Re: [PATCH v3 3/3] correct ce_compare_data() in a middle of a merge

["Torsten Bögershausen" <tboegi@web.de>]

On 07/07/16 20:43, Junio C Hamano wrote:
> Torsten Bögershausen <tboegi@web.de> writes:
>
>> This is the callstack:
>>
>> merge-recursive.c/add_cacheinfo(unsigned int mode, const unsigned char *sha1,
>> 		const char *path, int stage, int refresh, int options)
>> {
>> 	struct cache_entry *ce;
>> 	ce = make_cache_entry
>> 	if (!ce)
>> 		return error(_("addinfo_cache failed for path '%s'"), path);
>> 	return add_cache_entry(ce, options);
>> }
>> #Calls
>> read-cache.c/make_cache_entry(path=file sha1=0x99b633 stage=0)
>>
>>
>> struct cache_entry *make_cache_entry(unsigned int mode,
>> 		const unsigned char *sha1, const char *path, int stage,
>> 		unsigned int refresh_options)
>> {
>>          [snip]
>> 	ret = refresh_cache_entry(ce, refresh_options);
>> 	if (ret != ce)
>> 		free(ce);
>> 	return ret;
>> }
>>
>> #Calls
>> refresh_cache_ent(&the_index, ce, options, NULL, NULL);
>>
>> #Calls
>> ie_modified()
>>
>> #Calls
>> read-cache.c/ie_match_stat()
>>
>> #Calls
>> changed |= ce_modified_check_fs(ce, st);
>>
>> #Calls
>> ce_compare_data(path=file sha1=0x99b633)
>>
>> #Calls
>> index_fd(..., ..., ce->name, ...)
>> #Note the sha is lost here, the parameter sha is the output.
>>
>> Deep down, has_cr_in_index(path) will look at ad55e2 instead,
>> which is wrong.
> The call to add_cacheinfo() is made with 99b633 to record the 3-way
> merge result to the index in this callchain:
>
>   merge_trees()
>   -> git_merge_trees()
>   -> process_renames() # does nothing for this case
>   -> process_entry()
>      -> merge_content()
>         -> merge_file_special_markers()
>            -> merge_file_1()
>               -> merge_3way()
>                  -> ll_merge() # correctly handles the renormalization
>               -> write_sha1_file() # this is what gives us 99b633
>      -> update_file() # this is fed the 99b633 write_sha1_file() computed
>         -> update_file_flags()
>            -> read_sha1_file() # reads 99b633
>            -> convert_to_working_tree()
>            -> write_in_full() # updates the working tree
>            -> add_cacheinfo() # to record 99b633 at "file" in the index
>
> So refresh() may incorrectly find that 99b633 does not match what is
> in the filesystem if it looks at _any_ entry in the index.  We know
> we have written the file ourselves, so by definition it should match
> ;-) Even though I am not sure why that should affect the overall
> correctness of the program (because we have written the correct
> result to both working tree and to the index), this needs to be
> fixed.
>
> I am however not convinced passing the full SHA-1 is a good
> solution.
It seems at least that this is the correct way to do it.
> The make_cache_entry() function may be creating a cache
> entry to stuff in a non-default index (i.e. not "the_index"), but
> its caller does not have a way to tell it which index the resulting
> cache entry will go, and calls refresh_cache_entry(), which always
> assumes that the caller is interested in "the_index", so what
> add_cacheinfo() does by calling make_cache_entry() feels wrong in
> the first place.  Also, the call from update_file_flags() knows that
> the working tree is in sync with the resulting cache entry.
>
> Perhaps update_file_flags() should not even call add_cacheinfo()
> there in the first place?  I wonder if it should just call
> add_file_to_index()---after all, even though the current code
> already knows that 99b633 _is_ the result it wants in the index, it
> still goes to the filesystem and rehashes.
>
> I dunno.  I really do not like that extra sha1 argument added all
> over the callchain by this patch.
>
> Or did you mean other calls to add_cacheinfo()?
I didn't mean too much - the whole call chain touches code where I am not able to
comment on details.
I'm happy to test other implementations, if someone suggests a path, so to say.