Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
[muen/linux.git] / net / smc / af_smc.c
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 static void smc_connect_work(struct work_struct *);
49
50 static void smc_set_keepalive(struct sock *sk, int val)
51 {
52         struct smc_sock *smc = smc_sk(sk);
53
54         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
55 }
56
57 static struct smc_hashinfo smc_v4_hashinfo = {
58         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
59 };
60
61 static struct smc_hashinfo smc_v6_hashinfo = {
62         .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
63 };
64
65 int smc_hash_sk(struct sock *sk)
66 {
67         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
68         struct hlist_head *head;
69
70         head = &h->ht;
71
72         write_lock_bh(&h->lock);
73         sk_add_node(sk, head);
74         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
75         write_unlock_bh(&h->lock);
76
77         return 0;
78 }
79 EXPORT_SYMBOL_GPL(smc_hash_sk);
80
81 void smc_unhash_sk(struct sock *sk)
82 {
83         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
84
85         write_lock_bh(&h->lock);
86         if (sk_del_node_init(sk))
87                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
88         write_unlock_bh(&h->lock);
89 }
90 EXPORT_SYMBOL_GPL(smc_unhash_sk);
91
92 struct proto smc_proto = {
93         .name           = "SMC",
94         .owner          = THIS_MODULE,
95         .keepalive      = smc_set_keepalive,
96         .hash           = smc_hash_sk,
97         .unhash         = smc_unhash_sk,
98         .obj_size       = sizeof(struct smc_sock),
99         .h.smc_hash     = &smc_v4_hashinfo,
100         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
101 };
102 EXPORT_SYMBOL_GPL(smc_proto);
103
104 struct proto smc_proto6 = {
105         .name           = "SMC6",
106         .owner          = THIS_MODULE,
107         .keepalive      = smc_set_keepalive,
108         .hash           = smc_hash_sk,
109         .unhash         = smc_unhash_sk,
110         .obj_size       = sizeof(struct smc_sock),
111         .h.smc_hash     = &smc_v6_hashinfo,
112         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
113 };
114 EXPORT_SYMBOL_GPL(smc_proto6);
115
116 static int smc_release(struct socket *sock)
117 {
118         struct sock *sk = sock->sk;
119         struct smc_sock *smc;
120         int rc = 0;
121
122         if (!sk)
123                 goto out;
124
125         smc = smc_sk(sk);
126
127         /* cleanup for a dangling non-blocking connect */
128         flush_work(&smc->connect_work);
129         kfree(smc->connect_info);
130         smc->connect_info = NULL;
131
132         if (sk->sk_state == SMC_LISTEN)
133                 /* smc_close_non_accepted() is called and acquires
134                  * sock lock for child sockets again
135                  */
136                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
137         else
138                 lock_sock(sk);
139
140         if (!smc->use_fallback) {
141                 rc = smc_close_active(smc);
142                 sock_set_flag(sk, SOCK_DEAD);
143                 sk->sk_shutdown |= SHUTDOWN_MASK;
144         }
145         if (smc->clcsock) {
146                 sock_release(smc->clcsock);
147                 smc->clcsock = NULL;
148         }
149         if (smc->use_fallback) {
150                 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
151                         sock_put(sk); /* passive closing */
152                 sk->sk_state = SMC_CLOSED;
153                 sk->sk_state_change(sk);
154         }
155
156         /* detach socket */
157         sock_orphan(sk);
158         sock->sk = NULL;
159         if (!smc->use_fallback && sk->sk_state == SMC_CLOSED)
160                 smc_conn_free(&smc->conn);
161         release_sock(sk);
162
163         sk->sk_prot->unhash(sk);
164         sock_put(sk); /* final sock_put */
165 out:
166         return rc;
167 }
168
169 static void smc_destruct(struct sock *sk)
170 {
171         if (sk->sk_state != SMC_CLOSED)
172                 return;
173         if (!sock_flag(sk, SOCK_DEAD))
174                 return;
175
176         sk_refcnt_debug_dec(sk);
177 }
178
179 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
180                                    int protocol)
181 {
182         struct smc_sock *smc;
183         struct proto *prot;
184         struct sock *sk;
185
186         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
187         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
188         if (!sk)
189                 return NULL;
190
191         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
192         sk->sk_state = SMC_INIT;
193         sk->sk_destruct = smc_destruct;
194         sk->sk_protocol = protocol;
195         smc = smc_sk(sk);
196         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
197         INIT_WORK(&smc->connect_work, smc_connect_work);
198         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
199         INIT_LIST_HEAD(&smc->accept_q);
200         spin_lock_init(&smc->accept_q_lock);
201         spin_lock_init(&smc->conn.send_lock);
202         sk->sk_prot->hash(sk);
203         sk_refcnt_debug_inc(sk);
204
205         return sk;
206 }
207
208 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
209                     int addr_len)
210 {
211         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
212         struct sock *sk = sock->sk;
213         struct smc_sock *smc;
214         int rc;
215
216         smc = smc_sk(sk);
217
218         /* replicate tests from inet_bind(), to be safe wrt. future changes */
219         rc = -EINVAL;
220         if (addr_len < sizeof(struct sockaddr_in))
221                 goto out;
222
223         rc = -EAFNOSUPPORT;
224         if (addr->sin_family != AF_INET &&
225             addr->sin_family != AF_INET6 &&
226             addr->sin_family != AF_UNSPEC)
227                 goto out;
228         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
229         if (addr->sin_family == AF_UNSPEC &&
230             addr->sin_addr.s_addr != htonl(INADDR_ANY))
231                 goto out;
232
233         lock_sock(sk);
234
235         /* Check if socket is already active */
236         rc = -EINVAL;
237         if (sk->sk_state != SMC_INIT)
238                 goto out_rel;
239
240         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
241         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
242
243 out_rel:
244         release_sock(sk);
245 out:
246         return rc;
247 }
248
249 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
250                                    unsigned long mask)
251 {
252         /* options we don't get control via setsockopt for */
253         nsk->sk_type = osk->sk_type;
254         nsk->sk_sndbuf = osk->sk_sndbuf;
255         nsk->sk_rcvbuf = osk->sk_rcvbuf;
256         nsk->sk_sndtimeo = osk->sk_sndtimeo;
257         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
258         nsk->sk_mark = osk->sk_mark;
259         nsk->sk_priority = osk->sk_priority;
260         nsk->sk_rcvlowat = osk->sk_rcvlowat;
261         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
262         nsk->sk_err = osk->sk_err;
263
264         nsk->sk_flags &= ~mask;
265         nsk->sk_flags |= osk->sk_flags & mask;
266 }
267
268 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
269                              (1UL << SOCK_KEEPOPEN) | \
270                              (1UL << SOCK_LINGER) | \
271                              (1UL << SOCK_BROADCAST) | \
272                              (1UL << SOCK_TIMESTAMP) | \
273                              (1UL << SOCK_DBG) | \
274                              (1UL << SOCK_RCVTSTAMP) | \
275                              (1UL << SOCK_RCVTSTAMPNS) | \
276                              (1UL << SOCK_LOCALROUTE) | \
277                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
278                              (1UL << SOCK_RXQ_OVFL) | \
279                              (1UL << SOCK_WIFI_STATUS) | \
280                              (1UL << SOCK_NOFCS) | \
281                              (1UL << SOCK_FILTER_LOCKED))
282 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
283  * clc socket (since smc is not called for these options from net/core)
284  */
