Re: [RFC, PATCH 1/3] introduce SYS_CLONE_MASK

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

Oleg Nesterov <oleg@tv-sign.ru> writes:

> On 04/08, Eric W. Biederman wrote:
>
>> If we are going to have kernel only flags please use an additional
>> argument to do_fork and copy_process.
>
> Yes, we can do this. But we have a number of architectures which use
> sys_clone() to implement kernel_thread(). It would be nice to have an
> architecture neutral kernel_thread() implementation as you proposed.
> We should change all of them if we want to add a new parameter to
> do_fork().
>
> Perhaps it is better to add reparent_kthread() (next patch) to kthread()
> and forget about CLONE_KERNEL_THREAD.

Please. 

> Anyway, re-parenting to swapper breaks pstree, it doesn't show kernel
> threads. And if ->parent == /sbin/init, we can't remove us from ->children
> (unless we forbid sub-thread-of-init exec). So the only safe change is
> set ->exit_state = -1.

Yes.  We certainly need ->exit_state = -1.
Earlier I had forgotten about second the use of ->children to update
the parent pointer of processes when their parent exits.

There is a practical question how much we care about pstree being
confused (I assume it doesn't crash).  If this is just a confusion
issue then I say go for it.  PPID == 0 is a very legitimate way to say
the kernel is the parent process.

There are a few more cases where we are likely to get PPID == 0 in the
future and /sbin/init already has that now.  Plus there is a lot of
historic precedent.  The odd part is PPID = 0 having multiple
children.

If we decide maintaining a tree is important I would much rather put
init_task on the task_list so we can see it in /proc then go the other
way around.

I would like a confirmation that it PPID == 0 is what is confusing
pstree just to make certain we haven't half filled in some field
in init_task and are thus giving in correct /proc output.  But that is
all the double checking I would do.

>> Your current scheme also has the bad side that if user space supplied
>> a kernel flag it is hard to detect it and return -EINVAL.  Which
>> limits future expansion.  Silently dropping clone flags is a real
>> pain, if you are trying to detect if a new flag has been implemented.
>
> Yes. But that is what we are doing now. copy_process() just ignores
> unknown flags.

Agreed.  I fixed that in sys_unshare but I should really submit a
patch to do the same for sys_clone at some point.

When know flags aren't implemented we certainly return -EINVAL.

Given that this line of work looks to fix the race that messes allows
a threaded init to generate unkillable zombies I can probably find
some time in the next while to work on it.

Eric