't likely that we can
> > determine that reliably with the PMU. Given that, I think the first step is to
> > reproduce the result elsewhere, for which we will need some plublicly available
> > test-case.
> 
> As said before, I asked our collegues how they did the ARP test. In one word,
> a very small performance regression may bring the big change to the test
> result.
> 
> I feel very sorry for missing this important information. So, this patch may
> improve the performance a little and then lead to the big change to the
> test result.

No problem; thanks for confirming.

[...]

> OK, I see. What do you think if I send a v2 patch that only adds ifdeffery
> to
> 
> arch/arm64/kernel/entry.S and leave the rework to ALERNATIVE behind?

I think that would be reasonable given we do that for other bits in entry.S;
I'd be happy with such a patch.

Thanks,
Mark.