Re: RFC [patch 13/34] PID Virtualization Define new task_pid api

[ebiederm@xmission.com (Eric W. Biederman)]

Kirill Korotaev <dev@sw.ru> writes:

>>>In fact this is almost what OpenVZ does for half a year, both containers and
>>>VPIDs.
>>>But it is very usefull to see process tree from host system. To be able to use
>>>std tools to manage containers from host (i.e. ps, kill, top, etc.). So it is
>>>much more convinient to have 2 pids. One globally unique, and one for
> container.
>> There are two issues here.
>> 1) Monitoring.  (ps, top etc)
>> 2) Control (kill).
>> For monitoring you might need to patch ps/top a little but it is doable
> without
>> 2 pids.
>> For kill it is extremely rude to kill processes inside of a nested pid space.
>> There are other solutions to the problem.
> it is not always good idea to fix the tools everytime new functionality is
> involved. why do you think there are no more tools except for ps,top,kill? will
> you fix it all?

No.  In my implementation the nested pid space has one pid that essentially
looks like a threaded process.  So ps and top can see that something is there
but you aren't spammed with all of the pids.  In addition for top the cpu utilization
for all of the pids are shown for that one pid (just like the thread group leader of
on a threaded process).  So there is no confusion.

If you want the detailed information you can either chroot to an environment
where the appropriate version of /proc is available.  Or you can modify your
tools.

> Another example, when you have problems in your VPS it is very convinient to
> attach with strace/gdb/etc from the host. Will you patch it as well?
>
> OpenVZ big advantage is this ease of administering compared to VM approach and
> it is not good idea to forbid this. If you have broken VM you have problems, in
> OpenVZ you have control over VPSs.
>
>> It is undesireable to make it too easy to communicate through the barrier
> because
>> then applications may start to take advantage of it and then depend on
>> it and you will have lost the isolation that the container gives you.
> in OpenVZ we have VPSs fully isolated between each other.
> But host system has access to some of VPS resources such as files, processes,
> etc. I understand your concern which is related to checkpointing, yeah?

There areas.
1) Checkpointing.
2) Isolation for security purposes.
   There may be secrets that the sysadmin should not have access to.
3) Nesting of containers, (so they are general purpose and not special hacks).

The vserver way of solving some of these problems is to provide a way
to enter the guest.  I would rather have some explicit operation that puts
you into the guest context so there is a single point where we can tackle
the nested security issues, than to have hundreds of places we have to
look at individually.

Eric