rebase: handle --update-refs branch symrefs#2126
Open
sluongng wants to merge 2 commits into
Open
Conversation
sluongng
force-pushed
the
sl/rebase-update-refs-symrefs
branch
from
May 27, 2026 14:14
9938d7e to
0ab0a71
Compare
Author
|
/preview |
|
Preview email sent as pull.2126.git.1779908024.gitgitgadget@gmail.com |
Author
|
/preview |
|
Preview email sent as pull.2126.git.1779943633.gitgitgadget@gmail.com |
Author
|
/preview |
|
Preview email sent as pull.2126.git.1779943735.gitgitgadget@gmail.com |
Author
|
/preview |
|
Preview email sent as pull.2126.git.1779944585.gitgitgadget@gmail.com |
Author
|
/submit |
|
Submitted as pull.2126.git.1779946921.gitgitgadget@gmail.com To fetch this version into To fetch this version to local tag |
![gitgitgadget[bot]](https://cdn.github.com/in/12836?s=60&v=4){kind=small}
| @@ -6445,28 +6445,67 @@ static int add_decorations_to_list(const struct commit *commit, | |||
| struct todo_add_branch_context *ctx) | |||
There was a problem hiding this comment.
"Kristoffer Haugsbakk" wrote on the Git mailing list (how to reply to this email):
On Thu, May 28, 2026, at 07:42, Son Luong Ngoc via GitGitGadget wrote:
>[snip]
> -test_expect_failure '--update-refs skips branch symrefs to current branch' '
> +test_expect_success '--update-refs skips branch symrefs to current branch' '
> test_when_finished "
> test_might_fail git rebase --abort &&
> git checkout primary &&
This style of fixing a bug by:
• Add failing test `test_expect_failure` in the first commit
• Fix the bug in the next commit and flip to `test_expect_success`
Is legitimate and makes it easier to verify that the test really
exercises the regression. But in this project it is preferred to
just do the bug fix + regression test in one patch.
See https://lore.kernel.org/git/xmqqfrdk3aqy.fsf@gitster.g/
> --
> gitgitgadget
|
User |
|
Junio C Hamano wrote on the Git mailing list (how to reply to this email): "Son Luong Ngoc via GitGitGadget" <gitgitgadget@gmail.com> writes:
> git rebase --update-refs can fail after the normal rebase path has
> successfully updated the current branch when another local branch is a
> symbolic ref to it.
>
> One practical way to arrive at that setup is a default branch rename from
> master to main. While the migration is in progress, a user may keep
> refs/heads/main as a symbolic ref to refs/heads/master so that both names
> continue to work locally.
>
> If pull.rebase is enabled, a plain git pull can then finish the rebase of
> master and still fail while trying to update the main alias. The reported
> failure looked like this, with line breaks adjusted for the cover letter:
>
> Successfully rebased and updated refs/heads/master.
> error: update_ref failed for ref 'refs/heads/main':
> cannot lock ref 'refs/heads/main':
> is at fc2c7bd5f17abec7861ef759edcd33a1e16662a1
> but expected 531cabdfb49098d6ffa502ed4bf91d1b35edfcfa
> Updated the following refs with --update-refs:
> Failed to update the following refs with --update-refs:
> refs/heads/main
I vaguely recall we saw a different topic that dealt with a
situation somewhat similar to this topic (I think it was about
'describe' giving a name that is not a branch). How would this mesh
with what the other topic wanted to do? Instead of filtering out
non-branch names (which the other topic did), here we want to filter
out names that are not concrete branches but pointers to something
else. Would it mean that the logic here is more broad (i.e., both
wants to filter out names of non-branches), making the other topic
unnecessary?
|
|
This patch series was integrated into seen via git@d871d43. |
sluongng
force-pushed
the
sl/rebase-update-refs-symrefs
branch
from
May 29, 2026 13:41
0ab0a71 to
c1e8828
Compare
Author
|
/preview |
|
Preview email sent as pull.2126.v2.git.1780062147590.gitgitgadget@gmail.com |
sluongng
force-pushed
the
sl/rebase-update-refs-symrefs
branch
from
May 29, 2026 15:23
c1e8828 to
b59eb4e
Compare
Author
|
/preview |
|
Preview email sent as pull.2126.v2.git.1780068269833.gitgitgadget@gmail.com |
|
This branch is now known as |
|
This patch series was integrated into seen via git@eb72804. |
|
This patch series was integrated into seen via git@9ced507. |
| ) && | ||
| grep ^update-ref todo >actual && | ||
| test_write_lines "update-ref refs/heads/no-conflict-branch" >expect && | ||
| test_cmp expect actual |
There was a problem hiding this comment.
Phillip Wood wrote on the Git mailing list (how to reply to this email):
On 28/05/2026 06:42, Son Luong Ngoc via GitGitGadget wrote:
> From: Son Luong Ngoc <sluongng@gmail.com>
> > rebase --update-refs queues local branch decorations by their literal
> refnames. When a branch such as refs/heads/main is a symbolic ref to
> the current branch, the normal rebase path first updates the current
> branch and the queued symref update later tries to update the same
> referent with the old value it recorded before the rebase.
> > Add a known-breakage test that exercises this case so that the fix can
> flip it to test_expect_success. The expected behavior is that the branch
> symref keeps pointing at the rebased current branch.
Thanks for adding a test, I'd find it easier to review this series if the test was added in the same patch as the fix which is our usual practice.
> +test_expect_failure '--update-refs skips branch symrefs to current branch' '
> + test_when_finished "
> + test_might_fail git rebase --abort &&
> + git checkout primary &&
> + test_might_fail git symbolic-ref -d refs/heads/update-refs-symref-alias &&
> + test_might_fail git branch -D update-refs-symref update-refs-symref-base
> + " &&
> + git checkout -B update-refs-symref-base primary &&
> + test_commit --no-tag update-refs-symref-base symref-base.t &&
> + git checkout -B update-refs-symref &&
> + test_commit --no-tag update-refs-symref-topic symref-topic.t &&
> + git checkout update-refs-symref-base &&
> + test_commit --no-tag update-refs-symref-newbase symref-newbase.t &&
> + git checkout update-refs-symref &&
> + git symbolic-ref refs/heads/update-refs-symref-alias refs/heads/update-refs-symref &&
I think we want to test a symref that does not match HEAD as well. Rather than adding a new test, can we instead add a couple of symref branches to the test "--update-refs updates refs correctly"?
