xref: /openbmc/linux/net/smc/af_smc.c (revision 5a244f48)
1 /*
2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
3  *
4  *  AF_SMC protocol family socket handler keeping the AF_INET sock address type
5  *  applies to SOCK_STREAM sockets only
6  *  offers an alternative communication option for TCP-protocol sockets
7  *  applicable with RoCE-cards only
8  *
9  *  Initial restrictions:
10  *    - non-blocking connect postponed
11  *    - IPv6 support postponed
12  *    - support for alternate links postponed
13  *    - partial support for non-blocking sockets only
14  *    - support for urgent data postponed
15  *
16  *  Copyright IBM Corp. 2016
17  *
18  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
19  *              based on prototype from Frank Blaschka
20  */
21 
22 #define KMSG_COMPONENT "smc"
23 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
24 
25 #include <linux/module.h>
26 #include <linux/socket.h>
27 #include <linux/inetdevice.h>
28 #include <linux/workqueue.h>
29 #include <linux/in.h>
30 #include <linux/sched/signal.h>
31 
32 #include <net/sock.h>
33 #include <net/tcp.h>
34 #include <net/smc.h>
35 
36 #include "smc.h"
37 #include "smc_clc.h"
38 #include "smc_llc.h"
39 #include "smc_cdc.h"
40 #include "smc_core.h"
41 #include "smc_ib.h"
42 #include "smc_pnet.h"
43 #include "smc_tx.h"
44 #include "smc_rx.h"
45 #include "smc_close.h"
46 
47 static DEFINE_MUTEX(smc_create_lgr_pending);	/* serialize link group
48 						 * creation
49 						 */
50 
51 struct smc_lgr_list smc_lgr_list = {		/* established link groups */
52 	.lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
53 	.list = LIST_HEAD_INIT(smc_lgr_list.list),
54 };
55 
56 static void smc_tcp_listen_work(struct work_struct *);
57 
58 static void smc_set_keepalive(struct sock *sk, int val)
59 {
60 	struct smc_sock *smc = smc_sk(sk);
61 
62 	smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
63 }
64 
65 static struct smc_hashinfo smc_v4_hashinfo = {
66 	.lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
67 };
68 
69 int smc_hash_sk(struct sock *sk)
70 {
71 	struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
72 	struct hlist_head *head;
73 
74 	head = &h->ht;
75 
76 	write_lock_bh(&h->lock);
77 	sk_add_node(sk, head);
78 	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
79 	write_unlock_bh(&h->lock);
80 
81 	return 0;
82 }
83 EXPORT_SYMBOL_GPL(smc_hash_sk);
84 
85 void smc_unhash_sk(struct sock *sk)
86 {
87 	struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
88 
89 	write_lock_bh(&h->lock);
90 	if (sk_del_node_init(sk))
91 		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
92 	write_unlock_bh(&h->lock);
93 }
94 EXPORT_SYMBOL_GPL(smc_unhash_sk);
95 
96 struct proto smc_proto = {
97 	.name		= "SMC",
98 	.owner		= THIS_MODULE,
99 	.keepalive	= smc_set_keepalive,
100 	.hash		= smc_hash_sk,
101 	.unhash		= smc_unhash_sk,
102 	.obj_size	= sizeof(struct smc_sock),
103 	.h.smc_hash	= &smc_v4_hashinfo,
104 	.slab_flags	= SLAB_TYPESAFE_BY_RCU,
105 };
106 EXPORT_SYMBOL_GPL(smc_proto);
107 
108 static int smc_release(struct socket *sock)
109 {
110 	struct sock *sk = sock->sk;
111 	struct smc_sock *smc;
112 	int rc = 0;
113 
114 	if (!sk)
115 		goto out;
116 
117 	smc = smc_sk(sk);
118 	sock_hold(sk);
119 	if (sk->sk_state == SMC_LISTEN)
120 		/* smc_close_non_accepted() is called and acquires
121 		 * sock lock for child sockets again
122 		 */
123 		lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
124 	else
125 		lock_sock(sk);
126 
127 	if (smc->use_fallback) {
128 		sk->sk_state = SMC_CLOSED;
129 		sk->sk_state_change(sk);
130 	} else {
131 		rc = smc_close_active(smc);
132 		sock_set_flag(sk, SOCK_DEAD);
133 		sk->sk_shutdown |= SHUTDOWN_MASK;
134 	}
135 	if (smc->clcsock) {
136 		sock_release(smc->clcsock);
137 		smc->clcsock = NULL;
138 	}
139 
140 	/* detach socket */
141 	sock_orphan(sk);
142 	sock->sk = NULL;
143 	if (smc->use_fallback) {
144 		schedule_delayed_work(&smc->sock_put_work, TCP_TIMEWAIT_LEN);
145 	} else if (sk->sk_state == SMC_CLOSED) {
146 		smc_conn_free(&smc->conn);
147 		schedule_delayed_work(&smc->sock_put_work,
148 				      SMC_CLOSE_SOCK_PUT_DELAY);
149 	}
150 	release_sock(sk);
151 
152 	sock_put(sk);
153 out:
154 	return rc;
155 }
156 
157 static void smc_destruct(struct sock *sk)
158 {
159 	if (sk->sk_state != SMC_CLOSED)
160 		return;
161 	if (!sock_flag(sk, SOCK_DEAD))
162 		return;
163 
164 	sk_refcnt_debug_dec(sk);
165 }
166 
167 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock)
168 {
169 	struct smc_sock *smc;
170 	struct sock *sk;
171 
172 	sk = sk_alloc(net, PF_SMC, GFP_KERNEL, &smc_proto, 0);
173 	if (!sk)
174 		return NULL;
175 
176 	sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
177 	sk->sk_state = SMC_INIT;
178 	sk->sk_destruct = smc_destruct;
179 	sk->sk_protocol = SMCPROTO_SMC;
180 	smc = smc_sk(sk);
181 	INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
182 	INIT_LIST_HEAD(&smc->accept_q);
183 	spin_lock_init(&smc->accept_q_lock);
184 	INIT_DELAYED_WORK(&smc->sock_put_work, smc_close_sock_put_work);
185 	sk->sk_prot->hash(sk);
186 	sk_refcnt_debug_inc(sk);
187 
188 	return sk;
189 }
190 
