Re: [PATCH v2] kbuild: make all file references relative to source root

[Peter Oberparleiter <oberpar@linux.ibm.com>]

On 29.04.2025 18:12, Matthieu Baerts wrote:
> On 15/03/2025 14:20, Thomas Weißschuh wrote:
>> -fmacro-prefix-map only affects __FILE__ and __BASE_FILE__.
>> Other references, for example in debug information, are not affected.
>> This makes handling of file references in the compiler outputs harder to
>> use and creates problems for reproducible builds.
>>
>> Switch to -ffile-prefix map which affects all references.
>>
>> Also drop the documentation section advising manual specification of
>> -fdebug-prefix-map for reproducible builds, as it is not necessary
>> anymore.
>>
>> Suggested-by: Ben Hutchings <ben@decadent.org.uk>
>> Link: https://lore.kernel.org/lkml/c49cc967294f9a3a4a34f69b6a8727a6d3959ed8.camel@decadent.org.uk/
>> Acked-by: Borislav Petkov (AMD) <bp@alien8.de> # arch/x86/
>> Signed-off-by: Thomas Weißschuh <linux@weissschuh.net>
> 
> Thank you for having worked on that!
> 
> (...)
> 
>> diff --git a/Makefile b/Makefile
>> index 5c333682dc9142b1aacfe454a5c77f5923554b7d..4f920187cee658ae4d1b807fca365f6994274828 100644
>> --- a/Makefile
>> +++ b/Makefile
>> @@ -1067,7 +1067,7 @@ endif
>>  
>>  # change __FILE__ to the relative path to the source directory
>>  ifdef building_out_of_srctree
>> -KBUILD_CPPFLAGS += $(call cc-option,-fmacro-prefix-map=$(srcroot)/=)
>> +KBUILD_CPPFLAGS += $(call cc-option,-ffile-prefix-map=$(srcroot)/=)
>>  KBUILD_RUSTFLAGS += --remap-path-prefix=$(srcroot)/=
>>  endif
> 
> Today, I noticed that my CI for the MPTCP subsystem couldn't produce
> code coverage files like before: the source files are not found. A 'git
> bisect' pointed me to this patch. Reverting it seems to fix the issue.
> 
> 
> My CI is building the kernel out of the source tree, in ".virtme/build".
> Before and after this patch, GCOV seems to do its job properly.
> Capturing GCOV data with this lcov command seems OK too:
> 
>   lcov --capture --keep-going -j "${INPUT_CPUS}" \
>      --rc geninfo_unexecuted_blocks=1 \
>      --include '/net/mptcp/' \
>      --function-coverage --branch-coverage \
>      -b "${PWD}/.virtme/build" -o kernel.lcov
> 
> But after this patch, lcov complains some files are not found, e.g.
> 
>   ERROR: (source) unable to open
> ${WORKDIR}/.virtme/build/net/mptcp/ctrl.c: No such file or directory

I can confirm that the subject kernel commit breaks gcov-kernel for
out-of-srctree builds by making it impossible for a consumer of GCOV
data to determine the actual location of a source file without manually
specifying it.

Sample .gcno file content changes as seen with gcov-dump:

- cwd: /home/.../build
- /home/.../linux/kernel/workqueue.c:8049:19-8057:1
+ cwd: /home/.../build
+ kernel/workqueue.c:8049:19-8057:1

> The output file is different: the path to the source file is wrong
> because it points to the build dir. Instead of ...
> 
>   SF:${WORKDIR}/net/mptcp/ctrl.c
> 
> ... now I have ...
> 
>   SF:${WORKDIR}/.virtme/build/net/mptcp/ctrl.c
> 
> 
> Are there modifications needed on GCOV side to adapt to the behaviour
> change introduced by this patch? Or something else needed on the
> userspace side?

I don't see how this could be fixed by changes in userspace nor
gcov-kernel - the source tree directory information is missing from the
relevant data files.

-- 
Peter Oberparleiter
Linux on IBM Z Development - IBM Germany R&D