ip, esp, gre, icmp, icmpv6, ipmux, rudp, tcp, udp – network protocols over IP

bind –a #Ispec /net

The ip device provides the interface to Internet Protocol stacks. Spec is an integer from 0 to 15 identifying a stack. Each stack implements IPv4 and IPv6. Each stack is independent of all others: the only information transfer between them is via programs that mount multiple stacks. Normally a system uses only one stack. However multiple stacks can be used for debugging new IP networks or implementing firewalls or proxy services.

All addresses used are 16–byte IPv6 addresses. IPv4 addresses are a subset of the IPv6 addresses and both standard ASCII formats are accepted. In binary representation, all v4 addresses start with the 12 bytes, in hex:
00 00 00 00 00 00 00 00 00 00 ff ff

Configuring interfaces
Each stack may have multiple interfaces and each interface may have multiple addresses. The /net/ipifc directory contains a clone file, a stats file, and numbered subdirectories for each physical interface.

Opening the clone file reserves an interface. The file descriptor returned from the open(2) will point to the control file, ctl, of the newly allocated interface. Reading ctl returns a text string representing the number of the interface. Writing ctl alters aspects of the interface. The possible ctl messages are those described under Protocol directories below and these:
bind ether path
Treat the device mounted at path as an Ethernet medium carrying IP and ARP packets and associate it with this interface. The kernel will dial(2) path!0x800, path!0x806 and path!0x86dd and use the connections for IPv4, ARP and IPv6 respectively.
bind pkt       Treat this interface as a packet interface. Assume a user program will read and write the data file to receive and transmit IP packets to the kernel. This is used by programs such as ppp(8) to mediate IP packet transfer between the kernel and a PPP encoded device.
bind netdev path
Treat this interface as a packet interface. The kernel will open path and read and write the resulting file descriptor to receive and transmit IP packets.
bind loopbackTreat this interface as a local loopback. Anything written to it will be looped back.
unbind         Disassociate the physical device from an IP interface.
add local [ mask remote mtu proxy ]
try local [ mask remote mtu proxy ]
Add a local IP address to the interface. Try adds the local address as a tentative address if it's an IPv6 address. The mask, remote, mtu, and proxy arguments are all optional. The default mask is the class mask for the local address. The default remote address is local ANDed with mask. The default mtu (maximum transmission unit) is 1514 for Ethernet and 4096 for packet media. The mtu is the size in bytes of the largest packet that this interface can send. Proxy, if specified, means that this machine should answer ARP requests for the remote address. Ppp(8) does this to make remote machines appear to be connected to the local Ethernet.
remove local mask
Remove a local IP address from an interface.
mtu n          Set the maximum transfer unit for this device to n. The mtu is the maximum size of the packet including any medium–specific headers.
reassemble     Reassemble IP fragments before forwarding to this interface
iprouting n    Allow (n is missing or non–zero) or disallow (n is 0) forwarding packets between this interface and others.
bridge         Enable bridging (see bridge(3)).
promiscuous    Set the interface into promiscuous mode, which makes it accept all incoming packets, whether addressed to it or not.
connect type    marks the Ethernet packet type as being in use, if not already in use on this interface. A type of –1 means `all' but appears to be a no–op.
addmulti Media–addr
Treat the multicast Media–addr on this interface as a local address.
remmulti Media–addr
Remove the multicast address Media–addr from this interface.
scanbs         Make the wireless interface scan for base stations.
headersonly    Set the interface to pass only packet headers, not data too.
add6 v6addr pfx–len [onlink auto validlt preflt]
Add the local IPv6 address v6addr with prefix length pfx–len to this interface. See RFC 2461 §6.2.1 for more detail. The remaining arguments are optional:
onlinkflag: address is `on–link'
auto   flag: autonomous
validltvalid life–time in seconds
prefltpreferred life–time in seconds
ra6 keyword value ...
Set IPv6 router advertisement (RA) parameter keyword's value. Known keywords and the meanings of their values follow. See RFC 2461 §6.2.1 for more detail. Flags are true iff non–zero.
recvra     flag: receive and process RAs.
sendra     flag: generate and send RAs.
mflag      flag: ``Managed address configuration'', goes into RAs.
oflag      flag: ``Other stateful configuration'', goes into RAs.
maxraint   ``maximum time allowed between sending unsolicited multicast'' RAs from the interface, in ms.
minraint   ``minimum time allowed between sending unsolicited multicast'' RAs from the interface, in ms.
linkmtu    ``value to be placed in MTU options sent by the router.'' Zero indicates none.
reachtimesets the Reachable Time field in RAs sent by the router. ``Zero means unspecified (by this router).''
rxmitra    sets the Retrans Timer field in RAs sent by the router. ``Zero means unspecified (by this router).''
ttl        default value of the Cur Hop Limit field in RAs sent by the router. Should be set to the ``current diameter of the Internet.'' ``Zero means unspecified (by this router).''
routerlt   sets the Router Lifetime field of RAs sent from the interface, in ms. Zero means the router is not to be used as a default router.

