Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
From:   Keno Fischer <keno@juliacomputing.com>
Date:   Fri, 22 May 2020 21:01:01 -0400
Message-ID: <CABV8kRxA9mXPZwtYrjbAfOfFewhABHddipccgk-LQJO+ZYu4Xg@mail.gmail.com>
Subject: ptrace: seccomp: Return value when the call was already invalid
To:     Linux Kernel Mailing List <linux-kernel@vger.kernel.org>
Cc:     Oleg Nesterov <oleg@redhat.com>, Will Deacon <will@kernel.org>,
        Kees Cook <keescook@chromium.org>,
        Andy Lutomirski <luto@amacapital.net>,
        Will Drewry <wad@chromium.org>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

I'm seeing the following while porting a ptracer from
x86_64 to arm64 (cc'ing arm64 folks, but in this case
x86_64 is the odd one out, I think other archs would
be consistent with arm64).

Consider userspace code like the following:
```
int ret = syscall(-10, 0);
assert(ret == -ENOSYS);
```

(Never mind the fact that this is something userspace
shouldn't do, I saw this in our test suite that tests
corner cases where the ptracer shouldn't affect behavior).

Now, if we have a seccomp filter that simply does
SECCOMP_RET_TRACE, and a ptracer that simply
does PTRACE_CONT, then the assert will fire/fail on arm64,
but not on x86_64.

The reason this happens is that the return value gets set
early on x86_64, but this is not possible on arm64,
because doing so would clobber the first argument
register that it shares. As a result, no return value is
set and `ret` retains the value that the first syscall
argument used to have.

I can work around this of course, but I guess my
question is whether this is expected/ok,
or you would expect an active ptracer that does not
touch the registers not to affect behavior.

Interestingly, arm64 does do something different
if the syscall is -1 rather than -10, where early
in the ptrace stop it does.
```
/* set default errno for user-issued syscall(-1) */
if (scno == NO_SYSCALL)
    regs->regs[0] = -ENOSYS;
```

I'm not sure that's great either since the ptracer
may want to inspect x0 and arm64 does not
make orig_x0 available via ptrace. To me
this indicates that maybe this was intended
to apply to any syscall skipped here, not
just -1 (the different comes from the fact
that seccomp considers any negative
syscall a skip/fail, but on syscall-entry
stops arm64 only considers a literal -1
a skip).

On the other hand if this is deemed expected,
I'll go ahead and submit a man-page patch to at
least document this architecture difference.

Keno