2024-01-22 08:32:48

by Jingbo Xu

[permalink] [raw]
Subject: [MAYBE REGRESSION] kbuild time of kernel compiling

Hi,

I noticed a regression of kbuild time in v6.7, especially when running
`make` command when all images have already been compiled. In v6.6 it
takes ~3s, while in v6.7 it takes ~18s.

I'm not sure if it's a known issue, or an extra action configurable with
a "CONFIG_XX" option.


Following is the kbuild time in v6.6 versus v7.7, with
arch/x86/configs/x86_64_defconfig used here:


v6.6
```
make clean
make olddefconfig

# first full compiling
$time make bzImage -j128 -s

real 1m9.896s
user 36m56.153s
sys 4m21.748s

# second time with image already compiled
$time make bzImage -j128 -s

real 0m2.776s
user 0m13.823s
sys 0m3.936s
```


v6.7
```
make clean
make olddefconfig

# first full compiling
$time make bzImage -j128 -s

real 1m22.865s
user 37m25.977s
sys 4m22.094s

# second time with image already compiled
$time make bzImage -j128 -s

real 0m18.209s
user 0m29.243s
sys 0m4.330s
```


I tried to bisect, while commit d4e175f2c460 ("Merge tag 'vfs-6.7.super'
of gitolite.kernel.org:pub/scm/linux/kernel/git/vfs/vfs") is identified
as the first bad commit, which is obviously wrong. As I'm not familiar
with the structure of the merge commit, the further investigation has
not been made yet.


Besides, it seems that it will take most of the time before "CALL
scripts/checksyscalls.sh" is printed (with "INSTALL libsubcmd_headers"
has already printed).

$time make bzImage -j128
DESCEND objtool
INSTALL libsubcmd_headers

CALL scripts/checksyscalls.sh
BUILD arch/x86/boot/bzImage
Kernel: arch/x86/boot/bzImage is ready (#437)

--
Thanks,
Jingbo


2024-01-22 11:27:16

by Thorsten Leemhuis

[permalink] [raw]
Subject: Re: [MAYBE REGRESSION] kbuild time of kernel compiling

On 22.01.24 09:32, Jingbo Xu wrote:
>
> I noticed a regression of kbuild time in v6.7, especially when running
> `make` command when all images have already been compiled. In v6.6 it
> takes ~3s, while in v6.7 it takes ~18s.
>
> I'm not sure if it's a known issue, or an extra action configurable with
> a "CONFIG_XX" option.
> [...]

Not my area of expertise, so take this with a gain of salt, as I might
mislead you. Nevertheless:

One problem with compile times in 6.7 was noticed and fixed in mainline;
I wonder if that is the same issue as yours:

https://lore.kernel.org/lkml/[email protected]/
https://git.kernel.org/torvalds/c/31e97d7c9ae3de072d7b424b2cf706a03ec10720

I think there was another issue, but can't remember it from the top of
my head. Maybe simply checking if 6.8-rc1 fixes things might be the best.

Ciao, Thorsten (wearing his 'the Linux kernel's regression tracker' hat)
--
Everything you wanna know about Linux kernel regression tracking:
https://linux-regtracking.leemhuis.info/about/#tldr
If I did something stupid, please tell me, as explained on that page.

2024-01-22 13:00:31

by Masahiro Yamada

[permalink] [raw]
Subject: Re: [MAYBE REGRESSION] kbuild time of kernel compiling

On Mon, Jan 22, 2024 at 5:32 PM Jingbo Xu <[email protected]> wrote:
>
> Hi,
>
> I noticed a regression of kbuild time in v6.7, especially when running
> `make` command when all images have already been compiled. In v6.6 it
> takes ~3s, while in v6.7 it takes ~18s.
>
> I'm not sure if it's a known issue, or an extra action configurable with
> a "CONFIG_XX" option.
>
>
> Following is the kbuild time in v6.6 versus v7.7, with
> arch/x86/configs/x86_64_defconfig used here:
>
>
> v6.6
> ```
> make clean
> make olddefconfig
>
> # first full compiling
> $time make bzImage -j128 -s
>
> real 1m9.896s
> user 36m56.153s
> sys 4m21.748s
>
> # second time with image already compiled
> $time make bzImage -j128 -s
>
> real 0m2.776s
> user 0m13.823s
> sys 0m3.936s
> ```
>
>
> v6.7
> ```
> make clean
> make olddefconfig
>
> # first full compiling
> $time make bzImage -j128 -s
>
> real 1m22.865s
> user 37m25.977s
> sys 4m22.094s
>
> # second time with image already compiled
> $time make bzImage -j128 -s
>
> real 0m18.209s
> user 0m29.243s
> sys 0m4.330s
> ```
>
>
> I tried to bisect, while commit d4e175f2c460 ("Merge tag 'vfs-6.7.super'
> of gitolite.kernel.org:pub/scm/linux/kernel/git/vfs/vfs") is identified
> as the first bad commit, which is obviously wrong. As I'm not familiar
> with the structure of the merge commit, the further investigation has
> not been made yet.




If you use olddefconfig, git-bisect will not produce the correct result.

'make olddefconfig' updates the existing .config.
The resulting .config depends on what you had
before running 'make olddefconfig'.

You need to use a deterministic target (e.g. defconfig)
during the git-bisect.





>
>
> Besides, it seems that it will take most of the time before "CALL
> scripts/checksyscalls.sh" is printed (with "INSTALL libsubcmd_headers"
> has already printed).
>
> $time make bzImage -j128
> DESCEND objtool
> INSTALL libsubcmd_headers
>
> CALL scripts/checksyscalls.sh
> BUILD arch/x86/boot/bzImage
> Kernel: arch/x86/boot/bzImage is ready (#437)
>
> --
> Thanks,
> Jingbo
>


--
Best Regards
Masahiro Yamada

2024-01-22 14:07:36

by Jingbo Xu

[permalink] [raw]
Subject: Re: [MAYBE REGRESSION] kbuild time of kernel compiling



On 1/22/24 6:50 PM, Linux regression tracking (Thorsten Leemhuis) wrote:
> On 22.01.24 09:32, Jingbo Xu wrote:
>>
>> I noticed a regression of kbuild time in v6.7, especially when running
>> `make` command when all images have already been compiled. In v6.6 it
>> takes ~3s, while in v6.7 it takes ~18s.
>>
>> I'm not sure if it's a known issue, or an extra action configurable with
>> a "CONFIG_XX" option.
>> [...]
>
> Not my area of expertise, so take this with a gain of salt, as I might
> mislead you. Nevertheless:
>
> One problem with compile times in 6.7 was noticed and fixed in mainline;
> I wonder if that is the same issue as yours:
>
> https://lore.kernel.org/lkml/[email protected]/
> https://git.kernel.org/torvalds/c/31e97d7c9ae3de072d7b424b2cf706a03ec10720
>
> I think there was another issue, but can't remember it from the top of
> my head. Maybe simply checking if 6.8-rc1 fixes things might be the best.

Yeah this issue doesn't reproduce anymore after pulling 6.8-rc1. I
didn't test the latest 6.8-rc1 as our internal mirror repo is one day
later than the upstream, while my machine can not connect to github at
the time...

Anyway thanks for the feedback and sorry for the noise.


>
> Ciao, Thorsten (wearing his 'the Linux kernel's regression tracker' hat)
> --
> Everything you wanna know about Linux kernel regression tracking:
> https://linux-regtracking.leemhuis.info/about/#tldr
> If I did something stupid, please tell me, as explained on that page.

--
Thanks,
Jingbo