.. SPDX-License-Identifier: BSD-3-Clause ======================= Introduction to Netlink ======================= Netlink is often described as an ioctl() replacement. It aims to replace fixed-format C structures as supplied to ioctl() with a format which allows an easy way to add or extended the arguments. To achieve this Netlink uses a minimal fixed-format metadata header followed by multiple attributes in the TLV (type, length, value) format. Unfortunately the protocol has evolved over the years, in an organic and undocumented fashion, making it hard to coherently explain. To make the most practical sense this document starts by describing netlink as it is used today and dives into more "historical" uses in later sections. Opening a socket ================ Netlink communication happens over sockets, a socket needs to be opened first: .. code-block:: c fd = socket(AF_NETLINK, SOCK_RAW, NETLINK_GENERIC); The use of sockets allows for a natural way of exchanging information in both directions (to and from the kernel). The operations are still performed synchronously when applications send() the request but a separate recv() system call is needed to read the reply. A very simplified flow of a Netlink "call" will therefore look something like: .. code-block:: c fd = socket(AF_NETLINK, SOCK_RAW, NETLINK_GENERIC); /* format the request */ send(fd, &request, sizeof(request)); n = recv(fd, &response, RSP_BUFFER_SIZE); /* interpret the response */ Netlink also provides natural support for "dumping", i.e. communicating to user space all objects of a certain type (e.g. dumping all network interfaces). .. code-block:: c fd = socket(AF_NETLINK, SOCK_RAW, NETLINK_GENERIC); /* format the dump request */ send(fd, &request, sizeof(request)); while (1) { n = recv(fd, &buffer, RSP_BUFFER_SIZE); /* one recv() call can read multiple messages, hence the loop below */ for (nl_msg in buffer) { if (nl_msg.nlmsg_type == NLMSG_DONE) goto dump_finished; /* process the object */ } } dump_finished: The first two arguments of the socket() call require little explanation - it is opening a Netlink socket, with all headers provided by the user (hence NETLINK, RAW). The last argument is the protocol within Netlink. This field used to identify the subsystem with which the socket will communicate. Classic vs Generic Netlink -------------------------- Initial implementation of Netlink depended on a static allocation of IDs to subsystems and provided little supporting infrastructure. Let us refer to those protocols collectively as **Classic Netlink**. The list of them is defined on top of the ``include/uapi/linux/netlink.h`` file, they include among others - general networking (NETLINK_ROUTE), iSCSI (NETLINK_ISCSI), and audit (NETLINK_AUDIT). **Generic Netlink** (introduced in 2005) allows for dynamic registration of subsystems (and subsystem ID allocation), introspection and simplifies implementing the kernel side of the interface. The following section describes how to use Generic Netlink, as the number of subsystems using Generic Netlink outnumbers the older protocols by an order of magnitude. There are also no plans for adding more Classic Netlink protocols to the kernel. Basic information on how communicating with core networking parts of the Linux kernel (or another of the 20 subsystems using Classic Netlink) differs from Generic Netlink is provided later in this document. Generic Netlink =============== In addition to the Netlink fixed metadata header each Netlink protocol defines its own fixed metadata header. (Similarly to how network headers stack - Ethernet > IP > TCP we have Netlink > Generic N. > Family.) A Netlink message always starts with struct nlmsghdr, which is followed by a protocol-specific header. In case of Generic Netlink the protocol header is struct genlmsghdr. The practical meaning of the fields in case of Generic Netlink is as follows: .. code-block:: c struct nlmsghdr { __u32 nlmsg_len; /* Length of message including headers */ __u16 nlmsg_type; /* Generic Netlink Family (subsystem) ID */ __u16 nlmsg_flags; /* Flags - request or dump */ __u32 nlmsg_seq; /* Sequence number */ __u32 nlmsg_pid; /* Port ID, set to 0 */ }; struct genlmsghdr { __u8 cmd; /* Command, as defined by the Family */ __u8 version; /* Irrelevant, set to 1 */ __u16 reserved; /* Reserved, set to 0 */ }; /* TLV attributes follow... */ In Classic Netlink :c:member:`nlmsghdr.nlmsg_type` used to identify which operation within the subsystem the message was referring to (e.g. get information about a netdev). Generic Netlink needs to mux multiple subsystems in a single protocol so it uses this field to identify the subsystem, and :c:member:`genlmsghdr.cmd` identifies the operation instead. (See :ref:`res_fam` for information on how to find the Family ID of