Haiku Network Stack Architecture

The Haiku Network Stack is a modular and layered networking stack, very similar to what you may know as BONE.

The entry point when talking to the stack is through a dedicated device driver that publish itself in /dev/net. The userland library libnetwork.so (which combines libsocket.so, and libbind.so) directly talks to this driver, mostly via ioctl()1.

The driver either creates sockets, or passes on every command to the socket module2. Depending on the address family and type of the sockets, the lower layers will be loaded and connected.

For example, with a TCP/IP socket, the stack could look like this:

Socket

TCP

Protocols defined by the socket (address family, type)

(session, transport, network layers)

IPv4

Datalink

ARP

Datalink Protocols defined by the interface (IP address, device)

(datalink layer)

Ethernet framing

Ethernet device

(physical layer)

Where TCP, and IPv4 are net_protocol modules, and ARP, and the Ethernet framing are net_datalink_protocol modules. All modules are connected in a chain, even though the datalink layer introduces more than one path (one for each interface).

When sending data through a socket, a net_buffer is created in the socket module, and passed on to the lower levels where each protocol processes it, before passing it on to the next protocol in the chain. The last protocol in the chain is always a domain protocol - it will directly forward the buffers to the datalink module. When the buffer reaches the datalink level, an accompanied net_route object will determine for which interface (which determines the datalink protocols in the chain) the buffer is destined. The route has to be specified by the upper protocols before the buffer gets into the datalink level - if a buffer comes in without a valid route, it is discarded.

The protocol modules are loaded and unloaded as needed. The stack itself stays loaded as long as there are interfaces defined - as soon as the last interface is removed, the stack gets unloaded (which is, of course, not yet implemented).

The Structures and Classes

net_domain

Every supported address family gets its own domain. A domain comprises such a family, a net_protocol module that handles this domain, and a list of interfaces and routes. It also gets a name: for example, the IPv4 module registers the “internet” domain (AF_INET).

The domain protocol module is responsible for managing the domain; it has to register it when it’s loaded, and it has to unregister it when it is unloaded by the networking stack.

net_interface

An interface makes an underlying net_device accessible by the stack. When creating a new interface, you have to specify a domain, and a device to be used. The stack will then look through the registered datalink protocols, and builds a chain of them for that interface.

The interface usually gets a network address, and a route that directs buffers to be sent to it. If there is no route to an interface, it will never be used for outgoing data, but may well receive data from other hosts.

An interface can be “up” (when IFF_UP is set in its flags member) in which case it accepts data - when that flag is not set, it will discard all data it gets. The interface also specifies the maximum buffer size that can be sent over this interface (the mtu member, a.k.a. maximum transmission unit).

Interfaces are configured via ioctl()s (SIOCAIFADDR, …). You can use the command line tool “ifconfig” to do this for you.

net_device

A networking device is used to actually send and receive the buffers. It either points to an actual hardware device (in case of ethernet), or to a virtual device (in case of loopback). Every device has a unique name that identifies it. When creating a device, the name also decides which net_device module will be chosen; for example, everything that starts with “loop” will end up in the loopback device, while the ethernet device accepts names that start with “/dev/net/”.

A device can be shared by many interfaces at the same time. The device to be used by an interface is specified at the time an interface is created. It also has an mtu member that determines the upper limit of an interface’s mtu as well.

net_buffer

A buffer holds exactly one packet, and has a source as well as a destination address. The addresses may be changed in every layer the buffer passes through. For example, the datalink protocols usually use sockaddr_dl structures with family AF_DLI, while the upper levels may use sockaddr_in structures with family AF_INET. Every protocol only supports a small number of address types, and it’s the requirement of the upper protocols to prepare the address for use in the lower protocols (and that’s also a reason why it wouldn’t work to arbitrarily stack protocols onto each other).

The net_buffer module can be used to access the data within the buffer, append new data to the buffer, or remove chunks of data from it. Internally, the buffer consists of usually fixed size (2048 byte) buffers that can be shared or connected as needed.

net_socket

The socket is only of interest for the net_protocol modules, as it stores options that may have an effect on the protocol’s performance. It’s the direct counterpart to a socket file descriptor in userland, but it has only little logic bound to it.

When a socket is created, the networking stack creates a chain of net_protocol modules for the socket that will then do the real work. When the socket is closed, the net_protocol chain is freed, and the modules are eventually unloaded (if they are no longer in use).

net_protocol

The protocols are bound to a specific socket, process the outgoing buffers as needed (ie. add or remove headers, compute checksums, …), and pass it on to the next protocol. The last protocol in the chain is always a domain protocol that will forward the calls to the datalink module directly, if needed.

A domain protocol is a net_protocol that registered a domain, ie. IPv4. Other than usual protocols, domain protocols have some special requirements:

  • they need to be able to execute send_data(), and get_domain() without a pointer to its net_protocol object, as those may be called outside of the socket context.

  • as mentioned, they also don’t talk to the next protocol in the chain (as they are always the last one), but to the datalink module directly.

Similar to the need to perform send_data() outside of the socket context, all protocols that can receive data need to handle incoming data without the socket context: incoming data is always handled outside of the socket context, as the actual target socket is unknown during processing.

Only the top-most protocol will be able to forward the packet to the target socket(s). To receive incoming data, a protocol must register itself as receiving protocol with the networking stack. The domain protocol is usually registered automatically by a net_datalink_protocol module that knows about both ends (for example, the ARP module is both IPv4 and ethernet specific, and therefore registers the AF_INET domain to receive ethernet packets of type IP).

net_route

A route determines the target interface of an outgoing packet. A route is always owned by a specific domain, and the route is chosen by comparing the networking address of the outgoing buffer with the mask and address of the route.

A protocol will usually not use the routes directly, but use a net_route_info object (see below), that will make sure that the route is updated automatically whenever the routing table is changed.

net_route_info

A routing helper for protocol usage: it stores the target address as well as the route to be used, and has to be registered with the networking stack via register_route_info().

Then, the stack will automatically update the route as needed, whenever the routing table of the domain changes; it will always matches the address specified there. When the routing is no longer needed, you must unregister the net_route_info again.


1 You can find the definition of the driver interface in headers/private/net/net_stack_interface.h, as well as the driver itself at src/add-ons/kernel/drivers/network