blob: f2118c51cfa3b794c13177870e8966b600d667cf [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}
49
50void rds_inc_addref(struct rds_incoming *inc)
51{
52 rdsdebug("addref inc %p ref %d\n", inc, atomic_read(&inc->i_refcount));
53 atomic_inc(&inc->i_refcount);
54}
55
56void rds_inc_put(struct rds_incoming *inc)
57{
58 rdsdebug("put inc %p ref %d\n", inc, atomic_read(&inc->i_refcount));
59 if (atomic_dec_and_test(&inc->i_refcount)) {
60 BUG_ON(!list_empty(&inc->i_item));
61
62 inc->i_conn->c_trans->inc_free(inc);
63 }
64}
65
66static void rds_recv_rcvbuf_delta(struct rds_sock *rs, struct sock *sk,
67 struct rds_cong_map *map,
68 int delta, __be16 port)
69{
70 int now_congested;
71
72 if (delta == 0)
73 return;
74
75 rs->rs_rcv_bytes += delta;
76 now_congested = rs->rs_rcv_bytes > rds_sk_rcvbuf(rs);
77
78 rdsdebug("rs %p (%pI4:%u) recv bytes %d buf %d "
79 "now_cong %d delta %d\n",
80 rs, &rs->rs_bound_addr,
81 ntohs(rs->rs_bound_port), rs->rs_rcv_bytes,
82 rds_sk_rcvbuf(rs), now_congested, delta);
83
84 /* wasn't -> am congested */
85 if (!rs->rs_congested && now_congested) {
86 rs->rs_congested = 1;
87 rds_cong_set_bit(map, port);
88 rds_cong_queue_updates(map);
89 }
90 /* was -> aren't congested */
91 /* Require more free space before reporting uncongested to prevent
92 bouncing cong/uncong state too often */
93 else if (rs->rs_congested && (rs->rs_rcv_bytes < (rds_sk_rcvbuf(rs)/2))) {
94 rs->rs_congested = 0;
95 rds_cong_clear_bit(map, port);
96 rds_cong_queue_updates(map);
97 }
98
99 /* do nothing if no change in cong state */
100}
101
102/*
103 * Process all extension headers that come with this message.
104 */
105static void rds_recv_incoming_exthdrs(struct rds_incoming *inc, struct rds_sock *rs)
106{
107 struct rds_header *hdr = &inc->i_hdr;
108 unsigned int pos = 0, type, len;
109 union {
110 struct rds_ext_header_version version;
111 struct rds_ext_header_rdma rdma;
112 struct rds_ext_header_rdma_dest rdma_dest;
113 } buffer;
114
115 while (1) {
116 len = sizeof(buffer);
117 type = rds_message_next_extension(hdr, &pos, &buffer, &len);
118 if (type == RDS_EXTHDR_NONE)
119 break;
120 /* Process extension header here */
121 switch (type) {
122 case RDS_EXTHDR_RDMA:
123 rds_rdma_unuse(rs, be32_to_cpu(buffer.rdma.h_rdma_rkey), 0);
124 break;
125
126 case RDS_EXTHDR_RDMA_DEST:
127 /* We ignore the size for now. We could stash it
128 * somewhere and use it for error checking. */
129 inc->i_rdma_cookie = rds_rdma_make_cookie(
130 be32_to_cpu(buffer.rdma_dest.h_rdma_rkey),
131 be32_to_cpu(buffer.rdma_dest.h_rdma_offset));
132
133 break;
134 }
135 }
136}
137
138/*
139 * The transport must make sure that this is serialized against other
140 * rx and conn reset on this specific conn.
141 *
142 * We currently assert that only one fragmented message will be sent
143 * down a connection at a time. This lets us reassemble in the conn
144 * instead of per-flow which means that we don't have to go digging through
145 * flows to tear down partial reassembly progress on conn failure and
146 * we save flow lookup and locking for each frag arrival. It does mean
147 * that small messages will wait behind large ones. Fragmenting at all
148 * is only to reduce the memory consumption of pre-posted buffers.
149 *
150 * The caller passes in saddr and daddr instead of us getting it from the
151 * conn. This lets loopback, who only has one conn for both directions,
152 * tell us which roles the addrs in the conn are playing for this message.
