xref: /openbmc/linux/net/smc/af_smc.c (revision 77ab8d5d)
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  *    - support for alternate links postponed
11  *
12  *  Copyright IBM Corp. 2016, 2018
13  *
14  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
15  *              based on prototype from Frank Blaschka
16  */
17 
18 #define KMSG_COMPONENT "smc"
19 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
20 
21 #include <linux/module.h>
22 #include <linux/socket.h>
23 #include <linux/workqueue.h>
24 #include <linux/in.h>
25 #include <linux/sched/signal.h>
26 
27 #include <net/sock.h>
28 #include <net/tcp.h>
29 #include <net/smc.h>
30 #include <asm/ioctls.h>
31 
32 #include "smc.h"
33 #include "smc_clc.h"
34 #include "smc_llc.h"
35 #include "smc_cdc.h"
36 #include "smc_core.h"
37 #include "smc_ib.h"
38 #include "smc_pnet.h"
39 #include "smc_tx.h"
40 #include "smc_rx.h"
41 #include "smc_close.h"
42 
43 static DEFINE_MUTEX(smc_create_lgr_pending);	/* serialize link group
44 						 * creation
45 						 */
46 
47 static void smc_tcp_listen_work(struct work_struct *);
48 
49 static void smc_set_keepalive(struct sock *sk, int val)
50 {
51 	struct smc_sock *smc = smc_sk(sk);
52 
53 	smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
54 }
55 
56 static struct smc_hashinfo smc_v4_hashinfo = {
57 	.lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
58 };
59 
60 static struct smc_hashinfo smc_v6_hashinfo = {
61 	.lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
62 };
63 
64 int smc_hash_sk(struct sock *sk)
65 {
66 	struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
67 	struct hlist_head *head;
68 
69 	head = &h->ht;
70 
71 	write_lock_bh(&h->lock);
72 	sk_add_node(sk, head);
73 	sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
74 	write_unlock_bh(&h->lock);
75 
76 	return 0;
77 }
78 EXPORT_SYMBOL_GPL(smc_hash_sk);
79 
80 void smc_unhash_sk(struct sock *sk)
81 {
82 	struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
83 
84 	write_lock_bh(&h->lock);
85 	if (sk_del_node_init(sk))
86 		sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
87 	write_unlock_bh(&h->lock);
88 }
89 EXPORT_SYMBOL_GPL(smc_unhash_sk);
90 
91 struct proto smc_proto = {
92 	.name		= "SMC",
93 	.owner		= THIS_MODULE,
94 	.keepalive	= smc_set_keepalive,
95 	.hash		= smc_hash_sk,
96 	.unhash		= smc_unhash_sk,
97 	.obj_size	= sizeof(struct smc_sock),
98 	.h.smc_hash	= &smc_v4_hashinfo,
99 	.slab_flags	= SLAB_TYPESAFE_BY_RCU,
100 };
101 EXPORT_SYMBOL_GPL(smc_proto);
102 
103 struct proto smc_proto6 = {
104 	.name		= "SMC6",
105 	.owner		= THIS_MODULE,
106 	.keepalive	= smc_set_keepalive,
107 	.hash		= smc_hash_sk,
108 	.unhash		= smc_unhash_sk,
109 	.obj_size	= sizeof(struct smc_sock),
110 	.h.smc_hash	= &smc_v6_hashinfo,
111 	.slab_flags	= SLAB_TYPESAFE_BY_RCU,
112 };
113 EXPORT_SYMBOL_GPL(smc_proto6);
114 
115 static int smc_release(struct socket *sock)
116 {
117 	struct sock *sk = sock->sk;
118 	struct smc_sock *smc;
119 	int rc = 0;
120 
121 	if (!sk)
122 		goto out;
123 
124 	smc = smc_sk(sk);
125 	if (sk->sk_state == SMC_LISTEN)
126 		/* smc_close_non_accepted() is called and acquires
127 		 * sock lock for child sockets again
128 		 */
129 		lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
130 	else
131 		lock_sock(sk);
132 
133 	if (!smc->use_fallback) {
134 		rc = smc_close_active(smc);
135 		sock_set_flag(sk, SOCK_DEAD);
136 		sk->sk_shutdown |= SHUTDOWN_MASK;
137 	}
138 	if (smc->clcsock) {
139 		sock_release(smc->clcsock);
140 		smc->clcsock = NULL;
141 	}
142 	if (smc->use_fallback) {
143 		sock_put(sk); /* passive closing */
144 		sk->sk_state = SMC_CLOSED;
145 		sk->sk_state_change(sk);
146 	}
147 
148 	/* detach socket */
149 	sock_orphan(sk);
150 	sock->sk = NULL;
151 	if (!smc->use_fallback && sk->sk_state == SMC_CLOSED)
152 		smc_conn_free(&smc->conn);
153 	release_sock(sk);
154 
155 	sk->sk_prot->unhash(sk);
156 	sock_put(sk); /* final sock_put */
157 out:
158 	return rc;
159 }
160 
161 static void smc_destruct(struct sock *sk)
162 {
163 	if (sk->sk_state != SMC_CLOSED)
164 		return;
165 	if (!sock_flag(sk, SOCK_DEAD))
166 		return;
167 
168 	sk_refcnt_debug_dec(sk);
169 }
170 
171 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
172 				   int protocol)
173 {
174 	struct smc_sock *smc;
175 	struct proto *prot;
176 	struct sock *sk;
177 
178 	prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
179 	sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
180 	if (!sk)
181 		return NULL;
182 
183 	sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
184 	sk->sk_state = SMC_INIT;
185 	sk->sk_destruct = smc_destruct;
186 	sk->sk_protocol = protocol;
187 	smc = smc_sk(sk);
188 	INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
189 	INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
190 	INIT_LIST_HEAD(&smc->accept_q);
191 	spin_lock_init(&smc->accept_q_lock);
192 	spin_lock_init(&smc->conn.send_lock);
193 	sk->sk_prot->hash(sk);
194 	sk_refcnt_debug_inc(sk);
195 
196 	return sk;
197 }
198 
199 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
200 		    int addr_len)
201 {
202 	struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
203 	struct sock *sk = sock->sk;
204 	struct smc_sock *smc;
205 	int rc;
206 
207 	smc = smc_sk(sk);
208 
209 	/* replicate tests from inet_bind(), to be safe wrt. future changes */
210 	rc = -EINVAL;
211 	if (addr_len < sizeof(struct sockaddr_in))
212 		goto out;
213 
214 	rc = -EAFNOSUPPORT;
215 	if (addr->sin_family != AF_INET &&
216 	    addr->sin_family != AF_INET6 &&
217 	    addr->sin_family != AF_UNSPEC)
218 		goto out;
219 	/* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
220 	if (addr->sin_family == AF_UNSPEC &&
221 	    addr->sin_addr.s_addr != htonl(INADDR_ANY))
222 		goto out;
223 
224 	lock_sock(sk);
225 
226 	/* Check if socket is already active */
227 	rc = -EINVAL;
228 	if (sk->sk_state != SMC_INIT)
229 		goto out_rel;
230 
231 	smc->clcsock->sk->sk_reuse = sk->sk_reuse;
232 	rc = kernel_bind(smc->clcsock, uaddr, addr_len);
233 
234 out_rel:
235 	release_sock(sk);
236 out:
237 	return rc;
238 }
239 
240 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
241 				   unsigned long mask)
242 {
243 	/* options we don't get control via setsockopt for */
244 	nsk->sk_type = osk->sk_type;
245 	nsk->sk_sndbuf = osk->sk_sndbuf;
246 	nsk->sk_rcvbuf = osk->sk_rcvbuf;
247 	nsk->sk_sndtimeo = osk->sk_sndtimeo;
248 	nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
249 	nsk->sk_mark = osk->sk_mark;
250 	nsk->sk_priority = osk->sk_priority;
251 	nsk->sk_rcvlowat = osk->sk_rcvlowat;
