smc: remote memory buffers (RMBs)
[platform/kernel/linux-rpi.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  *    - non-blocking connect postponed
11  *    - IPv6 support postponed
12  *    - support for alternate links postponed
13  *    - partial support for non-blocking sockets only
14  *    - support for urgent data postponed
15  *
16  *  Copyright IBM Corp. 2016
17  *
18  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
19  *              based on prototype from Frank Blaschka
20  */
21
22 #define KMSG_COMPONENT "smc"
23 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
24
25 #include <linux/module.h>
26 #include <linux/socket.h>
27 #include <linux/inetdevice.h>
28 #include <linux/workqueue.h>
29 #include <net/sock.h>
30 #include <net/tcp.h>
31
32 #include "smc.h"
33 #include "smc_clc.h"
34 #include "smc_core.h"
35 #include "smc_ib.h"
36 #include "smc_pnet.h"
37
38 static DEFINE_MUTEX(smc_create_lgr_pending);    /* serialize link group
39                                                  * creation
40                                                  */
41
42 struct smc_lgr_list smc_lgr_list = {            /* established link groups */
43         .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
44         .list = LIST_HEAD_INIT(smc_lgr_list.list),
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 proto smc_proto = {
57         .name           = "SMC",
58         .owner          = THIS_MODULE,
59         .keepalive      = smc_set_keepalive,
60         .obj_size       = sizeof(struct smc_sock),
61         .slab_flags     = SLAB_DESTROY_BY_RCU,
62 };
63
64 static int smc_release(struct socket *sock)
65 {
66         struct sock *sk = sock->sk;
67         struct smc_sock *smc;
68
69         if (!sk)
70                 goto out;
71
72         smc = smc_sk(sk);
73         lock_sock(sk);
74
75         sk->sk_state = SMC_CLOSED;
76         if (smc->clcsock) {
77                 sock_release(smc->clcsock);
78                 smc->clcsock = NULL;
79         }
80
81         /* detach socket */
82         sock_orphan(sk);
83         sock->sk = NULL;
84         release_sock(sk);
85
86         sock_put(sk);
87 out:
88         return 0;
89 }
90
91 static void smc_destruct(struct sock *sk)
92 {
93         if (sk->sk_state != SMC_CLOSED)
94                 return;
95         if (!sock_flag(sk, SOCK_DEAD))
96                 return;
97
98         sk_refcnt_debug_dec(sk);
99 }
100
101 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock)
102 {
103         struct smc_sock *smc;
104         struct sock *sk;
105
106         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, &smc_proto, 0);
107         if (!sk)
108                 return NULL;
109
110         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
111         sk->sk_state = SMC_INIT;
112         sk->sk_destruct = smc_destruct;
113         sk->sk_protocol = SMCPROTO_SMC;
114         smc = smc_sk(sk);
115         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
116         INIT_LIST_HEAD(&smc->accept_q);
117         spin_lock_init(&smc->accept_q_lock);
118         sk_refcnt_debug_inc(sk);
119
120         return sk;
121 }
122
123 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
124                     int addr_len)
125 {
126         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
127         struct sock *sk = sock->sk;
128         struct smc_sock *smc;
129         int rc;
130
131         smc = smc_sk(sk);
132
133         /* replicate tests from inet_bind(), to be safe wrt. future changes */
134         rc = -EINVAL;
135         if (addr_len < sizeof(struct sockaddr_in))
136                 goto out;
137
138         rc = -EAFNOSUPPORT;
139         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
140         if ((addr->sin_family != AF_INET) &&
141             ((addr->sin_family != AF_UNSPEC) ||
142              (addr->sin_addr.s_addr != htonl(INADDR_ANY))))
143                 goto out;
144
145         lock_sock(sk);
146
147         /* Check if socket is already active */
148         rc = -EINVAL;
149         if (sk->sk_state != SMC_INIT)
150                 goto out_rel;
151
152         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
153         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
154
155 out_rel:
156         release_sock(sk);
157 out:
158         return rc;
159 }
160
161 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
162                                    unsigned long mask)
163 {
164         /* options we don't get control via setsockopt for */
165         nsk->sk_type = osk->sk_type;
166         nsk->sk_sndbuf = osk->sk_sndbuf;
167         nsk->sk_rcvbuf = osk->sk_rcvbuf;
168         nsk->sk_sndtimeo = osk->sk_sndtimeo;
169         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
170         nsk->sk_mark = osk->sk_mark;
171         nsk->sk_priority = osk->sk_priority;
172         nsk->sk_rcvlowat = osk->sk_rcvlowat;
173         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
174         nsk->sk_err = osk->sk_err;
175
176         nsk->sk_flags &= ~mask;
177         nsk->sk_flags |= osk->sk_flags & mask;
178 }
179
180 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
181                              (1UL << SOCK_KEEPOPEN) | \
182                              (1UL << SOCK_LINGER) | \
183                              (1UL << SOCK_BROADCAST) | \
184                              (1UL << SOCK_TIMESTAMP) | \
185                              (1UL << SOCK_DBG) | \
186                              (1UL << SOCK_RCVTSTAMP) | \
187                              (1UL << SOCK_RCVTSTAMPNS) | \
188                              (1UL << SOCK_LOCALROUTE) | \
189                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
190                              (1UL << SOCK_RXQ_OVFL) | \
191                              (1UL << SOCK_WIFI_STATUS) | \
192                              (1UL << SOCK_NOFCS) | \
193                              (1UL << SOCK_FILTER_LOCKED))
194 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
195  * clc socket (since smc is not called for these options from net/core)
196  */
197 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
198 {
199         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
200 }
201
202 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
203                              (1UL << SOCK_KEEPOPEN) | \
204                              (1UL << SOCK_LINGER) | \
205                              (1UL << SOCK_DBG))
206 /* copy only settings and flags relevant for smc from clc to smc socket */
207 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
208 {
209         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
210 }
211
212 /* determine subnet and mask of internal TCP socket */
213 int smc_netinfo_by_tcpsk(struct socket *clcsock,
214                          __be32 *subnet, u8 *prefix_len)
215 {
216         struct dst_entry *dst = sk_dst_get(clcsock->sk);
217         struct sockaddr_in addr;
218         int rc = -ENOENT;
219         int len;
220
221         if (!dst) {
222                 rc = -ENOTCONN;
223                 goto out;
224         }
225         if (!dst->dev) {
226                 rc = -ENODEV;
227                 goto out_rel;
228         }
229
230         /* get address to which the internal TCP socket is bound */
231         kernel_getsockname(clcsock, (struct sockaddr *)&addr, &len);
232         /* analyze IPv4 specific data of net_device belonging to TCP socket */
233         for_ifa(dst->dev->ip_ptr) {
234                 if (ifa->ifa_address != addr.sin_addr.s_addr)
235                         continue;
236                 *prefix_len = inet_mask_len(ifa->ifa_mask);
237                 *subnet = ifa->ifa_address & ifa->ifa_mask;
238                 rc = 0;
239                 break;
240         } endfor_ifa(dst->dev->ip_ptr);
241
242 out_rel:
243         dst_release(dst);
244 out:
245         return rc;
246 }
247
248 static void smc_conn_save_peer_info(struct smc_sock *smc,
249                                     struct smc_clc_msg_accept_confirm *clc)
250 {
251         smc->conn.peer_conn_idx = clc->conn_idx;
252         smc->conn.peer_rmbe_size = smc_uncompress_bufsize(clc->rmbe_size);
253         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
254 }
255
256 static void smc_link_save_peer_info(struct smc_link *link,
257                                     struct smc_clc_msg_accept_confirm *clc)
258 {
259         link->peer_qpn = ntoh24(clc->qpn);
260         memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
261         memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
262         link->peer_psn = ntoh24(clc->psn);
263         link->peer_mtu = clc->qp_mtu;
264 }
265
266 /* setup for RDMA connection of client */
267 static int smc_connect_rdma(struct smc_sock *smc)
268 {
269         struct sockaddr_in *inaddr = (struct sockaddr_in *)smc->addr;
270         struct smc_clc_msg_accept_confirm aclc;
271         int local_contact = SMC_FIRST_CONTACT;
272         struct smc_ib_device *smcibdev;
273         struct smc_link *link;
274         u8 srv_first_contact;
275         int reason_code = 0;
276         int rc = 0;
277         u8 ibport;
278
279         /* IPSec connections opt out of SMC-R optimizations */
280         if (using_ipsec(smc)) {
281                 reason_code = SMC_CLC_DECL_IPSEC;
282                 goto decline_rdma;
283         }
284
285         /* PNET table look up: search active ib_device and port
286          * within same PNETID that also contains the ethernet device
287          * used for the internal TCP socket
288          */
289         smc_pnet_find_roce_resource(smc->clcsock->sk, &smcibdev, &ibport);
290         if (!smcibdev) {
291                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
292                 goto decline_rdma;
293         }
294
295         /* do inband token exchange */
296         reason_code = smc_clc_send_proposal(smc, smcibdev, ibport);
