Re: ptrace: seccomp: Return value when the call was already invalid

[Will Deacon <will@kernel.org>]

On Sat, Jul 04, 2020 at 09:56:50PM -0700, Kees Cook wrote:
> On Sat, Jul 04, 2020 at 01:33:56PM +0100, Will Deacon wrote:
> > On Fri, Jul 03, 2020 at 08:52:05AM -0700, Kees Cook wrote:
> > > On Fri, Jul 03, 2020 at 04:44:27PM +0100, Will Deacon wrote:
> > > > On Fri, Jul 03, 2020 at 08:17:19AM -0700, Kees Cook wrote:
> > > > > On Fri, Jul 03, 2020 at 09:39:14AM +0100, Will Deacon wrote:
> > > > > > diff --git a/arch/arm64/kernel/syscall.c b/arch/arm64/kernel/syscall.c
> > > > > > index 5f5b868292f5..a13661f44818 100644
> > > > > > --- a/arch/arm64/kernel/syscall.c
> > > > > > +++ b/arch/arm64/kernel/syscall.c
> > > > > > @@ -121,12 +121,10 @@ static void el0_svc_common(struct pt_regs *regs, int scno, int sc_nr,
> > > > > >  	user_exit();
> > > > > >  
> > > > > >  	if (has_syscall_work(flags)) {
> > > > > > -		/* set default errno for user-issued syscall(-1) */
> > > > > > -		if (scno == NO_SYSCALL)
> > > > > > -			regs->regs[0] = -ENOSYS;
> > > > > > -		scno = syscall_trace_enter(regs);
> > > > > > -		if (scno == NO_SYSCALL)
> > > > > > +		if (syscall_trace_enter(regs))
> > > > > >  			goto trace_exit;
> > > > > > +
> > > > > > +		scno = regs->syscallno;
> > > > > >  	}
> > > > > >  
> > > > > >  	invoke_syscall(regs, scno, sc_nr, syscall_table);
> > > > > 
> > > > > What effect do either of these patches have on the existing seccomp
> > > > > selftests: tools/testing/selftests/seccomp/seccomp_bpf ?
> > > > 
> > > > Tests! Thanks, I'll have a look.
> > > 
> > > Thanks!
> > > 
> > > (And either way, that this behavioral difference went unnoticed means we
> > > need to add a test to the selftests for this patch.)
> > 
> > Unsurprisingly, I don't think the tests go near this. I get 75/77 passes
> > on arm64 defconfig with or without these changes.
> 
> (What doesn't pass for you? I tried to go find kernelci.org test output,
> but it doesn't appear to actually run selftests yet?)
> 
> Anyway, good that the test output doesn't change, bad that seccomp has
> missed a corner of this architecture interface. (i.e. the entire
> TRACE_syscall fixture is dedicated to exercising the changing/skipping
> interface, but I see now that it doesn't at all exercise any area of
> ENOSYS results.)
> 
> > We could add a test, but then we'd have to agree on what it's supposed to
> > be doing ;)
> 
> Well, if you look at change_syscall() in seccomp_bpf.c (once you stop
> screaming) you'll likely share my desire to have more things that are
> common across architectures. ;) So, to that end, yes, please, let's
> define what we'd like to see, and then build out the (likely wildly
> different per-architecture expectations). If I read this thread
> correctly, we need to test:
> 
> 	syscall(-1), direct,  returns ENOSYS
> 	syscall(-10), direct, returns ENOSYS
> 	syscall(-1), SECCOMP_RET_TRACE+PTRACE_CONT, returns ENOSYS
> 	syscall(-10), SECCOMP_RET_TRACE+PTRACE_CONT, returns ENOSYS
> 	syscall(-1), ptrace+PTRACE_SYSCALL, returns ENOSYS
> 	syscall(-10), ptrace+PTRACE_SYSCALL, returns ENOSYS
> 
> do we need to double-check that registers before/after are otherwise
> unchanged too? (I *think* just looking at syscall return should be
> sufficient to catch the visible results.)

There's also the case where the tracer sets the system call to -1 to skip
it.

Will