Re: [PATCH v2] scripts/setlocalversion: also consider annotated tags of the form vx.y.z-${file_localversion}

[Rasmus Villemoes]

On 08/09/2023 20.19, Sean Christopherson wrote:
> On Fri, Aug 04, 2023, Rasmus Villemoes wrote:
>> Commit 6ab7e1f95e96 ("setlocalversion: use only the correct release
>> tag for git-describe") was absolutely correct to limit which annotated
>> tags would be used to compute the -01234-gabcdef suffix. Otherwise, if
>> some random annotated tag exists closer to HEAD than the vX.Y.Z one,
>> the commit count would be too low.
>>
>> However, since the version string always includes the
>> ${file_localversion} part, now the problem is that the count can be
>> too high. For example, building an 6.4.6-rt8 kernel with a few patches
>> on top, I currently get
>>
>> $ make -s kernelrelease
>> 6.4.6-rt8-00128-gd78b7f406397
>>
>> But those 128 commits include the 100 commits that are in
>> v6.4.6..v6.4.6-rt8, so this is somewhat misleading.
>>
>> Amend the logic so that, in addition to the linux-next consideration,
>> the script also looks for a tag corresponding to the 6.4.6-rt8 part of
>> what will become the `uname -r` string. With this patch (so 29 patches
>> on top of v6.4.6-rt8), one instead gets
>>
>> $ make -s kernelrelease
>> 6.4.6-rt8-00029-gd533209291a2
>>
>> While there, note that the line
>>
>>   git describe --exact-match --match=$tag $tag 2>/dev/null
>>
>> obviously asks if $tag is an annotated tag, but it does not actually
>> tell if the commit pointed to has any relation to HEAD. So remove both
>> uses of --exact-match, and instead just ask if the description
>> generated is identical to the tag we provided. Since we then already
>> have the result of
>>
>>   git describe --match=$tag
>>
>> we also end up reducing the number of times we invoke "git describe".
> 
> Dropping "--exact-match" is resulting in unnacceptable latencies for me.  I don't
> understand what this is trying to do well enough to make a suggestion, but something
> has to change.

Hm, that's quite unexpected. I mean, before that commit, I think that
setlocalversion, especially when run on some dev branch, would _also_
end up doing at least one 'git describe --match=v6.5'. <goes digging>

Ah, so I assume that in your case you always end up in the

                # If only the short version is requested, don't bother
                # running further git commands
                if $short; then
                        echo "+"
                        return
                fi

case, i.e. you do not have CONFIG_LOCALVERSION_AUTO=y. Can you confirm that?

> E.g. on my build box, a single `git describe --match=v6.5` takes ~8.5 seconds,

That's a long time indeed. I don't really know why it can take that long.

> whereas a complete from-scratch kernel build takes <30 seconds, and an incremental
> build takes <2 seconds.  When build testing to-be-applied changes, I compile each
> commit ~15 times (different x86 configs plus one for each other KVM architecture),
> which makes the ~8.5 second delay beyond painful.

Sorry about that. I agree something needs to be done, but the commit
above fixed a real problem with -rt kernels, and the refactoring just
made it much easier to follow the logic IMO - and, for the
CONFIG_LOCALVERSION_AUTO=y case, did reduce the number of times git
describe gets invoked.

Until I have a bit more time to think, could you try setting the env var
LOCALVERSION when building - it can be set to anything, including empty.
Then if I'm reading the script right, the scm_version() function won't
be called at all.

Rasmus