git/list[1] front-page[2] threads[3] people[4] search[5] about
 

Re: [PATCH 1/2] branch: do not track a start point with no ref

From
D. Ben Knoble <ben.knoble@gmail.com>
Date
Nov 6, 2026, 14:05 UTC
Message-ID
<CALnO6CAxSHWzvB+jvWw8rFXtWuj9Upv+i3g5EA_b2PkvZ=LV+A@mail.gmail.com>
In-Reply-To
<20260822-vv-branch-recurse-no-start-ref-v1-1-46dc140acaa8@zitro.id>
On Fri, Aug 21, 2026 at 7:03 PM Volodymyr Vriukalo <0@zitro.id> wrote:
Show 19 quoted lines
>
> Forcing a branch to a commit that no ref points at aborts when both
>   `submodule.recurse` and `submodule.propagateBranches` are set and
>   the repository has a remote configured:
>
>     BUG: refspec.c:442: refspec_find_match: need either src or dst
>     Aborted (core dumped)
>
> `create_branches_recursively()` resolves the start point through
>   `dwim_branch_start()`, which leaves `branch_point` NULL when the
>   start point names no ref -- an object id, or a revision expression
>   such as `HEAD~0`.  That NULL becomes `tracking_name`, and the
>   `setup_tracking()` call below it is guarded on `track` alone.
>   `setup_tracking()` assigns it to `tracking.spec.dst` without
>   checking, then hands the spec to `for_each_remote()`, so
>   `refspec_find_match()` receives a query with neither src nor dst
>   and trips its assertion.
> `for_each_remote()` never reaches that callback where no remote is
>   configured, which is why the abort needs one.
I'm not sure this is quite easy to follow:
* for_each_remote() reads the config and calls its callback, so it's a
little less relevant (except to say that we need a remote to trigger
the bug)
* it's really find_tracked_branch() that passes along NULL
tracking.spec.dst to remote_find_tracking()
Show 30 quoted lines
> 961b130d20 (branch: add --recurse-submodules option for branch
>   creation, 2022-01-28) added the call with no guard at all.
> 75388bf5b4 (branch: support more tracking modes when recursing,
>   2022-03-29) added the guard on `track`.
>
> Updating the branch happens before the abort, so the command does
>   what was asked and then exits 134.
> Callers that check the exit status therefore see a failure that did
>   not happen, and one that rolls back on failure would undo a
>   successful update.
>
> `create_branch()` already declines this: it calls `setup_tracking()`
>   under `if (real_ref && track)`, leaving tracking unset when the
>   start point resolved to no ref.
> Make the recursive path agree.
> Checking for NULL inside `setup_tracking()` would also silence the
>   abort, but it would put the decision in the callee for one caller
>   that has the answer already, and leave the two creation paths
>   disagreeing about when tracking is set up.
>
> Reproducing it needs all four of:
>
>   - `submodule.recurse=true`
>   - `submodule.propagateBranches=true`
>   - a configured remote
>   - a start point that is not a ref name
>
> Submodules take no part, so the new test builds a repository with
>   neither a submodule nor a `.gitmodules`, where `propagateBranches`
>   is set and has nothing to propagate to.
Anyway, I think we can simplify the message to something like
    With submodule.propagateBranches, we recursively create branches in
    submodules. But when the target starting point is a commit instead of a
    branch name, we have no tracking information, yet we try to find some and
    bug:
        BUG: refspec.c:442: refspec_find_match: need either src or dst
    create_branches_recursively() resolves the start point through
    dwim_branch_start(), which is NULL when the start point is not a ref. That
    NULL becomes tracking_name, but the setup_tracking() call below it requires
    it to be valid. setup_tracking() assigns it to tracking.spec.dst without
    checking, which makes its way to find_tracked_branch() and
    remote_find_tracking() when there is a remote present.
    Make the recursive path agree with create_branch() and guard against bogus
    setup_tracking() calls.
    Since submodules are not necessary to reproduce the BUG, omit them from the
    test case. A remote configuration and recursing through submodules with
    branches are required.
perhaps?
Show 42 quoted lines
> Assisted-by: An LLM.
> Signed-off-by: Volodymyr Vriukalo <0@zitro.id>
> ---
>  branch.c                    |  2 +-
>  t/t3207-branch-submodule.sh | 17 +++++++++++++++++
>  2 files changed, 18 insertions(+), 1 deletion(-)
>
> diff --git a/branch.c b/branch.c
> index 243db7d0fc..182fc4a3dd 100644
> --- a/branch.c
> +++ b/branch.c
> @@ -806,7 +806,7 @@ void create_branches_recursively(struct repository *r, const char *name,
>          * tedious to determine whether or not tracking was set up in the
>          * superproject.
>          */
> -       if (track)
> +       if (tracking_name && track)
>                 setup_tracking(name, tracking_name, track, quiet);
>
>         for (i = 0; i < submodule_entry_list.entry_nr; i++) {
> diff --git a/t/t3207-branch-submodule.sh b/t/t3207-branch-submodule.sh
> index fe72b24716..54f7caeb2f 100755
> --- a/t/t3207-branch-submodule.sh
> +++ b/t/t3207-branch-submodule.sh
> @@ -98,6 +98,23 @@ test_expect_success 'should respect submodule.recurse when creating branches' '
>         )
>  '
>
> +test_expect_success 'should move a branch to a start point that names no ref' '
> +       test_when_finished "rm -rf no-submodules" &&
> +       git init no-submodules &&
> +       (
> +               cd no-submodules &&
> +               test_commit one &&
> +               test_commit two &&
> +               git remote add origin . &&
> +               git config submodule.propagateBranches true &&
> +               git config submodule.recurse true &&
> +               git branch branch-a HEAD~1 &&
> +               oid=$(git rev-parse HEAD) &&
> +               git branch -f branch-a "$oid" &&
> +               test_cmp_rev HEAD branch-a
I'm having a hard time understanding this test.