Thanks
Phillip
> +
> + git rebase --update-refs update-refs-symref-base 2>err &&
> +
> + test_cmp_rev update-refs-symref-base update-refs-symref^ &&
> + test_cmp_rev refs/heads/update-refs-symref refs/heads/update-refs-symref-alias &&
> + test_write_lines refs/heads/update-refs-symref >expect &&
> + git symbolic-ref refs/heads/update-refs-symref-alias >actual &&
> + test_cmp expect actual
> +'
> +
> test_expect_success '--update-refs updates refs correctly' '
> git checkout -B update-refs no-conflict-branch &&
> git branch -f base HEAD~4 &&
|
User |
| @@ -6445,28 +6445,67 @@ static int add_decorations_to_list(const struct commit *commit, | |||
| struct todo_add_branch_context *ctx) | |||
There was a problem hiding this comment.
Phillip Wood wrote on the Git mailing list (how to reply to this email):
On 28/05/2026 06:42, Son Luong Ngoc via GitGitGadget wrote:
> From: Son Luong Ngoc <sluongng@gmail.com>
> > rebase --update-refs records local branch decorations before replaying
> commits. If a decoration is a symbolic branch such as refs/heads/main
> pointing at refs/heads/master, updating it later dereferences back to
> master and can fail because the normal rebase path already moved that
> branch.
Good explanation, thanks for working on this.
> Resolve local branch symref decorations to their referents before
s/referents/targets/ ?
> queuing update-ref commands, and skip duplicates. This keeps branch
> aliases from scheduling a second update for the same underlying branch
> while still using the existing old-OID check for the single queued
> update.
That's not quite what the patch does though - it only checks that the target of the symref differs from the target of HEAD. If a symref points to another branch we still try to update it when we should skip it.
> Signed-off-by: Son Luong Ngoc <sluongng@gmail.com>
> ---
> sequencer.c | 63 +++++++++++++++++++++++++++++------
> t/t3404-rebase-interactive.sh | 2 +-
> 2 files changed, 53 insertions(+), 12 deletions(-)
> > diff --git a/sequencer.c b/sequencer.c
> index 1ee4b2875b..4a83d1337c 100644
> --- a/sequencer.c
> +++ b/sequencer.c
> @@ -6445,15 +6445,22 @@ static int add_decorations_to_list(const struct commit *commit,
> struct todo_add_branch_context *ctx)
> {
> const struct name_decoration *decoration = get_name_decoration(&commit->object);
> - const char *head_ref = refs_resolve_ref_unsafe(get_main_ref_store(the_repository),
> - "HEAD",
> + struct ref_store *refs = get_main_ref_store(the_repository);
> + const char *head_ref = refs_resolve_ref_unsafe(refs, "HEAD",
> RESOLVE_REF_READING,
> - NULL,
> - NULL);
> + NULL, NULL);
> + char *resolved_head_ref = refs_resolve_refdup(refs, "HEAD",
> + RESOLVE_REF_READING,
> + NULL, NULL);
We need to use refs_resolve_refdup() instead of refs_resolve_ref_unsafe() so that the return value is not overwritten by the later calls to refs_resolve_ref_unsafe() that are added below. But that is the only change that is needed - we do not need to add a new variable, we just replace refs_resolve_ref_unsafe() with refs_resole_refdup() and free "head_ref" before we return.
> + struct strbuf update_ref = STRBUF_INIT;
> > while (decoration) {
> struct todo_item *item;
> const char *path;
> + const char *ref = decoration->name;
> + const char *resolved_ref;
> + int is_symref = 0;
> + int flags = 0;
> size_t base_offset = ctx->buf->len;
> > /*
> @@ -6461,12 +6468,44 @@ static int add_decorations_to_list(const struct commit *commit,
> * updated by the default rebase behavior.
> * Exclude it from the list of refs to update,
> * as well as any non-branch decorations.
> + *
> + * Resolve branch symrefs after checking for the current HEAD so
> + * that aliases do not schedule duplicate updates for their
> + * referents.
> + *
> * Non-branch decorations may be present if the pretty format
> * includes "%d", which would have loaded all refs
> * into the global decoration table.
> */
> - if ((head_ref && !strcmp(head_ref, decoration->name)) ||
> - (decoration->type != DECORATION_REF_LOCAL)) {
> + if (decoration->type != DECORATION_REF_LOCAL) {
> + decoration = decoration->next;
> + continue;
> + }
> +
> + if (head_ref && !strcmp(head_ref, ref)) {
> + decoration = decoration->next;
> + continue;
> + }
This is just rewriting the existing if statement which has nothing to do with the stated aim of this patch - lets leave it as it was.
> +
> + strbuf_reset(&update_ref);
> + resolved_ref = refs_resolve_ref_unsafe(refs, ref,
> + RESOLVE_REF_READING |
> + RESOLVE_REF_NO_RECURSE,
Why are we passing RESOLVE_REF_NO_RECURSE here? I'd have thought we want to resolve the whole chain of symbolic refs to find out which ref is actually going to be updated.
> + NULL, &flags);
> + if ((flags & REF_ISSYMREF) && resolved_ref) {
I think it is generally safer to check the return value before using any of the "out" parameters from a function call. In this case the function unconditionally clears flags at the beginning so it is safe.
> + if (!starts_with(resolved_ref, "refs/heads/")) {
> + decoration = decoration->next;
> + continue;
This is the opposite of what I was expecting - if the decoration is a symref that resolves to a branch then that branch will also be in the list of decorations and so will be updated. If the decoration is a symref that resolves outside "refs/heads/" then we want to add the decoration to the list of refs to update to keep the current behavior.
If we do that then we skip all symbolic refs that point to another branch, instead of just skipping those that match HEAD and we don't need any of the changes below here.
