Problems with unrecognized headers in git bundles

Hi,

creating bundles from some repositories seems to lead to bundles with incorrectly formatted headers, at least with git >= 1.7.2. When cloning from such bundles, git prints the following error/warning:

  $ git clone perl-clone.bundle perl-clone
  Cloning into 'perl-clone'...
  warning: unrecognized header: --work around mangled archname on...

This can be reproduced easily with git from any version >= 1.7.2 or from master, using the following steps:

  git clone git://perl5.git.perl.org/perl.git perl
  GIT_DIR=perl/.git git bundle create perl-clone.bundle --all
  git clone perl-clone.bundle perl-clone

The content of the bundle is:

  # v2 git bundle
  -- work around mangled archname on win32 while finding...
  39ec54a59ce332fc44e553f4e5eeceef88e8369e refs/heads/blead
  39ec54a59ce332fc44e553f4e5eeceef88e8369e refs/remotes/origin/HEAD
  ...

The "--work around mangled archname..." line is rather long, so I've omitted most of it. What it contains is a series of commit messages all combined into a single line. It appears that this line is the problem because it's neither a comment (like '#v2 git bundle') nor a SHA1 followed by a ref name.

Note that this problem does not occur with all repositories. A bundle created from a test repository with a single text file and just one commit does not have this problem.

Note also that "git clone <bundle> <dest>" does not fail with corrupted bundles if the git version is something like 1.7.2 or 1.7.7.6. Those version print the above warning but still succeed in cloning. "git clone" from master however treats this as an error and fails.

Without any knowledge of how git works internally, I would assume that this is either a bug in how bundles are created, or a hint at a slightly broken perl repository (other's have the same thing though, perhaps it is because of a conversion from another SCM software to git in the past). Does this sound reasonable?

Is there a way to work around this or fix it properly? I'm not sure what lead to the decision to no longer ignore unrecognized headers in git master, would it be sensible to revert this change if nothing can be done to solve this during bundle creation?

- Jannis

Have researched this a bit, and I've found that all git versions back to when git-bundle was introduced (around v1.5.4) produces the same invalid line. The culprit is commit 3e8148feadabd0d0b1869fcc4d218a6475a5b0bc in perl.git, branch 'maint-5.005'. The log message of that commit contains email headers, maybe that's the reason git bundle gets confused?

Øyvind

...or maybe because the log message doesn't contain any empty lines, so they're joined together into an insanely long line. I've seen this behaviour before, in git-am or git-apply, I think. Anyway, when the bundle doesn't contain this commit, the line is not present.

Øyvind

For the lazy, the commit can be found here:

http://perl5.git.perl.org/perl.git/commit/3e8148feadabd0d0b1869fcc4d218a6475a5b0bc

The comment even said that it should eventually go there. While at it, match the calling convention and name of the function to the strbuf_get*line family. So it now is strbuf_getwholeline_fd.

---

I was left looking for strbuf_readline_fd() for a while, then tried to look for the readline() from the libc that it would presumably match. Hence this cleanup.

 bundle.c |   21 ++-------------------
 strbuf.c |   16 ++++++++++++++++
 strbuf.h |    1 +
 3 files changed, 19 insertions(+), 19 deletions(-)

index b8acf3c..313de42 100644
@@ -23,23 +23,6 @@ static void add_to_ref_list(const unsigned char *sha1, const char *name,
 	list->nr++;
 }
 
-/* Eventually this should go to strbuf.[ch] */
-static int strbuf_readline_fd(struct strbuf *sb, int fd)
-{
-	strbuf_reset(sb);
-
-	while (1) {
-		char ch;
-		ssize_t len = xread(fd, &ch, 1);
-		if (len <= 0)
-			return len;
-		strbuf_addch(sb, ch);
-		if (ch == '\n')
-			break;
-	}
-	return 0;
-}
-
 static int parse_bundle_header(int fd, struct bundle_header *header,
 			       const char *report_path)
 {

@@ -47,7 +30,7 @@ static int parse_bundle_header(int fd, struct bundle_header *header,
 	int status = 0;
 
