Re: [Bug] 'git submodule update --single-branch' still fetching master

Hi Sven,

Le 2021-01-05 à 19:19, - - a écrit :

Thanks for a nice scenario description. Just a note on wording though (to make sure we are on the same page): you can't have a submodule "track" a branch, at least not in the sense that the submodule is always "checked out" on a branch, and the superproject knows that the submodule should always be "on that branch"; that's not how submodules work. As specified in [1], the config `submodule.<name>.branch` is only ever used by `git submodule update --remote`, i.e. the superproject itself only records at which *commit* the submodule is at.

That's not a bug (in my opinion) since it's in a line with the documentation [1]. There is no way at the moment for the initial `git submodule add` to pass '--single-branch' and '--branch' to the underlying `git clone`.

  
I can partly reproduce that behaviour. The following script creates "lib" with two branches,
master and light, that *do not* share any history.

~~~ #!/bin/bash

TEST_AUTHOR_LOCALNAME=author TEST_AUTHOR_DOMAIN=example.com GIT_AUTHOR_EMAIL=${TEST_AUTHOR_LOCALNAME}@${TEST_AUTHOR_DOMAIN} GIT_AUTHOR_NAME='A U Thor' GIT_AUTHOR_DATE='1112354055 +0200' TEST_COMMITTER_LOCALNAME=committer TEST_COMMITTER_DOMAIN=example.com GIT_COMMITTER_EMAIL=${TEST_COMMITTER_LOCALNAME}@${TEST_COMMITTER_DOMAIN} GIT_COMMITTER_NAME='C O Mitter' GIT_COMMITTER_DATE='1112354055 +0200' export GIT_AUTHOR_EMAIL GIT_AUTHOR_NAME export GIT_COMMITTER_EMAIL GIT_COMMITTER_NAME export GIT_COMMITTER_DATE GIT_AUTHOR_DATE

mkdir test cd test

rm -rf project lib clone

git init lib && cd lib echo "content">file git add file git commit -m "initial on master in lib" git checkout --orphan light echo "content">file git add file git commit -m "initial on light in lib" # Switch HEAD back to master git checkout master cd ../

git init project && cd project echo "content">file git add file git commit -m "initial in proj" git submodule add -b light ../lib git commit -m "add lib to project" cd ../

git clone project clone && cd clone echo 'RUNNING: git submodule update --init --remote --single-branch' # This fails git submodule update --init --remote --single-branch echo 'RUNNING: ls -al lib' ls -al lib echo 'RUNNING: git -C lib ls-files -s' git -C lib ls-files -s echo 'RUNNING: git -C lib branch --all' git -C lib branch --all echo 'RUNNING: git -C lib config --get remote.origin.fetch' git -C lib config --get remote.origin.fetch # Try to update again echo 'RUNNING: git submodule update' GIT_TRACE2=1 git submodule update 2> >(grep -E 'start git fetch|^From|^ \*' ) ~~~

The output should end with:

RUNNING: git submodule update --init --remote --single-branch
Submodule 'lib' (/Users/Philippe/Code/GIT-devel/test/lib) registered for path 'lib'
Cloning into '/Users/Philippe/Code/GIT-devel/test/clone/lib'...
done.
fatal: Needed a single revision
Unable to find current origin/light revision in submodule path 'lib'
RUNNING: ls -al lib
total 8
drwxr-xr-x  3 Philippe  staff  102 Jan  6 14:03 .
drwxr-xr-x  6 Philippe  staff  204 Jan  6 14:03 ..
-rw-r--r--  1 Philippe  staff   28 Jan  6 14:03 .git
RUNNING: git -C lib ls-files -s
RUNNING: git -C lib branch --all
* master
   remotes/origin/HEAD -> origin/master
   remotes/origin/master
RUNNING: git -C lib config --get remote.origin.fetch
+refs/heads/master:refs/remotes/origin/master
RUNNING: git submodule update
14:14:09.745288 common-main.c:49                  start git fetch
14:14:09.892463 common-main.c:49                  start git fetch origin cc09362ae7314174785faed3e5ee72a8e2db5460
 From /Users/Philippe/Code/GIT-devel/test/lib
  * branch            cc09362ae7314174785faed3e5ee72a8e2db5460 -> FETCH_HEAD
Submodule path 'lib': checked out 'cc09362ae7314174785faed3e5ee72a8e2db5460'

You can see that the first 'git submodule update --init --remote --single-branch' fails. 'ls' and 'ls-files' tell us that the submodule is not checked out, and that the submodule index is empty. 'git branch' and 'git config' then reveal that indeed only the 'master' branch exists in the submodule, so '--single-branch' was passed correctly but not the branch itself. Afterwards, issuing 'git submodule update' a second time will issue a simple 'git fetch', which will not work because of the single-branch refspec, then it will try fetching the commit by its hash, which works, and then it will checkout that commit.

Looking at commit 132f600b06 (clone: pass --single-branch during,
  --recurse-submodules, 2020-02-20), which introduced the '--single-branch'
option for 'git submodule update', it seems that what we have at the moment
is partly a documentation bug, i.e. the doc should say
"Clone only the history leading up to the remote HEAD during update",
and not mention '--branch' at all. I'm CC-ing Emily and Peff who were involved
in that series. Looking at the mailing list thread where
this feature was discussed [2], it seems that mentioning 'branch' in the description
of 'single-branch' was just a leftover from v1 of the series that was forgotten
in later iterations.

Also,looking at the code, the '--remote' option has no effect at all for the initial clone of the submodule. It is only taken into account after the initial clone. I think it would make sense for it to also influence the initial clone, so that it would behave as you expect, i.e. the code looks at the branch in '.gitmodules' and this branch is passed to the underlying 'git clone' if '--single-branch' was passed to 'git submodule update'.

That would be a diffenret feature request, and I think it does make sense for (b) to work. I don't think that '--branch' should automatically imply '--single-branch' though (i.e. I don't think (a) should change behaviour).

Yes, this does a plain 'git clone' so all branches are fetched.

As my script above shows, this does fetch only the 'master' branch, and then *if* we issue a second 'git submodule update' it then fetches the required commmit by hash and checks it out. It would be nice if you could send a reproducer (i.e. a complete script like the above) that shows exactly the behaviour you are seeing, if indeed in your scenario it really checks out the 'light' branch. I'm not sure I understand how that could happen...

In the past, 'git submodule add' used to not clone the submodule at all, but this was deemed not very useful so now it does 'git clone' underneath.

Hope that helps,

Philippe.

[1] https://git-scm.com/docs/git-submodule#Documentation/git-submodule.txt--bltbranchgt [2] https://lore.kernel.org/git/20200108231900.192476-1-emilyshaffer@google.com/t/#u