We make a branch on HEAD~1, which is not a ref, so wouldn't it trip the bug prior to this fix? Anyway, let's assume it didn't---then we move the branch back to HEAD and check it. That part makes sense.

So as for the HEAD~1 business: the beginning of your mail did say "forcing" a branch, which I missed in my rewrite. Does this bug not trip when the creation is unforced? From the analysis you gave, I don't see why forcing would be a requirement, so wouldn't "git branch -f branch-a $oid" (where oid is resolved HEAD~1) do it? And then we could still keep the test_cmp_rev against HEAD~1 and branch-a.

And indeed, I can provoke the BUG in a repo with just
    git -c submodule.propagateBranches=true -c submodule.recurse=true
branch t @~3
for example, although "t" is still created.
Show 9 quoted lines
> +       )
> +'
> +
>  test_expect_success 'should ignore submodule.recurse when not creating branches' '
>         test_when_finished "reset_test" &&
>         (
>
> --
> 2.55.0.2.g927b4b9963
-- 
D. Ben Knoble
Previous: Volodymyr VriukaloNext: Volodymyr Vriukalo
Message 3 of 6 in “branch: fix --recurse-submodules with a nameless start point”
  1. 0/2 branch: fix --recurse-submodules with a nameless start pointVolodymyr Vriukalo, Aug 21, 2026
  2. 1/2 branch: do not track a start point with no refVolodymyr Vriukalo, Aug 21, 2026
  3. D. Ben KnobleNov 6, 2026
  4. 2/2 branch: allow recursion with no tracking nameVolodymyr Vriukalo, Aug 21, 2026
  5. D. Ben KnobleNov 6, 2026
  6. Junio C HamanoNov 2, 2026

Read the whole thread, see it on lore, or plain text.

$ cat FOOTERMessages come from the public archive at lore.kernel.org/git, fetched every hour. The front page is chosen and written each morning by an AI editor and can be wrong; the threads themselves are the record. About and API. For agents: an MCP server at https://gitlist.dev/mcp, and any thread, story or person page as Markdown by adding .md to its URL (or sending Accept: text/markdown). Details in /llms.txt.