Hosted Inferno provides a subset of the interface described here that gives to the TCP/IP and UDP/IP of the host system's own IP subsystem. See Hosted interfaces below for a summary of the differences.

Protocols

Within each instance, the IP device provides an interface to each IP protocol configured into the system, such as TCP/IP or UDP/IP.

Each of the protocols is served by the IP device, which represents a connection by a set of device files. The top level directory, proto in the SYNOPSIS above, is named after a protocol (eg, tcp, il, udp) and contains a clone file, a stats file, and subdirectories numbered from zero to the number of connections configured for this protocol.

The read-only stats file contains protocol-specific statistics as one or more lines of text. There is no particular format, but the values are often a superset of those required by the SNMP MIB.

Opening the clone file reserves a connection, represented by one of the numbered subdirectories. The resulting file descriptor will be open on the control file, ctl, of the newly allocated connection. Reading the ctl file returns a text string representing the number of the connection. Connections may be used either to listen for incoming calls or to initiate calls to other machines.

A connection is controlled by writing text strings to the associated ctl file. After a connection has been established data may be read from and written to the data file.

Before sending data, remote and local addresses must be set for the connection. For outgoing calls the local port number will be allocated randomly if none is set. Addresses are set by writing control messages to the ctl file of the connection. The connection is not established until the data file is opened. There are two models depending on the nature of the protocol. For connection-oriented protocols, the process will block on open until the remote host has acknowledged the connection, either accepting it, causing a successful return from open, or rejecting it, causing open to return an appropriate error. For connectionless protocols, the open always succeeds; the `connect' request sets local parameters for the source and destination fields for use by subsequent read and write requests.

The following control messages are provided by this interface to all protocols. A particular protocol can provide additional commands, or change the interpretation or even syntax of those below, as described in the manual page for that protocol. The description below shows the standard commands with the default argument syntax and interpretation:

connect ipaddress!port[!r] [lport]

Set the remote IP address and port number for the connection. If the r flag is supplied and the optional local port lport has not been specified the system will allocate a restricted port number (between 600 and 1024) for the connection to allow communication with Unix machines' login and exec services.

announce [ipaddress!]port

Set the local port number to port and accept calls to that port. Port is a decimal port number or *. If port is zero, assign a port number (the one assigned can be read from the local address file). If port is *, accept calls for any port that no process has explicitly announced. If the optional ipaddress is given, set the local IP address for the connection to that address, and accept only those incoming calls to port that are addressed to ipaddress. Announce fails if the connection is already announced or connected.

bind port

Port is a decimal port number or *. Set the local port number to port. This request exists to support emulation of of BSD sockets and is otherwise neither needed nor used in Inferno.

tos [ n ]

Set the type-of-service value in outgooing packets to n (default: 0).

ttl [ n ]

Set the time-to-live (TTL) value in packets transmitted on this conversation to n (default: 255).

Port numbers must be in the range 1 to 32767.

Several read-only files report the status of a connection. The remote and local files contain the IP address and port number for the remote and local side of the connection. The status file contains protocol-dependent information to help debug network connections. The first word on the first line gives the status of the connection.

Having announced, a process may accept incoming connections by calling open on the listen file. The open will block until a new connection request arrives; it will then return an open file descriptor that points to the control file of the newly accepted connection. Repeating this procedure will accept all calls for the given protocol.

In general it should not be necessary to use the file system interface to the networks. The dial, announce, and listen functions described in sys-dial(2) perform the necessary I/O to establish and manipulate network connections.

TCP protocol

The TCP protocol is the standard Internet protocol for reliable stream communication; it does not preserve read/write boundaries.

A connection is controlled by writing text strings to the associated ctl file. After a connection has been established data may be read from and written to the data file. The TCP protocol provides a stream connection that does not preserve read/write boundaries.

For outgoing calls the local port number will be allocated randomly if none is set. Addresses are set by writing control messages to the ctl file of the connection. The connection is not established until the data file is opened. For TCP the process will block until the remote host has acknowledged the connection.