Thanks
Phillip
> + }
> +
> + strbuf_addstr(&update_ref, resolved_ref);
> + ref = update_ref.buf;
> + is_symref = 1;
> + }
> +
> + if ((is_symref && resolved_head_ref &&
> + !strcmp(resolved_head_ref, ref)) ||
> + string_list_has_string(&ctx->refs_to_oids, ref)) {
> decoration = decoration->next;
> continue;
> }
> @@ -6478,19 +6517,19 @@ static int add_decorations_to_list(const struct commit *commit,
> memset(item, 0, sizeof(*item));
> > /* If the branch is checked out, then leave a comment instead. */
> - if ((path = branch_checked_out(decoration->name))) {
> + if ((path = branch_checked_out(ref))) {
> item->command = TODO_COMMENT;
> strbuf_commented_addf(ctx->buf, comment_line_str,
> "Ref %s checked out at '%s'\n",
> - decoration->name, path);
> + ref, path);
> } else {
> struct string_list_item *sti;
> item->command = TODO_UPDATE_REF;
> - strbuf_addf(ctx->buf, "%s\n", decoration->name);
> + strbuf_addf(ctx->buf, "%s\n", ref);
> > sti = string_list_insert(&ctx->refs_to_oids,
> - decoration->name);
> - sti->util = init_update_ref_record(decoration->name);
> + ref);
> + sti->util = init_update_ref_record(ref);
> }
> > item->offset_in_buf = base_offset;
> @@ -6501,6 +6540,8 @@ static int add_decorations_to_list(const struct commit *commit,
> decoration = decoration->next;
> }
> > + strbuf_release(&update_ref);
> + free(resolved_head_ref);
> return 0;
> }
> > diff --git a/t/t3404-rebase-interactive.sh b/t/t3404-rebase-interactive.sh
> index 42ba8cc313..29447c0fc3 100755
> --- a/t/t3404-rebase-interactive.sh
> +++ b/t/t3404-rebase-interactive.sh
> @@ -1978,7 +1978,7 @@ test_expect_success '--update-refs ignores non-branch decorations' '
> test_cmp expect actual
> '
> > -test_expect_failure '--update-refs skips branch symrefs to current branch' '
> +test_expect_success '--update-refs skips branch symrefs to current branch' '
> test_when_finished "
> test_might_fail git rebase --abort &&
> git checkout primary &&
|
This patch series was integrated into seen via git@b977dbc. |
|
This patch series was integrated into seen via git@389dbe0. |
|
There was a status update in the "Cooking" section about the branch 'git rebase --update-refs' has been taught to resolve local branch symrefs to their referents before queuing updates, ensuring aliases of the current branch are skipped and duplicate updates are avoided to prevent failures when branch aliases are present. Waiting for response(s) to review comment(s) for too long, stalled. cf. <f982c386-e329-4ab0-b695-e540bcb9de3d@gmail.com> source: <pull.2126.v2.git.1780482436865.gitgitgadget@gmail.com> |
|
There was a status update in the "Stalled" section about the branch 'git rebase --update-refs' has been taught to resolve local branch symrefs to their referents before queuing updates, ensuring aliases of the current branch are skipped and duplicate updates are avoided to prevent failures when branch aliases are present. Waiting for response(s) to review comment(s) for too long, stalled. cf. <f982c386-e329-4ab0-b695-e540bcb9de3d@gmail.com> source: <pull.2126.v2.git.1780482436865.gitgitgadget@gmail.com> |
sluongng
force-pushed
the
sl/rebase-update-refs-symrefs
branch
from
July 12, 2026 06:15
860ba6d to
07e863d
Compare
Author
|
/preview |
1 similar comment
Author
|
/preview |
|
Preview email sent as pull.2126.v3.git.1783837017299.gitgitgadget@gmail.com |
|
Preview email sent as pull.2126.v3.git.1783837035895.gitgitgadget@gmail.com |
|
There was a status update in the "Stalled" section about the branch 'git rebase --update-refs' has been taught to resolve local branch symrefs to their referents before queuing updates, ensuring aliases of the current branch are skipped and duplicate updates are avoided to prevent failures when branch aliases are present. Waiting for response(s) to review comment(s) for too long, stalled. cf. <f982c386-e329-4ab0-b695-e540bcb9de3d@gmail.com> source: <pull.2126.v2.git.1780482436865.gitgitgadget@gmail.com> |
|
There was a status update in the "Stalled" section about the branch 'git rebase --update-refs' has been taught to resolve local branch symrefs to their referents before queuing updates, ensuring aliases of the current branch are skipped and duplicate updates are avoided to prevent failures when branch aliases are present. Waiting for response(s) to review comment(s) for too long, stalled. cf. <f982c386-e329-4ab0-b695-e540bcb9de3d@gmail.com> source: <pull.2126.v2.git.1780482436865.gitgitgadget@gmail.com> |
|
There was a status update in the "Stalled" section about the branch 'git rebase --update-refs' has been taught to resolve local branch symrefs to their referents before queuing updates, ensuring aliases of the current branch are skipped and duplicate updates are avoided to prevent failures when branch aliases are present. Waiting for response for too long, stalled. cf. <f982c386-e329-4ab0-b695-e540bcb9de3d@gmail.com> source: <pull.2126.v2.git.1780482436865.gitgitgadget@gmail.com> |
|
There was a status update in the "Stalled" section about the branch 'git rebase --update-refs' has been taught to resolve local branch symrefs to their referents before queuing updates, ensuring aliases of the current branch are skipped and duplicate updates are avoided to prevent failures when branch aliases are present. Waiting for response for too long, stalled. cf. <f982c386-e329-4ab0-b695-e540bcb9de3d@gmail.com> source: <pull.2126.v2.git.1780482436865.gitgitgadget@gmail.com> |
|
There was a status update in the "Stalled" section about the branch 'git rebase --update-refs' has been taught to resolve local branch symrefs to their referents before queuing updates, ensuring aliases of the current branch are skipped and duplicate updates are avoided to prevent failures when branch aliases are present. Waiting for response for too long, stalled. cf. <f982c386-e329-4ab0-b695-e540bcb9de3d@gmail.com> source: <pull.2126.v2.git.1780482436865.gitgitgadget@gmail.com> |
sluongng
force-pushed
the
sl/rebase-update-refs-symrefs
branch
2 times, most recently
from
July 21, 2026 10:19
b4e2af3 to
e89289c
Compare
sluongng
changed the title
sluongng
force-pushed
the
sl/rebase-update-refs-symrefs
branch
from
July 21, 2026 12:26
e89289c to
7d27eba
Compare
sluongng
changed the title
sluongng
force-pushed
the
sl/rebase-update-refs-symrefs
branch
from
July 21, 2026 13:02
7d27eba to
a653f56
Compare
Author
|
/preview |
|
Preview email sent as pull.2126.v3.git.1784707485.gitgitgadget@gmail.com |
Author
|
/submit |
|
Submitted as pull.2126.v3.git.1784708107.gitgitgadget@gmail.com To fetch this version into To fetch this version to local tag |
|
Son Luong Ngoc wrote on the Git mailing list (how to reply to this email): On 04/06/2026 16:46, Phillip Wood wrote:
> A symref that points to another branch should always be skipped. When we
> look up which branches are checked out (see worktree.c:add_head_info()) we
> use
>
> refs_resolve_ref_unsafe(get_worktree_ref_store(wt),
> "HEAD",
> 0,
> &wt->head_oid, &flags);
>
> so it will never report a symref as being checked out - it always resolves
> any symrefs first.
