xref: /linux/net/rds/af_rds.c (revision b85d45947951d23cb22d90caecf4c1eb81342c96)
1 /*
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/module.h>
34 #include <linux/errno.h>
35 #include <linux/kernel.h>
36 #include <linux/gfp.h>
37 #include <linux/in.h>
38 #include <linux/poll.h>
39 #include <net/sock.h>
40 
41 #include "rds.h"
42 
43 /* this is just used for stats gathering :/ */
44 static DEFINE_SPINLOCK(rds_sock_lock);
45 static unsigned long rds_sock_count;
46 static LIST_HEAD(rds_sock_list);
47 DECLARE_WAIT_QUEUE_HEAD(rds_poll_waitq);
48 
49 /*
50  * This is called as the final descriptor referencing this socket is closed.
51  * We have to unbind the socket so that another socket can be bound to the
52  * address it was using.
53  *
54  * We have to be careful about racing with the incoming path.  sock_orphan()
55  * sets SOCK_DEAD and we use that as an indicator to the rx path that new
56  * messages shouldn't be queued.
57  */
58 static int rds_release(struct socket *sock)
59 {
60 	struct sock *sk = sock->sk;
61 	struct rds_sock *rs;
62 
63 	if (!sk)
64 		goto out;
65 
66 	rs = rds_sk_to_rs(sk);
67 
68 	sock_orphan(sk);
69 	/* Note - rds_clear_recv_queue grabs rs_recv_lock, so
70 	 * that ensures the recv path has completed messing
71 	 * with the socket. */
72 	rds_clear_recv_queue(rs);
73 	rds_cong_remove_socket(rs);
74 
75 	/*
76 	 * the binding lookup hash uses rcu, we need to
77 	 * make sure we synchronize_rcu before we free our
78 	 * entry
79 	 */
80 	rds_remove_bound(rs);
81 	synchronize_rcu();
82 
83 	rds_send_drop_to(rs, NULL);
84 	rds_rdma_drop_keys(rs);
85 	rds_notify_queue_get(rs, NULL);
86 
87 	spin_lock_bh(&rds_sock_lock);
88 	list_del_init(&rs->rs_item);
89 	rds_sock_count--;
90 	spin_unlock_bh(&rds_sock_lock);
91 
92 	rds_trans_put(rs->rs_transport);
93 
94 	sock->sk = NULL;
95 	sock_put(sk);
96 out:
97 	return 0;
98 }
99 
100 /*
101  * Careful not to race with rds_release -> sock_orphan which clears sk_sleep.
102  * _bh() isn't OK here, we're called from interrupt handlers.  It's probably OK
103  * to wake the waitqueue after sk_sleep is clear as we hold a sock ref, but
104  * this seems more conservative.
105  * NB - normally, one would use sk_callback_lock for this, but we can
106  * get here from interrupts, whereas the network code grabs sk_callback_lock
107  * with _lock_bh only - so relying on sk_callback_lock introduces livelocks.
108  */
109 void rds_wake_sk_sleep(struct rds_sock *rs)
110 {
111 	unsigned long flags;
112 
113 	read_lock_irqsave(&rs->rs_recv_lock, flags);
114 	__rds_wake_sk_sleep(rds_rs_to_sk(rs));
115 	read_unlock_irqrestore(&rs->rs_recv_lock, flags);
116 }
117 
118 static int rds_getname(struct socket *sock, struct sockaddr *uaddr,
119 		       int *uaddr_len, int peer)
120 {
121 	struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
122 	struct rds_sock *rs = rds_sk_to_rs(sock->sk);
123 
124 	memset(sin->sin_zero, 0, sizeof(sin->sin_zero));
125 
126 	/* racey, don't care */
127 	if (peer) {
128 		if (!rs->rs_conn_addr)
129 			return -ENOTCONN;
130 
131 		sin->sin_port = rs->rs_conn_port;
132 		sin->sin_addr.s_addr = rs->rs_conn_addr;
133 	} else {
134 		sin->sin_port = rs->rs_bound_port;
135 		sin->sin_addr.s_addr = rs->rs_bound_addr;
136 	}
137 
138 	sin->sin_family = AF_INET;
139 
140 	*uaddr_len = sizeof(*sin);
141 	return 0;
142 }
143 
144 /*
145  * RDS' poll is without a doubt the least intuitive part of the interface,
146  * as POLLIN and POLLOUT do not behave entirely as you would expect from
147  * a network protocol.