252 	nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
253 	nsk->sk_err = osk->sk_err;
254 
255 	nsk->sk_flags &= ~mask;
256 	nsk->sk_flags |= osk->sk_flags & mask;
257 }
258 
259 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
260 			     (1UL << SOCK_KEEPOPEN) | \
261 			     (1UL << SOCK_LINGER) | \
262 			     (1UL << SOCK_BROADCAST) | \
263 			     (1UL << SOCK_TIMESTAMP) | \
264 			     (1UL << SOCK_DBG) | \
265 			     (1UL << SOCK_RCVTSTAMP) | \
266 			     (1UL << SOCK_RCVTSTAMPNS) | \
267 			     (1UL << SOCK_LOCALROUTE) | \
268 			     (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
269 			     (1UL << SOCK_RXQ_OVFL) | \
270 			     (1UL << SOCK_WIFI_STATUS) | \
271 			     (1UL << SOCK_NOFCS) | \
272 			     (1UL << SOCK_FILTER_LOCKED))
273 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
274  * clc socket (since smc is not called for these options from net/core)
275  */
276 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
277 {
278 	smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
279 }
280 
281 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
282 			     (1UL << SOCK_KEEPOPEN) | \
283 			     (1UL << SOCK_LINGER) | \
284 			     (1UL << SOCK_DBG))
285 /* copy only settings and flags relevant for smc from clc to smc socket */
286 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
287 {
288 	smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
289 }
290 
291 /* register a new rmb, optionally send confirm_rkey msg to register with peer */
292 static int smc_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc,
293 		       bool conf_rkey)
294 {
295 	/* register memory region for new rmb */
296 	if (smc_wr_reg_send(link, rmb_desc->mr_rx[SMC_SINGLE_LINK])) {
297 		rmb_desc->regerr = 1;
298 		return -EFAULT;
299 	}
300 	if (!conf_rkey)
301 		return 0;
302 	/* exchange confirm_rkey msg with peer */
303 	if (smc_llc_do_confirm_rkey(link, rmb_desc)) {
304 		rmb_desc->regerr = 1;
305 		return -EFAULT;
306 	}
307 	return 0;
308 }
309 
310 static int smc_clnt_conf_first_link(struct smc_sock *smc)
311 {
312 	struct net *net = sock_net(smc->clcsock->sk);
313 	struct smc_link_group *lgr = smc->conn.lgr;
314 	struct smc_link *link;
315 	int rest;
316 	int rc;
317 
318 	link = &lgr->lnk[SMC_SINGLE_LINK];
319 	/* receive CONFIRM LINK request from server over RoCE fabric */
320 	rest = wait_for_completion_interruptible_timeout(
321 		&link->llc_confirm,
322 		SMC_LLC_WAIT_FIRST_TIME);
323 	if (rest <= 0) {
324 		struct smc_clc_msg_decline dclc;
325 
326 		rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
327 				      SMC_CLC_DECLINE);
328 		return rc;
329 	}
330 
331 	if (link->llc_confirm_rc)
332 		return SMC_CLC_DECL_RMBE_EC;
333 
334 	rc = smc_ib_modify_qp_rts(link);
335 	if (rc)
336 		return SMC_CLC_DECL_INTERR;
337 
338 	smc_wr_remember_qp_attr(link);
339 
340 	if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
341 		return SMC_CLC_DECL_INTERR;
342 
343 	/* send CONFIRM LINK response over RoCE fabric */
344 	rc = smc_llc_send_confirm_link(link,
345 				       link->smcibdev->mac[link->ibport - 1],
346 				       &link->smcibdev->gid[link->ibport - 1],
347 				       SMC_LLC_RESP);
348 	if (rc < 0)
349 		return SMC_CLC_DECL_TCL;
350 
351 	/* receive ADD LINK request from server over RoCE fabric */
352 	rest = wait_for_completion_interruptible_timeout(&link->llc_add,
353 							 SMC_LLC_WAIT_TIME);
354 	if (rest <= 0) {
355 		struct smc_clc_msg_decline dclc;
356 
357 		rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
358 				      SMC_CLC_DECLINE);
359 		return rc;
360 	}
361 
362 	/* send add link reject message, only one link supported for now */
363 	rc = smc_llc_send_add_link(link,
364 				   link->smcibdev->mac[link->ibport - 1],
365 				   &link->smcibdev->gid[link->ibport - 1],
366 				   SMC_LLC_RESP);
367 	if (rc < 0)
368 		return SMC_CLC_DECL_TCL;
369 
370 	smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
371 
372 	return 0;
373 }
374 
375 static void smc_conn_save_peer_info(struct smc_sock *smc,
376 				    struct smc_clc_msg_accept_confirm *clc)
377 {
378 	int bufsize = smc_uncompress_bufsize(clc->rmbe_size);
379 
380 	smc->conn.peer_rmbe_idx = clc->rmbe_idx;
381 	smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
382 	smc->conn.peer_rmbe_size = bufsize;
383 	atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
384 	smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
385 }
386 
387 static void smc_link_save_peer_info(struct smc_link *link,
388 				    struct smc_clc_msg_accept_confirm *clc)
389 {
390 	link->peer_qpn = ntoh24(clc->qpn);
391 	memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
392 	memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
393 	link->peer_psn = ntoh24(clc->psn);
394 	link->peer_mtu = clc->qp_mtu;
395 }
396 
397 /* fall back during connect */
398 static int smc_connect_fallback(struct smc_sock *smc)
399 {
400 	smc->use_fallback = true;
401 	smc_copy_sock_settings_to_clc(smc);
402 	if (smc->sk.sk_state == SMC_INIT)
403 		smc->sk.sk_state = SMC_ACTIVE;
404 	return 0;
405 }
406 
407 /* decline and fall back during connect */
408 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
409 {
410 	int rc;
411 
412 	if (reason_code < 0) /* error, fallback is not possible */
413 		return reason_code;
414 	if (reason_code != SMC_CLC_DECL_REPLY) {
415 		rc = smc_clc_send_decline(smc, reason_code);
416 		if (rc < 0)
417 			return rc;
418 	}
419 	return smc_connect_fallback(smc);
420 }
421 
422 /* abort connecting */
423 static int smc_connect_abort(struct smc_sock *smc, int reason_code,
424 			     int local_contact)
425 {
426 	if (local_contact == SMC_FIRST_CONTACT)
427 		smc_lgr_forget(smc->conn.lgr);
428 	mutex_unlock(&smc_create_lgr_pending);
429 	smc_conn_free(&smc->conn);
430 	if (reason_code < 0 && smc->sk.sk_state == SMC_INIT)
431 		sock_put(&smc->sk); /* passive closing */
432 	return reason_code;
433 }
434 
435 /* check if there is a rdma device available for this connection. */
436 /* called for connect and listen */
437 static int smc_check_rdma(struct smc_sock *smc, struct smc_ib_device **ibdev,
438 			  u8 *ibport)
439 {
440 	int reason_code = 0;
441 
442 	/* PNET table look up: search active ib_device and port
443 	 * within same PNETID that also contains the ethernet device
444 	 * used for the internal TCP socket
445 	 */
446 	smc_pnet_find_roce_resource(smc->clcsock->sk, ibdev, ibport);
447 	if (!(*ibdev))
448 		reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
449 
450 	return reason_code;
451 }
452 
453 /* CLC handshake during connect */
