[PATCH] read_gitfile_gently(): return non-repo path on error

[PATCH] read_gitfile_gently(): return non-repo path on error

[Jeff King]

This patch fixes a potential segfault when resolving a .git file that
points to an invalid path. The bug was introduced by 1dd27bfbfd (setup:
improve error diagnosis for invalid .git files, 2026-03-04).

In setup_git_directory_gently() we call read_gitfile_gently(), which may
return a numeric code to us on error. If die_on_error is set, we then
feed that code to read_gitfile_error_die(), which also wants the path to
the gitfile and, in the case of ERROR_NOT_A_REPO, the non-repo directory
that the gitfile pointed to.

But we don't have that pointed-to directory available, so we just pass
NULL. That ends up calling die("not a git repository: %s", NULL). This
may crash, though on many systems (like glibc) it will just print
"(null)". So even if we don't crash, we're generating nonsense output.

The problem comes from 1dd27bfbfd. Before that, when die_on_error was
set we'd pass NULL to read_gitfile_gently()'s return_error_code
parameter, which means it would call read_gitfile_error_die() itself.
And it _does_ have that pointed-to directory as a string, and correctly
passes it.  But since 1dd27bfbfd, we always get the numeric error code
back from read_gitfile_gently(), and then decide whether to call
read_gitfile_error_die() in the caller. And since we don't have the
"dir" parameter, we just pass NULL.

Unfortunately the fix is not a simple matter of passing the string to
the right function. We have to get it out of read_gitfile_gently() in
the first place, which means we have to return it as another
out-parameter. And because it involves allocating memory, we can't just
do so unconditionally; callers need to be ready to free it after
handling the error.

I've tried to make the minimally-invasive fix here:

  1. We only copy the string when we hit READ_GITFILE_ERR_NOT_A_REPO,
     so other error codes don't have to worry about freeing it.

  2. We'll turn read_gitfile_gently() into a wrapper which passes NULL
     by default, leaving other callers unaffected.

The result is kind of gross. There's an extra layer of macro
indirection, and the validity of the string is subtly tied to the
NOT_A_REPO error. A cleaner solution might be an error struct that
couples the code and the output string together, along with a function
to free the error struct. But then all callers would have to be modified
to call the free function. Alternatively, we could perhaps put a
large-ish fixed-size buffer in the struct, though that means potential
truncation and a larger stack footprint in each caller (even when they
don't have see an error).

So I've left that as possible work for the future, or maybe never. Some
of this gross-ness was already there. For example, the only other caller
of read_gitfile_error_die() is in submodule.c, and it also passes NULL
for the "dir" parameter. But it does so only when the code is not
NOT_A_REPO! So it is depending on the same subtle connection to avoid
triggering the bug.

There's an existing test in t0002 which triggers this case, but we
didn't notice the problem because it checks only that we said "not a
repository", and not the full string. So if we print "(null)" it is
happy. It will probably crash on some non-glibc platforms, but nobody
seems to have reported it yet (the breakage is recent-ish as of v2.54).
I'm also somewhat surprised that building with ASan/UBSan doesn't catch
this, but it doesn't seem to (and I found an open issue with somebody
asking for it to be implemented in the sanitizers).

We can beef up the test by checking for the full string, which does
demonstrate the bug.

Signed-off-by: Jeff King <peff@peff.net>
---
Two other points of interest.

One, I'm not sure how useful printing the pointed-to directory is. We
_could_ just say:

  fatal: gitfile does not point to a valid repository: /path/to/.git

which is enough for somebody to investigate themselves. That would
certainly make the patch smaller.

And two, I ran into this running doc-diff:

  $ ./doc-diff HEAD^ HEAD
  fatal: not a git repository: (null)

The correct output (which this patch produces) is:

  fatal: not a git repository: /home/peff/compile/git/.git/worktrees/worktree3

And indeed, that path is missing. But why? I feel like I've run into
this same problem occasionally over the last year or so, but never
before. Did we get more aggressive about removing worktrees at some
point? I haven't been able to reproduce whatever is killing off the
worktree directory, and by the time I see the error it is long gone.

Anyway, that's not strictly related to this bug, but just how I
happened across it.

 setup.c            | 19 +++++++++++++++----
 setup.h            |  3 ++-
 t/t0002-gitfile.sh |  2 +-
 3 files changed, 18 insertions(+), 6 deletions(-)

diff --git a/setup.c b/setup.c
index 075bf89fa9..2df6fbf595 100644
--- a/setup.c
+++ b/setup.c
@@ -955,8 +955,14 @@ void read_gitfile_error_die(int error_code, const char *path, const char *dir)
  * will be set to an error code and NULL will be returned. If
  * return_error_code is NULL the function will die instead (for most
  * cases).
+ *
+ * If the code is READ_GITFILE_ERR_NOT_A_REPO and return_error_dir is
+ * non-NULL, the directory to which the gitfile points will be returned
+ * there. The caller is responsible for freeing the resulting string.
  */
-const char *read_gitfile_gently(const char *path, int *return_error_code)
+const char *read_gitfile_gently_with_error_dir(const char *path,
+					       int *return_error_code,
+					       char **return_error_dir)
 {
 	const int max_file_size = 1 << 20;  /* 1MB */
 	int error_code = 0;
@@ -1021,6 +1027,8 @@ const char *read_gitfile_gently(const char *path, int *return_error_code)
 	}
 	if (!is_git_directory(dir)) {
 		error_code = READ_GITFILE_ERR_NOT_A_REPO;
+		if (return_error_dir)
+			*return_error_dir = xstrdup(dir);
 		goto cleanup_return;
 	}
 
