threads / patch / 66847

patchfor-each-ref: avoid segfault on --no-format

Subject: [PATCH] for-each-ref: avoid segfault on --no-format

## tl;dr

The crash dates to 2007, and reviewers are weighing whether --no-format should restore the default format or be rejected. Read the story.

replies: 5people: 3as markdown or json

aeron-gh· Nov 13, 2026, 16:14 UTC · lore

The "--format" option is declared with OPT_STRING(), so the parse-options machinery also accepts "--no-format" and sets the format string to NULL. The default format is assigned before the command line is parsed, so nothing puts it back, and the NULL is dereferenced in verify_ref_format():

    $ git for-each-ref --no-format
    Segmentation fault

"git refs list" shares the code and crashes the same way. The problem dates back to c3428da87f (Make builtin-for-each-ref.c use parse-opts., 2007-10-13).

"git branch" and "git tag" have the same option and fall back to their default format when it is NULL. Do the same here, and assign the default after parsing the command line when no format is left. "--no-format" then cancels an earlier "--format" and gives the default output.

Marking the option with PARSE_OPT_NONEG would also stop the crash, but then "--no-format" would work for "git branch" and "git tag" and be an error here.

Signed-off-by: aeron-gh <agab0323@gmail.com>
---
 builtin/for-each-ref.c  | 4 ++--
 t/for-each-ref-tests.sh | 6 ++++++
 2 files changed, 8 insertions(+), 2 deletions(-)
Show changes to 2 files +8 −2

builtin/for-each-ref.c, t/for-each-ref-tests.sh

diff --git a/builtin/for-each-ref.c b/builtin/for-each-ref.c
index 4a2fc421db..4390f3f5e9 100644
--- a/builtin/for-each-ref.c
+++ b/builtin/for-each-ref.c
@@ -52,14 +52,14 @@ int for_each_ref_core(int argc, const char **argv, const char *prefix, struct re
 		OPT_END(),
 	};
 
-	format.format = "%(objectname) %(objecttype)\t%(refname)";
-
 	repo_config(repo, git_default_config, NULL);
 
 	/* Set default (refname) sorting */
 	string_list_append(&sorting_options, "refname");
 
 	parse_options(argc, argv, prefix, opts, usage, 0);
+	if (!format.format)
+		format.format = "%(objectname) %(objecttype)\t%(refname)";
 	if (format.array_opts.max_count < 0) {
 		error("invalid --count argument: `%d'", format.array_opts.max_count);
 		usage_with_options(usage, opts);
diff --git a/t/for-each-ref-tests.sh b/t/for-each-ref-tests.sh
index 4991a82b91..3e0108a7db 100644
--- a/t/for-each-ref-tests.sh
+++ b/t/for-each-ref-tests.sh
@@ -285,6 +285,12 @@ test_expect_success 'Check invalid atoms names are errors' '
 	test_must_fail ${git_for_each_ref} --format="%(INVALID)" refs/heads
 '
 
+test_expect_success '--no-format restores the default format' '
+	${git_for_each_ref} refs/heads >expect &&
+	${git_for_each_ref} --format="%(refname)" --no-format refs/heads >actual &&
+	test_cmp expect actual
+'
+
 test_expect_success 'Check format specifiers are ignored in naming date atoms' '
 	${git_for_each_ref} --format="%(authordate)" refs/heads &&
 	${git_for_each_ref} --format="%(authordate:default) %(authordate)" refs/heads &&
-- 
2.56.0
Junio C Hamano· Nov 13, 2026, 22:10 UTC · re: aeron-gh · lore

Re: [PATCH] for-each-ref: avoid segfault on --no-format

aeron-gh <agab0323@gmail.com> writes:
Show 8 quoted lines
> The "--format" option is declared with OPT_STRING(), so the
> parse-options machinery also accepts "--no-format" and sets the
> format string to NULL. The default format is assigned before the
> command line is parsed, so nothing puts it back, and the NULL is
> dereferenced in verify_ref_format():
>
>     $ git for-each-ref --no-format
>     Segmentation fault

That is bad. PARSE_OPT_NONEG is necessary, I guess? Another plausible interpretation may be to exit without showing anything, but that does not sound like a very useful behaviour.

