5fda37decc55d7be8eb3e8ed01c6948f64f4c5e6
[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 }
253
254 static void smc_link_save_peer_info(struct smc_link *link,
255                                     struct smc_clc_msg_accept_confirm *clc)
256 {
257         link->peer_qpn = ntoh24(clc->qpn);
258         memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
259         memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
260         link->peer_psn = ntoh24(clc->psn);
261         link->peer_mtu = clc->qp_mtu;
262 }
263
264 /* setup for RDMA connection of client */
265 static int smc_connect_rdma(struct smc_sock *smc)
266 {
267         struct sockaddr_in *inaddr = (struct sockaddr_in *)smc->addr;
268         struct smc_clc_msg_accept_confirm aclc;
269         int local_contact = SMC_FIRST_CONTACT;
270         struct smc_ib_device *smcibdev;
271         struct smc_link *link;
272         u8 srv_first_contact;
273         int reason_code = 0;
274         int rc = 0;
275         u8 ibport;
276
277         /* IPSec connections opt out of SMC-R optimizations */
278         if (using_ipsec(smc)) {
279                 reason_code = SMC_CLC_DECL_IPSEC;
280                 goto decline_rdma;
281         }
282
283         /* PNET table look up: search active ib_device and port
284          * within same PNETID that also contains the ethernet device
285          * used for the internal TCP socket
286          */
287         smc_pnet_find_roce_resource(smc->clcsock->sk, &smcibdev, &ibport);
288         if (!smcibdev) {
289                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
290                 goto decline_rdma;
291         }
292
293         /* do inband token exchange */
294         reason_code = smc_clc_send_proposal(smc, smcibdev, ibport);
295         if (reason_code < 0) {
296                 rc = reason_code;
297                 goto out_err;
298         }
299         if (reason_code > 0) /* configuration error */
300                 goto decline_rdma;
301         /* receive SMC Accept CLC message */
302         reason_code = smc_clc_wait_msg(smc, &aclc, sizeof(aclc),
303                                        SMC_CLC_ACCEPT);
304         if (reason_code < 0) {
305                 rc = reason_code;
306                 goto out_err;
307         }
308         if (reason_code > 0)
309                 goto decline_rdma;
310
311         srv_first_contact = aclc.hdr.flag;
312         mutex_lock(&smc_create_lgr_pending);
313         local_contact = smc_conn_create(smc, inaddr->sin_addr.s_addr, smcibdev,
314                                         ibport, &aclc.lcl, srv_first_contact);
315         if (local_contact < 0) {
316                 rc = local_contact;
317                 if (rc == -ENOMEM)
318                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
319                 else if (rc == -ENOLINK)
320                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
321                 goto decline_rdma_unlock;
322         }
323         link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
324
325         smc_conn_save_peer_info(smc, &aclc);
326         if (local_contact == SMC_FIRST_CONTACT)
327                 smc_link_save_peer_info(link, &aclc);
328         /* tbd in follow-on patch: more steps to setup RDMA communcication,
329          * create rmbs, map rmbs, rtoken_handling, modify_qp
330          */
331
332         rc = smc_clc_send_confirm(smc);
333         if (rc)
334                 goto out_err_unlock;
335
336         /* tbd in follow-on patch: llc_confirm */
337
338         mutex_unlock(&smc_create_lgr_pending);
339 out_connected:
340         smc_copy_sock_settings_to_clc(smc);
341         smc->sk.sk_state = SMC_ACTIVE;
342
343         return rc ? rc : local_contact;
344
345 decline_rdma_unlock:
346         mutex_unlock(&smc_create_lgr_pending);
347         smc_conn_free(&smc->conn);
348 decline_rdma:
349         /* RDMA setup failed, switch back to TCP */
350         smc->use_fallback = true;
351         if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
352                 rc = smc_clc_send_decline(smc, reason_code, 0);
353                 if (rc < sizeof(struct smc_clc_msg_decline))
354                         goto out_err;
355         }
356         goto out_connected;
357
358 out_err_unlock:
359         mutex_unlock(&smc_create_lgr_pending);
360         smc_conn_free(&smc->conn);