285 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
286 {
287         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
288 }
289
290 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
291                              (1UL << SOCK_KEEPOPEN) | \
292                              (1UL << SOCK_LINGER) | \
293                              (1UL << SOCK_DBG))
294 /* copy only settings and flags relevant for smc from clc to smc socket */
295 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
296 {
297         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
298 }
299
300 /* register a new rmb, optionally send confirm_rkey msg to register with peer */
301 static int smc_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc,
302                        bool conf_rkey)
303 {
304         /* register memory region for new rmb */
305         if (smc_wr_reg_send(link, rmb_desc->mr_rx[SMC_SINGLE_LINK])) {
306                 rmb_desc->regerr = 1;
307                 return -EFAULT;
308         }
309         if (!conf_rkey)
310                 return 0;
311         /* exchange confirm_rkey msg with peer */
312         if (smc_llc_do_confirm_rkey(link, rmb_desc)) {
313                 rmb_desc->regerr = 1;
314                 return -EFAULT;
315         }
316         return 0;
317 }
318
319 static int smc_clnt_conf_first_link(struct smc_sock *smc)
320 {
321         struct net *net = sock_net(smc->clcsock->sk);
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         if (link->llc_confirm_rc)
341                 return SMC_CLC_DECL_RMBE_EC;
342
343         rc = smc_ib_modify_qp_rts(link);
344         if (rc)
345                 return SMC_CLC_DECL_INTERR;
346
347         smc_wr_remember_qp_attr(link);
348
349         if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
350                 return SMC_CLC_DECL_INTERR;
351
352         /* send CONFIRM LINK response over RoCE fabric */
353         rc = smc_llc_send_confirm_link(link,
354                                        link->smcibdev->mac[link->ibport - 1],
355                                        &link->smcibdev->gid[link->ibport - 1],
356                                        SMC_LLC_RESP);
357         if (rc < 0)
358                 return SMC_CLC_DECL_TCL;
359
360         /* receive ADD LINK request from server over RoCE fabric */
361         rest = wait_for_completion_interruptible_timeout(&link->llc_add,
362                                                          SMC_LLC_WAIT_TIME);
363         if (rest <= 0) {
364                 struct smc_clc_msg_decline dclc;
365
366                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
367                                       SMC_CLC_DECLINE);
368                 return rc;
369         }
370
371         /* send add link reject message, only one link supported for now */
372         rc = smc_llc_send_add_link(link,
373                                    link->smcibdev->mac[link->ibport - 1],
374                                    &link->smcibdev->gid[link->ibport - 1],
375                                    SMC_LLC_RESP);
376         if (rc < 0)
377                 return SMC_CLC_DECL_TCL;
378
379         smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
380
381         return 0;
382 }
383
384 static void smc_conn_save_peer_info(struct smc_sock *smc,
385                                     struct smc_clc_msg_accept_confirm *clc)
386 {
387         int bufsize = smc_uncompress_bufsize(clc->rmbe_size);
388
389         smc->conn.peer_rmbe_idx = clc->rmbe_idx;
390         smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
391         smc->conn.peer_rmbe_size = bufsize;
392         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
393         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
394 }
395
396 static void smc_link_save_peer_info(struct smc_link *link,
397                                     struct smc_clc_msg_accept_confirm *clc)
398 {
399         link->peer_qpn = ntoh24(clc->qpn);
400         memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
401         memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
402         link->peer_psn = ntoh24(clc->psn);
403         link->peer_mtu = clc->qp_mtu;
404 }
405
406 /* fall back during connect */
407 static int smc_connect_fallback(struct smc_sock *smc)
408 {
409         smc->use_fallback = true;
410         smc_copy_sock_settings_to_clc(smc);
411         if (smc->sk.sk_state == SMC_INIT)
412                 smc->sk.sk_state = SMC_ACTIVE;
413         return 0;
414 }
415
416 /* decline and fall back during connect */
417 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
418 {
419         int rc;
420
421         if (reason_code < 0) { /* error, fallback is not possible */
422                 if (smc->sk.sk_state == SMC_INIT)
423                         sock_put(&smc->sk); /* passive closing */
424                 return reason_code;
425         }
426         if (reason_code != SMC_CLC_DECL_REPLY) {
427                 rc = smc_clc_send_decline(smc, reason_code);
428                 if (rc < 0) {
429                         if (smc->sk.sk_state == SMC_INIT)
430                                 sock_put(&smc->sk); /* passive closing */
431                         return rc;
432                 }
433         }
434         return smc_connect_fallback(smc);
435 }
436
437 /* abort connecting */
438 static int smc_connect_abort(struct smc_sock *smc, int reason_code,
439                              int local_contact)
440 {
441         if (local_contact == SMC_FIRST_CONTACT)
442                 smc_lgr_forget(smc->conn.lgr);
443         mutex_unlock(&smc_create_lgr_pending);
444         smc_conn_free(&smc->conn);
445         return reason_code;
446 }
447
448 /* check if there is a rdma device available for this connection. */
449 /* called for connect and listen */
450 static int smc_check_rdma(struct smc_sock *smc, struct smc_ib_device **ibdev,
451                           u8 *ibport)
452 {
453         int reason_code = 0;
454
455         /* PNET table look up: search active ib_device and port
456          * within same PNETID that also contains the ethernet device
457          * used for the internal TCP socket
458          */
459         smc_pnet_find_roce_resource(smc->clcsock->sk, ibdev, ibport);
460         if (!(*ibdev))
461                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
462
463         return reason_code;
464 }
465
466 /* CLC handshake during connect */
467 static int smc_connect_clc(struct smc_sock *smc,
468                            struct smc_clc_msg_accept_confirm *aclc,
469                            struct smc_ib_device *ibdev, u8 ibport)
470 {
471         int rc = 0;
472
473         /* do inband token exchange */
474         rc = smc_clc_send_proposal(smc, ibdev, ibport);
475         if (rc)
476                 return rc;
477         /* receive SMC Accept CLC message */
478         return smc_clc_wait_msg(smc, aclc, sizeof(*aclc), SMC_CLC_ACCEPT);
479 }
480
481 /* setup for RDMA connection of client */
482 static int smc_connect_rdma(struct smc_sock *smc,
483                             struct smc_clc_msg_accept_confirm *aclc,
484                             struct smc_ib_device *ibdev, u8 ibport)
485 {
486         int local_contact = SMC_FIRST_CONTACT;
487         struct smc_link *link;
488         int reason_code = 0;
489
490         mutex_lock(&smc_create_lgr_pending);
491         local_contact = smc_conn_create(smc, ibdev, ibport, &aclc->lcl,
492                                         aclc->hdr.flag);
493         if (local_contact < 0) {
494                 if (local_contact == -ENOMEM)
495                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
496                 else if (local_contact == -ENOLINK)
497                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
498                 else
499                         reason_code = SMC_CLC_DECL_INTERR; /* other error */
500                 return smc_connect_abort(smc, reason_code, 0);
501         }
502         link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
503
504         smc_conn_save_peer_info(smc, aclc);
505
506         /* create send buffer and rmb */
507         if (smc_buf_create(smc))
508                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, local_contact);
509
510         if (local_contact == SMC_FIRST_CONTACT)
511                 smc_link_save_peer_info(link, aclc);