Reading the interface's status file returns information about the interface, one line for each local address on that interface. The first line has 9 white–space–separated fields: device, mtu, local address, mask, remote or network address, packets in, packets out, input errors, output errors. Each subsequent line contains all but the device and mtu. See readipifc in ip(2).

The file iproute controls information about IP routing. When read, it returns one line per routing entry. Each line contains six white–space–separated fields: target address, target IPv6 subnet mask, address of next hop, flags, tag, and interface number. The entry used for routing an IP packet is the one with the longest mask for which destination address ANDed with target mask equals the target address. The one–character flags are:
4   IPv4 route
6   IPv6 route
i   local interface
b   broadcast address
u   local unicast address
m   multicast route
p   point–to–point route

The tag is an arbitrary, up to 4 character, string. It is normally used to indicate what routing protocol originated the route.

Writing to /net/iproute changes the route table. The messages are:
flush       Remove all routes.
tag string    Associate the tag, string, with all subsequent routes added via this file descriptor.
add target mask nexthop
Add the route to the table. If one already exists with the same target and mask, replace it.
remove target mask
Remove a route with a matching target and mask.
route targetPrint on the console the route to address target, if any. Primarily a debugging aid.

Address resolution
The file /net/arp controls information about address resolution. The kernel automatically updates the v4 ARP and v6 Neighbour Discovery information for Ethernet interfaces. When read, the file returns one line per address containing the type of medium, the status of the entry (OK, WAIT), the IP address, and the medium address. Writing to /net/arp administers the ARP information. The control messages are:
flush      Remove all entries.
add type IP–addr Media–addr
Add an entry or replace an existing one for the same IP address.
del IP–addrDelete an individual entry.

ARP entries do not time out. The ARP table is a cache with an LRU replacement policy. The IP stack listens for all ARP requests and, if the requester is in the table, the entry is updated. Also, whenever a new address is configured onto an Ethernet, an ARP request is sent to help update the table on other systems.

Currently, the only medium type is ether.

Debugging and stack information
If any process is holding /net/log open, the IP stack queues debugging information to it. This is intended primarily for debugging the IP stack. The information provided is implementation–defined; see the source for details. Generally, what is returned is error messages about bad packets.

Writing to /net/log controls debugging. The control messages are:
set arglist    Arglist is a space–separated list of items for which to enable debugging. The possible items are: ppp, ip, fs, tcp, icmp, udp, compress, gre, tcpwin, tcprxmt, udpmsg, ipmsg, and esp.
clear arglistArglist is a space–separated list of items for which to disable debugging.
only addr    If addr is non–zero, restrict debugging to only those packets whose source or destination is that address.

The file /net/ndb can be read or written by programs. It is normally used by ipconfig(8) to leave configuration information for other programs such as dns and cs (see ndb(8)). /net/ndb may contain up to 1024 bytes.