153 */
154void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
155 struct rds_incoming *inc, gfp_t gfp, enum km_type km)
156{
157 struct rds_sock *rs = NULL;
158 struct sock *sk;
159 unsigned long flags;
160
161 inc->i_conn = conn;
162 inc->i_rx_jiffies = jiffies;
163
164 rdsdebug("conn %p next %llu inc %p seq %llu len %u sport %u dport %u "
165 "flags 0x%x rx_jiffies %lu\n", conn,
166 (unsigned long long)conn->c_next_rx_seq,
167 inc,
168 (unsigned long long)be64_to_cpu(inc->i_hdr.h_sequence),
169 be32_to_cpu(inc->i_hdr.h_len),
170 be16_to_cpu(inc->i_hdr.h_sport),
171 be16_to_cpu(inc->i_hdr.h_dport),
172 inc->i_hdr.h_flags,
173 inc->i_rx_jiffies);
174
175 /*
176 * Sequence numbers should only increase. Messages get their
177 * sequence number as they're queued in a sending conn. They
178 * can be dropped, though, if the sending socket is closed before
179 * they hit the wire. So sequence numbers can skip forward
180 * under normal operation. They can also drop back in the conn
181 * failover case as previously sent messages are resent down the
182 * new instance of a conn. We drop those, otherwise we have
183 * to assume that the next valid seq does not come after a
184 * hole in the fragment stream.
185 *
186 * The headers don't give us a way to realize if fragments of
187 * a message have been dropped. We assume that frags that arrive
188 * to a flow are part of the current message on the flow that is
189 * being reassembled. This means that senders can't drop messages
190 * from the sending conn until all their frags are sent.
191 *
192 * XXX we could spend more on the wire to get more robust failure
193 * detection, arguably worth it to avoid data corruption.
194 */
195 if (be64_to_cpu(inc->i_hdr.h_sequence) < conn->c_next_rx_seq
196 && (inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) {
197 rds_stats_inc(s_recv_drop_old_seq);
198 goto out;
199 }
200 conn->c_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1;
201
202 if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) {
203 rds_stats_inc(s_recv_ping);
204 rds_send_pong(conn, inc->i_hdr.h_sport);
205 goto out;
206 }
207
208 rs = rds_find_bound(daddr, inc->i_hdr.h_dport);
209 if (rs == NULL) {
210 rds_stats_inc(s_recv_drop_no_sock);
211 goto out;
212 }
213
214 /* Process extension headers */
215 rds_recv_incoming_exthdrs(inc, rs);
216
217 /* We can be racing with rds_release() which marks the socket dead. */
218 sk = rds_rs_to_sk(rs);
219
220 /* serialize with rds_release -> sock_orphan */
221 write_lock_irqsave(&rs->rs_recv_lock, flags);
222 if (!sock_flag(sk, SOCK_DEAD)) {
223 rdsdebug("adding inc %p to rs %p's recv queue\n", inc, rs);
224 rds_stats_inc(s_recv_queued);
225 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
226 be32_to_cpu(inc->i_hdr.h_len),
227 inc->i_hdr.h_dport);
228 rds_inc_addref(inc);
229 list_add_tail(&inc->i_item, &rs->rs_recv_queue);
230 __rds_wake_sk_sleep(sk);
231 } else {
232 rds_stats_inc(s_recv_drop_dead_sock);
233 }
234 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
235
236out:
237 if (rs)
238 rds_sock_put(rs);
239}
240
241/*
242 * be very careful here. This is being called as the condition in
243 * wait_event_*() needs to cope with being called many times.
244 */
245static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc)
246{
247 unsigned long flags;
248
249 if (*inc == NULL) {
250 read_lock_irqsave(&rs->rs_recv_lock, flags);
251 if (!list_empty(&rs->rs_recv_queue)) {
252 *inc = list_entry(rs->rs_recv_queue.next,
253 struct rds_incoming,
254 i_item);
255 rds_inc_addref(*inc);
256 }
257 read_unlock_irqrestore(&rs->rs_recv_lock, flags);
258 }
259
260 return *inc != NULL;
261}
262
263static int rds_still_queued(struct rds_sock *rs, struct rds_incoming *inc,
264 int drop)
265{
266 struct sock *sk = rds_rs_to_sk(rs);
267 int ret = 0;
268 unsigned long flags;
269
270 write_lock_irqsave(&rs->rs_recv_lock, flags);
271 if (!list_empty(&inc->i_item)) {
272 ret = 1;
273 if (drop) {
274 /* XXX make sure this i_conn is reliable */
275 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
276 -be32_to_cpu(inc->i_hdr.h_len),
277 inc->i_hdr.h_dport);
278 list_del_init(&inc->i_item);
279 rds_inc_put(inc);
280 }
281 }
282 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
283
284 rdsdebug("inc %p rs %p still %d dropped %d\n", inc, rs, ret, drop);
285 return ret;
286}
287
288/*
289 * Pull errors off the error queue.