191 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
192 		    int addr_len)
193 {
194 	struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
195 	struct sock *sk = sock->sk;
196 	struct smc_sock *smc;
197 	int rc;
198 
199 	smc = smc_sk(sk);
200 
201 	/* replicate tests from inet_bind(), to be safe wrt. future changes */
202 	rc = -EINVAL;
203 	if (addr_len < sizeof(struct sockaddr_in))
204 		goto out;
205 
206 	rc = -EAFNOSUPPORT;
207 	/* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
208 	if ((addr->sin_family != AF_INET) &&
209 	    ((addr->sin_family != AF_UNSPEC) ||
210 	     (addr->sin_addr.s_addr != htonl(INADDR_ANY))))
211 		goto out;
212 
213 	lock_sock(sk);
214 
215 	/* Check if socket is already active */
216 	rc = -EINVAL;
217 	if (sk->sk_state != SMC_INIT)
218 		goto out_rel;
219 
220 	smc->clcsock->sk->sk_reuse = sk->sk_reuse;
221 	rc = kernel_bind(smc->clcsock, uaddr, addr_len);
222 
223 out_rel:
224 	release_sock(sk);
225 out:
226 	return rc;
227 }
228 
229 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
230 				   unsigned long mask)
231 {
232 	/* options we don't get control via setsockopt for */
233 	nsk->sk_type = osk->sk_type;
234 	nsk->sk_sndbuf = osk->sk_sndbuf;
235 	nsk->sk_rcvbuf = osk->sk_rcvbuf;
236 	nsk->sk_sndtimeo = osk->sk_sndtimeo;
237 	nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
238 	nsk->sk_mark = osk->sk_mark;
239 	nsk->sk_priority = osk->sk_priority;
240 	nsk->sk_rcvlowat = osk->sk_rcvlowat;
241 	nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
242 	nsk->sk_err = osk->sk_err;
243 
244 	nsk->sk_flags &= ~mask;
245 	nsk->sk_flags |= osk->sk_flags & mask;
246 }
247 
248 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
249 			     (1UL << SOCK_KEEPOPEN) | \
250 			     (1UL << SOCK_LINGER) | \
251 			     (1UL << SOCK_BROADCAST) | \
252 			     (1UL << SOCK_TIMESTAMP) | \
253 			     (1UL << SOCK_DBG) | \
254 			     (1UL << SOCK_RCVTSTAMP) | \
255 			     (1UL << SOCK_RCVTSTAMPNS) | \
256 			     (1UL << SOCK_LOCALROUTE) | \
257 			     (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
258 			     (1UL << SOCK_RXQ_OVFL) | \
259 			     (1UL << SOCK_WIFI_STATUS) | \
260 			     (1UL << SOCK_NOFCS) | \
261 			     (1UL << SOCK_FILTER_LOCKED))
262 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
263  * clc socket (since smc is not called for these options from net/core)
264  */
265 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
266 {
267 	smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
268 }
269 
270 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
271 			     (1UL << SOCK_KEEPOPEN) | \
272 			     (1UL << SOCK_LINGER) | \
273 			     (1UL << SOCK_DBG))
274 /* copy only settings and flags relevant for smc from clc to smc socket */
275 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
276 {
277 	smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
278 }
279 
280 /* determine subnet and mask of internal TCP socket */
281 int smc_netinfo_by_tcpsk(struct socket *clcsock,
282 			 __be32 *subnet, u8 *prefix_len)
283 {
284 	struct dst_entry *dst = sk_dst_get(clcsock->sk);
285 	struct in_device *in_dev;
286 	struct sockaddr_in addr;
287 	int rc = -ENOENT;
288 	int len;
289 
290 	if (!dst) {
291 		rc = -ENOTCONN;
292 		goto out;
293 	}
294 	if (!dst->dev) {
295 		rc = -ENODEV;
296 		goto out_rel;
297 	}
298 
299 	/* get address to which the internal TCP socket is bound */
300 	kernel_getsockname(clcsock, (struct sockaddr *)&addr, &len);
301 	/* analyze IPv4 specific data of net_device belonging to TCP socket */
302 	rcu_read_lock();
303 	in_dev = __in_dev_get_rcu(dst->dev);
304 	for_ifa(in_dev) {
305 		if (!inet_ifa_match(addr.sin_addr.s_addr, ifa))
306 			continue;
307 		*prefix_len = inet_mask_len(ifa->ifa_mask);
308 		*subnet = ifa->ifa_address & ifa->ifa_mask;
309 		rc = 0;
310 		break;
311 	} endfor_ifa(in_dev);
312 	rcu_read_unlock();
313 
314 out_rel:
315 	dst_release(dst);
316 out:
317 	return rc;
318 }
319 
320 static int smc_clnt_conf_first_link(struct smc_sock *smc, union ib_gid *gid)
321 {
322 	struct smc_link_group *lgr = smc->conn.lgr;
323 	struct smc_link *link;
324 	int rest;
325 	int rc;
326 
327 	link = &lgr->lnk[SMC_SINGLE_LINK];
328 	/* receive CONFIRM LINK request from server over RoCE fabric */
329 	rest = wait_for_completion_interruptible_timeout(
330 		&link->llc_confirm,
331 		SMC_LLC_WAIT_FIRST_TIME);
332 	if (rest <= 0) {
333 		struct smc_clc_msg_decline dclc;
334 
335 		rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
336 				      SMC_CLC_DECLINE);
337 		return rc;
338 	}
339 
340 	rc = smc_ib_modify_qp_rts(link);
341 	if (rc)
342 		return SMC_CLC_DECL_INTERR;
343 
344 	smc_wr_remember_qp_attr(link);
345 
346 	rc = smc_wr_reg_send(link,
347 			     smc->conn.rmb_desc->mr_rx[SMC_SINGLE_LINK]);
348 	if (rc)
349 		return SMC_CLC_DECL_INTERR;
350 
351 	/* send CONFIRM LINK response over RoCE fabric */
352 	rc = smc_llc_send_confirm_link(link,
353 				       link->smcibdev->mac[link->ibport - 1],
354 				       gid, SMC_LLC_RESP);
355 	if (rc < 0)
356 		return SMC_CLC_DECL_TCL;
357 
358 	return rc;
359 }
360 
361 static void smc_conn_save_peer_info(struct smc_sock *smc,