@@ -1613,11 +1621,12 @@ static enum discovery_result setup_git_directory_gently_1(struct strbuf *dir,
 		int offset = dir->len, error_code = 0;
 		char *gitdir_path = NULL;
 		char *gitfile = NULL;
+		char *error_dst = NULL;
 
 		if (offset > min_offset)
 			strbuf_addch(dir, '/');
 		strbuf_addstr(dir, DEFAULT_GIT_DIR_ENVIRONMENT);
-		gitdirenv = read_gitfile_gently(dir->buf, &error_code);
+		gitdirenv = read_gitfile_gently_with_error_dir(dir->buf, &error_code, &error_dst);
 		if (!gitdirenv) {
 			switch (error_code) {
 			case READ_GITFILE_ERR_MISSING:
@@ -1641,9 +1650,11 @@ static enum discovery_result setup_git_directory_gently_1(struct strbuf *dir,
 					return GIT_DIR_INVALID_GITFILE;
 			default:
 				if (die_on_error)
-					read_gitfile_error_die(error_code, dir->buf, NULL);
-				else
+					read_gitfile_error_die(error_code, dir->buf, error_dst);
+				else {
+					free(error_dst);
 					return GIT_DIR_INVALID_GITFILE;
+				}
 			}
 		} else {
 			gitfile = xstrdup(dir->buf);
diff --git a/setup.h b/setup.h
index 7878c9d267..65f55d5268 100644
--- a/setup.h
+++ b/setup.h
@@ -39,7 +39,8 @@ int is_nonbare_repository_dir(struct strbuf *path);
 #define READ_GITFILE_ERR_MISSING 9
 #define READ_GITFILE_ERR_IS_A_DIR 10
 void read_gitfile_error_die(int error_code, const char *path, const char *dir);
-const char *read_gitfile_gently(const char *path, int *return_error_code);
+const char *read_gitfile_gently_with_error_dir(const char *path, int *return_error_code, char **return_error_dir);
+#define read_gitfile_gently(path, err) read_gitfile_gently_with_error_dir((path), (err), NULL)
 #define read_gitfile(path) read_gitfile_gently((path), NULL)
 const char *resolve_gitdir_gently(const char *suspect, int *return_error_code);
 #define resolve_gitdir(path) resolve_gitdir_gently((path), NULL)
diff --git a/t/t0002-gitfile.sh b/t/t0002-gitfile.sh
index dfbcdddbcc..6967e12b9f 100755
--- a/t/t0002-gitfile.sh
+++ b/t/t0002-gitfile.sh
@@ -27,7 +27,7 @@ test_expect_success 'bad setup: invalid .git file format' '
 test_expect_success 'bad setup: invalid .git file path' '
 	echo "gitdir: $REAL.not" >.git &&
 	test_must_fail git rev-parse 2>.err &&
-	test_grep "not a git repository" .err
+	test_grep "not a git repository: $REAL.not" .err
 '
 
 test_expect_success 'final setup + check rev-parse --git-dir' '
-- 
2.54.0.682.g2f9b59d445

Re: [PATCH] read_gitfile_gently(): return non-repo path on error

[Junio C Hamano]

Jeff King <peff@peff.net> writes:

> I've tried to make the minimally-invasive fix here:
>
>   1. We only copy the string when we hit READ_GITFILE_ERR_NOT_A_REPO,
>      so other error codes don't have to worry about freeing it.
>
>   2. We'll turn read_gitfile_gently() into a wrapper which passes NULL
>      by default, leaving other callers unaffected.

Nice, probably.  I do not know what to feel about the first point,
though, as it burdens those who add new callers in the future more.

> The result is kind of gross. There's an extra layer of macro
> indirection, and the validity of the string is subtly tied to the
> NOT_A_REPO error. A cleaner solution might be an error struct that
> couples the code and the output string together, along with a function
> to free the error struct. But then all callers would have to be modified
> to call the free function. Alternatively, we could perhaps put a
> large-ish fixed-size buffer in the struct, though that means potential
> truncation and a larger stack footprint in each caller (even when they
> don't have see an error).

None of thoese are particularly appetizing ;-).

> So I've left that as possible work for the future, or maybe never. Some
> of this gross-ness was already there. For example, the only other caller
> of read_gitfile_error_die() is in submodule.c, and it also passes NULL
> for the "dir" parameter. But it does so only when the code is not
> NOT_A_REPO! So it is depending on the same subtle connection to avoid
> triggering the bug.

Yup.  I can agree with this.

> ---
> Two other points of interest.
>
> One, I'm not sure how useful printing the pointed-to directory is. We
> _could_ just say:
>
>   fatal: gitfile does not point to a valid repository: /path/to/.git
>
> which is enough for somebody to investigate themselves. That would
> certainly make the patch smaller.

Thanks.  While reading the main explanation, it was the first thing
that came to me.

The implementation and the test are as expected in patches from you
and matches the intent explained in the log message exactly.

Thanks, will queue.

Re: [PATCH] read_gitfile_gently(): return non-repo path on error

[Jeff King]

On Tue, Jun 02, 2026 at 04:42:15PM +0900, Junio C Hamano wrote:

> > One, I'm not sure how useful printing the pointed-to directory is. We
> > _could_ just say:
> >
> >   fatal: gitfile does not point to a valid repository: /path/to/.git
> >
> > which is enough for somebody to investigate themselves. That would
> > certainly make the patch smaller.
> 
> Thanks.  While reading the main explanation, it was the first thing
> that came to me.

