[Devel] Re: [RFC][PATCH][v2] Define/use siginfo_from_ancestor_ns()

[Oleg Nesterov]

On 11/18, Sukadev Bhattiprolu wrote:
>
> |
> | > +static inline int siginfo_from_ancestor_ns(siginfo_t *info)
> | > +{
> | > +	return SI_FROMUSER(info) && (info->si_pid == 0);
> | > +}
> |
> | Yes, this is problem... I doubt we can rely on !si_pid here.
> | More on this later.
>
> BTW, rather than clearing SIG_FROM_USER in send_signal(), can we
> keep it till we dequeue the signal ? Yes, collect_signal() would
> need to consider this flag. But when we dequeue, we can note that
> it was from user and use that in the siginfo_from_ancestor() ?

Yes! I thought about this too. As a last resort this should work
afaics. But we should be carefull, we have to fix rm_from_queue_full()
for example as well.

Another note. We can split SIG_FROM_USER (if we are going to use this
hack) into 2 flags: SIG_KILL_SUB_NS and SIG_MANGLE_SI_PID. We can
even put "struct pid *pid" into si_signo along with these bits if
we find some strange user which sends the signal on behalve of
the different task.

But personally, I'd prefer to make 3 simple patches for the start.
Then we can continue with these complications if needed. Sukadev,
please feel free to disagree with me. I am just trying to make
the first step reviewable and simple. No changes on dequeue path,
no -ENOMEM in send_signal().

1. Introduce SIG_FROM_USER (or whatever). Basically, the patch I
   sent. Except I'd relly like to see this code under CONFIG_
   just for documentation, but please feel free to ignore.

   So, with this patch send_signal() has "bool from_ancestor", which
   is not used so far. And we the fixup code after copy_siginfo()
   which clears the flags, or better yet just sets .si_signo = sig.

2. Now we change send_signal()

	+	if (from_ancestor && sig == SIGKILL)
	+		t->signal->flags &= ~SIGNAL_UNKILLABLE;
		if (!prepare_signal(...))
			return;

   and change copy_signal() to set SIGNAL_UNKILLABLE for
   cinit.

   From now cinit is protected from unwanted signals from
   its namespace, and the parent can always kill it with
   SIGKILL.

   Actually, I think this is enough to solve most problems,
   the further changes can be discussed later. OK, the only
   "real" problem is SIGSTOP, afaics. This looks solveable.

3. mangle .si_pid in send_signal(). Again, it is not clear
   what should we do with sys_rt_sigqueueinfo(), but there
   is no "obviously right" solution.

And I am really sorry for delay.

Oleg.

_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers