threads / patch / 66832

patchref-filter: fix double free when an object is missing

Subject: [PATCH] ref-filter: fix double free when an object is missing

## tl;dr

A ref pointing at a missing object could abort git-for-each-ref with a double free instead of a proper error, a regression from the on-demand object parsing change. Read the story.

replies: 3people: 3as markdown or json

Mustafa Bayar via B4 Relay· Nov 12, 2026, 09:48 UTC · lore
From: Mustafa Bayar <mbayar@gitlab.com>

When a ref points to a missing object, git-for-each-ref(1) may abort with a double free instead of reporting the missing object:

    free(): double free detected in tcache 2

get_object() frees `oi->content` but leaves the pointer unchanged. Since the expand_data structures are reused across refs, a later call can start with `content` pointing to a freed buffer. Then, if the object lookup in that call fails before it overwrites the pointer, the cleanup path frees the buffer a second time.

a29e2e8fe (ref-filter: parse objects on demand, 2025-10-23) introduced this by making object lookup failures jump to the cleanup label. bea37f1d6 (ref-filter: fix stale parsed objects, 2025-11-04) later fixed a related stale-state bug by resetting `maybe_object`.

Similar to `maybe_object`, reset `content` at the start of get_object(). Add a regression test that exercises this path.

Signed-off-by: Mustafa Bayar <mbayar@gitlab.com>
---
This is based on maint, since the bug first appeared in v2.53.0.

We ran into this on GitLab.com, where a repository with a tag pointing to a missing object made `git for-each-ref` abort instead of reporting the missing object.

A note on the test: the double free only happens when a valid annotated
tag is processed before the broken ref, so that an earlier get_object()
call has already freed its content buffer. The test therefore names the
refs `annotated` and `missing`, which sort in that order.
---
 ref-filter.c                   |  1 +
 t/t6301-for-each-ref-errors.sh | 15 +++++++++++++++
 2 files changed, 16 insertions(+)
Show changes to 2 files +16 −0

ref-filter.c, t/t6301-for-each-ref-errors.sh

diff --git a/ref-filter.c b/ref-filter.c
index bdf54f6f..6c4df521 100644
--- a/ref-filter.c
+++ b/ref-filter.c
@@ -2349,6 +2349,7 @@ static int get_object(struct ref_array_item *ref, int deref,
 	int ret;
 
 	oi->maybe_object = NULL;
+	oi->content = NULL;
 
 	if (oi->info.contentp) {
 		/* We need to know that to use parse_object_buffer properly */
diff --git a/t/t6301-for-each-ref-errors.sh b/t/t6301-for-each-ref-errors.sh
index 72b27c8b..e955f66a 100755
--- a/t/t6301-for-each-ref-errors.sh
+++ b/t/t6301-for-each-ref-errors.sh
@@ -52,6 +52,21 @@ test_expect_success 'Missing objects are reported correctly' '
 	test_must_be_empty brief-err
 '
 
+test_expect_success 'missing objects after a tag are reported correctly' '
+	test_when_finished "rm -rf repo" &&
+	git init repo &&
+	(
+		cd repo &&
+		test_commit --no-tag initial &&
+		git tag -a -m annotated annotated &&
+		test-tool ref-store main update-ref msg refs/tags/missing "$MISSING" "$ZERO_OID" REF_SKIP_OID_VERIFICATION &&
+		git rev-parse HEAD >expect &&
+		test_must_fail git for-each-ref --format="%(*objectname)" refs/tags/ >actual 2>err &&
+		test_cmp expect actual &&
+		test_grep "fatal: missing object $MISSING for refs/tags/missing" err
+	)
+'
+
 test_expect_success 'missing ancestors are reported by contains filters' '
 	test_when_finished "git update-ref -d refs/heads/missing-parent" &&
 	{

---
base-commit: 03b41523c576f26ad34085378717a0c1b287470f
change-id: 20261112-mb-ref-filter-missing-object-double-free-maint-9eadbfe11e71
Karthik Nayak· Nov 12, 2026, 13:44 UTC · re: Mustafa Bayar via B4 Relay · lore

Re: [PATCH] ref-filter: fix double free when an object is missing

Mustafa Bayar via B4 Relay <devnull+mbayar.gitlab.com@kernel.org> writes:

Show 21 quoted lines
> From: Mustafa Bayar <mbayar@gitlab.com>
>
> When a ref points to a missing object, git-for-each-ref(1) may abort
> with a double free instead of reporting the missing object:
>
>     free(): double free detected in tcache 2
>
> get_object() frees `oi->content` but leaves the pointer unchanged. Since
> the expand_data structures are reused across refs, a later call can
> start with `content` pointing to a freed buffer. Then, if the object
> lookup in that call fails before it overwrites the pointer, the cleanup
> path frees the buffer a second time.
>
> a29e2e8fe (ref-filter: parse objects on demand, 2025-10-23) introduced
> this by making object lookup failures jump to the cleanup label.
> bea37f1d6 (ref-filter: fix stale parsed objects, 2025-11-04) later fixed
> a related stale-state bug by resetting `maybe_object`.
>
> Similar to `maybe_object`, reset `content` at the start of get_object().
> Add a regression test that exercises this path.
>

I had only reviewed this patch internally and we did some iterations there [1], so mostly just rubber-stamping here that it looks good to me.

Thanks for working on this!
[snip]
[1]: https://gitlab.com/gitlab-org/git/-/merge_requests/737
Junio C Hamano· Nov 12, 2026, 20:29 UTC · re: Karthik Nayak · lore

Re: [PATCH] ref-filter: fix double free when an object is missing

Karthik Nayak <karthik.188@gmail.com> writes:
Show 7 quoted lines
>> Similar to `maybe_object`, reset `content` at the start of get_object().
>> Add a regression test that exercises this path.
>
> I had only reviewed this patch internally and we did some iterations
> there [1], so mostly just rubber-stamping here that it looks good to me.
>
> Thanks for working on this!

I take the above with "only" -> "already". Thanks for writing and reviewing.

Will queue and mark the topic ready for 'next'. The patch is very well written, together with historical reference and all.

Thanks.
Karthik Nayak· Nov 12, 2026, 23:13 UTC · re: Junio C Hamano · lore

Re: [PATCH] ref-filter: fix double free when an object is missing

Junio C Hamano <gitster@pobox.com> writes:
Show 14 quoted lines
> Karthik Nayak <karthik.188@gmail.com> writes:
>
>>> Similar to `maybe_object`, reset `content` at the start of get_object().
>>> Add a regression test that exercises this path.
>>
>> I had only reviewed this patch internally and we did some iterations
>> there [1], so mostly just rubber-stamping here that it looks good to me.
>>
>> Thanks for working on this!
>
>
> I take the above with "only" -> "already".  Thanks for writing and
> reviewing.
>
Sigh! that's what I meant :)
> Will queue and mark the topic ready for 'next'.  The patch is very
> well written, together with historical reference and all.
>
> Thanks.

← back to recent threads