swh:1:snp:c3bf2749e3476071fa748f67b0ffa2fdc5fe49d9
Tip revision: 614124bea77e452aa6df7a8714e8bc820b489922 authored by Linus Torvalds on 06 June 2021, 22:47:27 UTC
Linux 5.13-rc5
Linux 5.13-rc5
Tip revision: 614124b
rds.rst
.. SPDX-License-Identifier: GPL-2.0
==
RDS
===
Overview
========
This readme tries to provide some background on the hows and whys of RDS,
and will hopefully help you find your way around the code.
In addition, please see this email about RDS origins:
http://oss.oracle.com/pipermail/rds-devel/2007-November/000228.html
RDS Architecture
================
RDS provides reliable, ordered datagram delivery by using a single
reliable connection between any two nodes in the cluster. This allows
applications to use a single socket to talk to any other process in the
cluster - so in a cluster with N processes you need N sockets, in contrast
to N*N if you use a connection-oriented socket transport like TCP.
RDS is not Infiniband-specific; it was designed to support different
transports. The current implementation used to support RDS over TCP as well
as IB.
The high-level semantics of RDS from the application's point of view are
* Addressing
RDS uses IPv4 addresses and 16bit port numbers to identify
the end point of a connection. All socket operations that involve
passing addresses between kernel and user space generally
use a struct sockaddr_in.
The fact that IPv4 addresses are used does not mean the underlying
transport has to be IP-based. In fact, RDS over IB uses a
reliable IB connection; the IP address is used exclusively to
locate the remote node's GID (by ARPing for the given IP).
The port space is entirely independent of UDP, TCP or any other
protocol.
* Socket interface
RDS sockets work *mostly* as you would expect from a BSD
socket. The next section will cover the details. At any rate,
all I/O is performed through the standard BSD socket API.
Some additions like zerocopy support are implemented through
control messages, while other extensions use the getsockopt/
setsockopt calls.
Sockets must be bound before you can send or receive data.
This is needed because binding also selects a transport and
attaches it to the socket. Once bound, the transport assignment
does not change. RDS will tolerate IPs moving around (eg in
a active-active HA scenario), but only as long as the address
doesn't move to a different transport.
* sysctls
RDS supports a number of sysctls in /proc/sys/net/rds
Socket Interface
================
AF_RDS, PF_RDS, SOL_RDS
AF_RDS and PF_RDS are the domain type to be used with socket(2)
to create RDS sockets. SOL_RDS is the socket-level to be used
with setsockopt(2) and getsockopt(2) for RDS specific socket
options.
fd = socket(PF_RDS, SOCK_SEQPACKET, 0);
This creates a new, unbound RDS socket.
setsockopt(SOL_SOCKET): send and receive buffer size
RDS honors the send and receive buffer size socket options.
You are not allowed to queue more than SO_SNDSIZE bytes to
a socket. A message is queued when sendmsg is called, and
it leaves the queue when the remote system acknowledges
its arrival.
The SO_RCVSIZE option controls the maximum receive queue length.
This is a soft limit rather than a hard limit - RDS will
continue to accept and queue incoming messages, even if that
takes the queue length over the limit. However, it will also
mark the port as "congested" and send a congestion update to
the source node. The source node is supposed to throttle any
processes sending to this congested port.
bind(fd, &sockaddr_in, ...)
This binds the socket to a local IP address and port, and a
transport, if one has not already been selected via the
SO_RDS_TRANSPORT socket option
sendmsg(fd, ...)
Sends a message to the indicated recipient. The kernel will
transparently establish the underlying reliable connection
if it isn't up yet.
An attempt to send a message that exceeds SO_SNDSIZE will
return with -EMSGSIZE
An attempt to send a message that would take the total number
of queued bytes over the SO_SNDSIZE threshold will return
EAGAIN.
An attempt to send a message to a destination that is marked
as "congested" will return ENOBUFS.
recvmsg(fd, ...)
Receives a message that was queued to this socket. The sockets
recv queue accounting is adjusted, and if the queue length
drops below SO_SNDSIZE, the port is marked uncongested, and
a congestion update is sent to all peers.
Applications can ask the RDS kernel module to receive
notifications via control messages (for instance, there is a
notification when a congestion update arrived, or when a RDMA
operation completes). These notifications are received through
the msg.msg_control buffer of struct msghdr. The format of the
messages is described in manpages.
poll(fd)
RDS supports the poll interface to allow the application
to implement async I/O.
POLLIN handling is pretty straightforward. When there's an
incoming message queued to the socket, or a pending notification,
we signal POLLIN.
POLLOUT is a little harder. Since you can essentially send
to any destination, RDS will always signal POLLOUT as long as
there's room on the send queue (ie the number of bytes queued
is less than the sendbuf size).
However, the kernel will refuse to accept messages to
a destination marked congested - in this case you will loop
forever if you rely on poll to tell you what to do.
