[RFC PATCH 0/2] Landlock network PoC implementation

[Konstantin Meskhidze]

Hi, Mickaёl
Thanks again for your opinion about minimal Landlock IPv4 network version.
I have already started refactoring the code.
Here are some additional thoughts about the design.

-----Original Message-----
From: Mickaël Salaün <mic at digikod.net> 
Sent: Friday, December 17, 2021 5:39 PM
To: Konstantin Meskhidze <konstantin.meskhidze at huawei.com>
Cc: yusongping <yusongping at huawei.com>; Artem Kuzin <artem.kuzin at huawei.com>; linux-security-module <linux-security-module at vger.kernel.org>; Network Development <netdev at vger.kernel.org>; netfilter at vger.kernel.org
Subject: Re: [RFC PATCH 0/2] Landlock network PoC implementation

New discussions and RFCs should also include netdev and netfilter mailing lists. For people new to Landlock, the goal is to enable unprivileged processes (and then potentially malicious ones) to limit their own network access (i.e. create a security sandbox for themselves).

Thinking more about network access control for Landlock use case, here are better suggestions:

On 14/12/2021 12:51, Mickaël Salaün wrote:
> 
> On 14/12/2021 04:49, Konstantin Meskhidze wrote:
>> Hi Mickaёl.
>> I've been thinking about your reply:
>>
>>> 4. Kernel objects.
>>> For filesystem restrictions inodes objects are used to tie landlock 
>>> rules.
>>> But for socket operations it's preferred to use task_struct object 
>>> of a process, cause sockets' inodes are created just after
>>> security_socket_create() hook is called, and if its needed to have 
>>> some restriction rule for creating sockets, this rule can't be tied 
>>> to a socket inode cause there is no any has been created at the 
>>> hook's catching moment, see the sock_create_lite() function below:
>>
>> - For the file system, we use inodes to identify hierarchies. We 
>> can't
>> - safely rely on stateless objects (e.g. path strings) because the 
>> file
>> - system changes, and then the rules must change with it.
>>
>> - To identify network objects (from the user point of view), we can 
>> rely
>> - on stateless rule definitions because they may be absolute (i.e. IP
>> - address), e.g. sandbox process creating a new connection or 
>> receveing an
>> - UDP packet. It is not be the case with UNIX socket if they are come 
>> from
>> - a path (i.e. inode) though. In this case we'll have to use the 
>> existing
>> - file system identification mechanism and probably extend the 
>> current FS
>> - access rights.
>> - A sandbox is a set of processes handled as "subjects". Generic inet
>> - rules should not be tied to processes (for now) but on 
>> subnets/protocols.
>>
>> In current Landlock version inodes are the objects to tie rules to.
>> For network you are saying that we can rely on stateless rule 
>> definitions and rules should be tied to subnets/protocols, not to 
>> processes'
>> task_struct objects.
>> Cause Landlock architecture requires all rules to be tied to a 
>> different kernel objects, and when LSM hooks are caught there must be 
>> search procedure completed in a ruleset's red-black tree structure:
>>     kernel_object -> landlock_object <- landlock_rule 
>> <-----landlock_ruleset
>>
>> What kind of kernel objects do you mean by subnets/protocols?
>> Do you suggest using sockets' inodes in this case or using network 
>> rules without to be tied to any kernel object?
> 
> The subnets/protocols is the definition provided when creating a rule 
> (i.e. the object from the user point of view), but the kernel may 
> relies on other internal representations. I guess datagram packets 
> would need to be matched against IP/port everytime they are received 
> by a sandboxed process, but tagging sockets or their underlying inodes 
> for stream connections make sense.
> 
> I don't have experience in the network LSM hooks though, any input is 
> welcome.
> 
>>     socket_inode -> landlock_object <- landlock_rule 
>> <-----landlock_ruleset
>>              OR
>>     landlock_object <- landlock_rule <-----landlock_ruleset
>>
>> -----Original Message-----
>> From: Mickaël Salaün <mic at digikod.net>
>> Sent: Monday, December 13, 2021 4:30 PM
>> To: Konstantin Meskhidze <konstantin.meskhidze at huawei.com>
>> Cc: linux-security-module at vger.kernel.org; yusongping 
>> <yusongping at huawei.com>; Artem Kuzin <artem.kuzin at huawei.com>
>> Subject: Re: [RFC PATCH 0/2] Landlock network PoC implementation
>>
>> Hi Konstantin,
>>
>> On 10/12/2021 08:21, Konstantin Meskhidze wrote:

[...]