The file /net/ipselftab is a read–only file containing all the IP addresses considered local. Each line in the file contains three white–space–separated fields: IP address, usage count, and flags. The usage count is the number of interfaces to which the address applies. The flags are the same as for routing entries. Note that the `IPv4 route' flag will never be set.

Protocol directories
The ip device supports IP as well as several protocols that run over it: TCP, UDP, RUDP, ICMP, GRE, and ESP. TCP and UDP provide the standard Internet protocols for reliable stream and unreliable datagram communication. RUDP is a locally–developed reliable datagram protocol based on UDP. ICMP is IP's catch–all control protocol used to send low level error messages and to implement ping(8). GRE is a general encapsulation protocol. ESP is the encapsulation protocol for IPsec. IL provided a reliable datagram service for communication between Plan 9 machines over IPv4, but is no longer part of the system.

Each protocol is a subdirectory of the IP stack. The top level directory of each protocol contains a clone file, a stats file, and subdirectories numbered from zero to the number of connections opened for this protocol.

Opening the clone file reserves a connection. The file descriptor returned from the open(2) will point to the control file, ctl, of the newly allocated connection. Reading ctl 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 data. A connection can be actively established using the connect message (see also dial(2)). A connection can be established passively by first using an announce message (see dial(2)) to bind to a local port and then opening the listen file (see dial(2)) to receive incoming calls.

The following control messages are supported:
connect ip–address!port!r local
Establish a connection to the remote ip–address and port. If local is specified, it is used as the local port number. If local is not specified but !r is, the system will allocate a restricted port number (less than 1024) for the connection to allow communication with Unix login and exec services. Otherwise a free port number starting at 5000 is chosen. The connect fails if the combination of local and remote address/port pairs are already assigned to another port.
announce X   X is a decimal port number or *. Set the local port number to X and accept calls to X. If X is *, accept calls for any port that no process has explicitly announced. The local IP address cannot be set. Announce fails if the connection is already announced or connected.
bind X      X is a decimal port number or *. Set the local port number to X. This exists to support emulation of BSD sockets by the APE libraries (see pcc(1)) and is not otherwise used.
ttl n       Set the time to live IP field in outgoing packets to n.
tos n       Set the service type IP field in outgoing packets to n.
Don't break (UDP) connections because of ICMP errors.
addmulti ifc–ip [ mcast–ip ]
Treat ifc–ip on this multicast interface as a local address. If mcast–ip is present, use it as the interface's multicast address.
remmulti ipRemove the address ip from this multicast interface.

Port numbers must be in the range 1 to 32767.

Several 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. On receiving and error or EOF reading or writing the data file, the err file contains the reason for error.

A process may accept incoming connections by open(2)ing the listen file. The open will block until a new connection request arrives. Then open will return an open file descriptor which points to the control file of the newly accepted connection. This procedure will accept all calls for the given protocol. See dial(2).

TCP connections are reliable point–to–point byte streams; there are no message delimiters. A connection is determined by the address and port numbers of the two ends. TCP ctl files support the following additional messages:
hangup      close down this TCP connection
keepalive nturn on keep alive messages. N, if given, is the milliseconds between keepalives (default 30000).
checksum n   emit TCP checksums of zero if n is zero; otherwise, and by default, TCP checksums are computed and sent normally.
tcpporthogdefense onoff
of on enables the TCP port–hog defense for all TCP connections; onoff of off disables it. The defense is a solution to hijacked systems staking out ports as a form of denial–of–service attack. To avoid stateless TCP conversation hogs, ip picks a TCP sequence number at random for keepalives. If that number gets acked by the other end, ip shuts down the connection. Some firewalls, notably ones that perform stateful inspection, discard such out–of–specification keepalives, so connections through such firewalls will be killed after five minutes by the lack of keepalives.

UDP connections carry unreliable and unordered datagrams. A read from data will return the next datagram, discarding anything that doesn't fit in the read buffer. A write is sent as a single datagram.

