Re: [PATCH 2.6.25-rc2 3/9] config: Improve init/Kconfig help descriptions - namespaces

[Pavel Emelyanov <xemul@openvz.org>]

Nick Andrew wrote:
> On Wed, Feb 20, 2008 at 03:23:05PM +0300, Pavel Emelyanov wrote:
>>> +	  This is used by container systems (i.e. vservers).
>>> +	  Tasks in the container are placed in the PID namespace
>>> +	  corresponding to the container, and can only see or
>>> +	  affect processes in the same PID namespace.
>> same of one of child namespaces. In other words when you create
>> a new pid namespace, you still see the tasks from this new one,
>> but the tasks from this one, doesn't see yours :)
> 
> Due to the hierarchial nature, I see. I'm still trying to grok
> it. Would it be adequate to describe what a process _cannot_
> do? i.e.
> 
>           This is used by container systems (i.e. vservers).
>           Tasks in the container are placed in the PID namespace
>           corresponding to the container, and cannot see or
>           affect processes in any parent PID namespaces.

This looks better.

> Or maybe I should say both what it cannot do and what it can,
> so readers don't have to use their imagination much :-)

I think this would be redundant.

> Let's see if I understand how it works with an example. Say we've
> got a hierarchy of PID namespaces ... pidA/pidB/pidC and a new
> process created in pidC. This new process may have pid 18925 in
> pidA, 2263 in pidB and 56 in pidC?

Yes.

> So if there's another process running in pidC, the first process
> may be signaled as pid 56, and if a process is running in pidB
> it would be 2263 and not 56. Can a process running in pidB see
> all processes running in pidC only with their pidB PIDs?

Yes it can.

> Now, a process A running in pidA can send a signal to a process
> C running in pidC but not vice-versa. 

Right.

> Process C cannot know where the signal came from. 

Signal delivery is not yet finished, but this is what we
plan to do.

> Is there any kernel mechanism
> which normally would provide that kind of information to
> process C but which breaks when PID namespaces are used,
> because there's no way to name process A from the context
> of pidC ?

If the process from pidC receives a signal from pidB or pidA
then the appropriate siginfo will (ok - should :)) contain the 
SI_KERNEL code and zero pid, as if the kernel shoot this process.

> Nick.
>