Re: [PATCH net-next] [RFC] netns: enable cross-ve Unix sockets

[Daniel Lezcano <dlezcano@fr.ibm.com>]

Denis V. Lunev wrote:
> On Wed, 2008-10-01 at 15:46 +0200, Daniel Lezcano wrote:
>> Denis V. Lunev wrote:
>>> On Wed, 2008-10-01 at 14:31 +0200, Daniel Lezcano wrote:
>>>> Pavel Emelyanov wrote:
>>>>>> So there are 2 cases:
>>>>>>   * full isolation : restriction on VPS
>>>>>>   * partial isolation : no restriction but *perhaps* problem when migrating
>>>>>>
>>>>>> Looks like we need an option per namespace to reduce the isolation for 
>>>>>> af_unix sockets :)
>>>>>>   - on (default): current behaviour => full isolation
>>>>>>   - off : partial isolation
>>>>> You mean some sysctl, that enables/disables this check in unix_find_socket_byinode?
>>>> Yes.
>>> I do not see much sense with sysctl as:
>>> - check (cross-connected sockets) is required as we can start namespace
>>>   with already opened socket
>> Check when checkpointing ? If you inherit a socket from your parent 
>> namespace, this socket belongs to your parent and you should not 
>> checkpoint it, no ?
>>
>> In case you allow cross-connected sockets, this check is mandatory I agree.
>>
>>> - this kind of sharing is not implicit but explicit as normal isolated
>>>   containers _must_ have separate filesystems. In this case this
>>>   sharing requires explicit host administrator action to link socket
>>>   between containers
>> What are "normal isolated containers" ? Are they OpenVZ containers ? 
>> These containers belong to the system containers family. What happens 
>> with application containers, if I want to share the filesystem without 
>> breaking the isolation of the afunix sockets ?
> 
> then you are doomed as you will have a FIFO opened from 2 namespaces and
> checking the absences of external references is still mandatory
>> The current code provides full isolation and this is in mainline. I 
>> don't think it is reasonable to change that. What I propose is to keep 
>> the current behaviour.
>>
>> When you create a network namespace, you can change the behaviour inside 
>> this namespace via /proc/sys/net/unix/isolated (for example).
>>
>> This option allows:
>> 1 - to connect to af_unix not belonging to the container
>> 2 - to accept af_unix connection from outside the container (avoid a 
>> container to forbid the checkpoint of another container);
> 
> this should be at least per/namespace option controlled from parent
> container from my POW

Yes per namespace, I agree.

If the option is controlled by the parent and it is done by sysctl, you 
will have to make proc/sys per namespace like Pavel did with /proc/net, no ?