blob: 596689e592727f93eaf749490cf6fb6af4956463 [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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090034#include <linux/slab.h>
Andy Groverbdbe6fb2009-02-24 15:30:28 +000035#include <net/sock.h>
36#include <linux/in.h>
37
38#include "rds.h"
Andy Groverbdbe6fb2009-02-24 15:30:28 +000039
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
stephen hemmingerff51bf82010-10-19 08:08:33 +000051static void rds_inc_addref(struct rds_incoming *inc)
Andy Groverbdbe6fb2009-02-24 15:30:28 +000052{
53 rdsdebug("addref inc %p ref %d\n", inc, atomic_read(&inc->i_refcount));
54 atomic_inc(&inc->i_refcount);
55}
56
57void rds_inc_put(struct rds_incoming *inc)
58{
59 rdsdebug("put inc %p ref %d\n", inc, atomic_read(&inc->i_refcount));
60 if (atomic_dec_and_test(&inc->i_refcount)) {
61 BUG_ON(!list_empty(&inc->i_item));
62
63 inc->i_conn->c_trans->inc_free(inc);
64 }
65}
Andy Grover616b7572009-08-21 12:28:32 +000066EXPORT_SYMBOL_GPL(rds_inc_put);
Andy Groverbdbe6fb2009-02-24 15:30:28 +000067
68static void rds_recv_rcvbuf_delta(struct rds_sock *rs, struct sock *sk,
69 struct rds_cong_map *map,
70 int delta, __be16 port)
71{
72 int now_congested;
73
74 if (delta == 0)
75 return;
76
77 rs->rs_rcv_bytes += delta;
78 now_congested = rs->rs_rcv_bytes > rds_sk_rcvbuf(rs);
79
80 rdsdebug("rs %p (%pI4:%u) recv bytes %d buf %d "
81 "now_cong %d delta %d\n",
82 rs, &rs->rs_bound_addr,
83 ntohs(rs->rs_bound_port), rs->rs_rcv_bytes,
84 rds_sk_rcvbuf(rs), now_congested, delta);
85
86 /* wasn't -> am congested */
87 if (!rs->rs_congested && now_congested) {
88 rs->rs_congested = 1;
89 rds_cong_set_bit(map, port);
90 rds_cong_queue_updates(map);
91 }
92 /* was -> aren't congested */
93 /* Require more free space before reporting uncongested to prevent
94 bouncing cong/uncong state too often */
95 else if (rs->rs_congested && (rs->rs_rcv_bytes < (rds_sk_rcvbuf(rs)/2))) {
96 rs->rs_congested = 0;
97 rds_cong_clear_bit(map, port);
98 rds_cong_queue_updates(map);
99 }
100
101 /* do nothing if no change in cong state */
102}
103
104/*
105 * Process all extension headers that come with this message.
106 */
107static void rds_recv_incoming_exthdrs(struct rds_incoming *inc, struct rds_sock *rs)
108{
109 struct rds_header *hdr = &inc->i_hdr;
110 unsigned int pos = 0, type, len;
111 union {
112 struct rds_ext_header_version version;
113 struct rds_ext_header_rdma rdma;
114 struct rds_ext_header_rdma_dest rdma_dest;
115 } buffer;
116
117 while (1) {
118 len = sizeof(buffer);
119 type = rds_message_next_extension(hdr, &pos, &buffer, &len);
120 if (type == RDS_EXTHDR_NONE)
121 break;
122 /* Process extension header here */
123 switch (type) {
124 case RDS_EXTHDR_RDMA:
125 rds_rdma_unuse(rs, be32_to_cpu(buffer.rdma.h_rdma_rkey), 0);
126 break;
127
128 case RDS_EXTHDR_RDMA_DEST:
129 /* We ignore the size for now. We could stash it
130 * somewhere and use it for error checking. */
131 inc->i_rdma_cookie = rds_rdma_make_cookie(
132 be32_to_cpu(buffer.rdma_dest.h_rdma_rkey),
133 be32_to_cpu(buffer.rdma_dest.h_rdma_offset));
134
135 break;
136 }
137 }
138}
139
140/*
141 * The transport must make sure that this is serialized against other
142 * rx and conn reset on this specific conn.