Here's what that looks like, for reference. It is nice and simple, if we
think the change in error message is acceptable. I hate to change
user-facing error messages because of internal code details, but I
really do wonder if the existing message is the most useful thing to
print in the first place.

diff --git a/setup.c b/setup.c
index 075bf89fa9..ed86671d84 100644
--- a/setup.c
+++ b/setup.c
@@ -920,7 +920,7 @@ int verify_repository_format(const struct repository_format *format,
 	return 0;
 }
 
-void read_gitfile_error_die(int error_code, const char *path, const char *dir)
+void read_gitfile_error_die(int error_code, const char *path)
 {
 	switch (error_code) {
 	case READ_GITFILE_ERR_NOT_A_FILE:
@@ -940,7 +940,8 @@ void read_gitfile_error_die(int error_code, const char *path, const char *dir)
 	case READ_GITFILE_ERR_NO_PATH:
 		die(_("no path in gitfile: %s"), path);
 	case READ_GITFILE_ERR_NOT_A_REPO:
-		die(_("not a git repository: %s"), dir);
+		die(_("gitfile does not point to a valid repository: %s"),
+		    path);
 	default:
 		BUG("unknown error code");
 	}
@@ -1031,7 +1032,7 @@ const char *read_gitfile_gently(const char *path, int *return_error_code)
 	if (return_error_code)
 		*return_error_code = error_code;
 	else if (error_code)
-		read_gitfile_error_die(error_code, path, dir);
+		read_gitfile_error_die(error_code, path);
 
 	free(buf);
 	return error_code ? NULL : path;
@@ -1641,7 +1642,7 @@ static enum discovery_result setup_git_directory_gently_1(struct strbuf *dir,
 					return GIT_DIR_INVALID_GITFILE;
 			default:
 				if (die_on_error)
-					read_gitfile_error_die(error_code, dir->buf, NULL);
+					read_gitfile_error_die(error_code, dir->buf);
 				else
 					return GIT_DIR_INVALID_GITFILE;
 			}
diff --git a/setup.h b/setup.h
index 7878c9d267..436aaa22c1 100644
--- a/setup.h
+++ b/setup.h
@@ -38,7 +38,7 @@ int is_nonbare_repository_dir(struct strbuf *path);
 #define READ_GITFILE_ERR_TOO_LARGE 8
 #define READ_GITFILE_ERR_MISSING 9
 #define READ_GITFILE_ERR_IS_A_DIR 10
-void read_gitfile_error_die(int error_code, const char *path, const char *dir);
+void read_gitfile_error_die(int error_code, const char *path);
 const char *read_gitfile_gently(const char *path, int *return_error_code);
 #define read_gitfile(path) read_gitfile_gently((path), NULL)
 const char *resolve_gitdir_gently(const char *suspect, int *return_error_code);
diff --git a/submodule.c b/submodule.c
index a939ff5072..c36732ca0b 100644
--- a/submodule.c
+++ b/submodule.c
@@ -2578,7 +2578,7 @@ void absorb_git_dir_into_superproject(const char *path,
 
 		if (err_code != READ_GITFILE_ERR_NOT_A_REPO)
 			/* We don't know what broke here. */
-			read_gitfile_error_die(err_code, path, NULL);
+			read_gitfile_error_die(err_code, path);
 
 		/*
 		* Maybe populated, but no git directory was found?
diff --git a/t/t0002-gitfile.sh b/t/t0002-gitfile.sh
index dfbcdddbcc..6356e9ec72 100755
--- a/t/t0002-gitfile.sh
+++ b/t/t0002-gitfile.sh
@@ -27,7 +27,7 @@ test_expect_success 'bad setup: invalid .git file format' '
 test_expect_success 'bad setup: invalid .git file path' '
 	echo "gitdir: $REAL.not" >.git &&
 	test_must_fail git rev-parse 2>.err &&
-	test_grep "not a git repository" .err
+	test_grep "gitfile does not point to a valid repository" .err
 '
 
 test_expect_success 'final setup + check rev-parse --git-dir' '

Re: [PATCH] read_gitfile_gently(): return non-repo path on error

[Patrick Steinhardt]

On Tue, Jun 02, 2026 at 02:11:59AM -0400, Jeff King wrote:
[snip]
> Two other points of interest.
> 
> One, I'm not sure how useful printing the pointed-to directory is. We
> _could_ just say:
> 
>   fatal: gitfile does not point to a valid repository: /path/to/.git
> 
> which is enough for somebody to investigate themselves. That would
> certainly make the patch smaller.

I have to agree that the patch is somewhat gross, and I myself don't
really see much of an issue to move to an error message like the above
if it ends up simplifying the logic.

> And two, I ran into this running doc-diff:
> 
>   $ ./doc-diff HEAD^ HEAD
>   fatal: not a git repository: (null)
> 
> The correct output (which this patch produces) is:
> 
>   fatal: not a git repository: /home/peff/compile/git/.git/worktrees/worktree3
> 
> And indeed, that path is missing. But why? I feel like I've run into
> this same problem occasionally over the last year or so, but never
> before. Did we get more aggressive about removing worktrees at some
> point? I haven't been able to reproduce whatever is killing off the
> worktree directory, and by the time I see the error it is long gone.

Both git-gc(1) and git-maintenance(1) prune orphaned worktrees that are
older than three months by default, which can be configured via
"gc.worktreePruneExpire". That logic has changed in 4dda60c9df (Merge
branch 'ps/maintenance-missing-tasks', 2025-05-15), which would kind of
match your timeline.

But rereading that patch series I cannot really see how it could result
in more aggressive pruning of worktrees. We used `git worktree prune
--expire <expiry>` before that series, and we still use that logic now.

Hum.