512
513         if (smc_rmb_rtoken_handling(&smc->conn, aclc))
514                 return smc_connect_abort(smc, SMC_CLC_DECL_INTERR,
515                                          local_contact);
516
517         smc_close_init(smc);
518         smc_rx_init(smc);
519
520         if (local_contact == SMC_FIRST_CONTACT) {
521                 if (smc_ib_ready_link(link))
522                         return smc_connect_abort(smc, SMC_CLC_DECL_INTERR,
523                                                  local_contact);
524         } else {
525                 if (!smc->conn.rmb_desc->reused &&
526                     smc_reg_rmb(link, smc->conn.rmb_desc, true))
527                         return smc_connect_abort(smc, SMC_CLC_DECL_INTERR,
528                                                  local_contact);
529         }
530         smc_rmb_sync_sg_for_device(&smc->conn);
531
532         reason_code = smc_clc_send_confirm(smc);
533         if (reason_code)
534                 return smc_connect_abort(smc, reason_code, local_contact);
535
536         smc_tx_init(smc);
537
538         if (local_contact == SMC_FIRST_CONTACT) {
539                 /* QP confirmation over RoCE fabric */
540                 reason_code = smc_clnt_conf_first_link(smc);
541                 if (reason_code)
542                         return smc_connect_abort(smc, reason_code,
543                                                  local_contact);
544         }
545         mutex_unlock(&smc_create_lgr_pending);
546
547         smc_copy_sock_settings_to_clc(smc);
548         if (smc->sk.sk_state == SMC_INIT)
549                 smc->sk.sk_state = SMC_ACTIVE;
550
551         return 0;
552 }
553
554 /* perform steps before actually connecting */
555 static int __smc_connect(struct smc_sock *smc)
556 {
557         struct smc_clc_msg_accept_confirm aclc;
558         struct smc_ib_device *ibdev;
559         int rc = 0;
560         u8 ibport;
561
562         sock_hold(&smc->sk); /* sock put in passive closing */
563
564         if (smc->use_fallback)
565                 return smc_connect_fallback(smc);
566
567         /* if peer has not signalled SMC-capability, fall back */
568         if (!tcp_sk(smc->clcsock->sk)->syn_smc)
569                 return smc_connect_fallback(smc);
570
571         /* IPSec connections opt out of SMC-R optimizations */
572         if (using_ipsec(smc))
573                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
574
575         /* check if a RDMA device is available; if not, fall back */
576         if (smc_check_rdma(smc, &ibdev, &ibport))
577                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_CNFERR);
578
579         /* perform CLC handshake */
580         rc = smc_connect_clc(smc, &aclc, ibdev, ibport);
581         if (rc)
582                 return smc_connect_decline_fallback(smc, rc);
583
584         /* connect using rdma */
585         rc = smc_connect_rdma(smc, &aclc, ibdev, ibport);
586         if (rc)
587                 return smc_connect_decline_fallback(smc, rc);
588
589         return 0;
590 }
591
592 static void smc_connect_work(struct work_struct *work)
593 {
594         struct smc_sock *smc = container_of(work, struct smc_sock,
595                                             connect_work);
596         int rc;
597
598         lock_sock(&smc->sk);
599         rc = kernel_connect(smc->clcsock, &smc->connect_info->addr,
600                             smc->connect_info->alen, smc->connect_info->flags);
601         if (smc->clcsock->sk->sk_err) {
602                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
603                 goto out;
604         }
605         if (rc < 0) {
606                 smc->sk.sk_err = -rc;
607                 goto out;
608         }
609
610         rc = __smc_connect(smc);
611         if (rc < 0)
612                 smc->sk.sk_err = -rc;
613
614 out:
615         smc->sk.sk_state_change(&smc->sk);
616         kfree(smc->connect_info);
617         smc->connect_info = NULL;
618         release_sock(&smc->sk);
619 }
620
621 static int smc_connect(struct socket *sock, struct sockaddr *addr,
622                        int alen, int flags)
623 {
624         struct sock *sk = sock->sk;
625         struct smc_sock *smc;
626         int rc = -EINVAL;
627
628         smc = smc_sk(sk);
629
630         /* separate smc parameter checking to be safe */
631         if (alen < sizeof(addr->sa_family))
632                 goto out_err;
633         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
634                 goto out_err;
635
636         lock_sock(sk);
637         switch (sk->sk_state) {
638         default:
639                 goto out;
640         case SMC_ACTIVE:
641                 rc = -EISCONN;
642                 goto out;
643         case SMC_INIT:
644                 rc = 0;
645                 break;
646         }
647
648         smc_copy_sock_settings_to_clc(smc);
649         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
650         if (flags & O_NONBLOCK) {
651                 if (smc->connect_info) {
652                         rc = -EALREADY;
653                         goto out;
654                 }
655                 smc->connect_info = kzalloc(alen + 2 * sizeof(int), GFP_KERNEL);
656                 if (!smc->connect_info) {
657                         rc = -ENOMEM;
658                         goto out;
659                 }
660                 smc->connect_info->alen = alen;
661                 smc->connect_info->flags = flags ^ O_NONBLOCK;
662                 memcpy(&smc->connect_info->addr, addr, alen);
663                 schedule_work(&smc->connect_work);
664                 rc = -EINPROGRESS;
665         } else {
666                 rc = kernel_connect(smc->clcsock, addr, alen, flags);
667                 if (rc)
668                         goto out;
669
670                 rc = __smc_connect(smc);
671                 if (rc < 0)
672                         goto out;
673                 else
674                         rc = 0; /* success cases including fallback */
675         }
676
677 out:
678         release_sock(sk);
679 out_err:
680         return rc;
681 }
682
683 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
684 {
685         struct socket *new_clcsock = NULL;
686         struct sock *lsk = &lsmc->sk;
687         struct sock *new_sk;
688         int rc;
689
690         release_sock(lsk);
691         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
692         if (!new_sk) {
693                 rc = -ENOMEM;
694                 lsk->sk_err = ENOMEM;
695                 *new_smc = NULL;
696                 lock_sock(lsk);
697                 goto out;
698         }
699         *new_smc = smc_sk(new_sk);
700
701         rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
702         lock_sock(lsk);
703         if  (rc < 0)
704                 lsk->sk_err = -rc;
705         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
706                 if (new_clcsock)
707                         sock_release(new_clcsock);
708                 new_sk->sk_state = SMC_CLOSED;
709                 sock_set_flag(new_sk, SOCK_DEAD);
710                 new_sk->sk_prot->unhash(new_sk);
711                 sock_put(new_sk); /* final */
712                 *new_smc = NULL;
713                 goto out;
714         }
715
716         (*new_smc)->clcsock = new_clcsock;
717 out:
718         return rc;
719 }
720
721 /* add a just created sock to the accept queue of the listen sock as
722  * candidate for a following socket accept call from user space
723  */
724 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
725 {
726         struct smc_sock *par = smc_sk(parent);
727
728         sock_hold(sk); /* sock_put in smc_accept_unlink () */
729         spin_lock(&par->accept_q_lock);
730         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
731         spin_unlock(&par->accept_q_lock);
732         sk_acceptq_added(parent);
733 }
734
735 /* remove a socket from the accept queue of its parental listening socket */
736 static void smc_accept_unlink(struct sock *sk)
737 {
738         struct smc_sock *par = smc_sk(sk)->listen_smc;
739
740         spin_lock(&par->accept_q_lock);
741         list_del_init(&smc_sk(sk)->accept_q);
742         spin_unlock(&par->accept_q_lock);
743         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
744         sock_put(sk); /* sock_hold in smc_accept_enqueue */
745 }