>>
>> To sum up, for IPv4 restrictions, we need a new rule type identified 
>> with LANDLOCK_RULE_NET_CIDR4. This will handle a new struct 
>> landlock_net_cidr4_attr {
>>       __u64 allowed_access;
>>       __u32 address; // IPv4
>>       __u8 prefix; // From 0 to 32
>>       __u8 type; // SOCK_DGRAM, SOCK_STREAM
>>       __u16 port;
>> } __attribute__((packed));
>> // https://datatracker.ietf.org/doc/html/rfc4632

>> IP addresses (and subnets) should not be part of a rule, at least for now. Indeed, IP addresses are tied either to the system architecture (e.g. container configuration), the local network or Internet, hence moving targets not >> controlled by application developers. Moreover, from a kernel point of view, it is more complex to check and handle subnets, which are most of the time tied to the Netfilter infrastructure, not suitable for Landlock because >> of its unprivileged nature.

>> On the other side, protocols such as TCP and their associated ports are normalized and are tied to an application semantic (e.g. TCP/443 for HTTPS).

>> There is other advantages to exclude subnets from this type of rules for now (e.g. they could be composed with protocols/ports), but that may come later.

>> I then think that a first MVP to bring network access control support to Landlock should focus only on TCP and related ports (i.e. services). I propose to not use my previous definition of landlock_net_cidr4_attr but to have a >> landlock_net_service_attr instead:

>> struct landlock_net_service_attr {
>>     __u64 allowed_access; // LANDLOCK_NET_*_TCP
>>     __u16 port;
>> } __attribute__((packed));

>> This attribute should handle IPv4 and IPv6 indistinguishably.

[...]

>>
>> Accesses/suffixes should be:
>> - CREATE
>> - ACCEPT
>> - BIND
>> - LISTEN
>> - CONNECT
>> - RECEIVE (RECEIVE_FROM and SEND_TO should not be needed)
>> - SEND
>> - SHUTDOWN
>> - GET_OPTION (GETSOCKOPT)
>> - SET_OPTION (SETSOCKOPT)

>> For now, the only access rights should be LANDLOCK_ACCESS_NET_BIND_TCP and LANDLOCK_ACCESS_NET_CONNECT_TCP (tie to two LSM hooks with struct sockaddr).

>> These attribute and access right changes reduce the scope of the network access control and make it simpler but still really useful. Datagram (e.g. UDP, which could add BIND_UDP and SEND_UDP) sockets will be more 
>>complex to restrict correctly and should then come in another patch series, once TCP is supported.

I think that having access rights like LANDLOCK_ACCESS_NET_CREATE_TCP_SOCKET_DENY/LANDLOCK_ACCESS_NET_CREATE_UDP_SOCKET_DENY might be useful during initialization phase of container/sandbox, cause a user could have the possibility to restrict the creation of some type of sockets at all, and to reduce the attack surface related to security aspect.
So the logic could be the following:
	1. Process restricts creation UDP sockets, allows TCP one.  
		- LANDLOCK_ACCESS_NET_CREATE_*_SOCKET_DENY rules are tied to process task_struct cause there are no sockets inodes created at this moment.
	2. Creates necessary number of sockets.
	3. Restricts sockets' access rights. 
		- LANDLOCK_ACCESS_NET_BIND_* / LANDLOCK_ACCESS_NET_CONNECT_* access rights are tied to sockets inodes individually.