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