362 				    struct smc_clc_msg_accept_confirm *clc)
363 {
364 	smc->conn.peer_conn_idx = clc->conn_idx;
365 	smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
366 	smc->conn.peer_rmbe_size = smc_uncompress_bufsize(clc->rmbe_size);
367 	atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
368 }
369 
370 static void smc_link_save_peer_info(struct smc_link *link,
371 				    struct smc_clc_msg_accept_confirm *clc)
372 {
373 	link->peer_qpn = ntoh24(clc->qpn);
374 	memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
375 	memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
376 	link->peer_psn = ntoh24(clc->psn);
377 	link->peer_mtu = clc->qp_mtu;
378 }
379 
380 /* setup for RDMA connection of client */
381 static int smc_connect_rdma(struct smc_sock *smc)
382 {
383 	struct sockaddr_in *inaddr = (struct sockaddr_in *)smc->addr;
384 	struct smc_clc_msg_accept_confirm aclc;
385 	int local_contact = SMC_FIRST_CONTACT;
386 	struct smc_ib_device *smcibdev;
387 	struct smc_link *link;
388 	u8 srv_first_contact;
389 	int reason_code = 0;
390 	int rc = 0;
391 	u8 ibport;
392 
393 	/* IPSec connections opt out of SMC-R optimizations */
394 	if (using_ipsec(smc)) {
395 		reason_code = SMC_CLC_DECL_IPSEC;
396 		goto decline_rdma;
397 	}
398 
399 	/* PNET table look up: search active ib_device and port
400 	 * within same PNETID that also contains the ethernet device
401 	 * used for the internal TCP socket
402 	 */
403 	smc_pnet_find_roce_resource(smc->clcsock->sk, &smcibdev, &ibport);
404 	if (!smcibdev) {
405 		reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
406 		goto decline_rdma;
407 	}
408 
409 	/* do inband token exchange */
410 	reason_code = smc_clc_send_proposal(smc, smcibdev, ibport);
411 	if (reason_code < 0) {
412 		rc = reason_code;
413 		goto out_err;
414 	}
415 	if (reason_code > 0) /* configuration error */
416 		goto decline_rdma;
417 	/* receive SMC Accept CLC message */
418 	reason_code = smc_clc_wait_msg(smc, &aclc, sizeof(aclc),
419 				       SMC_CLC_ACCEPT);
420 	if (reason_code < 0) {
421 		rc = reason_code;
422 		goto out_err;
423 	}
424 	if (reason_code > 0)
425 		goto decline_rdma;
426 
427 	srv_first_contact = aclc.hdr.flag;
428 	mutex_lock(&smc_create_lgr_pending);
429 	local_contact = smc_conn_create(smc, inaddr->sin_addr.s_addr, smcibdev,
430 					ibport, &aclc.lcl, srv_first_contact);
431 	if (local_contact < 0) {
432 		rc = local_contact;
433 		if (rc == -ENOMEM)
434 			reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
435 		else if (rc == -ENOLINK)
436 			reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
437 		goto decline_rdma_unlock;
438 	}
439 	link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
440 
441 	smc_conn_save_peer_info(smc, &aclc);
442 
443 	/* create send buffer and rmb */
444 	rc = smc_buf_create(smc);
445 	if (rc) {
446 		reason_code = SMC_CLC_DECL_MEM;
447 		goto decline_rdma_unlock;
448 	}
449 
450 	if (local_contact == SMC_FIRST_CONTACT)
451 		smc_link_save_peer_info(link, &aclc);
452 
453 	rc = smc_rmb_rtoken_handling(&smc->conn, &aclc);
454 	if (rc) {
455 		reason_code = SMC_CLC_DECL_INTERR;
456 		goto decline_rdma_unlock;
457 	}
458 
459 	smc_close_init(smc);
460 	smc_rx_init(smc);
461 
462 	if (local_contact == SMC_FIRST_CONTACT) {
463 		rc = smc_ib_ready_link(link);
464 		if (rc) {
465 			reason_code = SMC_CLC_DECL_INTERR;
466 			goto decline_rdma_unlock;
467 		}
468 	} else {
469 		struct smc_buf_desc *buf_desc = smc->conn.rmb_desc;
470 
471 		if (!buf_desc->reused) {
472 			/* register memory region for new rmb */
473 			rc = smc_wr_reg_send(link,
474 					     buf_desc->mr_rx[SMC_SINGLE_LINK]);
475 			if (rc) {
476 				reason_code = SMC_CLC_DECL_INTERR;
477 				goto decline_rdma_unlock;
478 			}
479 		}
480 	}
481 	smc_rmb_sync_sg_for_device(&smc->conn);
482 
483 	rc = smc_clc_send_confirm(smc);
484 	if (rc)
485 		goto out_err_unlock;
486 
487 	if (local_contact == SMC_FIRST_CONTACT) {
488 		/* QP confirmation over RoCE fabric */
489 		reason_code = smc_clnt_conf_first_link(
490 			smc, &smcibdev->gid[ibport - 1]);
491 		if (reason_code < 0) {
492 			rc = reason_code;
493 			goto out_err_unlock;
494 		}
495 		if (reason_code > 0)
496 			goto decline_rdma_unlock;
497 	}
498 
499 	mutex_unlock(&smc_create_lgr_pending);
500 	smc_tx_init(smc);
501 
502 out_connected:
503 	smc_copy_sock_settings_to_clc(smc);
504 	if (smc->sk.sk_state == SMC_INIT)
505 		smc->sk.sk_state = SMC_ACTIVE;
506 
507 	return rc ? rc : local_contact;
508 
509 decline_rdma_unlock:
510 	mutex_unlock(&smc_create_lgr_pending);
511 	smc_conn_free(&smc->conn);
512 decline_rdma:
513 	/* RDMA setup failed, switch back to TCP */
514 	smc->use_fallback = true;
515 	if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
516 		rc = smc_clc_send_decline(smc, reason_code);
517 		if (rc < sizeof(struct smc_clc_msg_decline))
518 			goto out_err;
519 	}
520 	goto out_connected;
521 
522 out_err_unlock:
523 	mutex_unlock(&smc_create_lgr_pending);
524 	smc_conn_free(&smc->conn);
525 out_err:
526 	return rc;
527 }
528 
529 static int smc_connect(struct socket *sock, struct sockaddr *addr,
530 		       int alen, int flags)
