Re: [PATCH 1/4] signals/sigaltstack: If SS_AUTODISARM, bypass on_sig_stack

[Stas Sergeev <stsp@list.ru>]

09.05.2016 04:32, Andy Lutomirski пишет:
> On May 7, 2016 7:38 AM, "Stas Sergeev" <stsp@list.ru> wrote:
>> 03.05.2016 20:31, Andy Lutomirski пишет:
>>
>>> If a signal stack is set up with SS_AUTODISARM, then the kernel
>>> inherently avoids incorrectly resetting the signal stack if signals
>>> recurse: the signal stack will be reset on the first signal
>>> delivery.  This means that we don't need check the stack pointer
>>> when delivering signals if SS_AUTODISARM is set.
>>>
>>> This will make segmented x86 programs more robust: currently there's
>>> a hole that could be triggered if ESP/RSP appears to point to the
>>> signal stack but actually doesn't due to a nonzero SS base.
>>>
>>> Signed-off-by: Stas Sergeev <stsp@list.ru>
>>> Cc: Al Viro <viro@zeniv.linux.org.uk>
>>> Cc: Aleksa Sarai <cyphar@cyphar.com>
>>> Cc: Amanieu d'Antras <amanieu@gmail.com>
>>> Cc: Andrea Arcangeli <aarcange@redhat.com>
>>> Cc: Andrew Morton <akpm@linux-foundation.org>
>>> Cc: Andy Lutomirski <luto@amacapital.net>
>>> Cc: Borislav Petkov <bp@alien8.de>
>>> Cc: Brian Gerst <brgerst@gmail.com>
>>> Cc: Denys Vlasenko <dvlasenk@redhat.com>
>>> Cc: Eric W. Biederman <ebiederm@xmission.com>
>>> Cc: Frederic Weisbecker <fweisbec@gmail.com>
>>> Cc: H. Peter Anvin <hpa@zytor.com>
>>> Cc: Heinrich Schuchardt <xypron.glpk@gmx.de>
>>> Cc: Jason Low <jason.low2@hp.com>
>>> Cc: Josh Triplett <josh@joshtriplett.org>
>>> Cc: Konstantin Khlebnikov <khlebnikov@yandex-team.ru>
>>> Cc: Linus Torvalds <torvalds@linux-foundation.org>
>>> Cc: Oleg Nesterov <oleg@redhat.com>
>>> Cc: Palmer Dabbelt <palmer@dabbelt.com>
>>> Cc: Paul Moore <pmoore@redhat.com>
>>> Cc: Pavel Emelyanov <xemul@parallels.com>
>>> Cc: Peter Zijlstra <peterz@infradead.org>
>>> Cc: Richard Weinberger <richard@nod.at>
>>> Cc: Sasha Levin <sasha.levin@oracle.com>
>>> Cc: Shuah Khan <shuahkh@osg.samsung.com>
>>> Cc: Tejun Heo <tj@kernel.org>
>>> Cc: Thomas Gleixner <tglx@linutronix.de>
>>> Cc: Vladimir Davydov <vdavydov@parallels.com>
>>> Cc: linux-api@vger.kernel.org
>>> Cc: linux-kernel@vger.kernel.org
>>> Signed-off-by: Andy Lutomirski <luto@kernel.org>
>>> ---
>>>    include/linux/sched.h | 12 ++++++++++++
>>>    1 file changed, 12 insertions(+)
>>>
>>> diff --git a/include/linux/sched.h b/include/linux/sched.h
>>> index 2950c5cd3005..8f03a93348b9 100644
>>> --- a/include/linux/sched.h
>>> +++ b/include/linux/sched.h
>>> @@ -2576,6 +2576,18 @@ static inline int kill_cad_pid(int sig, int priv)
>>>     */
>>>    static inline int on_sig_stack(unsigned long sp)
>>>    {
>>> +       /*
>>> +        * If the signal stack is AUTODISARM then, by construction, we
>>> +        * can't be on the signal stack unless user code deliberately set
>>> +        * SS_AUTODISARM when we were already on the it.
>> "on the it" -> "on it".
>>
>> Anyway, I am a bit puzzled with this patch.
>> You say "unless user code deliberately set
>>
>> SS_AUTODISARM when we were already on the it"
>> so what happens in case it actually does?
>>
> Stack corruption.  Don't do that.
Only after your change, I have to admit. :)

>> Without your patch: if user sets up the same sas - no stack switch.
>> if user sets up different sas - stack switch on nested signal.
>>
>> With your patch: stack switch in any case, so if user
>> set up same sas - stack corruption by nested signal.
>>
>> Or am I missing the intention?
> The intention is to make everything completely explicit.  With
> SS_AUTODISARM, the kernel knows directly whether you're on the signal
> stack, and there should be no need to look at sp.  If you set
> SS_AUTODISARM and get a signal, the signal stack gets disarmed.  If
> you take a nested signal, it's delivered normally.  When you return
> all the way out, the signal stack is re-armed.
>
> For DOSEMU, this means that no 16-bit register state can possibly
> cause a signal to be delivered wrong, because the register state when
> a signal is raised won't affect delivery, which seems like a good
> thing to me.
Yes, but doesn't affect dosemu1 which doesn't use SS_AUTODISARM.
So IMHO the SS check should still be added, even if not for dosemu2.

> If this behavior would be problematic for you, can you explain why?
Only theoretically: if someone sets SS_AUTODISARM inside a
sighandler. Since this doesn't give EPERM, I wouldn't deliberately
make it a broken scenario (esp if it wasn't before the particular change).
Ideally it would give EPERM, but we can't, so doesn't matter much.
I just wanted to warn about the possible regression.