Manual Page Search Parameters
NETLINK(4) Device Drivers Manual NETLINK(4)

Netlink — Kernel network configuration protocol

#include <netlink/netlink.h>
#include <netlink/netlink_route.h>

int
socket(AF_NETLINK, SOCK_RAW, int family);

Netlink is a user-kernel message-based communication protocol primarily used for network stack configuration. Netlink is easily extendable and supports large dumps and event notifications, all via a single socket. The protocol is fully asynchronous, allowing one to issue and track multiple requests at once. Netlink consists of multiple families, which commonly group the commands belonging to the particular kernel subsystem. Currently, the supported families are:

NETLINK_ROUTE	network configuration,
NETLINK_GENERIC	"container" family

The NETLINK_ROUTE family handles all interfaces, addresses, neighbors, routes, and VNETs configuration. More details can be found in rtnetlink(4). The NETLINK_GENERIC family serves as a “container”, allowing registering other families under the NETLINK_GENERIC umbrella. This approach allows using a single netlink socket to interact with multiple netlink families at once. More details can be found in genetlink(4).

Netlink has its own sockaddr structure:

struct sockaddr_nl {
	uint8_t		nl_len;		/* sizeof(sockaddr_nl) */
	sa_family_t	nl_family;	/* netlink family */
	uint16_t	nl_pad;		/* reserved, set to 0 */
	uint32_t	nl_pid;		/* automatically selected, set to 0 */
	uint32_t	nl_groups;	/* multicast groups mask to bind to */
};

Typically, filling this structure is not required for socket operations. It is presented here for completeness.

The protocol is message-based. Each message starts with the mandatory nlmsghdr header, followed by the family-specific header and the list of type-length-value pairs (TLVs). TLVs can be nested. All headers and TLVS are padded to 4-byte boundaries. Each send(2) or recv(2) system call may contain multiple messages.

struct nlmsghdr {
	uint32_t nlmsg_len;   /* Length of message including header */
	uint16_t nlmsg_type;  /* Message type identifier */
	uint16_t nlmsg_flags; /* Flags (NLM_F_) */
	uint32_t nlmsg_seq;   /* Sequence number */
	uint32_t nlmsg_pid;   /* Sending process port ID */
};

The nlmsg_len field stores the whole message length, in bytes, including the header. This length has to be rounded up to the nearest 4-byte boundary when iterating over messages. The nlmsg_type field represents the command/request type. This value is family-specific. The list of supported commands can be found in the relevant family header file. nlmsg_seq is a user-provided request identifier. An application can track the operation result using the NLMSG_ERROR messages and matching the nlmsg_seq The nlmsg_pid field is the message sender id. This field is optional for userland. The kernel sender id is zero. The nlmsg_flags field contains the message-specific flags. The following generic flags are defined:

NLM_F_REQUEST	Indicates that the message is an actual request to the kernel
NLM_F_ACK	Request an explicit ACK message with an operation result

The following generic flags are defined for the "GET" request types:

NLM_F_ROOT	Return the whole dataset
NLM_F_MATCH	Return all entries matching the criteria
These two flags are typically used together, aliased to NLM_F_DUMP

The following generic flags are defined for the "NEW" request types:

NLM_F_CREATE	Create an object if none exists
NLM_F_EXCL	Don't replace an object if it exists
NLM_F_REPLACE	Replace an existing matching object
NLM_F_APPEND	Append to an existing object

The following generic flags are defined for the replies:

NLM_F_MULTI	Indicates that the message is part of the message group
NLM_F_DUMP_INTR	Indicates that the state dump was not completed
NLM_F_DUMP_FILTERED	Indicates that the dump was filtered per request
NLM_F_CAPPED	Indicates the original message was capped to its header
NLM_F_ACK_TLVS	Indicates that extended ACK TLVs were included

Most messages encode their attributes as type-length-value pairs (TLVs). The base TLV header:

struct nlattr {
	uint16_t nla_len;	/* Total attribute length */
	uint16_t nla_type;	/* Attribute type */
};
The TLV type (nla_type) scope is typically the message type or group within a family. For example, the RTN_MULTICAST type value is only valid for RTM_NEWROUTE , RTM_DELROUTE and RTM_GETROUTE messages. TLVs can be nested; in that case internal TLVs may have their own sub-types. All TLVs are packed with 4-byte padding.

