[PATCH v2] diff: handle NULL filespecs in run_external_diff

From: Jinoh Kang <luke1337@theori.io>
Date: Nov 6, 2020, 17:02 UTC
Message-ID: <a0513d6f-1f69-683d-d6c5-75b17b8b6890@theori.io>
In-Reply-To: <xmqq4km4lppy.fsf@gitster.c.googlers.com>

[forwarding from git-security@googlegroups.com to git@vger.kernel.org]

On 11/4/20 10:30 PM, Junio C Hamano wrote:

Show 31 quoted lines

> Jinoh Kang <luke1337@theori.io> writes:
> 
>> `one` and `two` can be NULL when comparing an unmerged pair.
>> There is a conditional above that already tests for this.
>>
>> Fixes: 3aef54e8b8 ("diff: munmap() file contents before running external diff")
>> Signed-off-by: Jinoh Kang <luke1337@theori.io>
>> ---
>>  diff.c              |  6 ++++--
>>  t/t7800-difftool.sh | 23 +++++++++++++++++++++++
>>  2 files changed, 27 insertions(+), 2 deletions(-)
>>
>> diff --git a/diff.c b/diff.c
>> index d24f47df99..ae1ec2d6c8 100644
>> --- a/diff.c
>> +++ b/diff.c
>> @@ -4267,8 +4267,10 @@ static void run_external_diff(const char *pgm,
>>  	strvec_pushf(&env, "GIT_DIFF_PATH_COUNTER=%d", ++o->diff_path_counter);
>>  	strvec_pushf(&env, "GIT_DIFF_PATH_TOTAL=%d", q->nr);
>>  
>> -	diff_free_filespec_data(one);
>> -	diff_free_filespec_data(two);
>> +	if (one)
>> +		diff_free_filespec_data(one);
>> +	if (two)
>> +		diff_free_filespec_data(two);
>>  	if (run_command_v_opt_cd_env(argv.v, RUN_USING_SHELL, NULL, env.v))
>>  		die(_("external diff died, stopping at %s"), name);
> 
> Have you considered allowing diff_free_filespec_data() to take NULL
> and return safely without doing anything?

Yes, I have. In fact, this kind of behavior is exactly what I would expect from a function that "frees" something. However, I was not entirely sure if this applies here, for several reasons that follow:

> That models after free() and other "we are done with the resource
> and it is time to clean it up" functions,

This corresponds to `free_filespec`. In this particular case, I strongly agree that it makes perfect sense to do nothing when passed a NULL pointer.

However, I humbly opine that the free() semantics do not apply to `diff_free_filespec_data`; rather, I prefer to see it as a function that simply transitions a diff_filespec from one state to another.

For this reason, I'd consider the act of passing NULL to diff_free_filespec_data as a bug in the first place. Further, if it does not entail a security issue, why not just crash *right now* rather than (possibly) causing more obscure bugs later?

I would put the blame on its name, since "data" feels too generic and makes the function sound like freeing the filespec _itself_. diff_filespec carries a lot of other things besides just `data` and `cnt_data`.

Please feel free to correct me if I'm wrong; after all, I am not exactly one of the long-time maintainers.

> fixes this particular bug, and possibly simplifies existing
> callers that check the NULL-ness before calling it.

I was unable to find any callsites that explicitly check for NULL-ness _immediately_ before calling diff_free_filespec_data. Excluding 3aef54e8b8, `GCC -fanalyze` suggests to me that all the callers have already dereferenced the pointer by the time the call is made; therefore, a NULL pointer dereference would have already happened before the flow could even get to the diff_free_filespec_data function.

The NULL checks are usually placed in the beginning of caller functions rather than close to the exit path (which then calls diff_free_filespec_data).

Show 15 quoted lines

> 
>> diff --git a/t/t7800-difftool.sh b/t/t7800-difftool.sh
>> index 524f30f7dc..8cc1c9117c 100755
>> --- a/t/t7800-difftool.sh
>> +++ b/t/t7800-difftool.sh
>> @@ -728,6 +728,29 @@ test_expect_success 'add -N and difftool -d' '
>>  	git difftool --dir-diff --extcmd ls
>>  '
>>  
>> +test_expect_success 'difftool --cached with unmerged files' '
>> +	test_when_finished git reset --hard &&
>> +	echo base > file &&
> 
> Style.  "echo base >file &&" (no SP between redirect operator and
> its target).

My bad. Thanks!

> 
> I do not think this is "security" matter.  I do appreciate you
> erring on the side of being cautious and sending the patch first
> here, but please take it to the regular mailing list.

Yes, no sensible web-based git browser would prepare a working tree, start a conflicting merge *AND* run an external diff on it...

> 
> Thanks.
>

-- 
Jinoh Kang
Theori