143 *
144 * We currently assert that only one fragmented message will be sent
145 * down a connection at a time. This lets us reassemble in the conn
146 * instead of per-flow which means that we don't have to go digging through
147 * flows to tear down partial reassembly progress on conn failure and
148 * we save flow lookup and locking for each frag arrival. It does mean
149 * that small messages will wait behind large ones. Fragmenting at all
150 * is only to reduce the memory consumption of pre-posted buffers.
151 *
152 * The caller passes in saddr and daddr instead of us getting it from the
153 * conn. This lets loopback, who only has one conn for both directions,
154 * tell us which roles the addrs in the conn are playing for this message.
155 */
156void rds_recv_incoming(struct rds_connection *conn, __be32 saddr, __be32 daddr,
157 struct rds_incoming *inc, gfp_t gfp, enum km_type km)
158{
159 struct rds_sock *rs = NULL;
160 struct sock *sk;
161 unsigned long flags;
162
163 inc->i_conn = conn;
164 inc->i_rx_jiffies = jiffies;
165
166 rdsdebug("conn %p next %llu inc %p seq %llu len %u sport %u dport %u "
167 "flags 0x%x rx_jiffies %lu\n", conn,
168 (unsigned long long)conn->c_next_rx_seq,
169 inc,
170 (unsigned long long)be64_to_cpu(inc->i_hdr.h_sequence),
171 be32_to_cpu(inc->i_hdr.h_len),
172 be16_to_cpu(inc->i_hdr.h_sport),
173 be16_to_cpu(inc->i_hdr.h_dport),
174 inc->i_hdr.h_flags,
175 inc->i_rx_jiffies);
176
177 /*
178 * Sequence numbers should only increase. Messages get their
179 * sequence number as they're queued in a sending conn. They
180 * can be dropped, though, if the sending socket is closed before
181 * they hit the wire. So sequence numbers can skip forward
182 * under normal operation. They can also drop back in the conn
183 * failover case as previously sent messages are resent down the
184 * new instance of a conn. We drop those, otherwise we have
185 * to assume that the next valid seq does not come after a
186 * hole in the fragment stream.
187 *
188 * The headers don't give us a way to realize if fragments of
189 * a message have been dropped. We assume that frags that arrive
190 * to a flow are part of the current message on the flow that is
191 * being reassembled. This means that senders can't drop messages
192 * from the sending conn until all their frags are sent.
193 *
194 * XXX we could spend more on the wire to get more robust failure
195 * detection, arguably worth it to avoid data corruption.
196 */
Joe Perchesf64f9e72009-11-29 16:55:45 -0800197 if (be64_to_cpu(inc->i_hdr.h_sequence) < conn->c_next_rx_seq &&
198 (inc->i_hdr.h_flags & RDS_FLAG_RETRANSMITTED)) {
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000199 rds_stats_inc(s_recv_drop_old_seq);
200 goto out;
201 }
202 conn->c_next_rx_seq = be64_to_cpu(inc->i_hdr.h_sequence) + 1;
203
204 if (rds_sysctl_ping_enable && inc->i_hdr.h_dport == 0) {
205 rds_stats_inc(s_recv_ping);
206 rds_send_pong(conn, inc->i_hdr.h_sport);
207 goto out;
208 }
209
210 rs = rds_find_bound(daddr, inc->i_hdr.h_dport);
Andy Grover8690bfa2010-01-12 11:56:44 -0800211 if (!rs) {
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000212 rds_stats_inc(s_recv_drop_no_sock);
213 goto out;
214 }
215
216 /* Process extension headers */
217 rds_recv_incoming_exthdrs(inc, rs);
218
219 /* We can be racing with rds_release() which marks the socket dead. */
220 sk = rds_rs_to_sk(rs);
221
222 /* serialize with rds_release -> sock_orphan */
223 write_lock_irqsave(&rs->rs_recv_lock, flags);
224 if (!sock_flag(sk, SOCK_DEAD)) {
225 rdsdebug("adding inc %p to rs %p's recv queue\n", inc, rs);
226 rds_stats_inc(s_recv_queued);
227 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
228 be32_to_cpu(inc->i_hdr.h_len),
229 inc->i_hdr.h_dport);
230 rds_inc_addref(inc);
231 list_add_tail(&inc->i_item, &rs->rs_recv_queue);
232 __rds_wake_sk_sleep(sk);
233 } else {
234 rds_stats_inc(s_recv_drop_dead_sock);
235 }
236 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
237
238out:
239 if (rs)
240 rds_sock_put(rs);
241}
Andy Grover616b7572009-08-21 12:28:32 +0000242EXPORT_SYMBOL_GPL(rds_recv_incoming);
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000243
244/*
245 * be very careful here. This is being called as the condition in
246 * wait_event_*() needs to cope with being called many times.
