net/smc/af_smc.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  *  AF_SMC protocol family socket handler keeping the AF_INET sock address type
   6  *  applies to SOCK_STREAM sockets only
   7  *  offers an alternative communication option for TCP-protocol sockets
   8  *  applicable with RoCE-cards only
   9  *
  10  *  Initial restrictions:
  11  *    - support for alternate links postponed
  12  *
  13  *  Copyright IBM Corp. 2016, 2018
  14  *
  15  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  16  *              based on prototype from Frank Blaschka
  17  */
  18
  19 #define KMSG_COMPONENT "smc"
  20 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  21
  22 #include <linux/module.h>
  23 #include <linux/socket.h>
  24 #include <linux/workqueue.h>
  25 #include <linux/in.h>
  26 #include <linux/sched/signal.h>
  27 #include <linux/if_vlan.h>
  28 #include <linux/rcupdate_wait.h>
  29 #include <linux/ctype.h>
  30
  31 #include <net/sock.h>
  32 #include <net/tcp.h>
  33 #include <net/smc.h>
  34 #include <asm/ioctls.h>
  35
  36 #include <net/net_namespace.h>
  37 #include <net/netns/generic.h>
  38 #include "smc_netns.h"
  39
  40 #include "smc.h"
  41 #include "smc_clc.h"
  42 #include "smc_llc.h"
  43 #include "smc_cdc.h"
  44 #include "smc_core.h"
  45 #include "smc_ib.h"
  46 #include "smc_ism.h"
  47 #include "smc_pnet.h"
  48 #include "smc_netlink.h"
  49 #include "smc_tx.h"
  50 #include "smc_rx.h"
  51 #include "smc_close.h"
  52
  53 static DEFINE_MUTEX(smc_server_lgr_pending);    /* serialize link group
  54                                                  * creation on server
  55                                                  */
  56 static DEFINE_MUTEX(smc_client_lgr_pending);    /* serialize link group
  57                                                  * creation on client
  58                                                  */
  59
  60 struct workqueue_struct *smc_hs_wq;     /* wq for handshake work */
  61 struct workqueue_struct *smc_close_wq;  /* wq for close work */
  62
  63 static void smc_tcp_listen_work(struct work_struct *);
  64 static void smc_connect_work(struct work_struct *);
  65
  66 static void smc_set_keepalive(struct sock *sk, int val)
  67 {
  68         struct smc_sock *smc = smc_sk(sk);
  69
  70         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
  71 }
  72
  73 static struct smc_hashinfo smc_v4_hashinfo = {
  74         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
  75 };
  76
  77 static struct smc_hashinfo smc_v6_hashinfo = {
  78         .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
  79 };
  80
  81 int smc_hash_sk(struct sock *sk)
  82 {
  83         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  84         struct hlist_head *head;
  85
  86         head = &h->ht;
  87
  88         write_lock_bh(&h->lock);
  89         sk_add_node(sk, head);
  90         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
  91         write_unlock_bh(&h->lock);
  92
  93         return 0;
  94 }
  95 EXPORT_SYMBOL_GPL(smc_hash_sk);
  96
  97 void smc_unhash_sk(struct sock *sk)
  98 {
  99         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
 100
 101         write_lock_bh(&h->lock);
 102         if (sk_del_node_init(sk))
 103                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 104         write_unlock_bh(&h->lock);
 105 }
 106 EXPORT_SYMBOL_GPL(smc_unhash_sk);
 107
 108 struct proto smc_proto = {
 109         .name           = "SMC",
 110         .owner          = THIS_MODULE,
 111         .keepalive      = smc_set_keepalive,
 112         .hash           = smc_hash_sk,
 113         .unhash         = smc_unhash_sk,
 114         .obj_size       = sizeof(struct smc_sock),
 115         .h.smc_hash     = &smc_v4_hashinfo,
 116         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 117 };
 118 EXPORT_SYMBOL_GPL(smc_proto);
 119
 120 struct proto smc_proto6 = {
 121         .name           = "SMC6",
 122         .owner          = THIS_MODULE,
 123         .keepalive      = smc_set_keepalive,
 124         .hash           = smc_hash_sk,
 125         .unhash         = smc_unhash_sk,
 126         .obj_size       = sizeof(struct smc_sock),
 127         .h.smc_hash     = &smc_v6_hashinfo,
 128         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 129 };
 130 EXPORT_SYMBOL_GPL(smc_proto6);
 131
 132 static void smc_restore_fallback_changes(struct smc_sock *smc)
 133 {
 134         if (smc->clcsock->file) { /* non-accepted sockets have no file yet */
 135                 smc->clcsock->file->private_data = smc->sk.sk_socket;
 136                 smc->clcsock->file = NULL;
 137         }
 138 }
 139
 140 static int __smc_release(struct smc_sock *smc)
 141 {
 142         struct sock *sk = &smc->sk;
 143         int rc = 0;
 144
 145         if (!smc->use_fallback) {
 146                 rc = smc_close_active(smc);
 147                 sock_set_flag(sk, SOCK_DEAD);
 148                 sk->sk_shutdown |= SHUTDOWN_MASK;
 149         } else {
 150                 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
 151                         sock_put(sk); /* passive closing */
 152                 if (sk->sk_state == SMC_LISTEN) {
 153                         /* wake up clcsock accept */
 154                         rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
 155                 }
 156                 sk->sk_state = SMC_CLOSED;
 157                 sk->sk_state_change(sk);
 158                 smc_restore_fallback_changes(smc);
 159         }
 160
 161         sk->sk_prot->unhash(sk);
 162
 163         if (sk->sk_state == SMC_CLOSED) {
 164                 if (smc->clcsock) {
 165                         release_sock(sk);
 166                         smc_clcsock_release(smc);
 167                         lock_sock(sk);
 168                 }
 169                 if (!smc->use_fallback)
 170                         smc_conn_free(&smc->conn);
 171         }
 172
 173         return rc;
 174 }
 175
 176 static int smc_release(struct socket *sock)
 177 {
 178         struct sock *sk = sock->sk;
 179         struct smc_sock *smc;
 180         int rc = 0;
 181
 182         if (!sk)
 183                 goto out;
 184
 185         sock_hold(sk); /* sock_put below */
 186         smc = smc_sk(sk);
 187
 188         /* cleanup for a dangling non-blocking connect */
 189         if (smc->connect_nonblock && sk->sk_state == SMC_INIT)
 190                 tcp_abort(smc->clcsock->sk, ECONNABORTED);
 191         flush_work(&smc->connect_work);
 192
 193         if (sk->sk_state == SMC_LISTEN)
 194                 /* smc_close_non_accepted() is called and acquires
 195                  * sock lock for child sockets again
 196                  */
 197                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 198         else
 199                 lock_sock(sk);
 200
 201         rc = __smc_release(smc);
 202
 203         /* detach socket */
 204         sock_orphan(sk);
 205         sock->sk = NULL;
 206         release_sock(sk);
 207
 208         sock_put(sk); /* sock_hold above */
 209         sock_put(sk); /* final sock_put */
 210 out:
 211         return rc;
 212 }
 213
 214 static void smc_destruct(struct sock *sk)
 215 {
 216         if (sk->sk_state != SMC_CLOSED)
 217                 return;
 218         if (!sock_flag(sk, SOCK_DEAD))
 219                 return;
 220
 221         sk_refcnt_debug_dec(sk);
 222 }
 223
 224 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 225                                    int protocol)
 226 {
 227         struct smc_sock *smc;
 228         struct proto *prot;
 229         struct sock *sk;
 230
 231         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
 232         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
 233         if (!sk)
 234                 return NULL;
 235
 236         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 237         sk->sk_state = SMC_INIT;
 238         sk->sk_destruct = smc_destruct;
 239         sk->sk_protocol = protocol;
 240         smc = smc_sk(sk);
 241         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 242         INIT_WORK(&smc->connect_work, smc_connect_work);
 243         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 244         INIT_LIST_HEAD(&smc->accept_q);
 245         spin_lock_init(&smc->accept_q_lock);
 246         spin_lock_init(&smc->conn.send_lock);
 247         sk->sk_prot->hash(sk);
 248         sk_refcnt_debug_inc(sk);
 249         mutex_init(&smc->clcsock_release_lock);
 250
 251         return sk;
 252 }
 253
 254 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 255                     int addr_len)
 256 {
 257         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 258         struct sock *sk = sock->sk;
 259         struct smc_sock *smc;
 260         int rc;
 261
 262         smc = smc_sk(sk);
 263
 264         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 265         rc = -EINVAL;
 266         if (addr_len < sizeof(struct sockaddr_in))
 267                 goto out;
 268
 269         rc = -EAFNOSUPPORT;
 270         if (addr->sin_family != AF_INET &&
 271             addr->sin_family != AF_INET6 &&
 272             addr->sin_family != AF_UNSPEC)
 273                 goto out;
 274         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 275         if (addr->sin_family == AF_UNSPEC &&
 276             addr->sin_addr.s_addr != htonl(INADDR_ANY))
 277                 goto out;
 278
 279         lock_sock(sk);
 280
 281         /* Check if socket is already active */
 282         rc = -EINVAL;
 283         if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
 284                 goto out_rel;
 285
 286         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 287         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 288
 289 out_rel:
 290         release_sock(sk);
 291 out:
 292         return rc;
 293 }
 294
 295 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 296                                    unsigned long mask)
 297 {
 298         /* options we don't get control via setsockopt for */
 299         nsk->sk_type = osk->sk_type;
 300         nsk->sk_sndbuf = osk->sk_sndbuf;
 301         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 302         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 303         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 304         nsk->sk_mark = osk->sk_mark;
 305         nsk->sk_priority = osk->sk_priority;
 306         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 307         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 308         nsk->sk_err = osk->sk_err;
 309
 310         nsk->sk_flags &= ~mask;
 311         nsk->sk_flags |= osk->sk_flags & mask;
 312 }
 313
 314 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 315                              (1UL << SOCK_KEEPOPEN) | \
 316                              (1UL << SOCK_LINGER) | \
 317                              (1UL << SOCK_BROADCAST) | \
 318                              (1UL << SOCK_TIMESTAMP) | \
 319                              (1UL << SOCK_DBG) | \
 320                              (1UL << SOCK_RCVTSTAMP) | \
 321                              (1UL << SOCK_RCVTSTAMPNS) | \
 322                              (1UL << SOCK_LOCALROUTE) | \
 323                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 324                              (1UL << SOCK_RXQ_OVFL) | \
 325                              (1UL << SOCK_WIFI_STATUS) | \
 326                              (1UL << SOCK_NOFCS) | \
 327                              (1UL << SOCK_FILTER_LOCKED) | \
 328                              (1UL << SOCK_TSTAMP_NEW))
 329 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 330  * clc socket (since smc is not called for these options from net/core)
 331  */
 332 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 333 {
 334         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 335 }
 336
 337 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 338                              (1UL << SOCK_KEEPOPEN) | \
 339                              (1UL << SOCK_LINGER) | \
 340                              (1UL << SOCK_DBG))
 341 /* copy only settings and flags relevant for smc from clc to smc socket */
 342 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 343 {
 344         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 345 }
 346
 347 /* register the new rmb on all links */
 348 static int smcr_lgr_reg_rmbs(struct smc_link *link,
 349                              struct smc_buf_desc *rmb_desc)
 350 {
 351         struct smc_link_group *lgr = link->lgr;
 352         int i, rc = 0;
 353
 354         rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
 355         if (rc)
 356                 return rc;
 357         /* protect against parallel smc_llc_cli_rkey_exchange() and
 358          * parallel smcr_link_reg_rmb()
 359          */
 360         mutex_lock(&lgr->llc_conf_mutex);
 361         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 362                 if (!smc_link_active(&lgr->lnk[i]))
 363                         continue;
 364                 rc = smcr_link_reg_rmb(&lgr->lnk[i], rmb_desc);
 365                 if (rc)
 366                         goto out;