> diff --git a/setup.c b/setup.c
> index 075bf89fa9..2df6fbf595 100644
> --- a/setup.c
> +++ b/setup.c
> @@ -1641,9 +1650,11 @@ static enum discovery_result setup_git_directory_gently_1(struct strbuf *dir,
>  					return GIT_DIR_INVALID_GITFILE;
>  			default:
>  				if (die_on_error)
> -					read_gitfile_error_die(error_code, dir->buf, NULL);
> -				else
> +					read_gitfile_error_die(error_code, dir->buf, error_dst);
> +				else {
> +					free(error_dst);
>  					return GIT_DIR_INVALID_GITFILE;
> +				}

The `if` branch should also gain some curly braces here.

Patrick

Re: [PATCH] read_gitfile_gently(): return non-repo path on error

[Jeff King]

On Tue, Jun 02, 2026 at 10:36:34AM +0200, Patrick Steinhardt wrote:

> > The correct output (which this patch produces) is:
> > 
> >   fatal: not a git repository: /home/peff/compile/git/.git/worktrees/worktree3
> > 
> > And indeed, that path is missing. But why? I feel like I've run into
> > this same problem occasionally over the last year or so, but never
> > before. Did we get more aggressive about removing worktrees at some
> > point? I haven't been able to reproduce whatever is killing off the
> > worktree directory, and by the time I see the error it is long gone.
> 
> Both git-gc(1) and git-maintenance(1) prune orphaned worktrees that are
> older than three months by default, which can be configured via
> "gc.worktreePruneExpire". That logic has changed in 4dda60c9df (Merge
> branch 'ps/maintenance-missing-tasks', 2025-05-15), which would kind of
> match your timeline.
> 
> But rereading that patch series I cannot really see how it could result
> in more aggressive pruning of worktrees. We used `git worktree prune
> --expire <expiry>` before that series, and we still use that logic now.

Yeah, but this .git/worktrees/ directory shouldn't be pruned _at all_.
The worktree itself is still there (which is why I'm getting the error).
So perhaps there's a bug in checking that things are still there, or
perhaps something is corrupting .git/worktrees/*/gitdir.

Another option is "I moved my git checkout and the worktree prune
couldn't find the directory as an absolute path", but I'm sure I didn't
do that.

An even more exotic option is that I run Git's test suite a lot, and
very occasionally bugs in the test suite cause the script to escape the
trash directory. And some scripts do run "rm -r .git/worktrees". I find
it pretty unlikely for that to be the culprit though.

Oh well. I don't have any good leads, so I guess I'll see if it happens
again. But maybe now if somebody else sees it we can commiserate. :)

-Peff

Re: [PATCH] read_gitfile_gently(): return non-repo path on error

[Patrick Steinhardt]

On Thu, Jun 04, 2026 at 02:27:20AM -0400, Jeff King wrote:
> On Tue, Jun 02, 2026 at 10:36:34AM +0200, Patrick Steinhardt wrote:
> 
> > > The correct output (which this patch produces) is:
> > > 
> > >   fatal: not a git repository: /home/peff/compile/git/.git/worktrees/worktree3
> > > 
> > > And indeed, that path is missing. But why? I feel like I've run into
> > > this same problem occasionally over the last year or so, but never
> > > before. Did we get more aggressive about removing worktrees at some
> > > point? I haven't been able to reproduce whatever is killing off the
> > > worktree directory, and by the time I see the error it is long gone.
> > 
> > Both git-gc(1) and git-maintenance(1) prune orphaned worktrees that are
> > older than three months by default, which can be configured via
> > "gc.worktreePruneExpire". That logic has changed in 4dda60c9df (Merge
> > branch 'ps/maintenance-missing-tasks', 2025-05-15), which would kind of
> > match your timeline.
> > 
> > But rereading that patch series I cannot really see how it could result
> > in more aggressive pruning of worktrees. We used `git worktree prune
> > --expire <expiry>` before that series, and we still use that logic now.
> 
> Yeah, but this .git/worktrees/ directory shouldn't be pruned _at all_.
> The worktree itself is still there (which is why I'm getting the error).
> So perhaps there's a bug in checking that things are still there, or
> perhaps something is corrupting .git/worktrees/*/gitdir.

Oh, that sounds somewhat scary.

> Another option is "I moved my git checkout and the worktree prune
> couldn't find the directory as an absolute path", but I'm sure I didn't
> do that.
> 
> An even more exotic option is that I run Git's test suite a lot, and
> very occasionally bugs in the test suite cause the script to escape the
> trash directory. And some scripts do run "rm -r .git/worktrees". I find
> it pretty unlikely for that to be the culprit though.
> 
> Oh well. I don't have any good leads, so I guess I'll see if it happens
> again. But maybe now if somebody else sees it we can commiserate. :)

I'll certainly be on the watchout.

Patrick

Re: [PATCH] read_gitfile_gently(): return non-repo path on error

[Junio C Hamano]

Patrick Steinhardt <ps@pks.im> writes:

> On Tue, Jun 02, 2026 at 02:11:59AM -0400, Jeff King wrote:
> [snip]
>> Two other points of interest.
>> 
>> One, I'm not sure how useful printing the pointed-to directory is. We
>> _could_ just say:
>> 
>>   fatal: gitfile does not point to a valid repository: /path/to/.git
>> 
>> which is enough for somebody to investigate themselves. That would
>> certainly make the patch smaller.
>
> I have to agree that the patch is somewhat gross, and I myself don't
> really see much of an issue to move to an error message like the above
> if it ends up simplifying the logic.

So we are in agreement among three of us that simplifying the code
to lose error message with dubious value would be a good way
forward.

Peff, can we have a formal [v2] then?