454 static int smc_connect_clc(struct smc_sock *smc,
455 			   struct smc_clc_msg_accept_confirm *aclc,
456 			   struct smc_ib_device *ibdev, u8 ibport)
457 {
458 	int rc = 0;
459 
460 	/* do inband token exchange */
461 	rc = smc_clc_send_proposal(smc, ibdev, ibport);
462 	if (rc)
463 		return rc;
464 	/* receive SMC Accept CLC message */
465 	return smc_clc_wait_msg(smc, aclc, sizeof(*aclc), SMC_CLC_ACCEPT);
466 }
467 
468 /* setup for RDMA connection of client */
469 static int smc_connect_rdma(struct smc_sock *smc,
470 			    struct smc_clc_msg_accept_confirm *aclc,
471 			    struct smc_ib_device *ibdev, u8 ibport)
472 {
473 	int local_contact = SMC_FIRST_CONTACT;
474 	struct smc_link *link;
475 	int reason_code = 0;
476 
477 	mutex_lock(&smc_create_lgr_pending);
478 	local_contact = smc_conn_create(smc, ibdev, ibport, &aclc->lcl,
479 					aclc->hdr.flag);
480 	if (local_contact < 0) {
481 		if (local_contact == -ENOMEM)
482 			reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
483 		else if (local_contact == -ENOLINK)
484 			reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
485 		else
486 			reason_code = SMC_CLC_DECL_INTERR; /* other error */
487 		return smc_connect_abort(smc, reason_code, 0);
488 	}
489 	link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
490 
491 	smc_conn_save_peer_info(smc, aclc);
492 
493 	/* create send buffer and rmb */
494 	if (smc_buf_create(smc))
495 		return smc_connect_abort(smc, SMC_CLC_DECL_MEM, local_contact);
496 
497 	if (local_contact == SMC_FIRST_CONTACT)
498 		smc_link_save_peer_info(link, aclc);
499 
500 	if (smc_rmb_rtoken_handling(&smc->conn, aclc))
501 		return smc_connect_abort(smc, SMC_CLC_DECL_INTERR,
502 					 local_contact);
503 
504 	smc_close_init(smc);
505 	smc_rx_init(smc);
506 
507 	if (local_contact == SMC_FIRST_CONTACT) {
508 		if (smc_ib_ready_link(link))
509 			return smc_connect_abort(smc, SMC_CLC_DECL_INTERR,
510 						 local_contact);
511 	} else {
512 		if (!smc->conn.rmb_desc->reused &&
513 		    smc_reg_rmb(link, smc->conn.rmb_desc, true))
514 			return smc_connect_abort(smc, SMC_CLC_DECL_INTERR,
515 						 local_contact);
516 	}
517 	smc_rmb_sync_sg_for_device(&smc->conn);
518 
519 	reason_code = smc_clc_send_confirm(smc);
520 	if (reason_code)
521 		return smc_connect_abort(smc, reason_code, local_contact);
522 
523 	smc_tx_init(smc);
524 
525 	if (local_contact == SMC_FIRST_CONTACT) {
526 		/* QP confirmation over RoCE fabric */
527 		reason_code = smc_clnt_conf_first_link(smc);
528 		if (reason_code)
529 			return smc_connect_abort(smc, reason_code,
530 						 local_contact);
531 	}
532 	mutex_unlock(&smc_create_lgr_pending);
533 
534 	smc_copy_sock_settings_to_clc(smc);
535 	if (smc->sk.sk_state == SMC_INIT)
536 		smc->sk.sk_state = SMC_ACTIVE;
537 
538 	return 0;
539 }
540 
541 /* perform steps before actually connecting */
542 static int __smc_connect(struct smc_sock *smc)
543 {
544 	struct smc_clc_msg_accept_confirm aclc;
545 	struct smc_ib_device *ibdev;
546 	int rc = 0;
547 	u8 ibport;
548 
549 	sock_hold(&smc->sk); /* sock put in passive closing */
550 
551 	if (smc->use_fallback)
552 		return smc_connect_fallback(smc);
553 
554 	/* if peer has not signalled SMC-capability, fall back */
555 	if (!tcp_sk(smc->clcsock->sk)->syn_smc)
556 		return smc_connect_fallback(smc);
557 
558 	/* IPSec connections opt out of SMC-R optimizations */
559 	if (using_ipsec(smc))
560 		return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
561 
562 	/* check if a RDMA device is available; if not, fall back */
563 	if (smc_check_rdma(smc, &ibdev, &ibport))
564 		return smc_connect_decline_fallback(smc, SMC_CLC_DECL_CNFERR);
565 
566 	/* perform CLC handshake */
567 	rc = smc_connect_clc(smc, &aclc, ibdev, ibport);
568 	if (rc)
569 		return smc_connect_decline_fallback(smc, rc);
570 
571 	/* connect using rdma */
572 	rc = smc_connect_rdma(smc, &aclc, ibdev, ibport);
573 	if (rc)
574 		return smc_connect_decline_fallback(smc, rc);
575 
576 	return 0;
577 }
578 
579 static int smc_connect(struct socket *sock, struct sockaddr *addr,
580 		       int alen, int flags)
581 {
582 	struct sock *sk = sock->sk;
583 	struct smc_sock *smc;
584 	int rc = -EINVAL;
585 
586 	smc = smc_sk(sk);
587 
588 	/* separate smc parameter checking to be safe */
589 	if (alen < sizeof(addr->sa_family))
590 		goto out_err;
591 	if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
592 		goto out_err;
593 
594 	lock_sock(sk);
595 	switch (sk->sk_state) {
596 	default:
597 		goto out;
598 	case SMC_ACTIVE:
599 		rc = -EISCONN;
600 		goto out;
601 	case SMC_INIT:
602 		rc = 0;
603 		break;
604 	}
605 
606 	smc_copy_sock_settings_to_clc(smc);
607 	tcp_sk(smc->clcsock->sk)->syn_smc = 1;
608 	rc = kernel_connect(smc->clcsock, addr, alen, flags);
609 	if (rc)
610 		goto out;
611 
612 	rc = __smc_connect(smc);
613 	if (rc < 0)
614 		goto out;
615 	else
616 		rc = 0; /* success cases including fallback */
617 
618 out:
619 	release_sock(sk);
620 out_err:
621 	return rc;
622 }
623 
624 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
625 {
626 	struct socket *new_clcsock = NULL;
627 	struct sock *lsk = &lsmc->sk;
628 	struct sock *new_sk;
629 	int rc;
630 
631 	release_sock(lsk);
632 	new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
633 	if (!new_sk) {
634 		rc = -ENOMEM;
635 		lsk->sk_err = ENOMEM;
636 		*new_smc = NULL;
637 		lock_sock(lsk);
638 		goto out;
639 	}
640 	*new_smc = smc_sk(new_sk);
641 
642 	rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
643 	lock_sock(lsk);
644 	if  (rc < 0)
645 		lsk->sk_err = -rc;
646 	if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
647 		if (new_clcsock)
648 			sock_release(new_clcsock);
649 		new_sk->sk_state = SMC_CLOSED;
650 		sock_set_flag(new_sk, SOCK_DEAD);
651 		new_sk->sk_prot->unhash(new_sk);
652 		sock_put(new_sk); /* final */
653 		*new_smc = NULL;
654 		goto out;
655 	}
656 
657 	(*new_smc)->clcsock = new_clcsock;
658 out:
659 	return rc;
660 }
661 
662 /* add a just created sock to the accept queue of the listen sock as
663  * candidate for a following socket accept call from user space
664  */
665 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
666 {
667 	struct smc_sock *par = smc_sk(parent);
668 
669 	sock_hold(sk); /* sock_put in smc_accept_unlink () */
670 	spin_lock(&par->accept_q_lock);
671 	list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
672 	spin_unlock(&par->accept_q_lock);
673 	sk_acceptq_added(parent);
674 }
675 
676 /* remove a socket from the accept queue of its parental listening socket */
677 static void smc_accept_unlink(struct sock *sk)