746
747 /* remove a sock from the accept queue to bind it to a new socket created
748  * for a socket accept call from user space
749  */
750 struct sock *smc_accept_dequeue(struct sock *parent,
751                                 struct socket *new_sock)
752 {
753         struct smc_sock *isk, *n;
754         struct sock *new_sk;
755
756         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
757                 new_sk = (struct sock *)isk;
758
759                 smc_accept_unlink(new_sk);
760                 if (new_sk->sk_state == SMC_CLOSED) {
761                         if (isk->clcsock) {
762                                 sock_release(isk->clcsock);
763                                 isk->clcsock = NULL;
764                         }
765                         new_sk->sk_prot->unhash(new_sk);
766                         sock_put(new_sk); /* final */
767                         continue;
768                 }
769                 if (new_sock)
770                         sock_graft(new_sk, new_sock);
771                 return new_sk;
772         }
773         return NULL;
774 }
775
776 /* clean up for a created but never accepted sock */
777 void smc_close_non_accepted(struct sock *sk)
778 {
779         struct smc_sock *smc = smc_sk(sk);
780
781         lock_sock(sk);
782         if (!sk->sk_lingertime)
783                 /* wait for peer closing */
784                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
785         if (!smc->use_fallback) {
786                 smc_close_active(smc);
787                 sock_set_flag(sk, SOCK_DEAD);
788                 sk->sk_shutdown |= SHUTDOWN_MASK;
789         }
790         if (smc->clcsock) {
791                 struct socket *tcp;
792
793                 tcp = smc->clcsock;
794                 smc->clcsock = NULL;
795                 sock_release(tcp);
796         }
797         if (smc->use_fallback) {
798                 sock_put(sk); /* passive closing */
799                 sk->sk_state = SMC_CLOSED;
800         } else {
801                 if (sk->sk_state == SMC_CLOSED)
802                         smc_conn_free(&smc->conn);
803         }
804         release_sock(sk);
805         sk->sk_prot->unhash(sk);
806         sock_put(sk); /* final sock_put */
807 }
808
809 static int smc_serv_conf_first_link(struct smc_sock *smc)
810 {
811         struct net *net = sock_net(smc->clcsock->sk);
812         struct smc_link_group *lgr = smc->conn.lgr;
813         struct smc_link *link;
814         int rest;
815         int rc;
816
817         link = &lgr->lnk[SMC_SINGLE_LINK];
818
819         if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
820                 return SMC_CLC_DECL_INTERR;
821
822         /* send CONFIRM LINK request to client over the RoCE fabric */
823         rc = smc_llc_send_confirm_link(link,
824                                        link->smcibdev->mac[link->ibport - 1],
825                                        &link->smcibdev->gid[link->ibport - 1],
826                                        SMC_LLC_REQ);
827         if (rc < 0)
828                 return SMC_CLC_DECL_TCL;
829
830         /* receive CONFIRM LINK response from client over the RoCE fabric */
831         rest = wait_for_completion_interruptible_timeout(
832                 &link->llc_confirm_resp,
833                 SMC_LLC_WAIT_FIRST_TIME);
834         if (rest <= 0) {
835                 struct smc_clc_msg_decline dclc;
836
837                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
838                                       SMC_CLC_DECLINE);
839                 return rc;
840         }
841
842         if (link->llc_confirm_resp_rc)
843                 return SMC_CLC_DECL_RMBE_EC;
844
845         /* send ADD LINK request to client over the RoCE fabric */
846         rc = smc_llc_send_add_link(link,
847                                    link->smcibdev->mac[link->ibport - 1],
848                                    &link->smcibdev->gid[link->ibport - 1],
849                                    SMC_LLC_REQ);
850         if (rc < 0)
851                 return SMC_CLC_DECL_TCL;
852
853         /* receive ADD LINK response from client over the RoCE fabric */
854         rest = wait_for_completion_interruptible_timeout(&link->llc_add_resp,
855                                                          SMC_LLC_WAIT_TIME);
856         if (rest <= 0) {
857                 struct smc_clc_msg_decline dclc;
858
859                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
860                                       SMC_CLC_DECLINE);
861                 return rc;
862         }
863
864         smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
865
866         return 0;
867 }
868
869 /* listen worker: finish */
870 static void smc_listen_out(struct smc_sock *new_smc)
871 {
872         struct smc_sock *lsmc = new_smc->listen_smc;
873         struct sock *newsmcsk = &new_smc->sk;
874
875         lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
876         if (lsmc->sk.sk_state == SMC_LISTEN) {
877                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
878         } else { /* no longer listening */
879                 smc_close_non_accepted(newsmcsk);
880         }
881         release_sock(&lsmc->sk);
882
883         /* Wake up accept */
884         lsmc->sk.sk_data_ready(&lsmc->sk);
885         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
886 }
887
888 /* listen worker: finish in state connected */
889 static void smc_listen_out_connected(struct smc_sock *new_smc)
890 {
891         struct sock *newsmcsk = &new_smc->sk;
892
893         sk_refcnt_debug_inc(newsmcsk);
894         if (newsmcsk->sk_state == SMC_INIT)
895                 newsmcsk->sk_state = SMC_ACTIVE;
896
897         smc_listen_out(new_smc);
898 }
899
900 /* listen worker: finish in error state */
901 static void smc_listen_out_err(struct smc_sock *new_smc)
902 {
903         struct sock *newsmcsk = &new_smc->sk;
904
905         if (newsmcsk->sk_state == SMC_INIT)
906                 sock_put(&new_smc->sk); /* passive closing */
907         newsmcsk->sk_state = SMC_CLOSED;
908         smc_conn_free(&new_smc->conn);
909
910         smc_listen_out(new_smc);
911 }
912
913 /* listen worker: decline and fall back if possible */
914 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
915                                int local_contact)
916 {
917         /* RDMA setup failed, switch back to TCP */
918         if (local_contact == SMC_FIRST_CONTACT)
919                 smc_lgr_forget(new_smc->conn.lgr);
920         if (reason_code < 0) { /* error, no fallback possible */
921                 smc_listen_out_err(new_smc);
922                 return;
923         }
924         smc_conn_free(&new_smc->conn);
925         new_smc->use_fallback = true;
926         if (reason_code && reason_code != SMC_CLC_DECL_REPLY) {
927                 if (smc_clc_send_decline(new_smc, reason_code) < 0) {
928                         smc_listen_out_err(new_smc);
929                         return;
930                 }
931         }
932         smc_listen_out_connected(new_smc);
933 }
934
935 /* listen worker: check prefixes */
936 static int smc_listen_rdma_check(struct smc_sock *new_smc,
937                                  struct smc_clc_msg_proposal *pclc)
938 {
939         struct smc_clc_msg_proposal_prefix *pclc_prfx;
940         struct socket *newclcsock = new_smc->clcsock;
941
942         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
943         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
944                 return SMC_CLC_DECL_CNFERR;
945
946         return 0;
947 }
948
949 /* listen worker: initialize connection and buffers */
950 static int smc_listen_rdma_init(struct smc_sock *new_smc,
951                                 struct smc_clc_msg_proposal *pclc,
952                                 struct smc_ib_device *ibdev, u8 ibport,
953                                 int *local_contact)
954 {
955         /* allocate connection / link group */
956         *local_contact = smc_conn_create(new_smc, ibdev, ibport, &pclc->lcl, 0);
957         if (*local_contact < 0) {
958                 if (*local_contact == -ENOMEM)
959                         return SMC_CLC_DECL_MEM;/* insufficient memory*/
960                 return SMC_CLC_DECL_INTERR; /* other error */
961         }
962
963         /* create send buffer and rmb */
964         if (smc_buf_create(new_smc))