297         if (reason_code < 0) {
298                 rc = reason_code;
299                 goto out_err;
300         }
301         if (reason_code > 0) /* configuration error */
302                 goto decline_rdma;
303         /* receive SMC Accept CLC message */
304         reason_code = smc_clc_wait_msg(smc, &aclc, sizeof(aclc),
305                                        SMC_CLC_ACCEPT);
306         if (reason_code < 0) {
307                 rc = reason_code;
308                 goto out_err;
309         }
310         if (reason_code > 0)
311                 goto decline_rdma;
312
313         srv_first_contact = aclc.hdr.flag;
314         mutex_lock(&smc_create_lgr_pending);
315         local_contact = smc_conn_create(smc, inaddr->sin_addr.s_addr, smcibdev,
316                                         ibport, &aclc.lcl, srv_first_contact);
317         if (local_contact < 0) {
318                 rc = local_contact;
319                 if (rc == -ENOMEM)
320                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
321                 else if (rc == -ENOLINK)
322                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
323                 goto decline_rdma_unlock;
324         }
325         link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
326
327         smc_conn_save_peer_info(smc, &aclc);
328
329         rc = smc_sndbuf_create(smc);
330         if (rc) {
331                 reason_code = SMC_CLC_DECL_MEM;
332                 goto decline_rdma_unlock;
333         }
334         rc = smc_rmb_create(smc);
335         if (rc) {
336                 reason_code = SMC_CLC_DECL_MEM;
337                 goto decline_rdma_unlock;
338         }
339
340         if (local_contact == SMC_FIRST_CONTACT)
341                 smc_link_save_peer_info(link, &aclc);
342         /* tbd in follow-on patch: more steps to setup RDMA communcication,
343          * create rmbs, map rmbs, rtoken_handling, modify_qp
344          */
345
346         rc = smc_clc_send_confirm(smc);
347         if (rc)
348                 goto out_err_unlock;
349
350         /* tbd in follow-on patch: llc_confirm */
351
352         mutex_unlock(&smc_create_lgr_pending);
353 out_connected:
354         smc_copy_sock_settings_to_clc(smc);
355         smc->sk.sk_state = SMC_ACTIVE;
356
357         return rc ? rc : local_contact;
358
359 decline_rdma_unlock:
360         mutex_unlock(&smc_create_lgr_pending);
361         smc_conn_free(&smc->conn);
362 decline_rdma:
363         /* RDMA setup failed, switch back to TCP */
364         smc->use_fallback = true;
365         if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
366                 rc = smc_clc_send_decline(smc, reason_code, 0);
367                 if (rc < sizeof(struct smc_clc_msg_decline))
368                         goto out_err;
369         }
370         goto out_connected;
371
372 out_err_unlock:
373         mutex_unlock(&smc_create_lgr_pending);
374         smc_conn_free(&smc->conn);
375 out_err:
376         return rc;
377 }
378
379 static int smc_connect(struct socket *sock, struct sockaddr *addr,
380                        int alen, int flags)
381 {
382         struct sock *sk = sock->sk;
383         struct smc_sock *smc;
384         int rc = -EINVAL;
385
386         smc = smc_sk(sk);
387
388         /* separate smc parameter checking to be safe */
389         if (alen < sizeof(addr->sa_family))
390                 goto out_err;
391         if (addr->sa_family != AF_INET)
392                 goto out_err;
393         smc->addr = addr;       /* needed for nonblocking connect */
394
395         lock_sock(sk);
396         switch (sk->sk_state) {
397         default:
398                 goto out;
399         case SMC_ACTIVE:
400                 rc = -EISCONN;
401                 goto out;
402         case SMC_INIT:
403                 rc = 0;
404                 break;
405         }
406
407         smc_copy_sock_settings_to_clc(smc);
408         rc = kernel_connect(smc->clcsock, addr, alen, flags);
409         if (rc)
410                 goto out;
411
412         /* setup RDMA connection */
413         rc = smc_connect_rdma(smc);
414         if (rc < 0)
415                 goto out;
416         else
417                 rc = 0; /* success cases including fallback */
418
419 out:
420         release_sock(sk);
421 out_err:
422         return rc;
423 }
424
425 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
426 {
427         struct sock *sk = &lsmc->sk;
428         struct socket *new_clcsock;
429         struct sock *new_sk;
430         int rc;
431
432         release_sock(&lsmc->sk);
433         new_sk = smc_sock_alloc(sock_net(sk), NULL);
434         if (!new_sk) {
435                 rc = -ENOMEM;
436                 lsmc->sk.sk_err = ENOMEM;
437                 *new_smc = NULL;
438                 lock_sock(&lsmc->sk);
439                 goto out;
440         }
441         *new_smc = smc_sk(new_sk);
442