 367         }
 368
 369         /* exchange confirm_rkey msg with peer */
 370         rc = smc_llc_do_confirm_rkey(link, rmb_desc);
 371         if (rc) {
 372                 rc = -EFAULT;
 373                 goto out;
 374         }
 375         rmb_desc->is_conf_rkey = true;
 376 out:
 377         mutex_unlock(&lgr->llc_conf_mutex);
 378         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
 379         return rc;
 380 }
 381
 382 static int smcr_clnt_conf_first_link(struct smc_sock *smc)
 383 {
 384         struct smc_link *link = smc->conn.lnk;
 385         struct smc_llc_qentry *qentry;
 386         int rc;
 387
 388         /* receive CONFIRM LINK request from server over RoCE fabric */
 389         qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
 390                               SMC_LLC_CONFIRM_LINK);
 391         if (!qentry) {
 392                 struct smc_clc_msg_decline dclc;
 393
 394                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 395                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 396                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
 397         }
 398         smc_llc_save_peer_uid(qentry);
 399         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
 400         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
 401         if (rc)
 402                 return SMC_CLC_DECL_RMBE_EC;
 403
 404         rc = smc_ib_modify_qp_rts(link);
 405         if (rc)
 406                 return SMC_CLC_DECL_ERR_RDYLNK;
 407
 408         smc_wr_remember_qp_attr(link);
 409
 410         if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
 411                 return SMC_CLC_DECL_ERR_REGRMB;
 412
 413         /* confirm_rkey is implicit on 1st contact */
 414         smc->conn.rmb_desc->is_conf_rkey = true;
 415
 416         /* send CONFIRM LINK response over RoCE fabric */
 417         rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
 418         if (rc < 0)
 419                 return SMC_CLC_DECL_TIMEOUT_CL;
 420
 421         smc_llc_link_active(link);
 422         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
 423
 424         /* optional 2nd link, receive ADD LINK request from server */
 425         qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
 426                               SMC_LLC_ADD_LINK);
 427         if (!qentry) {
 428                 struct smc_clc_msg_decline dclc;
 429
 430                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 431                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 432                 if (rc == -EAGAIN)
 433                         rc = 0; /* no DECLINE received, go with one link */
 434                 return rc;
 435         }
 436         smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
 437         smc_llc_cli_add_link(link, qentry);
 438         return 0;
 439 }
 440
 441 static void smcr_conn_save_peer_info(struct smc_sock *smc,
 442                                      struct smc_clc_msg_accept_confirm *clc)
 443 {
 444         int bufsize = smc_uncompress_bufsize(clc->r0.rmbe_size);
 445
 446         smc->conn.peer_rmbe_idx = clc->r0.rmbe_idx;
 447         smc->conn.local_tx_ctrl.token = ntohl(clc->r0.rmbe_alert_token);
 448         smc->conn.peer_rmbe_size = bufsize;
 449         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 450         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
 451 }
 452
 453 static bool smc_isascii(char *hostname)
 454 {
 455         int i;
 456
 457         for (i = 0; i < SMC_MAX_HOSTNAME_LEN; i++)
 458                 if (!isascii(hostname[i]))
 459                         return false;
 460         return true;
 461 }
 462
 463 static void smcd_conn_save_peer_info(struct smc_sock *smc,
 464                                      struct smc_clc_msg_accept_confirm *clc)
 465 {
 466         int bufsize = smc_uncompress_bufsize(clc->d0.dmbe_size);
 467
 468         smc->conn.peer_rmbe_idx = clc->d0.dmbe_idx;
 469         smc->conn.peer_token = clc->d0.token;
 470         /* msg header takes up space in the buffer */
 471         smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
 472         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 473         smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
 474         if (clc->hdr.version > SMC_V1 &&
 475             (clc->hdr.typev2 & SMC_FIRST_CONTACT_MASK)) {
 476                 struct smc_clc_msg_accept_confirm_v2 *clc_v2 =
 477                         (struct smc_clc_msg_accept_confirm_v2 *)clc;
 478                 struct smc_clc_first_contact_ext *fce =
 479                         (struct smc_clc_first_contact_ext *)
 480                                 (((u8 *)clc_v2) + sizeof(*clc_v2));
 481
 482                 memcpy(smc->conn.lgr->negotiated_eid, clc_v2->eid,
 483                        SMC_MAX_EID_LEN);
 484                 smc->conn.lgr->peer_os = fce->os_type;
 485                 smc->conn.lgr->peer_smc_release = fce->release;
 486                 if (smc_isascii(fce->hostname))
 487                         memcpy(smc->conn.lgr->peer_hostname, fce->hostname,
 488                                SMC_MAX_HOSTNAME_LEN);
 489         }
 490 }
 491
 492 static void smc_conn_save_peer_info(struct smc_sock *smc,
 493                                     struct smc_clc_msg_accept_confirm *clc)
 494 {
 495         if (smc->conn.lgr->is_smcd)
 496                 smcd_conn_save_peer_info(smc, clc);
 497         else
 498                 smcr_conn_save_peer_info(smc, clc);
 499 }
 500
 501 static void smc_link_save_peer_info(struct smc_link *link,
 502                                     struct smc_clc_msg_accept_confirm *clc)
 503 {
 504         link->peer_qpn = ntoh24(clc->r0.qpn);
 505         memcpy(link->peer_gid, clc->r0.lcl.gid, SMC_GID_SIZE);
 506         memcpy(link->peer_mac, clc->r0.lcl.mac, sizeof(link->peer_mac));
 507         link->peer_psn = ntoh24(clc->r0.psn);
 508         link->peer_mtu = clc->r0.qp_mtu;
 509 }
 510
 511 static void smc_switch_to_fallback(struct smc_sock *smc)
 512 {
 513         smc->use_fallback = true;
 514         if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
 515                 smc->clcsock->file = smc->sk.sk_socket->file;
 516                 smc->clcsock->file->private_data = smc->clcsock;
 517                 smc->clcsock->wq.fasync_list =
 518                         smc->sk.sk_socket->wq.fasync_list;
 519         }
 520 }
 521
 522 /* fall back during connect */
 523 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
 524 {
 525         smc_switch_to_fallback(smc);
 526         smc->fallback_rsn = reason_code;
 527         smc_copy_sock_settings_to_clc(smc);
 528         smc->connect_nonblock = 0;
 529         if (smc->sk.sk_state == SMC_INIT)
 530                 smc->sk.sk_state = SMC_ACTIVE;
 531         return 0;
 532 }
 533
 534 /* decline and fall back during connect */
 535 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code,
 536                                         u8 version)
 537 {
 538         int rc;
 539
 540         if (reason_code < 0) { /* error, fallback is not possible */
 541                 if (smc->sk.sk_state == SMC_INIT)
 542                         sock_put(&smc->sk); /* passive closing */
 543                 return reason_code;
 544         }
 545         if (reason_code != SMC_CLC_DECL_PEERDECL) {
 546                 rc = smc_clc_send_decline(smc, reason_code, version);
 547                 if (rc < 0) {
 548                         if (smc->sk.sk_state == SMC_INIT)
 549                                 sock_put(&smc->sk); /* passive closing */
 550                         return rc;
 551                 }
 552         }
 553         return smc_connect_fallback(smc, reason_code);
 554 }
 555
 556 static void smc_conn_abort(struct smc_sock *smc, int local_first)
 557 {
 558         if (local_first)
 559                 smc_lgr_cleanup_early(&smc->conn);
 560         else
 561                 smc_conn_free(&smc->conn);
 562 }
 563
 564 /* check if there is a rdma device available for this connection. */
 565 /* called for connect and listen */
 566 static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
 567 {
 568         /* PNET table look up: search active ib_device and port
 569          * within same PNETID that also contains the ethernet device
 570          * used for the internal TCP socket
 571          */
 572         smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
 573         if (!ini->ib_dev)
 574                 return SMC_CLC_DECL_NOSMCRDEV;
 575         return 0;
 576 }
 577
 578 /* check if there is an ISM device available for this connection. */
 579 /* called for connect and listen */
 580 static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
 581 {
 582         /* Find ISM device with same PNETID as connecting interface  */
 583         smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
 584         if (!ini->ism_dev[0])
 585                 return SMC_CLC_DECL_NOSMCDDEV;
 586         else
 587                 ini->ism_chid[0] = smc_ism_get_chid(ini->ism_dev[0]);
 588         return 0;
 589 }
 590
 591 /* is chid unique for the ism devices that are already determined? */
 592 static bool smc_find_ism_v2_is_unique_chid(u16 chid, struct smc_init_info *ini,
 593                                            int cnt)
 594 {
 595         int i = (!ini->ism_dev[0]) ? 1 : 0;
 596
 597         for (; i < cnt; i++)
 598                 if (ini->ism_chid[i] == chid)
 599                         return false;
 600         return true;
 601 }
 602
 603 /* determine possible V2 ISM devices (either without PNETID or with PNETID plus
 604  * PNETID matching net_device)
 605  */
 606 static int smc_find_ism_v2_device_clnt(struct smc_sock *smc,
 607                                        struct smc_init_info *ini)
 608 {
 609         int rc = SMC_CLC_DECL_NOSMCDDEV;
 610         struct smcd_dev *smcd;
 611         int i = 1;
 612         u16 chid;
 613
 614         if (smcd_indicated(ini->smc_type_v1))
 615                 rc = 0;         /* already initialized for V1 */
 616         mutex_lock(&smcd_dev_list.mutex);
 617         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
 618                 if (smcd->going_away || smcd == ini->ism_dev[0])
 619                         continue;
 620                 chid = smc_ism_get_chid(smcd);
 621                 if (!smc_find_ism_v2_is_unique_chid(chid, ini, i))
 622                         continue;
 623                 if (!smc_pnet_is_pnetid_set(smcd->pnetid) ||
 624                     smc_pnet_is_ndev_pnetid(sock_net(&smc->sk), smcd->pnetid)) {
 625                         ini->ism_dev[i] = smcd;
 626                         ini->ism_chid[i] = chid;
 627                         ini->is_smcd = true;
 628                         rc = 0;
 629                         i++;
 630                         if (i > SMC_MAX_ISM_DEVS)
 631                                 break;
 632                 }
 633         }
 634         mutex_unlock(&smcd_dev_list.mutex);
 635         ini->ism_offered_cnt = i - 1;
 636         if (!ini->ism_dev[0] && !ini->ism_dev[1])
 637                 ini->smcd_version = 0;
 638
 639         return rc;
 640 }
 641
 642 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
 643 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
 644                                       struct smc_init_info *ini)
 645 {
 646         if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev[0], ini->vlan_id))
 647                 return SMC_CLC_DECL_ISMVLANERR;
 648         return 0;
 649 }
 650
 651 static int smc_find_proposal_devices(struct smc_sock *smc,
 652                                      struct smc_init_info *ini)
 653 {
 654         int rc = 0;
 655
 656         /* check if there is an ism device available */
 657         if (ini->smcd_version & SMC_V1) {
 658                 if (smc_find_ism_device(smc, ini) ||
 659                     smc_connect_ism_vlan_setup(smc, ini)) {
 660                         if (ini->smc_type_v1 == SMC_TYPE_B)
 661                                 ini->smc_type_v1 = SMC_TYPE_R;
 662                         else
 663                                 ini->smc_type_v1 = SMC_TYPE_N;
 664                 } /* else ISM V1 is supported for this connection */
 665                 if (smc_find_rdma_device(smc, ini)) {
 666                         if (ini->smc_type_v1 == SMC_TYPE_B)
 667                                 ini->smc_type_v1 = SMC_TYPE_D;
 668                         else
 669                                 ini->smc_type_v1 = SMC_TYPE_N;
 670                 } /* else RDMA is supported for this connection */
 671         }
 672         if (smc_ism_is_v2_capable() && smc_find_ism_v2_device_clnt(smc, ini))
 673                 ini->smc_type_v2 = SMC_TYPE_N;
 674
 675         /* if neither ISM nor RDMA are supported, fallback */
 676         if (!smcr_indicated(ini->smc_type_v1) &&
 677             ini->smc_type_v1 == SMC_TYPE_N && ini->smc_type_v2 == SMC_TYPE_N)
 678                 rc = SMC_CLC_DECL_NOSMCDEV;
 679
 680         return rc;
 681 }
 682
 683 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
 684  * used, the VLAN ID will be registered again during the connection setup.
