How much of a mess does OpenVZ make? ;) Was: What can OpenVZ do?

[Serge E. Hallyn]

Quoting Cedric Le Goater (legoater at free.fr):
> 
> > No, what you're suggesting does not suffice.
> 
> probably. I'm still trying to understand what you mean below :)
> 
> Man, I hate these hierarchicals pid_ns. one level would have been enough, 
> just one vpid attribute in 'struct pid*'

Well I don't mind - temporarily - saying that nested pid namespaces
are not checkpointable.  It's just that if we're going to need a new
syscall anyway, then why not go ahead and address the whole problem?
It's not hugely more complicated, and seems worth it.

> > Call
> > (5591,3,1) the task knows as 5591 in the init_pid_ns, 3 in a child pid
> > ns, and 1 in grandchild pid_ns created from there.  Now assume we are
> > checkpointing tasks T1=(5592,1), and T2=(5594,3,1).
> > 
> > We don't care about the first number in the tuples, so they will be
> > random numbers after the recreate. 
> 
> yes.
> 
> > But we do care about the second numbers.  
> 
> yes very much and we need a way set these numbers in alloc_pid()
> 
> > But specifying CLONE_NEWPID while recreating the process tree
> > in userspace does not allow you to specify the 3 in (5594,3,1).
> 
> I haven't looked closely at hierarchical pid namespaces but as we're
> using a an array of pid indexed but the pidns level, i don't see why 
> it shouldn't be possible. you might be right.
> 
> anyway, I think that some CLONE_NEW* should be forbidden. Daniel should
> send soon a little patch for the ns_cgroup restricting the clone flags
> being used in a container.

Uh, that feels a bit over the top.  We want to make this
uncheckpointable (if it remains so), not prevent the whole action.
After all I may be running a container which I don't plan on ever
checkpointing, and inside that container running a job which i do
want to migrate.

So depending on if we're doing the Dave or the rest-of-the-world
way :), we either clear_bit(pidns->may_checkpoint) on the parent
pid_ns when a child is created, or we walk every task being
checkpointed and make sure they each are in the same pid_ns.  Doesn't
that suffice?

-serge