 	/* The bundle header begins with the signature */
-	if (strbuf_readline_fd(&buf, fd) ||
+	if (strbuf_getwholeline_fd(&buf, fd, '\n') ||
 	    strcmp(buf.buf, bundle_signature)) {
 		if (report_path)
 			error("'%s' does not look like a v2 bundle file",

@@ -57,7 +40,7 @@ static int parse_bundle_header(int fd, struct bundle_header *header,
 	}
 
 	/* The bundle header ends with an empty line */
-	while (!strbuf_readline_fd(&buf, fd) &&
+	while (!strbuf_getwholeline_fd(&buf, fd, '\n') &&
 	       buf.len && buf.buf[0] != '\n') {
 		unsigned char sha1[20];
 		int is_prereq = 0;

index ff0b96b..5135d59 100644
@@ -383,6 +383,22 @@ int strbuf_getline(struct strbuf *sb, FILE *fp, int term)
 	return 0;
 }
 
+int strbuf_getwholeline_fd(struct strbuf *sb, int fd, int term)
+{
+	strbuf_reset(sb);
+
+	while (1) {
+		char ch;
+		ssize_t len = xread(fd, &ch, 1);
+		if (len <= 0)
+			return EOF;
+		strbuf_addch(sb, ch);
+		if (ch == term)
+			break;
+	}
+	return 0;
+}
+
 int strbuf_read_file(struct strbuf *sb, const char *path, size_t hint)
 {
 	int fd, len;

index fbf059f..3effaa8 100644
@@ -116,6 +116,7 @@ static inline void strbuf_complete_line(struct strbuf *sb)
 
 extern int strbuf_getwholeline(struct strbuf *, FILE *, int);
 extern int strbuf_getline(struct strbuf *, FILE *, int);
+extern int strbuf_getwholeline_fd(struct strbuf *, int, int);
 
 extern void stripspace(struct strbuf *buf, int skip_comments);
 extern int launch_editor(const char *path, struct strbuf *buffer, const char *const *env);

1.7.9.1.430.g4998543

Nit: no update to Documentation/technical/api-strbuf.txt.

You might want to also mention that this will read() one byte at a time, and is therefore only a good idea if you really care about the position of the fd. Otherwise, fdopen() + strbuf_getwholeline() is much more efficient.

-Peff

The first part of the bundle header contains the boundary commits, and could be approximated by

  # v2 git bundle
  $(git rev-list --pretty=oneline --boundary <ARGS> | grep ^-)

git-bundle actually spawns exactly this rev-list invocation, and does the grepping internally.

There was a subtle bug in the latter step: it used fgets() with a 1024-byte buffer. If the user has sufficiently long subjects (e.g., by not adhering to the git oneline-subject convention in the first place), the 'oneline' format can easily overflow the buffer. fgets() then returns the rest of the line in the next call(s). If one of these remaining parts started with '-', git-bundle would mistakenly insert it into the bundle thinking it was a boundary commit.

Fix it by using strbuf_getwholeline() instead, which handles arbitrary line lengths correctly.

Note that on the receiving side in parse_bundle_header() we were already using strbuf_getwholeline_fd(), so that part is safe.

---
 bundle.c          |   14 +++++++-------
 t/t5704-bundle.sh |   17 +++++++++++++++++
 2 files changed, 24 insertions(+), 7 deletions(-)

index 313de42..0dbd174 100644
@@ -234,7 +234,7 @@ int create_bundle(struct bundle_header *header, const char *path,
 	const char **argv_boundary = xmalloc((argc + 4) * sizeof(const char *));
 	const char **argv_pack = xmalloc(6 * sizeof(const char *));
 	int i, ref_count = 0;
-	char buffer[1024];
+	struct strbuf buf = STRBUF_INIT;
 	struct rev_info revs;
 	struct child_process rls;
 	FILE *rls_fout;