Yes, this is the right distinction. Patch 1 now resolves each local
branch decoration before deciding whether to queue it. If the target is
under refs/heads/, the alias is skipped unconditionally and the concrete
branch decoration remains the only update that is queued.
> If we have a symref pointing somewhere outside of "refs/heads" then we
> need to check whether the target is checked out, not the symref itself.
> I'm not sure how likely that is to happen in practice.
Patch 2 handles that case separately. It checks both the literal alias
and its resolved target ref against the checked-out reservations. I also
added a test with a non-branch target checked out in another worktree.
> If a decoration matches the current branch why don't we just skip it like
> we used to? (As an aside the existing code in wrong because if the user
> runs "git rebase --update-refs <upstream> <branch>" HEAD does not point to
> "<branch>" but lets not worry about that now)
Agreed. Patch 1 now skips the current branch before checked-out
handling, as the old code did.
The contrary expectation in t3400 came from head_ref pointing into a
buffer that was reused while resolving another decoration. That could
make the current-branch comparison fail. head_ref is now an owned copy.
The test expects the current branch to be omitted from the todo list.
While adding the non-branch coverage, I found that two aliases to the
same target ref could queue the same update twice and make the second
compare-and-swap fail. Patch 2 deduplicates those updates by target ref.
It also records resolved target refs from other worktrees' in-progress
update-refs state so that a different alias honors the same reservation.
I split the reroll into these two patches so that the branch-alias fix
and the non-branch safeguards can be reviewed independently.
Thanks for the review, and sorry for the slow response.
Thanks,
Son
On Thu, 4 Jun 2026 16:37:39 +0100, Phillip Wood
<phillip.wood123@gmail.com> wrote:
> On 03/06/2026 11:27, Son Luong Ngoc via GitGitGadget wrote:
> > From: Son Luong Ngoc <sluongng@gmail.com>
> >
> > git rebase --update-refs can fail after the normal rebase path has
> > updated the current branch when another local branch is a symref to it.
> > This can happen during a default-branch rename where refs/heads/main
> > points at refs/heads/master while users migrate.
> >
> > The sequencer queues update-ref commands from local branch decorations.
> > Commit 106b6885c7 (rebase: ignore non-branch update-refs) filters out
> > decorations that are not local branches, such as HEAD and tags. A branch
> > symref is different: it is still a local branch decoration, but if it
> > resolves to another branch then that target branch is itself present in
> > the decoration list and will be updated as a concrete branch.
> >
> > Skip branch decorations whose symrefs resolve to refs/heads/*, because
> > those targets are already represented by concrete branch decorations.
> > This prevents aliases from scheduling a second update for the same
> > branch. Keep symrefs to non-branch targets on the existing path.
>
> Makes sense
>
> > Preserve the existing checked-out branch handling before applying these
> > skips. Such refs still need a todo-list comment instead of an update-ref
> > command, even when the checked-out ref is the branch being rebased or a
> > branch symref alias. Use a copy of the resolved HEAD ref so later ref
> > resolution does not overwrite it.
>
> I don't quite understand this. A symref that points to another branch
> should always be skipped. When we look up which branches are checked out
> (see worktree.c:add_head_info()) we use
>
> refs_resolve_ref_unsafe(get_worktree_ref_store(wt),
> "HEAD",
> 0,
> &wt->head_oid, &flags);
>
> so it will never report a symref as being checked out - it always
> resolves any symrefs first.
>
> If we have a symref pointing somewhere outside of "refs/heads" then we
> need to check whether the target is checked out, not the symref itself.
> I'm not sure how likely that is to happen in practice.
>
> > diff --git a/sequencer.c b/sequencer.c
> > index 1ee4b2875b..6ab8b47108 100644
> > --- a/sequencer.c
> > +++ b/sequencer.c
> > @@ -6445,28 +6445,46 @@ static int add_decorations_to_list(const struct commit *commit,
> > struct todo_add_branch_context *ctx)
> > {
> > const struct name_decoration *decoration = get_name_decoration(&commit->object);
> > - const char *head_ref = refs_resolve_ref_unsafe(get_main_ref_store(the_repository),
> > - "HEAD",
> > - RESOLVE_REF_READING,
> > - NULL,
> > - NULL);
> > + struct ref_store *refs = get_main_ref_store(the_repository);
> > + char *head_ref = refs_resolve_refdup(refs, "HEAD",
> > + RESOLVE_REF_READING,
> > + NULL, NULL);
>
> This part and the test look good now
> > while (decoration) {
> > struct todo_item *item;
> > const char *path;
> > + const char *resolved_ref;
> > + int flags = 0;
> > size_t base_offset = ctx->buf->len;
> >
> > /*
> > - * If the branch is the current HEAD, then it will be
> > - * updated by the default rebase behavior.
> > - * Exclude it from the list of refs to update,
> > - * as well as any non-branch decorations.
> > * Non-branch decorations may be present if the pretty format
> > * includes "%d", which would have loaded all refs
> > * into the global decoration table.