the subsystem of interest.) Note that the first 16 values (0 - 15) of this field are reserved for control messages both in Classic Netlink and Generic Netlink. See :ref:`nl_msg_type` for more details. There are 3 usual types of message exchanges on a Netlink socket: - performing a single action (``do``); - dumping information (``dump``); - getting asynchronous notifications (``multicast``). Classic Netlink is very flexible and presumably allows other types of exchanges to happen, but in practice those are the three that get used. Asynchronous notifications are sent by the kernel and received by the user sockets which subscribed to them. ``do`` and ``dump`` requests are initiated by the user. :c:member:`nlmsghdr.nlmsg_flags` should be set as follows: - for ``do``: ``NLM_F_REQUEST | NLM_F_ACK`` - for ``dump``: ``NLM_F_REQUEST | NLM_F_ACK | NLM_F_DUMP`` :c:member:`nlmsghdr.nlmsg_seq` should be a set to a monotonically increasing value. The value gets echoed back in responses and doesn't matter in practice, but setting it to an increasing value for each message sent is considered good hygiene. The purpose of the field is matching responses to requests. Asynchronous notifications will have :c:member:`nlmsghdr.nlmsg_seq` of ``0``. :c:member:`nlmsghdr.nlmsg_pid` is the Netlink equivalent of an address. This field can be set to ``0`` when talking to the kernel. See :ref:`nlmsg_pid` for the (uncommon) uses of the field. The expected use for :c:member:`genlmsghdr.version` was to allow versioning of the APIs provided by the subsystems. No subsystem to date made significant use of this field, so setting it to ``1`` seems like a safe bet. .. _nl_msg_type: Netlink message types --------------------- As previously mentioned :c:member:`nlmsghdr.nlmsg_type` carries protocol specific values but the first 16 identifiers are reserved (first subsystem specific message type should be equal to ``NLMSG_MIN_TYPE`` which is ``0x10``). There are only 4 Netlink control messages defined: - ``NLMSG_NOOP`` - ignore the message, not used in practice; - ``NLMSG_ERROR`` - carries the return code of an operation; - ``NLMSG_DONE`` - marks the end of a dump; - ``NLMSG_OVERRUN`` - socket buffer has overflown, not used to date. ``NLMSG_ERROR`` and ``NLMSG_DONE`` are of practical importance. They carry return codes for operations. Note that unless the ``NLM_F_ACK`` flag is set on the request Netlink will not respond with ``NLMSG_ERROR`` if there is no error. To avoid having to special-case this quirk it is recommended to always set ``NLM_F_ACK``. The format of ``NLMSG_ERROR`` is described by struct nlmsgerr:: ---------------------------------------------- | struct nlmsghdr - response header | ---------------------------------------------- | int error | ---------------------------------------------- | struct nlmsghdr - original request header | ---------------------------------------------- | ** optionally (1) payload of the request | ---------------------------------------------- | ** optionally (2) extended ACK | ---------------------------------------------- There are two instances of struct nlmsghdr here, first of the response and second of the request. ``NLMSG_ERROR`` carries the information about the request which led to the error. This could be useful when trying to match requests to responses or re-parse the request to dump it into logs. The payload of the request is not echoed in messages reporting success (``error == 0``) or if ``NETLINK_CAP_ACK`` setsockopt() was set. The latter is common and perhaps recommended as having to read a copy of every request back from the kernel is rather wasteful. The absence of request payload is indicated by ``NLM_F_CAPPED`` in :c:member:`nlmsghdr.nlmsg_flags`. The second optional element of ``NLMSG_ERROR`` are the extended ACK attributes. See :ref:`ext_ack` for more details. The presence of extended ACK is indicated by ``NLM_F_ACK_TLVS`` in :c:member:`nlmsghdr.nlmsg_flags`. ``NLMSG_DONE`` is simpler, the request is never echoed but the extended ACK attributes may be present:: ---------------------------------------------- | struct nlmsghdr - response header | ---------------------------------------------- | int error | ---------------------------------------------- | ** optionally extended ACK | ---------------------------------------------- .. _res_fam: Resolving the Family ID ----------------------- This section explains how to find the Family ID of a subsystem. It also serves as an example of Generic Netlink communication. Generic Netlink is itself a subsystem exposed via the Generic Netlink API. To avoid a circular dependency Generic Netlink has a statically allocated Family ID (``GENL_ID_CTRL`` which is equal to ``NLMSG_MIN_TYPE``). The Generic Netlink family implements a command used to find out