247 */
248static int rds_next_incoming(struct rds_sock *rs, struct rds_incoming **inc)
249{
250 unsigned long flags;
251
Andy Grover8690bfa2010-01-12 11:56:44 -0800252 if (!*inc) {
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000253 read_lock_irqsave(&rs->rs_recv_lock, flags);
254 if (!list_empty(&rs->rs_recv_queue)) {
255 *inc = list_entry(rs->rs_recv_queue.next,
256 struct rds_incoming,
257 i_item);
258 rds_inc_addref(*inc);
259 }
260 read_unlock_irqrestore(&rs->rs_recv_lock, flags);
261 }
262
263 return *inc != NULL;
264}
265
266static int rds_still_queued(struct rds_sock *rs, struct rds_incoming *inc,
267 int drop)
268{
269 struct sock *sk = rds_rs_to_sk(rs);
270 int ret = 0;
271 unsigned long flags;
272
273 write_lock_irqsave(&rs->rs_recv_lock, flags);
274 if (!list_empty(&inc->i_item)) {
275 ret = 1;
276 if (drop) {
277 /* XXX make sure this i_conn is reliable */
278 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
279 -be32_to_cpu(inc->i_hdr.h_len),
280 inc->i_hdr.h_dport);
281 list_del_init(&inc->i_item);
282 rds_inc_put(inc);
283 }
284 }
285 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
286
287 rdsdebug("inc %p rs %p still %d dropped %d\n", inc, rs, ret, drop);
288 return ret;
289}
290
291/*
292 * Pull errors off the error queue.
293 * If msghdr is NULL, we will just purge the error queue.
294 */
295int rds_notify_queue_get(struct rds_sock *rs, struct msghdr *msghdr)
296{
297 struct rds_notifier *notifier;
Eric Dumazetf0375902010-08-16 03:25:00 +0000298 struct rds_rdma_notify cmsg = { 0 }; /* fill holes with zero */
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000299 unsigned int count = 0, max_messages = ~0U;
300 unsigned long flags;
301 LIST_HEAD(copy);
302 int err = 0;
303
304
305 /* put_cmsg copies to user space and thus may sleep. We can't do this
306 * with rs_lock held, so first grab as many notifications as we can stuff
307 * in the user provided cmsg buffer. We don't try to copy more, to avoid
308 * losing notifications - except when the buffer is so small that it wouldn't
309 * even hold a single notification. Then we give him as much of this single
310 * msg as we can squeeze in, and set MSG_CTRUNC.