> > */
> > - if ((head_ref && !strcmp(head_ref, decoration->name)) ||
> > - (decoration->type != DECORATION_REF_LOCAL)) {
> > + if (decoration->type != DECORATION_REF_LOCAL) {
> > + decoration = decoration->next;
> > + continue;
> > + }
>
> If a decoration matches the current branch why don't we just skip it
> like we used to? (As an aside the existing code in wrong because if the
> user runs "git rebase --update-refs <upstream> <branch>" HEAD does not
> point to "<branch>" but lets not worry about that now)
>
> > + path = branch_checked_out(decoration->name);
>
> As I said above if the symref target is anther branch we should skip it
> and if the target is not a branch then we need to check if the target is
> checked out so we need to resolve the ref before calling
> branch_checked_out().
>
> Thanks
>
> Phillip |
|
There was a status update in the "Cooking" section about the branch 'git rebase --update-refs' has been taught to resolve local branch symrefs to their referents before queuing updates, ensuring aliases of the current branch are skipped and duplicate updates are avoided to prevent failures when branch aliases are present. Needs review. source: <pull.2126.v3.git.1784708107.gitgitgadget@gmail.com> |
| @@ -6465,40 +6465,58 @@ static int add_decorations_to_list(const struct commit *commit, | |||
| struct todo_add_branch_context *ctx) | |||
There was a problem hiding this comment.
Phillip Wood wrote on the Git mailing list (how to reply to this email):
On 22/07/2026 09:15, Son Luong Ngoc via GitGitGadget wrote:
> From: Son Luong Ngoc <sluongng@gmail.com>
> > git rebase --update-refs can finish rewriting the current branch and
> then fail while updating a local branch that is a symbolic ref. This can
> happen during a default-branch rename where refs/heads/main points at
> refs/heads/master while users migrate.
> > The problem is a partially applied ref update: the main rebase has
> already succeeded when the later ref update fails.
> > The sequencer queues updates from local branch decorations. Commit
> 106b6885c7 (rebase: ignore non-branch update-refs) filters out
> decorations such as HEAD and tags. A branch symref is still a local
> branch decoration, but refs_update_ref() dereferences it, so an alias to
> another branch duplicates the concrete branch update.
> > Resolve local branch decorations before queuing them. Skip symrefs whose
> targets are under refs/heads/ so that only the concrete branch update is
> queued. Keep an owned copy of the resolved HEAD and skip the current
> branch before checked-out handling so later ref resolution cannot change
> the comparison.
> > This prevents a successful rebase from being followed by a failed,
> partially applied ref update while preserving each alias as a symref.
Thanks for re-rolling I'm pretty sure the logic is sound now but I'm a bit confused by a couple of things - see my comments below.
> Signed-off-by: Son Luong Ngoc <sluongng@gmail.com>
> ---
> sequencer.c | 44 +++++++++++++++++++++++++----------
> t/t3400-rebase.sh | 2 +-
> t/t3404-rebase-interactive.sh | 16 +++++++++++++
> 3 files changed, 49 insertions(+), 13 deletions(-)
> > diff --git a/sequencer.c b/sequencer.c
> index 1355a99a09..63aba60a08 100644
> --- a/sequencer.c
> +++ b/sequencer.c
> @@ -6465,32 +6465,50 @@ static int add_decorations_to_list(const struct commit *commit,
> struct todo_add_branch_context *ctx)
> {
> const struct name_decoration *decoration = get_name_decoration(&commit->object);
> - const char *head_ref = refs_resolve_ref_unsafe(get_main_ref_store(the_repository),
> - "HEAD",
> - RESOLVE_REF_READING,
> - NULL,
> - NULL);
> + struct ref_store *refs = get_main_ref_store(the_repository);
> + char *head_ref = refs_resolve_refdup(refs, "HEAD",
> + RESOLVE_REF_READING,
> + NULL, NULL);
> > while (decoration) {
> struct todo_item *item;
> const char *path;
> + char *resolved_ref;
> + int flags = 0;
> size_t base_offset = ctx->buf->len;
> > /*
> - * If the branch is the current HEAD, then it will be
> - * updated by the default rebase behavior.
> - * Exclude it from the list of refs to update,
> - * as well as any non-branch decorations.
> * Non-branch decorations may be present if the pretty format
> * includes "%d", which would have loaded all refs
> * into the global decoration table.
> */
> - if ((head_ref && !strcmp(head_ref, decoration->name)) ||
> - (decoration->type != DECORATION_REF_LOCAL)) {
> + if (decoration->type != DECORATION_REF_LOCAL) {
> + decoration = decoration->next;
> + continue;
> + }
It would be nice to have a comment here explaining what we're doing. Also I don't think we need to copy the refname so it would be more efficient to use refs_resolve_ref_unsafe().
> + resolved_ref = refs_resolve_refdup(refs, decoration->name,
> + RESOLVE_REF_READING,
> + NULL, &flags);
> + if (resolved_ref && (flags & REF_ISSYMREF) &&
> + starts_with(resolved_ref, "refs/heads/")) {
> + free(resolved_ref);
> + decoration = decoration->next;
> + continue;
> + }
We skip any symbolic refs that point to another branch which is good.
> + /*
> + * If the branch is the current HEAD, then it will be
> + * updated by the default rebase behavior.
> + */
> + if (head_ref && !strcmp(head_ref, decoration->name)) {
> + free(resolved_ref);
> decoration = decoration->next;
> continue;
> }
Then we check to see if the decoration matches HEAD which we used to do above - I'm not clear why we have moved this check.
> + path = branch_checked_out(decoration->name);
> +
This belongs in the next patch I think.
> diff --git a/t/t3400-rebase.sh b/t/t3400-rebase.sh
> index e62e07b894..1a02f6546b 100755
> --- a/t/t3400-rebase.sh
> +++ b/t/t3400-rebase.sh
> @@ -471,7 +471,7 @@ test_expect_success 'git rebase --update-ref with core.commentChar and branch on
Adding an extra context line shows
git checkout topic2> GIT_SEQUENCE_EDITOR="cat >actual" git -c core.commentChar=% \
> rebase -i --update-refs base &&
> test_grep "% Ref refs/heads/wt-topic checked out at" actual &&
> - test_grep "% Ref refs/heads/topic2 checked out at" actual
> + test_grep ! "% Ref refs/heads/topic2 checked out at" actual
As topic2 is checked out in the worktree where the rebase is running why did this line appear before?
