| blob | 0235ae69af2a862938e550e61453b4fe81d7b2d3 |
2 Overview
3 ========
5 This readme tries to provide some background on the hows and whys of RDS,
6 and will hopefully help you find your way around the code.
8 In addition, please see this email about RDS origins:
9 http://oss.oracle.com/pipermail/rds-devel/2007-November/000228.html
11 RDS Architecture
12 ================
14 RDS provides reliable, ordered datagram delivery by using a single
15 reliable connection between any two nodes in the cluster. This allows
16 applications to use a single socket to talk to any other process in the
17 cluster - so in a cluster with N processes you need N sockets, in contrast
18 to N*N if you use a connection-oriented socket transport like TCP.
20 RDS is not Infiniband-specific; it was designed to support different
21 transports. The current implementation used to support RDS over TCP as well
22 as IB.
24 The high-level semantics of RDS from the application's point of view are
26 * Addressing
27 RDS uses IPv4 addresses and 16bit port numbers to identify
28 the end point of a connection. All socket operations that involve
29 passing addresses between kernel and user space generally
30 use a struct sockaddr_in.
32 The fact that IPv4 addresses are used does not mean the underlying
33 transport has to be IP-based. In fact, RDS over IB uses a
34 reliable IB connection; the IP address is used exclusively to
35 locate the remote node's GID (by ARPing for the given IP).
37 The port space is entirely independent of UDP, TCP or any other
38 protocol.
40 * Socket interface
41 RDS sockets work *mostly* as you would expect from a BSD
42 socket. The next section will cover the details. At any rate,
43 all I/O is performed through the standard BSD socket API.
44 Some additions like zerocopy support are implemented through
45 control messages, while other extensions use the getsockopt/
46 setsockopt calls.
48 Sockets must be bound before you can send or receive data.
49 This is needed because binding also selects a transport and
50 attaches it to the socket. Once bound, the transport assignment
51 does not change. RDS will tolerate IPs moving around (eg in
52 a active-active HA scenario), but only as long as the address
53 doesn't move to a different transport.
55 * sysctls
56 RDS supports a number of sysctls in /proc/sys/net/rds
59 Socket Interface
60 ================
62 AF_RDS, PF_RDS, SOL_RDS
63 AF_RDS and PF_RDS are the domain type to be used with socket(2)
64 to create RDS sockets. SOL_RDS is the socket-level to be used
65 with setsockopt(2) and getsockopt(2) for RDS specific socket
66 options.
68 fd = socket(PF_RDS, SOCK_SEQPACKET, 0);
69 This creates a new, unbound RDS socket.
71 setsockopt(SOL_SOCKET): send and receive buffer size
72 RDS honors the send and receive buffer size socket options.
73 You are not allowed to queue more than SO_SNDSIZE bytes to
74 a socket. A message is queued when sendmsg is called, and
75 it leaves the queue when the remote system acknowledges
76 its arrival.
78 The SO_RCVSIZE option controls the maximum receive queue length.
79 This is a soft limit rather than a hard limit - RDS will
80 continue to accept and queue incoming messages, even if that
81 takes the queue length over the limit. However, it will also
82 mark the port as "congested" and send a congestion update to
83 the source node. The source node is supposed to throttle any
84 processes sending to this congested port.
86 bind(fd, &sockaddr_in, ...)
87 This binds the socket to a local IP address and port, and a
88 transport, if one has not already been selected via the
89 SO_RDS_TRANSPORT socket option
91 sendmsg(fd, ...)
92 Sends a message to the indicated recipient. The kernel will
93 transparently establish the underlying reliable connection
94 if it isn't up yet.
96 An attempt to send a message that exceeds SO_SNDSIZE will
97 return with -EMSGSIZE
99 An attempt to send a message that would take the total number
100 of queued bytes over the SO_SNDSIZE threshold will return
101 EAGAIN.
103 An attempt to send a message to a destination that is marked
104 as "congested" will return ENOBUFS.
106 recvmsg(fd, ...)
