Re: [musl] undefined behavior in fread.c

[Tamir Duberstein <tamird@google.com>]

[-- Attachment #1: Type: text/plain, Size: 1764 bytes --]

I agree, the caller's behavior is UB. I'll send them (freetype2) a patch.

That said, do we want to avoid internal UB here anyway?

- As mentioned earlier, glibc avoids the UB (and the lock).
- llvm-libc does the same starting with
https://github.com/llvm/llvm-project/commit/53c251b
- uclibc avoids the UB but still locks:
https://github.com/gittup/uClibc/blob/9dbf00b/libc/stdio/fread.c#L25
- FreeBSD avoids the UB but still locks:
https://svnweb.freebsd.org/base/head/lib/libc/stdio/fread.c?view=markup#l76
- Android (bionic) avoids the UB but still locks:
https://cs.android.com/android/platform/superproject/+/master:bionic/libc/stdio/stdio.cpp;l=1099;drc=4aa8f499f21ebf84101de34d68682d5388667001

Does this persuade?

On Fri, Feb 24, 2023 at 10:13 AM NRK <nrk@disroot.org> wrote:

> On Fri, Feb 24, 2023 at 09:42:00AM -0500, Tamir Duberstein wrote:
> > We could take the lock and still avoid UB with an early return.
>
> As Jens has pointed out, the UB in this case is the caller calling fread
> with NULL - not in musl.
>
> And on a sidenote, I've always found - especially for the various mem*
> functions - accepting 0 size but not accepting NULL arg (when n is 0) to
> be a poor choice. A lot of the value that accepting 0 size provides is
> diminished by not accepting NULL.
>
> And this affects more than just libc, too. Compilers like gcc/clang will
> see a call like `memcmp(p, q, 0)` and will ""determine"" `p` and `q` are
> non-null (which can lead to deleting any subsequent null-checks on those
> pointers).
>
> But anyways, that was just a small rant.
>
> As things currently are, *even if* musl deal with the NULL pointer - any
> caller calling fread with NULL is still in danger from compilers and
> needs to fix it on their side.
>
> - NRK
>

[-- Attachment #2: Type: text/html, Size: 2658 bytes --]