As well as the standard control messages above, TCP accepts the following:

hangup

Send a TCP reset (RST) to the remote side and end the conversation, without waiting for untransmitted data to be acknowledged, unlike a normal close of the device.

keepalive [n]

Enable `keep alive' mode: if no traffic crosses the link within a given period, send a packet to check that the remote party is still there, and remind it that the local connection is still live. The optional value n gives the keep-alive time in milliseconds (default: 120000).

The status file has many lines, each containing a labelled number, giving the values of parameters and statistics such as: maximum allowed connections, outgoing calls, incoming calls, established but later reset, active calls, input segments, output segments, retransmitted segments, retransmitted timeouts, input errors, transmitted reset.

UDP protocol

UDP provides the standard Internet protocol for unreliable datagram communication.

UDP opens always succeed. Before sending data, remote and local addresses must be set for the connection. Alternatively, the following special control requests can be used:

headers

Set the connection to use an address header with IPv6 addressing on reads and writes of the data file, allowing a single connection to send datagrams to converse with many different destination addresses and ports. The 52 byte binary header appears before the data read or written. It contains: remote IP address, local IP address, interface IP address, remote port, and local port. The IP addresses are 16 bytes each in IPv6 format, and the port addresses are 2 bytes each, all written in network (big-endian) order. On reads, the header gives the values from the incoming datagram, except that if the remote used a multicast destination address, the IP address of the receiving interface is substituted. On writes, the header provides the destination for the resulting datagram, and if the local IP address corresponds to a valid local unicast interface, that address is used, otherwise the IP address of the transmitting interface is substituted.

headers4

Set the connection to use an address header with IPv4 addresses on reads and writes of the data file, allowing a single connection to send datagrams to converse with many different destination addresses and ports. The 12 byte binary header appears before the data read or written. It contains: remote IP address, local IP address, remote port, and local port. The IP addresses are 4 bytes each, the port addresses are 2 bytes each, all written in network (big-endian) order. On reads, the header gives the values from the incoming datagram. On writes, the header provides the destination for the resulting datagram. This mode is obsolete and destined for oblivion.

A read of less than the size of the datagram will cause the entire datagram to be consumed. Each write to the data file will send a single datagram on the network.

In replies, in connection-oriented mode, if the remote address has not been set, the first arriving packet sets the following based on the source of the incoming datagram: the remote address and port for the conversation, and the local address is set to the destination address in the datagram unless that is a multicast address, and then the address of the receiving interface is used.

If a conversation is in headers mode, only the local port is relevant.

Connection-oriented UDP is hungup if an ICMP error (eg, host or port unreachable, or time exceeded) arrives with matching port.

The udp status file contains four lines, each containing a labelled number counting an event: input datagrams, datagrams on unannounced ports, datagrams with wrong checksum, and output datagrams.

IL Protocol

IL provides a reliable point-to-point datagram service for communication between Plan 9 and native Inferno machines. Each read and write transfers a single datagram, as for UDP. The datagrams are delivered reliably and in order. Conversations are addressed and established as for TCP.

Routing

The iproute file can be read and written. When read, it returns the contents of the IP routing tables, one line per entry, with six fields giving the destination host or network address, address mask, gateway address, route type, tag (see below), and the number of the ipifc interface owning the route (or `-' if none). The route type is up to four characters: 4 or 6 (IPv4 or IPv6 route); i (route is interface); one of u (unicast), b (broadcast), or m (multicast); and lastly p if the route is point-to-point.

Commands can also be written to control the routing:

add ip mask gw [ tag ]

Add a route via the gateway identified by IP address gw to the address specified by ip and subnet mask mask. Tag the resulting table entry with the tag provided, or the current tag (see tag below), or the tag none.

flush [ tag ]

Remove all routes with the given tag that do not correspond to a local interface. If tag is not given, flush all routes.

remove ip mask

Remove routes to the given address.

tag tag

Tag the routes generated by writes on the current file descriptor with the given tag of up to 4 characters. The default is none, set when iproute is opened.

The ipselftab file summarises the addresses and routes that refer to the local host. It gives an address, the number of logical interfaces, and the interface type in the same form as the route type of iproute.

The iprouter file is provided for use by a user-level application acting as an IP gateway. It is effective only when the kernel-level gateway is not enabled (see the iprouting interface control request below). Once opened, packets that are not addressed to a local address can be read from this device. The packet contents are preceded by a 16 byte binary header that gives the IPv6 address of the local interface that received the packet.

