Re: Possible build time regression affecting stable kernels

[Luiz Capitulino <luizcap@amazon.com>]

On 2023-06-01 02:06, Greg KH wrote:

> 
> 
> 
> On Wed, May 31, 2023 at 10:12:40PM -0400, Luiz Capitulino wrote:
>> Hi Paul,
>>
>> A number of stable kernels recently backported this upstream commit:
>>
>> """
>> commit 4ce1f694eb5d8ca607fed8542d32a33b4f1217a5
>> Author: Paul Moore <paul@paul-moore.com>
>> Date:   Wed Apr 12 13:29:11 2023 -0400
>>
>>      selinux: ensure av_permissions.h is built when needed
>> """
>>
>> We're seeing a build issue with this commit where the "crash" tool will fail
>> to start, it complains that the vmlinux image and /proc/version don't match.
>>
>> A minimum reproducer would be having "make" version before 4.3 and building
>> the kernel with:
>>
>> $ make bzImages
>> $ make modules
>>
>> Then compare the version strings in the bzImage and vmlinux images,
>> we can use "strings" for this. For example, in the 5.10.181 kernel I get:
>>
>> $ strings vmlinux | egrep '^Linux version'
>> Linux version 5.10.181 (ec2-user@ip-172-31-79-134.ec2.internal) (gcc (GCC) 7.3.1 20180712 (Red Hat 7.3.1-15), GNU ld version 2.29.1-31.amzn2) #2 SMP Thu Jun 1 01:26:38 UTC 2023
>>
>> $ strings ./arch/x86_64/boot/bzImage | egrep 'ld version'
>> 5.10.181 (ec2-user@ip-172-31-79-134.ec2.internal) (gcc (GCC) 7.3.1 20180712 (Red Hat 7.3.1-15), GNU ld version 2.29.1-31.amzn2) #1 SMP Thu Jun 1 01:23:59 UTC 2023
>>
>> The version string in the bzImage doesn't have the "Linux version" part, but
>> I think this is added by the kernel when printing. If you compare the strings,
>> you'll see that they have a different build date and the "#1" and "#2" are
>> different.
>>
>> This only happens with commit 4ce1f694eb5 applied and older "make", in my case I
>> have "make" version 3.82.
>>
>> If I revert 4ce1f694eb5 or use "make" version 4.3 I get identical strings (except
>> for the "Linux version" part):
>>
>> $ strings vmlinux | egrep '^Linux version'
>> Linux version 5.10.181+ (ec2-user@ip-172-31-79-134.ec2.internal) (gcc (GCC) 7.3.1 20180712 (Red Hat 7.3.1-15), GNU ld version 2.29.1-31.amzn2) #1 SMP Thu Jun 1 01:29:11 UTC 2023
>>
>> $ strings ./arch/x86_64/boot/bzImage | egrep 'ld version'
>> 5.10.181+ (ec2-user@ip-172-31-79-134.ec2.internal) (gcc (GCC) 7.3.1 20180712 (Red Hat 7.3.1-15), GNU ld version 2.29.1-31.amzn2) #1 SMP Thu Jun 1 01:29:11 UTC 2023
>>
>> Maybe the grouped target usage in 4ce1f694eb5 with older "make" is causing a
>> rebuild of the vmlinux image in "make modules"? If yes, is this expected?
>>
>> I'm afraid this issue could be high impact for distros with older user-space.
> 
> Is this issue also in 6.4-rc1 where this change came from?  

Yes. I'm reporting this here because I'm more concerned with -stable kernels since
they're more likely to be running on older user-space.

> What about
> the other stable releases?

I tested against: 6.1.31, 5.15.114, 5.10.181 they all reproduce it. I'd guess any
kernel having a backport of 4ce1f694eb5d8ca607fed8542d32a33b4f1217a5 will have this.

One detail though, I tested once with "make binrpm-pkg" and didn't reproduce it. I
think this could mean that it depends on the make targets you use to build the
kernel. But "make bzImage", "make modules", "make modules_install" is pretty
standard (or used to be).

- Luiz