> Marking the option with PARSE_OPT_NONEG would also stop the crash,
> but then "--no-format" would work for "git branch" and "git tag"
> and be an error here.
Shouldn't "git branch --no-format" also error out?
> Signed-off-by: aeron-gh <agab0323@gmail.com>

Please check [[dco]] and [[real-name]] A few sections in Documentation/SubmittingPatches.

Thanks.
Junio C Hamano· Nov 13, 2026, 23:11 UTC · re: Junio C Hamano · lore

Re: [PATCH] for-each-ref: avoid segfault on --no-format

Junio C Hamano <gitster@pobox.com> writes:
Show 5 quoted lines
>> Marking the option with PARSE_OPT_NONEG would also stop the crash,
>> but then "--no-format" would work for "git branch" and "git tag"
>> and be an error here.
>
> Shouldn't "git branch --no-format" also error out?

Hmph, "git branch --no-format" seems to take it as an instruction to revert to the default format. Making "--no-format" mean the same thing in "for-each-ref" may make sense.

The implementation of "git branch" side is quite clean; if format.format has a value, use it, or fill it with default. Perhaps mimicking it would give us the right behaviour? I didn't check.

Abhishek Agarwal· Nov 14, 2026, 04:53 UTC · re: Junio C Hamano · lore

Re: [PATCH] for-each-ref: avoid segfault on --no-format

Junio C Hamano <gitster@pobox.com> writes:
> The implementation of "git branch" side is quite clean; if
> format.format has a value, use it, or fill it with default.  Perhaps
> mimicking it would give us the right behaviour?  I didn't check.
Yes, the patch does the same thing.  After parse_options() it has
	if (!format.format)
		format.format = "%(objectname) %(objecttype)\t%(refname)";

so "--no-format" gives the default output, like it does for "git branch" and "git tag". The new test checks that "--format=<fmt> --no-format" prints the same as giving no option at all.

I will send a v2 with the name in the sign-off fixed. I do not plan any code change, unless u see something else.

Thanks.
Abhishek Agarwal· Nov 14, 2026, 18:37 UTC · re: aeron-gh · lore

[PATCH v2] for-each-ref: avoid segfault on --no-format

The "--format" option is declared with OPT_STRING(), so the parse-options machinery also accepts "--no-format" and sets the format string to NULL. The default format is assigned before the command line is parsed, so nothing puts it back, and the NULL is dereferenced in verify_ref_format():

    $ git for-each-ref --no-format
    Segmentation fault

"git refs list" shares the code and crashes the same way. The problem dates back to c3428da87f (Make builtin-for-each-ref.c use parse-opts., 2007-10-13).

"git branch" and "git tag" have the same option and fall back to their default format when it is NULL. Do the same here, and assign the default after parsing the command line when no format is left. "--no-format" then cancels an earlier "--format" and gives the default output.

Marking the option with PARSE_OPT_NONEG would also stop the crash, but then "--no-format" would work for "git branch" and "git tag" and be an error here.

Signed-off-by: Abhishek Agarwal <agab0323@gmail.com>
---
Notes:
    Changes since v1:
    - Use "Abhishek Agarwal" in the author and Signed-off-by lines
      instead of "aeron-gh".
    - No code change.
 builtin/for-each-ref.c  | 4 ++--
 t/for-each-ref-tests.sh | 6 ++++++
 2 files changed, 8 insertions(+), 2 deletions(-)
Show changes to 2 files +8 −2

builtin/for-each-ref.c, t/for-each-ref-tests.sh

diff --git a/builtin/for-each-ref.c b/builtin/for-each-ref.c
index 4a2fc421db..4390f3f5e9 100644
--- a/builtin/for-each-ref.c
+++ b/builtin/for-each-ref.c
@@ -52,14 +52,14 @@ int for_each_ref_core(int argc, const char **argv, const char *prefix, struct re
 		OPT_END(),
 	};
 
-	format.format = "%(objectname) %(objecttype)\t%(refname)";
-
 	repo_config(repo, git_default_config, NULL);
 