531 {
532 	struct sock *sk = sock->sk;
533 	struct smc_sock *smc;
534 	int rc = -EINVAL;
535 
536 	smc = smc_sk(sk);
537 
538 	/* separate smc parameter checking to be safe */
539 	if (alen < sizeof(addr->sa_family))
540 		goto out_err;
541 	if (addr->sa_family != AF_INET)
542 		goto out_err;
543 	smc->addr = addr;	/* needed for nonblocking connect */
544 
545 	lock_sock(sk);
546 	switch (sk->sk_state) {
547 	default:
548 		goto out;
549 	case SMC_ACTIVE:
550 		rc = -EISCONN;
551 		goto out;
552 	case SMC_INIT:
553 		rc = 0;
554 		break;
555 	}
556 
557 	smc_copy_sock_settings_to_clc(smc);
558 	rc = kernel_connect(smc->clcsock, addr, alen, flags);
559 	if (rc)
560 		goto out;
561 
562 	/* setup RDMA connection */
563 	rc = smc_connect_rdma(smc);
564 	if (rc < 0)
565 		goto out;
566 	else
567 		rc = 0; /* success cases including fallback */
568 
569 out:
570 	release_sock(sk);
571 out_err:
572 	return rc;
573 }
574 
575 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
576 {
577 	struct sock *sk = &lsmc->sk;
578 	struct socket *new_clcsock;
579 	struct sock *new_sk;
580 	int rc;
581 
582 	release_sock(&lsmc->sk);
583 	new_sk = smc_sock_alloc(sock_net(sk), NULL);
584 	if (!new_sk) {
585 		rc = -ENOMEM;
586 		lsmc->sk.sk_err = ENOMEM;
587 		*new_smc = NULL;
588 		lock_sock(&lsmc->sk);
589 		goto out;
590 	}
591 	*new_smc = smc_sk(new_sk);
592 
593 	rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
594 	lock_sock(&lsmc->sk);
595 	if  (rc < 0) {
596 		lsmc->sk.sk_err = -rc;
597 		new_sk->sk_state = SMC_CLOSED;
598 		sock_set_flag(new_sk, SOCK_DEAD);
599 		sk->sk_prot->unhash(new_sk);
600 		sock_put(new_sk);
601 		*new_smc = NULL;
602 		goto out;
603 	}
604 	if (lsmc->sk.sk_state == SMC_CLOSED) {
605 		if (new_clcsock)
606 			sock_release(new_clcsock);
607 		new_sk->sk_state = SMC_CLOSED;
608 		sock_set_flag(new_sk, SOCK_DEAD);
609 		sk->sk_prot->unhash(new_sk);
610 		sock_put(new_sk);
611 		*new_smc = NULL;
612 		goto out;
613 	}
614 
615 	(*new_smc)->clcsock = new_clcsock;
616 out:
617 	return rc;
618 }
619 
620 /* add a just created sock to the accept queue of the listen sock as
621  * candidate for a following socket accept call from user space
622  */
623 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
624 {
625 	struct smc_sock *par = smc_sk(parent);
626 
627 	sock_hold(sk);
628 	spin_lock(&par->accept_q_lock);
629 	list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
630 	spin_unlock(&par->accept_q_lock);
631 	sk_acceptq_added(parent);
632 }
633 
634 /* remove a socket from the accept queue of its parental listening socket */
635 static void smc_accept_unlink(struct sock *sk)
636 {
637 	struct smc_sock *par = smc_sk(sk)->listen_smc;
638 
639 	spin_lock(&par->accept_q_lock);
640 	list_del_init(&smc_sk(sk)->accept_q);
641 	spin_unlock(&par->accept_q_lock);
642 	sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
643 	sock_put(sk);
644 }
645 
646 /* remove a sock from the accept queue to bind it to a new socket created
647  * for a socket accept call from user space
648  */
649 struct sock *smc_accept_dequeue(struct sock *parent,
650 				struct socket *new_sock)
651 {
652 	struct smc_sock *isk, *n;
653 	struct sock *new_sk;
654 
655 	list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
656 		new_sk = (struct sock *)isk;
657 
658 		smc_accept_unlink(new_sk);
659 		if (new_sk->sk_state == SMC_CLOSED) {
660 			new_sk->sk_prot->unhash(new_sk);
661 			sock_put(new_sk);
662 			continue;
663 		}
664 		if (new_sock)
665 			sock_graft(new_sk, new_sock);
666 		return new_sk;
667 	}
668 	return NULL;
669 }
670 
671 /* clean up for a created but never accepted sock */
672 void smc_close_non_accepted(struct sock *sk)
673 {
674 	struct smc_sock *smc = smc_sk(sk);
675 
676 	sock_hold(sk);
677 	lock_sock(sk);
678 	if (!sk->sk_lingertime)
679 		/* wait for peer closing */
680 		sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
681 	if (smc->use_fallback) {
682 		sk->sk_state = SMC_CLOSED;
683 	} else {
684 		smc_close_active(smc);
685 		sock_set_flag(sk, SOCK_DEAD);
686 		sk->sk_shutdown |= SHUTDOWN_MASK;
687 	}
688 	if (smc->clcsock) {
689 		struct socket *tcp;
690 
691 		tcp = smc->clcsock;
692 		smc->clcsock = NULL;
693 		sock_release(tcp);
694 	}
695 	if (smc->use_fallback) {
696 		schedule_delayed_work(&smc->sock_put_work, TCP_TIMEWAIT_LEN);
697 	} else if (sk->sk_state == SMC_CLOSED) {
698 		smc_conn_free(&smc->conn);
699 		schedule_delayed_work(&smc->sock_put_work,
700 				      SMC_CLOSE_SOCK_PUT_DELAY);
701 	}
702 	release_sock(sk);
703 	sock_put(sk);
704 }
705 
706 static int smc_serv_conf_first_link(struct smc_sock *smc)
707 {
708 	struct smc_link_group *lgr = smc->conn.lgr;
709 	struct smc_link *link;
710 	int rest;
711 	int rc;
712 
713 	link = &lgr->lnk[SMC_SINGLE_LINK];
714 
715 	rc = smc_wr_reg_send(link,
716 			     smc->conn.rmb_desc->mr_rx[SMC_SINGLE_LINK]);
717 	if (rc)
718 		return SMC_CLC_DECL_INTERR;
719 
720 	/* send CONFIRM LINK request to client over the RoCE fabric */
721 	rc = smc_llc_send_confirm_link(link,
722 				       link->smcibdev->mac[link->ibport - 1],
723 				       &link->smcibdev->gid[link->ibport - 1],
724 				       SMC_LLC_REQ);