678 {
679 	struct smc_sock *par = smc_sk(sk)->listen_smc;
680 
681 	spin_lock(&par->accept_q_lock);
682 	list_del_init(&smc_sk(sk)->accept_q);
683 	spin_unlock(&par->accept_q_lock);
684 	sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
685 	sock_put(sk); /* sock_hold in smc_accept_enqueue */
686 }
687 
688 /* remove a sock from the accept queue to bind it to a new socket created
689  * for a socket accept call from user space
690  */
691 struct sock *smc_accept_dequeue(struct sock *parent,
692 				struct socket *new_sock)
693 {
694 	struct smc_sock *isk, *n;
695 	struct sock *new_sk;
696 
697 	list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
698 		new_sk = (struct sock *)isk;
699 
700 		smc_accept_unlink(new_sk);
701 		if (new_sk->sk_state == SMC_CLOSED) {
702 			if (isk->clcsock) {
703 				sock_release(isk->clcsock);
704 				isk->clcsock = NULL;
705 			}
706 			new_sk->sk_prot->unhash(new_sk);
707 			sock_put(new_sk); /* final */
708 			continue;
709 		}
710 		if (new_sock)
711 			sock_graft(new_sk, new_sock);
712 		return new_sk;
713 	}
714 	return NULL;
715 }
716 
717 /* clean up for a created but never accepted sock */
718 void smc_close_non_accepted(struct sock *sk)
719 {
720 	struct smc_sock *smc = smc_sk(sk);
721 
722 	lock_sock(sk);
723 	if (!sk->sk_lingertime)
724 		/* wait for peer closing */
725 		sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
726 	if (!smc->use_fallback) {
727 		smc_close_active(smc);
728 		sock_set_flag(sk, SOCK_DEAD);
729 		sk->sk_shutdown |= SHUTDOWN_MASK;
730 	}
731 	if (smc->clcsock) {
732 		struct socket *tcp;
733 
734 		tcp = smc->clcsock;
735 		smc->clcsock = NULL;
736 		sock_release(tcp);
737 	}
738 	if (smc->use_fallback) {
739 		sock_put(sk); /* passive closing */
740 		sk->sk_state = SMC_CLOSED;
741 	} else {
742 		if (sk->sk_state == SMC_CLOSED)
743 			smc_conn_free(&smc->conn);
744 	}
745 	release_sock(sk);
746 	sk->sk_prot->unhash(sk);
747 	sock_put(sk); /* final sock_put */
748 }
749 
750 static int smc_serv_conf_first_link(struct smc_sock *smc)
751 {
752 	struct net *net = sock_net(smc->clcsock->sk);
753 	struct smc_link_group *lgr = smc->conn.lgr;
754 	struct smc_link *link;
755 	int rest;
756 	int rc;
757 
758 	link = &lgr->lnk[SMC_SINGLE_LINK];
759 
760 	if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
761 		return SMC_CLC_DECL_INTERR;
762 
763 	/* send CONFIRM LINK request to client over the RoCE fabric */
764 	rc = smc_llc_send_confirm_link(link,
765 				       link->smcibdev->mac[link->ibport - 1],
766 				       &link->smcibdev->gid[link->ibport - 1],
767 				       SMC_LLC_REQ);
768 	if (rc < 0)
769 		return SMC_CLC_DECL_TCL;
770 
771 	/* receive CONFIRM LINK response from client over the RoCE fabric */
772 	rest = wait_for_completion_interruptible_timeout(
773 		&link->llc_confirm_resp,
774 		SMC_LLC_WAIT_FIRST_TIME);
775 	if (rest <= 0) {
776 		struct smc_clc_msg_decline dclc;
777 
778 		rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
779 				      SMC_CLC_DECLINE);
780 		return rc;
781 	}
782 
783 	if (link->llc_confirm_resp_rc)
784 		return SMC_CLC_DECL_RMBE_EC;
785 
786 	/* send ADD LINK request to client over the RoCE fabric */
787 	rc = smc_llc_send_add_link(link,
788 				   link->smcibdev->mac[link->ibport - 1],
789 				   &link->smcibdev->gid[link->ibport - 1],
790 				   SMC_LLC_REQ);
791 	if (rc < 0)
792 		return SMC_CLC_DECL_TCL;
793 
794 	/* receive ADD LINK response from client over the RoCE fabric */
795 	rest = wait_for_completion_interruptible_timeout(&link->llc_add_resp,
796 							 SMC_LLC_WAIT_TIME);
797 	if (rest <= 0) {
798 		struct smc_clc_msg_decline dclc;
799 
800 		rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
801 				      SMC_CLC_DECLINE);
802 		return rc;
803 	}
804 
805 	smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
806 
807 	return 0;
808 }
809 
810 /* listen worker: finish */
811 static void smc_listen_out(struct smc_sock *new_smc)
812 {
813 	struct smc_sock *lsmc = new_smc->listen_smc;
814 	struct sock *newsmcsk = &new_smc->sk;
815 
816 	lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
817 	if (lsmc->sk.sk_state == SMC_LISTEN) {
818 		smc_accept_enqueue(&lsmc->sk, newsmcsk);
819 	} else { /* no longer listening */
820 		smc_close_non_accepted(newsmcsk);
821 	}
822 	release_sock(&lsmc->sk);
823 
824 	/* Wake up accept */
825 	lsmc->sk.sk_data_ready(&lsmc->sk);
826 	sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
827 }
828 
829 /* listen worker: finish in state connected */
830 static void smc_listen_out_connected(struct smc_sock *new_smc)
831 {
832 	struct sock *newsmcsk = &new_smc->sk;
833 
834 	sk_refcnt_debug_inc(newsmcsk);
835 	if (newsmcsk->sk_state == SMC_INIT)
836 		newsmcsk->sk_state = SMC_ACTIVE;
837 
838 	smc_listen_out(new_smc);
839 }
840 
841 /* listen worker: finish in error state */
842 static void smc_listen_out_err(struct smc_sock *new_smc)
843 {
844 	struct sock *newsmcsk = &new_smc->sk;
845 
846 	if (newsmcsk->sk_state == SMC_INIT)
847 		sock_put(&new_smc->sk); /* passive closing */
848 	newsmcsk->sk_state = SMC_CLOSED;
849 	smc_conn_free(&new_smc->conn);
850 
851 	smc_listen_out(new_smc);
852 }
853 
854 /* listen worker: decline and fall back if possible */
855 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
856 			       int local_contact)
857 {
858 	/* RDMA setup failed, switch back to TCP */
859 	if (local_contact == SMC_FIRST_CONTACT)
860 		smc_lgr_forget(new_smc->conn.lgr);
861 	if (reason_code < 0) { /* error, no fallback possible */
862 		smc_listen_out_err(new_smc);
863 		return;
864 	}
865 	smc_conn_free(&new_smc->conn);
866 	new_smc->use_fallback = true;
867 	if (reason_code && reason_code != SMC_CLC_DECL_REPLY) {
868 		if (smc_clc_send_decline(new_smc, reason_code) < 0) {
869 			smc_listen_out_err(new_smc);
870 			return;
871 		}
872 	}
873 	smc_listen_out_connected(new_smc);
874 }
875 
876 /* listen worker: check prefixes */
877 static int smc_listen_rdma_check(struct smc_sock *new_smc,
878 				 struct smc_clc_msg_proposal *pclc)
879 {
880 	struct smc_clc_msg_proposal_prefix *pclc_prfx;
881 	struct socket *newclcsock = new_smc->clcsock;
882 
883 	pclc_prfx = smc_clc_proposal_get_prefix(pclc);
884 	if (smc_clc_prfx_match(newclcsock, pclc_prfx))
885 		return SMC_CLC_DECL_CNFERR;
886 
887 	return 0;
888 }
889 
890 /* listen worker: initialize connection and buffers */
891 static int smc_listen_rdma_init(struct smc_sock *new_smc,
892 				struct smc_clc_msg_proposal *pclc,
893 				struct smc_ib_device *ibdev, u8 ibport,
894 				int *local_contact)
895 {
896 	/* allocate connection / link group */
