Re: [musl] [Bug] Do not ignore membarrier return code

[Rich Felker <dalias@libc.org>]

On Mon, Mar 23, 2020 at 05:10:40PM +0100, Julio Guerra wrote:
> Hello,
> 
> The implementation of dlopen() uses membarrier() (
> https://git.musl-libc.org/cgit/musl/tree/ldso/dynlink.c#n1579) while
> currently forbidden by the default docker seccomp profile.
> 
> I perfectly understand that it's on docker's end and I suggested them to
> add it in this PR <https://github.com/moby/moby/pull/40731> but such a
> critical syscall shouldn't be silently ignored. And it for example leads to
> random segfaults on nodejs. I also saw opened qemu issues related to
> membarrier + alpine.
> 
> dlopen() should therefore fail when membarrier fails (ie. in this case
> when __membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED,
> 0) != 0).

At that point it's past the point where failure is possible; making it
possible would be rather nontrivial. But you missed that it can't
fail. musl has a very heavy fallback implemementation for the case
where it's not implemented or somehow fails; see
src/linux/membarrier.c.

However, the reason you're seeing the failure is something of a bug in
musl -- registration of intent to use membarrier is only done on first
pthread_create. That's okay because it's only needed at all if the
process is multithreaded. However, dlopen is calling it
unconditionally even if the process is not multithreaded, and thereby
getting a spurious failure since it wasn't registered yet. It should
just be fixed not to make the fall if it's not multithreaded.

Rich