> diff --git a/t/t3404-rebase-interactive.sh b/t/t3404-rebase-interactive.sh
> index e64816770a..11afa8be56 100755
> --- a/t/t3404-rebase-interactive.sh
> +++ b/t/t3404-rebase-interactive.sh
> @@ -1975,15 +1975,23 @@ test_expect_success '--update-refs ignores non-branch decorations' '
> ) &&
> grep ^update-ref todo >actual &&
> test_write_lines "update-ref refs/heads/no-conflict-branch" >expect &&
> + test_grep ! "^# Ref refs/heads/update-refs checked out" todo &&
Lets move this line below test_cmp so we keep that line next to the ones that create the files that are being compared. Is this another case where we used to add this comment and no longer do so?
> test_cmp expect actual
> '
> > test_expect_success '--update-refs updates refs correctly' '
> + test_when_finished "
> + test_might_fail git symbolic-ref -d refs/heads/no-conflict-branch-alias &&
> + test_might_fail git symbolic-ref -d refs/heads/second-alias
> + " &&
> git checkout -B update-refs no-conflict-branch &&
> git branch -f base HEAD~4 &&
> git branch -f first HEAD~3 &&
> git branch -f second HEAD~3 &&
> git branch -f third HEAD~1 &&
> + git symbolic-ref refs/heads/no-conflict-branch-alias \
> + refs/heads/no-conflict-branch &&
> + git symbolic-ref refs/heads/second-alias refs/heads/second &&
> test_commit extra2 fileX &&
> git commit --amend --fixup=L &&
> > @@ -1991,8 +1999,16 @@ test_expect_success '--update-refs updates refs correctly' '
> > test_cmp_rev HEAD~3 refs/heads/first &&
> test_cmp_rev HEAD~3 refs/heads/second &&
> + test_cmp_rev HEAD~3 refs/heads/second-alias &&
> test_cmp_rev HEAD~1 refs/heads/third &&
> test_cmp_rev HEAD refs/heads/no-conflict-branch &&
> + test_cmp_rev HEAD refs/heads/no-conflict-branch-alias &&
> + test_write_lines refs/heads/no-conflict-branch >expect &&
> + git symbolic-ref refs/heads/no-conflict-branch-alias >actual &&
> + test_cmp expect actual &&
> + test_write_lines refs/heads/second >expect &&
> + git symbolic-ref refs/heads/second-alias >actual &&
> + test_cmp expect actual &&
This looks good - we check that "rebase --update-refs" succeeds withh branches that are symrefs and also that those refs are untouched by the rebase.
Thanks
Phillip
> q_to_tab >expect <<-\EOF &&
> Successfully rebased and updated refs/heads/update-refs.There was a problem hiding this comment.
Phillip Wood wrote on the Git mailing list (how to reply to this email):
On 23/07/2026 19:58, Phillip Wood wrote:
> On 22/07/2026 09:15, Son Luong Ngoc via GitGitGadget wrote:
>> From: Son Luong Ngoc <sluongng@gmail.com>
>>
>> git rebase --update-refs can finish rewriting the current branch and
>> then fail while updating a local branch that is a symbolic ref. This can
>> happen during a default-branch rename where refs/heads/main points at
>> refs/heads/master while users migrate.
>>
>> The problem is a partially applied ref update: the main rebase has
>> already succeeded when the later ref update fails.
>>
>> The sequencer queues updates from local branch decorations. Commit
>> 106b6885c7 (rebase: ignore non-branch update-refs) filters out
>> decorations such as HEAD and tags. A branch symref is still a local
>> branch decoration, but refs_update_ref() dereferences it, so an alias to
>> another branch duplicates the concrete branch update.
>>
>> Resolve local branch decorations before queuing them. Skip symrefs whose
>> targets are under refs/heads/ so that only the concrete branch update is
>> queued. Keep an owned copy of the resolved HEAD and skip the current
>> branch before checked-out handling so later ref resolution cannot change
>> the comparison.
>>
>> This prevents a successful rebase from being followed by a failed,
>> partially applied ref update while preserving each alias as a symref.
> > Thanks for re-rolling I'm pretty sure the logic is sound now but I'm a > bit confused by a couple of things - see my comments below.
> >> Signed-off-by: Son Luong Ngoc <sluongng@gmail.com>
>> ---
>> sequencer.c | 44 +++++++++++++++++++++++++----------
>> t/t3400-rebase.sh | 2 +-
>> t/t3404-rebase-interactive.sh | 16 +++++++++++++
>> 3 files changed, 49 insertions(+), 13 deletions(-)
>>
>> diff --git a/sequencer.c b/sequencer.c
>> index 1355a99a09..63aba60a08 100644
>> --- a/sequencer.c
>> +++ b/sequencer.c
>> @@ -6465,32 +6465,50 @@ static int add_decorations_to_list(const >> struct commit *commit,
>> struct todo_add_branch_context *ctx)
>> {
>> const struct name_decoration *decoration = >> get_name_decoration(&commit->object);
>> - const char *head_ref = >> refs_resolve_ref_unsafe(get_main_ref_store(the_repository),
>> - "HEAD",
>> - RESOLVE_REF_READING,
>> - NULL,
>> - NULL);
>> + struct ref_store *refs = get_main_ref_store(the_repository);
>> + char *head_ref = refs_resolve_refdup(refs, "HEAD",
>> + RESOLVE_REF_READING,
>> + NULL, NULL);
>> while (decoration) {
>> struct todo_item *item;
>> const char *path;
>> + char *resolved_ref;
>> + int flags = 0;
>> size_t base_offset = ctx->buf->len;
>> /*
>> - * If the branch is the current HEAD, then it will be
>> - * updated by the default rebase behavior.
>> - * Exclude it from the list of refs to update,
>> - * as well as any non-branch decorations.
>> * Non-branch decorations may be present if the pretty format
>> * includes "%d", which would have loaded all refs
>> * into the global decoration table.
>> */
>> - if ((head_ref && !strcmp(head_ref, decoration->name)) ||
>> - (decoration->type != DECORATION_REF_LOCAL)) {
>> + if (decoration->type != DECORATION_REF_LOCAL) {
>> + decoration = decoration->next;
>> + continue;
>> + }
> > It would be nice to have a comment here explaining what we're doing. > Also I don't think we need to copy the refname so it would be more > efficient to use refs_resolve_ref_unsafe().
