2 * Copyright (c) 2006 Oracle. All rights reserved.
4 * This software is available to you under a choice of one of two
5 * licenses. You may choose to be licensed under the terms of the GNU
6 * General Public License (GPL) Version 2, available from the file
7 * COPYING in the main directory of this source tree, or the
8 * OpenIB.org BSD license below:
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
18 * - Redistributions in binary form must reproduce the above
19 * copyright notice, this list of conditions and the following
20 * disclaimer in the documentation and/or other materials
21 * provided with the distribution.
23 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
24 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
25 * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
26 * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
27 * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
28 * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
29 * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
33 #include <linux/kernel.h>
34 #include <linux/slab.h>
37 #include <linux/export.h>
38 #include <linux/time.h>
39 #include <linux/rds.h>
43 void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
46 atomic_set(&inc->i_refcount, 1);
47 INIT_LIST_HEAD(&inc->i_item);
50 inc->i_rdma_cookie = 0;
51 inc->i_rx_tstamp.tv_sec = 0;
52 inc->i_rx_tstamp.tv_usec = 0;
54 EXPORT_SYMBOL_GPL(rds_inc_init);
56 void rds_inc_path_init(struct rds_incoming *inc, struct rds_conn_path *cp,
59 atomic_set(&inc->i_refcount, 1);
60 INIT_LIST_HEAD(&inc->i_item);
61 inc->i_conn = cp->cp_conn;
62 inc->i_conn_path = cp;
64 inc->i_rdma_cookie = 0;
65 inc->i_rx_tstamp.tv_sec = 0;
66 inc->i_rx_tstamp.tv_usec = 0;
68 EXPORT_SYMBOL_GPL(rds_inc_path_init);
70 static void rds_inc_addref(struct rds_incoming *inc)
72 rdsdebug("addref inc %p ref %d\n", inc, atomic_read(&inc->i_refcount));
73 atomic_inc(&inc->i_refcount);
76 void rds_inc_put(struct rds_incoming *inc)
78 rdsdebug("put inc %p ref %d\n", inc, atomic_read(&inc->i_refcount));
79 if (atomic_dec_and_test(&inc->i_refcount)) {
80 BUG_ON(!list_empty(&inc->i_item));
82 inc->i_conn->c_trans->inc_free(inc);
85 EXPORT_SYMBOL_GPL(rds_inc_put);
87 static void rds_recv_rcvbuf_delta(struct rds_sock *rs, struct sock *sk,
88 struct rds_cong_map *map,
89 int delta, __be16 port)
96 rs->rs_rcv_bytes += delta;
97 now_congested = rs->rs_rcv_bytes > rds_sk_rcvbuf(rs);
99 rdsdebug("rs %p (%pI4:%u) recv bytes %d buf %d "
100 "now_cong %d delta %d\n",
101 rs, &rs->rs_bound_addr,
102 ntohs(rs->rs_bound_port), rs->rs_rcv_bytes,
103 rds_sk_rcvbuf(rs), now_congested, delta);
105 /* wasn't -> am congested */
106 if (!rs->rs_congested && now_congested) {
107 rs->rs_congested = 1;
108 rds_cong_set_bit(map, port);
109 rds_cong_queue_updates(map);
111 /* was -> aren't congested */
112 /* Require more free space before reporting uncongested to prevent
113 bouncing cong/uncong state too often */
114 else if (rs->rs_congested && (rs->rs_rcv_bytes < (rds_sk_rcvbuf(rs)/2))) {
115 rs->rs_congested = 0;
116 rds_cong_clear_bit(map, port);
117 rds_cong_queue_updates(map);
120 /* do nothing if no change in cong state */
124 * Process all extension headers that come with this message.