By default, a UDP connection is a point–to–point link. Either a connect establishes a local and remote address/port pair or after an announce, each datagram coming from a different remote address/port pair establishes a new incoming connection. However, many–to–one semantics is also possible.

If, after an announce, the message headers is written to ctl, then all messages sent to the announced port are received on the announced connection prefixed with the corresponding structure, declared in <ip.h>:
typedef struct Udphdr Udphdr;
struct Udphdr
uchar       raddr[16];       /* V6 remote address and port */
uchar       laddr[16];       /* V6 local address and port */
uchar       ifcaddr[16];     /* V6 interface address (receive only) */
uchar       rport[2]; /* remote port */
uchar       lport[2]; /* local port */

Before a write, a user must prefix a similar structure to each message. The system overrides the user specified local port with the announced one. If the user specifies an address that isn't a unicast address in /net/ipselftab, that too is overridden. Since the prefixed structure is the same in read and write, it is relatively easy to write a server that responds to client requests by just copying new data into the message body and then writing back the same buffer that was read.

In this case (writing headers to the ctl file), no listen nor accept is needed; otherwise, the usual sequence of announce, listen, accept must be executed before performing I/O on the corresponding data file.

RUDP is a reliable datagram protocol based on UDP, currently only for IPv4. Packets are delivered in order. RUDP does not support listen. One must write either connect or announce followed immediately by headers to ctl.

Unlike TCP, the reboot of one end of a connection does not force a closing of the connection. Communications will resume when the rebooted machine resumes talking. Any unacknowledged packets queued before the reboot will be lost. A reboot can be detected by reading the err file. It will contain the message
hangup address!port

where address and port are of the far side of the connection. Retransmitting a datagram more than 10 times is treated like a reboot: all queued messages are dropped, an error is queued to the err file, and the conversation resumes.

RUDP ctl files accept the following messages:
headers          Corresponds to the headers format of UDP.
hangup IP port      Drop the connection to address IP and port.
randdrop [ percent ]Randomly drop percent of outgoing packets. Default is 10%.

ICMP is a datagram protocol for IPv4 used to exchange control requests and their responses with other machines' IP implementations. ICMP is primarily a kernel–to–kernel protocol, but it is possible to generate `echo request' and read `echo reply' packets from user programs.

ICMPv6 is the IPv6 equivalent of ICMP. If, after an announce, the message headers is written to ctl, then before a write, a user must prefix each message with a corresponding structure, declared in <ip.h>:
*    user level icmpv6 with control message "headers"
typedef struct Icmp6hdr Icmp6hdr;
struct Icmp6hdr {
uchar       unused[8];
uchar       laddr[IPaddrlen];     /* local address */
uchar       raddr[IPaddrlen];     /* remote address */

In this case (writing headers to the ctl file), no listen nor accept is needed; otherwise, the usual sequence of announce, listen, accept must be executed before performing I/O on the corresponding data file.

GRE is the encapsulation protocol used by PPTP. The kernel implements just enough of the protocol to multiplex it. Our implementation encapsulates in IPv4, per RFC 1702. Announce is not allowed in GRE, only connect. Since GRE has no port numbers, the port number in the connect is actually the 16 bit eproto field in the GRE header.

Reads and writes transfer a GRE datagram starting at the GRE header. On write, the kernel fills in the eproto field with the port number specified in the connect message.

ESP is the Encapsulating Security Payload (RFC 1827, obsoleted by RFC 4303) for IPsec (RFC 4301). We currently implement only tunnel mode, not transport mode. It is used to set up an encrypted tunnel between machines. Like GRE, ESP has no port numbers. Instead, the port number in the connect message is the SPI (Security Association Identifier (sic)). IP packets are written to and read from data. The kernel encrypts any packets written to data, appends a MAC, and prefixes an ESP header before sending to the other end of the tunnel. Received packets are checked against their MAC's, decrypted, and queued for reading from data. In the following, secret is the hexadecimal encoding of a key, without a leading 0x. The control messages are:
esp alg secretEncrypt with the algorithm, alg, using secret as the key. Possible algorithms are: null, des_56_cbc, des3_cbc, and eventually aes_128_cbc, and aes_ctr.
ah alg secret   Use the hash algorithm, alg, with secret as the key for generating the MAC. Possible algorithms are: null, hmac_sha1_96, hmac_md5_96, and eventually aes_xcbc_mac_96.
header      Turn on header mode. Every buffer read from data starts with 4 unused bytes, and the first 4 bytes of every buffer written to data are ignored.
noheader     Turn off header mode.