>> diff --git a/setup.c b/setup.c
>> index 075bf89fa9..2df6fbf595 100644
>> --- a/setup.c
>> +++ b/setup.c
>> @@ -1641,9 +1650,11 @@ static enum discovery_result setup_git_directory_gently_1(struct strbuf *dir,
>>  					return GIT_DIR_INVALID_GITFILE;
>>  			default:
>>  				if (die_on_error)
>> -					read_gitfile_error_die(error_code, dir->buf, NULL);
>> -				else
>> +					read_gitfile_error_die(error_code, dir->buf, error_dst);
>> +				else {
>> +					free(error_dst);
>>  					return GIT_DIR_INVALID_GITFILE;
>> +				}
>
> The `if` branch should also gain some curly braces here.

True.

Thanks.

[PATCH v2] read_gitfile(): simplify NOT_A_REPO error message

[Jeff King]

On Wed, Jun 10, 2026 at 06:01:03AM -0700, Junio C Hamano wrote:

> > I have to agree that the patch is somewhat gross, and I myself don't
> > really see much of an issue to move to an error message like the above
> > if it ends up simplifying the logic.
> 
> So we are in agreement among three of us that simplifying the code
> to lose error message with dubious value would be a good way
> forward.
> 
> Peff, can we have a formal [v2] then?

Here it is.

-- >8 --
Subject: read_gitfile(): simplify NOT_A_REPO error message

If a .git file is well-formed but points to a directory that is not
itself a valid repository, then we say:

  fatal: not a git repository: <pointed-to-repo>

without mentioning the .git file that pointed us there in the first
place. Doing so could better help the user understand the source of the
problem.

In theory the most helpful thing we could do is mention both paths,
like:

  gitfile '<gitfile>' points to invalid repository: <pointed-to-repo>

But there's another catch: when we generate the error, we don't always
know the pointed-to repository! This leads to a potential segfault.

The message comes from read_gitfile_error_die(). Originally we only
called that function from inside read_gitfile_gently(), passing in both
the gitfile path and the pointed-to path. But that changed in 1dd27bfbfd
(setup: improve error diagnosis for invalid .git files, 2026-03-04).
Since then, the caller in setup_git_directory_gently(), even if it wants
to die on error, always passes in the "return_error_code" flag, asking
the function to instead return a numeric error code. And then it calls
read_gitfile_error_die() itself, passing NULL for the pointed-to path.

If we get the READ_GITFILE_ERR_NOT_A_REPO code, we form a message using
that NULL pointer, and either segfault or get garbage like "not a git
repository: (null)", depending on the platform.

We could fix this by having the function pass out both the numeric error
code and the pointed-to path. But that creates a new headache: we have
to allocate that string on the heap and pass ownership back to the
caller. So now every caller has to be aware of it (and either free the
result, or signal that they are not interested by using an extra
parameter).

Instead, let's just drop the pointed-to path from the error message
entirely, and mention only the gitfile. This fixes the NULL dereference
without introducing any more complexity. The user-facing error message
is not as detailed as it could be, but is better than the original.
Since it mentions the gitfile, a user investigating the situation can
look there to find the pointed-to path (whereas you could not go the
other way from the original message).

There's an existing test in t0002 which triggers this case, but we
didn't notice the problem because it checks only that we said "not a
repository", and not the full string. So if we print "(null)" it is
happy. It will probably crash on some non-glibc platforms, but nobody
seems to have reported it yet (the breakage is recent-ish as of v2.54).
I'm also somewhat surprised that building with ASan/UBSan doesn't catch
this, but it doesn't seem to (and I found an open issue with somebody
asking for NULL printf checks to be implemented in the sanitizers).

We'll tweak the test to match the new error, but there's no need to beef
it up further, since we're not showing the pointed-to path at all.

Signed-off-by: Jeff King <peff@peff.net>
---
 setup.c            | 9 +++++----
 setup.h            | 2 +-
 submodule.c        | 2 +-
 t/t0002-gitfile.sh | 2 +-
 4 files changed, 8 insertions(+), 7 deletions(-)

diff --git a/setup.c b/setup.c
index b4652651df..b1d9249d91 100644
--- a/setup.c
+++ b/setup.c
@@ -917,7 +917,7 @@ int verify_repository_format(const struct repository_format *format,
 	return 0;
 }
 
-void read_gitfile_error_die(int error_code, const char *path, const char *dir)
+void read_gitfile_error_die(int error_code, const char *path)
 {
 	switch (error_code) {
 	case READ_GITFILE_ERR_NOT_A_FILE:
@@ -937,7 +937,8 @@ void read_gitfile_error_die(int error_code, const char *path, const char *dir)
 	case READ_GITFILE_ERR_NO_PATH:
 		die(_("no path in gitfile: %s"), path);
 	case READ_GITFILE_ERR_NOT_A_REPO:
-		die(_("not a git repository: %s"), dir);
+		die(_("gitfile does not point to a valid repository: %s"),
+		    path);
 	default:
 		BUG("unknown error code");
 	}
@@ -1028,7 +1029,7 @@ const char *read_gitfile_gently(const char *path, int *return_error_code)
 	if (return_error_code)
 		*return_error_code = error_code;
 	else if (error_code)
-		read_gitfile_error_die(error_code, path, dir);
+		read_gitfile_error_die(error_code, path);
 