107 Receives a message that was queued to this socket. The sockets
108 recv queue accounting is adjusted, and if the queue length
109 drops below SO_SNDSIZE, the port is marked uncongested, and
110 a congestion update is sent to all peers.
112 Applications can ask the RDS kernel module to receive
113 notifications via control messages (for instance, there is a
114 notification when a congestion update arrived, or when a RDMA
115 operation completes). These notifications are received through
116 the msg.msg_control buffer of struct msghdr. The format of the
117 messages is described in manpages.
119 poll(fd)
120 RDS supports the poll interface to allow the application
121 to implement async I/O.
123 POLLIN handling is pretty straightforward. When there's an
124 incoming message queued to the socket, or a pending notification,
125 we signal POLLIN.
127 POLLOUT is a little harder. Since you can essentially send
128 to any destination, RDS will always signal POLLOUT as long as
129 there's room on the send queue (ie the number of bytes queued
130 is less than the sendbuf size).
132 However, the kernel will refuse to accept messages to
133 a destination marked congested - in this case you will loop
134 forever if you rely on poll to tell you what to do.
135 This isn't a trivial problem, but applications can deal with
136 this - by using congestion notifications, and by checking for
137 ENOBUFS errors returned by sendmsg.
139 setsockopt(SOL_RDS, RDS_CANCEL_SENT_TO, &sockaddr_in)
140 This allows the application to discard all messages queued to a
141 specific destination on this particular socket.
143 This allows the application to cancel outstanding messages if
144 it detects a timeout. For instance, if it tried to send a message,
145 and the remote host is unreachable, RDS will keep trying forever.
146 The application may decide it's not worth it, and cancel the
147 operation. In this case, it would use RDS_CANCEL_SENT_TO to
148 nuke any pending messages.
150 setsockopt(fd, SOL_RDS, SO_RDS_TRANSPORT, (int *)&transport ..)
151 getsockopt(fd, SOL_RDS, SO_RDS_TRANSPORT, (int *)&transport ..)
152 Set or read an integer defining the underlying
153 encapsulating transport to be used for RDS packets on the
154 socket. When setting the option, integer argument may be
155 one of RDS_TRANS_TCP or RDS_TRANS_IB. When retrieving the
156 value, RDS_TRANS_NONE will be returned on an unbound socket.
157 This socket option may only be set exactly once on the socket,
158 prior to binding it via the bind(2) system call. Attempts to
159 set SO_RDS_TRANSPORT on a socket for which the transport has
160 been previously attached explicitly (by SO_RDS_TRANSPORT) or
161 implicitly (via bind(2)) will return an error of EOPNOTSUPP.
162 An attempt to set SO_RDS_TRANSPPORT to RDS_TRANS_NONE will
163 always return EINVAL.
165 RDMA for RDS
166 ============
168 see rds-rdma(7) manpage (available in rds-tools)
171 Congestion Notifications
172 ========================
174 see rds(7) manpage
177 RDS Protocol
178 ============
180 Message header
182 The message header is a 'struct rds_header' (see rds.h):
183 Fields:
184 h_sequence:
185 per-packet sequence number
186 h_ack:
187 piggybacked acknowledgment of last packet received
188 h_len:
189 length of data, not including header
190 h_sport:
191 source port
192 h_dport:
193 destination port
194 h_flags:
195 CONG_BITMAP - this is a congestion update bitmap
196 ACK_REQUIRED - receiver must ack this packet
197 RETRANSMITTED - packet has previously been sent
198 h_credit:
199 indicate to other end of connection that
200 it has more credits available (i.e. there is
201 more send room)
202 h_padding[4]:
203 unused, for future use
204 h_csum:
205 header checksum
206 h_exthdr:
207 optional data can be passed here. This is currently used for
208 passing RDMA-related information.
210 ACK and retransmit handling
212 One might think that with reliable IB connections you wouldn't need
213 to ack messages that have been received. The problem is that IB
214 hardware generates an ack message before it has DMAed the message
215 into memory. This creates a potential message loss if the HCA is
216 disabled for any reason between when it sends the ack and before
217 the message is DMAed and processed. This is only a potential issue
218 if another HCA is available for fail-over.