 685  */
 686 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc,
 687                                         struct smc_init_info *ini)
 688 {
 689         if (!smcd_indicated(ini->smc_type_v1))
 690                 return 0;
 691         if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev[0], ini->vlan_id))
 692                 return SMC_CLC_DECL_CNFERR;
 693         return 0;
 694 }
 695
 696 #define SMC_CLC_MAX_ACCEPT_LEN \
 697         (sizeof(struct smc_clc_msg_accept_confirm_v2) + \
 698          sizeof(struct smc_clc_first_contact_ext) + \
 699          sizeof(struct smc_clc_msg_trail))
 700
 701 /* CLC handshake during connect */
 702 static int smc_connect_clc(struct smc_sock *smc,
 703                            struct smc_clc_msg_accept_confirm_v2 *aclc2,
 704                            struct smc_init_info *ini)
 705 {
 706         int rc = 0;
 707
 708         /* do inband token exchange */
 709         rc = smc_clc_send_proposal(smc, ini);
 710         if (rc)
 711                 return rc;
 712         /* receive SMC Accept CLC message */
 713         return smc_clc_wait_msg(smc, aclc2, SMC_CLC_MAX_ACCEPT_LEN,
 714                                 SMC_CLC_ACCEPT, CLC_WAIT_TIME);
 715 }
 716
 717 /* setup for RDMA connection of client */
 718 static int smc_connect_rdma(struct smc_sock *smc,
 719                             struct smc_clc_msg_accept_confirm *aclc,
 720                             struct smc_init_info *ini)
 721 {
 722         int i, reason_code = 0;
 723         struct smc_link *link;
 724
 725         ini->is_smcd = false;
 726         ini->ib_lcl = &aclc->r0.lcl;
 727         ini->ib_clcqpn = ntoh24(aclc->r0.qpn);
 728         ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
 729
 730         mutex_lock(&smc_client_lgr_pending);
 731         reason_code = smc_conn_create(smc, ini);
 732         if (reason_code) {
 733                 mutex_unlock(&smc_client_lgr_pending);
 734                 return reason_code;
 735         }
 736
 737         smc_conn_save_peer_info(smc, aclc);
 738
 739         if (ini->first_contact_local) {
 740                 link = smc->conn.lnk;
 741         } else {
 742                 /* set link that was assigned by server */
 743                 link = NULL;
 744                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 745                         struct smc_link *l = &smc->conn.lgr->lnk[i];
 746
 747                         if (l->peer_qpn == ntoh24(aclc->r0.qpn) &&
 748                             !memcmp(l->peer_gid, &aclc->r0.lcl.gid,
 749                                     SMC_GID_SIZE) &&
 750                             !memcmp(l->peer_mac, &aclc->r0.lcl.mac,
 751                                     sizeof(l->peer_mac))) {
 752                                 link = l;
 753                                 break;
 754                         }
 755                 }
 756                 if (!link) {
 757                         reason_code = SMC_CLC_DECL_NOSRVLINK;
 758                         goto connect_abort;
 759                 }
 760                 smc->conn.lnk = link;
 761         }
 762
 763         /* create send buffer and rmb */
 764         if (smc_buf_create(smc, false)) {
 765                 reason_code = SMC_CLC_DECL_MEM;
 766                 goto connect_abort;
 767         }
 768
 769         if (ini->first_contact_local)
 770                 smc_link_save_peer_info(link, aclc);
 771
 772         if (smc_rmb_rtoken_handling(&smc->conn, link, aclc)) {
 773                 reason_code = SMC_CLC_DECL_ERR_RTOK;
 774                 goto connect_abort;
 775         }
 776
 777         smc_close_init(smc);
 778         smc_rx_init(smc);
 779
 780         if (ini->first_contact_local) {
 781                 if (smc_ib_ready_link(link)) {
 782                         reason_code = SMC_CLC_DECL_ERR_RDYLNK;
 783                         goto connect_abort;
 784                 }
 785         } else {
 786                 if (smcr_lgr_reg_rmbs(link, smc->conn.rmb_desc)) {
 787                         reason_code = SMC_CLC_DECL_ERR_REGRMB;
 788                         goto connect_abort;
 789                 }
 790         }
 791         smc_rmb_sync_sg_for_device(&smc->conn);
 792
 793         reason_code = smc_clc_send_confirm(smc, ini->first_contact_local,
 794                                            SMC_V1);
 795         if (reason_code)
 796                 goto connect_abort;
 797
 798         smc_tx_init(smc);
 799
 800         if (ini->first_contact_local) {
 801                 /* QP confirmation over RoCE fabric */
 802                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
 803                 reason_code = smcr_clnt_conf_first_link(smc);
 804                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
 805                 if (reason_code)
 806                         goto connect_abort;
 807         }
 808         mutex_unlock(&smc_client_lgr_pending);
 809
 810         smc_copy_sock_settings_to_clc(smc);
 811         smc->connect_nonblock = 0;
 812         if (smc->sk.sk_state == SMC_INIT)
 813                 smc->sk.sk_state = SMC_ACTIVE;
 814
 815         return 0;
 816 connect_abort:
 817         smc_conn_abort(smc, ini->first_contact_local);
 818         mutex_unlock(&smc_client_lgr_pending);
 819         smc->connect_nonblock = 0;
 820
 821         return reason_code;
 822 }
 823
 824 /* The server has chosen one of the proposed ISM devices for the communication.
 825  * Determine from the CHID of the received CLC ACCEPT the ISM device chosen.
 826  */
 827 static int
 828 smc_v2_determine_accepted_chid(struct smc_clc_msg_accept_confirm_v2 *aclc,
 829                                struct smc_init_info *ini)
 830 {
 831         int i;
 832
 833         for (i = 0; i < ini->ism_offered_cnt + 1; i++) {
 834                 if (ini->ism_chid[i] == ntohs(aclc->chid)) {
 835                         ini->ism_selected = i;
 836                         return 0;
 837                 }
 838         }
 839
 840         return -EPROTO;
 841 }
 842
 843 /* setup for ISM connection of client */
 844 static int smc_connect_ism(struct smc_sock *smc,
 845                            struct smc_clc_msg_accept_confirm *aclc,
 846                            struct smc_init_info *ini)
 847 {
 848         int rc = 0;
 849
 850         ini->is_smcd = true;
 851         ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
 852
 853         if (aclc->hdr.version == SMC_V2) {
 854                 struct smc_clc_msg_accept_confirm_v2 *aclc_v2 =
 855                         (struct smc_clc_msg_accept_confirm_v2 *)aclc;
 856
 857                 rc = smc_v2_determine_accepted_chid(aclc_v2, ini);
 858                 if (rc)
 859                         return rc;
 860         }
 861         ini->ism_peer_gid[ini->ism_selected] = aclc->d0.gid;
 862
 863         /* there is only one lgr role for SMC-D; use server lock */
 864         mutex_lock(&smc_server_lgr_pending);
 865         rc = smc_conn_create(smc, ini);
 866         if (rc) {
 867                 mutex_unlock(&smc_server_lgr_pending);
 868                 return rc;
 869         }
 870
 871         /* Create send and receive buffers */
 872         rc = smc_buf_create(smc, true);
 873         if (rc) {
 874                 rc = (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB : SMC_CLC_DECL_MEM;
 875                 goto connect_abort;
 876         }
 877
 878         smc_conn_save_peer_info(smc, aclc);
 879         smc_close_init(smc);
 880         smc_rx_init(smc);
 881         smc_tx_init(smc);
 882
 883         rc = smc_clc_send_confirm(smc, ini->first_contact_local,
 884                                   aclc->hdr.version);
 885         if (rc)
 886                 goto connect_abort;
 887         mutex_unlock(&smc_server_lgr_pending);
 888
 889         smc_copy_sock_settings_to_clc(smc);
 890         smc->connect_nonblock = 0;
 891         if (smc->sk.sk_state == SMC_INIT)
 892                 smc->sk.sk_state = SMC_ACTIVE;
 893
 894         return 0;
 895 connect_abort:
 896         smc_conn_abort(smc, ini->first_contact_local);
 897         mutex_unlock(&smc_server_lgr_pending);
 898         smc->connect_nonblock = 0;
 899
 900         return rc;
 901 }
 902
 903 /* check if received accept type and version matches a proposed one */
 904 static int smc_connect_check_aclc(struct smc_init_info *ini,
 905                                   struct smc_clc_msg_accept_confirm *aclc)
 906 {
 907         if ((aclc->hdr.typev1 == SMC_TYPE_R &&
 908              !smcr_indicated(ini->smc_type_v1)) ||
 909             (aclc->hdr.typev1 == SMC_TYPE_D &&
 910              ((!smcd_indicated(ini->smc_type_v1) &&
 911                !smcd_indicated(ini->smc_type_v2)) ||
 912               (aclc->hdr.version == SMC_V1 &&
 913                !smcd_indicated(ini->smc_type_v1)) ||
 914               (aclc->hdr.version == SMC_V2 &&
 915                !smcd_indicated(ini->smc_type_v2)))))
 916                 return SMC_CLC_DECL_MODEUNSUPP;
 917
 918         return 0;
 919 }
 920
 921 /* perform steps before actually connecting */
 922 static int __smc_connect(struct smc_sock *smc)
 923 {
 924         u8 version = smc_ism_is_v2_capable() ? SMC_V2 : SMC_V1;
 925         struct smc_clc_msg_accept_confirm_v2 *aclc2;
 926         struct smc_clc_msg_accept_confirm *aclc;
 927         struct smc_init_info *ini = NULL;
 928         u8 *buf = NULL;
 929         int rc = 0;
 930
 931         if (smc->use_fallback)
 932                 return smc_connect_fallback(smc, smc->fallback_rsn);
 933
 934         /* if peer has not signalled SMC-capability, fall back */
 935         if (!tcp_sk(smc->clcsock->sk)->syn_smc)
 936                 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
 937
 938         /* IPSec connections opt out of SMC optimizations */
 939         if (using_ipsec(smc))
 940                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC,
 941                                                     version);
 942
 943         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
 944         if (!ini)
 945                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_MEM,
 946                                                     version);
 947
 948         ini->smcd_version = SMC_V1;
 949         ini->smcd_version |= smc_ism_is_v2_capable() ? SMC_V2 : 0;
 950         ini->smc_type_v1 = SMC_TYPE_B;
 951         ini->smc_type_v2 = smc_ism_is_v2_capable() ? SMC_TYPE_D : SMC_TYPE_N;
 952
 953         /* get vlan id from IP device */
 954         if (smc_vlan_by_tcpsk(smc->clcsock, ini)) {
 955                 ini->smcd_version &= ~SMC_V1;
 956                 ini->smc_type_v1 = SMC_TYPE_N;
 957                 if (!ini->smcd_version) {
 958                         rc = SMC_CLC_DECL_GETVLANERR;
 959                         goto fallback;
 960                 }
 961         }
 962
 963         rc = smc_find_proposal_devices(smc, ini);
 964         if (rc)
 965                 goto fallback;
 966
 967         buf = kzalloc(SMC_CLC_MAX_ACCEPT_LEN, GFP_KERNEL);
 968         if (!buf) {
 969                 rc = SMC_CLC_DECL_MEM;
 970                 goto fallback;
 971         }
 972         aclc2 = (struct smc_clc_msg_accept_confirm_v2 *)buf;
 973         aclc = (struct smc_clc_msg_accept_confirm *)aclc2;
 974
 975         /* perform CLC handshake */
 976         rc = smc_connect_clc(smc, aclc2, ini);
 977         if (rc)
 978                 goto vlan_cleanup;
 979
 980         /* check if smc modes and versions of CLC proposal and accept match */
 981         rc = smc_connect_check_aclc(ini, aclc);
 982         version = aclc->hdr.version == SMC_V1 ? SMC_V1 : SMC_V2;
 983         ini->smcd_version = version;
 984         if (rc)
 985                 goto vlan_cleanup;
 986
 987         /* depending on previous steps, connect using rdma or ism */
 988         if (aclc->hdr.typev1 == SMC_TYPE_R)
 989                 rc = smc_connect_rdma(smc, aclc, ini);
 990         else if (aclc->hdr.typev1 == SMC_TYPE_D)
 991                 rc = smc_connect_ism(smc, aclc, ini);
 992         if (rc)
 993                 goto vlan_cleanup;
 994
 995         smc_connect_ism_vlan_cleanup(smc, ini);
