RDS - Reliable Datagram Sockets
This is an implementation of the RDS socket API. It provides reliable,
in-order datagram delivery between sockets over a variety of
Currently, RDS can be transported over Infiniband, and loopback. RDS
over TCP is disabled, but will be re-enabled in the near future.
RDS uses standard AF_INET addresses as described in ip(7) to identify
RDS is still in development and as such does not have a reserved
protocol family constant. Applications must read the string
representation of the protocol family value from the pf_rds sysctl
parameter file described below.
rds_socket = socket(pf_rds, SOCK_SEQPACKET, 0);
RDS sockets support a number of socket options through the
setsockopt(2) and getsockopt(2) calls. The following generic options
(with socket level SOL_SOCKET) are of specific importance:
Specifies the size of the receive buffer. See section on
"Congestion Control" below.
Specifies the size of the send buffer. See "Message
Specifies the send timeout when trying to enqueue a message on a
socket with a full queue in blocking mode.
In addition to these, RDS supports a number of protocol specific
options (with socket level SOL_RDS). Just as with the RDS protocol
family, an official value has not been assigned yet, so the kernel will
assign a value dynamically. The assigned value can be retrieved from
the sol_rds sysctl parameter file.
RDS specific socket options will be described in a separate section
A new RDS socket has no local address when it is first returned from
socket(2). It must be bound to a local address by calling bind(2)
before any messages can be sent or received. This will also attach the
socket to a specific transport, based on the type of interface the
local address is attached to. From that point on, the socket can only
reach destinations which are available through this transport.
For instance, when binding to the address of an Infiniband interface
such as ib0, the socket will use the Infiniband transport. If RDS is
not able to associate a transport with the given address, it will
An RDS socket can only be bound to one address and only one socket can
be bound to a given address/port pair. If no port is specified in the
binding address then an unbound port is selected at random.
RDS does not allow the application to bind a previously bound socket to
another address. Binding to the wildcard address INADDR_ANY is not
The default mode of operation for RDS is to use unconnected socket, and
specify a destination address as an argument to sendmsg. However, RDS
allows sockets to be connected to a remote end point using connect(2).
If a socket is connected, calling sendmsg without specifying a
destination address will use the previously given remote address.
RDS does not have explicit congestion control like common streaming
protocols such as TCP. However, sockets have two queue limits
associated with them; the send queue size and the receive queue size.
Messages are accounted based on the number of bytes of payload.
The send queue size limits how much data local processes can queue on a
local socket (see the following section). If that limit is exceeded,
the kernel will not accept further messages until the queue is drained
and messages have been delivered to and acknowledged by the remote
The receive queue size limits how much data RDS will put on the receive
queue of a socket before marking the socket as congested. When a
socket becomes congested, RDS will send a congestion map update to the
other participating hosts, who are then expected to stop sending more
messages to this port.
There is a timing window during which a remote host can still continue
to send messages to a congested port; RDS solves this by accepting
these messages even if the socket’s receive queue is already over the
As the application pulls incoming messages off the receive queue using
recvmsg(2), the number of bytes on the receive queue will eventually
drop below the receive queue size, at which point the port is then
marked uncongested, and another congestion update is sent to all
participating hosts. This tells them to allow applications to send
additional messages to this port.
The default values for the send and receive buffer size are controlled
by the A given RDS socket has limited transmit buffer space. It
defaults to the system wide socket send buffer size set in the
wmem_default and rmem_default sysctls, respectively. They can be tuned
by the application through the SO_SNDBUF and SO_RCVBUF socket options.
The sendmsg(2) and recvmsg(2) calls can block in a variety of
situations. Whether a call blocks or returns with an error depends on
the non-blocking setting of the file descriptor and the MSG_DONTWAIT
message flag. If the file descriptor is set to blocking mode (which is
the default), and the MSG_DONTWAIT flag is not given, the call will
In addition, the SO_SNDTIMEO and SO_RCVTIMEO socket options can be used
to specify a timeout (in seconds) after which the call will abort
waiting, and return an error. The default timeout is 0, which tells RDS
to block indefinitely.
Messages may be sent using sendmsg(2) once the RDS socket is bound.
Message length cannot exceed 4 gigabytes as the wire protocol uses an
unsigned 32 bit integer to express the message length.
RDS does not support out of band data. Applications are allowed to send
to unicast addresses only; broadcast or multicast are not supported.
A successful sendmsg(2) call puts the message in the socket’s transmit
queue where it will remain until either the destination acknowledges
that the message is no longer in the network or the application removes
the message from the send queue.
Messages can be removed from the send queue with the RDS_CANCEL_SENT_TO
socket option described below.
While a message is in the transmit queue its payload bytes are
accounted for. If an attempt is made to send a message while there is
not sufficient room on the transmit queue, the call will either block
or return EAGAIN.
Trying to send to a destination that is marked congested (see above),
the call will either block or return ENOBUFS.
A message sent with no payload bytes will not consume any space in the
destination’s send buffer but will result in a message receipt on the
destination. The receiver will not get any payload data but will be
able to see the sender’s address.
Messages sent to a port to which no socket is bound will be silently
discarded by the destination host. No error messages are reported to
Messages may be received with recvmsg(2) on an RDS socket once it is
bound to a source address. RDS will return messages in-order, i.e.
messages from the same sender will arrive in the same order in which
they were be sent.
The address of the sender will be returned in the sockaddr_in structure
pointed to by the msg_name field, if set.
If the MSG_PEEK flag is given, the first message on the receive is
returned without removing it from the queue.
The memory consumed by messages waiting for delivery does not limit the
number of messages that can be queued for receive. RDS does attempt to
perform congestion control as described in the section above.