725 	if (rc < 0)
726 		return SMC_CLC_DECL_TCL;
727 
728 	/* receive CONFIRM LINK response from client over the RoCE fabric */
729 	rest = wait_for_completion_interruptible_timeout(
730 		&link->llc_confirm_resp,
731 		SMC_LLC_WAIT_FIRST_TIME);
732 	if (rest <= 0) {
733 		struct smc_clc_msg_decline dclc;
734 
735 		rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
736 				      SMC_CLC_DECLINE);
737 	}
738 
739 	return rc;
740 }
741 
742 /* setup for RDMA connection of server */
743 static void smc_listen_work(struct work_struct *work)
744 {
745 	struct smc_sock *new_smc = container_of(work, struct smc_sock,
746 						smc_listen_work);
747 	struct socket *newclcsock = new_smc->clcsock;
748 	struct smc_sock *lsmc = new_smc->listen_smc;
749 	struct smc_clc_msg_accept_confirm cclc;
750 	int local_contact = SMC_REUSE_CONTACT;
751 	struct sock *newsmcsk = &new_smc->sk;
752 	struct smc_clc_msg_proposal pclc;
753 	struct smc_ib_device *smcibdev;
754 	struct sockaddr_in peeraddr;
755 	struct smc_link *link;
756 	int reason_code = 0;
757 	int rc = 0, len;
758 	__be32 subnet;
759 	u8 prefix_len;
760 	u8 ibport;
761 
762 	/* do inband token exchange -
763 	 *wait for and receive SMC Proposal CLC message
764 	 */
765 	reason_code = smc_clc_wait_msg(new_smc, &pclc, sizeof(pclc),
766 				       SMC_CLC_PROPOSAL);
767 	if (reason_code < 0)
768 		goto out_err;
769 	if (reason_code > 0)
770 		goto decline_rdma;
771 
772 	/* IPSec connections opt out of SMC-R optimizations */
773 	if (using_ipsec(new_smc)) {
774 		reason_code = SMC_CLC_DECL_IPSEC;
775 		goto decline_rdma;
776 	}
777 
778 	/* PNET table look up: search active ib_device and port
779 	 * within same PNETID that also contains the ethernet device
780 	 * used for the internal TCP socket
781 	 */
782 	smc_pnet_find_roce_resource(newclcsock->sk, &smcibdev, &ibport);
783 	if (!smcibdev) {
784 		reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
785 		goto decline_rdma;
786 	}
787 
788 	/* determine subnet and mask from internal TCP socket */
789 	rc = smc_netinfo_by_tcpsk(newclcsock, &subnet, &prefix_len);
790 	if (rc) {
791 		reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
792 		goto decline_rdma;
793 	}
794 	if ((pclc.outgoing_subnet != subnet) ||
795 	    (pclc.prefix_len != prefix_len)) {
796 		reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
797 		goto decline_rdma;
798 	}
799 
800 	/* get address of the peer connected to the internal TCP socket */
801 	kernel_getpeername(newclcsock, (struct sockaddr *)&peeraddr, &len);
802 
803 	/* allocate connection / link group */
804 	mutex_lock(&smc_create_lgr_pending);
805 	local_contact = smc_conn_create(new_smc, peeraddr.sin_addr.s_addr,
806 					smcibdev, ibport, &pclc.lcl, 0);
807 	if (local_contact < 0) {
808 		rc = local_contact;
809 		if (rc == -ENOMEM)
810 			reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
811 		goto decline_rdma;
812 	}
813 	link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
814 
815 	/* create send buffer and rmb */
816 	rc = smc_buf_create(new_smc);
817 	if (rc) {
818 		reason_code = SMC_CLC_DECL_MEM;
819 		goto decline_rdma;
820 	}
821 
822 	smc_close_init(new_smc);
823 	smc_rx_init(new_smc);
824 
825 	if (local_contact != SMC_FIRST_CONTACT) {
826 		struct smc_buf_desc *buf_desc = new_smc->conn.rmb_desc;
827 
828 		if (!buf_desc->reused) {
829 			/* register memory region for new rmb */
830 			rc = smc_wr_reg_send(link,
831 					     buf_desc->mr_rx[SMC_SINGLE_LINK]);
832 			if (rc) {
833 				reason_code = SMC_CLC_DECL_INTERR;
834 				goto decline_rdma;
835 			}
836 		}
837 	}
838 	smc_rmb_sync_sg_for_device(&new_smc->conn);
839 
840 	rc = smc_clc_send_accept(new_smc, local_contact);
841 	if (rc)
842 		goto out_err;
843 
844 	/* receive SMC Confirm CLC message */
845 	reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
846 				       SMC_CLC_CONFIRM);
847 	if (reason_code < 0)
848 		goto out_err;
849 	if (reason_code > 0)
850 		goto decline_rdma;
851 	smc_conn_save_peer_info(new_smc, &cclc);
852 	if (local_contact == SMC_FIRST_CONTACT)
853 		smc_link_save_peer_info(link, &cclc);
854 
855 	rc = smc_rmb_rtoken_handling(&new_smc->conn, &cclc);
856 	if (rc) {
857 		reason_code = SMC_CLC_DECL_INTERR;
858 		goto decline_rdma;
859 	}
860 
861 	if (local_contact == SMC_FIRST_CONTACT) {
862 		rc = smc_ib_ready_link(link);
863 		if (rc) {
864 			reason_code = SMC_CLC_DECL_INTERR;
865 			goto decline_rdma;
866 		}
867 		/* QP confirmation over RoCE fabric */
868 		reason_code = smc_serv_conf_first_link(new_smc);
869 		if (reason_code < 0) {
870 			/* peer is not aware of a problem */
871 			rc = reason_code;
872 			goto out_err;
873 		}
874 		if (reason_code > 0)
875 			goto decline_rdma;
876 	}
877 
878 	smc_tx_init(new_smc);
879 
880 out_connected:
881 	sk_refcnt_debug_inc(newsmcsk);
882 	if (newsmcsk->sk_state == SMC_INIT)
883 		newsmcsk->sk_state = SMC_ACTIVE;
884 enqueue:
885 	mutex_unlock(&smc_create_lgr_pending);
886 	lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
887 	if (lsmc->sk.sk_state == SMC_LISTEN) {
888 		smc_accept_enqueue(&lsmc->sk, newsmcsk);
889 	} else { /* no longer listening */