361 out_err:
362         return rc;
363 }
364
365 static int smc_connect(struct socket *sock, struct sockaddr *addr,
366                        int alen, int flags)
367 {
368         struct sock *sk = sock->sk;
369         struct smc_sock *smc;
370         int rc = -EINVAL;
371
372         smc = smc_sk(sk);
373
374         /* separate smc parameter checking to be safe */
375         if (alen < sizeof(addr->sa_family))
376                 goto out_err;
377         if (addr->sa_family != AF_INET)
378                 goto out_err;
379         smc->addr = addr;       /* needed for nonblocking connect */
380
381         lock_sock(sk);
382         switch (sk->sk_state) {
383         default:
384                 goto out;
385         case SMC_ACTIVE:
386                 rc = -EISCONN;
387                 goto out;
388         case SMC_INIT:
389                 rc = 0;
390                 break;
391         }
392
393         smc_copy_sock_settings_to_clc(smc);
394         rc = kernel_connect(smc->clcsock, addr, alen, flags);
395         if (rc)
396                 goto out;
397
398         /* setup RDMA connection */
399         rc = smc_connect_rdma(smc);
400         if (rc < 0)
401                 goto out;
402         else
403                 rc = 0; /* success cases including fallback */
404
405 out:
406         release_sock(sk);
407 out_err:
408         return rc;
409 }
410
411 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
412 {
413         struct sock *sk = &lsmc->sk;
414         struct socket *new_clcsock;
415         struct sock *new_sk;
416         int rc;
417
418         release_sock(&lsmc->sk);
419         new_sk = smc_sock_alloc(sock_net(sk), NULL);
420         if (!new_sk) {
421                 rc = -ENOMEM;
422                 lsmc->sk.sk_err = ENOMEM;
423                 *new_smc = NULL;
424                 lock_sock(&lsmc->sk);
425                 goto out;
426         }
427         *new_smc = smc_sk(new_sk);
428
429         rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
430         lock_sock(&lsmc->sk);
431         if  (rc < 0) {
432                 lsmc->sk.sk_err = -rc;
433                 new_sk->sk_state = SMC_CLOSED;
434                 sock_set_flag(new_sk, SOCK_DEAD);
435                 sock_put(new_sk);
436                 *new_smc = NULL;
437                 goto out;
438         }
439         if (lsmc->sk.sk_state == SMC_CLOSED) {
440                 if (new_clcsock)
441                         sock_release(new_clcsock);
442                 new_sk->sk_state = SMC_CLOSED;
443                 sock_set_flag(new_sk, SOCK_DEAD);
444                 sock_put(new_sk);
445                 *new_smc = NULL;
446                 goto out;
447         }
448
449         (*new_smc)->clcsock = new_clcsock;
450 out:
451         return rc;
452 }
453
454 /* add a just created sock to the accept queue of the listen sock as
455  * candidate for a following socket accept call from user space
456  */
457 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
458 {
459         struct smc_sock *par = smc_sk(parent);
460
461         sock_hold(sk);
462         spin_lock(&par->accept_q_lock);
463         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
464         spin_unlock(&par->accept_q_lock);
465         sk_acceptq_added(parent);
466 }
467
468 /* remove a socket from the accept queue of its parental listening socket */
469 static void smc_accept_unlink(struct sock *sk)
470 {
471         struct smc_sock *par = smc_sk(sk)->listen_smc;
472
473         spin_lock(&par->accept_q_lock);
474         list_del_init(&smc_sk(sk)->accept_q);
475         spin_unlock(&par->accept_q_lock);
476         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
477         sock_put(sk);
478 }
479
480 /* remove a sock from the accept queue to bind it to a new socket created
481  * for a socket accept call from user space
482  */
483 static struct sock *smc_accept_dequeue(struct sock *parent,
484                                        struct socket *new_sock)
485 {
486         struct smc_sock *isk, *n;
487         struct sock *new_sk;
488
489         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
490                 new_sk = (struct sock *)isk;
491
492                 smc_accept_unlink(new_sk);
493                 if (new_sk->sk_state == SMC_CLOSED) {
494                         /* tbd in follow-on patch: close this sock */
495                         continue;