This isn't a trivial problem, but applications can deal with
this - by using congestion notifications, and by checking for
ENOBUFS errors returned by sendmsg.
setsockopt(SOL_RDS, RDS_CANCEL_SENT_TO, &sockaddr_in)
This allows the application to discard all messages queued to a
specific destination on this particular socket.
This allows the application to cancel outstanding messages if
it detects a timeout. For instance, if it tried to send a message,
and the remote host is unreachable, RDS will keep trying forever.
The application may decide it's not worth it, and cancel the
operation. In this case, it would use RDS_CANCEL_SENT_TO to
nuke any pending messages.
``setsockopt(fd, SOL_RDS, SO_RDS_TRANSPORT, (int *)&transport ..), getsockopt(fd, SOL_RDS, SO_RDS_TRANSPORT, (int *)&transport ..)``
Set or read an integer defining the underlying
encapsulating transport to be used for RDS packets on the
socket. When setting the option, integer argument may be
one of RDS_TRANS_TCP or RDS_TRANS_IB. When retrieving the
value, RDS_TRANS_NONE will be returned on an unbound socket.
This socket option may only be set exactly once on the socket,
prior to binding it via the bind(2) system call. Attempts to
set SO_RDS_TRANSPORT on a socket for which the transport has
been previously attached explicitly (by SO_RDS_TRANSPORT) or
implicitly (via bind(2)) will return an error of EOPNOTSUPP.
An attempt to set SO_RDS_TRANSPORT to RDS_TRANS_NONE will
always return EINVAL.
RDMA for RDS
============
see rds-rdma(7) manpage (available in rds-tools)
Congestion Notifications
========================
see rds(7) manpage
RDS Protocol
============
Message header
The message header is a 'struct rds_header' (see rds.h):
Fields:
h_sequence:
per-packet sequence number
h_ack:
piggybacked acknowledgment of last packet received
h_len:
length of data, not including header
h_sport:
source port
h_dport:
destination port
h_flags:
Can be:
============= ==================================
CONG_BITMAP this is a congestion update bitmap
ACK_REQUIRED receiver must ack this packet
RETRANSMITTED packet has previously been sent
============= ==================================
h_credit:
indicate to other end of connection that
it has more credits available (i.e. there is
more send room)
h_padding[4]:
unused, for future use
h_csum:
header checksum
h_exthdr:
optional data can be passed here. This is currently used for
passing RDMA-related information.
ACK and retransmit handling
One might think that with reliable IB connections you wouldn't need
to ack messages that have been received. The problem is that IB
hardware generates an ack message before it has DMAed the message
into memory. This creates a potential message loss if the HCA is
disabled for any reason between when it sends the ack and before
the message is DMAed and processed. This is only a potential issue
if another HCA is available for fail-over.
Sending an ack immediately would allow the sender to free the sent
message from their send queue quickly, but could cause excessive
traffic to be used for acks. RDS piggybacks acks on sent data
packets. Ack-only packets are reduced by only allowing one to be
in flight at a time, and by the sender only asking for acks when
its send buffers start to fill up. All retransmissions are also
acked.
Flow Control
RDS's IB transport uses a credit-based mechanism to verify that
there is space in the peer's receive buffers for more data. This
eliminates the need for hardware retries on the connection.
Congestion
Messages waiting in the receive queue on the receiving socket
are accounted against the sockets SO_RCVBUF option value. Only
the payload bytes in the message are accounted for. If the
number of bytes queued equals or exceeds rcvbuf then the socket
is congested. All sends attempted to this socket's address
should return block or return -EWOULDBLOCK.
Applications are expected to be reasonably tuned such that this
situation very rarely occurs. An application encountering this
"back-pressure" is considered a bug.
This is implemented by having each node maintain bitmaps which
indicate which ports on bound addresses are congested. As the
bitmap changes it is sent through all the connections which
terminate in the local address of the bitmap which changed.
The bitmaps are allocated as connections are brought up. This
avoids allocation in the interrupt handling path which queues
sages on sockets. The dense bitmaps let transports send the
entire bitmap on any bitmap change reasonably efficiently. This
is much easier to implement than some finer-grained
communication of per-port congestion. The sender does a very
inexpensive bit test to test if the port it's about to send to
is congested or not.
RDS Transport Layer
===================
As mentioned above, RDS is not IB-specific. Its code is divided
into a general RDS layer and a transport layer.
The general layer handles the socket API, congestion handling,
loopback, stats, usermem pinning, and the connection state machine.
The transport layer handles the details of the transport. The IB
transport, for example, handles all the queue pairs, work requests,
CM event handlers, and other Infiniband details.
RDS Kernel Structures
=====================
struct rds_message
aka possibly "rds_outgoing", the generic RDS layer copies data to
be sent and sets header fields as needed, based on the socket API.
