[Devel] Re: [RFC][v3][PATCH 7/7] Define clone_with_pids syscall

[Oren Laadan]

Serge E. Hallyn wrote:
> Quoting Sukadev Bhattiprolu (sukadev at linux.vnet.ibm.com):
> 
> Bear with me as I try to get this straight...
> 
> Let's say we have the following task being checkpointed (btw, the
> actual checkpoint code hopefully will, as Oren mentioned, not
> be always checkpointing all of a task's pids, but only the
> pids from the checkpointer's pidns level an up - but that doesn't
> matter to you):
> 
> Either:
> 
> level no    |    pid
> 0               1001
> 1                50
> 2                 3
> 
> or
> 
> level no    |    pid
> 0               1001
> 1                50
> 2                 1
> 
> Now we want to restart that inside a container.  So for the
> first task we have a parent task:
> 
> level no       |     pid
> 0                    5009
> 1                    1000
> 2                     49
> 3                      2
> 
> calling clone_with_pid with upids {1001,50,3} to produce task:
> 
> level no       |     pid
> 0                   5010
> 1                   1001
> 2                    50
> 3                     3
> 
> So the numbers in your code for this case become:
> 
>> +	unum_pids = pid_set.num_pids;
>         unum_pids = 3
>> +	knum_pids = task_pid(current)->level + 1;
>         knum_pids = 3 + 1 = 4
> 
>> +	target_pids = kzalloc((knum_pids + 1) * sizeof(pid_t), GFP_KERNEL);
> 
> XXX First problem:
>         target_pids gets room for 5 pids, one more than we need.

The two issues are related, and this is intentional. The idea
was that when you use CLONE_NEWPID, you imply a new nesting level
and a pid==1 there. So you are not supposed to specify the pid
of that new level.

IOW, the parent specify the pid's of the child from the _parent's_
level and up (of the desired depth). CLONE_NEWPID creates a new
pidns level below the parent's, and that is not covered in the
array of pids.

By allocation an extra slot and forcing it to be 0, we ensure that
the case of CLONE_NEWPID is covered correctly. Clearly, if this
flag isn't set, then the extra slot is redundant (but doesn't hurt).

> 
>> +	j = knum_pids - unum_pids;
>         j = 1, so we copy the 3 pids in the right place.
> 
>> +	rc = copy_from_user(&target_pids[j], pid_set.target_pids, size);
>> +	if (rc) {
>> +		rc = -EFAULT;
>> +		goto out_free;
>> +	}
>> +
>> +	return target_pids;
> 
> 
> For the second one, we have a parent task
> 
> level no       |     pid
> 0                   5009
> 1                   1000
> 2                    49
> 
> calling clone_with_pid with CLONE_NEWPID and {1001,50,1} to produce:
> 
> level no       |     pid
> 0                   5010
> 1                   1001
> 2                    50
> 3                    1
> 
> So the numbers in your code become:
> 
>> +	unum_pids = pid_set.num_pids;
>         unum_pids = 3

This is a "bug" of the parent. The parent should specify the pids
from the parent's level only and up, and not include the new level
below that will be created. (After all, it will have to be 1).

So unum_pids = 3 will not do what you want; instead it will try to
create the process:

0	1001
1	50
2	1
3	1

And will fail, of course, because pid==1 at level 2 is already
taken.

Instead, parent should say use {1001, 50}.

>> +	knum_pids = task_pid(current)->level + 1;
>         knum_pids = 2 + 1 = 3
> 
>> +	target_pids = kzalloc((knum_pids + 1) * sizeof(pid_t), GFP_KERNEL);
> 
>         target_pids gets room for 4 pids
> 
>> +	j = knum_pids - unum_pids;
> 
> XXX Second problem:
>         j = 0, which is not right, should be 1?
> 
>> +	rc = copy_from_user(&target_pids[j], pid_set.target_pids, size);
>> +	if (rc) {
>> +		rc = -EFAULT;
>> +		goto out_free;
>> +	}
>> +
>> +	return target_pids;
> 
> So it looks like your usage of knum_pids isn't quite right.
> 
> (right?)

The main difference is that in the first case the parent forks
from level 3 and in the second case is forks from level 2.  And
the syscall is relative to the parent's nesting level.

> Or are you expecting userspace to not specify the 1, only 2
> upids?  In that case, you're requiring different behavior for
> pid 1 than any others, which is kind of weird?  Or is that just
> me?

Either that, or it's just Suka and myself :p

Oren.

_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers