[bug?] clone(CLONE_IO) failing after kernel commit commit ef2c41cf38a7

[Christian Brauner]

On Tue, May 05, 2020 at 09:59:38AM +0200, Christian Brauner wrote:
> On Tue, May 05, 2020 at 09:49:50AM +0200, Florian Weimer wrote:
> > * Jan Stancek via Libc-alpha:
> > 
> > > I'm seeing an issue with CLONE_IO and libc' clone() on ppc64le,
> > > where flags parameter appears to be sign extended before it's passed
> > > to kernel syscall.
> > 
> > Interesting, thanks for reporting this.  The manual page clearly
> > documents the interface as;
> > 
> >        int clone(int (*fn)(void *), void *child_stack,
> >                  int flags, void *arg, ...
> >                  /* pid_t *ptid, void *newtls, pid_t *ctid */ );
> > 
> > But the kernel uses unsigned long for clone_flags.  This looks like an
> > unintended userspace ABI breakage.
> > 
> > Rather than dropping the invalid flags check in the kernel (having the
> > check is valuable), I think the parameter should be changed to int or
> > unsigned int, or the flags check should be written in such a way that
> > it disregards bits that result from sign extensions: fail if
> > clone_flags != (int) clone_flags, otherwise set clone_flags = 0xFFFFFFFF.
> 
> Let me take a quick look. This just affects legacy clone it seems. Maybe
> we can do something to avoid that breakage without having to change
> extensions. Legacy clone shouldn't care about invalid flag checking
> since it never used to do that anyway, only clone3() does.

Oh I see. Ugh.

Christian