311 */
312 if (msghdr) {
313 max_messages = msghdr->msg_controllen / CMSG_SPACE(sizeof(cmsg));
314 if (!max_messages)
315 max_messages = 1;
316 }
317
318 spin_lock_irqsave(&rs->rs_lock, flags);
319 while (!list_empty(&rs->rs_notify_queue) && count < max_messages) {
320 notifier = list_entry(rs->rs_notify_queue.next,
321 struct rds_notifier, n_list);
322 list_move(&notifier->n_list, &copy);
323 count++;
324 }
325 spin_unlock_irqrestore(&rs->rs_lock, flags);
326
327 if (!count)
328 return 0;
329
330 while (!list_empty(&copy)) {
331 notifier = list_entry(copy.next, struct rds_notifier, n_list);
332
333 if (msghdr) {
334 cmsg.user_token = notifier->n_user_token;
Andy Grover6200ed72010-01-12 10:53:05 -0800335 cmsg.status = notifier->n_status;
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000336
337 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_RDMA_STATUS,
Andy Grover6200ed72010-01-12 10:53:05 -0800338 sizeof(cmsg), &cmsg);
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000339 if (err)
340 break;
341 }
342
343 list_del_init(&notifier->n_list);
344 kfree(notifier);
345 }
346
347 /* If we bailed out because of an error in put_cmsg,
348 * we may be left with one or more notifications that we
349 * didn't process. Return them to the head of the list. */
350 if (!list_empty(&copy)) {
351 spin_lock_irqsave(&rs->rs_lock, flags);
352 list_splice(&copy, &rs->rs_notify_queue);
353 spin_unlock_irqrestore(&rs->rs_lock, flags);
354 }
355
356 return err;
357}
358
359/*
360 * Queue a congestion notification
361 */
362static int rds_notify_cong(struct rds_sock *rs, struct msghdr *msghdr)
363{
364 uint64_t notify = rs->rs_cong_notify;
365 unsigned long flags;
366 int err;
367
368 err = put_cmsg(msghdr, SOL_RDS, RDS_CMSG_CONG_UPDATE,
369 sizeof(notify), &notify);
370 if (err)
371 return err;
372
373 spin_lock_irqsave(&rs->rs_lock, flags);
374 rs->rs_cong_notify &= ~notify;
375 spin_unlock_irqrestore(&rs->rs_lock, flags);
376
377 return 0;
378}
379
380/*
381 * Receive any control messages.
382 */
383static int rds_cmsg_recv(struct rds_incoming *inc, struct msghdr *msg)
384{
385 int ret = 0;
386
387 if (inc->i_rdma_cookie) {
388 ret = put_cmsg(msg, SOL_RDS, RDS_CMSG_RDMA_DEST,
389 sizeof(inc->i_rdma_cookie), &inc->i_rdma_cookie);
390 if (ret)
391 return ret;
392 }
393
394 return 0;
395}
396
397int rds_recvmsg(struct kiocb *iocb, struct socket *sock, struct msghdr *msg,
398 size_t size, int msg_flags)
399{
400 struct sock *sk = sock->sk;
401 struct rds_sock *rs = rds_sk_to_rs(sk);
402 long timeo;
403 int ret = 0, nonblock = msg_flags & MSG_DONTWAIT;
404 struct sockaddr_in *sin;
405 struct rds_incoming *inc = NULL;
406
407 /* udp_recvmsg()->sock_recvtimeo() gets away without locking too.. */
408 timeo = sock_rcvtimeo(sk, nonblock);
409
410 rdsdebug("size %zu flags 0x%x timeo %ld\n", size, msg_flags, timeo);
411
412 if (msg_flags & MSG_OOB)
413 goto out;
414
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000415 while (1) {
Andy Groveredacaea2009-07-17 13:13:32 +0000416 /* If there are pending notifications, do those - and nothing else */
417 if (!list_empty(&rs->rs_notify_queue)) {
418 ret = rds_notify_queue_get(rs, msg);
419 break;
420 }
421
422 if (rs->rs_cong_notify) {
423 ret = rds_notify_cong(rs, msg);
424 break;
425 }
426
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000427 if (!rds_next_incoming(rs, &inc)) {
428 if (nonblock) {
429 ret = -EAGAIN;
430 break;
431 }
432
Eric Dumazetaa395142010-04-20 13:03:51 +0000433 timeo = wait_event_interruptible_timeout(*sk_sleep(sk),
Joe Perchesf64f9e72009-11-29 16:55:45 -0800434 (!list_empty(&rs->rs_notify_queue) ||
435 rs->rs_cong_notify ||
436 rds_next_incoming(rs, &inc)), timeo);
Andy Groverbdbe6fb2009-02-24 15:30:28 +0000437 rdsdebug("recvmsg woke inc %p timeo %ld\n", inc,
438 timeo);
439 if (timeo > 0 || timeo == MAX_SCHEDULE_TIMEOUT)
440 continue;
441
442 ret = timeo;
443 if (ret == 0)
444 ret = -ETIMEDOUT;
445 break;
446 }
447
448 rdsdebug("copying inc %p from %pI4:%u to user\n", inc,
449 &inc->i_conn->c_faddr,
450 ntohs(inc->i_hdr.h_sport));
451 ret = inc->i_conn->c_trans->inc_copy_to_user(inc, msg->msg_iov,
452 size);
453 if (ret < 0)
454 break;
455
456 /*
457 * if the message we just copied isn't at the head of the
458 * recv queue then someone else raced us to return it, try
459 * to get the next message.