@@ -266,16 +266,16 @@ int create_bundle(struct bundle_header *header, const char *path,
 	if (start_command(&rls))
 		return -1;
 	rls_fout = xfdopen(rls.out, "r");
-	while (fgets(buffer, sizeof(buffer), rls_fout)) {
+	while (strbuf_getwholeline(&buf, rls_fout, '\n') != EOF) {
 		unsigned char sha1[20];
-		if (buffer[0] == '-') {
-			write_or_die(bundle_fd, buffer, strlen(buffer));
-			if (!get_sha1_hex(buffer + 1, sha1)) {
+		if (buf.len > 0 && buf.buf[0] == '-') {
+			write_or_die(bundle_fd, buf.buf, buf.len);
+			if (!get_sha1_hex(buf.buf + 1, sha1)) {
 				struct object *object = parse_object(sha1);
 				object->flags |= UNINTERESTING;
-				add_pending_object(&revs, object, buffer);
+				add_pending_object(&revs, object, buf.buf);
 			}
-		} else if (!get_sha1_hex(buffer, sha1)) {
+		} else if (!get_sha1_hex(buf.buf, sha1)) {
 			struct object *object = parse_object(sha1);
 			object->flags |= SHOWN;
 		}

index 4ae127d..7c2f307 100755
@@ -59,4 +59,21 @@ test_expect_success 'empty bundle file is rejected' '
 
 '
 
+# If "ridiculous" is at least 1004 chars, this traps a bug in old
+# versions where the resulting 1025-char line (with --pretty=oneline)
+# was longer than a 1024-char buffer
+test_expect_success 'ridiculously long subject in boundary' '
+
+	: > file4 &&
+	test_tick &&
+	git add file4 &&
+	printf "abcdefghijkl %s\n" $(seq 1 100) | git commit -F - &&
+	test_commit fifth &&
+	git bundle create long-subject-bundle.bdl HEAD^..HEAD &&
+	git fetch long-subject-bundle.bdl &&
+	sed -n "/^-/{p;q}" long-subject-bundle.bdl > boundary &&
+	grep "^-$_x40 " boundary
+
+'
+
 test_done

1.7.9.1.430.g4998543

Am 22.02.2012 21:55, schrieb Jeff King:

Thanks for pointing this out.

If a sequence of ASCII zeros is good enough, you can also do:

printf %02134d 0

-- Hannes

Thanks for diagnosing this, but I wonder if it even needs --pretty=oneline to begin with, except for debugging purposes.

Do we ever use the subject string read from the rev-list output in any way?

In other words, I am wondering if the right patch to minimally fix the issue starting from older releases is something along this line instead:

index b8acf3c..339dbb0 100644
@@ -248,7 +248,7 @@ int create_bundle(struct bundle_header *header, const char *path,
 	static struct lock_file lock;
 	int bundle_fd = -1;
 	int bundle_to_stdout;
-	const char **argv_boundary = xmalloc((argc + 4) * sizeof(const char *));
+	const char **argv_boundary = xmalloc((argc + 3) * sizeof(const char *));
 	const char **argv_pack = xmalloc(6 * sizeof(const char *));
 	int i, ref_count = 0;
 	char buffer[1024];

@@ -271,11 +271,10 @@ int create_bundle(struct bundle_header *header, const char *path,
 	init_revisions(&revs, NULL);
 
 	/* write prerequisites */
-	memcpy(argv_boundary + 3, argv + 1, argc * sizeof(const char *));
+	memcpy(argv_boundary + 2, argv + 1, argc * sizeof(const char *));
 	argv_boundary[0] = "rev-list";
 	argv_boundary[1] = "--boundary";
-	argv_boundary[2] = "--pretty=oneline";
-	argv_boundary[argc + 2] = NULL;
+	argv_boundary[argc + 1] = NULL;
 	memset(&rls, 0, sizeof(rls));
 	rls.argv = argv_boundary;
 	rls.out = -1;

Not sure. The only use I could think of would be to google for the subjects, in the hope of finding some repository that has the commit you are looking for. Other than that...