890 		smc_close_non_accepted(newsmcsk);
891 	}
892 	release_sock(&lsmc->sk);
893 
894 	/* Wake up accept */
895 	lsmc->sk.sk_data_ready(&lsmc->sk);
896 	sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
897 	return;
898 
899 decline_rdma:
900 	/* RDMA setup failed, switch back to TCP */
901 	smc_conn_free(&new_smc->conn);
902 	new_smc->use_fallback = true;
903 	if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
904 		rc = smc_clc_send_decline(new_smc, reason_code);
905 		if (rc < sizeof(struct smc_clc_msg_decline))
906 			goto out_err;
907 	}
908 	goto out_connected;
909 
910 out_err:
911 	newsmcsk->sk_state = SMC_CLOSED;
912 	smc_conn_free(&new_smc->conn);
913 	goto enqueue; /* queue new sock with sk_err set */
914 }
915 
916 static void smc_tcp_listen_work(struct work_struct *work)
917 {
918 	struct smc_sock *lsmc = container_of(work, struct smc_sock,
919 					     tcp_listen_work);
920 	struct smc_sock *new_smc;
921 	int rc = 0;
922 
923 	lock_sock(&lsmc->sk);
924 	while (lsmc->sk.sk_state == SMC_LISTEN) {
925 		rc = smc_clcsock_accept(lsmc, &new_smc);
926 		if (rc)
927 			goto out;
928 		if (!new_smc)
929 			continue;
930 
931 		new_smc->listen_smc = lsmc;
932 		new_smc->use_fallback = false; /* assume rdma capability first*/
933 		sock_hold(&lsmc->sk); /* sock_put in smc_listen_work */
934 		INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
935 		smc_copy_sock_settings_to_smc(new_smc);
936 		schedule_work(&new_smc->smc_listen_work);
937 	}
938 
939 out:
940 	release_sock(&lsmc->sk);
941 	lsmc->sk.sk_data_ready(&lsmc->sk); /* no more listening, wake accept */
942 }
943 
944 static int smc_listen(struct socket *sock, int backlog)
945 {
946 	struct sock *sk = sock->sk;
947 	struct smc_sock *smc;
948 	int rc;
949 
950 	smc = smc_sk(sk);
951 	lock_sock(sk);
952 
953 	rc = -EINVAL;
954 	if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
955 		goto out;
956 
957 	rc = 0;
958 	if (sk->sk_state == SMC_LISTEN) {
959 		sk->sk_max_ack_backlog = backlog;
960 		goto out;
961 	}
962 	/* some socket options are handled in core, so we could not apply
963 	 * them to the clc socket -- copy smc socket options to clc socket
964 	 */
965 	smc_copy_sock_settings_to_clc(smc);
966 
967 	rc = kernel_listen(smc->clcsock, backlog);
968 	if (rc)
969 		goto out;
970 	sk->sk_max_ack_backlog = backlog;
971 	sk->sk_ack_backlog = 0;
972 	sk->sk_state = SMC_LISTEN;
973 	INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
974 	schedule_work(&smc->tcp_listen_work);
975 
976 out:
977 	release_sock(sk);
978 	return rc;
979 }
980 
981 static int smc_accept(struct socket *sock, struct socket *new_sock,
982 		      int flags, bool kern)
983 {
984 	struct sock *sk = sock->sk, *nsk;
985 	DECLARE_WAITQUEUE(wait, current);
986 	struct smc_sock *lsmc;
987 	long timeo;
988 	int rc = 0;
989 
990 	lsmc = smc_sk(sk);
991 	lock_sock(sk);
992 
993 	if (lsmc->sk.sk_state != SMC_LISTEN) {
994 		rc = -EINVAL;
995 		goto out;
996 	}
997 
998 	/* Wait for an incoming connection */
999 	timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1000 	add_wait_queue_exclusive(sk_sleep(sk), &wait);
1001 	while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1002 		set_current_state(TASK_INTERRUPTIBLE);
1003 		if (!timeo) {
1004 			rc = -EAGAIN;
1005 			break;
1006 		}
1007 		release_sock(sk);
1008 		timeo = schedule_timeout(timeo);
1009 		/* wakeup by sk_data_ready in smc_listen_work() */
1010 		sched_annotate_sleep();
1011 		lock_sock(sk);
1012 		if (signal_pending(current)) {
1013 			rc = sock_intr_errno(timeo);
1014 			break;
1015 		}
1016 	}
1017 	set_current_state(TASK_RUNNING);
1018 	remove_wait_queue(sk_sleep(sk), &wait);
1019 
1020 	if (!rc)
1021 		rc = sock_error(nsk);
1022 
1023 out:
1024 	release_sock(sk);
1025 	return rc;
1026 }
1027 
1028 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1029 		       int *len, int peer)
1030 {
1031 	struct smc_sock *smc;
1032 
1033 	if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1034 	    (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1035 		return -ENOTCONN;
1036 
1037 	smc = smc_sk(sock->sk);
1038 
1039 	return smc->clcsock->ops->getname(smc->clcsock, addr, len, peer);
1040 }
1041 
1042 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1043 {
1044 	struct sock *sk = sock->sk;
1045 	struct smc_sock *smc;
1046 	int rc = -EPIPE;
1047 
1048 	smc = smc_sk(sk);
1049 	lock_sock(sk);
1050 	if ((sk->sk_state != SMC_ACTIVE) &&
1051 	    (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1052 	    (sk->sk_state != SMC_INIT))
1053 		goto out;
1054 	if (smc->use_fallback)
1055 		rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1056 	else
1057 		rc = smc_tx_sendmsg(smc, msg, len);
1058 out:
1059 	release_sock(sk);
1060 	return rc;
1061 }
1062 
1063 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1064 		       int flags)
1065 {
1066 	struct sock *sk = sock->sk;
1067 	struct smc_sock *smc;
1068 	int rc = -ENOTCONN;
1069 
1070 	smc = smc_sk(sk);
1071 	lock_sock(sk);
1072 	if ((sk->sk_state == SMC_INIT) ||
1073 	    (sk->sk_state == SMC_LISTEN) ||
1074 	    (sk->sk_state == SMC_CLOSED))
1075 		goto out;
1076 
1077 	if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1078 		rc = 0;