 996         kfree(buf);
 997         kfree(ini);
 998         return 0;
 999
1000 vlan_cleanup:
1001         smc_connect_ism_vlan_cleanup(smc, ini);
1002         kfree(buf);
1003 fallback:
1004         kfree(ini);
1005         return smc_connect_decline_fallback(smc, rc, version);
1006 }
1007
1008 static void smc_connect_work(struct work_struct *work)
1009 {
1010         struct smc_sock *smc = container_of(work, struct smc_sock,
1011                                             connect_work);
1012         long timeo = smc->sk.sk_sndtimeo;
1013         int rc = 0;
1014
1015         if (!timeo)
1016                 timeo = MAX_SCHEDULE_TIMEOUT;
1017         lock_sock(smc->clcsock->sk);
1018         if (smc->clcsock->sk->sk_err) {
1019                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
1020         } else if ((1 << smc->clcsock->sk->sk_state) &
1021                                         (TCPF_SYN_SENT | TCP_SYN_RECV)) {
1022                 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
1023                 if ((rc == -EPIPE) &&
1024                     ((1 << smc->clcsock->sk->sk_state) &
1025                                         (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
1026                         rc = 0;
1027         }
1028         release_sock(smc->clcsock->sk);
1029         lock_sock(&smc->sk);
1030         if (rc != 0 || smc->sk.sk_err) {
1031                 smc->sk.sk_state = SMC_CLOSED;
1032                 if (rc == -EPIPE || rc == -EAGAIN)
1033                         smc->sk.sk_err = EPIPE;
1034                 else if (signal_pending(current))
1035                         smc->sk.sk_err = -sock_intr_errno(timeo);
1036                 sock_put(&smc->sk); /* passive closing */
1037                 goto out;
1038         }
1039
1040         rc = __smc_connect(smc);
1041         if (rc < 0)
1042                 smc->sk.sk_err = -rc;
1043
1044 out:
1045         if (!sock_flag(&smc->sk, SOCK_DEAD)) {
1046                 if (smc->sk.sk_err) {
1047                         smc->sk.sk_state_change(&smc->sk);
1048                 } else { /* allow polling before and after fallback decision */
1049                         smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
1050                         smc->sk.sk_write_space(&smc->sk);
1051                 }
1052         }
1053         release_sock(&smc->sk);
1054 }
1055
1056 static int smc_connect(struct socket *sock, struct sockaddr *addr,
1057                        int alen, int flags)
1058 {
1059         struct sock *sk = sock->sk;
1060         struct smc_sock *smc;
1061         int rc = -EINVAL;
1062
1063         smc = smc_sk(sk);
1064
1065         /* separate smc parameter checking to be safe */
1066         if (alen < sizeof(addr->sa_family))
1067                 goto out_err;
1068         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
1069                 goto out_err;
1070
1071         lock_sock(sk);
1072         switch (sk->sk_state) {
1073         default:
1074                 goto out;
1075         case SMC_ACTIVE:
1076                 rc = -EISCONN;
1077                 goto out;
1078         case SMC_INIT:
1079                 break;
1080         }
1081
1082         smc_copy_sock_settings_to_clc(smc);
1083         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1084         if (smc->connect_nonblock) {
1085                 rc = -EALREADY;
1086                 goto out;
1087         }
1088         rc = kernel_connect(smc->clcsock, addr, alen, flags);
1089         if (rc && rc != -EINPROGRESS)
1090                 goto out;
1091
1092         sock_hold(&smc->sk); /* sock put in passive closing */
1093         if (smc->use_fallback)
1094                 goto out;
1095         if (flags & O_NONBLOCK) {
1096                 if (queue_work(smc_hs_wq, &smc->connect_work))
1097                         smc->connect_nonblock = 1;
1098                 rc = -EINPROGRESS;
1099         } else {
1100                 rc = __smc_connect(smc);
1101                 if (rc < 0)
1102                         goto out;
1103                 else
1104                         rc = 0; /* success cases including fallback */
1105         }
1106
1107 out:
1108         release_sock(sk);
1109 out_err:
1110         return rc;
1111 }
1112
1113 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
1114 {
1115         struct socket *new_clcsock = NULL;
1116         struct sock *lsk = &lsmc->sk;
1117         struct sock *new_sk;
1118         int rc = -EINVAL;
1119
1120         release_sock(lsk);
1121         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
1122         if (!new_sk) {
1123                 rc = -ENOMEM;
1124                 lsk->sk_err = ENOMEM;
1125                 *new_smc = NULL;
1126                 lock_sock(lsk);
1127                 goto out;
1128         }
1129         *new_smc = smc_sk(new_sk);
1130
1131         mutex_lock(&lsmc->clcsock_release_lock);
1132         if (lsmc->clcsock)
1133                 rc = kernel_accept(lsmc->clcsock, &new_clcsock, SOCK_NONBLOCK);
1134         mutex_unlock(&lsmc->clcsock_release_lock);
1135         lock_sock(lsk);
1136         if  (rc < 0 && rc != -EAGAIN)
1137                 lsk->sk_err = -rc;
1138         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
1139                 new_sk->sk_prot->unhash(new_sk);
1140                 if (new_clcsock)
1141                         sock_release(new_clcsock);
1142                 new_sk->sk_state = SMC_CLOSED;
1143                 sock_set_flag(new_sk, SOCK_DEAD);
1144                 sock_put(new_sk); /* final */
1145                 *new_smc = NULL;
1146                 goto out;
1147         }
1148
1149         /* new clcsock has inherited the smc listen-specific sk_data_ready
1150          * function; switch it back to the original sk_data_ready function
1151          */
1152         new_clcsock->sk->sk_data_ready = lsmc->clcsk_data_ready;
1153         (*new_smc)->clcsock = new_clcsock;
1154 out:
1155         return rc;
1156 }
1157
1158 /* add a just created sock to the accept queue of the listen sock as
1159  * candidate for a following socket accept call from user space
1160  */
1161 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
1162 {
1163         struct smc_sock *par = smc_sk(parent);
1164
1165         sock_hold(sk); /* sock_put in smc_accept_unlink () */
1166         spin_lock(&par->accept_q_lock);
1167         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
1168         spin_unlock(&par->accept_q_lock);
1169         sk_acceptq_added(parent);
1170 }
1171
1172 /* remove a socket from the accept queue of its parental listening socket */
1173 static void smc_accept_unlink(struct sock *sk)
1174 {
1175         struct smc_sock *par = smc_sk(sk)->listen_smc;
1176
1177         spin_lock(&par->accept_q_lock);
1178         list_del_init(&smc_sk(sk)->accept_q);
1179         spin_unlock(&par->accept_q_lock);
1180         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
1181         sock_put(sk); /* sock_hold in smc_accept_enqueue */
1182 }
1183
1184 /* remove a sock from the accept queue to bind it to a new socket created
1185  * for a socket accept call from user space
1186  */
1187 struct sock *smc_accept_dequeue(struct sock *parent,
1188                                 struct socket *new_sock)
1189 {
1190         struct smc_sock *isk, *n;
1191         struct sock *new_sk;
1192
1193         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1194                 new_sk = (struct sock *)isk;
1195
1196                 smc_accept_unlink(new_sk);
1197                 if (new_sk->sk_state == SMC_CLOSED) {
1198                         new_sk->sk_prot->unhash(new_sk);
1199                         if (isk->clcsock) {
1200                                 sock_release(isk->clcsock);
1201                                 isk->clcsock = NULL;
1202                         }
1203                         sock_put(new_sk); /* final */
1204                         continue;
1205                 }
1206                 if (new_sock) {
1207                         sock_graft(new_sk, new_sock);
1208                         if (isk->use_fallback) {
1209                                 smc_sk(new_sk)->clcsock->file = new_sock->file;
1210                                 isk->clcsock->file->private_data = isk->clcsock;
1211                         }
1212                 }
1213                 return new_sk;
1214         }
1215         return NULL;
1216 }
1217
1218 /* clean up for a created but never accepted sock */
1219 void smc_close_non_accepted(struct sock *sk)
1220 {
1221         struct smc_sock *smc = smc_sk(sk);
1222
1223         sock_hold(sk); /* sock_put below */
1224         lock_sock(sk);
1225         if (!sk->sk_lingertime)
1226                 /* wait for peer closing */
1227                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
1228         __smc_release(smc);
1229         release_sock(sk);
1230         sock_put(sk); /* sock_hold above */
1231         sock_put(sk); /* final sock_put */
1232 }
1233
1234 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1235 {
1236         struct smc_link *link = smc->conn.lnk;
1237         struct smc_llc_qentry *qentry;
1238         int rc;
1239
1240         if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
1241                 return SMC_CLC_DECL_ERR_REGRMB;
1242
1243         /* send CONFIRM LINK request to client over the RoCE fabric */
1244         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1245         if (rc < 0)
1246                 return SMC_CLC_DECL_TIMEOUT_CL;
1247
1248         /* receive CONFIRM LINK response from client over the RoCE fabric */
1249         qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1250                               SMC_LLC_CONFIRM_LINK);
1251         if (!qentry) {
1252                 struct smc_clc_msg_decline dclc;
1253
1254                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1255                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1256                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1257         }
1258         smc_llc_save_peer_uid(qentry);
1259         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1260         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1261         if (rc)
1262                 return SMC_CLC_DECL_RMBE_EC;
1263
1264         /* confirm_rkey is implicit on 1st contact */
1265         smc->conn.rmb_desc->is_conf_rkey = true;
1266
1267         smc_llc_link_active(link);
1268         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1269
1270         /* initial contact - try to establish second link */
1271         smc_llc_srv_add_link(link);
1272         return 0;
1273 }
1274
1275 /* listen worker: finish */
1276 static void smc_listen_out(struct smc_sock *new_smc)
1277 {
1278         struct smc_sock *lsmc = new_smc->listen_smc;
1279         struct sock *newsmcsk = &new_smc->sk;
1280
1281         if (lsmc->sk.sk_state == SMC_LISTEN) {
1282                 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1283                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1284                 release_sock(&lsmc->sk);
1285         } else { /* no longer listening */
1286                 smc_close_non_accepted(newsmcsk);
1287         }
1288
1289         /* Wake up accept */
1290         lsmc->sk.sk_data_ready(&lsmc->sk);
1291         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1292 }
1293
1294 /* listen worker: finish in state connected */
1295 static void smc_listen_out_connected(struct smc_sock *new_smc)
1296 {
1297         struct sock *newsmcsk = &new_smc->sk;
1298
1299         sk_refcnt_debug_inc(newsmcsk);
1300         if (newsmcsk->sk_state == SMC_INIT)
1301                 newsmcsk->sk_state = SMC_ACTIVE;
1302
1303         smc_listen_out(new_smc);
1304 }
1305
1306 /* listen worker: finish in error state */
1307 static void smc_listen_out_err(struct smc_sock *new_smc)
1308 {
1309         struct sock *newsmcsk = &new_smc->sk;
1310
1311         if (newsmcsk->sk_state == SMC_INIT)
1312                 sock_put(&new_smc->sk); /* passive closing */
1313         newsmcsk->sk_state = SMC_CLOSED;
1314
1315         smc_listen_out(new_smc);
1316 }
1317