148  *
149  * POLLIN is asserted if
150  *  -	there is data on the receive queue.
151  *  -	to signal that a previously congested destination may have become
152  *	uncongested
153  *  -	A notification has been queued to the socket (this can be a congestion
154  *	update, or a RDMA completion).
155  *
156  * POLLOUT is asserted if there is room on the send queue. This does not mean
157  * however, that the next sendmsg() call will succeed. If the application tries
158  * to send to a congested destination, the system call may still fail (and
159  * return ENOBUFS).
160  */
161 static unsigned int rds_poll(struct file *file, struct socket *sock,
162 			     poll_table *wait)
163 {
164 	struct sock *sk = sock->sk;
165 	struct rds_sock *rs = rds_sk_to_rs(sk);
166 	unsigned int mask = 0;
167 	unsigned long flags;
168 
169 	poll_wait(file, sk_sleep(sk), wait);
170 
171 	if (rs->rs_seen_congestion)
172 		poll_wait(file, &rds_poll_waitq, wait);
173 
174 	read_lock_irqsave(&rs->rs_recv_lock, flags);
175 	if (!rs->rs_cong_monitor) {
176 		/* When a congestion map was updated, we signal POLLIN for
177 		 * "historical" reasons. Applications can also poll for
178 		 * WRBAND instead. */
179 		if (rds_cong_updated_since(&rs->rs_cong_track))
180 			mask |= (POLLIN | POLLRDNORM | POLLWRBAND);
181 	} else {
182 		spin_lock(&rs->rs_lock);
183 		if (rs->rs_cong_notify)
184 			mask |= (POLLIN | POLLRDNORM);
185 		spin_unlock(&rs->rs_lock);
186 	}
187 	if (!list_empty(&rs->rs_recv_queue) ||
188 	    !list_empty(&rs->rs_notify_queue))
189 		mask |= (POLLIN | POLLRDNORM);
190 	if (rs->rs_snd_bytes < rds_sk_sndbuf(rs))
191 		mask |= (POLLOUT | POLLWRNORM);
192 	read_unlock_irqrestore(&rs->rs_recv_lock, flags);
193 
194 	/* clear state any time we wake a seen-congested socket */
195 	if (mask)
196 		rs->rs_seen_congestion = 0;
197 
198 	return mask;
199 }
200 
201 static int rds_ioctl(struct socket *sock, unsigned int cmd, unsigned long arg)
202 {
203 	return -ENOIOCTLCMD;
204 }
205 
206 static int rds_cancel_sent_to(struct rds_sock *rs, char __user *optval,
207 			      int len)
208 {
209 	struct sockaddr_in sin;
210 	int ret = 0;
211 
212 	/* racing with another thread binding seems ok here */
213 	if (rs->rs_bound_addr == 0) {
214 		ret = -ENOTCONN; /* XXX not a great errno */
215 		goto out;
216 	}
217 
218 	if (len < sizeof(struct sockaddr_in)) {
219 		ret = -EINVAL;
220 		goto out;
221 	}
222 
223 	if (copy_from_user(&sin, optval, sizeof(sin))) {
224 		ret = -EFAULT;
225 		goto out;
226 	}
227 
228 	rds_send_drop_to(rs, &sin);
229 out:
230 	return ret;
231 }
232 
233 static int rds_set_bool_option(unsigned char *optvar, char __user *optval,
234 			       int optlen)
235 {
236 	int value;
237 
238 	if (optlen < sizeof(int))
239 		return -EINVAL;
240 	if (get_user(value, (int __user *) optval))
241 		return -EFAULT;
242 	*optvar = !!value;
243 	return 0;
244 }
245 
246 static int rds_cong_monitor(struct rds_sock *rs, char __user *optval,
247 			    int optlen)
248 {
249 	int ret;
250 
251 	ret = rds_set_bool_option(&rs->rs_cong_monitor, optval, optlen);
252 	if (ret == 0) {
253 		if (rs->rs_cong_monitor) {
254 			rds_cong_add_socket(rs);
255 		} else {
256 			rds_cong_remove_socket(rs);
257 			rs->rs_cong_mask = 0;