965                 return SMC_CLC_DECL_MEM;
966
967         return 0;
968 }
969
970 /* listen worker: register buffers */
971 static int smc_listen_rdma_reg(struct smc_sock *new_smc, int local_contact)
972 {
973         struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
974
975         if (local_contact != SMC_FIRST_CONTACT) {
976                 if (!new_smc->conn.rmb_desc->reused) {
977                         if (smc_reg_rmb(link, new_smc->conn.rmb_desc, true))
978                                 return SMC_CLC_DECL_INTERR;
979                 }
980         }
981         smc_rmb_sync_sg_for_device(&new_smc->conn);
982
983         return 0;
984 }
985
986 /* listen worker: finish RDMA setup */
987 static void smc_listen_rdma_finish(struct smc_sock *new_smc,
988                                    struct smc_clc_msg_accept_confirm *cclc,
989                                    int local_contact)
990 {
991         struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
992         int reason_code = 0;
993
994         if (local_contact == SMC_FIRST_CONTACT)
995                 smc_link_save_peer_info(link, cclc);
996
997         if (smc_rmb_rtoken_handling(&new_smc->conn, cclc)) {
998                 reason_code = SMC_CLC_DECL_INTERR;
999                 goto decline;
1000         }
1001
1002         if (local_contact == SMC_FIRST_CONTACT) {
1003                 if (smc_ib_ready_link(link)) {
1004                         reason_code = SMC_CLC_DECL_INTERR;
1005                         goto decline;
1006                 }
1007                 /* QP confirmation over RoCE fabric */
1008                 reason_code = smc_serv_conf_first_link(new_smc);
1009                 if (reason_code)
1010                         goto decline;
1011         }
1012         return;
1013
1014 decline:
1015         mutex_unlock(&smc_create_lgr_pending);
1016         smc_listen_decline(new_smc, reason_code, local_contact);
1017 }
1018
1019 /* setup for RDMA connection of server */
1020 static void smc_listen_work(struct work_struct *work)
1021 {
1022         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1023                                                 smc_listen_work);
1024         struct socket *newclcsock = new_smc->clcsock;
1025         struct smc_clc_msg_accept_confirm cclc;
1026         struct smc_clc_msg_proposal *pclc;
1027         struct smc_ib_device *ibdev;
1028         u8 buf[SMC_CLC_MAX_LEN];
1029         int local_contact = 0;
1030         int reason_code = 0;
1031         int rc = 0;
1032         u8 ibport;
1033
1034         if (new_smc->use_fallback) {
1035                 smc_listen_out_connected(new_smc);
1036                 return;
1037         }
1038
1039         /* check if peer is smc capable */
1040         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1041                 new_smc->use_fallback = true;
1042                 smc_listen_out_connected(new_smc);
1043                 return;
1044         }
1045
1046         /* do inband token exchange -
1047          * wait for and receive SMC Proposal CLC message
1048          */
1049         pclc = (struct smc_clc_msg_proposal *)&buf;
1050         reason_code = smc_clc_wait_msg(new_smc, pclc, SMC_CLC_MAX_LEN,
1051                                        SMC_CLC_PROPOSAL);
1052         if (reason_code) {
1053                 smc_listen_decline(new_smc, reason_code, 0);
1054                 return;
1055         }
1056
1057         /* IPSec connections opt out of SMC-R optimizations */
1058         if (using_ipsec(new_smc)) {
1059                 smc_listen_decline(new_smc, SMC_CLC_DECL_IPSEC, 0);
1060                 return;
1061         }
1062
1063         mutex_lock(&smc_create_lgr_pending);
1064         smc_close_init(new_smc);
1065         smc_rx_init(new_smc);
1066         smc_tx_init(new_smc);
1067
1068         /* check if RDMA is available */
1069         if (smc_check_rdma(new_smc, &ibdev, &ibport) ||
1070             smc_listen_rdma_check(new_smc, pclc) ||
1071             smc_listen_rdma_init(new_smc, pclc, ibdev, ibport,
1072                                  &local_contact) ||
1073             smc_listen_rdma_reg(new_smc, local_contact)) {
1074                 /* SMC not supported, decline */
1075                 mutex_unlock(&smc_create_lgr_pending);
1076                 smc_listen_decline(new_smc, SMC_CLC_DECL_CNFERR, local_contact);
1077                 return;
1078         }
1079
1080         /* send SMC Accept CLC message */
1081         rc = smc_clc_send_accept(new_smc, local_contact);
1082         if (rc) {
1083                 mutex_unlock(&smc_create_lgr_pending);
1084                 smc_listen_decline(new_smc, rc, local_contact);
1085                 return;
1086         }
1087
1088         /* receive SMC Confirm CLC message */
1089         reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
1090                                        SMC_CLC_CONFIRM);
1091         if (reason_code) {
1092                 mutex_unlock(&smc_create_lgr_pending);
1093                 smc_listen_decline(new_smc, reason_code, local_contact);
1094                 return;
1095         }
1096
1097         /* finish worker */
1098         smc_listen_rdma_finish(new_smc, &cclc, local_contact);
1099         smc_conn_save_peer_info(new_smc, &cclc);
1100         mutex_unlock(&smc_create_lgr_pending);
1101         smc_listen_out_connected(new_smc);
1102 }
1103
1104 static void smc_tcp_listen_work(struct work_struct *work)
1105 {
1106         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1107                                              tcp_listen_work);
1108         struct sock *lsk = &lsmc->sk;
1109         struct smc_sock *new_smc;
1110         int rc = 0;
1111
1112         lock_sock(lsk);
1113         while (lsk->sk_state == SMC_LISTEN) {
1114                 rc = smc_clcsock_accept(lsmc, &new_smc);
1115                 if (rc)
1116                         goto out;
1117                 if (!new_smc)
1118                         continue;
1119
1120                 new_smc->listen_smc = lsmc;
1121                 new_smc->use_fallback = lsmc->use_fallback;
1122                 sock_hold(lsk); /* sock_put in smc_listen_work */
1123                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1124                 smc_copy_sock_settings_to_smc(new_smc);
1125                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1126                 if (!schedule_work(&new_smc->smc_listen_work))
1127                         sock_put(&new_smc->sk);
1128         }
1129
1130 out:
1131         release_sock(lsk);
1132         sock_put(&lsmc->sk); /* sock_hold in smc_listen */
1133 }
1134
1135 static int smc_listen(struct socket *sock, int backlog)
1136 {
1137         struct sock *sk = sock->sk;
1138         struct smc_sock *smc;
1139         int rc;
1140
1141         smc = smc_sk(sk);
1142         lock_sock(sk);
1143
1144         rc = -EINVAL;
1145         if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
1146                 goto out;
1147
1148         rc = 0;
1149         if (sk->sk_state == SMC_LISTEN) {
1150                 sk->sk_max_ack_backlog = backlog;
1151                 goto out;
1152         }
1153         /* some socket options are handled in core, so we could not apply
1154          * them to the clc socket -- copy smc socket options to clc socket
1155          */
1156         smc_copy_sock_settings_to_clc(smc);
1157         if (!smc->use_fallback)
1158                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1159
1160         rc = kernel_listen(smc->clcsock, backlog);
1161         if (rc)
1162                 goto out;
1163         sk->sk_max_ack_backlog = backlog;
1164         sk->sk_ack_backlog = 0;
1165         sk->sk_state = SMC_LISTEN;
1166         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
1167         sock_hold(sk); /* sock_hold in tcp_listen_worker */
1168         if (!schedule_work(&smc->tcp_listen_work))
1169                 sock_put(sk);
1170
1171 out:
1172         release_sock(sk);
1173         return rc;
1174 }
1175
1176 static int smc_accept(struct socket *sock, struct socket *new_sock,
1177                       int flags, bool kern)
1178 {
1179         struct sock *sk = sock->sk, *nsk;
1180         DECLARE_WAITQUEUE(wait, current);
1181         struct smc_sock *lsmc;
1182         long timeo;
1183         int rc = 0;
1184
1185         lsmc = smc_sk(sk);