1318 /* listen worker: decline and fall back if possible */
1319 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1320                                int local_first, u8 version)
1321 {
1322         /* RDMA setup failed, switch back to TCP */
1323         smc_conn_abort(new_smc, local_first);
1324         if (reason_code < 0) { /* error, no fallback possible */
1325                 smc_listen_out_err(new_smc);
1326                 return;
1327         }
1328         smc_switch_to_fallback(new_smc);
1329         new_smc->fallback_rsn = reason_code;
1330         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1331                 if (smc_clc_send_decline(new_smc, reason_code, version) < 0) {
1332                         smc_listen_out_err(new_smc);
1333                         return;
1334                 }
1335         }
1336         smc_listen_out_connected(new_smc);
1337 }
1338
1339 /* listen worker: version checking */
1340 static int smc_listen_v2_check(struct smc_sock *new_smc,
1341                                struct smc_clc_msg_proposal *pclc,
1342                                struct smc_init_info *ini)
1343 {
1344         struct smc_clc_smcd_v2_extension *pclc_smcd_v2_ext;
1345         struct smc_clc_v2_extension *pclc_v2_ext;
1346         int rc = SMC_CLC_DECL_PEERNOSMC;
1347
1348         ini->smc_type_v1 = pclc->hdr.typev1;
1349         ini->smc_type_v2 = pclc->hdr.typev2;
1350         ini->smcd_version = ini->smc_type_v1 != SMC_TYPE_N ? SMC_V1 : 0;
1351         if (pclc->hdr.version > SMC_V1)
1352                 ini->smcd_version |=
1353                                 ini->smc_type_v2 != SMC_TYPE_N ? SMC_V2 : 0;
1354         if (!(ini->smcd_version & SMC_V2)) {
1355                 rc = SMC_CLC_DECL_PEERNOSMC;
1356                 goto out;
1357         }
1358         if (!smc_ism_is_v2_capable()) {
1359                 ini->smcd_version &= ~SMC_V2;
1360                 rc = SMC_CLC_DECL_NOISM2SUPP;
1361                 goto out;
1362         }
1363         pclc_v2_ext = smc_get_clc_v2_ext(pclc);
1364         if (!pclc_v2_ext) {
1365                 ini->smcd_version &= ~SMC_V2;
1366                 rc = SMC_CLC_DECL_NOV2EXT;
1367                 goto out;
1368         }
1369         pclc_smcd_v2_ext = smc_get_clc_smcd_v2_ext(pclc_v2_ext);
1370         if (!pclc_smcd_v2_ext) {
1371                 ini->smcd_version &= ~SMC_V2;
1372                 rc = SMC_CLC_DECL_NOV2DEXT;
1373         }
1374
1375 out:
1376         if (!ini->smcd_version)
1377                 return rc;
1378
1379         return 0;
1380 }
1381
1382 /* listen worker: check prefixes */
1383 static int smc_listen_prfx_check(struct smc_sock *new_smc,
1384                                  struct smc_clc_msg_proposal *pclc)
1385 {
1386         struct smc_clc_msg_proposal_prefix *pclc_prfx;
1387         struct socket *newclcsock = new_smc->clcsock;
1388
1389         if (pclc->hdr.typev1 == SMC_TYPE_N)
1390                 return 0;
1391         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1392         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1393                 return SMC_CLC_DECL_DIFFPREFIX;
1394
1395         return 0;
1396 }
1397
1398 /* listen worker: initialize connection and buffers */
1399 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1400                                 struct smc_init_info *ini)
1401 {
1402         int rc;
1403
1404         /* allocate connection / link group */
1405         rc = smc_conn_create(new_smc, ini);
1406         if (rc)
1407                 return rc;
1408
1409         /* create send buffer and rmb */
1410         if (smc_buf_create(new_smc, false))
1411                 return SMC_CLC_DECL_MEM;
1412
1413         return 0;
1414 }
1415
1416 /* listen worker: initialize connection and buffers for SMC-D */
1417 static int smc_listen_ism_init(struct smc_sock *new_smc,
1418                                struct smc_init_info *ini)
1419 {
1420         int rc;
1421
1422         rc = smc_conn_create(new_smc, ini);
1423         if (rc)
1424                 return rc;
1425
1426         /* Create send and receive buffers */
1427         rc = smc_buf_create(new_smc, true);
1428         if (rc) {
1429                 smc_conn_abort(new_smc, ini->first_contact_local);
1430                 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
1431                                          SMC_CLC_DECL_MEM;
1432         }
1433
1434         return 0;
1435 }
1436
1437 static bool smc_is_already_selected(struct smcd_dev *smcd,
1438                                     struct smc_init_info *ini,
1439                                     int matches)
1440 {
1441         int i;
1442
1443         for (i = 0; i < matches; i++)
1444                 if (smcd == ini->ism_dev[i])
1445                         return true;
1446
1447         return false;
1448 }
1449
1450 /* check for ISM devices matching proposed ISM devices */
1451 static void smc_check_ism_v2_match(struct smc_init_info *ini,
1452                                    u16 proposed_chid, u64 proposed_gid,
1453                                    unsigned int *matches)
1454 {
1455         struct smcd_dev *smcd;
1456
1457         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1458                 if (smcd->going_away)
1459                         continue;
1460                 if (smc_is_already_selected(smcd, ini, *matches))
1461                         continue;
1462                 if (smc_ism_get_chid(smcd) == proposed_chid &&
1463                     !smc_ism_cantalk(proposed_gid, ISM_RESERVED_VLANID, smcd)) {
1464                         ini->ism_peer_gid[*matches] = proposed_gid;
1465                         ini->ism_dev[*matches] = smcd;
1466                         (*matches)++;
1467                         break;
1468                 }
1469         }
1470 }
1471
1472 static void smc_find_ism_store_rc(u32 rc, struct smc_init_info *ini)
1473 {
1474         if (!ini->rc)
1475                 ini->rc = rc;
1476 }
1477
1478 static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
1479                                         struct smc_clc_msg_proposal *pclc,
1480                                         struct smc_init_info *ini)
1481 {
1482         struct smc_clc_smcd_v2_extension *smcd_v2_ext;
1483         struct smc_clc_v2_extension *smc_v2_ext;
1484         struct smc_clc_msg_smcd *pclc_smcd;
1485         unsigned int matches = 0;
1486         u8 smcd_version;
1487         u8 *eid = NULL;
1488         int i, rc;
1489
1490         if (!(ini->smcd_version & SMC_V2) || !smcd_indicated(ini->smc_type_v2))
1491                 goto not_found;
1492
1493         pclc_smcd = smc_get_clc_msg_smcd(pclc);
1494         smc_v2_ext = smc_get_clc_v2_ext(pclc);
1495         smcd_v2_ext = smc_get_clc_smcd_v2_ext(smc_v2_ext);
1496         if (!smcd_v2_ext ||
1497             !smc_v2_ext->hdr.flag.seid) { /* no system EID support for SMCD */
1498                 smc_find_ism_store_rc(SMC_CLC_DECL_NOSEID, ini);
1499                 goto not_found;
1500         }
1501
1502         mutex_lock(&smcd_dev_list.mutex);
1503         if (pclc_smcd->ism.chid)
1504                 /* check for ISM device matching proposed native ISM device */
1505                 smc_check_ism_v2_match(ini, ntohs(pclc_smcd->ism.chid),
1506                                        ntohll(pclc_smcd->ism.gid), &matches);
1507         for (i = 1; i <= smc_v2_ext->hdr.ism_gid_cnt; i++) {
1508                 /* check for ISM devices matching proposed non-native ISM
1509                  * devices
1510                  */
1511                 smc_check_ism_v2_match(ini,
1512                                        ntohs(smcd_v2_ext->gidchid[i - 1].chid),
1513                                        ntohll(smcd_v2_ext->gidchid[i - 1].gid),
1514                                        &matches);
1515         }
1516         mutex_unlock(&smcd_dev_list.mutex);
1517
1518         if (ini->ism_dev[0]) {
1519                 smc_ism_get_system_eid(ini->ism_dev[0], &eid);
1520                 if (memcmp(eid, smcd_v2_ext->system_eid, SMC_MAX_EID_LEN))
1521                         goto not_found;
1522         } else {
1523                 goto not_found;
1524         }
1525
1526         /* separate - outside the smcd_dev_list.lock */
1527         smcd_version = ini->smcd_version;
1528         for (i = 0; i < matches; i++) {
1529                 ini->smcd_version = SMC_V2;
1530                 ini->is_smcd = true;
1531                 ini->ism_selected = i;
1532                 rc = smc_listen_ism_init(new_smc, ini);
1533                 if (rc) {
1534                         smc_find_ism_store_rc(rc, ini);
1535                         /* try next active ISM device */
1536                         continue;
1537                 }
1538                 return; /* matching and usable V2 ISM device found */
1539         }
1540         /* no V2 ISM device could be initialized */
1541         ini->smcd_version = smcd_version;       /* restore original value */
1542
1543 not_found:
1544         ini->smcd_version &= ~SMC_V2;
1545         ini->ism_dev[0] = NULL;
1546         ini->is_smcd = false;
1547 }
1548
1549 static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
1550                                         struct smc_clc_msg_proposal *pclc,
1551                                         struct smc_init_info *ini)
1552 {
1553         struct smc_clc_msg_smcd *pclc_smcd = smc_get_clc_msg_smcd(pclc);
1554         int rc = 0;
1555
1556         /* check if ISM V1 is available */
1557         if (!(ini->smcd_version & SMC_V1) || !smcd_indicated(ini->smc_type_v1))
1558                 goto not_found;
1559         ini->is_smcd = true; /* prepare ISM check */
1560         ini->ism_peer_gid[0] = ntohll(pclc_smcd->ism.gid);
1561         rc = smc_find_ism_device(new_smc, ini);
1562         if (rc)
1563                 goto not_found;
1564         ini->ism_selected = 0;
1565         rc = smc_listen_ism_init(new_smc, ini);
1566         if (!rc)
1567                 return;         /* V1 ISM device found */
1568
1569 not_found:
1570         smc_find_ism_store_rc(rc, ini);
1571         ini->ism_dev[0] = NULL;
1572         ini->is_smcd = false;
1573 }
1574
1575 /* listen worker: register buffers */
1576 static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
1577 {
1578         struct smc_connection *conn = &new_smc->conn;
1579
1580         if (!local_first) {
1581                 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
1582                         return SMC_CLC_DECL_ERR_REGRMB;
1583         }
1584         smc_rmb_sync_sg_for_device(&new_smc->conn);
1585
1586         return 0;
1587 }
1588
1589 static int smc_find_rdma_v1_device_serv(struct smc_sock *new_smc,
1590                                         struct smc_clc_msg_proposal *pclc,
1591                                         struct smc_init_info *ini)
1592 {
1593         int rc;
1594
1595         if (!smcr_indicated(ini->smc_type_v1))
1596                 return SMC_CLC_DECL_NOSMCDEV;
1597
1598         /* prepare RDMA check */
1599         ini->ib_lcl = &pclc->lcl;
1600         rc = smc_find_rdma_device(new_smc, ini);
1601         if (rc) {
1602                 /* no RDMA device found */
1603                 if (ini->smc_type_v1 == SMC_TYPE_B)
1604                         /* neither ISM nor RDMA device found */
1605                         rc = SMC_CLC_DECL_NOSMCDEV;
1606                 return rc;
1607         }
1608         rc = smc_listen_rdma_init(new_smc, ini);
1609         if (rc)
1610                 return rc;
1611         return smc_listen_rdma_reg(new_smc, ini->first_contact_local);
1612 }
1613
1614 /* determine the local device matching to proposal */
1615 static int smc_listen_find_device(struct smc_sock *new_smc,
1616                                   struct smc_clc_msg_proposal *pclc,
1617                                   struct smc_init_info *ini)
1618 {
1619         int rc;
1620
1621         /* check for ISM device matching V2 proposed device */
1622         smc_find_ism_v2_device_serv(new_smc, pclc, ini);
1623         if (ini->ism_dev[0])
1624                 return 0;
1625
1626         if (!(ini->smcd_version & SMC_V1))
1627                 return ini->rc ?: SMC_CLC_DECL_NOSMCD2DEV;