496                 }
497                 if (new_sock)
498                         sock_graft(new_sk, new_sock);
499                 return new_sk;
500         }
501         return NULL;
502 }
503
504 /* clean up for a created but never accepted sock */
505 static void smc_close_non_accepted(struct sock *sk)
506 {
507         struct smc_sock *smc = smc_sk(sk);
508
509         sock_hold(sk);
510         if (smc->clcsock) {
511                 struct socket *tcp;
512
513                 tcp = smc->clcsock;
514                 smc->clcsock = NULL;
515                 sock_release(tcp);
516         }
517         /* more closing stuff to be added with socket closing patch */
518         sock_put(sk);
519 }
520
521 /* setup for RDMA connection of server */
522 static void smc_listen_work(struct work_struct *work)
523 {
524         struct smc_sock *new_smc = container_of(work, struct smc_sock,
525                                                 smc_listen_work);
526         struct socket *newclcsock = new_smc->clcsock;
527         struct smc_sock *lsmc = new_smc->listen_smc;
528         struct smc_clc_msg_accept_confirm cclc;
529         int local_contact = SMC_REUSE_CONTACT;
530         struct sock *newsmcsk = &new_smc->sk;
531         struct smc_clc_msg_proposal pclc;
532         struct smc_ib_device *smcibdev;
533         struct sockaddr_in peeraddr;
534         struct smc_link *link;
535         int reason_code = 0;
536         int rc = 0, len;
537         __be32 subnet;
538         u8 prefix_len;
539         u8 ibport;
540
541         /* do inband token exchange -
542          *wait for and receive SMC Proposal CLC message
543          */
544         reason_code = smc_clc_wait_msg(new_smc, &pclc, sizeof(pclc),
545                                        SMC_CLC_PROPOSAL);
546         if (reason_code < 0)
547                 goto out_err;
548         if (reason_code > 0)
549                 goto decline_rdma;
550
551         /* IPSec connections opt out of SMC-R optimizations */
552         if (using_ipsec(new_smc)) {
553                 reason_code = SMC_CLC_DECL_IPSEC;
554                 goto decline_rdma;
555         }
556
557         /* PNET table look up: search active ib_device and port
558          * within same PNETID that also contains the ethernet device
559          * used for the internal TCP socket
560          */
561         smc_pnet_find_roce_resource(newclcsock->sk, &smcibdev, &ibport);
562         if (!smcibdev) {
563                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
564                 goto decline_rdma;
565         }
566
567         /* determine subnet and mask from internal TCP socket */
568         rc = smc_netinfo_by_tcpsk(newclcsock, &subnet, &prefix_len);
569         if (rc) {
570                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
571                 goto decline_rdma;
572         }
573         if ((pclc.outgoing_subnet != subnet) ||
574             (pclc.prefix_len != prefix_len)) {
575                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
576                 goto decline_rdma;
577         }
578
579         /* get address of the peer connected to the internal TCP socket */
580         kernel_getpeername(newclcsock, (struct sockaddr *)&peeraddr, &len);
581
582         /* allocate connection / link group */
583         mutex_lock(&smc_create_lgr_pending);
584         local_contact = smc_conn_create(new_smc, peeraddr.sin_addr.s_addr,
585                                         smcibdev, ibport, &pclc.lcl, 0);
586         if (local_contact == SMC_REUSE_CONTACT)
587                 /* lock no longer needed, free it due to following
588                  * smc_clc_wait_msg() call
589                  */
590                 mutex_unlock(&smc_create_lgr_pending);
591         if (local_contact < 0) {
592                 rc = local_contact;
593                 if (rc == -ENOMEM)
594                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
595                 else if (rc == -ENOLINK)
596                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
597                 goto decline_rdma;
598         }
599         link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
600
601         /* tbd in follow-on patch: more steps to setup RDMA communcication,
602          * create rmbs, map rmbs
603          */
604
605         rc = smc_clc_send_accept(new_smc, local_contact);
