blob: b426d67f760c61f77717a40057d71fd0090da46e [file] [log] [blame]
Andy Groverbdbe6fb2009-02-24 15:30:28 +00001/*
2 * Copyright (c) 2006 Oracle. All rights reserved.
3 *
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:
9 *
10 * Redistribution and use in source and binary forms, with or
11 * without modification, are permitted provided that the following
12 * conditions are met:
13 *
14 * - Redistributions of source code must retain the above
15 * copyright notice, this list of conditions and the following
16 * disclaimer.
17 *
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.
22 *
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
30 * SOFTWARE.
31 *
32 */
33#include <linux/kernel.h>
34#include <net/sock.h>
35#include <linux/in.h>
36
37#include "rds.h"
38#include "rdma.h"
39
40void rds_inc_init(struct rds_incoming *inc, struct rds_connection *conn,
41 __be32 saddr)
42{
43 atomic_set(&inc->i_refcount, 1);
44 INIT_LIST_HEAD(&inc->i_item);
45 inc->i_conn = conn;
46 inc->i_saddr = saddr;
47 inc->i_rdma_cookie = 0;
48}
Andy Grover616b7572009-08-21 12:28:32 +000049EXPORT_SYMBOL_GPL(rds_inc_init);
Andy Groverbdbe6fb2009-02-24 15:30:28 +000050
51void rds_inc_addref(struct rds_incoming *inc)
52{
53 rdsdebug("addref inc %p ref %d\n", inc, atomic_read(&inc->i_refcount));
54 atomic_inc(&inc->i_refcount);
55}
Andy Grover616b7572009-08-21 12:28:32 +000056EXPORT_SYMBOL_GPL(rds_inc_addref);
Andy Groverbdbe6fb2009-02-24 15:30:28 +000057
58void rds_inc_put(struct rds_incoming *inc)
59{
60 rdsdebug("put inc %p ref %d\n", inc, atomic_read(&inc->i_refcount));
61 if (atomic_dec_and_test(&inc->i_refcount)) {
62 BUG_ON(!list_empty(&inc->i_item));
63
64 inc->i_conn->c_trans->inc_free(inc);
65 }
66}
Andy Grover616b7572009-08-21 12:28:32 +000067EXPORT_SYMBOL_GPL(rds_inc_put);
Andy Groverbdbe6fb2009-02-24 15:30:28 +000068
69static void rds_recv_rcvbuf_delta(struct rds_sock *rs, struct sock *sk,
70 struct rds_cong_map *map,
71 int delta, __be16 port)
72{
73 int now_congested;
74
75 if (delta == 0)
76 return;
77
78 rs->rs_rcv_bytes += delta;
79 now_congested = rs->rs_rcv_bytes > rds_sk_rcvbuf(rs);
80
81 rdsdebug("rs %p (%pI4:%u) recv bytes %d buf %d "
82 "now_cong %d delta %d\n",
83 rs, &rs->rs_bound_addr,
84 ntohs(rs->rs_bound_port), rs->rs_rcv_bytes,
85 rds_sk_rcvbuf(rs), now_congested, delta);
86
87 /* wasn't -> am congested */
88 if (!rs->rs_congested && now_congested) {
89 rs->rs_congested = 1;
90 rds_cong_set_bit(map, port);
91 rds_cong_queue_updates(map);
92 }
93 /* was -> aren't congested */
94 /* Require more free space before reporting uncongested to prevent
95 bouncing cong/uncong state too often */
96 else if (rs->rs_congested && (rs->rs_rcv_bytes < (rds_sk_rcvbuf(rs)/2))) {
97 rs->rs_congested = 0;
98 rds_cong_clear_bit(map, port);
99 rds_cong_queue_updates(map);
100 }
101
102 /* do nothing if no change in cong state */
103}
104
105/*
106 * Process all extension headers that come with this message.