443         rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
444         lock_sock(&lsmc->sk);
445         if  (rc < 0) {
446                 lsmc->sk.sk_err = -rc;
447                 new_sk->sk_state = SMC_CLOSED;
448                 sock_set_flag(new_sk, SOCK_DEAD);
449                 sock_put(new_sk);
450                 *new_smc = NULL;
451                 goto out;
452         }
453         if (lsmc->sk.sk_state == SMC_CLOSED) {
454                 if (new_clcsock)
455                         sock_release(new_clcsock);
456                 new_sk->sk_state = SMC_CLOSED;
457                 sock_set_flag(new_sk, SOCK_DEAD);
458                 sock_put(new_sk);
459                 *new_smc = NULL;
460                 goto out;
461         }
462
463         (*new_smc)->clcsock = new_clcsock;
464 out:
465         return rc;
466 }
467
468 /* add a just created sock to the accept queue of the listen sock as
469  * candidate for a following socket accept call from user space
470  */
471 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
472 {
473         struct smc_sock *par = smc_sk(parent);
474
475         sock_hold(sk);
476         spin_lock(&par->accept_q_lock);
477         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
478         spin_unlock(&par->accept_q_lock);
479         sk_acceptq_added(parent);
480 }
481
482 /* remove a socket from the accept queue of its parental listening socket */
483 static void smc_accept_unlink(struct sock *sk)
484 {
485         struct smc_sock *par = smc_sk(sk)->listen_smc;
486
487         spin_lock(&par->accept_q_lock);
488         list_del_init(&smc_sk(sk)->accept_q);
489         spin_unlock(&par->accept_q_lock);
490         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
491         sock_put(sk);
492 }
493
494 /* remove a sock from the accept queue to bind it to a new socket created
495  * for a socket accept call from user space
496  */
497 static struct sock *smc_accept_dequeue(struct sock *parent,
498                                        struct socket *new_sock)
499 {
500         struct smc_sock *isk, *n;
501         struct sock *new_sk;
502
503         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
504                 new_sk = (struct sock *)isk;
505
506                 smc_accept_unlink(new_sk);
507                 if (new_sk->sk_state == SMC_CLOSED) {
508                         /* tbd in follow-on patch: close this sock */
509                         continue;
510                 }
511                 if (new_sock)
512                         sock_graft(new_sk, new_sock);
513                 return new_sk;
514         }
515         return NULL;
516 }
517
518 /* clean up for a created but never accepted sock */
519 static void smc_close_non_accepted(struct sock *sk)
520 {
521         struct smc_sock *smc = smc_sk(sk);
522
523         sock_hold(sk);
524         if (smc->clcsock) {
525                 struct socket *tcp;
526
527                 tcp = smc->clcsock;
528                 smc->clcsock = NULL;
529                 sock_release(tcp);
530         }
531         /* more closing stuff to be added with socket closing patch */
532         sock_put(sk);
533 }
534
535 /* setup for RDMA connection of server */
536 static void smc_listen_work(struct work_struct *work)
537 {
538         struct smc_sock *new_smc = container_of(work, struct smc_sock,
539                                                 smc_listen_work);
540         struct socket *newclcsock = new_smc->clcsock;
541         struct smc_sock *lsmc = new_smc->listen_smc;
542         struct smc_clc_msg_accept_confirm cclc;
543         int local_contact = SMC_REUSE_CONTACT;
544         struct sock *newsmcsk = &new_smc->sk;
545         struct smc_clc_msg_proposal pclc;
546         struct smc_ib_device *smcibdev;
547         struct sockaddr_in peeraddr;
548         struct smc_link *link;
549         int reason_code = 0;
550         int rc = 0, len;
551         __be32 subnet;
552         u8 prefix_len;
553         u8 ibport;
554
555         /* do inband token exchange -
556          *wait for and receive SMC Proposal CLC message
557          */
558         reason_code = smc_clc_wait_msg(new_smc, &pclc, sizeof(pclc),
559                                        SMC_CLC_PROPOSAL);
560         if (reason_code < 0)
561                 goto out_err;
562         if (reason_code > 0)
563                 goto decline_rdma;
564
565         /* IPSec connections opt out of SMC-R optimizations */
566         if (using_ipsec(new_smc)) {
567                 reason_code = SMC_CLC_DECL_IPSEC;
568                 goto decline_rdma;
569         }
570
571         /* PNET table look up: search active ib_device and port
572          * within same PNETID that also contains the ethernet device
573          * used for the internal TCP socket
574          */
575         smc_pnet_find_roce_resource(newclcsock->sk, &smcibdev, &ibport);
576         if (!smcibdev) {
577                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