290 * If msghdr is NULL, we will just purge the error queue.
291 */
292int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr)
293{
294 struct rds_notifier *notifier;
295 struct rds_rdma_notify cmsg;
296 unsigned int count = 0, max_messages = ~0U;
297 unsigned long flags;
298 LIST_HEAD(copy);
299 int err = 0;
300
301
302 /* put_cmsg copies to user space and thus may sleep. We can't do this
303 * with rs_lock held, so first grab as many notifications as we can stuff
304 * in the user provided cmsg buffer. We don't try to copy more, to avoid
305 * losing notifications - except when the buffer is so small that it wouldn't
306 * even hold a single notification. Then we give him as much of this single
307 * msg as we can squeeze in, and set MSG_CTRUNC.
308 */
309 if (msghdr) {
310 max_messages = msghdr->msg_controllen / CMSG_SPACE(sizeof(cmsg));
311 if (!max_messages)
312 max_messages = 1;
313 }
314
315 spin_lock_irqsave(&rs->rs_lock, flags);
316 while (!list_empty(&rs->rs_notify_queue) && count < max_messages) {
317 notifier = list_entry(rs->rs_notify_queue.next,
318 struct rds_notifier, n_list);
319 list_move(&notifier->n_list, &copy);
320 count++;
321 }
322 spin_unlock_irqrestore(&rs->rs_lock, flags);
323
324 if (!count)
325 return 0;
326
327 while (!list_empty(&copy)) {
328 notifier = list_entry(copy.next, struct rds_notifier, n_list);
329
330 if (msghdr) {
331 cmsg.user_token = notifier->n_user_token;
332 cmsg.status = notifier->n_status;
333
334 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS,
335 sizeof(cmsg), &cmsg);
336 if (err)
337 break;
338 }
339
340 list_del_init(&notifier->n_list);
341 kfree(notifier);
342 }
343
344 /* If we bailed out because of an error in put_cmsg,
345 * we may be left with one or more notifications that we
346 * didn't process. Return them to the head of the list. */
347 if (!list_empty(&copy)) {
348 spin_lock_irqsave(&rs->rs_lock, flags);
349 list_splice(&copy, &rs->rs_notify_queue);
350 spin_unlock_irqrestore(&rs->rs_lock, flags);
351 }
352
353 return err;
354}
355
356/*
357 * Queue a congestion notification
358 */
359static int rds_notify_cong(struct rds_sock *rs, struct msghdr *msghdr)
360{
361 uint64_t notify = rs->rs_cong_notify;
362 unsigned long flags;
363 int err;
364
365 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_CONG_UPDATE,
366 sizeof(notify), &notify);
367 if (err)
368 return err;
369
370 spin_lock_irqsave(&rs->rs_lock, flags);
371 rs->rs_cong_notify &= ~notify;
372 spin_unlock_irqrestore(&rs->rs_lock, flags);
373
374 return 0;
375}
376
377/*
378 * Receive any control messages.