If the length of the message exceeds the size of the buffer provided to
recvmsg(2), then the remainder of the bytes in the message are
discarded and the MSG_TRUNC flag is set in the msg_flags field. In this
truncating case recvmsg(2) will still return the number of bytes
copied, not the length of entire messge. If MSG_TRUNC is set in the
flags argument to recvmsg(2), then it will return the number of bytes
in the entire message. Thus one can examine the size of the next
message in the receive queue without incurring a copying overhead by
providing a zero length buffer and setting MSG_PEEK and MSG_TRUNC in
the flags argument.
The sending address of a zero-length message will still be provided in
the msg_name field.
RDS uses control messages (a.k.a. ancillary data) through the
msg_control and msg_controllen fields in sendmsg(2) and recvmsg(2).
Control messages generated by RDS have a cmsg_level value of sol_rds.
Most control messages are related to the zerocopy interface added in
RDS version 3, and are described in rds-rdma(7).
The only exception is the RDS_CMSG_CONG_UPDATE message, which is
described in the following section.
RDS supports the poll(2) interface in a limited fashion. POLLIN is
returned when there is a message (either a proper RDS message, or a
control message) waiting in the socket’s receive queue. POLLOUT is
always returned while there is room on the socket’s send queue.
Sending to congested ports requires special handling. When an
application tries to send to a congested destination, the system call
will return ENOBUFS. However, it cannot poll for POLLOUT, as there is
probably still room on the transmit queue, so the call to poll(2) would
return immediately, even though the destination is still congested.
There are two ways of dealing with this situation. The first is to
simply poll for POLLIN. By default, a process sleeping in poll(2) is
always woken up when the congestion map is updated, and thus the
application can retry any previously congested sends.
The second option is explicit congestion monitoring, which gives the
application more fine-grained control.
With explicit monitoring, the application polls for POLLIN as before,
and additionally uses the RDS_CONG_MONITOR socket option to install a
64bit mask value in the socket, where each bit corresponds to a group
of ports. When a congestion update arrives, RDS checks the set of ports
that became uncongested against the bit mask installed in the socket.
If they overlap, a control messages is enqueued on the socket, and the
application is woken up. When it calls recvmsg(2), it will be given the
control message containing the bitmap. on the socket.
The congestion monitor bitmask can be set and queried using
setsockopt(2) with RDS_CONG_MONITOR, and a pointer to the 64bit mask
Congestion updates are delivered to the application via
RDS_CMSG_CONG_UPDATE control messages. These control messages are
always delivered by themselves (or possibly additional control
messages), but never along with a RDS data message. The cmsg_data field
of the control message is an 8 byte datum containing the 64bit mask
Applications can use the following macros to test for and set bits in
#define RDS_CONG_MONITOR_SIZE 64
#define RDS_CONG_MONITOR_BIT(port) (((unsigned int) port) % RDS_CONG_MONITOR_SIZE)
#define RDS_CONG_MONITOR_MASK(port) (1 << RDS_CONG_MONITOR_BIT(port))
An application can cancel (flush) messages from the send queue using
the RDS_CANCEL_SENT_TO socket option with setsockopt(2). This call
takes an optional sockaddr_in address structure as argument. If given,
only messages to the destination specified by this address are
discarded. If no address is given, all pending messages are discarded.
Note that this affects messages that have not yet been transmitted as
well as messages that have been transmitted, but for which no
acknowledgment from the remote host has been received yet.
If sendmsg(2) succeeds, RDS guarantees that the message will be
visible to recvmsg(2) on a socket bound to the destination address as
long as that destination socket remains open.
If there is no socket bound on the destination, the message is
silently dropped. If the sending RDS can’t be sure that there is no
socket bound then it will try to send the message indefinitely until it
can be sure or the sent message is canceled.
If a socket is closed then all pending sent messages on the socket are
canceled and may or may not be seen by the receiver.
The RDS_CANCEL_SENT_TO socket option can be used to cancel all pending
messages to a given destination.
If a receiving socket is closed with pending messages then the sender
considers those messages as having left the network and will not
A message will only be seen by recvmsg(2) once, unless MSG_PEEK was
specified. Once the message has been delivered it is removed from the
sending socket’s transmit queue.
All messages sent from the same socket to the same destination will be
delivered in the order they’re sent. Messages sent from different
sockets, or to different destinations, may be delivered in any order.
These parameteres may only be accessed through their files in
/proc/sys/net/rds. Access through sysctl(2) is not supported.
pf_rds This file contains the string representation of the protocol
family constant passed to socket(2) to create a new RDS socket.
This file contains the string representation of the socket level
parameter that is passed to getsockopt(2) and setsockopt(2) to
manipulate RDS socket options.
max_unacked_bytes and max_unacked_packets
These parameters are used to tune the generation of
acknowledgements. By default, the system receiving RDS messages
does not send back explicit acknowledgements unless it transmits
a message of its own (in which case the ACK is piggybacked onto
the outgoing message), or when the sending system requests an
However, the sender needs to see an ACK from time to time so
that it can purge old messages from the send queue. The unacked
bytes and packet counters are used to keep track of how much
data has been sent without requesting an ACK. The default is to
request an acknowledgement every 16 packets, or every 16 MB,
whichever comes first.
reconnect_delay_min_ms and reconnect_delay_max_ms
RDS uses host-to-host connections to transport RDS messages
(both for the TCP and the Infiniband transport). If this
connection breaks, RDS will try to re-establish the connection.
Because this reconnect may be triggered by both hosts at the
same time and fail, RDS uses a random backoff before attempting
a reconnect. These two parameters specify the minimum and
maximum delay in milliseconds. The default values are 1 and
rds-rdma(7), socket(2), bind(2), sendmsg(2), recvmsg(2), getsockopt(2),