Re: [PATCH] config.c: NULL check when reading protected config

["Ævar Arnfjörð Bjarmason" <avarab@gmail.com>]

On Tue, Jul 26 2022, Glen Choo wrote:

> Ævar Arnfjörð Bjarmason <avarab@gmail.com> writes:
>
>> On Tue, Jul 26 2022, Glen Choo via GitGitGadget wrote:
>>
>>> From: Glen Choo <chooglen@google.com>
>>>
>>> In read_protected_config(), check whether each file name is NULL before
>>> attempting to read it. This mirrors do_git_config_sequence() (which
>>> read_protected_config() is modelled after).
>>>
>>> Without these NULL checks,
>>>
>>>  make SANITIZE=address test T=t0410*.sh
>>>
>>> fails because xdg_config is NULL, causing us to call fopen(NULL).
>>
>> FWIW a lot more than that fails, that's just the test I focused on for
>> the bug report, the others ones (I didn't check out all of them) all
>> variants of that.
>>
>> See https://github.com/avar/git/runs/7519070124?check_suite_focus=true
>> for the current failing run with that "[2]" patch you quoted. We fail a
>> total of 14 test files (and many more tests within those files).
>
> Ah thanks, I'll amend the message accordingly.
>
>>> Reported-by: Ævar Arnfjörð Bjarmason <avarab@gmail.com>
>>> Signed-off-by: Glen Choo <chooglen@google.com>
>>> ---
>>>     config.c: NULL check when reading protected config
>>>     
>>>     This fixes the SANITIZE=address failure on master, That was introduced
>>>     by gc/bare-repo-discovery. Thanks again to Ævar for the original report
>>>     [1] and for proposing a way to catch this in CI [2].
>>>     
>>>     [1]
>>>     https://lore.kernel.org/git/220725.861qu9oxl4.gmgdl@evledraar.gmail.com
>>>     [2]
>>>     https://lore.kernel.org/git/patch-1.1-e48b6853dd5-20220726T110716Z-avarab@gmail.com
>>>
>>> Published-As: https://github.com/gitgitgadget/git/releases/tag/pr-git-1299%2Fchooglen%2Fconfig%2Ffix-sanitize-address-v1
>>> Fetch-It-Via: git fetch https://github.com/gitgitgadget/git pr-git-1299/chooglen/config/fix-sanitize-address-v1
>>> Pull-Request: https://github.com/git/git/pull/1299
>>>
>>>  config.c | 9 ++++++---
>>>  1 file changed, 6 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/config.c b/config.c
>>> index 015bec360f5..b0ba7f439a4 100644
>>> --- a/config.c
>>> +++ b/config.c
>>> @@ -2645,9 +2645,12 @@ static void read_protected_config(void)
>>>  	system_config = git_system_config();
>>>  	git_global_config(&user_config, &xdg_config);
>>>  
>>> -	git_configset_add_file(&protected_config, system_config);
>>> -	git_configset_add_file(&protected_config, xdg_config);
>>> -	git_configset_add_file(&protected_config, user_config);
>>> +	if (system_config)
>>> +		git_configset_add_file(&protected_config, system_config);
>>> +	if (xdg_config)
>>> +		git_configset_add_file(&protected_config, xdg_config);
>>> +	if (user_config)
>>> +		git_configset_add_file(&protected_config, user_config);
>>>  	git_configset_add_parameters(&protected_config);
>>>  
>>>  	free(system_config);
>>>
>>> base-commit: 6a475b71f8c4ce708d69fdc9317aefbde3769e25
>>
>> Re your claim in
>> https://lore.kernel.org/git/kl6lzggwsyh1.fsf@chooglen-macbookpro.roam.corp.google.com/
>> I tried testing this, and came up with the below.
>>
>> I wonder if we should work in here for general paranoia, but I'm not too
>> familiar with the this part of the config reading, maybe we're confident
>> enough that these are invariants within the process.
>>
>> This will BUG() out if these variables change within the process, which
>> would mean that our caching assumptions are no longer true, which would
>> cause you to return the wrong data here.
>>
>> Of course you'd have segfaulted or similar before, but this should
>> demonstrate that not only are these sometimes NULL, but that they stay
>> that way.
>
> Interesting, this is worth proposing, but I suspect that the
> conversation will be long enough for this to be its own thread. Surely
> someone must have given some thought to this, especially for long-lived
> processes (git-daemon?).
>
> There's also the general question of config cache freshness, e.g. what
> if another git process writes to a shared config file? (We don't worry
> about the single process case because the process will invalidate its
> own cache).
>
> Perhaps we should also worry about that (probably more common) case in
> addition to this one? At any rate, that seems like a bigger topic than
> this fix here.

We can leave it for later, I've run it as a one-off and didn't have any
failures.

But FWIW I think it's tied up in this fix here, i.e. your original code
both added caching, and implicitly assumed that these were never NULL,
so it was "obvious" that it didn't need such assertions.

But now we have 3x if's in a code path that's cached, and the cache is
*not* guarded by the same 3x checks.

So we can leave it for later, but it really seems worth adding some
self-documentation here sooner than later.

This BUG() method I came up with is one way, another would be to strdup
it and use a "static" variable in the function, i.e. stick with whatever
value(s) we start out with.

But in any case, this fix seems correct, and fixes the current issues
SANITIZE=address is spotting for us, thanks!