 	free(buf);
 	return error_code ? NULL : path;
@@ -1629,7 +1630,7 @@ static enum discovery_result setup_git_directory_gently_1(struct strbuf *dir,
 					return GIT_DIR_INVALID_GITFILE;
 			default:
 				if (die_on_error)
-					read_gitfile_error_die(error_code, dir->buf, NULL);
+					read_gitfile_error_die(error_code, dir->buf);
 				else
 					return GIT_DIR_INVALID_GITFILE;
 			}
diff --git a/setup.h b/setup.h
index 705d1d6ff7..df8c93687a 100644
--- a/setup.h
+++ b/setup.h
@@ -38,7 +38,7 @@ int is_nonbare_repository_dir(struct strbuf *path);
 #define READ_GITFILE_ERR_TOO_LARGE 8
 #define READ_GITFILE_ERR_MISSING 9
 #define READ_GITFILE_ERR_IS_A_DIR 10
-void read_gitfile_error_die(int error_code, const char *path, const char *dir);
+void read_gitfile_error_die(int error_code, const char *path);
 const char *read_gitfile_gently(const char *path, int *return_error_code);
 #define read_gitfile(path) read_gitfile_gently((path), NULL)
 const char *resolve_gitdir_gently(const char *suspect, int *return_error_code);
diff --git a/submodule.c b/submodule.c
index fd91201a92..93d0361072 100644
--- a/submodule.c
+++ b/submodule.c
@@ -2579,7 +2579,7 @@ void absorb_git_dir_into_superproject(const char *path,
 
 		if (err_code != READ_GITFILE_ERR_NOT_A_REPO)
 			/* We don't know what broke here. */
-			read_gitfile_error_die(err_code, path, NULL);
+			read_gitfile_error_die(err_code, path);
 
 		/*
 		* Maybe populated, but no git directory was found?
diff --git a/t/t0002-gitfile.sh b/t/t0002-gitfile.sh
index dfbcdddbcc..6356e9ec72 100755
--- a/t/t0002-gitfile.sh
+++ b/t/t0002-gitfile.sh
@@ -27,7 +27,7 @@ test_expect_success 'bad setup: invalid .git file format' '
 test_expect_success 'bad setup: invalid .git file path' '
 	echo "gitdir: $REAL.not" >.git &&
 	test_must_fail git rev-parse 2>.err &&
-	test_grep "not a git repository" .err
+	test_grep "gitfile does not point to a valid repository" .err
 '
 
 test_expect_success 'final setup + check rev-parse --git-dir' '
-- 
2.55.0.rc0.342.g45e27e83e2

[PATCH v3] read_gitfile(): simplify NOT_A_REPO error message

[Jeff King]

On Tue, Jun 16, 2026 at 07:19:20AM -0400, Jeff King wrote:

> Here it is.
> 
> -- >8 --
> Subject: read_gitfile(): simplify NOT_A_REPO error message

<sigh> This triggered a failure in CI after passing tests locally.
Turned out to be a race in t7450.

Here's an update, with range-diff.

1:  daf7f99511 ! 1:  67d42141e9 read_gitfile(): simplify NOT_A_REPO error message
    @@ Commit message
         We'll tweak the test to match the new error, but there's no need to beef
         it up further, since we're not showing the pointed-to path at all.
     
    +    We also racily trigger this in t7450. During parallel cloning we might
    +    see one of several errors, including this one. And so we must update
    +    that message, too (you can otherwise find the failure pretty quickly by
    +    running t7450 with --stress).
    +
         Signed-off-by: Jeff King <peff@peff.net>
     
      ## setup.c ##
    @@ t/t0002-gitfile.sh: test_expect_success 'bad setup: invalid .git file format' '
      '
      
      test_expect_success 'final setup + check rev-parse --git-dir' '
    +
    + ## t/t7450-bad-git-dotfiles.sh ##
    +@@ t/t7450-bad-git-dotfiles.sh: test_expect_success 'git dirs of sibling submodules must not be nested' '
    + test_expect_success 'submodule git dir nesting detection must work with parallel cloning' '
    + 	test_must_fail git clone --recurse-submodules --jobs=2 nested clone_parallel 2>err &&
    + 	cat err &&
    +-	grep -E "(already exists|is inside git dir|not a git repository)" err &&
    ++	grep -E "(already exists|is inside git dir|does not point to a valid repository)" err &&
    + 	{
    + 		test_path_is_missing .git/modules/hippo/HEAD ||
    + 		test_path_is_missing .git/modules/hippo/hooks/HEAD

-- >8 --
Subject: [PATCH] read_gitfile(): simplify NOT_A_REPO error message

If a .git file is well-formed but points to a directory that is not
itself a valid repository, then we say:

  fatal: not a git repository: <pointed-to-repo>

without mentioning the .git file that pointed us there in the first
place. Doing so could better help the user understand the source of the
problem.

In theory the most helpful thing we could do is mention both paths,
like:

  gitfile '<gitfile>' points to invalid repository: <pointed-to-repo>

But there's another catch: when we generate the error, we don't always
know the pointed-to repository! This leads to a potential segfault.

The message comes from read_gitfile_error_die(). Originally we only
called that function from inside read_gitfile_gently(), passing in both
the gitfile path and the pointed-to path. But that changed in 1dd27bfbfd
(setup: improve error diagnosis for invalid .git files, 2026-03-04).
Since then, the caller in setup_git_directory_gently(), even if it wants
to die on error, always passes in the "return_error_code" flag, asking
the function to instead return a numeric error code. And then it calls
read_gitfile_error_die() itself, passing NULL for the pointed-to path.

If we get the READ_GITFILE_ERR_NOT_A_REPO code, we form a message using
that NULL pointer, and either segfault or get garbage like "not a git
repository: (null)", depending on the platform.

We could fix this by having the function pass out both the numeric error
code and the pointed-to path. But that creates a new headache: we have
to allocate that string on the heap and pass ownership back to the
caller. So now every caller has to be aware of it (and either free the
result, or signal that they are not interested by using an extra
parameter).