578                 goto decline_rdma;
579         }
580
581         /* determine subnet and mask from internal TCP socket */
582         rc = smc_netinfo_by_tcpsk(newclcsock, &subnet, &prefix_len);
583         if (rc) {
584                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
585                 goto decline_rdma;
586         }
587         if ((pclc.outgoing_subnet != subnet) ||
588             (pclc.prefix_len != prefix_len)) {
589                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
590                 goto decline_rdma;
591         }
592
593         /* get address of the peer connected to the internal TCP socket */
594         kernel_getpeername(newclcsock, (struct sockaddr *)&peeraddr, &len);
595
596         /* allocate connection / link group */
597         mutex_lock(&smc_create_lgr_pending);
598         local_contact = smc_conn_create(new_smc, peeraddr.sin_addr.s_addr,
599                                         smcibdev, ibport, &pclc.lcl, 0);
600         if (local_contact == SMC_REUSE_CONTACT)
601                 /* lock no longer needed, free it due to following
602                  * smc_clc_wait_msg() call
603                  */
604                 mutex_unlock(&smc_create_lgr_pending);
605         if (local_contact < 0) {
606                 rc = local_contact;
607                 if (rc == -ENOMEM)
608                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
609                 else if (rc == -ENOLINK)
610                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
611                 goto decline_rdma;
612         }
613         link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
614
615         rc = smc_sndbuf_create(new_smc);
616         if (rc) {
617                 reason_code = SMC_CLC_DECL_MEM;
618                 goto decline_rdma;
619         }
620         rc = smc_rmb_create(new_smc);
621         if (rc) {
622                 reason_code = SMC_CLC_DECL_MEM;
623                 goto decline_rdma;
624         }
625
626         rc = smc_clc_send_accept(new_smc, local_contact);
627         if (rc)
628                 goto out_err;
629
630         /* receive SMC Confirm CLC message */
631         reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
632                                        SMC_CLC_CONFIRM);
633         if (reason_code < 0)
634                 goto out_err;
635         if (reason_code > 0)
636                 goto decline_rdma;
637         smc_conn_save_peer_info(new_smc, &cclc);
638         if (local_contact == SMC_FIRST_CONTACT)
639                 smc_link_save_peer_info(link, &cclc);
640
641         /* tbd in follow-on patch: more steps to setup RDMA communcication,
642          * rtoken_handling, modify_qp
643          */
644
645 out_connected:
646         sk_refcnt_debug_inc(newsmcsk);
647         newsmcsk->sk_state = SMC_ACTIVE;
648 enqueue:
649         if (local_contact == SMC_FIRST_CONTACT)
650                 mutex_unlock(&smc_create_lgr_pending);
651         lock_sock(&lsmc->sk);
652         if (lsmc->sk.sk_state == SMC_LISTEN) {
653                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
654         } else { /* no longer listening */
655                 smc_close_non_accepted(newsmcsk);
656         }
657         release_sock(&lsmc->sk);
658
659         /* Wake up accept */
660         lsmc->sk.sk_data_ready(&lsmc->sk);
661         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
662         return;
663
664 decline_rdma:
665         /* RDMA setup failed, switch back to TCP */
666         smc_conn_free(&new_smc->conn);
667         new_smc->use_fallback = true;
668         if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
669                 rc = smc_clc_send_decline(new_smc, reason_code, 0);
670                 if (rc < sizeof(struct smc_clc_msg_decline))
671                         goto out_err;
672         }
673         goto out_connected;
674
675 out_err:
676         newsmcsk->sk_state = SMC_CLOSED;
677         goto enqueue; /* queue new sock with sk_err set */
678 }
679
680 static void smc_tcp_listen_work(struct work_struct *work)
681 {
682         struct smc_sock *lsmc = container_of(work, struct smc_sock,
683                                              tcp_listen_work);
684         struct smc_sock *new_smc;
685         int rc = 0;
686
687         lock_sock(&lsmc->sk);
688         while (lsmc->sk.sk_state == SMC_LISTEN) {
689                 rc = smc_clcsock_accept(lsmc, &new_smc);
690                 if (rc)
691                         goto out;
692                 if (!new_smc)
693                         continue;
694
695                 new_smc->listen_smc = lsmc;
696                 new_smc->use_fallback = false; /* assume rdma capability first*/
697                 sock_hold(&lsmc->sk); /* sock_put in smc_listen_work */
698                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