In any case the --pretty=oneline is very deliberate, as we can see from the commit below. It just doesn't give a reason :-)

commit 239296770dae75e21c179733785731ec6ffae1f5

Author: Johannes Schindelin <[email protected]>
Date:   Fri Feb 23 03:17:51 2007 +0100

git-bundle: record commit summary in the prerequisite data

index 191ec55..d74afaa 100644
@@ -267,7 +267,7 @@ static int create_bundle(struct bundle_header *header, const char *path,
 		int argc, const char **argv)
 {
 	int bundle_fd = -1;
-	const char **argv_boundary = xmalloc((argc + 3) * sizeof(const char *));
+	const char **argv_boundary = xmalloc((argc + 4) * sizeof(const char *));
 	const char **argv_pack = xmalloc(4 * sizeof(const char *));
 	int pid, in, out, i, status;
 	char buffer[1024];

@@ -282,10 +282,11 @@ static int create_bundle(struct bundle_header *header, const char *path,
 	write_or_die(bundle_fd, bundle_signature, strlen(bundle_signature));
 
 	/* write prerequisites */
-	memcpy(argv_boundary + 2, argv + 1, argc * sizeof(const char *));
+	memcpy(argv_boundary + 3, argv + 1, argc * sizeof(const char *));
 	argv_boundary[0] = "rev-list";
 	argv_boundary[1] = "--boundary";
-	argv_boundary[argc + 1] = NULL;
+	argv_boundary[2] = "--pretty=oneline";
+	argv_boundary[argc + 2] = NULL;
 	out = -1;
 	pid = fork_with_pipe(argv_boundary, NULL, &out);
 	if (pid < 0)

Thomas Rast
trast@{inf,student}.ethz.ch

Or because the bundle creator (which may even be you on a different day) did not correctly guess at the negative refs when specifying a cutoff. Assuming you no longer have access to the original repo (not unreasonable, since you are using a bundle), the messages could help you understand where the error occurred.

Of course, once you figure out "aha! I expected the bundle to include foo, but it is actually a cutoff point. I should have said XYZ^ as the cutoff", I'm not sure what you do then. It's not like there is an easy way to salvage the data that is in the bundle.

So I think it is a debugging aid, but one that ultimately doesn't help you that much.

-Peff

Thanks for digging and I hope everybody on this list who would ever send a patch learns the importance of describing _why_ in the log from this lesson.

I'll amend your 2/2 with a note that keeping the subject material is not justified and we may want to remove (or at least reduce) it in the future.

So you all made very good points, and I don't want to repeat them. Junio's analysis

is of course also correct for the case where the overlong line is in a boundary commit. (The analysis about the accidentally-leading '-' is also correct and can create a boundary line where there was none, like in the perl case.)

The patches in this round are:

[1] new; following up on Peff's complaint that I should document strbuf_getwholeline_fd, I found that the strbuf_get*line documentation was also lacking.

[2] previously (1/2), now with documentation added

[3] new; Junio corrected me on the ': > file' spelling, which I added because the start of the test used it too. So let's just fix it outright to match the Git style all over.

[4] previously (2/2), now incorporating a strbuf_release and the suggestions for the tests as sent by Junio. I upped the %0982 a bit as it felt *too* tailored for the old bug, and this way everyone can see immediately that it will exceed the 1024 char limit (without thinking about the length of a sha and accounting for \0 and \n).

Thomas Rast (4):
strbuf: improve strbuf_get*line documentation
bundle: put strbuf_readline_fd in strbuf.c with adjustments
t5704: match tests to modern style
bundle: use a strbuf to scan the log for boundary commits

 Documentation/technical/api-strbuf.txt |   19 +++++++++++--
 bundle.c                               |   36 +++++++-----------------
 strbuf.c                               |   16 +++++++++++
 strbuf.h                               |    1 +
 t/t5704-bundle.sh                      |   47 ++++++++++++++++----------------
 5 files changed, 67 insertions(+), 52 deletions(-)

1.7.9.1.430.g4998543

I really wish you did not make an elephant out of this, or if you must, kept the minimum bugfix part that must apply to older codebase and making an elephant part clearly separable.

Anyway, I've reordered so that the bugfix part would apply to an older maintenance release, and then the remainder applied on top would merge cleanly to the bleeding edge, so there is no need to resend.

Thanks for working on this, everybody.

strbuf_getline() was not documented very clearly, though a reader familiar with getline() would not have had any questions about it. strbuf_getwholeline() was not documented at all.