1186         sock_hold(sk); /* sock_put below */
1187         lock_sock(sk);
1188
1189         if (lsmc->sk.sk_state != SMC_LISTEN) {
1190                 rc = -EINVAL;
1191                 release_sock(sk);
1192                 goto out;
1193         }
1194
1195         /* Wait for an incoming connection */
1196         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1197         add_wait_queue_exclusive(sk_sleep(sk), &wait);
1198         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1199                 set_current_state(TASK_INTERRUPTIBLE);
1200                 if (!timeo) {
1201                         rc = -EAGAIN;
1202                         break;
1203                 }
1204                 release_sock(sk);
1205                 timeo = schedule_timeout(timeo);
1206                 /* wakeup by sk_data_ready in smc_listen_work() */
1207                 sched_annotate_sleep();
1208                 lock_sock(sk);
1209                 if (signal_pending(current)) {
1210                         rc = sock_intr_errno(timeo);
1211                         break;
1212                 }
1213         }
1214         set_current_state(TASK_RUNNING);
1215         remove_wait_queue(sk_sleep(sk), &wait);
1216
1217         if (!rc)
1218                 rc = sock_error(nsk);
1219         release_sock(sk);
1220         if (rc)
1221                 goto out;
1222
1223         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1224                 /* wait till data arrives on the socket */
1225                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1226                                                                 MSEC_PER_SEC);
1227                 if (smc_sk(nsk)->use_fallback) {
1228                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1229
1230                         lock_sock(clcsk);
1231                         if (skb_queue_empty(&clcsk->sk_receive_queue))
1232                                 sk_wait_data(clcsk, &timeo, NULL);
1233                         release_sock(clcsk);
1234                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1235                         lock_sock(nsk);
1236                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1237                         release_sock(nsk);
1238                 }
1239         }
1240
1241 out:
1242         sock_put(sk); /* sock_hold above */
1243         return rc;
1244 }
1245
1246 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1247                        int peer)
1248 {
1249         struct smc_sock *smc;
1250
1251         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1252             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1253                 return -ENOTCONN;
1254
1255         smc = smc_sk(sock->sk);
1256
1257         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1258 }
1259
1260 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1261 {
1262         struct sock *sk = sock->sk;
1263         struct smc_sock *smc;
1264         int rc = -EPIPE;
1265
1266         smc = smc_sk(sk);
1267         lock_sock(sk);
1268         if ((sk->sk_state != SMC_ACTIVE) &&
1269             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1270             (sk->sk_state != SMC_INIT))
1271                 goto out;
1272
1273         if (msg->msg_flags & MSG_FASTOPEN) {
1274                 if (sk->sk_state == SMC_INIT) {
1275                         smc->use_fallback = true;
1276                 } else {
1277                         rc = -EINVAL;
1278                         goto out;
1279                 }
1280         }
1281
1282         if (smc->use_fallback)
1283                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1284         else
1285                 rc = smc_tx_sendmsg(smc, msg, len);
1286 out:
1287         release_sock(sk);
1288         return rc;
1289 }
1290
1291 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1292                        int flags)
1293 {
1294         struct sock *sk = sock->sk;
1295         struct smc_sock *smc;
1296         int rc = -ENOTCONN;
1297
1298         smc = smc_sk(sk);
1299         lock_sock(sk);
1300         if ((sk->sk_state == SMC_INIT) ||
1301             (sk->sk_state == SMC_LISTEN) ||
1302             (sk->sk_state == SMC_CLOSED))
1303                 goto out;
1304
1305         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1306                 rc = 0;
1307                 goto out;
1308         }
1309
1310         if (smc->use_fallback) {
1311                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1312         } else {
1313                 msg->msg_namelen = 0;
1314                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1315         }
1316
1317 out:
1318         release_sock(sk);
1319         return rc;
1320 }
1321
1322 static __poll_t smc_accept_poll(struct sock *parent)
1323 {
1324         struct smc_sock *isk = smc_sk(parent);
1325         __poll_t mask = 0;
1326
1327         spin_lock(&isk->accept_q_lock);
1328         if (!list_empty(&isk->accept_q))
1329                 mask = EPOLLIN | EPOLLRDNORM;
1330         spin_unlock(&isk->accept_q_lock);
1331
1332         return mask;
1333 }
1334
1335 static __poll_t smc_poll(struct file *file, struct socket *sock,
1336                              poll_table *wait)
1337 {
1338         struct sock *sk = sock->sk;
1339         __poll_t mask = 0;
1340         struct smc_sock *smc;
1341
1342         if (!sk)
1343                 return EPOLLNVAL;
1344
1345         smc = smc_sk(sock->sk);
1346         if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
1347                 /* delegate to CLC child sock */
1348                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1349                 sk->sk_err = smc->clcsock->sk->sk_err;
1350                 if (sk->sk_err)
1351                         mask |= EPOLLERR;
1352         } else {
1353                 if (sk->sk_state != SMC_CLOSED)
1354                         sock_poll_wait(file, sk_sleep(sk), wait);
1355                 if (sk->sk_err)
1356                         mask |= EPOLLERR;
1357                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1358                     (sk->sk_state == SMC_CLOSED))
1359                         mask |= EPOLLHUP;
1360                 if (sk->sk_state == SMC_LISTEN) {
1361                         /* woken up by sk_data_ready in smc_listen_work() */
1362                         mask = smc_accept_poll(sk);
1363                 } else {
1364                         if (atomic_read(&smc->conn.sndbuf_space) ||
1365                             sk->sk_shutdown & SEND_SHUTDOWN) {
1366                                 mask |= EPOLLOUT | EPOLLWRNORM;
1367                         } else {
1368                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1369                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1370                         }
1371                         if (atomic_read(&smc->conn.bytes_to_rcv))
1372                                 mask |= EPOLLIN | EPOLLRDNORM;
1373                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1374                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1375                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
1376                                 mask |= EPOLLIN;
1377                 }
1378                 if (smc->conn.urg_state == SMC_URG_VALID)
1379                         mask |= EPOLLPRI;
1380         }
1381
1382         return mask;
1383 }
1384
1385 static int smc_shutdown(struct socket *sock, int how)
1386 {
1387         struct sock *sk = sock->sk;
1388         struct smc_sock *smc;
1389         int rc = -EINVAL;
1390         int rc1 = 0;
1391
1392         smc = smc_sk(sk);
1393
1394         if ((how < SHUT_RD) || (how > SHUT_RDWR))
1395                 return rc;
1396
1397         lock_sock(sk);
1398
1399         rc = -ENOTCONN;
1400         if ((sk->sk_state != SMC_LISTEN) &&
1401             (sk->sk_state != SMC_ACTIVE) &&
1402             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1403             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1404             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1405             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1406             (sk->sk_state != SMC_APPFINCLOSEWAIT))
1407                 goto out;
1408         if (smc->use_fallback) {