460 */
461 if (!rds_still_queued(rs, inc, !(msg_flags & MSG_PEEK))) {
462 rds_inc_put(inc);
463 inc = NULL;
464 rds_stats_inc(s_recv_deliver_raced);
465 continue;
466 }
467
468 if (ret < be32_to_cpu(inc->i_hdr.h_len)) {
469 if (msg_flags & MSG_TRUNC)
470 ret = be32_to_cpu(inc->i_hdr.h_len);
471 msg->msg_flags |= MSG_TRUNC;
472 }
473
474 if (rds_cmsg_recv(inc, msg)) {
475 ret = -EFAULT;
476 goto out;
477 }
478
479 rds_stats_inc(s_recv_delivered);
480
481 sin = (struct sockaddr_in *)msg->msg_name;
482 if (sin) {
483 sin->sin_family = AF_INET;
484 sin->sin_port = inc->i_hdr.h_sport;
485 sin->sin_addr.s_addr = inc->i_saddr;
486 memset(sin->sin_zero, 0, sizeof(sin->sin_zero));
487 }
488 break;
489 }
490
491 if (inc)
492 rds_inc_put(inc);
493
494out:
495 return ret;
496}
497
498/*
499 * The socket is being shut down and we're asked to drop messages that were
500 * queued for recvmsg. The caller has unbound the socket so the receive path
501 * won't queue any more incoming fragments or messages on the socket.
502 */
503void rds_clear_recv_queue(struct rds_sock *rs)
504{
505 struct sock *sk = rds_rs_to_sk(rs);
506 struct rds_incoming *inc, *tmp;
507 unsigned long flags;
508
509 write_lock_irqsave(&rs->rs_recv_lock, flags);
510 list_for_each_entry_safe(inc, tmp, &rs->rs_recv_queue, i_item) {
511 rds_recv_rcvbuf_delta(rs, sk, inc->i_conn->c_lcong,
512 -be32_to_cpu(inc->i_hdr.h_len),
513 inc->i_hdr.h_dport);
514 list_del_init(&inc->i_item);
515 rds_inc_put(inc);
516 }
517 write_unlock_irqrestore(&rs->rs_recv_lock, flags);
518}
519
520/*
521 * inc->i_saddr isn't used here because it is only set in the receive
522 * path.
523 */
524void rds_inc_info_copy(struct rds_incoming *inc,
525 struct rds_info_iterator *iter,
526 __be32 saddr, __be32 daddr, int flip)
527{
528 struct rds_info_message minfo;
529
530 minfo.seq = be64_to_cpu(inc->i_hdr.h_sequence);
531 minfo.len = be32_to_cpu(inc->i_hdr.h_len);
532
533 if (flip) {
534 minfo.laddr = daddr;
535 minfo.faddr = saddr;
536 minfo.lport = inc->i_hdr.h_dport;
537 minfo.fport = inc->i_hdr.h_sport;
538 } else {
539 minfo.laddr = saddr;
540 minfo.faddr = daddr;
541 minfo.lport = inc->i_hdr.h_sport;
542 minfo.fport = inc->i_hdr.h_dport;
543 }
544
545 rds_info_copy(iter, &minfo, sizeof(minfo));
546}