information about other families (``CTRL_CMD_GETFAMILY``). To get information about the Generic Netlink family named for example ``"test1"`` we need to send a message on the previously opened Generic Netlink socket. The message should target the Generic Netlink Family (1), be a ``do`` (2) call to ``CTRL_CMD_GETFAMILY`` (3). A ``dump`` version of this call would make the kernel respond with information about *all* the families it knows about. Last but not least the name of the family in question has to be specified (4) as an attribute with the appropriate type:: struct nlmsghdr: __u32 nlmsg_len: 32 __u16 nlmsg_type: GENL_ID_CTRL // (1) __u16 nlmsg_flags: NLM_F_REQUEST | NLM_F_ACK // (2) __u32 nlmsg_seq: 1 __u32 nlmsg_pid: 0 struct genlmsghdr: __u8 cmd: CTRL_CMD_GETFAMILY // (3) __u8 version: 2 /* or 1, doesn't matter */ __u16 reserved: 0 struct nlattr: // (4) __u16 nla_len: 10 __u16 nla_type: CTRL_ATTR_FAMILY_NAME char data: test1\0 (padding:) char data: \0\0 The length fields in Netlink (:c:member:`nlmsghdr.nlmsg_len` and :c:member:`nlattr.nla_len`) always *include* the header. Attribute headers in netlink must be aligned to 4 bytes from the start of the message, hence the extra ``\0\0`` after ``CTRL_ATTR_FAMILY_NAME``. The attribute lengths *exclude* the padding. If the family is found kernel will reply with two messages, the response with all the information about the family:: /* Message #1 - reply */ struct nlmsghdr: __u32 nlmsg_len: 136 __u16 nlmsg_type: GENL_ID_CTRL __u16 nlmsg_flags: 0 __u32 nlmsg_seq: 1 /* echoed from our request */ __u32 nlmsg_pid: 5831 /* The PID of our user space process */ struct genlmsghdr: __u8 cmd: CTRL_CMD_GETFAMILY __u8 version: 2 __u16 reserved: 0 struct nlattr: __u16 nla_len: 10 __u16 nla_type: CTRL_ATTR_FAMILY_NAME char data: test1\0 (padding:) data: \0\0 struct nlattr: __u16 nla_len: 6 __u16 nla_type: CTRL_ATTR_FAMILY_ID __u16: 123 /* The Family ID we are after */ (padding:) char data: \0\0 struct nlattr: __u16 nla_len: 9 __u16 nla_type: CTRL_ATTR_FAMILY_VERSION __u16: 1 /* ... etc, more attributes will follow. */ And the error code (success) since ``NLM_F_ACK`` had been set on the request:: /* Message #2 - the ACK */ struct nlmsghdr: __u32 nlmsg_len: 36 __u16 nlmsg_type: NLMSG_ERROR __u16 nlmsg_flags: NLM_F_CAPPED /* There won't be a payload */ __u32 nlmsg_seq: 1 /* echoed from our request */ __u32 nlmsg_pid: 5831 /* The PID of our user space process */ int error: 0 struct nlmsghdr: /* Copy of the request header as we sent it */ __u32 nlmsg_len: 32 __u16 nlmsg_type: GENL_ID_CTRL __u16 nlmsg_flags: NLM_F_REQUEST | NLM_F_ACK __u32 nlmsg_seq: 1 __u32 nlmsg_pid: 0 The order of attributes (struct nlattr) is not guaranteed so the user has to walk the attributes and parse them. Note that Generic Netlink sockets are not associated or bound to a single family. A socket can be used to exchange messages with many different families, selecting the recipient family on message-by-message basis using the :c:member:`nlmsghdr.nlmsg_type` field. .. _ext_ack: Extended ACK ------------ Extended ACK controls reporting of additional error/warning TLVs in ``NLMSG_ERROR`` and ``NLMSG_DONE`` messages. To maintain backward compatibility this feature has to be explicitly enabled by setting the ``NETLINK_EXT_ACK`` setsockopt() to ``1``. Types of extended ack attributes are defined in enum nlmsgerr_attrs. The most commonly used attributes are ``NLMSGERR_ATTR_MSG``, ``NLMSGERR_ATTR_OFFS`` and ``NLMSGERR_ATTR_MISS_*``. ``NLMSGERR_ATTR_MSG`` carries a message in English describing the encountered problem. These messages are far more detailed than what can be expressed thru standard UNIX error codes. ``NLMSGERR_ATTR_OFFS`` points to the attribute which caused the problem. ``NLMSGERR_ATTR_MISS_TYPE`` and ``NLMSGERR_ATTR_MISS_NEST`` inform about a missing attribute. Extended ACKs can be reported on errors as well as in case of success. The latter should be treated as a warning. Extended ACKs greatly improve the usability of Netlink and should always be enabled, appropriately parsed and reported to the user. Advanced topics =============== Dump consistency ---------------- Some of the data structures kernel uses for storing objects make it hard to provide an atomic snapshot of all the objects in a dump (without impacting the fast-paths updating them). Kernel may set the ``NLM_F_DUMP_INTR`` flag on any message in a dump (including the ``NLMSG_DONE`` message) if the dump was interrupted and may be inconsistent (e.g. missing objects). User space should retry the dump if it sees the flag set. Introspection ------------- The basic introspection abilities are enabled by access to the Family object as reported in :ref:`res_fam`. User can query information about the Generic Netlink family, including which operations