258 			rs->rs_cong_notify = 0;
259 		}
260 	}
261 	return ret;
262 }
263 
264 static int rds_set_transport(struct rds_sock *rs, char __user *optval,
265 			     int optlen)
266 {
267 	int t_type;
268 
269 	if (rs->rs_transport)
270 		return -EOPNOTSUPP; /* previously attached to transport */
271 
272 	if (optlen != sizeof(int))
273 		return -EINVAL;
274 
275 	if (copy_from_user(&t_type, (int __user *)optval, sizeof(t_type)))
276 		return -EFAULT;
277 
278 	if (t_type < 0 || t_type >= RDS_TRANS_COUNT)
279 		return -EINVAL;
280 
281 	rs->rs_transport = rds_trans_get(t_type);
282 
283 	return rs->rs_transport ? 0 : -ENOPROTOOPT;
284 }
285 
286 static int rds_setsockopt(struct socket *sock, int level, int optname,
287 			  char __user *optval, unsigned int optlen)
288 {
289 	struct rds_sock *rs = rds_sk_to_rs(sock->sk);
290 	int ret;
291 
292 	if (level != SOL_RDS) {
293 		ret = -ENOPROTOOPT;
294 		goto out;
295 	}
296 
297 	switch (optname) {
298 	case RDS_CANCEL_SENT_TO:
299 		ret = rds_cancel_sent_to(rs, optval, optlen);
300 		break;
301 	case RDS_GET_MR:
302 		ret = rds_get_mr(rs, optval, optlen);
303 		break;
304 	case RDS_GET_MR_FOR_DEST:
305 		ret = rds_get_mr_for_dest(rs, optval, optlen);
306 		break;
307 	case RDS_FREE_MR:
308 		ret = rds_free_mr(rs, optval, optlen);
309 		break;
310 	case RDS_RECVERR:
311 		ret = rds_set_bool_option(&rs->rs_recverr, optval, optlen);
312 		break;
313 	case RDS_CONG_MONITOR:
314 		ret = rds_cong_monitor(rs, optval, optlen);
315 		break;
316 	case SO_RDS_TRANSPORT:
317 		lock_sock(sock->sk);
318 		ret = rds_set_transport(rs, optval, optlen);
319 		release_sock(sock->sk);
320 		break;
321 	default:
322 		ret = -ENOPROTOOPT;
323 	}
324 out:
325 	return ret;
326 }
327 
328 static int rds_getsockopt(struct socket *sock, int level, int optname,
329 			  char __user *optval, int __user *optlen)
330 {
331 	struct rds_sock *rs = rds_sk_to_rs(sock->sk);
332 	int ret = -ENOPROTOOPT, len;
333 	int trans;
334 
335 	if (level != SOL_RDS)
336 		goto out;
337 
338 	if (get_user(len, optlen)) {
339 		ret = -EFAULT;
340 		goto out;
341 	}
342 
343 	switch (optname) {
344 	case RDS_INFO_FIRST ... RDS_INFO_LAST:
345 		ret = rds_info_getsockopt(sock, optname, optval,
346 					  optlen);
347 		break;
348 
349 	case RDS_RECVERR:
350 		if (len < sizeof(int))
351 			ret = -EINVAL;
352 		else
353 		if (put_user(rs->rs_recverr, (int __user *) optval) ||
354 		    put_user(sizeof(int), optlen))
355 			ret = -EFAULT;
356 		else
357 			ret = 0;
358 		break;
359 	case SO_RDS_TRANSPORT:
360 		if (len < sizeof(int)) {
361 			ret = -EINVAL;
362 			break;
363 		}
364 		trans = (rs->rs_transport ? rs->rs_transport->t_type :
365 			 RDS_TRANS_NONE); /* unbound */
366 		if (put_user(trans, (int __user *)optval) ||
367 		    put_user(sizeof(int), optlen))
368 			ret = -EFAULT;
369 		else
370 			ret = 0;
371 		break;
372 	default:
373 		break;
374 	}
375 
376 out:
377 	return ret;
378 
379 }
380 
381 static int rds_connect(struct socket *sock, struct sockaddr *uaddr,
382 		       int addr_len, int flags)
383 {
384 	struct sock *sk = sock->sk;
385 	struct sockaddr_in *sin = (struct sockaddr_in *)uaddr;
386 	struct rds_sock *rs = rds_sk_to_rs(sk);
387 	int ret = 0;
388 
