Re: [PATCH 0/2] Send a SIGCHLD to the init's pid namespace parent when reboot

[Oleg Nesterov <oleg@redhat.com>]

On 08/14, Serge E. Hallyn wrote:
>
> Quoting Oleg Nesterov (oleg@redhat.com):
> > On 08/11, Daniel Lezcano wrote:
> > >
> > > In the case of a VPS, when we shutdown/halt/reboot the container, the
> > > reboot utility will invoke the sys_reboot syscall which has the bad
> > > effect to reboot the host.
> >
> > Stupid question. Can't sys_reboot() simply kill init (and thus the whole
> > pid_ns) in this case?
>
> The goal is to be able to distinguish a request for reboot from shutdown.
> If we just kill the init, then the parent of init (the container monitor)
> cannot restart the container to emulate reboot.

OK, thanks.

What if init reports the reason it was killed?

Ignoring LINUX_REBOOT_CMD_CAD_/etc, I mean, roughly,

	- add "int reboot_cmd" into struct pid_namespace

	- sys_reboot(cmd) does

		if (!global_namespace) {
			task_active_pid_ns(current)->reboot_cmd = cmd;
			sigkill_my_init();
		}

	- change zap_pid_ns_processes() to do

		if (pid_ns->reboot_cmd)	// approximately
			current->exit_state = pid_ns->reboot_cmd;

Then its parent can look at status after wait(&status).


Not that I think this is very nice, but signals are not reliable.
And once again, SIGCHLD doesn't queue. And, perhaps this doesn't
matter, but sys_reboot() sends SIGCHLD and returns -EPERM, this
can confuse the container.


In any case. If you want to send a signal, please do not introduce
SA_CLDREBOOT. Please do not play with ptrace or __wake_up_parent, this
is meaningless. Just fill siginfo and send SIGCHLD unconditionally.

Oleg.