Looking at this again we cannot use refs_resolve_ref_unsafe() because the result would be overwritten by the call to refs_resolve_refdup() in branch_checked_out().
>> + resolved_ref = refs_resolve_refdup(refs, decoration->name,
>> + RESOLVE_REF_READING,
>> + NULL, &flags);
>> + if (resolved_ref && (flags & REF_ISSYMREF) &&
>> + starts_with(resolved_ref, "refs/heads/")) {
>> + free(resolved_ref);
>> + decoration = decoration->next;
>> + continue;
>> + }
> > We skip any symbolic refs that point to another branch which is good.
> >> + /*
>> + * If the branch is the current HEAD, then it will be
>> + * updated by the default rebase behavior.
>> + */
>> + if (head_ref && !strcmp(head_ref, decoration->name)) {
>> + free(resolved_ref);
>> decoration = decoration->next;
>> continue;
>> }
> > Then we check to see if the decoration matches HEAD which we used to do > above - I'm not clear why we have moved this check.
Should we be using "resolved_ref" instead of "decoration->name"? That would explain why this was moved and would makes sense as we resolve symrefs when reading HEAD. When HEAD points outside "refs/heads/" we'd then skip updating any symrefs under "refs/heads/" that pointed to the same ref as HEAD.
Thanks
Phillip
> >> + path = branch_checked_out(decoration->name);
>> +
> > This belongs in the next patch I think.
> >> diff --git a/t/t3400-rebase.sh b/t/t3400-rebase.sh
>> index e62e07b894..1a02f6546b 100755
>> --- a/t/t3400-rebase.sh
>> +++ b/t/t3400-rebase.sh
>> @@ -471,7 +471,7 @@ test_expect_success 'git rebase --update-ref with >> core.commentChar and branch on
> > Adding an extra context line shows
> > git checkout topic2> GIT_SEQUENCE_EDITOR="cat >actual" git -c > core.commentChar=% \
>> rebase -i --update-refs base &&
>> test_grep "% Ref refs/heads/wt-topic checked out at" actual &&
>> - test_grep "% Ref refs/heads/topic2 checked out at" actual
>> + test_grep ! "% Ref refs/heads/topic2 checked out at" actual
> > As topic2 is checked out in the worktree where the rebase is running why > did this line appear before?
> >> diff --git a/t/t3404-rebase-interactive.sh b/t/t3404-rebase- >> interactive.sh
>> index e64816770a..11afa8be56 100755
>> --- a/t/t3404-rebase-interactive.sh
>> +++ b/t/t3404-rebase-interactive.sh
>> @@ -1975,15 +1975,23 @@ test_expect_success '--update-refs ignores >> non-branch decorations' '
>> ) &&
>> grep ^update-ref todo >actual &&
>> test_write_lines "update-ref refs/heads/no-conflict-branch" >> >expect &&
>> + test_grep ! "^# Ref refs/heads/update-refs checked out" todo &&
> > Lets move this line below test_cmp so we keep that line next to the ones > that create the files that are being compared. Is this another case > where we used to add this comment and no longer do so?
> >> test_cmp expect actual
>> '
>> test_expect_success '--update-refs updates refs correctly' '
>> + test_when_finished "
>> + test_might_fail git symbolic-ref -d refs/heads/no-conflict- >> branch-alias &&
>> + test_might_fail git symbolic-ref -d refs/heads/second-alias
>> + " &&
>> git checkout -B update-refs no-conflict-branch &&
>> git branch -f base HEAD~4 &&
>> git branch -f first HEAD~3 &&
>> git branch -f second HEAD~3 &&
>> git branch -f third HEAD~1 &&
>> + git symbolic-ref refs/heads/no-conflict-branch-alias \
>> + refs/heads/no-conflict-branch &&
>> + git symbolic-ref refs/heads/second-alias refs/heads/second &&
>> test_commit extra2 fileX &&
>> git commit --amend --fixup=L &&
>> @@ -1991,8 +1999,16 @@ test_expect_success '--update-refs updates refs >> correctly' '
>> test_cmp_rev HEAD~3 refs/heads/first &&
>> test_cmp_rev HEAD~3 refs/heads/second &&
>> + test_cmp_rev HEAD~3 refs/heads/second-alias &&
>> test_cmp_rev HEAD~1 refs/heads/third &&
>> test_cmp_rev HEAD refs/heads/no-conflict-branch &&
>> + test_cmp_rev HEAD refs/heads/no-conflict-branch-alias &&
>> + test_write_lines refs/heads/no-conflict-branch >expect &&
>> + git symbolic-ref refs/heads/no-conflict-branch-alias >actual &&
>> + test_cmp expect actual &&
>> + test_write_lines refs/heads/second >expect &&
>> + git symbolic-ref refs/heads/second-alias >actual &&
>> + test_cmp expect actual &&
> > This looks good - we check that "rebase --update-refs" succeeds withh > branches that are symrefs and also that those refs are untouched by the > rebase.
> > Thanks
> > Phillip
> >> q_to_tab >expect <<-\EOF &&
>> Successfully rebased and updated refs/heads/update-refs.
> There was a problem hiding this comment.
Erik Cervin-Edin wrote on the Git mailing list (how to reply to this email):
I ran into a minor regression that I tracked down to v2 of this patch
series. When I was running
git rebase --interactive --update-refs
I got a comment in my git-rebase-todo for the branch I was rebasing,
even though it's just the current branch in the current worktree:
# Ref refs/heads/main checked out at '/private/tmp/gittest'
where I'd expect no entry at all, since the branch is updated by the
rebase itself. After applying v3 of the series, the regression went
away. Nevertheless, I thought I ought to share what I found.