Instead, let's just drop the pointed-to path from the error message
entirely, and mention only the gitfile. This fixes the NULL dereference
without introducing any more complexity. The user-facing error message
is not as detailed as it could be, but is better than the original.
Since it mentions the gitfile, a user investigating the situation can
look there to find the pointed-to path (whereas you could not go the
other way from the original message).

There's an existing test in t0002 which triggers this case, but we
didn't notice the problem because it checks only that we said "not a
repository", and not the full string. So if we print "(null)" it is
happy. It will probably crash on some non-glibc platforms, but nobody
seems to have reported it yet (the breakage is recent-ish as of v2.54).
I'm also somewhat surprised that building with ASan/UBSan doesn't catch
this, but it doesn't seem to (and I found an open issue with somebody
asking for NULL printf checks to be implemented in the sanitizers).

We'll tweak the test to match the new error, but there's no need to beef
it up further, since we're not showing the pointed-to path at all.

We also racily trigger this in t7450. During parallel cloning we might
see one of several errors, including this one. And so we must update
that message, too (you can otherwise find the failure pretty quickly by
running t7450 with --stress).

Signed-off-by: Jeff King <peff@peff.net>
---
 setup.c                     | 9 +++++----
 setup.h                     | 2 +-
 submodule.c                 | 2 +-
 t/t0002-gitfile.sh          | 2 +-
 t/t7450-bad-git-dotfiles.sh | 2 +-
 5 files changed, 9 insertions(+), 8 deletions(-)

diff --git a/setup.c b/setup.c
index b4652651df..b1d9249d91 100644
--- a/setup.c
+++ b/setup.c
@@ -917,7 +917,7 @@ int verify_repository_format(const struct repository_format *format,
 	return 0;
 }
 
-void read_gitfile_error_die(int error_code, const char *path, const char *dir)
+void read_gitfile_error_die(int error_code, const char *path)
 {
 	switch (error_code) {
 	case READ_GITFILE_ERR_NOT_A_FILE:
@@ -937,7 +937,8 @@ void read_gitfile_error_die(int error_code, const char *path, const char *dir)
 	case READ_GITFILE_ERR_NO_PATH:
 		die(_("no path in gitfile: %s"), path);
 	case READ_GITFILE_ERR_NOT_A_REPO:
-		die(_("not a git repository: %s"), dir);
+		die(_("gitfile does not point to a valid repository: %s"),
+		    path);
 	default:
 		BUG("unknown error code");
 	}
@@ -1028,7 +1029,7 @@ const char *read_gitfile_gently(const char *path, int *return_error_code)
 	if (return_error_code)
 		*return_error_code = error_code;
 	else if (error_code)
-		read_gitfile_error_die(error_code, path, dir);
+		read_gitfile_error_die(error_code, path);
 
 	free(buf);
 	return error_code ? NULL : path;
@@ -1629,7 +1630,7 @@ static enum discovery_result setup_git_directory_gently_1(struct strbuf *dir,
 					return GIT_DIR_INVALID_GITFILE;
 			default:
 				if (die_on_error)
-					read_gitfile_error_die(error_code, dir->buf, NULL);
+					read_gitfile_error_die(error_code, dir->buf);
 				else
 					return GIT_DIR_INVALID_GITFILE;
 			}
diff --git a/setup.h b/setup.h
index 705d1d6ff7..df8c93687a 100644
--- a/setup.h
+++ b/setup.h
@@ -38,7 +38,7 @@ int is_nonbare_repository_dir(struct strbuf *path);
 #define READ_GITFILE_ERR_TOO_LARGE 8
 #define READ_GITFILE_ERR_MISSING 9
 #define READ_GITFILE_ERR_IS_A_DIR 10
-void read_gitfile_error_die(int error_code, const char *path, const char *dir);
+void read_gitfile_error_die(int error_code, const char *path);
 const char *read_gitfile_gently(const char *path, int *return_error_code);
 #define read_gitfile(path) read_gitfile_gently((path), NULL)
 const char *resolve_gitdir_gently(const char *suspect, int *return_error_code);
diff --git a/submodule.c b/submodule.c
index fd91201a92..93d0361072 100644
--- a/submodule.c
+++ b/submodule.c
@@ -2579,7 +2579,7 @@ void absorb_git_dir_into_superproject(const char *path,
 
 		if (err_code != READ_GITFILE_ERR_NOT_A_REPO)
 			/* We don't know what broke here. */
-			read_gitfile_error_die(err_code, path, NULL);
+			read_gitfile_error_die(err_code, path);
 
 		/*
 		* Maybe populated, but no git directory was found?
diff --git a/t/t0002-gitfile.sh b/t/t0002-gitfile.sh
index dfbcdddbcc..6356e9ec72 100755
--- a/t/t0002-gitfile.sh
+++ b/t/t0002-gitfile.sh
@@ -27,7 +27,7 @@ test_expect_success 'bad setup: invalid .git file format' '
 test_expect_success 'bad setup: invalid .git file path' '
 	echo "gitdir: $REAL.not" >.git &&
 	test_must_fail git rev-parse 2>.err &&
-	test_grep "not a git repository" .err
+	test_grep "gitfile does not point to a valid repository" .err
 '
 
 test_expect_success 'final setup + check rev-parse --git-dir' '
diff --git a/t/t7450-bad-git-dotfiles.sh b/t/t7450-bad-git-dotfiles.sh
index 8cc86522b2..69a17a9d13 100755
--- a/t/t7450-bad-git-dotfiles.sh
+++ b/t/t7450-bad-git-dotfiles.sh
@@ -350,7 +350,7 @@ test_expect_success 'git dirs of sibling submodules must not be nested' '
 test_expect_success 'submodule git dir nesting detection must work with parallel cloning' '
 	test_must_fail git clone --recurse-submodules --jobs=2 nested clone_parallel 2>err &&
 	cat err &&
-	grep -E "(already exists|is inside git dir|not a git repository)" err &&
+	grep -E "(already exists|is inside git dir|does not point to a valid repository)" err &&
 	{
 		test_path_is_missing .git/modules/hippo/HEAD ||
 		test_path_is_missing .git/modules/hippo/hooks/HEAD
-- 
2.55.0.rc0.346.g4c7eff6ddc

Re: [PATCH v3] read_gitfile(): simplify NOT_A_REPO error message

[Junio C Hamano]

Jeff King <peff@peff.net> writes:

> On Tue, Jun 16, 2026 at 07:19:20AM -0400, Jeff King wrote:
>
>> Here it is.