1628
1629         /* check for matching IP prefix and subnet length */
1630         rc = smc_listen_prfx_check(new_smc, pclc);
1631         if (rc)
1632                 return ini->rc ?: rc;
1633
1634         /* get vlan id from IP device */
1635         if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
1636                 return ini->rc ?: SMC_CLC_DECL_GETVLANERR;
1637
1638         /* check for ISM device matching V1 proposed device */
1639         smc_find_ism_v1_device_serv(new_smc, pclc, ini);
1640         if (ini->ism_dev[0])
1641                 return 0;
1642
1643         if (pclc->hdr.typev1 == SMC_TYPE_D)
1644                 /* skip RDMA and decline */
1645                 return ini->rc ?: SMC_CLC_DECL_NOSMCDDEV;
1646
1647         /* check if RDMA is available */
1648         rc = smc_find_rdma_v1_device_serv(new_smc, pclc, ini);
1649         smc_find_ism_store_rc(rc, ini);
1650
1651         return (!rc) ? 0 : ini->rc;
1652 }
1653
1654 /* listen worker: finish RDMA setup */
1655 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1656                                   struct smc_clc_msg_accept_confirm *cclc,
1657                                   bool local_first)
1658 {
1659         struct smc_link *link = new_smc->conn.lnk;
1660         int reason_code = 0;
1661
1662         if (local_first)
1663                 smc_link_save_peer_info(link, cclc);
1664
1665         if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc))
1666                 return SMC_CLC_DECL_ERR_RTOK;
1667
1668         if (local_first) {
1669                 if (smc_ib_ready_link(link))
1670                         return SMC_CLC_DECL_ERR_RDYLNK;
1671                 /* QP confirmation over RoCE fabric */
1672                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1673                 reason_code = smcr_serv_conf_first_link(new_smc);
1674                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1675         }
1676         return reason_code;
1677 }
1678
1679 /* setup for connection of server */
1680 static void smc_listen_work(struct work_struct *work)
1681 {
1682         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1683                                                 smc_listen_work);
1684         u8 version = smc_ism_is_v2_capable() ? SMC_V2 : SMC_V1;
1685         struct socket *newclcsock = new_smc->clcsock;
1686         struct smc_clc_msg_accept_confirm *cclc;
1687         struct smc_clc_msg_proposal_area *buf;
1688         struct smc_clc_msg_proposal *pclc;
1689         struct smc_init_info *ini = NULL;
1690         int rc = 0;
1691
1692         if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1693                 return smc_listen_out_err(new_smc);
1694
1695         if (new_smc->use_fallback) {
1696                 smc_listen_out_connected(new_smc);
1697                 return;
1698         }
1699
1700         /* check if peer is smc capable */
1701         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1702                 smc_switch_to_fallback(new_smc);
1703                 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1704                 smc_listen_out_connected(new_smc);
1705                 return;
1706         }
1707
1708         /* do inband token exchange -
1709          * wait for and receive SMC Proposal CLC message
1710          */
1711         buf = kzalloc(sizeof(*buf), GFP_KERNEL);
1712         if (!buf) {
1713                 rc = SMC_CLC_DECL_MEM;
1714                 goto out_decl;
1715         }
1716         pclc = (struct smc_clc_msg_proposal *)buf;
1717         rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
1718                               SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1719         if (rc)
1720                 goto out_decl;
1721         version = pclc->hdr.version == SMC_V1 ? SMC_V1 : version;
1722
1723         /* IPSec connections opt out of SMC optimizations */
1724         if (using_ipsec(new_smc)) {
1725                 rc = SMC_CLC_DECL_IPSEC;
1726                 goto out_decl;
1727         }
1728
1729         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1730         if (!ini) {
1731                 rc = SMC_CLC_DECL_MEM;
1732                 goto out_decl;
1733         }
1734
1735         /* initial version checking */
1736         rc = smc_listen_v2_check(new_smc, pclc, ini);
1737         if (rc)
1738                 goto out_decl;
1739
1740         mutex_lock(&smc_server_lgr_pending);
1741         smc_close_init(new_smc);
1742         smc_rx_init(new_smc);
1743         smc_tx_init(new_smc);
1744
1745         /* determine ISM or RoCE device used for connection */
1746         rc = smc_listen_find_device(new_smc, pclc, ini);
1747         if (rc)
1748                 goto out_unlock;
1749
1750         /* send SMC Accept CLC message */
1751         rc = smc_clc_send_accept(new_smc, ini->first_contact_local,
1752                                  ini->smcd_version == SMC_V2 ? SMC_V2 : SMC_V1);
1753         if (rc)
1754                 goto out_unlock;
1755
1756         /* SMC-D does not need this lock any more */
1757         if (ini->is_smcd)
1758                 mutex_unlock(&smc_server_lgr_pending);
1759
1760         /* receive SMC Confirm CLC message */
1761         memset(buf, 0, sizeof(*buf));
1762         cclc = (struct smc_clc_msg_accept_confirm *)buf;
1763         rc = smc_clc_wait_msg(new_smc, cclc, sizeof(*buf),
1764                               SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1765         if (rc) {
1766                 if (!ini->is_smcd)
1767                         goto out_unlock;
1768                 goto out_decl;
1769         }
1770
1771         /* finish worker */
1772         if (!ini->is_smcd) {
1773                 rc = smc_listen_rdma_finish(new_smc, cclc,
1774                                             ini->first_contact_local);
1775                 if (rc)
1776                         goto out_unlock;
1777                 mutex_unlock(&smc_server_lgr_pending);
1778         }
1779         smc_conn_save_peer_info(new_smc, cclc);
1780         smc_listen_out_connected(new_smc);
1781         goto out_free;
1782
1783 out_unlock:
1784         mutex_unlock(&smc_server_lgr_pending);
1785 out_decl:
1786         smc_listen_decline(new_smc, rc, ini ? ini->first_contact_local : 0,
1787                            version);
1788 out_free:
1789         kfree(ini);
1790         kfree(buf);
1791 }
1792
1793 static void smc_tcp_listen_work(struct work_struct *work)
1794 {
1795         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1796                                              tcp_listen_work);
1797         struct sock *lsk = &lsmc->sk;
1798         struct smc_sock *new_smc;
1799         int rc = 0;
1800
1801         lock_sock(lsk);
1802         while (lsk->sk_state == SMC_LISTEN) {
1803                 rc = smc_clcsock_accept(lsmc, &new_smc);
1804                 if (rc) /* clcsock accept queue empty or error */
1805                         goto out;
1806                 if (!new_smc)
1807                         continue;
1808
1809                 new_smc->listen_smc = lsmc;
1810                 new_smc->use_fallback = lsmc->use_fallback;
1811                 new_smc->fallback_rsn = lsmc->fallback_rsn;
1812                 sock_hold(lsk); /* sock_put in smc_listen_work */
1813                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1814                 smc_copy_sock_settings_to_smc(new_smc);
1815                 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1816                 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1817                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1818                 if (!queue_work(smc_hs_wq, &new_smc->smc_listen_work))
1819                         sock_put(&new_smc->sk);
1820         }
1821
1822 out:
1823         release_sock(lsk);
1824         sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
1825 }
1826
1827 static void smc_clcsock_data_ready(struct sock *listen_clcsock)
1828 {
1829         struct smc_sock *lsmc;
1830
1831         lsmc = (struct smc_sock *)
1832                ((uintptr_t)listen_clcsock->sk_user_data & ~SK_USER_DATA_NOCOPY);
1833         if (!lsmc)
1834                 return;
1835         lsmc->clcsk_data_ready(listen_clcsock);
1836         if (lsmc->sk.sk_state == SMC_LISTEN) {
1837                 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
1838                 if (!queue_work(smc_hs_wq, &lsmc->tcp_listen_work))
1839                         sock_put(&lsmc->sk);
1840         }
1841 }
1842
1843 static int smc_listen(struct socket *sock, int backlog)
1844 {
1845         struct sock *sk = sock->sk;
1846         struct smc_sock *smc;
1847         int rc;
1848
1849         smc = smc_sk(sk);
1850         lock_sock(sk);
1851
1852         rc = -EINVAL;
1853         if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
1854             smc->connect_nonblock)
1855                 goto out;
1856
1857         rc = 0;
1858         if (sk->sk_state == SMC_LISTEN) {
1859                 sk->sk_max_ack_backlog = backlog;
1860                 goto out;
1861         }
1862         /* some socket options are handled in core, so we could not apply
1863          * them to the clc socket -- copy smc socket options to clc socket
1864          */
1865         smc_copy_sock_settings_to_clc(smc);
1866         if (!smc->use_fallback)
1867                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1868
1869         /* save original sk_data_ready function and establish
1870          * smc-specific sk_data_ready function
1871          */
1872         smc->clcsk_data_ready = smc->clcsock->sk->sk_data_ready;
1873         smc->clcsock->sk->sk_data_ready = smc_clcsock_data_ready;
1874         smc->clcsock->sk->sk_user_data =
1875                 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
1876         rc = kernel_listen(smc->clcsock, backlog);
1877         if (rc)
1878                 goto out;
1879         sk->sk_max_ack_backlog = backlog;
1880         sk->sk_ack_backlog = 0;
1881         sk->sk_state = SMC_LISTEN;
1882
1883 out:
1884         release_sock(sk);
1885         return rc;
1886 }
1887
1888 static int smc_accept(struct socket *sock, struct socket *new_sock,
1889                       int flags, bool kern)
1890 {
1891         struct sock *sk = sock->sk, *nsk;
1892         DECLARE_WAITQUEUE(wait, current);
1893         struct smc_sock *lsmc;
1894         long timeo;
1895         int rc = 0;
1896
1897         lsmc = smc_sk(sk);
1898         sock_hold(sk); /* sock_put below */
1899         lock_sock(sk);
1900
1901         if (lsmc->sk.sk_state != SMC_LISTEN) {
1902                 rc = -EINVAL;
1903                 release_sock(sk);
1904                 goto out;
1905         }
1906
1907         /* Wait for an incoming connection */
1908         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1909         add_wait_queue_exclusive(sk_sleep(sk), &wait);
1910         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1911                 set_current_state(TASK_INTERRUPTIBLE);
1912                 if (!timeo) {
1913                         rc = -EAGAIN;
1914                         break;
1915                 }
1916                 release_sock(sk);
1917                 timeo = schedule_timeout(timeo);
1918                 /* wakeup by sk_data_ready in smc_listen_work() */
1919                 sched_annotate_sleep();
1920                 lock_sock(sk);
1921                 if (signal_pending(current)) {
1922                         rc = sock_intr_errno(timeo);
1923                         break;
1924                 }
1925         }
1926         set_current_state(TASK_RUNNING);
1927         remove_wait_queue(sk_sleep(sk), &wait);
1928
1929         if (!rc)
1930                 rc = sock_error(nsk);
1931         release_sock(sk);
1932         if (rc)
1933                 goto out;
1934
1935         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1936                 /* wait till data arrives on the socket */
1937                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1938                                                                 MSEC_PER_SEC);