126 static void rds_recv_incoming_exthdrs(struct rds_incoming *inc, struct rds_sock *rs)
128 struct rds_header *hdr = &inc->i_hdr;
129 unsigned int pos = 0, type, len;
131 struct rds_ext_header_version version;
132 struct rds_ext_header_rdma rdma;
133 struct rds_ext_header_rdma_dest rdma_dest;
137 len = sizeof(buffer);
138 type = rds_message_next_extension(hdr, &pos, &buffer, &len);
139 if (type == RDS_EXTHDR_NONE)
141 /* Process extension header here */
143 case RDS_EXTHDR_RDMA:
144 rds_rdma_unuse(rs, be32_to_cpu(buffer.rdma.h_rdma_rkey), 0);
147 case RDS_EXTHDR_RDMA_DEST:
148 /* We ignore the size for now. We could stash it
149 * somewhere and use it for error checking. */
150 inc->i_rdma_cookie = rds_rdma_make_cookie(
151 be32_to_cpu(buffer.rdma_dest.h_rdma_rkey),
152 be32_to_cpu(buffer.rdma_dest.h_rdma_offset));
159 static void rds_recv_hs_exthdrs(struct rds_header *hdr,
160 struct rds_connection *conn)
162 unsigned int pos = 0, type, len;
164 struct rds_ext_header_version version;
169 len = sizeof(buffer);
170 type = rds_message_next_extension(hdr, &pos, &buffer, &len);
171 if (type == RDS_EXTHDR_NONE)
173 /* Process extension header here */
175 case RDS_EXTHDR_NPATHS:
176 conn->c_npaths = min_t(int, RDS_MPATH_WORKERS,
180 pr_warn_ratelimited("ignoring unknown exthdr type "
184 /* if RDS_EXTHDR_NPATHS was not found, default to a single-path */
185 conn->c_npaths = max_t(int, conn->c_npaths, 1);
188 /* rds_start_mprds() will synchronously start multiple paths when appropriate.
189 * The scheme is based on the following rules:
191 * 1. rds_sendmsg on first connect attempt sends the probe ping, with the
192 * sender's npaths (s_npaths)
193 * 2. rcvr of probe-ping knows the mprds_paths = min(s_npaths, r_npaths). It
194 * sends back a probe-pong with r_npaths. After that, if rcvr is the
195 * smaller ip addr, it starts rds_conn_path_connect_if_down on all
197 * 3. sender gets woken up, and can move to rds_conn_path_connect_if_down.
198 * If it is the smaller ipaddr, rds_conn_path_connect_if_down can be
199 * called after reception of the probe-pong on all mprds_paths.
200 * Otherwise (sender of probe-ping is not the smaller ip addr): just call
201 * rds_conn_path_connect_if_down on the hashed path. (see rule 4)
202 * 4. when cp_index > 0, rds_connect_worker must only trigger
203 * a connection if laddr < faddr.
204 * 5. sender may end up queuing the packet on the cp. will get sent out later.
205 * when connection is completed.
207 static void rds_start_mprds(struct rds_connection *conn)
210 struct rds_conn_path *cp;
212 if (conn->c_npaths > 1 && conn->c_laddr < conn->c_faddr) {
213 for (i = 1; i < conn->c_npaths; i++) {
214 cp = &conn->c_path[i];
215 rds_conn_path_connect_if_down(cp);
221 * The transport must make sure that this is serialized against other
222 * rx and conn reset on this specific conn.
224 * We currently assert that only one fragmented message will be sent
225 * down a connection at a time. This lets us reassemble in the conn
226 * instead of per-flow which means that we don't have to go digging through
227 * flows to tear down partial reassembly progress on conn failure and
228 * we save flow lookup and locking for each frag arrival. It does mean
229 * that small messages will wait behind large ones. Fragmenting at all
230 * is only to reduce the memory consumption of pre-posted buffers.
232 * The caller passes in saddr and daddr instead of us getting it from the
233 * conn. This lets loopback, who only has one conn for both directions,
234 * tell us which roles the addrs in the conn are playing for this message.
