ext4: Copied preprocessed code when it shouldn't

[marcosps]

When trying to create a livepatch for function ext4_reconfigure, file fs/ext4/super.c, on upstream, it generates the following code:

__attribute__((gnu_inline)) __attribute__((unused)) __attribute__((no_instrument_function)) static inline bool ext4_has_feature_metadata_csum(struct super_block *sb) {                                                                                                   
    return ((EXT4_SB(sb)->s_es->s_feature_ro_compat & ((__le32)(__u32)(1024))) != 0);                                                
}  

The original code, a macro in this case, has the following code on file fs/ext4/ext4.h:

...
static inline bool ext4_has_feature_##name(struct super_block *sb) \                                                                 
{ \                                                                                                                                  
        return ((EXT4_SB(sb)->s_es->s_feature_compat & \                                                                             
                cpu_to_le32(EXT4_FEATURE_COMPAT_##flagname)) != 0); \                                                                
} \     
...

For this livepatch specifically it should be good to import the code as it was written, instead of copying the preprocessed one. When a livepatch is generated for multiple architectures, resolving cpu_to_le32 is problematic, since it changes according the machine architecture it's being generated on.

[giulianobelinassi]

Just a note:

It is not safe to extract code from x86_64 architecture to run on PPC, as the define paths would differ. The correct way of doing this would be to extract from both arch and compare the differences. If there are none, then it is safe to use on both, else you may need to merge the things together.

[marcosps]

I agree. As we are using clang-extract as an experiment for kernel livepatches, maybe it would be nice to add some info about possible differences on the clang-extract README.md. Would it be safe to target PPC for example when generating code using clang-extract using cross compilation on a x86_64 machine? This would nice to have in the README as well.

[giulianobelinassi]

Cross-compilation from x86_64 to another arch is safe.

In any case, yes. The readme should have instructions about that.