606         if (rc)
607                 goto out_err;
608
609         /* receive SMC Confirm CLC message */
610         reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
611                                        SMC_CLC_CONFIRM);
612         if (reason_code < 0)
613                 goto out_err;
614         if (reason_code > 0)
615                 goto decline_rdma;
616         smc_conn_save_peer_info(new_smc, &cclc);
617         if (local_contact == SMC_FIRST_CONTACT)
618                 smc_link_save_peer_info(link, &cclc);
619
620         /* tbd in follow-on patch: more steps to setup RDMA communcication,
621          * rtoken_handling, modify_qp
622          */
623
624 out_connected:
625         sk_refcnt_debug_inc(newsmcsk);
626         newsmcsk->sk_state = SMC_ACTIVE;
627 enqueue:
628         if (local_contact == SMC_FIRST_CONTACT)
629                 mutex_unlock(&smc_create_lgr_pending);
630         lock_sock(&lsmc->sk);
631         if (lsmc->sk.sk_state == SMC_LISTEN) {
632                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
633         } else { /* no longer listening */
634                 smc_close_non_accepted(newsmcsk);
635         }
636         release_sock(&lsmc->sk);
637
638         /* Wake up accept */
639         lsmc->sk.sk_data_ready(&lsmc->sk);
640         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
641         return;
642
643 decline_rdma:
644         /* RDMA setup failed, switch back to TCP */
645         smc_conn_free(&new_smc->conn);
646         new_smc->use_fallback = true;
647         if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
648                 rc = smc_clc_send_decline(new_smc, reason_code, 0);
649                 if (rc < sizeof(struct smc_clc_msg_decline))
650                         goto out_err;
651         }
652         goto out_connected;
653
654 out_err:
655         newsmcsk->sk_state = SMC_CLOSED;
656         goto enqueue; /* queue new sock with sk_err set */
657 }
658
659 static void smc_tcp_listen_work(struct work_struct *work)
660 {
661         struct smc_sock *lsmc = container_of(work, struct smc_sock,
662                                              tcp_listen_work);
663         struct smc_sock *new_smc;
664         int rc = 0;
665
666         lock_sock(&lsmc->sk);
667         while (lsmc->sk.sk_state == SMC_LISTEN) {
668                 rc = smc_clcsock_accept(lsmc, &new_smc);
669                 if (rc)
670                         goto out;
671                 if (!new_smc)
672                         continue;
673
674                 new_smc->listen_smc = lsmc;
675                 new_smc->use_fallback = false; /* assume rdma capability first*/
676                 sock_hold(&lsmc->sk); /* sock_put in smc_listen_work */
677                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
678                 smc_copy_sock_settings_to_smc(new_smc);
679                 schedule_work(&new_smc->smc_listen_work);
680         }
681
682 out:
683         release_sock(&lsmc->sk);
684         lsmc->sk.sk_data_ready(&lsmc->sk); /* no more listening, wake accept */
685 }
686
687 static int smc_listen(struct socket *sock, int backlog)
688 {
689         struct sock *sk = sock->sk;
690         struct smc_sock *smc;
691         int rc;
692
693         smc = smc_sk(sk);
694         lock_sock(sk);
695
696         rc = -EINVAL;
697         if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
698                 goto out;
699
700         rc = 0;
701         if (sk->sk_state == SMC_LISTEN) {
702                 sk->sk_max_ack_backlog = backlog;
703                 goto out;
704         }
705         /* some socket options are handled in core, so we could not apply
706          * them to the clc socket -- copy smc socket options to clc socket
707          */
708         smc_copy_sock_settings_to_clc(smc);
709
710         rc = kernel_listen(smc->clcsock, backlog);
711         if (rc)
712                 goto out;
713         sk->sk_max_ack_backlog = backlog;
714         sk->sk_ack_backlog = 0;
715         sk->sk_state = SMC_LISTEN;
716         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
717         schedule_work(&smc->tcp_listen_work);
718
719 out:
720         release_sock(sk);
721         return rc;
722 }
723
724 static int smc_accept(struct socket *sock, struct socket *new_sock,
725                       int flags)
726 {