699                 smc_copy_sock_settings_to_smc(new_smc);
700                 schedule_work(&new_smc->smc_listen_work);
701         }
702
703 out:
704         release_sock(&lsmc->sk);
705         lsmc->sk.sk_data_ready(&lsmc->sk); /* no more listening, wake accept */
706 }
707
708 static int smc_listen(struct socket *sock, int backlog)
709 {
710         struct sock *sk = sock->sk;
711         struct smc_sock *smc;
712         int rc;
713
714         smc = smc_sk(sk);
715         lock_sock(sk);
716
717         rc = -EINVAL;
718         if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
719                 goto out;
720
721         rc = 0;
722         if (sk->sk_state == SMC_LISTEN) {
723                 sk->sk_max_ack_backlog = backlog;
724                 goto out;
725         }
726         /* some socket options are handled in core, so we could not apply
727          * them to the clc socket -- copy smc socket options to clc socket
728          */
729         smc_copy_sock_settings_to_clc(smc);
730
731         rc = kernel_listen(smc->clcsock, backlog);
732         if (rc)
733                 goto out;
734         sk->sk_max_ack_backlog = backlog;
735         sk->sk_ack_backlog = 0;
736         sk->sk_state = SMC_LISTEN;
737         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
738         schedule_work(&smc->tcp_listen_work);
739
740 out:
741         release_sock(sk);
742         return rc;
743 }
744
745 static int smc_accept(struct socket *sock, struct socket *new_sock,
746                       int flags)
747 {
748         struct sock *sk = sock->sk, *nsk;
749         DECLARE_WAITQUEUE(wait, current);
750         struct smc_sock *lsmc;
751         long timeo;
752         int rc = 0;
753
754         lsmc = smc_sk(sk);
755         lock_sock(sk);
756
757         if (lsmc->sk.sk_state != SMC_LISTEN) {
758                 rc = -EINVAL;
759                 goto out;
760         }
761
762         /* Wait for an incoming connection */
763         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
764         add_wait_queue_exclusive(sk_sleep(sk), &wait);
765         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
766                 set_current_state(TASK_INTERRUPTIBLE);
767                 if (!timeo) {
768                         rc = -EAGAIN;
769                         break;
770                 }
771                 release_sock(sk);
772                 timeo = schedule_timeout(timeo);
773                 /* wakeup by sk_data_ready in smc_listen_work() */
774                 sched_annotate_sleep();
775                 lock_sock(sk);
776                 if (signal_pending(current)) {
777                         rc = sock_intr_errno(timeo);
778                         break;
779                 }
780         }
781         set_current_state(TASK_RUNNING);
782         remove_wait_queue(sk_sleep(sk), &wait);
783
784         if (!rc)
785                 rc = sock_error(nsk);
786
787 out:
788         release_sock(sk);
789         return rc;
790 }
791
792 static int smc_getname(struct socket *sock, struct sockaddr *addr,
793                        int *len, int peer)
794 {
795         struct smc_sock *smc;
796
797         if (peer && (sock->sk->sk_state != SMC_ACTIVE))
798                 return -ENOTCONN;
799
800         smc = smc_sk(sock->sk);
801
802         return smc->clcsock->ops->getname(smc->clcsock, addr, len, peer);
803 }
804
805 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
806 {
807         struct sock *sk = sock->sk;
808         struct smc_sock *smc;
809         int rc = -EPIPE;
810
811         smc = smc_sk(sk);
812         lock_sock(sk);
813         if (sk->sk_state != SMC_ACTIVE)
814                 goto out;
815         if (smc->use_fallback)
816                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
817         else
818                 rc = sock_no_sendmsg(sock, msg, len);
819 out:
820         release_sock(sk);
821         return rc;
822 }
823
824 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
825                        int flags)
826 {
827         struct sock *sk = sock->sk;
828         struct smc_sock *smc;
829         int rc = -ENOTCONN;
830
831         smc = smc_sk(sk);
832         lock_sock(sk);
833         if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
834                 goto out;
835
836         if (smc->use_fallback)
837                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
838         else
839                 rc = sock_no_recvmsg(sock, msg, len, flags);
840 out:
841         release_sock(sk);
842         return rc;
843 }
844
845 static unsigned int smc_accept_poll(struct sock *parent)
846 {
847         struct smc_sock *isk;
848         struct sock *sk;
849
850         lock_sock(parent);
851         list_for_each_entry(isk, &smc_sk(parent)->accept_q, accept_q) {
852                 sk = (struct sock *)isk;
853
854                 if (sk->sk_state == SMC_ACTIVE) {
855                         release_sock(parent);