IP packet filter
The directory /net/ipmux looks like another protocol directory. It is a packet filter built on top of IP. Each numbered subdirectory represents a different filter. The connect messages written to the ctl file describe the filter. Packets matching the filter can be read on the data file. Packets written to the data file are routed to an interface and transmitted.

A filter is a semicolon–separated list of relations. Each relation describes a portion of a packet to match. The possible relations are:
proto=n        the IP protocol number must be n.
data[n:m]=exprbytes n through m following the IP packet must match expr.
iph[n:m]=expr   bytes n through m of the IP packet header must match expr.
ifc=expr        the packet must have been received on an interface whose address matches expr.
src=expr        The source address in the packet must match expr.
dst=expr        The destination address in the packet must match expr.

Expr is of the form:

If a mask is given, the relevant field is first ANDed with the mask. The result is compared against the value or list of values for a match. In the case of ifc, dst, and src the value is a dot–formatted IP address and the mask is a dot–formatted IP mask. In the case of data, iph and proto, both value and mask are strings of 2 hexadecimal digits representing 8–bit values.

A packet is delivered to only one filter. The filters are merged into a single comparison tree. If two filters match the same packet, the following rules apply in order (here '>' means is preferred to):
1)    protocol > data > source > destination > interface
2)    lower data offsets > higher data offsets
3)    longer matches > shorter matches
4)    older > younger

So far this has just been used to implement a version of OSPF in Inferno and 6to4 tunnelling.

The stats files are read only and contain statistics useful to network monitoring.

Reading /net/ipifc/stats returns a list of 19 tagged and newline–separated fields representing:
forwarding status (0 and 2 mean forwarding off,
1 means on)
default TTL
input packets
input header errors
input address errors
packets forwarded
input packets for unknown protocols
input packets discarded
input packets delivered to higher level protocols
output packets
output packets discarded
output packets with no route
timed out fragments in reassembly queue
requested reassemblies
successful reassemblies
failed reassemblies
successful fragmentations
unsuccessful fragmentations
fragments created

Reading /net/icmp/stats returns a list of 26 tagged and newline–separated fields representing:
messages received
bad received messages
unreachables received
time exceededs received
input parameter problems received
source quenches received
redirects received
echo requests received
echo replies received
timestamps received
timestamp replies received
address mask requests received
address mask replies received
messages sent
transmission errors
unreachables sent
time exceededs sent
input parameter problems sent
source quenches sent
redirects sent
echo requests sent
echo replies sent
timestamps sent
timestamp replies sent
address mask requests sent
address mask replies sent

Reading /net/tcp/stats returns a list of 11 tagged and newline–separated fields representing:
maximum number of connections
total outgoing calls
total incoming calls
number of established connections to be reset
number of currently established connections
segments received
segments sent
segments retransmitted
retransmit timeouts
bad received segments
transmission failures

Reading /net/udp/stats returns a list of 4 tagged and newline–separated fields representing:
datagrams received
datagrams received for bad ports
malformed datagrams received
datagrams sent

Reading /net/gre/stats returns a list of 1 tagged number representing:
header length errors

dial(2), ip(2), bridge(3), ndb(6), listen(8)
/lib/rfc/rfc2460   IPv6
/lib/rfc/rfc4291   IPv6 address architecture
/lib/rfc/rfc4443   ICMPv6


Ipmux has not been heavily used and should be considered experimental. It may disappear in favor of a more traditional packet filter in the future.
Copyright © 2017 Alcatel-Lucent. All rights reserved.