This is then queued for the individual connection and sent by the
connection's transport.
struct rds_incoming
a generic struct referring to incoming data that can be handed from
the transport to the general code and queued by the general code
while the socket is awoken. It is then passed back to the transport
code to handle the actual copy-to-user.
struct rds_socket
per-socket information
struct rds_connection
per-connection information
struct rds_transport
pointers to transport-specific functions
struct rds_statistics
non-transport-specific statistics
struct rds_cong_map
wraps the raw congestion bitmap, contains rbnode, waitq, etc.
Connection management
=====================
Connections may be in UP, DOWN, CONNECTING, DISCONNECTING, and
ERROR states.
The first time an attempt is made by an RDS socket to send data to
a node, a connection is allocated and connected. That connection is
then maintained forever -- if there are transport errors, the
connection will be dropped and re-established.
Dropping a connection while packets are queued will cause queued or
partially-sent datagrams to be retransmitted when the connection is
re-established.
The send path
=============
rds_sendmsg()
- struct rds_message built from incoming data
- CMSGs parsed (e.g. RDMA ops)
- transport connection alloced and connected if not already
- rds_message placed on send queue
- send worker awoken
rds_send_worker()
- calls rds_send_xmit() until queue is empty
rds_send_xmit()
- transmits congestion map if one is pending
- may set ACK_REQUIRED
- calls transport to send either non-RDMA or RDMA message
(RDMA ops never retransmitted)
rds_ib_xmit()
- allocs work requests from send ring
- adds any new send credits available to peer (h_credits)
- maps the rds_message's sg list
- piggybacks ack
- populates work requests
- post send to connection's queue pair
The recv path
=============
rds_ib_recv_cq_comp_handler()
- looks at write completions
- unmaps recv buffer from device
- no errors, call rds_ib_process_recv()
- refill recv ring
rds_ib_process_recv()
- validate header checksum
- copy header to rds_ib_incoming struct if start of a new datagram
- add to ibinc's fraglist
- if competed datagram:
- update cong map if datagram was cong update
- call rds_recv_incoming() otherwise
- note if ack is required
rds_recv_incoming()
- drop duplicate packets
- respond to pings
- find the sock associated with this datagram
- add to sock queue
- wake up sock
- do some congestion calculations
rds_recvmsg
- copy data into user iovec
- handle CMSGs
- return to application
Multipath RDS (mprds)
=====================
Mprds is multipathed-RDS, primarily intended for RDS-over-TCP
(though the concept can be extended to other transports). The classical
implementation of RDS-over-TCP is implemented by demultiplexing multiple
PF_RDS sockets between any 2 endpoints (where endpoint == [IP address,
port]) over a single TCP socket between the 2 IP addresses involved. This
has the limitation that it ends up funneling multiple RDS flows over a
single TCP flow, thus it is
(a) upper-bounded to the single-flow bandwidth,
(b) suffers from head-of-line blocking for all the RDS sockets.
Better throughput (for a fixed small packet size, MTU) can be achieved
by having multiple TCP/IP flows per rds/tcp connection, i.e., multipathed
RDS (mprds). Each such TCP/IP flow constitutes a path for the rds/tcp
connection. RDS sockets will be attached to a path based on some hash
(e.g., of local address and RDS port number) and packets for that RDS
socket will be sent over the attached path using TCP to segment/reassemble
RDS datagrams on that path.
Multipathed RDS is implemented by splitting the struct rds_connection into
a common (to all paths) part, and a per-path struct rds_conn_path. All
I/O workqs and reconnect threads are driven from the rds_conn_path.
Transports such as TCP that are multipath capable may then set up a
TCP socket per rds_conn_path, and this is managed by the transport via
the transport privatee cp_transport_data pointer.
Transports announce themselves as multipath capable by setting the
t_mp_capable bit during registration with the rds core module. When the
transport is multipath-capable, rds_sendmsg() hashes outgoing traffic
across multiple paths. The outgoing hash is computed based on the
local address and port that the PF_RDS socket is bound to.
Additionally, even if the transport is MP capable, we may be
peering with some node that does not support mprds, or supports
a different number of paths. As a result, the peering nodes need
to agree on the number of paths to be used for the connection.
This is done by sending out a control packet exchange before the
first data packet. The control packet exchange must have completed
prior to outgoing hash completion in rds_sendmsg() when the transport
is mutlipath capable.
The control packet is an RDS ping packet (i.e., packet to rds dest
port 0) with the ping packet having a rds extension header option of
type RDS_EXTHDR_NPATHS, length 2 bytes, and the value is the
number of paths supported by the sender. The "probe" ping packet will
get sent from some reserved port, RDS_FLAG_PROBE_PORT (in <linux/rds.h>)
The receiver of a ping from RDS_FLAG_PROBE_PORT will thus immediately
be able to compute the min(sender_paths, rcvr_paths). The pong
sent in response to a probe-ping should contain the rcvr's npaths
when the rcvr is mprds-capable.
If the rcvr is not mprds-capable, the exthdr in the ping will be
ignored. In this case the pong will not have any exthdrs, so the sender
of the probe-ping can default to single-path mprds.