Scroll to navigation

RXRPC(7) kafs-client RXRPC(7)


rxrpc - Linux RxRPC (AFS) protocol implementation


      #include <sys/socket.h>
      #include <linux/rxrpc.h>
      socket = socket(AF_RXRPC, SOCK_DGRAM, PF_INET);
      socket = socket(AF_RXRPC, SOCK_DGRAM, PF_INET6);


Linux optionally implements the RxRPC transport protocol, as used by the AFS network filesystem. Both client and server ends are support - even on the same socket - and authentication is supported at both ends.

This can be used through the BSD socket interface, using the sendmsg(2) and recvmsg(2) system calls with control data to multiplex calls over the socket and to provide or retrieve call metadata. request_key(2) is used to find the authentication keys to use in the calling process's keyrings.

The AF_RXRPC driver uses udp(7) sockets underneath, either IPv4-only or IPv6 (with IPv4), for transport. Under certain circumstances, the underlying transport sockets may be shared between client-only sockets (but are never shared if a socket is implementing a server).

Address format

      struct sockaddr_rxrpc {
          sa_family_t     srx_family;     /* AF_RXRPC */
          uint16_t        srx_service;    /* The service identifier */
          uint16_t        transport_type; /* The type of transport */
          uint16_t        transport_len;  /* Transport address length */
          union {
              sa_family_t family;         /* Transport address family */
              struct sockaddr_in sin;     /* IPv4 transport address */
              struct sockaddr_in6 sin6;   /* IPv6 transport address */
          } transport;

Where "srx_family" is always set to "AF_RXRPC" ; "srx_service" is set to the service ID of the desired service; transport_type is set to the transport type, which is always SOCK_DGRAM for now; transport_len says how big the address in transport is.

Inside the transport address part, and appropriate address for the underlying socket should be set, including things like family, port and address as appropriate. Note that it may be permissible to use IPv4 addresses on an IPv6 socket.

Socket Options

AF_RXRPC provides a number of socket options that can be set with setsockopt(2) and read with getsockopt(2). The socket option level for IPv6 is SOL_RXRPC.

The option value is a string that specifies the name of a key to pass to request_key(2) to get an appropriate authentication key. Such keys are expected to be of rxrpc type.

If this isn't set, AF_RXRPC will perform an unauthenticated, unencrypted call to the server.

The option value is a string that specifies the name of a keyring to pass to request_key(2) to specify the keys used by the server end to authenticate connections.

The service keys in the ring should be of type rxrpc_s and their descriptions should be of the form "<service-id>:<security-index>" and each should be given an 8-byte secret.

The option value should be empty. This causes each call made on this socket to get its own virtual connection and thus its own negotiated security context.
The option value should be a 4-byte unsigned integer. This can be one of the following constants: RXRPC_SECURITY_PLAIN, RXRPC_SECURITY_AUTH, or RXRPC_SECURITY_ENCRYPT ; the first indicating the packets should be securely checksummed only, the second that packets should be authenticated and the third that full encryption should be employed.
The option value should be a 2-slot array of 2-byte unsigned integers. To use this, the socket must be a server socket and must have been bound to more than one address with different srx_service specifiers.

Slot[0] in the array specified the service ID to upgrade from; slot[1] specifies the service ID to upgrade to. This allows a client to find out if there's a 'better' version of the service available on the same address, but a different service ID.

If the client follows the correct protocol for probing an upgradeable service, the kernel will automatically upgrade the service ID on the connection and this will be reflected in the address returned by recvmsg(2).

The option buffer should have room for a 4-byte integer. The maximum control buffer message type supported by the kernel is written into the buffer. This allows an application to find out what control messages it may use so that it can avoid getting an error if it tries to use something unsupported.

Message flags

AF_RXRPC communicates certain information by way of the message flags passed to and received from sendmsg(2) and recvmsg(2).

