[PATCH] shallow: verify shallow file after taking lock

[Jeff King <peff@peff.net>]

Before writing the shallow file, we stat() the existing file
to make sure it has not been updated since our operation
began. However, we do not do so under a lock, so there is a
possible race:

  1. Process A takes the lock.

  2. Process B calls check_shallow_file_for_update and finds
     no update.

  3. Process A commits the lockfile.

  4. Process B takes the lock, then overwrite's process A's
     changes.

We can fix this by doing our check while we hold the lock.

Signed-off-by: Jeff King <peff@peff.net>
---
On Thu, Feb 27, 2014 at 04:04:26AM -0500, Jeff King wrote:

> by the way, shouldn't we do that check under the lock? I think
> setup_alternate_shallow has a race condition).

Here is the fix. I didn't actually reproduce the race experimentally,
but it seems pretty straightforward.

I also notice that check_shallow_file_for_update returns early if
!is_shallow. Is that safe? Is it possible for another process to have
made us shallow since the program began? In that case, we would have to
stat() the file always, then complain if it exists and !is_shallow.

 shallow.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/shallow.c b/shallow.c
index bbc98b5..75da07a 100644
--- a/shallow.c
+++ b/shallow.c
@@ -246,9 +246,9 @@ void setup_alternate_shallow(struct lock_file *shallow_lock,
 	struct strbuf sb = STRBUF_INIT;
 	int fd;
 
-	check_shallow_file_for_update();
 	fd = hold_lock_file_for_update(shallow_lock, git_path("shallow"),
 				       LOCK_DIE_ON_ERROR);
+	check_shallow_file_for_update();
 	if (write_shallow_commits(&sb, 0, extra)) {
 		if (write_in_full(fd, sb.buf, sb.len) != sb.len)
 			die_errno("failed to write to %s",
@@ -293,9 +293,9 @@ void prune_shallow(int show_only)
 		strbuf_release(&sb);
 		return;
 	}
-	check_shallow_file_for_update();
 	fd = hold_lock_file_for_update(&shallow_lock, git_path("shallow"),
 				       LOCK_DIE_ON_ERROR);
+	check_shallow_file_for_update();
 	if (write_shallow_commits_1(&sb, 0, NULL, SEEN_ONLY)) {
 		if (write_in_full(fd, sb.buf, sb.len) != sb.len)
 			die_errno("failed to write to %s",
-- 
1.8.5.2.500.g8060133