1409                 rc = kernel_sock_shutdown(smc->clcsock, how);
1410                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1411                 if (sk->sk_shutdown == SHUTDOWN_MASK)
1412                         sk->sk_state = SMC_CLOSED;
1413                 goto out;
1414         }
1415         switch (how) {
1416         case SHUT_RDWR:         /* shutdown in both directions */
1417                 rc = smc_close_active(smc);
1418                 break;
1419         case SHUT_WR:
1420                 rc = smc_close_shutdown_write(smc);
1421                 break;
1422         case SHUT_RD:
1423                 rc = 0;
1424                 /* nothing more to do because peer is not involved */
1425                 break;
1426         }
1427         if (smc->clcsock)
1428                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1429         /* map sock_shutdown_cmd constants to sk_shutdown value range */
1430         sk->sk_shutdown |= how + 1;
1431
1432 out:
1433         release_sock(sk);
1434         return rc ? rc : rc1;
1435 }
1436
1437 static int smc_setsockopt(struct socket *sock, int level, int optname,
1438                           char __user *optval, unsigned int optlen)
1439 {
1440         struct sock *sk = sock->sk;
1441         struct smc_sock *smc;
1442         int val, rc;
1443
1444         smc = smc_sk(sk);
1445
1446         /* generic setsockopts reaching us here always apply to the
1447          * CLC socket
1448          */
1449         rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1450                                            optval, optlen);
1451         if (smc->clcsock->sk->sk_err) {
1452                 sk->sk_err = smc->clcsock->sk->sk_err;
1453                 sk->sk_error_report(sk);
1454         }
1455         if (rc)
1456                 return rc;
1457
1458         if (optlen < sizeof(int))
1459                 return -EINVAL;
1460         if (get_user(val, (int __user *)optval))
1461                 return -EFAULT;
1462
1463         lock_sock(sk);
1464         switch (optname) {
1465         case TCP_ULP:
1466         case TCP_FASTOPEN:
1467         case TCP_FASTOPEN_CONNECT:
1468         case TCP_FASTOPEN_KEY:
1469         case TCP_FASTOPEN_NO_COOKIE:
1470                 /* option not supported by SMC */
1471                 if (sk->sk_state == SMC_INIT) {
1472                         smc->use_fallback = true;
1473                 } else {
1474                         if (!smc->use_fallback)
1475                                 rc = -EINVAL;
1476                 }
1477                 break;
1478         case TCP_NODELAY:
1479                 if (sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) {
1480                         if (val && !smc->use_fallback)
1481                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1482                                                  0);
1483                 }
1484                 break;
1485         case TCP_CORK:
1486                 if (sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) {
1487                         if (!val && !smc->use_fallback)
1488                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1489                                                  0);
1490                 }
1491                 break;
1492         case TCP_DEFER_ACCEPT:
1493                 smc->sockopt_defer_accept = val;
1494                 break;
1495         default:
1496                 break;
1497         }
1498         release_sock(sk);
1499
1500         return rc;
1501 }
1502
1503 static int smc_getsockopt(struct socket *sock, int level, int optname,
1504                           char __user *optval, int __user *optlen)
1505 {
1506         struct smc_sock *smc;
1507
1508         smc = smc_sk(sock->sk);
1509         /* socket options apply to the CLC socket */
1510         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1511                                              optval, optlen);
1512 }
1513
1514 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1515                      unsigned long arg)
1516 {
1517         union smc_host_cursor cons, urg;
1518         struct smc_connection *conn;
1519         struct smc_sock *smc;
1520         int answ;
1521
1522         smc = smc_sk(sock->sk);
1523         conn = &smc->conn;
1524         if (smc->use_fallback) {
1525                 if (!smc->clcsock)
1526                         return -EBADF;
1527                 return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1528         }
1529         lock_sock(&smc->sk);
1530         switch (cmd) {
1531         case SIOCINQ: /* same as FIONREAD */
1532                 if (smc->sk.sk_state == SMC_LISTEN) {
1533                         release_sock(&smc->sk);
1534                         return -EINVAL;
1535                 }
1536                 if (smc->sk.sk_state == SMC_INIT ||
1537                     smc->sk.sk_state == SMC_CLOSED)
1538                         answ = 0;
1539                 else
1540                         answ = atomic_read(&smc->conn.bytes_to_rcv);
1541                 break;
1542         case SIOCOUTQ:
1543                 /* output queue size (not send + not acked) */
1544                 if (smc->sk.sk_state == SMC_LISTEN) {
1545                         release_sock(&smc->sk);
1546                         return -EINVAL;
1547                 }
1548                 if (smc->sk.sk_state == SMC_INIT ||
1549                     smc->sk.sk_state == SMC_CLOSED)
1550                         answ = 0;
1551                 else
1552                         answ = smc->conn.sndbuf_desc->len -
1553                                         atomic_read(&smc->conn.sndbuf_space);
1554                 break;
1555         case SIOCOUTQNSD:
1556                 /* output queue size (not send only) */
1557                 if (smc->sk.sk_state == SMC_LISTEN) {
1558                         release_sock(&smc->sk);
1559                         return -EINVAL;
1560                 }
1561                 if (smc->sk.sk_state == SMC_INIT ||
1562                     smc->sk.sk_state == SMC_CLOSED)
1563                         answ = 0;
1564                 else
1565                         answ = smc_tx_prepared_sends(&smc->conn);
1566                 break;
1567         case SIOCATMARK:
1568                 if (smc->sk.sk_state == SMC_LISTEN) {
1569                         release_sock(&smc->sk);
1570                         return -EINVAL;
1571                 }
1572                 if (smc->sk.sk_state == SMC_INIT ||
1573                     smc->sk.sk_state == SMC_CLOSED) {
1574                         answ = 0;
1575                 } else {
1576                         smc_curs_write(&cons,
1577                                smc_curs_read(&conn->local_tx_ctrl.cons, conn),
1578                                        conn);
1579                         smc_curs_write(&urg,
1580                                        smc_curs_read(&conn->urg_curs, conn),
1581                                        conn);
1582                         answ = smc_curs_diff(conn->rmb_desc->len,
1583                                              &cons, &urg) == 1;
1584                 }
1585                 break;
1586         default:
1587                 release_sock(&smc->sk);
1588                 return -ENOIOCTLCMD;
1589         }
1590         release_sock(&smc->sk);
1591
1592         return put_user(answ, (int __user *)arg);
1593 }
1594
1595 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1596                             int offset, size_t size, int flags)
1597 {
1598         struct sock *sk = sock->sk;
1599         struct smc_sock *smc;
1600         int rc = -EPIPE;
1601
1602         smc = smc_sk(sk);
1603         lock_sock(sk);
1604         if (sk->sk_state != SMC_ACTIVE) {
1605                 release_sock(sk);
1606                 goto out;
1607         }
1608         release_sock(sk);
1609         if (smc->use_fallback)
1610                 rc = kernel_sendpage(smc->clcsock, page, offset,
1611                                      size, flags);
1612         else
1613                 rc = sock_no_sendpage(sock, page, offset, size, flags);
1614
1615 out:
1616         return rc;
1617 }
1618
1619 /* Map the affected portions of the rmbe into an spd, note the number of bytes
1620  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
1621  * updates till whenever a respective page has been fully processed.