236 void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
237 struct rds_incoming *inc, gfp_t gfp)
239 struct rds_sock *rs = NULL;
242 struct rds_conn_path *cp;
245 inc->i_rx_jiffies = jiffies;
246 if (conn->c_trans->t_mp_capable)
247 cp = inc->i_conn_path;
249 cp = &conn->c_path[0];
251 rdsdebug("conn %p next %llu inc %p seq %llu len %u sport %u dport %u "
252 "flags 0x%x rx_jiffies %lu\n", conn,
253 (unsigned long long)cp->cp_next_rx_seq,
255 (unsigned long long)be64_to_cpu(inc->i_hdr.h_sequence),
256 be32_to_cpu(inc->i_hdr.h_len),
257 be16_to_cpu(inc->i_hdr.h_sport),
258 be16_to_cpu(inc->i_hdr.h_dport),
263 * Sequence numbers should only increase. Messages get their
264 * sequence number as they're queued in a sending conn. They
265 * can be dropped, though, if the sending socket is closed before
266 * they hit the wire. So sequence numbers can skip forward
267 * under normal operation. They can also drop back in the conn
268 * failover case as previously sent messages are resent down the
269 * new instance of a conn. We drop those, otherwise we have
270 * to assume that the next valid seq does not come after a
271 * hole in the fragment stream.
273 * The headers don't give us a way to realize if fragments of
274 * a message have been dropped. We assume that frags that arrive
275 * to a flow are part of the current message on the flow that is
276 * being reassembled. This means that senders can't drop messages
277 * from the sending conn until all their frags are sent.
279 * XXX we could spend more on the wire to get more robust failure
280 * detection, arguably worth it to avoid data corruption.
282 if (be64_to_cpu(inc->i_hdr.h_sequence) < cp->cp_next_rx_seq &&
283 (inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) {
284 rds_stats_inc(s_recv_drop_old_seq);
287 cp->cp_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1;
289 if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) {
290 if (inc->i_hdr.h_sport == 0) {
291 rdsdebug("ignore ping with 0 sport from 0x%x\n", saddr);
294 rds_stats_inc(s_recv_ping);
295 rds_send_pong(cp, inc->i_hdr.h_sport);
296 /* if this is a handshake ping, start multipath if necessary */
297 if (RDS_HS_PROBE(inc->i_hdr.h_sport, inc->i_hdr.h_dport)) {
298 rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn);
299 rds_start_mprds(cp->cp_conn);
304 if (inc->i_hdr.h_dport == RDS_FLAG_PROBE_PORT &&
305 inc->i_hdr.h_sport == 0) {
306 rds_recv_hs_exthdrs(&inc->i_hdr, cp->cp_conn);
307 /* if this is a handshake pong, start multipath if necessary */
308 rds_start_mprds(cp->cp_conn);
309 wake_up(&cp->cp_conn->c_hs_waitq);
313 rs = rds_find_bound(daddr, inc->i_hdr.h_dport);
315 rds_stats_inc(s_recv_drop_no_sock);
319 /* Process extension headers */
320 rds_recv_incoming_exthdrs(inc, rs);
322 /* We can be racing with rds_release() which marks the socket dead. */
323 sk = rds_rs_to_sk(rs);
325 /* serialize with rds_release -> sock_orphan */
326 write_lock_irqsave(&rs->rs_recv_lock, flags);
327 if (!sock_flag(sk, SOCK_DEAD)) {
328 rdsdebug("adding inc %p to rs %p's recv queue\n", inc, rs);
329 rds_stats_inc(s_recv_queued);
330 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
331 be32_to_cpu(inc->i_hdr.h_len),
333 if (sock_flag(sk, SOCK_RCVTSTAMP))
334 do_gettimeofday(&inc->i_rx_tstamp);
336 list_add_tail(&inc->i_item, &rs->rs_recv_queue);
337 __rds_wake_sk_sleep(sk);
339 rds_stats_inc(s_recv_drop_dead_sock);
341 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
347 EXPORT_SYMBOL_GPL(rds_recv_incoming);
350 * be very careful here. This is being called as the condition in
351 * wait_event_*() needs to cope with being called many times.
