Re: [PATCH 3/4] sh: Add SECCOMP_FILTER

[Rich Felker]

On Wed, Sep 02, 2020 at 11:56:04PM -0400, Rich Felker wrote:
> On Sat, Aug 29, 2020 at 01:09:43PM +0200, John Paul Adrian Glaubitz wrote:
> > Hi!
> > 
> > On 8/29/20 2:49 AM, Rich Felker wrote:
> > > This restored my ability to use strace
> > 
> > I can confirm that. However ...
> > 
> > > and I've written and tested a minimal strace-like hack using
> > > SECCOMP_RET_USER_NOTIF that works as
> > > expected on both j2 and qemu-system-sh4, so I think the above is
> > > correct.
> > 
> > The seccomp live testsuite has regressed.
> > 
> > [...]
> > Test 58-live-tsync_notify%%001-00001 result:   FAILURE 58-live-tsync_notify 6 ALLOW rc=14
> 
> This is similar to 51.
> 
> I think the commonality of all the failures is that they deal with
> return values set by seccomp filters for blocked syscalls, which are
> getting clobbered by ENOSYS from the failed syscall here. So I do need
> to keep the code path that jumps over the actual syscall if
> do_syscall_trace_enter returns -1, but that means
> do_syscall_trace_enter must now be responsible for setting the return
> value in non-seccomp failure paths.
> 
> I'll experiment to see what's still needed if that change is made.

OK, I think I have an explanation for the mechanism of the bug, and it
really is a combination of the 2008 bug (confusion of r0 vs r3) and
the SECCOMP_FILTER commit. When the syscall_trace_entry code path is
in use, a syscall with argument 5 having value -1 causes
do_syscall_trace_enter to return -1 (because it returns regs[0], which
contains argument 5), which the change in entry-common.S interprets as
a sign to skip the syscall and jump to syscall_exit, and things blow
up from there. In particular, SYS_mmap2 is almost always called with
-1 as the 5th argument (fd), and this is even more common on nommu
where SYS_brk does not work.

I'll follow up with a new proposed patch.

Rich