Re: [patch 5/9] x86: Cure per CPU madness on UP

[Arnd Bergmann]

On Mon, Mar 18, 2024, at 18:27, Thomas Gleixner wrote:
> On Sat, Mar 16 2024 at 02:11, Thomas Gleixner wrote:
>> On Fri, Mar 15 2024 at 16:23, Linus Torvalds wrote:
>> The amount of subtle SMP=n fallout has been kinda exponentially
>> increasing over the years and it's just putting burden on the wrong
>> people. TBH, I'm tired of this nonsense.
>
> And for the fun of it I hacked Kconfig to allow a SMP=y NR_CPUS=1 build
> and checked the size of vmlinux:
>
>                 64-bit          32-bit
> SMP, NCPUS=1    38438400        22110177
> UP              38393703        21682041
> Delta              44697          428076
>                      0.1%              2%              
>
> The UP savings are not really impressive...
>
> Let me look what it actually takes to do that.

FWIW, I did some experiments a few weeks ago on 32-bit ARM,
using a fairly minimal kernel in a virtual machine, and
checking the runtime memory consumption rather than compile-time.
In a kvm guest with 32MiB RAM, I saw a difference of multiple
megabytes in memory usage:

Linux testvm 6.8.0-rc4-00410-gc02197fc9076-dirty #1 SMP PREEMPT armv7l
root@testvm:~# free
           	total   used    free  	shared  buff/cache   available
Mem:       	26932   14956   1732   	    52       12800   	11976
Swap:      	16360    3632   12728

Linux testvm 6.8.0-rc4-00410-gc02197fc9076-dirty #2 PREEMPT armv7l
root@testvm:~# free
           	total    used  	free  	shared  buff/cache   available
Mem:       	26932   13744  	5648        32       10092   	13188
Swap:      	16360    3880  	12480

There is a little difference between runs, but this does seem
significant enough to keep it. The SMP build was with
CONFIG_NR_CPUS=2 (the smallest supported compile-time number),
but running on a single-CPU qemu instance.

      Arnd