1939                 if (smc_sk(nsk)->use_fallback) {
1940                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1941
1942                         lock_sock(clcsk);
1943                         if (skb_queue_empty(&clcsk->sk_receive_queue))
1944                                 sk_wait_data(clcsk, &timeo, NULL);
1945                         release_sock(clcsk);
1946                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1947                         lock_sock(nsk);
1948                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1949                         release_sock(nsk);
1950                 }
1951         }
1952
1953 out:
1954         sock_put(sk); /* sock_hold above */
1955         return rc;
1956 }
1957
1958 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1959                        int peer)
1960 {
1961         struct smc_sock *smc;
1962
1963         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1964             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1965                 return -ENOTCONN;
1966
1967         smc = smc_sk(sock->sk);
1968
1969         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1970 }
1971
1972 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1973 {
1974         struct sock *sk = sock->sk;
1975         struct smc_sock *smc;
1976         int rc = -EPIPE;
1977
1978         smc = smc_sk(sk);
1979         lock_sock(sk);
1980         if ((sk->sk_state != SMC_ACTIVE) &&
1981             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1982             (sk->sk_state != SMC_INIT))
1983                 goto out;
1984
1985         if (msg->msg_flags & MSG_FASTOPEN) {
1986                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1987                         smc_switch_to_fallback(smc);
1988                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1989                 } else {
1990                         rc = -EINVAL;
1991                         goto out;
1992                 }
1993         }
1994
1995         if (smc->use_fallback)
1996                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1997         else
1998                 rc = smc_tx_sendmsg(smc, msg, len);
1999 out:
2000         release_sock(sk);
2001         return rc;
2002 }
2003
2004 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
2005                        int flags)
2006 {
2007         struct sock *sk = sock->sk;
2008         struct smc_sock *smc;
2009         int rc = -ENOTCONN;
2010
2011         smc = smc_sk(sk);
2012         lock_sock(sk);
2013         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2014                 /* socket was connected before, no more data to read */
2015                 rc = 0;
2016                 goto out;
2017         }
2018         if ((sk->sk_state == SMC_INIT) ||
2019             (sk->sk_state == SMC_LISTEN) ||
2020             (sk->sk_state == SMC_CLOSED))
2021                 goto out;
2022
2023         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2024                 rc = 0;
2025                 goto out;
2026         }
2027
2028         if (smc->use_fallback) {
2029                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
2030         } else {
2031                 msg->msg_namelen = 0;
2032                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
2033         }
2034
2035 out:
2036         release_sock(sk);
2037         return rc;
2038 }
2039
2040 static __poll_t smc_accept_poll(struct sock *parent)
2041 {
2042         struct smc_sock *isk = smc_sk(parent);
2043         __poll_t mask = 0;
2044
2045         spin_lock(&isk->accept_q_lock);
2046         if (!list_empty(&isk->accept_q))
2047                 mask = EPOLLIN | EPOLLRDNORM;
2048         spin_unlock(&isk->accept_q_lock);
2049
2050         return mask;
2051 }
2052
2053 static __poll_t smc_poll(struct file *file, struct socket *sock,
2054                              poll_table *wait)
2055 {
2056         struct sock *sk = sock->sk;
2057         struct smc_sock *smc;
2058         __poll_t mask = 0;
2059
2060         if (!sk)
2061                 return EPOLLNVAL;
2062
2063         smc = smc_sk(sock->sk);
2064         if (smc->use_fallback) {
2065                 /* delegate to CLC child sock */
2066                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
2067                 sk->sk_err = smc->clcsock->sk->sk_err;
2068         } else {
2069                 if (sk->sk_state != SMC_CLOSED)
2070                         sock_poll_wait(file, sock, wait);
2071                 if (sk->sk_err)
2072                         mask |= EPOLLERR;
2073                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
2074                     (sk->sk_state == SMC_CLOSED))
2075                         mask |= EPOLLHUP;
2076                 if (sk->sk_state == SMC_LISTEN) {
2077                         /* woken up by sk_data_ready in smc_listen_work() */
2078                         mask |= smc_accept_poll(sk);
2079                 } else if (smc->use_fallback) { /* as result of connect_work()*/
2080                         mask |= smc->clcsock->ops->poll(file, smc->clcsock,
2081                                                            wait);
2082                         sk->sk_err = smc->clcsock->sk->sk_err;
2083                 } else {
2084                         if ((sk->sk_state != SMC_INIT &&
2085                              atomic_read(&smc->conn.sndbuf_space)) ||
2086                             sk->sk_shutdown & SEND_SHUTDOWN) {
2087                                 mask |= EPOLLOUT | EPOLLWRNORM;
2088                         } else {
2089                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
2090                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2091                         }
2092                         if (atomic_read(&smc->conn.bytes_to_rcv))
2093                                 mask |= EPOLLIN | EPOLLRDNORM;
2094                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2095                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2096                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
2097                                 mask |= EPOLLIN;
2098                         if (smc->conn.urg_state == SMC_URG_VALID)
2099                                 mask |= EPOLLPRI;
2100                 }
2101         }
2102
2103         return mask;
2104 }
2105
2106 static int smc_shutdown(struct socket *sock, int how)
2107 {
2108         struct sock *sk = sock->sk;
2109         struct smc_sock *smc;
2110         int rc = -EINVAL;
2111         int rc1 = 0;
2112
2113         smc = smc_sk(sk);
2114
2115         if ((how < SHUT_RD) || (how > SHUT_RDWR))
2116                 return rc;
2117
2118         lock_sock(sk);
2119
2120         rc = -ENOTCONN;
2121         if ((sk->sk_state != SMC_ACTIVE) &&
2122             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
2123             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
2124             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2125             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
2126             (sk->sk_state != SMC_APPFINCLOSEWAIT))
2127                 goto out;
2128         if (smc->use_fallback) {
2129                 rc = kernel_sock_shutdown(smc->clcsock, how);
2130                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
2131                 if (sk->sk_shutdown == SHUTDOWN_MASK)
2132                         sk->sk_state = SMC_CLOSED;
2133                 goto out;
2134         }
2135         switch (how) {
2136         case SHUT_RDWR:         /* shutdown in both directions */
2137                 rc = smc_close_active(smc);
2138                 break;
2139         case SHUT_WR:
2140                 rc = smc_close_shutdown_write(smc);
2141                 break;
2142         case SHUT_RD:
2143                 rc = 0;
2144                 /* nothing more to do because peer is not involved */
2145                 break;
2146         }
2147         if (smc->clcsock)
2148                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
2149         /* map sock_shutdown_cmd constants to sk_shutdown value range */
2150         sk->sk_shutdown |= how + 1;
2151
2152 out:
2153         release_sock(sk);
2154         return rc ? rc : rc1;
2155 }
2156
2157 static int smc_setsockopt(struct socket *sock, int level, int optname,
2158                           sockptr_t optval, unsigned int optlen)
2159 {
2160         struct sock *sk = sock->sk;
2161         struct smc_sock *smc;
2162         int val, rc;
2163
2164         if (level == SOL_TCP && optname == TCP_ULP)
2165                 return -EOPNOTSUPP;
2166
2167         smc = smc_sk(sk);
2168
2169         /* generic setsockopts reaching us here always apply to the
2170          * CLC socket
2171          */
2172         if (unlikely(!smc->clcsock->ops->setsockopt))
2173                 rc = -EOPNOTSUPP;
2174         else
2175                 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
2176                                                    optval, optlen);
2177         if (smc->clcsock->sk->sk_err) {
2178                 sk->sk_err = smc->clcsock->sk->sk_err;
2179                 sk->sk_error_report(sk);
2180         }
2181
2182         if (optlen < sizeof(int))
2183                 return -EINVAL;
2184         if (copy_from_sockptr(&val, optval, sizeof(int)))
2185                 return -EFAULT;
2186
2187         lock_sock(sk);
2188         if (rc || smc->use_fallback)
2189                 goto out;
2190         switch (optname) {
2191         case TCP_FASTOPEN:
2192         case TCP_FASTOPEN_CONNECT:
2193         case TCP_FASTOPEN_KEY:
2194         case TCP_FASTOPEN_NO_COOKIE:
2195                 /* option not supported by SMC */
2196                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2197                         smc_switch_to_fallback(smc);
2198                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
2199                 } else {
2200                         rc = -EINVAL;
2201                 }
2202                 break;
2203         case TCP_NODELAY:
2204                 if (sk->sk_state != SMC_INIT &&
2205                     sk->sk_state != SMC_LISTEN &&
2206                     sk->sk_state != SMC_CLOSED) {
2207                         if (val)
2208                                 mod_delayed_work(smc->conn.lgr->tx_wq,
2209                                                  &smc->conn.tx_work, 0);
2210                 }
2211                 break;
2212         case TCP_CORK:
2213                 if (sk->sk_state != SMC_INIT &&
2214                     sk->sk_state != SMC_LISTEN &&
2215                     sk->sk_state != SMC_CLOSED) {
2216                         if (!val)
2217                                 mod_delayed_work(smc->conn.lgr->tx_wq,
2218                                                  &smc->conn.tx_work, 0);
2219                 }
2220                 break;
2221         case TCP_DEFER_ACCEPT:
2222                 smc->sockopt_defer_accept = val;
2223                 break;
2224         default:
2225                 break;
2226         }
2227 out:
2228         release_sock(sk);
2229
2230         return rc;
2231 }
2232
2233 static int smc_getsockopt(struct socket *sock, int level, int optname,
2234                           char __user *optval, int __user *optlen)
2235 {
2236         struct smc_sock *smc;
2237
2238         smc = smc_sk(sock->sk);
2239         /* socket options apply to the CLC socket */
2240         if (unlikely(!smc->clcsock->ops->getsockopt))
2241                 return -EOPNOTSUPP;
2242         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
2243                                              optval, optlen);
2244 }
2245
2246 static int smc_ioctl(struct socket *sock, unsigned int cmd,
2247                      unsigned long arg)
2248 {
2249         union smc_host_cursor cons, urg;
2250         struct smc_connection *conn;
2251         struct smc_sock *smc;
2252         int answ;
2253
2254         smc = smc_sk(sock->sk);
2255         conn = &smc->conn;
2256         lock_sock(&smc->sk);
2257         if (smc->use_fallback) {
2258                 if (!smc->clcsock) {