are supported by the kernel and what attributes the kernel understands. Family information includes the highest ID of an attribute kernel can parse, a separate command (``CTRL_CMD_GETPOLICY``) provides detailed information about supported attributes, including ranges of values the kernel accepts. Querying family information is useful in cases when user space needs to make sure that the kernel has support for a feature before issuing a request. .. _nlmsg_pid: nlmsg_pid --------- :c:member:`nlmsghdr.nlmsg_pid` is the Netlink equivalent of an address. It is referred to as Port ID, sometimes Process ID because for historical reasons if the application does not select (bind() to) an explicit Port ID kernel will automatically assign it the ID equal to its Process ID (as reported by the getpid() system call). Similarly to the bind() semantics of the TCP/IP network protocols the value of zero means "assign automatically", hence it is common for applications to leave the :c:member:`nlmsghdr.nlmsg_pid` field initialized to ``0``. The field is still used today in rare cases when kernel needs to send a unicast notification. User space application can use bind() to associate its socket with a specific PID, it then communicates its PID to the kernel. This way the kernel can reach the specific user space process. This sort of communication is utilized in UMH (User Mode Helper)-like scenarios when kernel needs to trigger user space processing or ask user space for a policy decision. Multicast notifications ----------------------- One of the strengths of Netlink is the ability to send event notifications to user space. This is a unidirectional form of communication (kernel -> user) and does not involve any control messages like ``NLMSG_ERROR`` or ``NLMSG_DONE``. For example the Generic Netlink family itself defines a set of multicast notifications about registered families. When a new family is added the sockets subscribed to the notifications will get the following message:: struct nlmsghdr: __u32 nlmsg_len: 136 __u16 nlmsg_type: GENL_ID_CTRL __u16 nlmsg_flags: 0 __u32 nlmsg_seq: 0 __u32 nlmsg_pid: 0 struct genlmsghdr: __u8 cmd: CTRL_CMD_NEWFAMILY __u8 version: 2 __u16 reserved: 0 struct nlattr: __u16 nla_len: 10 __u16 nla_type: CTRL_ATTR_FAMILY_NAME char data: test1\0 (padding:) data: \0\0 struct nlattr: __u16 nla_len: 6 __u16 nla_type: CTRL_ATTR_FAMILY_ID __u16: 123 /* The Family ID we are after */ (padding:) char data: \0\0 struct nlattr: __u16 nla_len: 9 __u16 nla_type: CTRL_ATTR_FAMILY_VERSION __u16: 1 /* ... etc, more attributes will follow. */ The notification contains the same information as the response to the ``CTRL_CMD_GETFAMILY`` request. The Netlink headers of the notification are mostly 0 and irrelevant. The :c:member:`nlmsghdr.nlmsg_seq` may be either zero or a monotonically increasing notification sequence number maintained by the family. To receive notifications the user socket must subscribe to the relevant notification group. Much like the Family ID, the Group ID for a given multicast group is dynamic and can be found inside the Family information. The ``CTRL_ATTR_MCAST_GROUPS`` attribute contains nests with names (``CTRL_ATTR_MCAST_GRP_NAME``) and IDs (``CTRL_ATTR_MCAST_GRP_ID``) of the groups family. Once the Group ID is known a setsockopt() call adds the socket to the group: .. code-block:: c unsigned int group_id; /* .. find the group ID... */ setsockopt(fd, SOL_NETLINK, NETLINK_ADD_MEMBERSHIP, &group_id, sizeof(group_id)); The socket will now receive notifications. It is recommended to use separate sockets for receiving notifications and sending requests to the kernel. The asynchronous nature of notifications means that they may get mixed in with the responses making the message handling much harder. Buffer sizing ------------- Netlink sockets are datagram sockets rather than stream sockets, meaning that each message must be received in its entirety by a single recv()/recvmsg() system call. If the buffer provided by the user is too short, the message will be truncated and the ``MSG_TRUNC`` flag set in struct msghdr (struct msghdr is the second argument of the recvmsg() system call, *not* a Netlink header). Upon truncation the remaining part of the message is discarded. Netlink expects that the user buffer will be at least 8kB or a page size of the CPU architecture, whichever is bigger. Particular Netlink families may, however, require a larger buffer. 