Re: [PATCH 11/16] time: optimize tick_check_preferred()

[Rasmus Villemoes]

On 08/08/2022 18.38, Yury Norov wrote:
> On Mon, Aug 08, 2022 at 01:42:54PM +0200, Thomas Gleixner wrote:
>> On Sat, Aug 06 2022 at 10:30, Thomas Gleixner wrote:
>>> On Mon, Jul 18 2022 at 12:28, Yury Norov wrote:
>>>
>>>> tick_check_preferred() calls cpumask_equal() even if
>>>> curdev->cpumask == newdev->cpumask. Fix it.
>>>
>>> What's to fix here? It's a pointless operation in a slow path and all
>>> your "fix' is doing is to make the code larger.
> 
> Pointless operation in a slow path is still pointless.
>  
>> In fact cpumask_equal() should have the ptr1 == ptr2 check, so you don't
>> have to add it all over the place.
> 
> This adds to the image size:
> add/remove: 1/1 grow/shrink: 24/3 up/down: 507/-46 (461)
> 
> The more important, cpumask shouldn't check parameters because this is
> an internal function. This whole series point is about adding such checks
> under DEBUG_BITMAP config, and not affecting general case.

Yury, calling bitmap_equal (and by extension cpumask_equal) with
something that happens in some cases to be the same pointer for both
operands is not a bug.

If you want to optimize that case, add a check in __bitmap_equal(), it
will add a few bytes (maybe 2-4 instructions) to the kernel image, much
less than what this whole series does by open-coding that check all
over, and since it's comparing two registers, it won't in any way affect
the performance of the case where the pointers are distinct.

Rasmus