 	/* Set default (refname) sorting */
 	string_list_append(&sorting_options, "refname");
 
 	parse_options(argc, argv, prefix, opts, usage, 0);
+	if (!format.format)
+		format.format = "%(objectname) %(objecttype)\t%(refname)";
 	if (format.array_opts.max_count < 0) {
 		error("invalid --count argument: `%d'", format.array_opts.max_count);
 		usage_with_options(usage, opts);
diff --git a/t/for-each-ref-tests.sh b/t/for-each-ref-tests.sh
index 4991a82b91..3e0108a7db 100644
--- a/t/for-each-ref-tests.sh
+++ b/t/for-each-ref-tests.sh
@@ -285,6 +285,12 @@ test_expect_success 'Check invalid atoms names are errors' '
 	test_must_fail ${git_for_each_ref} --format="%(INVALID)" refs/heads
 '
 
+test_expect_success '--no-format restores the default format' '
+	${git_for_each_ref} refs/heads >expect &&
+	${git_for_each_ref} --format="%(refname)" --no-format refs/heads >actual &&
+	test_cmp expect actual
+'
+
 test_expect_success 'Check format specifiers are ignored in naming date atoms' '
 	${git_for_each_ref} --format="%(authordate)" refs/heads &&
 	${git_for_each_ref} --format="%(authordate:default) %(authordate)" refs/heads &&
-- 
2.56.0
Patrick Steinhardt· Nov 16, 2026, 06:45 UTC · re: Abhishek Agarwal · lore

Re: [PATCH v2] for-each-ref: avoid segfault on --no-format

On Sun, Nov 15, 2026 at 12:07:18AM +0530, Abhishek Agarwal wrote:
Show 18 quoted lines
> The "--format" option is declared with OPT_STRING(), so the
> parse-options machinery also accepts "--no-format" and sets the
> format string to NULL. The default format is assigned before the
> command line is parsed, so nothing puts it back, and the NULL is
> dereferenced in verify_ref_format():
> 
>     $ git for-each-ref --no-format
>     Segmentation fault
> 
> "git refs list" shares the code and crashes the same way. The
> problem dates back to c3428da87f (Make builtin-for-each-ref.c use
> parse-opts., 2007-10-13).
> 
> "git branch" and "git tag" have the same option and fall back to
> their default format when it is NULL. Do the same here, and assign
> the default after parsing the command line when no format is left.
> "--no-format" then cancels an earlier "--format" and gives the
> default output.

That behaviour is a bit on the weird side if you ask me, but so be it. It's definitely better compared to crashing.

> Marking the option with PARSE_OPT_NONEG would also stop the crash,
> but then "--no-format" would work for "git branch" and "git tag"
> and be an error here.
Yeah, being consistent is the better pick, I guess.
Show 21 quoted lines
> diff --git a/builtin/for-each-ref.c b/builtin/for-each-ref.c
> index 4a2fc421db..4390f3f5e9 100644
> --- a/builtin/for-each-ref.c
> +++ b/builtin/for-each-ref.c
> @@ -52,14 +52,14 @@ int for_each_ref_core(int argc, const char **argv, const char *prefix, struct re
>  		OPT_END(),
>  	};
>  
> -	format.format = "%(objectname) %(objecttype)\t%(refname)";
> -
>  	repo_config(repo, git_default_config, NULL);
>  
>  	/* Set default (refname) sorting */
>  	string_list_append(&sorting_options, "refname");
>  
>  	parse_options(argc, argv, prefix, opts, usage, 0);
> +	if (!format.format)
> +		format.format = "%(objectname) %(objecttype)\t%(refname)";
>  	if (format.array_opts.max_count < 0) {
>  		error("invalid --count argument: `%d'", format.array_opts.max_count);
>  		usage_with_options(usage, opts);
Makes sense.
Show 13 quoted lines
> diff --git a/t/for-each-ref-tests.sh b/t/for-each-ref-tests.sh
> index 4991a82b91..3e0108a7db 100644
> --- a/t/for-each-ref-tests.sh
> +++ b/t/for-each-ref-tests.sh
> @@ -285,6 +285,12 @@ test_expect_success 'Check invalid atoms names are errors' '
>  	test_must_fail ${git_for_each_ref} --format="%(INVALID)" refs/heads
>  '
>  
> +test_expect_success '--no-format restores the default format' '
> +	${git_for_each_ref} refs/heads >expect &&
> +	${git_for_each_ref} --format="%(refname)" --no-format refs/heads >actual &&
> +	test_cmp expect actual
> +'

And the test looks obviously good to me, as well. So this patch looks good to me overall, thanks!

Patrick

← back to recent threads