Re: [PATCH] attr: fix msan issue in read_attr_from_index

From: Kyle Lippincott <spectral@google.com>
Date: Jun 17, 2024, 21:17 UTC
Message-ID: <CAO_smVg1GuyreBw7Dw0RjLPjD1KhH-NmFMkkC-D4hB7kfPqjCQ@mail.gmail.com>
In-Reply-To: <xmqqcyof5n2t.fsf@gitster.g>

On Mon, Jun 17, 2024 at 1:30 PM Junio C Hamano <gitster@pobox.com> wrote:

Show 48 quoted lines

>
> "Kyle Lippincott via GitGitGadget" <gitgitgadget@gmail.com> writes:
>
> > The issue exists because `size` is an output parameter from
> > `read_blob_data_from_index`, but it's only modified if
> > `read_blob_data_from_index` returns non-NULL.
>
> Correct.
>
> > The read of `size` when
> > calling `read_attr_from_buf` unconditionally may read from an
> > uninitialized value. `read_attr_from_buf` checks that `buf` is non-NULL
> > before reading from `size`, but by then it's already too late: the
> > uninitialized read will have happened already.
>
> Yes, but it is dubious that reading an uninitialized value that we
> know will not be used is a problem, so I am inclined to say that
> MSAN is giving a false positive here.
>
> > Furthermore, there's no
> > guarantee that the compiler won't reorder things so that it checks
> > `size` before checking `!buf`.
>
> This I do not understand.  Are you talking about buf vs length here
> in the callee?
>
>         static struct attr_stack *read_attr_from_buf(char *buf, size_t length,
>                                                      const char *path, unsigned flags)
>         {
>                 struct attr_stack *res;
>                 char *sp;
>                 int lineno = 0;
>
>                 if (!buf)
>                         return NULL;
>                 if (length >= ATTR_MAX_FILE_SIZE) {
>                         warning(_("ignoring overly large gitattributes blob '%s'"), path);
>                         free(buf);
>                         return NULL;
>                 }
>
> At the machine level, a prefetch may happen from both buf and
> length, but the program ought to behave the same way as the code is
> executed serially as written.  If the compiler allows the outside
> world to observe that resulting code checks length even when buf is
> NULL, such a compiler is broken.  So I do not think that is what you
> are referring to, but then I do not know what problem you are
> describing.

Once there's an uninitialized read, we're in undefined behavior territory. There's no requirement that the compiler keep the code operating the way we'd logically expect once there's undefined behavior, especially with the optimizer involved.

I think that you're right though: when I wrote this I'd convinced myself that this wasn't guaranteed to work, but taking a look now I can't think of a way for this to go wrong, because afaik size_t is such a simple type, conceptually. When compiling `read_attr_from_buf`, it can't actually assume any relationship between `buf` and `length`. Maybe I was thinking that with link-time optimizations (whole-program optimizations) it can go wrong? I'm not remembering what I was thinking about when I wrote that, sorry.

Show 19 quoted lines

>
> Having said all that ...
>
> > Make the call to `read_attr_from_buf` conditional on `buf` being
> > non-NULL, ensuring that `size` is not read if it's never set.
>
> ... this makes the logic at the caller crystal clear, so even if
> there are suboptimal checker that bothers us with false positives,
> the change itself justifies itself, I would say.
>
> >       } else {
> >               buf = read_blob_data_from_index(istate, path, &size);
> > -             stack = read_attr_from_buf(buf, size, path, flags);
> > +             if (buf)
> > +                     stack = read_attr_from_buf(buf, size, path, flags);
> >       }
> >       return stack;
>
> Thanks.