Re: [RFC][PATCH 6/6][v3] Protect cinit from blocked fatal signals

[Oleg Nesterov <oleg@redhat.com>]

On 12/20, Sukadev Bhattiprolu wrote:
> 
> +static int sig_unkillable(struct signal_struct *signal, int sig)
> +{
> +	if (signal->flags & SIGNAL_UNKILLABLE_FROM_NS)
> +		return !sig_kernel_only(sig);
> +
> +	/*
> +	 * We must have dropped SIGKILL/SIGSTOP in sig_ignored()
> +	 * TODO: Remove BUG_ON().
> +	 */
> +	BUG_ON(signal->flags & SIGNAL_UNKILLABLE && sig_kernel_only(sig));
> +
> +	return (signal->flags & SIGNAL_UNKILLABLE);
> +}
> +
>  int get_signal_to_deliver(siginfo_t *info, struct k_sigaction *return_ka,
>  			  struct pt_regs *regs, void *cookie)
>  {
> @@ -1907,9 +1943,10 @@ relock:
>
>  		/*
>  		 * Global init gets no signals it doesn't want.
> +		 * Container-init gets no signals it doesn't want from same
> +		 * container.
>  		 */
> -		if (unlikely(signal->flags & SIGNAL_UNKILLABLE) &&
> -		    !signal_group_exit(signal))
> +		if (sig_unkillable(signal, signr) && !signal_group_exit(signal))
>  			continue;

Again, I do not understand why do we need SIGNAL_UNKILLABLE_FROM_NS.

I thought about the change in get_signal_to_deliver() during the
previous discussion, and I think what we need is:

		if (unlikely(signal->flags & SIGNAL_UNKILLABLE) &&
		    !sig_kernel_only(sig))
			continue;

and this was yet another reason for "protect init from unwanted signals more".

Because, if we see SIGKILL/SIGSTOP here, this means that the signal
was sent from the parent ns, or it was generated "internally", for
example by sys_exit_group().

Oleg.