1079 		goto out;
1080 	}
1081 
1082 	if (smc->use_fallback)
1083 		rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1084 	else
1085 		rc = smc_rx_recvmsg(smc, msg, len, flags);
1086 
1087 out:
1088 	release_sock(sk);
1089 	return rc;
1090 }
1091 
1092 static unsigned int smc_accept_poll(struct sock *parent)
1093 {
1094 	struct smc_sock *isk;
1095 	struct sock *sk;
1096 
1097 	lock_sock(parent);
1098 	list_for_each_entry(isk, &smc_sk(parent)->accept_q, accept_q) {
1099 		sk = (struct sock *)isk;
1100 
1101 		if (sk->sk_state == SMC_ACTIVE) {
1102 			release_sock(parent);
1103 			return POLLIN | POLLRDNORM;
1104 		}
1105 	}
1106 	release_sock(parent);
1107 
1108 	return 0;
1109 }
1110 
1111 static unsigned int smc_poll(struct file *file, struct socket *sock,
1112 			     poll_table *wait)
1113 {
1114 	struct sock *sk = sock->sk;
1115 	unsigned int mask = 0;
1116 	struct smc_sock *smc;
1117 	int rc;
1118 
1119 	smc = smc_sk(sock->sk);
1120 	if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
1121 		/* delegate to CLC child sock */
1122 		mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1123 		/* if non-blocking connect finished ... */
1124 		lock_sock(sk);
1125 		if ((sk->sk_state == SMC_INIT) && (mask & POLLOUT)) {
1126 			sk->sk_err = smc->clcsock->sk->sk_err;
1127 			if (sk->sk_err) {
1128 				mask |= POLLERR;
1129 			} else {
1130 				rc = smc_connect_rdma(smc);
1131 				if (rc < 0)
1132 					mask |= POLLERR;
1133 				else
1134 					/* success cases including fallback */
1135 					mask |= POLLOUT | POLLWRNORM;
1136 			}
1137 		}
1138 		release_sock(sk);
1139 	} else {
1140 		sock_poll_wait(file, sk_sleep(sk), wait);
1141 		if (sk->sk_state == SMC_LISTEN)
1142 			/* woken up by sk_data_ready in smc_listen_work() */
1143 			mask |= smc_accept_poll(sk);
1144 		if (sk->sk_err)
1145 			mask |= POLLERR;
1146 		if (atomic_read(&smc->conn.sndbuf_space) ||
1147 		    (sk->sk_shutdown & SEND_SHUTDOWN)) {
1148 			mask |= POLLOUT | POLLWRNORM;
1149 		} else {
1150 			sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1151 			set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1152 		}
1153 		if (atomic_read(&smc->conn.bytes_to_rcv))
1154 			mask |= POLLIN | POLLRDNORM;
1155 		if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1156 		    (sk->sk_state == SMC_CLOSED))
1157 			mask |= POLLHUP;
1158 		if (sk->sk_shutdown & RCV_SHUTDOWN)
1159 			mask |= POLLIN | POLLRDNORM | POLLRDHUP;
1160 		if (sk->sk_state == SMC_APPCLOSEWAIT1)
1161 			mask |= POLLIN;
1162 
1163 	}
1164 
1165 	return mask;
1166 }
1167 
1168 static int smc_shutdown(struct socket *sock, int how)
1169 {
1170 	struct sock *sk = sock->sk;
1171 	struct smc_sock *smc;
1172 	int rc = -EINVAL;
1173 	int rc1 = 0;
1174 
1175 	smc = smc_sk(sk);
1176 
1177 	if ((how < SHUT_RD) || (how > SHUT_RDWR))
1178 		return rc;
1179 
1180 	lock_sock(sk);
1181 
1182 	rc = -ENOTCONN;
1183 	if ((sk->sk_state != SMC_LISTEN) &&
1184 	    (sk->sk_state != SMC_ACTIVE) &&
1185 	    (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1186 	    (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1187 	    (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1188 	    (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1189 	    (sk->sk_state != SMC_APPFINCLOSEWAIT))
1190 		goto out;
1191 	if (smc->use_fallback) {
1192 		rc = kernel_sock_shutdown(smc->clcsock, how);
1193 		sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1194 		if (sk->sk_shutdown == SHUTDOWN_MASK)
1195 			sk->sk_state = SMC_CLOSED;
1196 		goto out;
1197 	}
1198 	switch (how) {
1199 	case SHUT_RDWR:		/* shutdown in both directions */
1200 		rc = smc_close_active(smc);
1201 		break;
1202 	case SHUT_WR:
1203 		rc = smc_close_shutdown_write(smc);
1204 		break;
1205 	case SHUT_RD:
1206 		if (sk->sk_state == SMC_LISTEN)
1207 			rc = smc_close_active(smc);
1208 		else
1209 			rc = 0;
1210 			/* nothing more to do because peer is not involved */
1211 		break;
1212 	}
1213 	rc1 = kernel_sock_shutdown(smc->clcsock, how);
1214 	/* map sock_shutdown_cmd constants to sk_shutdown value range */
1215 	sk->sk_shutdown |= how + 1;
1216 
1217 out:
1218 	release_sock(sk);
1219 	return rc ? rc : rc1;
1220 }
1221 
1222 static int smc_setsockopt(struct socket *sock, int level, int optname,
1223 			  char __user *optval, unsigned int optlen)
1224 {
1225 	struct sock *sk = sock->sk;
1226 	struct smc_sock *smc;
1227 
1228 	smc = smc_sk(sk);
1229 
1230 	/* generic setsockopts reaching us here always apply to the
1231 	 * CLC socket
1232 	 */
1233 	return smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1234 					     optval, optlen);
1235 }
1236 
1237 static int smc_getsockopt(struct socket *sock, int level, int optname,
1238 			  char __user *optval, int __user *optlen)
1239 {
1240 	struct smc_sock *smc;
1241 
1242 	smc = smc_sk(sock->sk);
1243 	/* socket options apply to the CLC socket */
1244 	return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1245 					     optval, optlen);
1246 }
1247 
1248 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1249 		     unsigned long arg)
1250 {
1251 	struct smc_sock *smc;
1252 
