Re: Replacing a_crash() ?

[Rich Felker <dalias@libc.org>]

On Mon, Sep 17, 2018 at 05:24:15PM +0200, Markus Wichmann wrote:
> On Sun, Sep 16, 2018 at 11:23:17PM -0400, Rich Felker wrote:
> > Now that we have an abort() that reliably terminates with uncatchable
> > SIGABRT, I've been thinking about replacing the a_crash() calls in
> > musl (which are usually an instruction generating SIGILL or SIGSEGV)
> > with calls to the uncatchable tail of abort(), which I would factor
> > off as a __forced_abort() function.
> > 
> > In case it's not clear, the reason for not just calling abort() is
> > that too many programs catch it, and catching it is even encouraged.
> > Catchability is a problem with the current approach too, since
> > a_crash() is used in places where process state is known to be
> > dangerously corrupt and likely under attacker control; eliminating it
> > is one of the potential goals of switching to __forced_abort().
> > 
> > Are there any objections to making such a change? So far I've gotten
> > mostly positive feedback -- SIGABRT is more telling of what's happened
> > than SIGSEGV/SIGILL. It would also get rid of the ugly misplacement of
> > a_crash() (no longer needed) in "atomic.h" and the inclusion of
> > "atomic.h" in some files where it makes no sense without knowing it's
> > where a_crash() is defined.
> > 
> > For i386, some nontrivial work would be needed to make abort's tail
> > perform syscalls with int $128 rather than the vdso, which is unsafe
> > since the pointer to it may have been subverted. On other archs,
> > inline syscalls are fully inline. I'd probably add a
> > NEED_FAILSAFE_SYSCALL macro to define before including "syscall.h" and
> > have arch/i386/syscall_arch.h adjust the asm string based on it; this
> > is more maintainable than writing an asm version of the function.
> 
> Simple checklist for whether to perform a change or not:
> 
> 1. Does the change fix problems? Check (namely, maintainability,
> legibility, understandability of problems).

It slightly reduces amount of per-arch asm needed. (Actually not,
because there's a "generic" a_crash() that writes to a volatile null
pointer, but it doesn't work on nommu.) It also gets rid of atomic.h
dependencies.

> 2. Does the change introduce problems? Unlikely. Someone might subvert
> __forced_abort(), but then, someone might catch SIGILL, so we haven't
> gone anywhere.

I was thinking more like friendliness to debugging workflows; that's
the motivation for not using SIGKILL, which always would have been
easy. Subverting __forced_abort() for static linking is of course
easy; for dynamic you'd have to modify the mapped libc.so since it
would be a direct call.

> 3. Is the change compatible with old programs? No, but a_crash() was
> never a defined interface, so any program catching it was walking on
> thin ice, anyway.
> 
> So that's two green lights and a don't care, so please go ahead.

Indeed, catching it was never intended to be supported usage.

Rich