389 	lock_sock(sk);
390 
391 	if (addr_len != sizeof(struct sockaddr_in)) {
392 		ret = -EINVAL;
393 		goto out;
394 	}
395 
396 	if (sin->sin_family != AF_INET) {
397 		ret = -EAFNOSUPPORT;
398 		goto out;
399 	}
400 
401 	if (sin->sin_addr.s_addr == htonl(INADDR_ANY)) {
402 		ret = -EDESTADDRREQ;
403 		goto out;
404 	}
405 
406 	rs->rs_conn_addr = sin->sin_addr.s_addr;
407 	rs->rs_conn_port = sin->sin_port;
408 
409 out:
410 	release_sock(sk);
411 	return ret;
412 }
413 
414 static struct proto rds_proto = {
415 	.name	  = "RDS",
416 	.owner	  = THIS_MODULE,
417 	.obj_size = sizeof(struct rds_sock),
418 };
419 
420 static const struct proto_ops rds_proto_ops = {
421 	.family =	AF_RDS,
422 	.owner =	THIS_MODULE,
423 	.release =	rds_release,
424 	.bind =		rds_bind,
425 	.connect =	rds_connect,
426 	.socketpair =	sock_no_socketpair,
427 	.accept =	sock_no_accept,
428 	.getname =	rds_getname,
429 	.poll =		rds_poll,
430 	.ioctl =	rds_ioctl,
431 	.listen =	sock_no_listen,
432 	.shutdown =	sock_no_shutdown,
433 	.setsockopt =	rds_setsockopt,
434 	.getsockopt =	rds_getsockopt,
435 	.sendmsg =	rds_sendmsg,
436 	.recvmsg =	rds_recvmsg,
437 	.mmap =		sock_no_mmap,
438 	.sendpage =	sock_no_sendpage,
439 };
440 
441 static void rds_sock_destruct(struct sock *sk)
442 {
443 	struct rds_sock *rs = rds_sk_to_rs(sk);
444 
445 	WARN_ON((&rs->rs_item != rs->rs_item.next ||
446 		 &rs->rs_item != rs->rs_item.prev));
447 }
448 
449 static int __rds_create(struct socket *sock, struct sock *sk, int protocol)
450 {
451 	struct rds_sock *rs;
452 
453 	sock_init_data(sock, sk);
454 	sock->ops		= &rds_proto_ops;
455 	sk->sk_protocol		= protocol;
456 	sk->sk_destruct		= rds_sock_destruct;
457 
458 	rs = rds_sk_to_rs(sk);
459 	spin_lock_init(&rs->rs_lock);
460 	rwlock_init(&rs->rs_recv_lock);
461 	INIT_LIST_HEAD(&rs->rs_send_queue);
462 	INIT_LIST_HEAD(&rs->rs_recv_queue);
463 	INIT_LIST_HEAD(&rs->rs_notify_queue);
464 	INIT_LIST_HEAD(&rs->rs_cong_list);
465 	spin_lock_init(&rs->rs_rdma_lock);
466 	rs->rs_rdma_keys = RB_ROOT;
467 
468 	spin_lock_bh(&rds_sock_lock);
469 	list_add_tail(&rs->rs_item, &rds_sock_list);
470 	rds_sock_count++;
471 	spin_unlock_bh(&rds_sock_lock);
472 
473 	return 0;
474 }
475 
476 static int rds_create(struct net *net, struct socket *sock, int protocol,
477 		      int kern)
478 {
479 	struct sock *sk;
480 
481 	if (sock->type != SOCK_SEQPACKET || protocol)
482 		return -ESOCKTNOSUPPORT;
483 
484 	sk = sk_alloc(net, AF_RDS, GFP_ATOMIC, &rds_proto, kern);
485 	if (!sk)
486 		return -ENOMEM;
487 
488 	return __rds_create(sock, sk, protocol);
489 }
490 
491 void rds_sock_addref(struct rds_sock *rs)
492 {
493 	sock_hold(rds_rs_to_sk(rs));
494 }
495 
496 void rds_sock_put(struct rds_sock *rs)
497 {
498 	sock_put(rds_rs_to_sk(rs));
499 }
500 
501 static const struct net_proto_family rds_family_ops = {
502 	.family =	AF_RDS,
503 	.create =	rds_create,
504 	.owner	=	THIS_MODULE,
505 };
506 
507 static void rds_sock_inc_info(struct socket *sock, unsigned int len,
508 			      struct rds_info_iterator *iter,
509 			      struct rds_info_lengths *lens)
510 {
511 	struct rds_sock *rs;
512 	struct rds_incoming *inc;
513 	unsigned int total = 0;
514 
515 	len /= sizeof(struct rds_info_message);
516 