727         struct sock *sk = sock->sk, *nsk;
728         DECLARE_WAITQUEUE(wait, current);
729         struct smc_sock *lsmc;
730         long timeo;
731         int rc = 0;
732
733         lsmc = smc_sk(sk);
734         lock_sock(sk);
735
736         if (lsmc->sk.sk_state != SMC_LISTEN) {
737                 rc = -EINVAL;
738                 goto out;
739         }
740
741         /* Wait for an incoming connection */
742         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
743         add_wait_queue_exclusive(sk_sleep(sk), &wait);
744         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
745                 set_current_state(TASK_INTERRUPTIBLE);
746                 if (!timeo) {
747                         rc = -EAGAIN;
748                         break;
749                 }
750                 release_sock(sk);
751                 timeo = schedule_timeout(timeo);
752                 /* wakeup by sk_data_ready in smc_listen_work() */
753                 sched_annotate_sleep();
754                 lock_sock(sk);
755                 if (signal_pending(current)) {
756                         rc = sock_intr_errno(timeo);
757                         break;
758                 }
759         }
760         set_current_state(TASK_RUNNING);
761         remove_wait_queue(sk_sleep(sk), &wait);
762
763         if (!rc)
764                 rc = sock_error(nsk);
765
766 out:
767         release_sock(sk);
768         return rc;
769 }
770
771 static int smc_getname(struct socket *sock, struct sockaddr *addr,
772                        int *len, int peer)
773 {
774         struct smc_sock *smc;
775
776         if (peer && (sock->sk->sk_state != SMC_ACTIVE))
777                 return -ENOTCONN;
778
779         smc = smc_sk(sock->sk);
780
781         return smc->clcsock->ops->getname(smc->clcsock, addr, len, peer);
782 }
783
784 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
785 {
786         struct sock *sk = sock->sk;
787         struct smc_sock *smc;
788         int rc = -EPIPE;
789
790         smc = smc_sk(sk);
791         lock_sock(sk);
792         if (sk->sk_state != SMC_ACTIVE)
793                 goto out;
794         if (smc->use_fallback)
795                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
796         else
797                 rc = sock_no_sendmsg(sock, msg, len);
798 out:
799         release_sock(sk);
800         return rc;
801 }
802
803 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
804                        int flags)
805 {
806         struct sock *sk = sock->sk;
807         struct smc_sock *smc;
808         int rc = -ENOTCONN;
809
810         smc = smc_sk(sk);
811         lock_sock(sk);
812         if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
813                 goto out;
814
815         if (smc->use_fallback)
816                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
817         else
818                 rc = sock_no_recvmsg(sock, msg, len, flags);
819 out:
820         release_sock(sk);
821         return rc;
822 }
823
824 static unsigned int smc_accept_poll(struct sock *parent)
825 {
826         struct smc_sock *isk;
827         struct sock *sk;
828
829         lock_sock(parent);
830         list_for_each_entry(isk, &smc_sk(parent)->accept_q, accept_q) {
831                 sk = (struct sock *)isk;
832
833                 if (sk->sk_state == SMC_ACTIVE) {
834                         release_sock(parent);
835                         return POLLIN | POLLRDNORM;
836                 }
837         }
838         release_sock(parent);
839
840         return 0;
841 }
842
843 static unsigned int smc_poll(struct file *file, struct socket *sock,
844                              poll_table *wait)
845 {
846         struct sock *sk = sock->sk;
847         unsigned int mask = 0;
848         struct smc_sock *smc;
849         int rc;
850
851         smc = smc_sk(sock->sk);
852         if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
853                 /* delegate to CLC child sock */
854                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
855                 /* if non-blocking connect finished ... */
856                 lock_sock(sk);
857                 if ((sk->sk_state == SMC_INIT) && (mask & POLLOUT)) {
858                         sk->sk_err = smc->clcsock->sk->sk_err;
859                         if (sk->sk_err) {
860                                 mask |= POLLERR;