897 	*local_contact = smc_conn_create(new_smc, ibdev, ibport, &pclc->lcl, 0);
898 	if (*local_contact < 0) {
899 		if (*local_contact == -ENOMEM)
900 			return SMC_CLC_DECL_MEM;/* insufficient memory*/
901 		return SMC_CLC_DECL_INTERR; /* other error */
902 	}
903 
904 	/* create send buffer and rmb */
905 	if (smc_buf_create(new_smc))
906 		return SMC_CLC_DECL_MEM;
907 
908 	return 0;
909 }
910 
911 /* listen worker: register buffers */
912 static int smc_listen_rdma_reg(struct smc_sock *new_smc, int local_contact)
913 {
914 	struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
915 
916 	if (local_contact != SMC_FIRST_CONTACT) {
917 		if (!new_smc->conn.rmb_desc->reused) {
918 			if (smc_reg_rmb(link, new_smc->conn.rmb_desc, true))
919 				return SMC_CLC_DECL_INTERR;
920 		}
921 	}
922 	smc_rmb_sync_sg_for_device(&new_smc->conn);
923 
924 	return 0;
925 }
926 
927 /* listen worker: finish RDMA setup */
928 static void smc_listen_rdma_finish(struct smc_sock *new_smc,
929 				   struct smc_clc_msg_accept_confirm *cclc,
930 				   int local_contact)
931 {
932 	struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
933 	int reason_code = 0;
934 
935 	if (local_contact == SMC_FIRST_CONTACT)
936 		smc_link_save_peer_info(link, cclc);
937 
938 	if (smc_rmb_rtoken_handling(&new_smc->conn, cclc)) {
939 		reason_code = SMC_CLC_DECL_INTERR;
940 		goto decline;
941 	}
942 
943 	if (local_contact == SMC_FIRST_CONTACT) {
944 		if (smc_ib_ready_link(link)) {
945 			reason_code = SMC_CLC_DECL_INTERR;
946 			goto decline;
947 		}
948 		/* QP confirmation over RoCE fabric */
949 		reason_code = smc_serv_conf_first_link(new_smc);
950 		if (reason_code)
951 			goto decline;
952 	}
953 	return;
954 
955 decline:
956 	mutex_unlock(&smc_create_lgr_pending);
957 	smc_listen_decline(new_smc, reason_code, local_contact);
958 }
959 
960 /* setup for RDMA connection of server */
961 static void smc_listen_work(struct work_struct *work)
962 {
963 	struct smc_sock *new_smc = container_of(work, struct smc_sock,
964 						smc_listen_work);
965 	struct socket *newclcsock = new_smc->clcsock;
966 	struct smc_clc_msg_accept_confirm cclc;
967 	struct smc_clc_msg_proposal *pclc;
968 	struct smc_ib_device *ibdev;
969 	u8 buf[SMC_CLC_MAX_LEN];
970 	int local_contact = 0;
971 	int reason_code = 0;
972 	int rc = 0;
973 	u8 ibport;
974 
975 	if (new_smc->use_fallback) {
976 		smc_listen_out_connected(new_smc);
977 		return;
978 	}
979 
980 	/* check if peer is smc capable */
981 	if (!tcp_sk(newclcsock->sk)->syn_smc) {
982 		new_smc->use_fallback = true;
983 		smc_listen_out_connected(new_smc);
984 		return;
985 	}
986 
987 	/* do inband token exchange -
988 	 * wait for and receive SMC Proposal CLC message
989 	 */
990 	pclc = (struct smc_clc_msg_proposal *)&buf;
991 	reason_code = smc_clc_wait_msg(new_smc, pclc, SMC_CLC_MAX_LEN,
992 				       SMC_CLC_PROPOSAL);
993 	if (reason_code) {
994 		smc_listen_decline(new_smc, reason_code, 0);
995 		return;
996 	}
997 
998 	/* IPSec connections opt out of SMC-R optimizations */
999 	if (using_ipsec(new_smc)) {
1000 		smc_listen_decline(new_smc, SMC_CLC_DECL_IPSEC, 0);
1001 		return;
1002 	}
1003 
1004 	mutex_lock(&smc_create_lgr_pending);
1005 	smc_close_init(new_smc);
1006 	smc_rx_init(new_smc);
1007 	smc_tx_init(new_smc);
1008 
1009 	/* check if RDMA is available */
1010 	if (smc_check_rdma(new_smc, &ibdev, &ibport) ||
1011 	    smc_listen_rdma_check(new_smc, pclc) ||
1012 	    smc_listen_rdma_init(new_smc, pclc, ibdev, ibport,
1013 				 &local_contact) ||
1014 	    smc_listen_rdma_reg(new_smc, local_contact)) {
1015 		/* SMC not supported, decline */
1016 		mutex_unlock(&smc_create_lgr_pending);
1017 		smc_listen_decline(new_smc, SMC_CLC_DECL_CNFERR, local_contact);
1018 		return;
1019 	}
1020 
1021 	/* send SMC Accept CLC message */
1022 	rc = smc_clc_send_accept(new_smc, local_contact);
1023 	if (rc) {
1024 		mutex_unlock(&smc_create_lgr_pending);
1025 		smc_listen_decline(new_smc, rc, local_contact);
1026 		return;
1027 	}
1028 
1029 	/* receive SMC Confirm CLC message */
1030 	reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
1031 				       SMC_CLC_CONFIRM);
1032 	if (reason_code) {
1033 		mutex_unlock(&smc_create_lgr_pending);
1034 		smc_listen_decline(new_smc, reason_code, local_contact);
1035 		return;
1036 	}
1037 
1038 	/* finish worker */
1039 	smc_listen_rdma_finish(new_smc, &cclc, local_contact);
1040 	smc_conn_save_peer_info(new_smc, &cclc);
1041 	mutex_unlock(&smc_create_lgr_pending);
1042 	smc_listen_out_connected(new_smc);
1043 }
1044 
1045 static void smc_tcp_listen_work(struct work_struct *work)
1046 {
1047 	struct smc_sock *lsmc = container_of(work, struct smc_sock,
1048 					     tcp_listen_work);
1049 	struct sock *lsk = &lsmc->sk;
1050 	struct smc_sock *new_smc;
1051 	int rc = 0;
1052 
1053 	lock_sock(lsk);
1054 	while (lsk->sk_state == SMC_LISTEN) {
1055 		rc = smc_clcsock_accept(lsmc, &new_smc);
1056 		if (rc)
1057 			goto out;
1058 		if (!new_smc)
1059 			continue;
1060 
1061 		new_smc->listen_smc = lsmc;
1062 		new_smc->use_fallback = lsmc->use_fallback;
1063 		sock_hold(lsk); /* sock_put in smc_listen_work */
1064 		INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1065 		smc_copy_sock_settings_to_smc(new_smc);
1066 		sock_hold(&new_smc->sk); /* sock_put in passive closing */
1067 		if (!schedule_work(&new_smc->smc_listen_work))
1068 			sock_put(&new_smc->sk);
1069 	}
1070 
1071 out:
1072 	release_sock(lsk);
1073 	sock_put(&lsmc->sk); /* sock_hold in smc_listen */
1074 }
1075 
1076 static int smc_listen(struct socket *sock, int backlog)
1077 {
1078 	struct sock *sk = sock->sk;
1079 	struct smc_sock *smc;
1080 	int rc;
1081 
1082 	smc = smc_sk(sk);
1083 	lock_sock(sk);
1084 
1085 	rc = -EINVAL;
1086 	if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
1087 		goto out;
1088 
1089 	rc = 0;
1090 	if (sk->sk_state == SMC_LISTEN) {
1091 		sk->sk_max_ack_backlog = backlog;
1092 		goto out;
1093 	}
1094 	/* some socket options are handled in core, so we could not apply
1095 	 * them to the clc socket -- copy smc socket options to clc socket
1096 	 */
1097 	smc_copy_sock_settings_to_clc(smc);
1098 	if (!smc->use_fallback)
1099 		tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1100 
1101 	rc = kernel_listen(smc->clcsock, backlog);
1102 	if (rc)
1103 		goto out;
1104 	sk->sk_max_ack_backlog = backlog;
1105 	sk->sk_ack_backlog = 0;
1106 	sk->sk_state = SMC_LISTEN;
1107 	INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
1108 	sock_hold(sk); /* sock_hold in tcp_listen_worker */
1109 	if (!schedule_work(&smc->tcp_listen_work))