517 	spin_lock_bh(&rds_sock_lock);
518 
519 	list_for_each_entry(rs, &rds_sock_list, rs_item) {
520 		read_lock(&rs->rs_recv_lock);
521 
522 		/* XXX too lazy to maintain counts.. */
523 		list_for_each_entry(inc, &rs->rs_recv_queue, i_item) {
524 			total++;
525 			if (total <= len)
526 				rds_inc_info_copy(inc, iter, inc->i_saddr,
527 						  rs->rs_bound_addr, 1);
528 		}
529 
530 		read_unlock(&rs->rs_recv_lock);
531 	}
532 
533 	spin_unlock_bh(&rds_sock_lock);
534 
535 	lens->nr = total;
536 	lens->each = sizeof(struct rds_info_message);
537 }
538 
539 static void rds_sock_info(struct socket *sock, unsigned int len,
540 			  struct rds_info_iterator *iter,
541 			  struct rds_info_lengths *lens)
542 {
543 	struct rds_info_socket sinfo;
544 	struct rds_sock *rs;
545 
546 	len /= sizeof(struct rds_info_socket);
547 
548 	spin_lock_bh(&rds_sock_lock);
549 
550 	if (len < rds_sock_count)
551 		goto out;
552 
553 	list_for_each_entry(rs, &rds_sock_list, rs_item) {
554 		sinfo.sndbuf = rds_sk_sndbuf(rs);
555 		sinfo.rcvbuf = rds_sk_rcvbuf(rs);
556 		sinfo.bound_addr = rs->rs_bound_addr;
557 		sinfo.connected_addr = rs->rs_conn_addr;
558 		sinfo.bound_port = rs->rs_bound_port;
559 		sinfo.connected_port = rs->rs_conn_port;
560 		sinfo.inum = sock_i_ino(rds_rs_to_sk(rs));
561 
562 		rds_info_copy(iter, &sinfo, sizeof(sinfo));
563 	}
564 
565 out:
566 	lens->nr = rds_sock_count;
567 	lens->each = sizeof(struct rds_info_socket);
568 
569 	spin_unlock_bh(&rds_sock_lock);
570 }
571 
572 static void rds_exit(void)
573 {
574 	sock_unregister(rds_family_ops.family);
575 	proto_unregister(&rds_proto);
576 	rds_conn_exit();
577 	rds_cong_exit();
578 	rds_sysctl_exit();
579 	rds_threads_exit();
580 	rds_stats_exit();
581 	rds_page_exit();
582 	rds_info_deregister_func(RDS_INFO_SOCKETS, rds_sock_info);
583 	rds_info_deregister_func(RDS_INFO_RECV_MESSAGES, rds_sock_inc_info);
584 }
585 module_exit(rds_exit);
586 
587 static int rds_init(void)
588 {
589 	int ret;
590 
591 	ret = rds_conn_init();
592 	if (ret)
593 		goto out;
594 	ret = rds_threads_init();
595 	if (ret)
596 		goto out_conn;
597 	ret = rds_sysctl_init();
598 	if (ret)
599 		goto out_threads;
600 	ret = rds_stats_init();
601 	if (ret)
602 		goto out_sysctl;
603 	ret = proto_register(&rds_proto, 1);
604 	if (ret)
605 		goto out_stats;
606 	ret = sock_register(&rds_family_ops);
607 	if (ret)
608 		goto out_proto;
609 
610 	rds_info_register_func(RDS_INFO_SOCKETS, rds_sock_info);
611 	rds_info_register_func(RDS_INFO_RECV_MESSAGES, rds_sock_inc_info);
612 
613 	goto out;
614 
615 out_proto:
616 	proto_unregister(&rds_proto);
617 out_stats:
618 	rds_stats_exit();
619 out_sysctl:
620 	rds_sysctl_exit();
621 out_threads:
622 	rds_threads_exit();
623 out_conn:
624 	rds_conn_exit();
625 	rds_cong_exit();
626 	rds_page_exit();
627 out:
628 	return ret;
629 }
630 module_init(rds_init);
631 
632 #define DRV_VERSION     "4.0"
633 #define DRV_RELDATE     "Feb 12, 2009"
634 
635 MODULE_AUTHOR("Oracle Corporation <rds-devel@oss.oracle.com>");
636 MODULE_DESCRIPTION("RDS: Reliable Datagram Sockets"
637 		   " v" DRV_VERSION " (" DRV_RELDATE ")");
638 MODULE_VERSION(DRV_VERSION);
639 MODULE_LICENSE("Dual BSD/GPL");
640 MODULE_ALIAS_NETPROTO(PF_RDS);
641