861                         } else {
862                                 rc = smc_connect_rdma(smc);
863                                 if (rc < 0)
864                                         mask |= POLLERR;
865                                 else
866                                         /* success cases including fallback */
867                                         mask |= POLLOUT | POLLWRNORM;
868                         }
869                 }
870                 release_sock(sk);
871         } else {
872                 sock_poll_wait(file, sk_sleep(sk), wait);
873                 if (sk->sk_state == SMC_LISTEN)
874                         /* woken up by sk_data_ready in smc_listen_work() */
875                         mask |= smc_accept_poll(sk);
876                 if (sk->sk_err)
877                         mask |= POLLERR;
878                 /* for now - to be enhanced in follow-on patch */
879         }
880
881         return mask;
882 }
883
884 static int smc_shutdown(struct socket *sock, int how)
885 {
886         struct sock *sk = sock->sk;
887         struct smc_sock *smc;
888         int rc = -EINVAL;
889
890         smc = smc_sk(sk);
891
892         if ((how < SHUT_RD) || (how > SHUT_RDWR))
893                 goto out_err;
894
895         lock_sock(sk);
896
897         rc = -ENOTCONN;
898         if (sk->sk_state == SMC_CLOSED)
899                 goto out;
900         if (smc->use_fallback) {
901                 rc = kernel_sock_shutdown(smc->clcsock, how);
902                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
903                 if (sk->sk_shutdown == SHUTDOWN_MASK)
904                         sk->sk_state = SMC_CLOSED;
905         } else {
906                 rc = sock_no_shutdown(sock, how);
907         }
908
909 out:
910         release_sock(sk);
911
912 out_err:
913         return rc;
914 }
915
916 static int smc_setsockopt(struct socket *sock, int level, int optname,
917                           char __user *optval, unsigned int optlen)
918 {
919         struct sock *sk = sock->sk;
920         struct smc_sock *smc;
921
922         smc = smc_sk(sk);
923
924         /* generic setsockopts reaching us here always apply to the
925          * CLC socket
926          */
927         return smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
928                                              optval, optlen);
929 }
930
931 static int smc_getsockopt(struct socket *sock, int level, int optname,
932                           char __user *optval, int __user *optlen)
933 {
934         struct smc_sock *smc;
935
936         smc = smc_sk(sock->sk);
937         /* socket options apply to the CLC socket */
938         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
939                                              optval, optlen);
940 }
941
942 static int smc_ioctl(struct socket *sock, unsigned int cmd,
943                      unsigned long arg)
944 {
945         struct smc_sock *smc;
946
947         smc = smc_sk(sock->sk);
948         if (smc->use_fallback)
949                 return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
950         else
951                 return sock_no_ioctl(sock, cmd, arg);
952 }
953
954 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
955                             int offset, size_t size, int flags)
956 {
957         struct sock *sk = sock->sk;
958         struct smc_sock *smc;
959         int rc = -EPIPE;
960
961         smc = smc_sk(sk);
962         lock_sock(sk);
963         if (sk->sk_state != SMC_ACTIVE)
964                 goto out;
965         if (smc->use_fallback)
966                 rc = kernel_sendpage(smc->clcsock, page, offset,
967                                      size, flags);
968         else
969                 rc = sock_no_sendpage(sock, page, offset, size, flags);
970
971 out:
972         release_sock(sk);
973         return rc;
974 }
975
976 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
977                                struct pipe_inode_info *pipe, size_t len,
978                                     unsigned int flags)
979 {
980         struct sock *sk = sock->sk;
981         struct smc_sock *smc;
982         int rc = -ENOTCONN;
983
984         smc = smc_sk(sk);
985         lock_sock(sk);
986         if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
987                 goto out;
988         if (smc->use_fallback) {