1110 		sock_put(sk);
1111 
1112 out:
1113 	release_sock(sk);
1114 	return rc;
1115 }
1116 
1117 static int smc_accept(struct socket *sock, struct socket *new_sock,
1118 		      int flags, bool kern)
1119 {
1120 	struct sock *sk = sock->sk, *nsk;
1121 	DECLARE_WAITQUEUE(wait, current);
1122 	struct smc_sock *lsmc;
1123 	long timeo;
1124 	int rc = 0;
1125 
1126 	lsmc = smc_sk(sk);
1127 	sock_hold(sk); /* sock_put below */
1128 	lock_sock(sk);
1129 
1130 	if (lsmc->sk.sk_state != SMC_LISTEN) {
1131 		rc = -EINVAL;
1132 		release_sock(sk);
1133 		goto out;
1134 	}
1135 
1136 	/* Wait for an incoming connection */
1137 	timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1138 	add_wait_queue_exclusive(sk_sleep(sk), &wait);
1139 	while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1140 		set_current_state(TASK_INTERRUPTIBLE);
1141 		if (!timeo) {
1142 			rc = -EAGAIN;
1143 			break;
1144 		}
1145 		release_sock(sk);
1146 		timeo = schedule_timeout(timeo);
1147 		/* wakeup by sk_data_ready in smc_listen_work() */
1148 		sched_annotate_sleep();
1149 		lock_sock(sk);
1150 		if (signal_pending(current)) {
1151 			rc = sock_intr_errno(timeo);
1152 			break;
1153 		}
1154 	}
1155 	set_current_state(TASK_RUNNING);
1156 	remove_wait_queue(sk_sleep(sk), &wait);
1157 
1158 	if (!rc)
1159 		rc = sock_error(nsk);
1160 	release_sock(sk);
1161 	if (rc)
1162 		goto out;
1163 
1164 	if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1165 		/* wait till data arrives on the socket */
1166 		timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1167 								MSEC_PER_SEC);
1168 		if (smc_sk(nsk)->use_fallback) {
1169 			struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1170 
1171 			lock_sock(clcsk);
1172 			if (skb_queue_empty(&clcsk->sk_receive_queue))
1173 				sk_wait_data(clcsk, &timeo, NULL);
1174 			release_sock(clcsk);
1175 		} else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1176 			lock_sock(nsk);
1177 			smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1178 			release_sock(nsk);
1179 		}
1180 	}
1181 
1182 out:
1183 	sock_put(sk); /* sock_hold above */
1184 	return rc;
1185 }
1186 
1187 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1188 		       int peer)
1189 {
1190 	struct smc_sock *smc;
1191 
1192 	if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1193 	    (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1194 		return -ENOTCONN;
1195 
1196 	smc = smc_sk(sock->sk);
1197 
1198 	return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1199 }
1200 
1201 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1202 {
1203 	struct sock *sk = sock->sk;
1204 	struct smc_sock *smc;
1205 	int rc = -EPIPE;
1206 
1207 	smc = smc_sk(sk);
1208 	lock_sock(sk);
1209 	if ((sk->sk_state != SMC_ACTIVE) &&
1210 	    (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1211 	    (sk->sk_state != SMC_INIT))
1212 		goto out;
1213 
1214 	if (msg->msg_flags & MSG_FASTOPEN) {
1215 		if (sk->sk_state == SMC_INIT) {
1216 			smc->use_fallback = true;
1217 		} else {
1218 			rc = -EINVAL;
1219 			goto out;
1220 		}
1221 	}
1222 
1223 	if (smc->use_fallback)
1224 		rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1225 	else
1226 		rc = smc_tx_sendmsg(smc, msg, len);
1227 out:
1228 	release_sock(sk);
1229 	return rc;
1230 }
1231 
1232 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1233 		       int flags)
1234 {
1235 	struct sock *sk = sock->sk;
1236 	struct smc_sock *smc;
1237 	int rc = -ENOTCONN;
1238 
1239 	smc = smc_sk(sk);
1240 	lock_sock(sk);
1241 	if ((sk->sk_state == SMC_INIT) ||
1242 	    (sk->sk_state == SMC_LISTEN) ||
1243 	    (sk->sk_state == SMC_CLOSED))
1244 		goto out;
1245 
1246 	if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1247 		rc = 0;
1248 		goto out;
1249 	}
1250 
1251 	if (smc->use_fallback) {
1252 		rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1253 	} else {
1254 		msg->msg_namelen = 0;
1255 		rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1256 	}
1257 
1258 out:
1259 	release_sock(sk);
1260 	return rc;
1261 }
1262 
1263 static __poll_t smc_accept_poll(struct sock *parent)
1264 {
1265 	struct smc_sock *isk = smc_sk(parent);
1266 	__poll_t mask = 0;
1267 
1268 	spin_lock(&isk->accept_q_lock);
1269 	if (!list_empty(&isk->accept_q))
1270 		mask = EPOLLIN | EPOLLRDNORM;
1271 	spin_unlock(&isk->accept_q_lock);
1272 
1273 	return mask;
1274 }
1275 
1276 static __poll_t smc_poll(struct file *file, struct socket *sock,
1277 			     poll_table *wait)
1278 {
1279 	struct sock *sk = sock->sk;
1280 	__poll_t mask = 0;
1281 	struct smc_sock *smc;
1282 	int rc;
1283 
1284 	if (!sk)
1285 		return EPOLLNVAL;
1286 
1287 	smc = smc_sk(sock->sk);
1288 	sock_hold(sk);
1289 	lock_sock(sk);
1290 	if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
1291 		/* delegate to CLC child sock */
1292 		release_sock(sk);
1293 		mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1294 		lock_sock(sk);
1295 		sk->sk_err = smc->clcsock->sk->sk_err;
1296 		if (sk->sk_err) {
1297 			mask |= EPOLLERR;
1298 		} else {
1299 			/* if non-blocking connect finished ... */
1300 			if (sk->sk_state == SMC_INIT &&
1301 			    mask & EPOLLOUT &&
1302 			    smc->clcsock->sk->sk_state != TCP_CLOSE) {
1303 				rc = __smc_connect(smc);
1304 				if (rc < 0)
1305 					mask |= EPOLLERR;
1306 				/* success cases including fallback */
1307 				mask |= EPOLLOUT | EPOLLWRNORM;
1308 			}
1309 		}
1310 	} else {
1311 		if (sk->sk_state != SMC_CLOSED) {
1312 			release_sock(sk);
1313 			sock_poll_wait(file, sk_sleep(sk), wait);
1314 			lock_sock(sk);
1315 		}
1316 		if (sk->sk_err)
1317 			mask |= EPOLLERR;
1318 		if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1319 		    (sk->sk_state == SMC_CLOSED))
1320 			mask |= EPOLLHUP;
1321 		if (sk->sk_state == SMC_LISTEN) {
1322 			/* woken up by sk_data_ready in smc_listen_work() */
1323 			mask = smc_accept_poll(sk);
1324 		} else {
1325 			if (atomic_read(&smc->conn.sndbuf_space) ||
1326 			    sk->sk_shutdown & SEND_SHUTDOWN) {
1327 				mask |= EPOLLOUT | EPOLLWRNORM;
1328 			} else {
1329 				sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1330 				set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1331 			}
1332 			if (atomic_read(&smc->conn.bytes_to_rcv))
1333 				mask |= EPOLLIN | EPOLLRDNORM;
1334 			if (sk->sk_shutdown & RCV_SHUTDOWN)
1335 				mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1336 			if (sk->sk_state == SMC_APPCLOSEWAIT1)
1337 				mask |= EPOLLIN;
1338 		}
1339 		if (smc->conn.urg_state == SMC_URG_VALID)
1340 			mask |= EPOLLPRI;
1341 
1342 	}
1343 	release_sock(sk);