On 26/07/24 10:55AM, Phillip Wood wrote:
> > > + if (head_ref && !strcmp(head_ref, decoration->name)) {
> > > + free(resolved_ref);
> > > decoration = decoration->next;
> > > continue;
> > > }
> > > +
> > > + path = branch_checked_out(decoration->name);
> >
> > Then we check to see if the decoration matches HEAD which we used to do
> > above - I'm not clear why we have moved this check.
branch_checked_out() can't tell "checked out in another worktree"
apart from "checked out right here", so `path` is never NULL for the
branch actually being rebased. In v2, the check above was instead:
if (!path && head_ref && !strcmp(head_ref, decoration->name))
continue;
which made it a no-op for exactly that branch -- the regression I
observed. v3 drops the `!path` gate and moves `path` below it, which
is why it's fixed.
> > As topic2 is checked out in the worktree where the rebase is running
> > why did this line appear before?
This might be the same symptom from another cause: on master, head_ref
comes from refs_resolve_ref_unsafe(), and as you note its buffer is
overwritten inside branch_checked_out(). So by the time topic2 is
compared, wt-topic's lookup may already have clobbered head_ref,
letting topic2 fall through to the comment. I haven't run the test to
confirm, though.
Thanks,
ErikThere was a problem hiding this comment.
Junio C Hamano wrote on the Git mailing list (how to reply to this email):
Phillip Wood <phillip.wood123@gmail.com> writes:
>> Thanks for re-rolling I'm pretty sure the logic is sound now but I'm a
>> bit confused by a couple of things - see my comments below.
>> ...
>> It would be nice to have a comment here explaining what we're doing.
>> Also I don't think we need to copy the refname so it would be more
>> efficient to use refs_resolve_ref_unsafe().
>
> Looking at this again we cannot use refs_resolve_ref_unsafe() because
> the result would be overwritten by the call to refs_resolve_refdup() in
> branch_checked_out().
Makes sense. Thanks for raising a possible alternative and then
clarifying that it is not quite workable.
>>> + /*
>>> + * If the branch is the current HEAD, then it will be
>>> + * updated by the default rebase behavior.
>>> + */
>>> + if (head_ref && !strcmp(head_ref, decoration->name)) {
>>> + free(resolved_ref);
>>> decoration = decoration->next;
>>> continue;
>>> }
>>
>> Then we check to see if the decoration matches HEAD which we used to do
>> above - I'm not clear why we have moved this check.
>
> Should we be using "resolved_ref" instead of "decoration->name"? That
> would explain why this was moved and would makes sense as we resolve
> symrefs when reading HEAD. When HEAD points outside "refs/heads/" we'd
> then skip updating any symrefs under "refs/heads/" that pointed to the
> same ref as HEAD.
Yeah, decoration is very much end-user facing and if we can make
behavioural decision based on a more stable resolved_ref that would
make it easier to reason about.
But stepping back a bit, is having a HEAD that is a symref and
points outside "refs/heads/" an invalid state? Why are we catering
to such a configuration to begin with?
|
There was a status update in the "Cooking" section about the branch 'git rebase --update-refs' has been taught to resolve local branch symrefs to their referents before queuing updates, ensuring aliases of the current branch are skipped and duplicate updates are avoided to prevent failures when branch aliases are present. Waiting for a response. cf. <f982c386-e329-4ab0-b695-e540bcb9de3d@gmail.com> cf. <00e529b6-7ae7-463f-a4b3-0991e9411aba@gmail.com> source: <pull.2126.v3.git.1784708107.gitgitgadget@gmail.com> |
|
User |
git rebase --update-refs can finish rewriting the current branch and then fail while updating a local branch that is a symbolic ref. This can happen during a default-branch rename where refs/heads/main points at refs/heads/master while users migrate. The problem is a partially applied ref update: the main rebase has already succeeded when the later ref update fails. The sequencer queues updates from local branch decorations. Commit 106b688 (rebase: ignore non-branch update-refs) filters out decorations such as HEAD and tags. A branch symref is still a local branch decoration, but refs_update_ref() dereferences it, so an alias to another branch duplicates the concrete branch update. Resolve local branch decorations before queuing them. Skip symrefs whose targets are under refs/heads/ so that only the concrete branch update is queued. Keep an owned copy of the resolved HEAD. The previous pointer could be overwritten while checking another decoration, causing the current branch to be mistaken for a branch checked out in a worktree and added as a comment to the todo list. Skip the current branch before checked-out handling so the normal rebase remains its only updater. Compare the resolved names so this also covers a local branch alias whose non-branch target is HEAD. This prevents a successful rebase from being followed by a failed, partially applied ref update while preserving each alias as a symref. Signed-off-by: Son Luong Ngoc <sluongng@gmail.com>
A local branch symbolic ref may point outside refs/heads/. Such an alias cannot be skipped like a branch-to-branch alias because its concrete target ref is absent from the local branch decoration list. However, queuing each alias independently can update the same target ref more than once and make the second compare-and-swap fail. A reservation from another worktree can also name either an alias or its resolved target ref, so checking only one form can miss an in-progress update. Fix these cases by checking both the literal alias and its resolved target ref against checked-out reservations. Deduplicate updates by target ref. Also reserve both forms when loading another worktree's update-refs state. This makes different aliases honor the same in-progress update. This keeps non-branch symrefs supported without allowing duplicate or cross-worktree ref updates. Signed-off-by: Son Luong Ngoc <sluongng@gmail.com>
sluongng
force-pushed
the
sl/rebase-update-refs-symrefs
branch
from
July 26, 2026 08:37
a653f56 to
34e27ba
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
git rebase --update-refs can finish rewriting the current branch and
then fail while updating a local branch that is a symbolic ref. This can
happen during a default-branch rename where refs/heads/main points at
refs/heads/master while users migrate. The failure leaves refs partially
updated even though the main rebase has succeeded.
Resolve local branch decorations before adding update-ref commands. The
first patch skips aliases whose targets are other branches and preserves
the existing handling of the current branch. The second patch keeps
aliases to non-branch refs supported while preventing duplicate and
cross-worktree updates to their resolved targets.
Changes since v3:
branch alias that shares the current non-branch target.
branch_checked_out()refactor in the non-branch safeguardspatch where it is needed.
and scratch-buffer clobbering it prevents.
intentionally strips todo comments.
The focused t3400 and t3404 test suites pass with both the files and
reftable backends.
cc: "Kristoffer Haugsbakk" kristofferhaugsbakk@fastmail.com
cc: Phillip Wood phillip.wood123@gmail.com
cc: Erik Cervin-Edin erik@cervined.in