1253 	smc = smc_sk(sock->sk);
1254 	if (smc->use_fallback)
1255 		return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1256 	else
1257 		return sock_no_ioctl(sock, cmd, arg);
1258 }
1259 
1260 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1261 			    int offset, size_t size, int flags)
1262 {
1263 	struct sock *sk = sock->sk;
1264 	struct smc_sock *smc;
1265 	int rc = -EPIPE;
1266 
1267 	smc = smc_sk(sk);
1268 	lock_sock(sk);
1269 	if (sk->sk_state != SMC_ACTIVE)
1270 		goto out;
1271 	if (smc->use_fallback)
1272 		rc = kernel_sendpage(smc->clcsock, page, offset,
1273 				     size, flags);
1274 	else
1275 		rc = sock_no_sendpage(sock, page, offset, size, flags);
1276 
1277 out:
1278 	release_sock(sk);
1279 	return rc;
1280 }
1281 
1282 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1283 			       struct pipe_inode_info *pipe, size_t len,
1284 				    unsigned int flags)
1285 {
1286 	struct sock *sk = sock->sk;
1287 	struct smc_sock *smc;
1288 	int rc = -ENOTCONN;
1289 
1290 	smc = smc_sk(sk);
1291 	lock_sock(sk);
1292 	if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
1293 		goto out;
1294 	if (smc->use_fallback) {
1295 		rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1296 						    pipe, len, flags);
1297 	} else {
1298 		rc = -EOPNOTSUPP;
1299 	}
1300 out:
1301 	release_sock(sk);
1302 	return rc;
1303 }
1304 
1305 /* must look like tcp */
1306 static const struct proto_ops smc_sock_ops = {
1307 	.family		= PF_SMC,
1308 	.owner		= THIS_MODULE,
1309 	.release	= smc_release,
1310 	.bind		= smc_bind,
1311 	.connect	= smc_connect,
1312 	.socketpair	= sock_no_socketpair,
1313 	.accept		= smc_accept,
1314 	.getname	= smc_getname,
1315 	.poll		= smc_poll,
1316 	.ioctl		= smc_ioctl,
1317 	.listen		= smc_listen,
1318 	.shutdown	= smc_shutdown,
1319 	.setsockopt	= smc_setsockopt,
1320 	.getsockopt	= smc_getsockopt,
1321 	.sendmsg	= smc_sendmsg,
1322 	.recvmsg	= smc_recvmsg,
1323 	.mmap		= sock_no_mmap,
1324 	.sendpage	= smc_sendpage,
1325 	.splice_read	= smc_splice_read,
1326 };
1327 
1328 static int smc_create(struct net *net, struct socket *sock, int protocol,
1329 		      int kern)
1330 {
1331 	struct smc_sock *smc;
1332 	struct sock *sk;
1333 	int rc;
1334 
1335 	rc = -ESOCKTNOSUPPORT;
1336 	if (sock->type != SOCK_STREAM)
1337 		goto out;
1338 
1339 	rc = -EPROTONOSUPPORT;
1340 	if ((protocol != IPPROTO_IP) && (protocol != IPPROTO_TCP))
1341 		goto out;
1342 
1343 	rc = -ENOBUFS;
1344 	sock->ops = &smc_sock_ops;
1345 	sk = smc_sock_alloc(net, sock);
1346 	if (!sk)
1347 		goto out;
1348 
1349 	/* create internal TCP socket for CLC handshake and fallback */
1350 	smc = smc_sk(sk);
1351 	smc->use_fallback = false; /* assume rdma capability first */
1352 	rc = sock_create_kern(net, PF_INET, SOCK_STREAM,
1353 			      IPPROTO_TCP, &smc->clcsock);
1354 	if (rc)
1355 		sk_common_release(sk);
1356 	smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1357 	smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1358 
1359 out:
1360 	return rc;
1361 }
1362 
1363 static const struct net_proto_family smc_sock_family_ops = {
1364 	.family	= PF_SMC,
1365 	.owner	= THIS_MODULE,
1366 	.create	= smc_create,
1367 };
1368 
1369 static int __init smc_init(void)
1370 {
1371 	int rc;
1372 
1373 	rc = smc_pnet_init();
1374 	if (rc)
1375 		return rc;
1376 
1377 	rc = smc_llc_init();
1378 	if (rc) {
1379 		pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
1380 		goto out_pnet;
1381 	}
1382 
1383 	rc = smc_cdc_init();
1384 	if (rc) {
1385 		pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
1386 		goto out_pnet;
1387 	}
1388 
1389 	rc = proto_register(&smc_proto, 1);
1390 	if (rc) {
1391 		pr_err("%s: proto_register fails with %d\n", __func__, rc);
1392 		goto out_pnet;
1393 	}
1394 
1395 	rc = sock_register(&smc_sock_family_ops);
1396 	if (rc) {
1397 		pr_err("%s: sock_register fails with %d\n", __func__, rc);
1398 		goto out_proto;
1399 	}
1400 	INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
1401 
1402 	rc = smc_ib_register_client();
1403 	if (rc) {
1404 		pr_err("%s: ib_register fails with %d\n", __func__, rc);
1405 		goto out_sock;
1406 	}
1407 
1408 	return 0;
1409 
1410 out_sock:
1411 	sock_unregister(PF_SMC);
1412 out_proto:
1413 	proto_unregister(&smc_proto);
1414 out_pnet:
1415 	smc_pnet_exit();
1416 	return rc;
1417 }
1418 
1419 static void __exit smc_exit(void)
1420 {
1421 	struct smc_link_group *lgr, *lg;
1422 	LIST_HEAD(lgr_freeing_list);
1423 
1424 	spin_lock_bh(&smc_lgr_list.lock);
1425 	if (!list_empty(&smc_lgr_list.list))
1426 		list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1427 	spin_unlock_bh(&smc_lgr_list.lock);
1428 	list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1429 		list_del_init(&lgr->list);
1430 		smc_lgr_free(lgr); /* free link group */
1431 	}
1432 	smc_ib_unregister_client();
1433 	sock_unregister(PF_SMC);
1434 	proto_unregister(&smc_proto);
1435 	smc_pnet_exit();
1436 }
1437 
1438 module_init(smc_init);
1439 module_exit(smc_exit);
1440 
1441 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1442 MODULE_DESCRIPTION("smc socket address family");
1443 MODULE_LICENSE("GPL");
1444 MODULE_ALIAS_NETPROTO(PF_SMC);
1445