1344 	sock_put(sk);
1345 
1346 	return mask;
1347 }
1348 
1349 static int smc_shutdown(struct socket *sock, int how)
1350 {
1351 	struct sock *sk = sock->sk;
1352 	struct smc_sock *smc;
1353 	int rc = -EINVAL;
1354 	int rc1 = 0;
1355 
1356 	smc = smc_sk(sk);
1357 
1358 	if ((how < SHUT_RD) || (how > SHUT_RDWR))
1359 		return rc;
1360 
1361 	lock_sock(sk);
1362 
1363 	rc = -ENOTCONN;
1364 	if ((sk->sk_state != SMC_LISTEN) &&
1365 	    (sk->sk_state != SMC_ACTIVE) &&
1366 	    (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1367 	    (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1368 	    (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1369 	    (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1370 	    (sk->sk_state != SMC_APPFINCLOSEWAIT))
1371 		goto out;
1372 	if (smc->use_fallback) {
1373 		rc = kernel_sock_shutdown(smc->clcsock, how);
1374 		sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1375 		if (sk->sk_shutdown == SHUTDOWN_MASK)
1376 			sk->sk_state = SMC_CLOSED;
1377 		goto out;
1378 	}
1379 	switch (how) {
1380 	case SHUT_RDWR:		/* shutdown in both directions */
1381 		rc = smc_close_active(smc);
1382 		break;
1383 	case SHUT_WR:
1384 		rc = smc_close_shutdown_write(smc);
1385 		break;
1386 	case SHUT_RD:
1387 		rc = 0;
1388 		/* nothing more to do because peer is not involved */
1389 		break;
1390 	}
1391 	if (smc->clcsock)
1392 		rc1 = kernel_sock_shutdown(smc->clcsock, how);
1393 	/* map sock_shutdown_cmd constants to sk_shutdown value range */
1394 	sk->sk_shutdown |= how + 1;
1395 
1396 out:
1397 	release_sock(sk);
1398 	return rc ? rc : rc1;
1399 }
1400 
1401 static int smc_setsockopt(struct socket *sock, int level, int optname,
1402 			  char __user *optval, unsigned int optlen)
1403 {
1404 	struct sock *sk = sock->sk;
1405 	struct smc_sock *smc;
1406 	int val, rc;
1407 
1408 	smc = smc_sk(sk);
1409 
1410 	/* generic setsockopts reaching us here always apply to the
1411 	 * CLC socket
1412 	 */
1413 	rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1414 					   optval, optlen);
1415 	if (smc->clcsock->sk->sk_err) {
1416 		sk->sk_err = smc->clcsock->sk->sk_err;
1417 		sk->sk_error_report(sk);
1418 	}
1419 	if (rc)
1420 		return rc;
1421 
1422 	if (optlen < sizeof(int))
1423 		return rc;
1424 	get_user(val, (int __user *)optval);
1425 
1426 	lock_sock(sk);
1427 	switch (optname) {
1428 	case TCP_ULP:
1429 	case TCP_FASTOPEN:
1430 	case TCP_FASTOPEN_CONNECT:
1431 	case TCP_FASTOPEN_KEY:
1432 	case TCP_FASTOPEN_NO_COOKIE:
1433 		/* option not supported by SMC */
1434 		if (sk->sk_state == SMC_INIT) {
1435 			smc->use_fallback = true;
1436 		} else {
1437 			if (!smc->use_fallback)
1438 				rc = -EINVAL;
1439 		}
1440 		break;
1441 	case TCP_NODELAY:
1442 		if (sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) {
1443 			if (val && !smc->use_fallback)
1444 				mod_delayed_work(system_wq, &smc->conn.tx_work,
1445 						 0);
1446 		}
1447 		break;
1448 	case TCP_CORK:
1449 		if (sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) {
1450 			if (!val && !smc->use_fallback)
1451 				mod_delayed_work(system_wq, &smc->conn.tx_work,
1452 						 0);
1453 		}
1454 		break;
1455 	case TCP_DEFER_ACCEPT:
1456 		smc->sockopt_defer_accept = val;
1457 		break;
1458 	default:
1459 		break;
1460 	}
1461 	release_sock(sk);
1462 
1463 	return rc;
1464 }
1465 
1466 static int smc_getsockopt(struct socket *sock, int level, int optname,
1467 			  char __user *optval, int __user *optlen)
1468 {
1469 	struct smc_sock *smc;
1470 
1471 	smc = smc_sk(sock->sk);
1472 	/* socket options apply to the CLC socket */
1473 	return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1474 					     optval, optlen);
1475 }
1476 
1477 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1478 		     unsigned long arg)
1479 {
1480 	union smc_host_cursor cons, urg;
1481 	struct smc_connection *conn;
1482 	struct smc_sock *smc;
1483 	int answ;
1484 
1485 	smc = smc_sk(sock->sk);
1486 	conn = &smc->conn;
1487 	if (smc->use_fallback) {
1488 		if (!smc->clcsock)
1489 			return -EBADF;
1490 		return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1491 	}
1492 	switch (cmd) {
1493 	case SIOCINQ: /* same as FIONREAD */
1494 		if (smc->sk.sk_state == SMC_LISTEN)
1495 			return -EINVAL;
1496 		if (smc->sk.sk_state == SMC_INIT ||
1497 		    smc->sk.sk_state == SMC_CLOSED)
1498 			answ = 0;
1499 		else
1500 			answ = atomic_read(&smc->conn.bytes_to_rcv);
1501 		break;
1502 	case SIOCOUTQ:
1503 		/* output queue size (not send + not acked) */
1504 		if (smc->sk.sk_state == SMC_LISTEN)
1505 			return -EINVAL;
1506 		if (smc->sk.sk_state == SMC_INIT ||
1507 		    smc->sk.sk_state == SMC_CLOSED)
1508 			answ = 0;
1509 		else
1510 			answ = smc->conn.sndbuf_desc->len -
1511 					atomic_read(&smc->conn.sndbuf_space);
1512 		break;
1513 	case SIOCOUTQNSD:
1514 		/* output queue size (not send only) */
1515 		if (smc->sk.sk_state == SMC_LISTEN)
1516 			return -EINVAL;
1517 		if (smc->sk.sk_state == SMC_INIT ||
1518 		    smc->sk.sk_state == SMC_CLOSED)
1519 			answ = 0;
1520 		else
1521 			answ = smc_tx_prepared_sends(&smc->conn);
1522 		break;
1523 	case SIOCATMARK:
1524 		if (smc->sk.sk_state == SMC_LISTEN)
1525 			return -EINVAL;
1526 		if (smc->sk.sk_state == SMC_INIT ||
1527 		    smc->sk.sk_state == SMC_CLOSED) {
1528 			answ = 0;
1529 		} else {
1530 			smc_curs_write(&cons,
1531 			       smc_curs_read(&conn->local_tx_ctrl.cons, conn),
1532 				       conn);
1533 			smc_curs_write(&urg,
1534 				       smc_curs_read(&conn->urg_curs, conn),
1535 				       conn);
1536 			answ = smc_curs_diff(conn->rmb_desc->len,
1537 					     &cons, &urg) == 1;
1538 		}
1539 		break;
1540 	default:
1541 		return -ENOIOCTLCMD;
1542 	}
1543 
1544 	return put_user(answ, (int __user *)arg);
1545 }
1546 
1547 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1548 			    int offset, size_t size, int flags)
1549 {
1550 	struct sock *sk = sock->sk;
1551 	struct smc_sock *smc;
1552 	int rc = -EPIPE;
1553 
1554 	smc = smc_sk(sk);
1555 	lock_sock(sk);
1556 	if (sk->sk_state != SMC_ACTIVE) {
1557 		release_sock(sk);
1558 		goto out;
1559 	}
1560 	release_sock(sk);
1561 	if (smc->use_fallback)
1562 		rc = kernel_sendpage(smc->clcsock, page, offset,
1563 				     size, flags);
1564 	else
1565 		rc = sock_no_sendpage(sock, page, offset, size, flags);
1566 
1567 out:
1568 	return rc;
1569 }
1570 
1571 /* Map the affected portions of the rmbe into an spd, note the number of bytes
1572  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
1573  * updates till whenever a respective page has been fully processed.