2259                         release_sock(&smc->sk);
2260                         return -EBADF;
2261                 }
2262                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
2263                 release_sock(&smc->sk);
2264                 return answ;
2265         }
2266         switch (cmd) {
2267         case SIOCINQ: /* same as FIONREAD */
2268                 if (smc->sk.sk_state == SMC_LISTEN) {
2269                         release_sock(&smc->sk);
2270                         return -EINVAL;
2271                 }
2272                 if (smc->sk.sk_state == SMC_INIT ||
2273                     smc->sk.sk_state == SMC_CLOSED)
2274                         answ = 0;
2275                 else
2276                         answ = atomic_read(&smc->conn.bytes_to_rcv);
2277                 break;
2278         case SIOCOUTQ:
2279                 /* output queue size (not send + not acked) */
2280                 if (smc->sk.sk_state == SMC_LISTEN) {
2281                         release_sock(&smc->sk);
2282                         return -EINVAL;
2283                 }
2284                 if (smc->sk.sk_state == SMC_INIT ||
2285                     smc->sk.sk_state == SMC_CLOSED)
2286                         answ = 0;
2287                 else
2288                         answ = smc->conn.sndbuf_desc->len -
2289                                         atomic_read(&smc->conn.sndbuf_space);
2290                 break;
2291         case SIOCOUTQNSD:
2292                 /* output queue size (not send only) */
2293                 if (smc->sk.sk_state == SMC_LISTEN) {
2294                         release_sock(&smc->sk);
2295                         return -EINVAL;
2296                 }
2297                 if (smc->sk.sk_state == SMC_INIT ||
2298                     smc->sk.sk_state == SMC_CLOSED)
2299                         answ = 0;
2300                 else
2301                         answ = smc_tx_prepared_sends(&smc->conn);
2302                 break;
2303         case SIOCATMARK:
2304                 if (smc->sk.sk_state == SMC_LISTEN) {
2305                         release_sock(&smc->sk);
2306                         return -EINVAL;
2307                 }
2308                 if (smc->sk.sk_state == SMC_INIT ||
2309                     smc->sk.sk_state == SMC_CLOSED) {
2310                         answ = 0;
2311                 } else {
2312                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
2313                         smc_curs_copy(&urg, &conn->urg_curs, conn);
2314                         answ = smc_curs_diff(conn->rmb_desc->len,
2315                                              &cons, &urg) == 1;
2316                 }
2317                 break;
2318         default:
2319                 release_sock(&smc->sk);
2320                 return -ENOIOCTLCMD;
2321         }
2322         release_sock(&smc->sk);
2323
2324         return put_user(answ, (int __user *)arg);
2325 }
2326
2327 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
2328                             int offset, size_t size, int flags)
2329 {
2330         struct sock *sk = sock->sk;
2331         struct smc_sock *smc;
2332         int rc = -EPIPE;
2333
2334         smc = smc_sk(sk);
2335         lock_sock(sk);
2336         if (sk->sk_state != SMC_ACTIVE) {
2337                 release_sock(sk);
2338                 goto out;
2339         }
2340         release_sock(sk);
2341         if (smc->use_fallback)
2342                 rc = kernel_sendpage(smc->clcsock, page, offset,
2343                                      size, flags);
2344         else
2345                 rc = sock_no_sendpage(sock, page, offset, size, flags);
2346
2347 out:
2348         return rc;
2349 }
2350
2351 /* Map the affected portions of the rmbe into an spd, note the number of bytes
2352  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
2353  * updates till whenever a respective page has been fully processed.
2354  * Note that subsequent recv() calls have to wait till all splice() processing
2355  * completed.
2356  */
2357 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
2358                                struct pipe_inode_info *pipe, size_t len,
2359                                unsigned int flags)
2360 {
2361         struct sock *sk = sock->sk;
2362         struct smc_sock *smc;
2363         int rc = -ENOTCONN;
2364
2365         smc = smc_sk(sk);
2366         lock_sock(sk);
2367         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2368                 /* socket was connected before, no more data to read */
2369                 rc = 0;
2370                 goto out;
2371         }
2372         if (sk->sk_state == SMC_INIT ||
2373             sk->sk_state == SMC_LISTEN ||
2374             sk->sk_state == SMC_CLOSED)
2375                 goto out;
2376
2377         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2378                 rc = 0;
2379                 goto out;
2380         }
2381
2382         if (smc->use_fallback) {
2383                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
2384                                                     pipe, len, flags);
2385         } else {
2386                 if (*ppos) {
2387                         rc = -ESPIPE;
2388                         goto out;
2389                 }
2390                 if (flags & SPLICE_F_NONBLOCK)
2391                         flags = MSG_DONTWAIT;
2392                 else
2393                         flags = 0;
2394                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
2395         }
2396 out:
2397         release_sock(sk);
2398
2399         return rc;
2400 }
2401
2402 /* must look like tcp */
2403 static const struct proto_ops smc_sock_ops = {
2404         .family         = PF_SMC,
2405         .owner          = THIS_MODULE,
2406         .release        = smc_release,
2407         .bind           = smc_bind,
2408         .connect        = smc_connect,
2409         .socketpair     = sock_no_socketpair,
2410         .accept         = smc_accept,
2411         .getname        = smc_getname,
2412         .poll           = smc_poll,
2413         .ioctl          = smc_ioctl,
2414         .listen         = smc_listen,
2415         .shutdown       = smc_shutdown,
2416         .setsockopt     = smc_setsockopt,
2417         .getsockopt     = smc_getsockopt,
2418         .sendmsg        = smc_sendmsg,
2419         .recvmsg        = smc_recvmsg,
2420         .mmap           = sock_no_mmap,
2421         .sendpage       = smc_sendpage,
2422         .splice_read    = smc_splice_read,
2423 };
2424
2425 static int smc_create(struct net *net, struct socket *sock, int protocol,
2426                       int kern)
2427 {
2428         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
2429         struct smc_sock *smc;
2430         struct sock *sk;
2431         int rc;
2432
2433         rc = -ESOCKTNOSUPPORT;
2434         if (sock->type != SOCK_STREAM)
2435                 goto out;
2436
2437         rc = -EPROTONOSUPPORT;
2438         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
2439                 goto out;
2440
2441         rc = -ENOBUFS;
2442         sock->ops = &smc_sock_ops;
2443         sk = smc_sock_alloc(net, sock, protocol);
2444         if (!sk)
2445                 goto out;
2446
2447         /* create internal TCP socket for CLC handshake and fallback */
2448         smc = smc_sk(sk);
2449         smc->use_fallback = false; /* assume rdma capability first */
2450         smc->fallback_rsn = 0;
2451         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2452                               &smc->clcsock);
2453         if (rc) {
2454                 sk_common_release(sk);
2455                 goto out;
2456         }
2457         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2458         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
2459
2460 out:
2461         return rc;
2462 }
2463
2464 static const struct net_proto_family smc_sock_family_ops = {
2465         .family = PF_SMC,
2466         .owner  = THIS_MODULE,
2467         .create = smc_create,
2468 };
2469
2470 unsigned int smc_net_id;
2471
2472 static __net_init int smc_net_init(struct net *net)
2473 {
2474         return smc_pnet_net_init(net);
2475 }
2476
2477 static void __net_exit smc_net_exit(struct net *net)
2478 {
2479         smc_pnet_net_exit(net);
2480 }
2481
2482 static struct pernet_operations smc_net_ops = {
2483         .init = smc_net_init,
2484         .exit = smc_net_exit,
2485         .id   = &smc_net_id,
2486         .size = sizeof(struct smc_net),
2487 };
2488
2489 static int __init smc_init(void)
2490 {
2491         int rc;
2492
2493         rc = register_pernet_subsys(&smc_net_ops);
2494         if (rc)
2495                 return rc;
2496
2497         smc_ism_init();
2498         smc_clc_init();
2499
2500         rc = smc_nl_init();
2501         if (rc)
2502                 goto out_pernet_subsys;
2503
2504         rc = smc_pnet_init();
2505         if (rc)
2506                 goto out_nl;
2507
2508         rc = -ENOMEM;
2509         smc_hs_wq = alloc_workqueue("smc_hs_wq", 0, 0);
2510         if (!smc_hs_wq)
2511                 goto out_pnet;
2512
2513         smc_close_wq = alloc_workqueue("smc_close_wq", 0, 0);
2514         if (!smc_close_wq)
2515                 goto out_alloc_hs_wq;
2516
2517         rc = smc_core_init();
2518         if (rc) {
2519                 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
2520                 goto out_alloc_wqs;
2521         }
2522
2523         rc = smc_llc_init();
2524         if (rc) {
2525                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2526                 goto out_core;
2527         }
2528
2529         rc = smc_cdc_init();
2530         if (rc) {
2531                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2532                 goto out_core;
2533         }
2534
2535         rc = proto_register(&smc_proto, 1);
2536         if (rc) {
2537                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2538                 goto out_core;
2539         }
2540
2541         rc = proto_register(&smc_proto6, 1);
2542         if (rc) {
2543                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2544                 goto out_proto;
2545         }
2546
2547         rc = sock_register(&smc_sock_family_ops);
2548         if (rc) {
2549                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2550                 goto out_proto6;
2551         }
2552         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2553         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2554
2555         rc = smc_ib_register_client();
2556         if (rc) {
2557                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2558                 goto out_sock;
2559         }
2560
2561         static_branch_enable(&tcp_have_smc);
2562         return 0;
2563
2564 out_sock:
2565         sock_unregister(PF_SMC);
2566 out_proto6:
2567         proto_unregister(&smc_proto6);
2568 out_proto:
2569         proto_unregister(&smc_proto);
2570 out_core:
2571         smc_core_exit();
2572 out_alloc_wqs:
2573         destroy_workqueue(smc_close_wq);
2574 out_alloc_hs_wq:
2575         destroy_workqueue(smc_hs_wq);
2576 out_pnet:
2577         smc_pnet_exit();
2578 out_nl:
2579         smc_nl_exit();
2580 out_pernet_subsys:
2581         unregister_pernet_subsys(&smc_net_ops);
2582
2583         return rc;
2584 }
2585
2586 static void __exit smc_exit(void)
2587 {
2588         static_branch_disable(&tcp_have_smc);
2589         sock_unregister(PF_SMC);
2590         smc_core_exit();
2591         smc_ib_unregister_client();
2592         destroy_workqueue(smc_close_wq);
2593         destroy_workqueue(smc_hs_wq);
2594         proto_unregister(&smc_proto6);
2595         proto_unregister(&smc_proto);
2596         smc_pnet_exit();
2597         smc_nl_exit();
2598         unregister_pernet_subsys(&smc_net_ops);
2599         rcu_barrier();
2600 }
2601
2602 module_init(smc_init);
2603 module_exit(smc_exit);
2604
2605 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2606 MODULE_DESCRIPTION("smc socket address family");
2607 MODULE_LICENSE("GPL");
2608 MODULE_ALIAS_NETPROTO(PF_SMC);