Re: [BUG] commit-graph path normalization reads uninitialized bytes

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On 10/27/2019 12:21 AM, Jeff King wrote:
> I noticed an interesting bug related to 16110c9348 (commit-graph:
> normalize commit-graph filenames, 2019-06-18). It calls
> normalize_path_copy() on the object directory without checking the
> result. That function can fail when the path can't be normalized (e.g.,
> a relative path that goes above the cwd, like "../foo"). That leaves
> nonsense in the destination buffer, which we then call open() on.
> 
> Try this from a git.git build:
> 
>   (cd t && GIT_DIR=../.git valgrind ./git rev-parse HEAD^)
> 
> which yields:
> 
>   ==5907== Syscall param openat(filename) points to uninitialised byte(s)
>   ==5907==    at 0x4914A35: open (open64.c:48)
>   ==5907==    by 0x31C58B: git_open_cloexec (sha1-file.c:1021)
>   ==5907==    by 0x217BF1: open_commit_graph (commit-graph.c:107)
>   ==5907==    by 0x2182BF: load_commit_graph_one (commit-graph.c:314)
>   ...etc...
> 
> This is even triggered in the test suite, but we don't notice because
> the exit code from that invocation is thrown away. Doing this:
> 
> diff --git a/t/t7406-submodule-update.sh b/t/t7406-submodule-update.sh
> index df34c994d2..eebb49ecc5 100755
> --- a/t/t7406-submodule-update.sh
> +++ b/t/t7406-submodule-update.sh
> @@ -222,7 +222,9 @@ test_expect_success 'submodule update --remote should fetch upstream changes' '
>  	(cd super &&
>  	 git submodule update --remote --force submodule &&
>  	 cd submodule &&
> -	 test "$(git log -1 --oneline)" = "$(GIT_DIR=../../submodule/.git git log -1 --oneline)"
> +	 git log -1 --oneline >expect &&
> +	 GIT_DIR=../../submodule/.git git log -1 --oneline >actual &&
> +	 test_cmp expect actual
>  	)
>  '
>  
> 
> and then running t7406 with --valgrind-only=10 yields a similar error
> (and a test failure). Interestingly, neither ASan nor MSan seem to catch
> it. But even more interesting is that it generally does something
> sensible! If I "strace -e openat" that test, I get:
> 
>   openat(AT_FDCWD, "ZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZZB\245\245\245\245\245\245\245\10\261\v", O_RDONLY|O_CLOEXEC) = -1 ENOENT (No such file or directory)
>   openat(AT_FDCWD, "../../submodule/.git/objects/info/commit-graphs/commit-graph-chain", O_RDONLY) = -1 ENOENT (No such file or directory)
> 
> Whoops. That's obviously nonsense, but the kernel tells us only ENOENT
> (I suppose if there was no NUL in the uninitialized bytes, it'd perhaps
> run off the buffer and return EFAULT). But either way, we decide there
> is no commit-graph file and fall back to operating without it. So we
> produce the correct output either way. But obviously using uninitialized
> bytes is bad, and of course it could end up opening some random file
> (less likely than you might think, if we reuse a heap buffer).

These all sound bad. Thanks for finding this issue!

> We could "fix" it like this (there are a few other sites, too, so this
> is just illustrating the idea):
> 
> diff --git a/commit-graph.c b/commit-graph.c
> index fc4a43b8d6..422092d3a9 100644
> --- a/commit-graph.c
> +++ b/commit-graph.c
> @@ -45,7 +45,10 @@ char *get_commit_graph_filename(const char *obj_dir)
>  {
>  	char *filename = xstrfmt("%s/info/commit-graph", obj_dir);
>  	char *normalized = xmalloc(strlen(filename) + 1);
> -	normalize_path_copy(normalized, filename);
> +	if (normalize_path_copy(normalized, filename) < 0) {
> +		free(normalized);
> +		return filename;
> +	}
>  	free(filename);
>  	return normalized;
>  }
> 
> But it really makes me wonder what we're trying to get out of the
> normalization. If you want to compare two paths reliably, then you
> really need to create an absolute path and normalize that. And then
> you'd probably want to handle symlinks, too, so you'd end using
> real_path().
>
> So it's not clear to me if this is a kind of "best effort"
> normalization, and the patch above is the right approach. Or if we need
> to be doing it more carefully.

The reason for the normalization is _mostly_ so we know when to expire a
commit-graph that is no longer referenced by the current commit-graph-chain
file. There are some issues around if the commit-graph is in an alternate,
too, which is likely the reason around this example (it only uses "info/commit-graph"
instead of "info/commit-graphs/graph-{hash}.graph").

In the case that someone runs `git commit-graph write --split` in a repo
with an alternate, and the alternate has a commit-graph at
"{alternate}/info/commit-graph", we want to make sure that file does not
get moved to ".git/objects/info/commit-graphs/graph-{hash}.graph" to be
the base of a new split commit-graph in the current repo. The normalization
seemed to be necessary for that.

With that in mind, I'm not sure your proposed fix is the best solution,
but the "create an absolute path and normalize that" is really what we
want. Or maybe I just misunderstood what normalizing did (a canonical
absolute path would have been sufficient, I think).

Thanks,
-Stolee



[Index of Archives]     [Linux Kernel Development]     [Gcc Help]     [IETF Annouce]     [DCCP]     [Netdev]     [Networking]     [Security]     [V4L]     [Bugtraq]     [Yosemite]     [MIPS Linux]     [ARM Linux]     [Linux Security]     [Linux RAID]     [Linux SCSI]     [Fedora Users]

  Powered by Linux