1574  * Note that subsequent recv() calls have to wait till all splice() processing
1575  * completed.
1576  */
1577 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1578 			       struct pipe_inode_info *pipe, size_t len,
1579 			       unsigned int flags)
1580 {
1581 	struct sock *sk = sock->sk;
1582 	struct smc_sock *smc;
1583 	int rc = -ENOTCONN;
1584 
1585 	smc = smc_sk(sk);
1586 	lock_sock(sk);
1587 
1588 	if (sk->sk_state == SMC_INIT ||
1589 	    sk->sk_state == SMC_LISTEN ||
1590 	    sk->sk_state == SMC_CLOSED)
1591 		goto out;
1592 
1593 	if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1594 		rc = 0;
1595 		goto out;
1596 	}
1597 
1598 	if (smc->use_fallback) {
1599 		rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1600 						    pipe, len, flags);
1601 	} else {
1602 		if (*ppos) {
1603 			rc = -ESPIPE;
1604 			goto out;
1605 		}
1606 		if (flags & SPLICE_F_NONBLOCK)
1607 			flags = MSG_DONTWAIT;
1608 		else
1609 			flags = 0;
1610 		rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
1611 	}
1612 out:
1613 	release_sock(sk);
1614 
1615 	return rc;
1616 }
1617 
1618 /* must look like tcp */
1619 static const struct proto_ops smc_sock_ops = {
1620 	.family		= PF_SMC,
1621 	.owner		= THIS_MODULE,
1622 	.release	= smc_release,
1623 	.bind		= smc_bind,
1624 	.connect	= smc_connect,
1625 	.socketpair	= sock_no_socketpair,
1626 	.accept		= smc_accept,
1627 	.getname	= smc_getname,
1628 	.poll		= smc_poll,
1629 	.ioctl		= smc_ioctl,
1630 	.listen		= smc_listen,
1631 	.shutdown	= smc_shutdown,
1632 	.setsockopt	= smc_setsockopt,
1633 	.getsockopt	= smc_getsockopt,
1634 	.sendmsg	= smc_sendmsg,
1635 	.recvmsg	= smc_recvmsg,
1636 	.mmap		= sock_no_mmap,
1637 	.sendpage	= smc_sendpage,
1638 	.splice_read	= smc_splice_read,
1639 };
1640 
1641 static int smc_create(struct net *net, struct socket *sock, int protocol,
1642 		      int kern)
1643 {
1644 	int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
1645 	struct smc_sock *smc;
1646 	struct sock *sk;
1647 	int rc;
1648 
1649 	rc = -ESOCKTNOSUPPORT;
1650 	if (sock->type != SOCK_STREAM)
1651 		goto out;
1652 
1653 	rc = -EPROTONOSUPPORT;
1654 	if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
1655 		goto out;
1656 
1657 	rc = -ENOBUFS;
1658 	sock->ops = &smc_sock_ops;
1659 	sk = smc_sock_alloc(net, sock, protocol);
1660 	if (!sk)
1661 		goto out;
1662 
1663 	/* create internal TCP socket for CLC handshake and fallback */
1664 	smc = smc_sk(sk);
1665 	smc->use_fallback = false; /* assume rdma capability first */
1666 	rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
1667 			      &smc->clcsock);
1668 	if (rc) {
1669 		sk_common_release(sk);
1670 		goto out;
1671 	}
1672 	smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1673 	smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1674 
1675 out:
1676 	return rc;
1677 }
1678 
1679 static const struct net_proto_family smc_sock_family_ops = {
1680 	.family	= PF_SMC,
1681 	.owner	= THIS_MODULE,
1682 	.create	= smc_create,
1683 };
1684 
1685 static int __init smc_init(void)
1686 {
1687 	int rc;
1688 
1689 	rc = smc_pnet_init();
1690 	if (rc)
1691 		return rc;
1692 
1693 	rc = smc_llc_init();
1694 	if (rc) {
1695 		pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
1696 		goto out_pnet;
1697 	}
1698 
1699 	rc = smc_cdc_init();
1700 	if (rc) {
1701 		pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
1702 		goto out_pnet;
1703 	}
1704 
1705 	rc = proto_register(&smc_proto, 1);
1706 	if (rc) {
1707 		pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
1708 		goto out_pnet;
1709 	}
1710 
1711 	rc = proto_register(&smc_proto6, 1);
1712 	if (rc) {
1713 		pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
1714 		goto out_proto;
1715 	}
1716 
1717 	rc = sock_register(&smc_sock_family_ops);
1718 	if (rc) {
1719 		pr_err("%s: sock_register fails with %d\n", __func__, rc);
1720 		goto out_proto6;
1721 	}
1722 	INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
1723 	INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
1724 
1725 	rc = smc_ib_register_client();
1726 	if (rc) {
1727 		pr_err("%s: ib_register fails with %d\n", __func__, rc);
1728 		goto out_sock;
1729 	}
1730 
1731 	static_branch_enable(&tcp_have_smc);
1732 	return 0;
1733 
1734 out_sock:
1735 	sock_unregister(PF_SMC);
1736 out_proto6:
1737 	proto_unregister(&smc_proto6);
1738 out_proto:
1739 	proto_unregister(&smc_proto);
1740 out_pnet:
1741 	smc_pnet_exit();
1742 	return rc;
1743 }
1744 
1745 static void __exit smc_exit(void)
1746 {
1747 	smc_core_exit();
1748 	static_branch_disable(&tcp_have_smc);
1749 	smc_ib_unregister_client();
1750 	sock_unregister(PF_SMC);
1751 	proto_unregister(&smc_proto6);
1752 	proto_unregister(&smc_proto);
1753 	smc_pnet_exit();
1754 }
1755 
1756 module_init(smc_init);
1757 module_exit(smc_exit);
1758 
1759 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1760 MODULE_DESCRIPTION("smc socket address family");
1761 MODULE_LICENSE("GPL");
1762 MODULE_ALIAS_NETPROTO(PF_SMC);
1763