---
 Documentation/technical/api-strbuf.txt |   12 ++++++++++--
 1 file changed, 10 insertions(+), 2 deletions(-)

index afe2759..cc6db5e 100644
@@ -255,8 +255,16 @@ same behaviour as well.
 
 `strbuf_getline`::
 
-	Read a line from a FILE* pointer. The second argument specifies the line
-	terminator character, typically `'\n'`.
+	Read a line from a FILE*. The second argument specifies the
+	line terminator character, typically `'\n'`.  Reading stops
+	after the terminator or at EOF.  The terminator is removed
+	from the buffer before returning.  Returns 0 unless there was
+	nothing left before EOF, in which case it returns `EOF`.
+
+`strbuf_getwholeline`::
+	
+	Like `strbuf_getline`, but keeps the trailing terminator (if
+	any) in the buffer.
 
 `stripspace`::
 

1.7.9.1.430.g4998543

Thanks for improving the existing docs. One suggestion:

The comment even said that it should eventually go there. While at it, match the calling convention and name of the function to the strbuf_get*line family. So it now is strbuf_getwholeline_fd.

---
 Documentation/technical/api-strbuf.txt |    7 +++++++
 bundle.c                               |   21 ++-------------------
 strbuf.c                               |   16 ++++++++++++++++
 strbuf.h                               |    1 +
 4 files changed, 26 insertions(+), 19 deletions(-)

index cc6db5e..827c4bd 100644
@@ -266,6 +266,13 @@ same behaviour as well.
 	Like `strbuf_getline`, but keeps the trailing terminator (if
 	any) in the buffer.
 
+`strbuf_getwholeline_fd`::
+
+	Like `strbuf_getwholeline`, but operates on a file descriptor.
+	It reads one character at a time, so it is very slow.  Do not
+	use it unless you need the correct position in the file
+	descriptor!
+
 `stripspace`::
 
 	Strip whitespace from a buffer. The second parameter controls if

index b8acf3c..313de42 100644
@@ -23,23 +23,6 @@ static void add_to_ref_list(const unsigned char *sha1, const char *name,
 	list->nr++;
 }
 
-/* Eventually this should go to strbuf.[ch] */
-static int strbuf_readline_fd(struct strbuf *sb, int fd)
-{
-	strbuf_reset(sb);
-
-	while (1) {
-		char ch;
-		ssize_t len = xread(fd, &ch, 1);
-		if (len <= 0)
-			return len;
-		strbuf_addch(sb, ch);
-		if (ch == '\n')
-			break;
-	}
-	return 0;
-}
-
 static int parse_bundle_header(int fd, struct bundle_header *header,
 			       const char *report_path)
 {

@@ -47,7 +30,7 @@ static int parse_bundle_header(int fd, struct bundle_header *header,
 	int status = 0;
 
 	/* The bundle header begins with the signature */
-	if (strbuf_readline_fd(&buf, fd) ||
+	if (strbuf_getwholeline_fd(&buf, fd, '\n') ||
 	    strcmp(buf.buf, bundle_signature)) {
 		if (report_path)
 			error("'%s' does not look like a v2 bundle file",

@@ -57,7 +40,7 @@ static int parse_bundle_header(int fd, struct bundle_header *header,
 	}
 
 	/* The bundle header ends with an empty line */
-	while (!strbuf_readline_fd(&buf, fd) &&
+	while (!strbuf_getwholeline_fd(&buf, fd, '\n') &&
 	       buf.len && buf.buf[0] != '\n') {
 		unsigned char sha1[20];
 		int is_prereq = 0;

index ff0b96b..5135d59 100644
@@ -383,6 +383,22 @@ int strbuf_getline(struct strbuf *sb, FILE *fp, int term)
 	return 0;
 }
 