Bootstrap

The read-only bootp file contains the results of the last BOOTP request transmitted on any interface (see Physical and logical interfaces below) as several lines of text, with two fields each. The first field names an entity and the second field gives its value in IPv4 address format. The current entities are:

auip

Authentication server address

fsip

File server address

gwip

Address of an IP gateway out of this (sub)net.

ipaddr

Local IP address

ipmask

Subnet mask for the local IP address

If any value is unknown (no reply to BOOTP, or value unspecified), the value will be zero, represented as 0.0.0.0.

Address resolution

The arp file can be read and written. When read, it returns the contents of the current ARP cache as a sequence of lines, one per map entry, giving type, state, IP address and corresponding MAC address. Several textual commands can be written to it:

add [ medium ] ip mac

Add a mapping from IP address ip to the given mac address (a sequence of bytes in hexadecimal) on the given medium. It must support address resolution (eg, Ethernet). If the medium is not specified, find the one associated with a route to ip (which must be IPv4).

flush

Clear the cache.

Logging

The log file provides protocol tracing and debugging data. While the file is held open, the system saves, in a small circular buffer, error messages logged by selected protocols. When read, it returns data not previously read, blocking until there is data to read. The following commands can be written to determine what is logged:

set proto ...

Enable logging of messages from each source proto, one or more of: ppp, ip, fs, tcp, il, icmp, udp, compress, ilmsg, gre, tcpmsg, udpmsg, ipmsg and esp.

clear proto ...

Disable logging of messages from the given sources.

Physical and logical interfaces

The configuration of the physical and logical IP interfaces in a given instance of #I uses a virtual protocol ipifc within that instance, that adds, controls and removes IP interfaces. It is represented by the protocol directory ipifc. Each connection corresponds to an interface to a physical or virtual medium on which IP packets can be sent and received. It has a set of associated values: minimum and maximum transfer unit, MAC address, and a set of logical IP interfaces. Each logical IP interface has local and remote addresses and an address mask.

Opening the clone file returns a file descriptor open on the ctl file for a new connection. A medium is then attached using a bind request; logical interfaces are associated by connect or add; they are removed by remove; and finally unbind detaches the medium from the connection. For certain types of media, the unbind is automatic when the connection itself is closed. With most media, including Ethernet, the ipifc connection files can be closed after configuration, and later reopened if need be to add or remove logical interfaces, or set other parameters.

The ctl file responds to the following text commands, including interface-specific variants of standard IP device requests:

bind medium [ name [ arg ... ]

Attach device medium to the interface, which must not already be bound to a device. The name and subsequent arguments are interpreted by the driver for the medium. The device name associated with the interface is name, if given, or a generated name otherwise.

connect ip [mask [remote [mtu ]]]

Remove all existing logical interfaces and create a new one as if by add (see below). The connection must be bound to a medium.

add ip [ mask [ remote [ mtu ] ] ]

Add a logical interface with local IP address ip. The default for mask is the mask for ip's address class; for the remote address, ip's network; and for mtu, the largest MTU allowed by the medium. The new interface is registered in the IP routing tables.

bootp

Broadcast a BOOTP packet (using udp). If a valid response is received, set the interface's IP address and mask, and the IP stack's default gateway to the results obtained from BOOTP. The results are also available to applications by reading the bootp file above. Note that this mechanism is now deprecated in favour of dhcpclient(2).

remove ip mask

Remove the logical interface determined by ip and mask.

iprouting [n]

Control the use of IP routing on this ip(3) instance. If n is missing or non-zero, allow use as a gateway, rerouting via one interface packets received on another. By default, or if n is zero, use as a gateway is not allowed: if a packet received is not addressed to any local interface, either pass it to a gateway application if active (see iprouter in ip(3)), and otherwise drop the packet.

mtu n

Set the maximum transmit unit (MTU) on this interface to n bytes, which must be valid for the medium.

addmulti multi

Add the multicast address multi to the interface.

remmulti multi

Remove the multicast address multi from the interface.

unbind

Remove any association between the current medium (device) and the connection: remove all routes using this interface, detach the device, stop packet transport, and remove all logical interfaces. The connection is ready for re-use.