This is passed to sendmsg() to indicate that there is more data to be transmitted as part of the request phase of a client call or the reply phase of a service operation. MSG_EOR recvmsg() sets this to indicate that the call has been terminated (the control messages must be parsed for information as to why) and that the kernel has discarded the user call ID tag. The tag may now be reused. MSG_PEEK This is passed to recvmsg() to look at the front of the message queue without removing any messages or changing the state of any outstanding calls. MSG_WAITALL This is passed to sendmsg() to instruct it not to return for a signal if it is still loading up the message queue and progress is being made at the other side in emptying it. This works around the problem of sendmsg() getting interrupted after partially queuing its data, but not then being able to return how much it has consumed. MSG_DONTWAIT This is passed to recvmsg() to indicate that it shouldn't wait if the message queue is empty.

Control messages

AF_RXRPC communicates metadata to the caller through the ancillary data buffer (msg_control) in the messages passed to and fro using sendmsg(2) and recvmsg(2). When building a control message buffer for sendmsg(), the RXRPC_SUPPORTED_CMSG value should be consulted to make sure that the control message type is supported.

The data for this is an arbitrary long integer/pointer-sized tag that represents the call to the kernel. It may, for example, hold a pointer to some userspace structure representing the call to the process.

[sendmsg] This is passed to sendmsg() when the message proposed will create a client call. It must thereafter be included in all future sendmsg() calls pertaining to that call.

[recvmsg] recvmsg() includes the tag in all messages pertaining to a call until the final termination message is reached - which recvmsg() will mark by setting MSG_EOR.

The data for this is a 32-bit integer that is the abort code.

[sendmsg] When passed to sendmsg(), this causes the operation matching the tag to be aborted; this will be followed up by recvmsg() indicating MSG_EOR and a local error of ECONNABORTED, thereby terminating the tag.

[recvmsg] When obtained from recvmsg(), this indicates that a remote abort was received from the peer and the data gives the code for that abort.

[recvmsg] This conveys no data. It indicates the final acknowledgement to a service call has been received.
[recvmsg] This conveys a 32-bit integer into which the network error that terminated a call will have been placed.
[recvmsg] This conveys no data. It indicates that the operation has been rejected because the server is busy.
[recvmsg] This conveys a 32-bit integer into which the local error that terminated a call will have been placed.
[recvmsg] This conveys no data. It indicates that a new service call has arrived at a server socket and is in need of a tag. RXRPC_ACCEPT is must be used for that.
The data for this is an arbitrary long integer/pointer-sized tag that represents the call to the kernel with the same semantics as for RXRPC_USER_CALL_ID.

[sendmsg] Supply a user call ID tag to a new service call.

[sendmsg] Indicate that this particular call should be made on its own connection with an unshared negotiated security context. This requires no additional data.
[sendmsg] Indicate that this call should attempt to probe the service ID on the other side to see if it gets upgraded. The answer can be found in the srx_service value of the peer address recvmsg() returns for this call. This requires no additional data.
The data for this is a signed 64-bit integer.

[sendmsg] Specify the exact total transmit size. This allows AF_RXRPC to work out in advance how big encrypted packets are going to be (under some circumstances, there's a data length encrypted inside the packet).

If this is set, it may allow AF_RXRPC to be more efficient at filling packets. If the wrong amount of data is given (too little or too much), then the call will be aborted.

The data for this is an array of 1-3 32-bit integers.

[sendmsg] Specify various call timeouts. The first timeout is the hard timeout for the call in seconds: the call will be aborted if it takes longer than this amount of time in total.

The second timeout is the \fIidle\fP timeout for the call in milliseconds: the call will be aborted if we don't receive the next DATA packet within that amount of time during the reception phase.

The third timeout is the \fInormal\fP timeout for the call in milliseconds: the call will be aborted if we go for that amount of time without receiving any type of packet pertaining to the call.


kafs(7), request_key(2)


Copyright (C) 2019 Red Hat, Inc. All Rights Reserved.

Written by David Howells (

This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version.

2023-02-05 0.5-4