220 Sending an ack immediately would allow the sender to free the sent
221 message from their send queue quickly, but could cause excessive
222 traffic to be used for acks. RDS piggybacks acks on sent data
223 packets. Ack-only packets are reduced by only allowing one to be
224 in flight at a time, and by the sender only asking for acks when
225 its send buffers start to fill up. All retransmissions are also
226 acked.
228 Flow Control
230 RDS's IB transport uses a credit-based mechanism to verify that
231 there is space in the peer's receive buffers for more data. This
232 eliminates the need for hardware retries on the connection.
234 Congestion
236 Messages waiting in the receive queue on the receiving socket
237 are accounted against the sockets SO_RCVBUF option value. Only
238 the payload bytes in the message are accounted for. If the
239 number of bytes queued equals or exceeds rcvbuf then the socket
240 is congested. All sends attempted to this socket's address
241 should return block or return -EWOULDBLOCK.
243 Applications are expected to be reasonably tuned such that this
244 situation very rarely occurs. An application encountering this
245 "back-pressure" is considered a bug.
247 This is implemented by having each node maintain bitmaps which
248 indicate which ports on bound addresses are congested. As the
249 bitmap changes it is sent through all the connections which
250 terminate in the local address of the bitmap which changed.
252 The bitmaps are allocated as connections are brought up. This
253 avoids allocation in the interrupt handling path which queues
254 sages on sockets. The dense bitmaps let transports send the
255 entire bitmap on any bitmap change reasonably efficiently. This
256 is much easier to implement than some finer-grained
257 communication of per-port congestion. The sender does a very
258 inexpensive bit test to test if the port it's about to send to
259 is congested or not.
262 RDS Transport Layer
263 ==================
265 As mentioned above, RDS is not IB-specific. Its code is divided
266 into a general RDS layer and a transport layer.
268 The general layer handles the socket API, congestion handling,
269 loopback, stats, usermem pinning, and the connection state machine.
271 The transport layer handles the details of the transport. The IB
272 transport, for example, handles all the queue pairs, work requests,
273 CM event handlers, and other Infiniband details.
276 RDS Kernel Structures
277 =====================
279 struct rds_message
280 aka possibly "rds_outgoing", the generic RDS layer copies data to
281 be sent and sets header fields as needed, based on the socket API.
282 This is then queued for the individual connection and sent by the
283 connection's transport.
284 struct rds_incoming
285 a generic struct referring to incoming data that can be handed from
286 the transport to the general code and queued by the general code
287 while the socket is awoken. It is then passed back to the transport
288 code to handle the actual copy-to-user.
289 struct rds_socket
290 per-socket information
291 struct rds_connection
292 per-connection information
293 struct rds_transport
294 pointers to transport-specific functions
295 struct rds_statistics
296 non-transport-specific statistics
297 struct rds_cong_map
298 wraps the raw congestion bitmap, contains rbnode, waitq, etc.
300 Connection management
301 =====================
303 Connections may be in UP, DOWN, CONNECTING, DISCONNECTING, and
304 ERROR states.
306 The first time an attempt is made by an RDS socket to send data to
307 a node, a connection is allocated and connected. That connection is
308 then maintained forever -- if there are transport errors, the
309 connection will be dropped and re-established.
311 Dropping a connection while packets are queued will cause queued or
312 partially-sent datagrams to be retransmitted when the connection is
313 re-established.
