On 12/2/25 6:56 PM, Bobby Eshleman wrote:
> On Tue, Dec 02, 2025 at 11:18:14AM +0100, Paolo Abeni wrote:
>> On 11/27/25 8:47 AM, Bobby Eshleman wrote:
>>> @@ -674,6 +689,17 @@ static int vhost_vsock_dev_open(struct inode *inode, 
>>> struct file *file)
>>>             goto out;
>>>     }
>>>  
>>> +   net = current->nsproxy->net_ns;
>>> +   vsock->net = get_net_track(net, &vsock->ns_tracker, GFP_KERNEL);
>>> +
>>> +   /* Store the mode of the namespace at the time of creation. If this
>>> +    * namespace later changes from "global" to "local", we want this vsock
>>> +    * to continue operating normally and not suddenly break. For that
>>> +    * reason, we save the mode here and later use it when performing
>>> +    * socket lookups with vsock_net_check_mode() (see vhost_vsock_get()).
>>> +    */
>>> +   vsock->net_mode = vsock_net_mode(net);
>>
>> I'm sorry for the very late feedback. I think that at very least the
>> user-space needs a way to query if the given transport is in local or
>> global mode, as AFAICS there is no way to tell that when socket creation
>> races with mode change.
> 
> Are you thinking something along the lines of sockopt?

I'd like to see a way for the user-space to query the socket 'namespace
mode'.

sockopt could be an option; a possibly better one could be sock_diag. Or
you could do both using dumping the info with a shared helper invoked by
both code paths, alike what TCP is doing.
>> Also I'm a bit uneasy with the model implemented here, as 'local' socket
>> may cross netns boundaris and connect to 'local' socket in other netns
>> (if I read correctly patch 2/12). That in turns AFAICS break the netns
>> isolation.
> 
> Local mode sockets are unable to communicate with local mode (and global
> mode too) sockets that are in other namespaces. The key piece of code
> for that is vsock_net_check_mode(), where if either modes is local the
> namespaces must be the same.

Sorry, I likely misread the large comment in patch 2:

https://lore.kernel.org/netdev/[email protected]/

>> Have you considered instead a slightly different model, where the
>> local/global model is set in stone at netns creation time - alike what
>> /proc/sys/net/ipv4/tcp_child_ehash_entries is doing[1] - and
>> inter-netns connectivity is explicitly granted by the admin (I guess
>> you will need new transport operations for that)?
>>
>> /P
>>
>> [1] tcp allows using per-netns established socket lookup tables - as
>> opposed to the default global lookup table (even if match always takes
>> in account the netns obviously). The mentioned sysctl specify such
>> configuration for the children namespaces, if any.
> 
> I'll save this discussion if the above doesn't resolve your concerns.
I still have some concern WRT the dynamic mode change after netns
creation. I fear some 'unsolvable' (or very hard to solve) race I can't
see now. A tcp_child_ehash_entries-like model will avoid completely the
issue, but I understand it would be a significant change over the
current status.

"Luckily" the merge window is on us and we have some time to discuss. Do
you have a specific use-case for the ability to change the netns mode
after creation?

/P


Reply via email to