989                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
990                                                     pipe, len, flags);
991         } else {
992                 rc = -EOPNOTSUPP;
993         }
994 out:
995         release_sock(sk);
996         return rc;
997 }
998
999 /* must look like tcp */
1000 static const struct proto_ops smc_sock_ops = {
1001         .family         = PF_SMC,
1002         .owner          = THIS_MODULE,
1003         .release        = smc_release,
1004         .bind           = smc_bind,
1005         .connect        = smc_connect,
1006         .socketpair     = sock_no_socketpair,
1007         .accept         = smc_accept,
1008         .getname        = smc_getname,
1009         .poll           = smc_poll,
1010         .ioctl          = smc_ioctl,
1011         .listen         = smc_listen,
1012         .shutdown       = smc_shutdown,
1013         .setsockopt     = smc_setsockopt,
1014         .getsockopt     = smc_getsockopt,
1015         .sendmsg        = smc_sendmsg,
1016         .recvmsg        = smc_recvmsg,
1017         .mmap           = sock_no_mmap,
1018         .sendpage       = smc_sendpage,
1019         .splice_read    = smc_splice_read,
1020 };
1021
1022 static int smc_create(struct net *net, struct socket *sock, int protocol,
1023                       int kern)
1024 {
1025         struct smc_sock *smc;
1026         struct sock *sk;
1027         int rc;
1028
1029         rc = -ESOCKTNOSUPPORT;
1030         if (sock->type != SOCK_STREAM)
1031                 goto out;
1032
1033         rc = -EPROTONOSUPPORT;
1034         if ((protocol != IPPROTO_IP) && (protocol != IPPROTO_TCP))
1035                 goto out;
1036
1037         rc = -ENOBUFS;
1038         sock->ops = &smc_sock_ops;
1039         sk = smc_sock_alloc(net, sock);
1040         if (!sk)
1041                 goto out;
1042
1043         /* create internal TCP socket for CLC handshake and fallback */
1044         smc = smc_sk(sk);
1045         smc->use_fallback = false; /* assume rdma capability first */
1046         rc = sock_create_kern(net, PF_INET, SOCK_STREAM,
1047                               IPPROTO_TCP, &smc->clcsock);
1048         if (rc)
1049                 sk_common_release(sk);
1050
1051 out:
1052         return rc;
1053 }
1054
1055 static const struct net_proto_family smc_sock_family_ops = {
1056         .family = PF_SMC,
1057         .owner  = THIS_MODULE,
1058         .create = smc_create,
1059 };
1060
1061 static int __init smc_init(void)
1062 {
1063         int rc;
1064
1065         rc = smc_pnet_init();
1066         if (rc)
1067                 return rc;
1068
1069         rc = proto_register(&smc_proto, 1);
1070         if (rc) {
1071                 pr_err("%s: proto_register fails with %d\n", __func__, rc);
1072                 goto out_pnet;
1073         }
1074
1075         rc = sock_register(&smc_sock_family_ops);
1076         if (rc) {
1077                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1078                 goto out_proto;
1079         }
1080
1081         rc = smc_ib_register_client();
1082         if (rc) {
1083                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1084                 goto out_sock;
1085         }
1086
1087         return 0;
1088
1089 out_sock:
1090         sock_unregister(PF_SMC);
1091 out_proto:
1092         proto_unregister(&smc_proto);
1093 out_pnet:
1094         smc_pnet_exit();
1095         return rc;
1096 }
1097
1098 static void __exit smc_exit(void)
1099 {
1100         struct smc_link_group *lgr, *lg;
1101         LIST_HEAD(lgr_freeing_list);
1102
1103         spin_lock_bh(&smc_lgr_list.lock);
1104         if (!list_empty(&smc_lgr_list.list))
1105                 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1106         spin_unlock_bh(&smc_lgr_list.lock);
1107         list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1108                 list_del_init(&lgr->list);
1109                 smc_lgr_free(lgr); /* free link group */
1110         }
1111         smc_ib_unregister_client();
1112         sock_unregister(PF_SMC);
1113         proto_unregister(&smc_proto);
1114         smc_pnet_exit();
1115 }
1116
1117 module_init(smc_init);
1118 module_exit(smc_exit);
1119
1120 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1121 MODULE_DESCRIPTION("smc socket address family");
1122 MODULE_LICENSE("GPL");
1123 MODULE_ALIAS_NETPROTO(PF_SMC);