107 */
108static void rds_recv_incoming_exthdrs(struct rds_incoming *inc, struct rds_sock *rs)
109{
110 struct rds_header *hdr = &inc->i_hdr;
111 unsigned int pos = 0, type, len;
112 union {
113 struct rds_ext_header_version version;
114 struct rds_ext_header_rdma rdma;
115 struct rds_ext_header_rdma_dest rdma_dest;
116 } buffer;
117
118 while (1) {
119 len = sizeof(buffer);
120 type = rds_message_next_extension(hdr, &pos, &buffer, &len);
121 if (type == RDS_EXTHDR_NONE)
122 break;
123 /* Process extension header here */
124 switch (type) {
125 case RDS_EXTHDR_RDMA:
126 rds_rdma_unuse(rs, be32_to_cpu(buffer.rdma.h_rdma_rkey), 0);
127 break;
128
129 case RDS_EXTHDR_RDMA_DEST:
130 /* We ignore the size for now. We could stash it
131 * somewhere and use it for error checking. */
132 inc->i_rdma_cookie = rds_rdma_make_cookie(
133 be32_to_cpu(buffer.rdma_dest.h_rdma_rkey),
134 be32_to_cpu(buffer.rdma_dest.h_rdma_offset));
135
136 break;
137 }
138 }
139}
140
141/*
142 * The transport must make sure that this is serialized against other
143 * rx and conn reset on this specific conn.
144 *
145 * We currently assert that only one fragmented message will be sent
146 * down a connection at a time. This lets us reassemble in the conn
147 * instead of per-flow which means that we don't have to go digging through
148 * flows to tear down partial reassembly progress on conn failure and
149 * we save flow lookup and locking for each frag arrival. It does mean
150 * that small messages will wait behind large ones. Fragmenting at all
151 * is only to reduce the memory consumption of pre-posted buffers.
152 *
153 * The caller passes in saddr and daddr instead of us getting it from the
154 * conn. This lets loopback, who only has one conn for both directions,
155 * tell us which roles the addrs in the conn are playing for this message.
156 */
157void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
158 struct rds_incoming *inc, gfp_t gfp, enum km_type km)
159{
160 struct rds_sock *rs = NULL;
161 struct sock *sk;
162 unsigned long flags;
163
164 inc->i_conn = conn;
165 inc->i_rx_jiffies = jiffies;
166
167 rdsdebug("conn %p next %llu inc %p seq %llu len %u sport %u dport %u "
168 "flags 0x%x rx_jiffies %lu\n", conn,
169 (unsigned long long)conn->c_next_rx_seq,
170 inc,
171 (unsigned long long)be64_to_cpu(inc->i_hdr.h_sequence),
172 be32_to_cpu(inc->i_hdr.h_len),
173 be16_to_cpu(inc->i_hdr.h_sport),
174 be16_to_cpu(inc->i_hdr.h_dport),
175 inc->i_hdr.h_flags,
176 inc->i_rx_jiffies);
177
178 /*
179 * Sequence numbers should only increase. Messages get their
180 * sequence number as they're queued in a sending conn. They
181 * can be dropped, though, if the sending socket is closed before
182 * they hit the wire. So sequence numbers can skip forward
183 * under normal operation. They can also drop back in the conn
184 * failover case as previously sent messages are resent down the
185 * new instance of a conn. We drop those, otherwise we have
186 * to assume that the next valid seq does not come after a
187 * hole in the fragment stream.
188 *
189 * The headers don't give us a way to realize if fragments of
190 * a message have been dropped. We assume that frags that arrive
191 * to a flow are part of the current message on the flow that is
192 * being reassembled. This means that senders can't drop messages
193 * from the sending conn until all their frags are sent.
194 *
195 * XXX we could spend more on the wire to get more robust failure
196 * detection, arguably worth it to avoid data corruption.
