sparc/ppc/arm compat siginfo ABI regressions: sending SIGFPE via kill() returns wrong values in si_pid and si_uid

[Dave Martin]

On Fri, Apr 13, 2018 at 07:50:17PM +0100, Russell King - ARM Linux wrote:
> On Fri, Apr 13, 2018 at 07:35:38PM +0100, Dave Martin wrote:
> > If that's the case though, I don't see how a userspace testsuite is
> > hitting this code path.  Maybe I've misunderstood the context of this
> > thread.
> 
> It isn't hitting this exact case.
> 
> The userspace testsuite is hitting an entirely different case:
> 
> 	kill(getpid(), SIGFPE);
> 
> As one expects, this generates a SIGFPE to the current process, which
> then inspects the siginfo structure.  Being a userspace generated
> signal, si_code is set to SI_USER, which has the value 0.
> 
> With FPE_FIXME defined to zero, as Eric has done:
> 
> enum siginfo_layout siginfo_layout(int sig, int si_code)
> {
>         enum siginfo_layout layout = SIL_KILL;
>         if ((si_code > SI_USER) && (si_code < SI_KERNEL)) {
> ...
>         } else {
> ...
> #ifdef FPE_FIXME
>                 if ((sig == SIGFPE) && (si_code == FPE_FIXME))
>                         layout = SIL_FAULT;
> #endif
>         }
>         return layout;
> }
> 
> This causes siginfo_layout() to return SIL_FAULT for this userspace
> generated signal, rather than the correct SIL_KILL.
> 
> This affects which fields we copy to userspace.
> 
> SI_USER is defined to pass si_pid and si_uid to the userspace process,
> which on ARM are the first two consecutive 32-bit quantities in the
> union, which is done when siginfo_layout() returns SIL_KILL.  However,
> when SIL_FAULT is returned, we only copy si_addr in the union, which
> on ARM is just one 32-bit quantity.
> 
> Consequently, userspace sees a correct value for si_pid, and si_uid
> remains set to whatever was there in userspace.  In the case of the
> strace program, that's zero.  This means if you run the strace
> testsuite as root, the problem doesn't appear, but if you run it as
> a non-root user, it will.
> 
> So, the testsuite case has little to do with the behaviour of the VFP
> handling - it's to do with the behaviour of the kernel's signal handling.

Oh, right.  So, going back to the unhandled VFP bounce question,
is it reasonable for that to be a SIGKILL?  That avoids the question
of what si_code userspace should see, because userspace doesn't get
to see siginfo at all in that case: it's dead.

Or do we hit this in real situations that we want userspace to bail out
of more gracefully?

Cheers
---Dave