A number of generic control messages are reserved in each family.

NLMSG_ERROR reports the operation result if requested, optionally followed by the metadata TLVs. The value of nlmsg_seq is set to its value in the original messages, while nlmsg_pid is set to the socket pid of the original socket. The operation result is reported via struct nlmsgerr:

struct nlmsgerr {
	int	error;		/* Standard errno */
	struct	nlmsghdr msg;	/* Original message header */
};
If the NETLINK_CAP_ACK socket option is not set, the remainder of the original message will follow. If the NETLINK_EXT_ACK socket option is set, the kernel may add a NLMSGERR_ATTR_MSG string TLV with the textual error description, optionally followed by the NLMSGERR_ATTR_OFFS TLV, indicating the offset from the message start that triggered an error. Some operations may return additional metadata encapsulated in the NLMSGERR_ATTR_COOKIE TLV. The metadata format is specific to the operation. If the operation reply is a multipart message, then no NLMSG_ERROR reply is generated, only a NLMSG_DONE message, closing multipart sequence.

NLMSG_DONE indicates the end of the message group: typically, the end of the dump. It contains a single int field, describing the dump result as a standard errno value.

Netlink supports a number of custom socket options, which can be set with setsockopt(2) with the SOL_NETLINK level:

Subscribes to the notifications for the specific group (int).
Unsubscribes from the notifications for the specific group (int).
Lists the memberships as a bitmask.
Instructs the kernel to send the original message header in the reply without the message body.
Acknowledges ability to receive additional TLVs in the ACK message.

Additionally, netlink overrides the following socket options from the SOL_SOCKET level:

Sets the maximum size of the socket receive buffer. If the caller has PRIV_NET_ROUTE permission, the value can exceed the currently-set kern.ipc.maxsockbuf value.

A set of sysctl(8) variables is available to tweak run-time parameters:

net.netlink.sendspace
Default send buffer for the netlink socket. Note that the socket sendspace has to be at least as long as the longest message that can be transmitted via this socket.
net.netlink.recvspace
Default receive buffer for the netlink socket. Note that the socket recvspace has to be least as long as the longest message that can be received from this socket.
net.netlink.nl_maxsockbuf
Maximum receive buffer for the netlink socket that can be set via SO_RCVBUF socket option.

Netlink implements per-functional-unit debugging, with different severities controllable via the net.netlink.debug branch. These messages are logged in the kernel message buffer and can be seen in dmesg(8) The following severity levels are defined:

Rare events or per-socket errors are reported here. This is the default level, not impacting production performance.
Socket events such as groups memberships, privilege checks, commands and dumps are logged. This level does not incur significant performance overhead.
All socket events, each dumped or modified entities are logged. Turning it on may result in significant performance overhead.

Netlink reports operation results, including errors and error metadata, by sending a NLMSG_ERROR message for each request message. The following errors can be returned:

[]
when the current privileges are insufficient to perform the required operation;
[ENOBUFS] or [ENOMEM]
when the system runs out of memory for an internal data structure;
[]
when the requested command is not supported by the family or the family is not supported;
[]
when some necessary TLVs are missing or invalid, detailed info may be provided in NLMSGERR_ATTR_MSG and NLMSGERR_ATTR_OFFS TLVs;
[]
when trying to delete a non-existent object.

Additionally, a socket operation itself may fail with one of the errors specified in socket(2) , recv(2) or send(2)

genetlink(4), rtnetlink(4)

J. Salim, H. Khosravi, A. Kleen, and A. Kuznetsov, Linux Netlink as an IP Services Protocol, RFC 3549.

The netlink protocol appeared in FreeBSD 13.2.

The netlink was implemented by Alexander Chernikov <melifaro@FreeBSD.org>. It was derived from the Google Summer of Code 2021 project by Ng Peng Nam Sean.

November 30, 2022 dev