+int strbuf_getwholeline_fd(struct strbuf *sb, int fd, int term)
+{
+	strbuf_reset(sb);
+
+	while (1) {
+		char ch;
+		ssize_t len = xread(fd, &ch, 1);
+		if (len <= 0)
+			return EOF;
+		strbuf_addch(sb, ch);
+		if (ch == term)
+			break;
+	}
+	return 0;
+}
+
 int strbuf_read_file(struct strbuf *sb, const char *path, size_t hint)
 {
 	int fd, len;

index fbf059f..3effaa8 100644
@@ -116,6 +116,7 @@ static inline void strbuf_complete_line(struct strbuf *sb)
 
 extern int strbuf_getwholeline(struct strbuf *, FILE *, int);
 extern int strbuf_getline(struct strbuf *, FILE *, int);
+extern int strbuf_getwholeline_fd(struct strbuf *, int, int);
 
 extern void stripspace(struct strbuf *buf, int skip_comments);
 extern int launch_editor(const char *path, struct strbuf *buffer, const char *const *env);

1.7.9.1.430.g4998543

The test did not adhere to the current style on several counts:

- empty lines around the test blocks, but within the test string

- ': > file' or even just '> file' with an extra space

- inconsistent indentation

- hand-rolled commits instead of using test_commit

Fix all of them. There's a catch to the last point: test_commit creates a tag. We still change it to test_commit, and explicitly delete the tags, so as to highlight that the test relies on not having them.

---
 t/t5704-bundle.sh |   33 ++++++++-------------------------
 1 file changed, 8 insertions(+), 25 deletions(-)

index 4ae127d..f35f559 100755
@@ -4,59 +4,42 @@ test_description='some bundle related tests'
 . ./test-lib.sh
 
 test_expect_success 'setup' '
-
-	: > file &&
-	git add file &&
-	test_tick &&
-	git commit -m initial &&
+	test_commit initial &&
 	test_tick &&
 	git tag -m tag tag &&
-	: > file2 &&
-	git add file2 &&
-	: > file3 &&
-	test_tick &&
-	git commit -m second &&
-	git add file3 &&
-	test_tick &&
-	git commit -m third
-
+	test_commit second &&
+	test_commit third &&
+	git tag -d initial &&
+	git tag -d second &&
+	git tag -d third
 '
 
 test_expect_success 'tags can be excluded by rev-list options' '
-
 	git bundle create bundle --all --since=7.Apr.2005.15:16:00.-0700 &&
 	git ls-remote bundle > output &&
 	! grep tag output
-
 '
 
 test_expect_success 'die if bundle file cannot be created' '
-
 	mkdir adir &&
 	test_must_fail git bundle create adir --all
-
 '
 
 test_expect_failure 'bundle --stdin' '
-
 	echo master | git bundle create stdin-bundle.bdl --stdin &&
 	git ls-remote stdin-bundle.bdl >output &&
 	grep master output
-
 '
 
 test_expect_failure 'bundle --stdin <rev-list options>' '
-
 	echo master | git bundle create hybrid-bundle.bdl --stdin tag &&
 	git ls-remote hybrid-bundle.bdl >output &&
 	grep master output
-
 '
 
 test_expect_success 'empty bundle file is rejected' '
-
-    >empty-bundle && test_must_fail git fetch empty-bundle
-
+	: >empty-bundle &&
+	test_must_fail git fetch empty-bundle
 '
 
 test_done

1.7.9.1.430.g4998543

I do not have an objection to the use of these three test_commits, and I do not have an objection to delete the extra three tags they create so that the set of refs in the repository matches what the later test expects, either.

But I found the explanation a bit iffy.

index a51c8b0..3c436e7 100755
@@ -6,18 +6,15 @@ test_description='some bundle related tests'
 test_expect_success 'setup' '
 	test_commit initial &&
 	test_tick &&
-	git tag -m tag tag &&
+	git tag -m tag antienttag &&
 	test_commit second &&
-	test_commit third &&
-	git tag -d initial &&
-	git tag -d second &&
-	git tag -d third
+	test_commit third
 '
 
 test_expect_success 'tags can be excluded by rev-list options' '
 	git bundle create bundle --all --since=7.Apr.2005.15:16:00.-0700 &&
 	git ls-remote bundle > output &&
-	! grep tag output
+	! grep antienttag output
 '
 
 test_expect_success 'die if bundle file cannot be created' '

The first part of the bundle header contains the boundary commits, and could be approximated by

  # v2 git bundle
  $(git rev-list --pretty=oneline --boundary <ARGS> | grep ^-)

git-bundle actually spawns exactly this rev-list invocation, and does the grepping internally.

