[RFD] reboot / shutdown of a container

Daniel Lezcano daniel.lezcano at free.fr
Fri Jan 14 23:54:22 PST 2011

On 01/15/2011 12:11 AM, Bruno Prémont wrote:
> On Thu, 13 January 2011 Daniel Lezcano<daniel.lezcano at free.fr>  wrote:
>> On 01/13/2011 10:50 PM, Bruno Prémont wrote:
>>> On Thu, 13 January 2011 Daniel Lezcano<daniel.lezcano at free.fr>   wrote:
>>>> On 01/13/2011 09:09 PM, Bruno Prémont wrote:
>>>>> On Thu, 13 January 2011 Daniel Lezcano<daniel.lezcano at free.fr>    wrote:
>>>>>> in the container implementation, we are facing the problem of a process
>>>>>> calling the sys_reboot syscall which of course makes the host to
>>>>>> poweroff/reboot.
>>>>>> If we drop the cap_sys_reboot capability, sys_reboot fails and the
>>>>>> container reach a shutdown state but the init process stay there, hence
>>>>>> the container becomes stuck waiting indefinitely the process '1' to exit.
>>>>>> The current implementation to make the shutdown / reboot of the
>>>>>> container to work is we watch, from a process outside of the container,
>>>>>> the<rootfs>/var/run/utmp file and check the runlevel each time the file
>>>>>> changes. When the 'reboot' or 'shutdown' level is detected, we wait for
>>>>>> a single remaining in the container and then we kill it.
>>>>>> That works but this is not efficient in case of a large number of
>>>>>> containers as we will have to watch a lot of utmp files. In addition,
>>>>>> the /var/run directory must *not* mounted as tmpfs in the distro.
>>>>>> Unfortunately, it is the default setup on most of the distros and tends
>>>>>> to generalize. That implies, the rootfs init's scripts must be modified
>>>>>> for the container when we put in place its rootfs and as /var/run is
>>>>>> supposed to be a tmpfs, most of the applications do not cleanup the
>>>>>> directory, so we need to add extra services to wipeout the files.
>>>>>> More problems arise when we do an upgrade of the distro inside the
>>>>>> container, because all the setup we made at creation time will be lost.
>>>>>> The upgrade overwrite the scripts, the fstab and so on.
>>>>>> We did what was possible to solve the problem from userspace but we
>>>>>> reach always a limit because there are different implementations of the
>>>>>> 'init' process and the init's scripts differ from a distro to another
>>>>>> and the same with the versions.
>>>>>> We think this problem can only be solved from the kernel.
>>>>>> The idea was to send a signal SIGPWR to the parent of the pid '1' of the
>>>>>> pid namespace when the sys_reboot is called. Of course that won't occur
>>>>>> for the init pid namespace.
>>>>> Wouldn't sending SIGKILL to the pid '1' process of the originating PID
>>>>> namespace be sufficient (that would trigger a SIGCHLD for the parent
>>>>> process in the outer PID namespace.
>>>> This is already the case. The question is : when do we send this signal ?
>>>> We have to wait for the container system shutdown before killing it.
>>> I meant that sys_reboot() would kill the namespace's init if it's not
>>> called from boot namespace.
>>> See below
>>>>> (as far as I remember the PID namespace is killed when its 'init' exits,
>>>>> if this is not the case all other processes in the given namespace would
>>>>> have to be killed as well)
>>>> Yes, absolutely but this is not the point, reaping the container is not
>>>> a problem.
>>>> What we are trying to achieve is to shutdown properly the container from
>>>> inside (from outside will be possible too with the setns syscall).
>>>> Assuming the process '1234' creates a new process in a new namespace set
>>>> and wait for it.
>>>> The new process '1' will exec /sbin/init and the system will boot up.
>>>> But, when the system is shutdown or rebooted, after the down scripts are
>>>> executed the kill -15 -1 will be invoked, killing all the processes
>>>> expect the process '1' and the caller. This one will then call
>>>> 'sys_reboot' and exit. Hence we still have the init process idle and its
>>>> parent '1234' waiting for it to die.
>>> This call to sys_reboot() would kill "new process '1'" instead of trying to
>>> operate on the HW box.
>>> This also has the advantage that a container would not require an informed
>>> parent "monitoring" it from outside (though it would not be restarted even if
>>> requested without such informed outside parent).
>> Oh, ok. Sorry I misunderstood.
>> Yes, that could be better than crossing the namespace boundaries.
>>>> If we are able to receive the information in the process '1234' : "the
>>>> sys_reboot was called in the child pid namespace", we can take then kill
>>>> our child pid.  If this information is raised via a signal sent by the
>>>> kernel with the proper information in the siginfo_t (eg. si_code
>>>> solution will be generic for all the shutdown/reboot of any kind of
>>>> container and init version.
>>> Could this be passed for a SIGCHLD? (when namespace is reaped, and received
>>> by 1234 from above example assuming sys_reboot() kills the "new process '1'")
>> Yes, that sounds a good idea.
>>> Looks like yes, but with the need to define new values for si_code (reusing
>>> LINUX_REBOOT_CMD_* would certainly clash, no matter which signal is choosen).
> I would just map both to the same thing...
>> CLD_REBOOT_CMD_RESTART2 (what about the cmd buffer, shall we ignore it ?)
> The cmd buffer could be passed via si_ptr if we want it, otherwise it would
> be the same as for CLD_REBOOT_CMD_RESTART (which would have si_ptr set to NULL
> in case no si_code differentiation is needed)
> I don't think kexec makes any sense inside a container, such a sys_reboot()
> call should probably fail or fallback to _RESTART
>> CLD_REBOOT_CMD_SW_SUSPEND (useful for the future checkpoint/restart)
> Looks reasonable
>> for a non-init pid namespace, no ?
> I haven't looked at how/when the state set by these is checked, but it could
> keep its meaning and a CAD shortcut would act on the container to which the
> active task on the given tty belongs. (so as if the process which would have
> gotten SIGINT had issued sys_reboot(LINUX_REBOOT_CMD_RESTART), permissions
> set aside)

That makes sense.

Thanks Bruno !

   -- Daniel (cooking a patch ... ;)

More information about the Containers mailing list