1622  * Note that subsequent recv() calls have to wait till all splice() processing
1623  * completed.
1624  */
1625 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1626                                struct pipe_inode_info *pipe, size_t len,
1627                                unsigned int flags)
1628 {
1629         struct sock *sk = sock->sk;
1630         struct smc_sock *smc;
1631         int rc = -ENOTCONN;
1632
1633         smc = smc_sk(sk);
1634         lock_sock(sk);
1635
1636         if (sk->sk_state == SMC_INIT ||
1637             sk->sk_state == SMC_LISTEN ||
1638             sk->sk_state == SMC_CLOSED)
1639                 goto out;
1640
1641         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1642                 rc = 0;
1643                 goto out;
1644         }
1645
1646         if (smc->use_fallback) {
1647                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1648                                                     pipe, len, flags);
1649         } else {
1650                 if (*ppos) {
1651                         rc = -ESPIPE;
1652                         goto out;
1653                 }
1654                 if (flags & SPLICE_F_NONBLOCK)
1655                         flags = MSG_DONTWAIT;
1656                 else
1657                         flags = 0;
1658                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
1659         }
1660 out:
1661         release_sock(sk);
1662
1663         return rc;
1664 }
1665
1666 /* must look like tcp */
1667 static const struct proto_ops smc_sock_ops = {
1668         .family         = PF_SMC,
1669         .owner          = THIS_MODULE,
1670         .release        = smc_release,
1671         .bind           = smc_bind,
1672         .connect        = smc_connect,
1673         .socketpair     = sock_no_socketpair,
1674         .accept         = smc_accept,
1675         .getname        = smc_getname,
1676         .poll           = smc_poll,
1677         .ioctl          = smc_ioctl,
1678         .listen         = smc_listen,
1679         .shutdown       = smc_shutdown,
1680         .setsockopt     = smc_setsockopt,
1681         .getsockopt     = smc_getsockopt,
1682         .sendmsg        = smc_sendmsg,
1683         .recvmsg        = smc_recvmsg,
1684         .mmap           = sock_no_mmap,
1685         .sendpage       = smc_sendpage,
1686         .splice_read    = smc_splice_read,
1687 };
1688
1689 static int smc_create(struct net *net, struct socket *sock, int protocol,
1690                       int kern)
1691 {
1692         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
1693         struct smc_sock *smc;
1694         struct sock *sk;
1695         int rc;
1696
1697         rc = -ESOCKTNOSUPPORT;
1698         if (sock->type != SOCK_STREAM)
1699                 goto out;
1700
1701         rc = -EPROTONOSUPPORT;
1702         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
1703                 goto out;
1704
1705         rc = -ENOBUFS;
1706         sock->ops = &smc_sock_ops;
1707         sk = smc_sock_alloc(net, sock, protocol);
1708         if (!sk)
1709                 goto out;
1710
1711         /* create internal TCP socket for CLC handshake and fallback */
1712         smc = smc_sk(sk);
1713         smc->use_fallback = false; /* assume rdma capability first */
1714         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
1715                               &smc->clcsock);
1716         if (rc) {
1717                 sk_common_release(sk);
1718                 goto out;
1719         }
1720         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1721         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1722
1723 out:
1724         return rc;
1725 }
1726
1727 static const struct net_proto_family smc_sock_family_ops = {
1728         .family = PF_SMC,
1729         .owner  = THIS_MODULE,
1730         .create = smc_create,
1731 };
1732
1733 static int __init smc_init(void)
1734 {
1735         int rc;
1736
1737         rc = smc_pnet_init();
1738         if (rc)
1739                 return rc;
1740
1741         rc = smc_llc_init();
1742         if (rc) {
1743                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
1744                 goto out_pnet;
1745         }
1746
1747         rc = smc_cdc_init();
1748         if (rc) {
1749                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
1750                 goto out_pnet;
1751         }
1752
1753         rc = proto_register(&smc_proto, 1);
1754         if (rc) {
1755                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
1756                 goto out_pnet;
1757         }
1758
1759         rc = proto_register(&smc_proto6, 1);
1760         if (rc) {
1761                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
1762                 goto out_proto;
1763         }
1764
1765         rc = sock_register(&smc_sock_family_ops);
1766         if (rc) {
1767                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1768                 goto out_proto6;
1769         }
1770         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
1771         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
1772
1773         rc = smc_ib_register_client();
1774         if (rc) {
1775                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1776                 goto out_sock;
1777         }
1778
1779         static_branch_enable(&tcp_have_smc);
1780         return 0;
1781
1782 out_sock:
1783         sock_unregister(PF_SMC);
1784 out_proto6:
1785         proto_unregister(&smc_proto6);
1786 out_proto:
1787         proto_unregister(&smc_proto);
1788 out_pnet:
1789         smc_pnet_exit();
1790         return rc;
1791 }
1792
1793 static void __exit smc_exit(void)
1794 {
1795         smc_core_exit();
1796         static_branch_disable(&tcp_have_smc);
1797         smc_ib_unregister_client();
1798         sock_unregister(PF_SMC);
1799         proto_unregister(&smc_proto6);
1800         proto_unregister(&smc_proto);
1801         smc_pnet_exit();
1802 }
1803
1804 module_init(smc_init);
1805 module_exit(smc_exit);
1806
1807 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1808 MODULE_DESCRIPTION("smc socket address family");
1809 MODULE_LICENSE("GPL");
1810 MODULE_ALIAS_NETPROTO(PF_SMC);