Thanks.

>     +@@ t/t7450-bad-git-dotfiles.sh: test_expect_success 'git dirs of sibling submodules must not be nested' '
>     + test_expect_success 'submodule git dir nesting detection must work with parallel cloning' '
>     + 	test_must_fail git clone --recurse-submodules --jobs=2 nested clone_parallel 2>err &&
>     + 	cat err &&
>     +-	grep -E "(already exists|is inside git dir|not a git repository)" err &&
>     ++	grep -E "(already exists|is inside git dir|does not point to a valid repository)" err &&

A few things.

 * Will we be happy to see only one of these possibilities, or do we
   expect to see these once for each kind?

 * a recently started in-flight topic tries to catch bare "grep" and
   fails until you write test_grep X-<.

> We also racily trigger this in t7450. During parallel cloning we might
> see one of several errors, including this one. And so we must update
> that message, too (you can otherwise find the failure pretty quickly by
> running t7450 with --stress).

Re: [PATCH v3] read_gitfile(): simplify NOT_A_REPO error message

[Jeff King]

On Tue, Jun 16, 2026 at 07:25:02AM -0700, Junio C Hamano wrote:

> >     +@@ t/t7450-bad-git-dotfiles.sh: test_expect_success 'git dirs of sibling submodules must not be nested' '
> >     + test_expect_success 'submodule git dir nesting detection must work with parallel cloning' '
> >     + 	test_must_fail git clone --recurse-submodules --jobs=2 nested clone_parallel 2>err &&
> >     + 	cat err &&
> >     +-	grep -E "(already exists|is inside git dir|not a git repository)" err &&
> >     ++	grep -E "(already exists|is inside git dir|does not point to a valid repository)" err &&
> 
> A few things.
> 
>  * Will we be happy to see only one of these possibilities, or do we
>    expect to see these once for each kind?

I imagine it is only one. This all comes from 9cf8547320 (clone: prevent
clashing git dirs when cloning submodule in parallel, 2024-01-28), and
it is expecting the nested path to cause a failure. Which failure I
guess depends on the racy ordering. If we create the inner one first,
then we probably get "already exists", and if the outer one, then "is
inside git dir". I don't know exactly what sequence yields the
NOT_A_REPO message.

But none of that is changing in this patch, just what the user-visible
text is for the NOT_A_REPO case.

I did briefly wonder if we might see "not a git repository" from a
_different_ code path, and need to catch it along with the new message.
But running successfully with --stress implies that we never see the old
one anymore.

>  * a recently started in-flight topic tries to catch bare "grep" and
>    fails until you write test_grep X-<.

Yeah. This will create a merge conflict for you, but hopefully the
resolution should be obvious. I don't think it makes sense to fix here,
as it's orthogonal to the purpose of the patch.

-Peff

Re: [PATCH v3] read_gitfile(): simplify NOT_A_REPO error message

[Junio C Hamano]

Jeff King <peff@peff.net> writes:

> On Tue, Jun 16, 2026 at 07:25:02AM -0700, Junio C Hamano wrote:
>
>> >     +@@ t/t7450-bad-git-dotfiles.sh: test_expect_success 'git dirs of sibling submodules must not be nested' '
>> >     + test_expect_success 'submodule git dir nesting detection must work with parallel cloning' '
>> >     + 	test_must_fail git clone --recurse-submodules --jobs=2 nested clone_parallel 2>err &&
>> >     + 	cat err &&
>> >     +-	grep -E "(already exists|is inside git dir|not a git repository)" err &&
>> >     ++	grep -E "(already exists|is inside git dir|does not point to a valid repository)" err &&
>> 
>> A few things.
>> 
>>  * Will we be happy to see only one of these possibilities, or do we
>>    expect to see these once for each kind?
>
> I imagine it is only one. This all comes from 9cf8547320 (clone: prevent
> clashing git dirs when cloning submodule in parallel, 2024-01-28), and
> it is expecting the nested path to cause a failure. Which failure I
> guess depends on the racy ordering. If we create the inner one first,
> then we probably get "already exists", and if the outer one, then "is
> inside git dir". I don't know exactly what sequence yields the
> NOT_A_REPO message.
>
> But none of that is changing in this patch, just what the user-visible
> text is for the NOT_A_REPO case.
>
> I did briefly wonder if we might see "not a git repository" from a
> _different_ code path, and need to catch it along with the new message.
> But running successfully with --stress implies that we never see the old
> one anymore.

I see.  Thanks.

>
>>  * a recently started in-flight topic tries to catch bare "grep" and
>>    fails until you write test_grep X-<.
>
> Yeah. This will create a merge conflict for you, but hopefully the
> resolution should be obvious. I don't think it makes sense to fix here,
> as it's orthogonal to the purpose of the patch.

Yup.  I agree that, given that others in the same script will be
updated by that other topic to conflict with this change, it would
not make sense to do the same changes here.

Thanks.