32kB buffer is recommended for most efficient handling of dumps (larger buffer fits more dumped objects and therefore fewer recvmsg() calls are needed). Classic Netlink =============== The main differences between Classic and Generic Netlink are the dynamic allocation of subsystem identifiers and availability of introspection. In theory the protocol does not differ significantly, however, in practice Classic Netlink experimented with concepts which were abandoned in Generic Netlink (really, they usually only found use in a small corner of a single subsystem). This section is meant as an explainer of a few of such concepts, with the explicit goal of giving the Generic Netlink users the confidence to ignore them when reading the uAPI headers. Most of the concepts and examples here refer to the ``NETLINK_ROUTE`` family, which covers much of the configuration of the Linux networking stack. Real documentation of that family, deserves a chapter (or a book) of its own. Families -------- Netlink refers to subsystems as families. This is a remnant of using sockets and the concept of protocol families, which are part of message demultiplexing in ``NETLINK_ROUTE``. Sadly every layer of encapsulation likes to refer to whatever it's carrying as "families" making the term very confusing: 1. AF_NETLINK is a bona fide socket protocol family 2. AF_NETLINK's documentation refers to what comes after its own header (struct nlmsghdr) in a message as a "Family Header" 3. Generic Netlink is a family for AF_NETLINK (struct genlmsghdr follows struct nlmsghdr), yet it also calls its users "Families". Note that the Generic Netlink Family IDs are in a different "ID space" and overlap with Classic Netlink protocol numbers (e.g. ``NETLINK_CRYPTO`` has the Classic Netlink protocol ID of 21 which Generic Netlink will happily allocate to one of its families as well). Strict checking --------------- The ``NETLINK_GET_STRICT_CHK`` socket option enables strict input checking in ``NETLINK_ROUTE``. It was needed because historically kernel did not validate the fields of structures it didn't process. This made it impossible to start using those fields later without risking regressions in applications which initialized them incorrectly or not at all. ``NETLINK_GET_STRICT_CHK`` declares that the application is initializing all fields correctly. It also opts into validating that message does not contain trailing data and requests that kernel rejects attributes with type higher than largest attribute type known to the kernel. ``NETLINK_GET_STRICT_CHK`` is not used outside of ``NETLINK_ROUTE``. Unknown attributes ------------------ Historically Netlink ignored all unknown attributes. The thinking was that it would free the application from having to probe what kernel supports. The application could make a request to change the state and check which parts of the request "stuck". This is no longer the case for new Generic Netlink families and those opting in to strict checking. See enum netlink_validation for validation types performed. Fixed metadata and structures ----------------------------- Classic Netlink made liberal use of fixed-format structures within the messages. Messages would commonly have a structure with a considerable number of fields after struct nlmsghdr. It was also common to put structures with multiple members inside attributes, without breaking each member into an attribute of its own. This has caused problems with validation and extensibility and therefore using binary structures is actively discouraged for new attributes. Request types ------------- ``NETLINK_ROUTE`` categorized requests into 4 types ``NEW``, ``DEL``, ``GET``, and ``SET``. Each object can handle all or some of those requests (objects being netdevs, routes, addresses, qdiscs etc.) Request type is defined by the 2 lowest bits of the message type, so commands for new objects would always be allocated with a stride of 4. Each object would also have it's own fixed metadata shared by all request types (e.g. struct ifinfomsg for netdev requests, struct ifaddrmsg for address requests, struct tcmsg for qdisc requests). Even though other protocols and Generic Netlink commands often use the same verbs in their message names (``GET``, ``SET``) the concept of request types did not find wider adoption. Message flags ------------- The earlier section has already covered the basic request flags (``NLM_F_REQUEST``, ``NLM_F_ACK``, ``NLM_F_DUMP``) and the ``NLMSG_ERROR`` / ``NLMSG_DONE`` flags (``NLM_F_CAPPED``, ``NLM_F_ACK_TLVS``). Dump flags were also mentioned (``NLM_F_MULTI``, ``NLM_F_DUMP_INTR``). Those are the main flags of note, with a small exception (of ``ieee802154``) Generic Netlink does not make use of other flags. If the protocol needs to communicate special constraints for a request it should use an attribute, not the flags in struct nlmsghdr. Classic Netlink, however, defined various flags for its ``GET``, ``NEW`` and ``DEL`` requests. Since request types have not been generalized the request type specific flags should not be used either. uAPI reference ============== .. kernel-doc:: include/uapi/linux/netlink.h