Re: [PATCH] x86: Optimize load_mm_cr4 to load_mm_cr4_irqsoff

[Jan Kiszka]

On 18.06.19 16:41, Dave Hansen wrote:
> On 6/18/19 12:32 AM, Jan Kiszka wrote:
> > Thus, we can avoid disabling interrupts again in cr4_set/clear_bits.
>
> Seems reasonable.
>
> Your *_irqsoff() variants need lockdep_assert_irqs_disabled(), at least,
> though.

It inherits this check via __cr4_set, so I didn't add that to the outer path.

> Can you talk a bit about the motivation here, though?  Did you encounter
> some performance issue that led you to make this patch, or was it simply
> an improvement you realized you could make from code inspection?

I ran into a downstream issue adjusting patches to this code. In a nutshell, 
Xenomai has something like an NMI context over most Linux code that shares some 
codepaths with the kernel, though. One of them is switch_mm_irqs_off, and there 
our checking logic warned. The rest was code inspection.

For upstream, this is a micro-optimization. But given that something like 
switch_mm is in the path, I thought it's worth to propose.

Jan

--
Siemens AG, Corporate Technology, CT RDA IOT SES-DE
Corporate Competence Center Embedded Linux