197 */
Joe Perchesf64f9e72009-11-29 16:55:45 -0800198 if (be64_to_cpu(inc->i_hdr.h_sequence) < conn->c_next_rx_seq &&
199 (inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) {
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000200 rds_stats_inc(s_recv_drop_old_seq);
201 goto out;
202 }
203 conn->c_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1;
204
205 if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) {
206 rds_stats_inc(s_recv_ping);
207 rds_send_pong(conn, inc->i_hdr.h_sport);
208 goto out;
209 }
210
211 rs = rds_find_bound(daddr, inc->i_hdr.h_dport);
212 if (rs == NULL) {
213 rds_stats_inc(s_recv_drop_no_sock);
214 goto out;
215 }
216
217 /* Process extension headers */
218 rds_recv_incoming_exthdrs(inc, rs);
219
220 /* We can be racing with rds_release() which marks the socket dead. */
221 sk = rds_rs_to_sk(rs);
222
223 /* serialize with rds_release -> sock_orphan */
224 write_lock_irqsave(&rs->rs_recv_lock, flags);
225 if (!sock_flag(sk, SOCK_DEAD)) {
226 rdsdebug("adding inc %p to rs %p's recv queue\n", inc, rs);
227 rds_stats_inc(s_recv_queued);
228 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
229 be32_to_cpu(inc->i_hdr.h_len),
230 inc->i_hdr.h_dport);
231 rds_inc_addref(inc);
232 list_add_tail(&inc->i_item, &rs->rs_recv_queue);
233 __rds_wake_sk_sleep(sk);
234 } else {
235 rds_stats_inc(s_recv_drop_dead_sock);
236 }
237 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
238
239out:
240 if (rs)
241 rds_sock_put(rs);
242}
Andy Grover616b7572009-08-21 12:28:32 +0000243EXPORT_SYMBOL_GPL(rds_recv_incoming);
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000244
245/*
246 * be very careful here. This is being called as the condition in
247 * wait_event_*() needs to cope with being called many times.
248 */
249static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc)
250{
251 unsigned long flags;
252
253 if (*inc == NULL) {
254 read_lock_irqsave(&rs->rs_recv_lock, flags);
255 if (!list_empty(&rs->rs_recv_queue)) {
256 *inc = list_entry(rs->rs_recv_queue.next,
257 struct rds_incoming,
258 i_item);
259 rds_inc_addref(*inc);
260 }
261 read_unlock_irqrestore(&rs->rs_recv_lock, flags);
262 }
263
264 return *inc != NULL;
265}
266
267static int rds_still_queued(struct rds_sock *rs, struct rds_incoming *inc,
268 int drop)
269{
270 struct sock *sk = rds_rs_to_sk(rs);
271 int ret = 0;
272 unsigned long flags;
273
274 write_lock_irqsave(&rs->rs_recv_lock, flags);
275 if (!list_empty(&inc->i_item)) {
276 ret = 1;
277 if (drop) {
278 /* XXX make sure this i_conn is reliable */
279 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
280 -be32_to_cpu(inc->i_hdr.h_len),
281 inc->i_hdr.h_dport);
282 list_del_init(&inc->i_item);
283 rds_inc_put(inc);
284 }
285 }
286 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
287
288 rdsdebug("inc %p rs %p still %d dropped %d\n", inc, rs, ret, drop);
289 return ret;
290}
291
292/*
293 * Pull errors off the error queue.
294 * If msghdr is NULL, we will just purge the error queue.
295 */
296int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr)
297{
298 struct rds_notifier *notifier;
299 struct rds_rdma_notify cmsg;
300 unsigned int count = 0, max_messages = ~0U;
301 unsigned long flags;
302 LIST_HEAD(copy);
303 int err = 0;
304
305
306 /* put_cmsg copies to user space and thus may sleep. We can't do this
307 * with rs_lock held, so first grab as many notifications as we can stuff
308 * in the user provided cmsg buffer. We don't try to copy more, to avoid
309 * losing notifications - except when the buffer is so small that it wouldn't
310 * even hold a single notification. Then we give him as much of this single
311 * msg as we can squeeze in, and set MSG_CTRUNC.