The local file contains one line for each logical interface, of the form:

local->self...

where local is the local address associated with the interface and each self is a broadcast or multicast address that can address that interface, including subnet addresses, if any.

The status file contains many fields: the first two give the device name and the value of the current MTU, followed by 7 fields per line for each logical interface: local address, address mask, remote address, packets in, packets out, input errors, and output errors.

The following sections describe the media drivers available. Each is separately configurable into a kernel.

Ethernet medium

Ethernet devices as described in ether(3) can be bound to an IP interface. The bind request has the form:

bind ether device

The interface opens two conversations on the given Ethernet device, for instance ether0, using an internal version of dial, with the addresses device!0x800 (IPv4) and device!0x806 (ARP). See sys-dial(2) for the interpretation of such addresses. The interface runs until a process does an explicit unbind. Multicast settings made on the interface are propagated to the device.

Point-to-point medium

An asynchronous serial device as described in eia(3) can be bound to an interface as a Point-to-Point protocol (PPP) device. The bind request has the form:

bind ppp serial ip remote mtu framing username secret

All parameters except serial are optional. The character `-' can appear as a placeholder for any parameter. Except for authentication data, an attempt is made to negotiate suitable values for any missing parameter values, including network addresses. The parameters are interpreted as follows:

serial

Name of the device that will run PPP.

ip

Local IP address for the interface.

remote

IP address of the other end of the link.

mtu

Initial MTU value for negotiation (default: 1450)

framing

If framing is zero, do not provide asynch. framing (on by default). Unimplemented.

username

Identification string used in PAP or CHAP authentication.

secret

Secret used in authentication; with CHAP it never crosses the link.

If the name serial contains `!' a connection will be opened using dial (see sys-dial(2)). Otherwise the name will be opened as-is; usually it is the name of a serial device (eg, #t/eia0). In the latter case, a companion ctl file will also be opened if possible, to set serial characteristics for PPP (flow control, 64kbyte queue size, nonblocking writes). An attempt is made to start the PPP link immediately. The write of the bind control message returns with an error if the link cannot be started, or if negotiation fails. The PPP link is automatically unbound if the line hangs up (eg, modem drops carrier), or an unrecoverable error occurs when reading or writing the connection.

The PPP implementation can use either PAP and CHAP authentication, as negotiated, provided an appropriate username and secret is given in the bind request. It does not yet support the Microsoft authentication scheme.

Packet medium

The packet medium allows an application to be source and sink for IP packets. It is bound to an interface by the simple request:

bind pkt

All other interface parameters including its IP address are set using the standard ipifc requests described above. Once that has been done, the application reads the data file of the interface to receive packets addressed to the interface, and it writes to the file to inject packets into the IP network. The interface is automatically unbound when all interface files are closed.

Hosted interfaces

Native Inferno and Plan 9 have related IP implementations. Plan 9 emu therefore simply imports Plan 9's /net, and in the absence of version-specific differences, what is described above still applies.

On all other hosted platforms, the IP device gives applications within emu(1) a portable interface to TCP/IP and UDP/IP, even through it is ultimately using the host system's own TCP/IP and UDP/IP implementations (usually but not always socket based). The interface remains the same: for instance by /net/tcp and /net/udp, but is currently more limited in the set of services and control requests. Both IPv4 and IPv6 address syntax may be used, but the IPv6 form must still map to the IPv4 address space. Only TCP and UDP are generally available, and a limited interface to ARP on some platforms (see below). The set of TCP/UDP control requests is limited to: connect, announce, bind, ttl, tos, ignoreadvice, headers4, oldheaders, headers, hangup and keepalive.

The write-only arp file is implemented only on some Unix systems, and is intended to allow the implementation of the BOOTP protocol using Inferno, on hosted systems. It accepts a single textual control request:

add ip ether

Add a new ARP map entry, or replace an existing one, for IP address ip, associating it with the given ether MAC address. The ip address is expressed in the usual dotted address notation; ether is a 12 digit hexadecimal number.

An error results if the host system does not allow the ARP map to be set, or the current user lacks the privileges to set it.