856                         return POLLIN | POLLRDNORM;
857                 }
858         }
859         release_sock(parent);
860
861         return 0;
862 }
863
864 static unsigned int smc_poll(struct file *file, struct socket *sock,
865                              poll_table *wait)
866 {
867         struct sock *sk = sock->sk;
868         unsigned int mask = 0;
869         struct smc_sock *smc;
870         int rc;
871
872         smc = smc_sk(sock->sk);
873         if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
874                 /* delegate to CLC child sock */
875                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
876                 /* if non-blocking connect finished ... */
877                 lock_sock(sk);
878                 if ((sk->sk_state == SMC_INIT) && (mask & POLLOUT)) {
879                         sk->sk_err = smc->clcsock->sk->sk_err;
880                         if (sk->sk_err) {
881                                 mask |= POLLERR;
882                         } else {
883                                 rc = smc_connect_rdma(smc);
884                                 if (rc < 0)
885                                         mask |= POLLERR;
886                                 else
887                                         /* success cases including fallback */
888                                         mask |= POLLOUT | POLLWRNORM;
889                         }
890                 }
891                 release_sock(sk);
892         } else {
893                 sock_poll_wait(file, sk_sleep(sk), wait);
894                 if (sk->sk_state == SMC_LISTEN)
895                         /* woken up by sk_data_ready in smc_listen_work() */
896                         mask |= smc_accept_poll(sk);
897                 if (sk->sk_err)
898                         mask |= POLLERR;
899                 /* for now - to be enhanced in follow-on patch */
900         }
901
902         return mask;
903 }
904
905 static int smc_shutdown(struct socket *sock, int how)
906 {
907         struct sock *sk = sock->sk;
908         struct smc_sock *smc;
909         int rc = -EINVAL;
910
911         smc = smc_sk(sk);
912
913         if ((how < SHUT_RD) || (how > SHUT_RDWR))
914                 goto out_err;
915
916         lock_sock(sk);
917
918         rc = -ENOTCONN;
919         if (sk->sk_state == SMC_CLOSED)
920                 goto out;
921         if (smc->use_fallback) {
922                 rc = kernel_sock_shutdown(smc->clcsock, how);
923                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
924                 if (sk->sk_shutdown == SHUTDOWN_MASK)
925                         sk->sk_state = SMC_CLOSED;
926         } else {
927                 rc = sock_no_shutdown(sock, how);
928         }
929
930 out:
931         release_sock(sk);
932
933 out_err:
934         return rc;
935 }
936
937 static int smc_setsockopt(struct socket *sock, int level, int optname,
938                           char __user *optval, unsigned int optlen)
939 {
940         struct sock *sk = sock->sk;
941         struct smc_sock *smc;
942
943         smc = smc_sk(sk);
944
945         /* generic setsockopts reaching us here always apply to the
946          * CLC socket
947          */
948         return smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
949                                              optval, optlen);
950 }
951
952 static int smc_getsockopt(struct socket *sock, int level, int optname,
953                           char __user *optval, int __user *optlen)
954 {
955         struct smc_sock *smc;
956
957         smc = smc_sk(sock->sk);
958         /* socket options apply to the CLC socket */
959         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
960                                              optval, optlen);
961 }
962
963 static int smc_ioctl(struct socket *sock, unsigned int cmd,
964                      unsigned long arg)
965 {
966         struct smc_sock *smc;
967
968         smc = smc_sk(sock->sk);
969         if (smc->use_fallback)
970                 return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
971         else
972                 return sock_no_ioctl(sock, cmd, arg);
973 }
974
975 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
976                             int offset, size_t size, int flags)
977 {
978         struct sock *sk = sock->sk;
979         struct smc_sock *smc;
980         int rc = -EPIPE;
981
982         smc = smc_sk(sk);
983         lock_sock(sk);
984         if (sk->sk_state != SMC_ACTIVE)
985                 goto out;
986         if (smc->use_fallback)
987                 rc = kernel_sendpage(smc->clcsock, page, offset,
988                                      size, flags);
989         else
990                 rc = sock_no_sendpage(sock, page, offset, size, flags);
991
992 out:
993         release_sock(sk);
994         return rc;
995 }
996
997 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
998                                struct pipe_inode_info *pipe, size_t len,