312 */
313 if (msghdr) {
314 max_messages = msghdr->msg_controllen / CMSG_SPACE(sizeof(cmsg));
315 if (!max_messages)
316 max_messages = 1;
317 }
318
319 spin_lock_irqsave(&rs->rs_lock, flags);
320 while (!list_empty(&rs->rs_notify_queue) && count < max_messages) {
321 notifier = list_entry(rs->rs_notify_queue.next,
322 struct rds_notifier, n_list);
323 list_move(&notifier->n_list, &copy);
324 count++;
325 }
326 spin_unlock_irqrestore(&rs->rs_lock, flags);
327
328 if (!count)
329 return 0;
330
331 while (!list_empty(&copy)) {
332 notifier = list_entry(copy.next, struct rds_notifier, n_list);
333
334 if (msghdr) {
335 cmsg.user_token = notifier->n_user_token;
336 cmsg.status = notifier->n_status;
337
338 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS,
339 sizeof(cmsg), &cmsg);
340 if (err)
341 break;
342 }
343
344 list_del_init(&notifier->n_list);
345 kfree(notifier);
346 }
347
348 /* If we bailed out because of an error in put_cmsg,
349 * we may be left with one or more notifications that we
350 * didn't process. Return them to the head of the list. */
351 if (!list_empty(&copy)) {
352 spin_lock_irqsave(&rs->rs_lock, flags);
353 list_splice(&copy, &rs->rs_notify_queue);
354 spin_unlock_irqrestore(&rs->rs_lock, flags);
355 }
356
357 return err;
358}
359
360/*
361 * Queue a congestion notification
362 */
363static int rds_notify_cong(struct rds_sock *rs, struct msghdr *msghdr)
364{
365 uint64_t notify = rs->rs_cong_notify;
366 unsigned long flags;
367 int err;
368
369 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_CONG_UPDATE,
370 sizeof(notify), &notify);
371 if (err)
372 return err;
373
374 spin_lock_irqsave(&rs->rs_lock, flags);
375 rs->rs_cong_notify &= ~notify;
376 spin_unlock_irqrestore(&rs->rs_lock, flags);
377
378 return 0;
379}
380
381/*
382 * Receive any control messages.
383 */
384static int rds_cmsg_recv(struct rds_incoming *inc, struct msghdr *msg)
385{
386 int ret = 0;
387
388 if (inc->i_rdma_cookie) {
389 ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RDMA_DEST,
390 sizeof(inc->i_rdma_cookie), &inc->i_rdma_cookie);
391 if (ret)
392 return ret;
393 }
394
395 return 0;
396}
397
398int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
399 size_t size, int msg_flags)
400{
401 struct sock *sk = sock->sk;
402 struct rds_sock *rs = rds_sk_to_rs(sk);
403 long timeo;
404 int ret = 0, nonblock = msg_flags & MSG_DONTWAIT;
405 struct sockaddr_in *sin;
406 struct rds_incoming *inc = NULL;
407
408 /* udp_recvmsg()->sock_recvtimeo() gets away without locking too.. */
409 timeo = sock_rcvtimeo(sk, nonblock);
410
411 rdsdebug("size %zu flags 0x%x timeo %ld\n", size, msg_flags, timeo);
412
413 if (msg_flags & MSG_OOB)
414 goto out;
415
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000416 while (1) {
Andy Groveredacaea2009-07-17 13:13:32 +0000417 /* If there are pending notifications, do those - and nothing else */
418 if (!list_empty(&rs->rs_notify_queue)) {
419 ret = rds_notify_queue_get(rs, msg);
420 break;
421 }
422
423 if (rs->rs_cong_notify) {
424 ret = rds_notify_cong(rs, msg);
425 break;
426 }
427
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000428 if (!rds_next_incoming(rs, &inc)) {
429 if (nonblock) {
430 ret = -EAGAIN;
431 break;
432 }
433
434 timeo = wait_event_interruptible_timeout(*sk->sk_sleep,
Joe Perchesf64f9e72009-11-29 16:55:45 -0800435 (!list_empty(&rs->rs_notify_queue) ||
436 rs->rs_cong_notify ||
437 rds_next_incoming(rs, &inc)), timeo);
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000438 rdsdebug("recvmsg woke inc %p timeo %ld\n", inc,
439 timeo);
440 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
441 continue;
442
443 ret = timeo;
444 if (ret == 0)
445 ret = -ETIMEDOUT;
446 break;
447 }
448
449 rdsdebug("copying inc %p from %pI4:%u to user\n", inc,
450 &inc->i_conn->c_faddr,
451 ntohs(inc->i_hdr.h_sport));
452 ret = inc->i_conn->c_trans->inc_copy_to_user(inc, msg->msg_iov,
453 size);
454 if (ret < 0)
455 break;
456
457 /*
458 * if the message we just copied isn't at the head of the
459 * recv queue then someone else raced us to return it, try
460 * to get the next message.