353 static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc)
358 read_lock_irqsave(&rs->rs_recv_lock, flags);
359 if (!list_empty(&rs->rs_recv_queue)) {
360 *inc = list_entry(rs->rs_recv_queue.next,
363 rds_inc_addref(*inc);
365 read_unlock_irqrestore(&rs->rs_recv_lock, flags);
371 static int rds_still_queued(struct rds_sock *rs, struct rds_incoming *inc,
374 struct sock *sk = rds_rs_to_sk(rs);
378 write_lock_irqsave(&rs->rs_recv_lock, flags);
379 if (!list_empty(&inc->i_item)) {
382 /* XXX make sure this i_conn is reliable */
383 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
384 -be32_to_cpu(inc->i_hdr.h_len),
386 list_del_init(&inc->i_item);
390 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
392 rdsdebug("inc %p rs %p still %d dropped %d\n", inc, rs, ret, drop);
397 * Pull errors off the error queue.
398 * If msghdr is NULL, we will just purge the error queue.
400 int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr)
402 struct rds_notifier *notifier;
403 struct rds_rdma_notify cmsg = { 0 }; /* fill holes with zero */
404 unsigned int count = 0, max_messages = ~0U;
410 /* put_cmsg copies to user space and thus may sleep. We can't do this
411 * with rs_lock held, so first grab as many notifications as we can stuff
412 * in the user provided cmsg buffer. We don't try to copy more, to avoid
413 * losing notifications - except when the buffer is so small that it wouldn't
414 * even hold a single notification. Then we give him as much of this single
415 * msg as we can squeeze in, and set MSG_CTRUNC.
418 max_messages = msghdr->msg_controllen / CMSG_SPACE(sizeof(cmsg));
423 spin_lock_irqsave(&rs->rs_lock, flags);
424 while (!list_empty(&rs->rs_notify_queue) && count < max_messages) {
425 notifier = list_entry(rs->rs_notify_queue.next,
426 struct rds_notifier, n_list);
427 list_move(¬ifier->n_list, ©);
430 spin_unlock_irqrestore(&rs->rs_lock, flags);
435 while (!list_empty(©)) {
436 notifier = list_entry(copy.next, struct rds_notifier, n_list);
439 cmsg.user_token = notifier->n_user_token;
440 cmsg.status = notifier->n_status;
442 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS,
443 sizeof(cmsg), &cmsg);
448 list_del_init(¬ifier->n_list);
452 /* If we bailed out because of an error in put_cmsg,
453 * we may be left with one or more notifications that we
454 * didn't process. Return them to the head of the list. */
455 if (!list_empty(©)) {
456 spin_lock_irqsave(&rs->rs_lock, flags);
457 list_splice(©, &rs->rs_notify_queue);
458 spin_unlock_irqrestore(&rs->rs_lock, flags);
465 * Queue a congestion notification
467 static int rds_notify_cong(struct rds_sock *rs, struct msghdr *msghdr)
469 uint64_t notify = rs->rs_cong_notify;
473 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_CONG_UPDATE,
474 sizeof(notify), ¬ify);
478 spin_lock_irqsave(&rs->rs_lock, flags);
479 rs->rs_cong_notify &= ~notify;
480 spin_unlock_irqrestore(&rs->rs_lock, flags);
486 * Receive any control messages.
