[PATCH] show-branch: use strbuf instead of static buffer

When we generate relative names (e.g., "master~20^2"), we
format the name into a static buffer, then xstrdup the
result to attach it to the commit. Since the first thing we
add into the static buffer is the already-computed name of
the child commit, the names may get longer and longer as
the traversal gets deeper, and we may eventually overflow
the fixed-size buffer.

Fix this by converting the fixed-size buffer into a dynamic
strbuf.  The performance implications should be minimal, as
we end up allocating a heap copy of the name anyway (and now
we can just detach the heap copy from the strbuf).

Reported-by: Eric Roman <redacted>
Signed-off-by: Jeff King <redacted>
---
This is a fix for a bug report that came to me off-list.  A real-world
example can be seen by running "git show-branch --all" on a fresh clone
of:

  https://chromium.googlesource.com/chromium/src.git

(but that repo is 1.7G, so I don't recommend cloning it unless you're
really interested). Its master branch consists of a strange sequence of
merges that results in naming commits like master^2^2^2^2... and so on
(it's unclear to me why, but it looks like maybe syncing up separate svn
and git repositories?).  Which is odd, but looking at graph, I think the
names show-branch is generating are correct; they're just really long.
And of course odd history is no excuse to overflow a buffer.

Though this is a stack overflow, I don't know that it's exploitable for
anything interesting; an attacker does not get to write arbitrary data,
but rather only a sequence of "^%d" and "~%d" relative history markers.
Perhaps in theory one could devise a history such that the sequence
markers spelled out some malicious code, but it would be quite a
challenge (and given that you have only ascii [^~0-9] to work with,
probably impossible).

I prepared this on "master", but it should be suitable for "maint"; the
code dates all the way back to git v0.99.

 builtin/show-branch.c | 17 ++++++++---------
 1 file changed, 8 insertions(+), 9 deletions(-)

diff --git a/builtin/show-branch.c b/builtin/show-branch.c
index d208fd6..90fc6b1 100644
--- a/builtin/show-branch.c
+++ b/builtin/show-branch.c

@@ -162,29 +162,28 @@ static void name_commits(struct commit_list *list,
 			nth = 0;
 			while (parents) {
 				struct commit *p = parents->item;
-				char newname[1000], *en;
+				struct strbuf newname = STRBUF_INIT;
 				parents = parents->next;
 				nth++;
 				if (p->util)
 					continue;
-				en = newname;
 				switch (n->generation) {
 				case 0:
-					en += sprintf(en, "%s", n->head_name);
+					strbuf_addstr(&newname, n->head_name);
 					break;
 				case 1:
-					en += sprintf(en, "%s^", n->head_name);
+					strbuf_addf(&newname, "%s^", n->head_name);
 					break;
 				default:
-					en += sprintf(en, "%s~%d",
-						n->head_name, n->generation);
+					strbuf_addf(&newname, "%s~%d",
+						    n->head_name, n->generation);
 					break;
 				}
 				if (nth == 1)
-					en += sprintf(en, "^");
+					strbuf_addch(&newname, '^');
 				else
-					en += sprintf(en, "^%d", nth);
-				name_commit(p, xstrdup(newname), 0);
+					strbuf_addf(&newname, "^%d", nth);
+				name_commit(p, strbuf_detach(&newname, NULL), 0);
 				i++;
 				name_first_parent_chain(p);
 			}

-- 
1.8.2.rc0.33.gd915649

On Fri, Apr 05, 2013 at 04:49:15PM -0700, Jonathan Nieder wrote:

quoted

Though this is a stack overflow, I don't know that it's exploitable for
anything interesting; an attacker does not get to write arbitrary data,
but rather only a sequence of "^%d" and "~%d" relative history markers.
Perhaps in theory one could devise a history such that the sequence
markers spelled out some malicious code, but it would be quite a
challenge

Overwrite the return address and return-to-libc?

Still hard, since you need to construct a usable address (and arguments)
out of sequences of "^[0-9]+" and "~[0-9]+". But I'd love to see a
working exploit if somebody thinks they can do it. :)

Very clean and obviously correct.  Thanks.

Reviewed-by: Jonathan Nieder <redacted>

Thanks.

A test would be nice, though.

What should it be testing? That a giant chain of second-parent merges
that exceeds 1000 bytes doesn't segfault? Tests like that are not all
that interesting, because they do not catch real-world regressions. We
have closed this barn door; it is not impossible that it will be
re-opened, but it is not likely. A test that checks only for a very
specific type of failure is only ever going to see that failure.

If you want to design a suite of tests that check that show-branch gives
correct output for particular brands of large repo, that would be
generic and potentially useful. But I don't think it's actually worth
spending a lot of time on (reviewing the code for more static buffers
and sprintfs would probably be a much more fruitful use of time).

-Peff

Jeff King wrote:

When we generate relative names (e.g., "master~20^2"), we
format the name into a static buffer, then xstrdup the
result to attach it to the commit. Since the first thing we
add into the static buffer is the already-computed name of
the child commit, the names may get longer and longer as
the traversal gets deeper, and we may eventually overflow
the fixed-size buffer.

Good catch.

[...]

Though this is a stack overflow, I don't know that it's exploitable for
anything interesting; an attacker does not get to write arbitrary data,
but rather only a sequence of "^%d" and "~%d" relative history markers.
Perhaps in theory one could devise a history such that the sequence
markers spelled out some malicious code, but it would be quite a
challenge

Overwrite the return address and return-to-libc?

[...]

Very clean and obviously correct.  Thanks.

Reviewed-by: Jonathan Nieder <redacted>

A test would be nice, though.

Hope that helps,
Jonathan