999                                     unsigned int flags)
1000 {
1001         struct sock *sk = sock->sk;
1002         struct smc_sock *smc;
1003         int rc = -ENOTCONN;
1004
1005         smc = smc_sk(sk);
1006         lock_sock(sk);
1007         if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
1008                 goto out;
1009         if (smc->use_fallback) {
1010                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1011                                                     pipe, len, flags);
1012         } else {
1013                 rc = -EOPNOTSUPP;
1014         }
1015 out:
1016         release_sock(sk);
1017         return rc;
1018 }
1019
1020 /* must look like tcp */
1021 static const struct proto_ops smc_sock_ops = {
1022         .family         = PF_SMC,
1023         .owner          = THIS_MODULE,
1024         .release        = smc_release,
1025         .bind           = smc_bind,
1026         .connect        = smc_connect,
1027         .socketpair     = sock_no_socketpair,
1028         .accept         = smc_accept,
1029         .getname        = smc_getname,
1030         .poll           = smc_poll,
1031         .ioctl          = smc_ioctl,
1032         .listen         = smc_listen,
1033         .shutdown       = smc_shutdown,
1034         .setsockopt     = smc_setsockopt,
1035         .getsockopt     = smc_getsockopt,
1036         .sendmsg        = smc_sendmsg,
1037         .recvmsg        = smc_recvmsg,
1038         .mmap           = sock_no_mmap,
1039         .sendpage       = smc_sendpage,
1040         .splice_read    = smc_splice_read,
1041 };
1042
1043 static int smc_create(struct net *net, struct socket *sock, int protocol,
1044                       int kern)
1045 {
1046         struct smc_sock *smc;
1047         struct sock *sk;
1048         int rc;
1049
1050         rc = -ESOCKTNOSUPPORT;
1051         if (sock->type != SOCK_STREAM)
1052                 goto out;
1053
1054         rc = -EPROTONOSUPPORT;
1055         if ((protocol != IPPROTO_IP) && (protocol != IPPROTO_TCP))
1056                 goto out;
1057
1058         rc = -ENOBUFS;
1059         sock->ops = &smc_sock_ops;
1060         sk = smc_sock_alloc(net, sock);
1061         if (!sk)
1062                 goto out;
1063
1064         /* create internal TCP socket for CLC handshake and fallback */
1065         smc = smc_sk(sk);
1066         smc->use_fallback = false; /* assume rdma capability first */
1067         rc = sock_create_kern(net, PF_INET, SOCK_STREAM,
1068                               IPPROTO_TCP, &smc->clcsock);
1069         if (rc)
1070                 sk_common_release(sk);
1071         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1072         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1073
1074 out:
1075         return rc;
1076 }
1077
1078 static const struct net_proto_family smc_sock_family_ops = {
1079         .family = PF_SMC,
1080         .owner  = THIS_MODULE,
1081         .create = smc_create,
1082 };
1083
1084 static int __init smc_init(void)
1085 {
1086         int rc;
1087
1088         rc = smc_pnet_init();
1089         if (rc)
1090                 return rc;
1091
1092         rc = proto_register(&smc_proto, 1);
1093         if (rc) {
1094                 pr_err("%s: proto_register fails with %d\n", __func__, rc);
1095                 goto out_pnet;
1096         }
1097
1098         rc = sock_register(&smc_sock_family_ops);
1099         if (rc) {
1100                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1101                 goto out_proto;
1102         }
1103
1104         rc = smc_ib_register_client();
1105         if (rc) {
1106                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1107                 goto out_sock;
1108         }
1109
1110         return 0;
1111
1112 out_sock:
1113         sock_unregister(PF_SMC);
1114 out_proto:
1115         proto_unregister(&smc_proto);
1116 out_pnet:
1117         smc_pnet_exit();
1118         return rc;
1119 }
1120
1121 static void __exit smc_exit(void)
1122 {
1123         struct smc_link_group *lgr, *lg;
1124         LIST_HEAD(lgr_freeing_list);
1125
1126         spin_lock_bh(&smc_lgr_list.lock);
1127         if (!list_empty(&smc_lgr_list.list))
1128                 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1129         spin_unlock_bh(&smc_lgr_list.lock);
1130         list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1131                 list_del_init(&lgr->list);
1132                 smc_lgr_free(lgr); /* free link group */
1133         }
1134         smc_ib_unregister_client();
1135         sock_unregister(PF_SMC);
1136         proto_unregister(&smc_proto);
1137         smc_pnet_exit();
1138 }
1139
1140 module_init(smc_init);
1141 module_exit(smc_exit);
1142
1143 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1144 MODULE_DESCRIPTION("smc socket address family");
1145 MODULE_LICENSE("GPL");
1146 MODULE_ALIAS_NETPROTO(PF_SMC);