488 static int rds_cmsg_recv(struct rds_incoming *inc, struct msghdr *msg,
493 if (inc->i_rdma_cookie) {
494 ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RDMA_DEST,
495 sizeof(inc->i_rdma_cookie), &inc->i_rdma_cookie);
500 if ((inc->i_rx_tstamp.tv_sec != 0) &&
501 sock_flag(rds_rs_to_sk(rs), SOCK_RCVTSTAMP)) {
502 ret = put_cmsg(msg, SOL_SOCKET, SCM_TIMESTAMP,
503 sizeof(struct timeval),
512 int rds_recvmsg(struct socket *sock, struct msghdr *msg, size_t size,
515 struct sock *sk = sock->sk;
516 struct rds_sock *rs = rds_sk_to_rs(sk);
518 int ret = 0, nonblock = msg_flags & MSG_DONTWAIT;
519 DECLARE_SOCKADDR(struct sockaddr_in *, sin, msg->msg_name);
520 struct rds_incoming *inc = NULL;
522 /* udp_recvmsg()->sock_recvtimeo() gets away without locking too.. */
523 timeo = sock_rcvtimeo(sk, nonblock);
525 rdsdebug("size %zu flags 0x%x timeo %ld\n", size, msg_flags, timeo);
527 if (msg_flags & MSG_OOB)
531 struct iov_iter save;
532 /* If there are pending notifications, do those - and nothing else */
533 if (!list_empty(&rs->rs_notify_queue)) {
534 ret = rds_notify_queue_get(rs, msg);
538 if (rs->rs_cong_notify) {
539 ret = rds_notify_cong(rs, msg);
543 if (!rds_next_incoming(rs, &inc)) {
549 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
550 (!list_empty(&rs->rs_notify_queue) ||
551 rs->rs_cong_notify ||
552 rds_next_incoming(rs, &inc)), timeo);
553 rdsdebug("recvmsg woke inc %p timeo %ld\n", inc,
555 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
564 rdsdebug("copying inc %p from %pI4:%u to user\n", inc,
565 &inc->i_conn->c_faddr,
566 ntohs(inc->i_hdr.h_sport));
567 save = msg->msg_iter;
568 ret = inc->i_conn->c_trans->inc_copy_to_user(inc, &msg->msg_iter);
573 * if the message we just copied isn't at the head of the
574 * recv queue then someone else raced us to return it, try
575 * to get the next message.
577 if (!rds_still_queued(rs, inc, !(msg_flags & MSG_PEEK))) {
580 rds_stats_inc(s_recv_deliver_raced);
581 msg->msg_iter = save;
585 if (ret < be32_to_cpu(inc->i_hdr.h_len)) {
586 if (msg_flags & MSG_TRUNC)
587 ret = be32_to_cpu(inc->i_hdr.h_len);
588 msg->msg_flags |= MSG_TRUNC;
591 if (rds_cmsg_recv(inc, msg, rs)) {
596 rds_stats_inc(s_recv_delivered);
599 sin->sin_family = AF_INET;
600 sin->sin_port = inc->i_hdr.h_sport;
601 sin->sin_addr.s_addr = inc->i_saddr;
602 memset(sin->sin_zero, 0, sizeof(sin->sin_zero));
603 msg->msg_namelen = sizeof(*sin);
616 * The socket is being shut down and we're asked to drop messages that were
617 * queued for recvmsg. The caller has unbound the socket so the receive path
618 * won't queue any more incoming fragments or messages on the socket.
620 void rds_clear_recv_queue(struct rds_sock *rs)
622 struct sock *sk = rds_rs_to_sk(rs);
623 struct rds_incoming *inc, *tmp;
626 write_lock_irqsave(&rs->rs_recv_lock, flags);
627 list_for_each_entry_safe(inc, tmp, &rs->rs_recv_queue, i_item) {
628 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
629 -be32_to_cpu(inc->i_hdr.h_len),
631 list_del_init(&inc->i_item);
634 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
638 * inc->i_saddr isn't used here because it is only set in the receive
641 void rds_inc_info_copy(struct rds_incoming *inc,
642 struct rds_info_iterator *iter,
643 __be32 saddr, __be32 daddr, int flip)
645 struct rds_info_message minfo;
647 minfo.seq = be64_to_cpu(inc->i_hdr.h_sequence);
648 minfo.len = be32_to_cpu(inc->i_hdr.h_len);
653 minfo.lport = inc->i_hdr.h_dport;
654 minfo.fport = inc->i_hdr.h_sport;
658 minfo.lport = inc->i_hdr.h_sport;
659 minfo.fport = inc->i_hdr.h_dport;
664 rds_info_copy(iter, &minfo, sizeof(minfo));