316 The send path
317 =============
319 rds_sendmsg()
320 struct rds_message built from incoming data
321 CMSGs parsed (e.g. RDMA ops)
322 transport connection alloced and connected if not already
323 rds_message placed on send queue
324 send worker awoken
325 rds_send_worker()
326 calls rds_send_xmit() until queue is empty
327 rds_send_xmit()
328 transmits congestion map if one is pending
329 may set ACK_REQUIRED
330 calls transport to send either non-RDMA or RDMA message
331 (RDMA ops never retransmitted)
332 rds_ib_xmit()
333 allocs work requests from send ring
334 adds any new send credits available to peer (h_credits)
335 maps the rds_message's sg list
336 piggybacks ack
337 populates work requests
338 post send to connection's queue pair
340 The recv path
341 =============
343 rds_ib_recv_cq_comp_handler()
344 looks at write completions
345 unmaps recv buffer from device
346 no errors, call rds_ib_process_recv()
347 refill recv ring
348 rds_ib_process_recv()
349 validate header checksum
350 copy header to rds_ib_incoming struct if start of a new datagram
351 add to ibinc's fraglist
352 if competed datagram:
353 update cong map if datagram was cong update
354 call rds_recv_incoming() otherwise
355 note if ack is required
356 rds_recv_incoming()
357 drop duplicate packets
358 respond to pings
359 find the sock associated with this datagram
360 add to sock queue
361 wake up sock
362 do some congestion calculations
363 rds_recvmsg
364 copy data into user iovec
365 handle CMSGs
366 return to application
368 Multipath RDS (mprds)
369 =====================
370 Mprds is multipathed-RDS, primarily intended for RDS-over-TCP
371 (though the concept can be extended to other transports). The classical
372 implementation of RDS-over-TCP is implemented by demultiplexing multiple
373 PF_RDS sockets between any 2 endpoints (where endpoint == [IP address,
374 port]) over a single TCP socket between the 2 IP addresses involved. This
375 has the limitation that it ends up funneling multiple RDS flows over a
376 single TCP flow, thus it is
377 (a) upper-bounded to the single-flow bandwidth,
378 (b) suffers from head-of-line blocking for all the RDS sockets.
380 Better throughput (for a fixed small packet size, MTU) can be achieved
381 by having multiple TCP/IP flows per rds/tcp connection, i.e., multipathed
382 RDS (mprds). Each such TCP/IP flow constitutes a path for the rds/tcp
383 connection. RDS sockets will be attached to a path based on some hash
384 (e.g., of local address and RDS port number) and packets for that RDS
385 socket will be sent over the attached path using TCP to segment/reassemble
386 RDS datagrams on that path.
388 Multipathed RDS is implemented by splitting the struct rds_connection into
389 a common (to all paths) part, and a per-path struct rds_conn_path. All
390 I/O workqs and reconnect threads are driven from the rds_conn_path.
391 Transports such as TCP that are multipath capable may then set up a
392 TPC socket per rds_conn_path, and this is managed by the transport via
393 the transport privatee cp_transport_data pointer.
395 Transports announce themselves as multipath capable by setting the
396 t_mp_capable bit during registration with the rds core module. When the
397 transport is multipath-capable, rds_sendmsg() hashes outgoing traffic
398 across multiple paths. The outgoing hash is computed based on the
399 local address and port that the PF_RDS socket is bound to.
401 Additionally, even if the transport is MP capable, we may be
402 peering with some node that does not support mprds, or supports
403 a different number of paths. As a result, the peering nodes need
404 to agree on the number of paths to be used for the connection.
405 This is done by sending out a control packet exchange before the
406 first data packet. The control packet exchange must have completed
407 prior to outgoing hash completion in rds_sendmsg() when the transport
408 is mutlipath capable.
410 The control packet is an RDS ping packet (i.e., packet to rds dest
411 port 0) with the ping packet having a rds extension header option of
412 type RDS_EXTHDR_NPATHS, length 2 bytes, and the value is the
413 number of paths supported by the sender. The "probe" ping packet will
414 get sent from some reserved port, RDS_FLAG_PROBE_PORT (in <linux/rds.h>)
415 The receiver of a ping from RDS_FLAG_PROBE_PORT will thus immediately
416 be able to compute the min(sender_paths, rcvr_paths). The pong
417 sent in response to a probe-ping should contain the rcvr's npaths
418 when the rcvr is mprds-capable.
420 If the rcvr is not mprds-capable, the exthdr in the ping will be
421 ignored. In this case the pong will not have any exthdrs, so the sender
422 of the probe-ping can default to single-path mprds.