Shourya Shukla [off-list ref] writes:
+ if (git_config_rename_section_in_file(index_only ? GITMODULES_INDEX :
+ GITMODULES_FILE, sect.buf, NULL) < 0) {
Also, is it really sufficient to pass GITMODULES_INDEX as the first
argument to this function to tweak what is in the index?
git_config_copy_or_rename_section_in_file() which is the
implementation of that helper seems to always want to work with a
file that is on disk, by making unconditional calls to
hold_lock_file_for_update(), fopen(), fstat(), chmod(), etc.
So I suspect that there are much more work needed.
It seems to me that the config editing API is one of the older and
hackier parts of the system and requires quite a lot of work to
teach it to work with anything but a on-disk file. In the longer
term, it may be a good thing to clean it up, but I suspect that it
is way too much work for too little benefit to do so as a part of
this topic, so an easier way out for now would be to:
- write out the .gitmodules in the index to a temporary file (learn
how to correctly call entry.c::checkout_entry() by studying how
builtin/checkout-index.c::checkout_file() calls it, especially to
a temporary file with the --temp option).
- use git_config_rename_section_in_file() on that temporary file to
remove the section about the submodule.
- read that temporary file back into memory and write it out as a
blob object by calling sha1-file.c::write_object_file().
- add that back to the index as .gitmodules (studying how
builtin/update-index.c::add_cacheinfo() calls add_cache_entry()
would be a good way to learn how to do this).
The working tree side can stay as is, but as I said in the earlier
message, I think you need to update the .gitmodules in the working
tree and .gitmodules in the index separately (and without doing any
equivalent of "git add .gitmodules").
On 22/02 11:29, Junio C Hamano wrote:
Shourya Shukla [off-list ref] writes:
quoted
+ if (git_config_rename_section_in_file(index_only ? GITMODULES_INDEX :
+ GITMODULES_FILE, sect.buf, NULL) < 0) {
Also, is it really sufficient to pass GITMODULES_INDEX as the first
argument to this function to tweak what is in the index?
git_config_copy_or_rename_section_in_file() which is the
implementation of that helper seems to always want to work with a
file that is on disk, by making unconditional calls to
hold_lock_file_for_update(), fopen(), fstat(), chmod(), etc.
So I suspect that there are much more work needed.
I am not able to comprehend _why_ we need so much more work. To me it
seems to work fine.
The flow now is something like:
1. If !index_only i.e., '--cached' is not passed then remove the entry
of the SM from the working tree copy of '.gitmodules' i.e.,
GITMODULES_FILE. If there are any unstaged mods in '.gitmodules', we do
not proceed with 'git rm'.
2. Now, delete the entry of the above SM from the index copy of the
'.gitmodules' i.e., GITMODULES_INDEX (irrespective of the value of
'index_only'). If there are any unstaged mods in '.gitmodules', we do
not proceed with 'git rm'.
3. Finally, after the deletion of the SM entry, we stage the changes
using 'stage_updated_gitmodules()'.
Also, before any of the above thing happens, we check if it is OK to
write the '.gitmodules' using 'is_staging_gitmodules_ok()'. All the
above behaviour is in line with the current behaviour of 'git rm'.
What exactly do we need to change then?
It seems to me that the config editing API is one of the older and
hackier parts of the system and requires quite a lot of work to
teach it to work with anything but a on-disk file. In the longer
term, it may be a good thing to clean it up, but I suspect that it
is way too much work for too little benefit to do so as a part of
this topic, so an easier way out for now would be to:
- write out the .gitmodules in the index to a temporary file (learn
how to correctly call entry.c::checkout_entry() by studying how
builtin/checkout-index.c::checkout_file() calls it, especially to
a temporary file with the --temp option).
- use git_config_rename_section_in_file() on that temporary file to
remove the section about the submodule.
- read that temporary file back into memory and write it out as a
blob object by calling sha1-file.c::write_object_file().
- add that back to the index as .gitmodules (studying how
builtin/update-index.c::add_cacheinfo() calls add_cache_entry()
would be a good way to learn how to do this).
The working tree side can stay as is, but as I said in the earlier
message, I think you need to update the .gitmodules in the working
tree and .gitmodules in the index separately (and without doing any
equivalent of "git add .gitmodules").
But 'git rm' itself used to stage the changes i.e., 'git add'-ing them.
Regards,
Shourya Shukla