461 */
462 if (!rds_still_queued(rs, inc, !(msg_flags & MSG_PEEK))) {
463 rds_inc_put(inc);
464 inc = NULL;
465 rds_stats_inc(s_recv_deliver_raced);
466 continue;
467 }
468
469 if (ret < be32_to_cpu(inc->i_hdr.h_len)) {
470 if (msg_flags & MSG_TRUNC)
471 ret = be32_to_cpu(inc->i_hdr.h_len);
472 msg->msg_flags |= MSG_TRUNC;
473 }
474
475 if (rds_cmsg_recv(inc, msg)) {
476 ret = -EFAULT;
477 goto out;
478 }
479
480 rds_stats_inc(s_recv_delivered);
481
482 sin = (struct sockaddr_in *)msg->msg_name;
483 if (sin) {
484 sin->sin_family = AF_INET;
485 sin->sin_port = inc->i_hdr.h_sport;
486 sin->sin_addr.s_addr = inc->i_saddr;
487 memset(sin->sin_zero, 0, sizeof(sin->sin_zero));
488 }
489 break;
490 }
491
492 if (inc)
493 rds_inc_put(inc);
494
495out:
496 return ret;
497}
498
499/*
500 * The socket is being shut down and we're asked to drop messages that were
501 * queued for recvmsg. The caller has unbound the socket so the receive path
502 * won't queue any more incoming fragments or messages on the socket.
503 */
504void rds_clear_recv_queue(struct rds_sock *rs)
505{
506 struct sock *sk = rds_rs_to_sk(rs);
507 struct rds_incoming *inc, *tmp;
508 unsigned long flags;
509
510 write_lock_irqsave(&rs->rs_recv_lock, flags);
511 list_for_each_entry_safe(inc, tmp, &rs->rs_recv_queue, i_item) {
512 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
513 -be32_to_cpu(inc->i_hdr.h_len),
514 inc->i_hdr.h_dport);
515 list_del_init(&inc->i_item);
516 rds_inc_put(inc);
517 }
518 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
519}
520
521/*
522 * inc->i_saddr isn't used here because it is only set in the receive
523 * path.
524 */
525void rds_inc_info_copy(struct rds_incoming *inc,
526 struct rds_info_iterator *iter,
527 __be32 saddr, __be32 daddr, int flip)
528{
529 struct rds_info_message minfo;
530
531 minfo.seq = be64_to_cpu(inc->i_hdr.h_sequence);
532 minfo.len = be32_to_cpu(inc->i_hdr.h_len);
533
534 if (flip) {
535 minfo.laddr = daddr;
536 minfo.faddr = saddr;
537 minfo.lport = inc->i_hdr.h_dport;
538 minfo.fport = inc->i_hdr.h_sport;
539 } else {
540 minfo.laddr = saddr;
541 minfo.faddr = daddr;
542 minfo.lport = inc->i_hdr.h_sport;
543 minfo.fport = inc->i_hdr.h_dport;
544 }
545
546 rds_info_copy(iter, &minfo, sizeof(minfo));
547}