379 */
380static int rds_cmsg_recv(struct rds_incoming *inc, struct msghdr *msg)
381{
382 int ret = 0;
383
384 if (inc->i_rdma_cookie) {
385 ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RDMA_DEST,
386 sizeof(inc->i_rdma_cookie), &inc->i_rdma_cookie);
387 if (ret)
388 return ret;
389 }
390
391 return 0;
392}
393
394int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
395 size_t size, int msg_flags)
396{
397 struct sock *sk = sock->sk;
398 struct rds_sock *rs = rds_sk_to_rs(sk);
399 long timeo;
400 int ret = 0, nonblock = msg_flags & MSG_DONTWAIT;
401 struct sockaddr_in *sin;
402 struct rds_incoming *inc = NULL;
403
404 /* udp_recvmsg()->sock_recvtimeo() gets away without locking too.. */
405 timeo = sock_rcvtimeo(sk, nonblock);
406
407 rdsdebug("size %zu flags 0x%x timeo %ld\n", size, msg_flags, timeo);
408
409 if (msg_flags & MSG_OOB)
410 goto out;
411
412 /* If there are pending notifications, do those - and nothing else */
413 if (!list_empty(&rs->rs_notify_queue)) {
414 ret = rds_notify_queue_get(rs, msg);
415 goto out;
416 }
417
418 if (rs->rs_cong_notify) {
419 ret = rds_notify_cong(rs, msg);
420 goto out;
421 }
422
423 while (1) {
424 if (!rds_next_incoming(rs, &inc)) {
425 if (nonblock) {
426 ret = -EAGAIN;
427 break;
428 }
429
430 timeo = wait_event_interruptible_timeout(*sk->sk_sleep,
431 rds_next_incoming(rs, &inc),
432 timeo);
433 rdsdebug("recvmsg woke inc %p timeo %ld\n", inc,
434 timeo);
435 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
436 continue;
437
438 ret = timeo;
439 if (ret == 0)
440 ret = -ETIMEDOUT;
441 break;
442 }
443
444 rdsdebug("copying inc %p from %pI4:%u to user\n", inc,
445 &inc->i_conn->c_faddr,
446 ntohs(inc->i_hdr.h_sport));
447 ret = inc->i_conn->c_trans->inc_copy_to_user(inc, msg->msg_iov,
448 size);
449 if (ret < 0)
450 break;
451
452 /*
453 * if the message we just copied isn't at the head of the
454 * recv queue then someone else raced us to return it, try
455 * to get the next message.
456 */
457 if (!rds_still_queued(rs, inc, !(msg_flags & MSG_PEEK))) {
458 rds_inc_put(inc);
459 inc = NULL;
460 rds_stats_inc(s_recv_deliver_raced);
461 continue;
462 }
463
464 if (ret < be32_to_cpu(inc->i_hdr.h_len)) {
465 if (msg_flags & MSG_TRUNC)
466 ret = be32_to_cpu(inc->i_hdr.h_len);
467 msg->msg_flags |= MSG_TRUNC;
468 }
469
470 if (rds_cmsg_recv(inc, msg)) {
471 ret = -EFAULT;
472 goto out;
473 }
474
475 rds_stats_inc(s_recv_delivered);
476
477 sin = (struct sockaddr_in *)msg->msg_name;
478 if (sin) {
479 sin->sin_family = AF_INET;
480 sin->sin_port = inc->i_hdr.h_sport;
481 sin->sin_addr.s_addr = inc->i_saddr;
482 memset(sin->sin_zero, 0, sizeof(sin->sin_zero));
483 }
484 break;
485 }
486
487 if (inc)
488 rds_inc_put(inc);
489
490out:
491 return ret;
492}
493
494/*
495 * The socket is being shut down and we're asked to drop messages that were
496 * queued for recvmsg. The caller has unbound the socket so the receive path
497 * won't queue any more incoming fragments or messages on the socket.
498 */
499void rds_clear_recv_queue(struct rds_sock *rs)
500{
501 struct sock *sk = rds_rs_to_sk(rs);
502 struct rds_incoming *inc, *tmp;
503 unsigned long flags;
504
505 write_lock_irqsave(&rs->rs_recv_lock, flags);
506 list_for_each_entry_safe(inc, tmp, &rs->rs_recv_queue, i_item) {
507 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
508 -be32_to_cpu(inc->i_hdr.h_len),
509 inc->i_hdr.h_dport);
510 list_del_init(&inc->i_item);
511 rds_inc_put(inc);
512 }
513 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
514}
515
516/*
517 * inc->i_saddr isn't used here because it is only set in the receive
518 * path.
519 */
520void rds_inc_info_copy(struct rds_incoming *inc,
521 struct rds_info_iterator *iter,
522 __be32 saddr, __be32 daddr, int flip)
523{
524 struct rds_info_message minfo;
525
526 minfo.seq = be64_to_cpu(inc->i_hdr.h_sequence);
527 minfo.len = be32_to_cpu(inc->i_hdr.h_len);
528
529 if (flip) {
530 minfo.laddr = daddr;
531 minfo.faddr = saddr;
532 minfo.lport = inc->i_hdr.h_dport;
533 minfo.fport = inc->i_hdr.h_sport;
534 } else {
535 minfo.laddr = saddr;
536 minfo.faddr = daddr;
537 minfo.lport = inc->i_hdr.h_sport;
538 minfo.fport = inc->i_hdr.h_dport;
539 }
540
541 rds_info_copy(iter, &minfo, sizeof(minfo));
542}