There was a subtle bug in the latter step: it used fgets() with a 1024-byte buffer. If the user has sufficiently long subjects (e.g., by not adhering to the git oneline-subject convention in the first place), the 'oneline' format can easily overflow the buffer. fgets() then returns only the first part of the line _without a \n_, and it would be printed as such, turning the next line into trailing garbage. On top of that, fgets() returns the rest of the line in the next call(s). If one of these remaining parts starts with '-', git-bundle would mistakenly insert it into the boundary list.

Fix it by using strbuf_getwholeline() instead, which handles arbitrary line lengths correctly.

Note that on the receiving side in parse_bundle_header() we were already using strbuf_getwholeline_fd(), so that part is safe.

Thanks to Peff, Johannes Sixt and Junio for suggestions.

---
 bundle.c          |   15 ++++++++-------
 t/t5704-bundle.sh |   16 ++++++++++++++++
 2 files changed, 24 insertions(+), 7 deletions(-)

index 313de42..7a760db 100644
@@ -234,7 +234,7 @@ int create_bundle(struct bundle_header *header, const char *path,
 	const char **argv_boundary = xmalloc((argc + 4) * sizeof(const char *));
 	const char **argv_pack = xmalloc(6 * sizeof(const char *));
 	int i, ref_count = 0;
-	char buffer[1024];
+	struct strbuf buf = STRBUF_INIT;
 	struct rev_info revs;
 	struct child_process rls;
 	FILE *rls_fout;

@@ -266,20 +266,21 @@ int create_bundle(struct bundle_header *header, const char *path,
 	if (start_command(&rls))
 		return -1;
 	rls_fout = xfdopen(rls.out, "r");
-	while (fgets(buffer, sizeof(buffer), rls_fout)) {
+	while (strbuf_getwholeline(&buf, rls_fout, '\n') != EOF) {
 		unsigned char sha1[20];
-		if (buffer[0] == '-') {
-			write_or_die(bundle_fd, buffer, strlen(buffer));
-			if (!get_sha1_hex(buffer + 1, sha1)) {
+		if (buf.len > 0 && buf.buf[0] == '-') {
+			write_or_die(bundle_fd, buf.buf, buf.len);
+			if (!get_sha1_hex(buf.buf + 1, sha1)) {
 				struct object *object = parse_object(sha1);
 				object->flags |= UNINTERESTING;
-				add_pending_object(&revs, object, buffer);
+				add_pending_object(&revs, object, buf.buf);
 			}
-		} else if (!get_sha1_hex(buffer, sha1)) {
+		} else if (!get_sha1_hex(buf.buf, sha1)) {
 			struct object *object = parse_object(sha1);
 			object->flags |= SHOWN;
 		}
 	}
+	strbuf_release(&buf);
 	fclose(rls_fout);
 	if (finish_command(&rls))
 		return error("rev-list died");

index f35f559..a51c8b0 100755
@@ -42,4 +42,20 @@ test_expect_success 'empty bundle file is rejected' '
 	test_must_fail git fetch empty-bundle
 '
 
+# This triggers a bug in older versions where the resulting line (with
+# --pretty=oneline) was longer than a 1024-char buffer.
+test_expect_success 'ridiculously long subject in boundary' '
+	: >file4 &&
+	test_tick &&
+	git add file4 &&
+	printf "%01200d\n" 0 | git commit -F - &&
+	test_commit fifth &&
+	git bundle create long-subject-bundle.bdl HEAD^..HEAD &&
+	git bundle list-heads long-subject-bundle.bdl >heads &&
+	test -s heads &&
+	git fetch long-subject-bundle.bdl &&
+	sed -n "/^-/{p;q}" long-subject-bundle.bdl >boundary &&
+	grep "^-$_x40 " boundary
+'
+
 test_done

1.7.9.1.430.g4998543