net/smc/af_smc.c

   1 /*
   2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   3  *
   4  *  AF_SMC protocol family socket handler keeping the AF_INET sock address type
   5  *  applies to SOCK_STREAM sockets only
   6  *  offers an alternative communication option for TCP-protocol sockets
   7  *  applicable with RoCE-cards only
   8  *
   9  *  Initial restrictions:
  10  *    - non-blocking connect postponed
  11  *    - IPv6 support postponed
  12  *    - support for alternate links postponed
  13  *    - partial support for non-blocking sockets only
  14  *    - support for urgent data postponed
  15  *
  16  *  Copyright IBM Corp. 2016
  17  *
  18  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  19  *              based on prototype from Frank Blaschka
  20  */
  21
  22 #define KMSG_COMPONENT "smc"
  23 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  24
  25 #include <linux/module.h>
  26 #include <linux/socket.h>
  27 #include <linux/inetdevice.h>
  28 #include <linux/workqueue.h>
  29 #include <linux/in.h>
  30 #include <linux/sched/signal.h>
  31
  32 #include <net/sock.h>
  33 #include <net/tcp.h>
  34 #include <net/smc.h>
  35
  36 #include "smc.h"
  37 #include "smc_clc.h"
  38 #include "smc_llc.h"
  39 #include "smc_cdc.h"
  40 #include "smc_core.h"
  41 #include "smc_ib.h"
  42 #include "smc_pnet.h"
  43 #include "smc_tx.h"
  44 #include "smc_rx.h"
  45 #include "smc_close.h"
  46
  47 static DEFINE_MUTEX(smc_create_lgr_pending);    /* serialize link group
  48                                                  * creation
  49                                                  */
  50
  51 struct smc_lgr_list smc_lgr_list = {            /* established link groups */
  52         .lock = __SPIN_LOCK_UNLOCKED(smc_lgr_list.lock),
  53         .list = LIST_HEAD_INIT(smc_lgr_list.list),
  54 };
  55
  56 static void smc_tcp_listen_work(struct work_struct *);
  57
  58 static void smc_set_keepalive(struct sock *sk, int val)
  59 {
  60         struct smc_sock *smc = smc_sk(sk);
  61
  62         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
  63 }
  64
  65 static struct smc_hashinfo smc_v4_hashinfo = {
  66         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
  67 };
  68
  69 int smc_hash_sk(struct sock *sk)
  70 {
  71         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  72         struct hlist_head *head;
  73
  74         head = &h->ht;
  75
  76         write_lock_bh(&h->lock);
  77         sk_add_node(sk, head);
  78         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
  79         write_unlock_bh(&h->lock);
  80
  81         return 0;
  82 }
  83 EXPORT_SYMBOL_GPL(smc_hash_sk);
  84
  85 void smc_unhash_sk(struct sock *sk)
  86 {
  87         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  88
  89         write_lock_bh(&h->lock);
  90         if (sk_del_node_init(sk))
  91                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
  92         write_unlock_bh(&h->lock);
  93 }
  94 EXPORT_SYMBOL_GPL(smc_unhash_sk);
  95
  96 struct proto smc_proto = {
  97         .name           = "SMC",
  98         .owner          = THIS_MODULE,
  99         .keepalive      = smc_set_keepalive,
 100         .hash           = smc_hash_sk,
 101         .unhash         = smc_unhash_sk,
 102         .obj_size       = sizeof(struct smc_sock),
 103         .h.smc_hash     = &smc_v4_hashinfo,
 104         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 105 };
 106 EXPORT_SYMBOL_GPL(smc_proto);
 107
 108 static int smc_release(struct socket *sock)
 109 {
 110         struct sock *sk = sock->sk;
 111         struct smc_sock *smc;
 112         int rc = 0;
 113
 114         if (!sk)
 115                 goto out;
 116
 117         smc = smc_sk(sk);
 118         if (sk->sk_state == SMC_LISTEN)
 119                 /* smc_close_non_accepted() is called and acquires
 120                  * sock lock for child sockets again
 121                  */
 122                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 123         else
 124                 lock_sock(sk);
 125
 126         if (!smc->use_fallback) {
 127                 rc = smc_close_active(smc);
 128                 sock_set_flag(sk, SOCK_DEAD);
 129                 sk->sk_shutdown |= SHUTDOWN_MASK;
 130         }
 131         if (smc->clcsock) {
 132                 sock_release(smc->clcsock);
 133                 smc->clcsock = NULL;
 134         }
 135         if (smc->use_fallback) {
 136                 sock_put(sk); /* passive closing */
 137                 sk->sk_state = SMC_CLOSED;
 138                 sk->sk_state_change(sk);
 139         }
 140
 141         /* detach socket */
 142         sock_orphan(sk);
 143         sock->sk = NULL;
 144         if (!smc->use_fallback && sk->sk_state == SMC_CLOSED)
 145                 smc_conn_free(&smc->conn);
 146         release_sock(sk);
 147
 148         sk->sk_prot->unhash(sk);
 149         sock_put(sk); /* final sock_put */
 150 out:
 151         return rc;
 152 }
 153
 154 static void smc_destruct(struct sock *sk)
 155 {
 156         if (sk->sk_state != SMC_CLOSED)
 157                 return;
 158         if (!sock_flag(sk, SOCK_DEAD))
 159                 return;
 160
 161         sk_refcnt_debug_dec(sk);
 162 }
 163
 164 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock)
 165 {
 166         struct smc_sock *smc;
 167         struct sock *sk;
 168
 169         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, &smc_proto, 0);
 170         if (!sk)
 171                 return NULL;
 172
 173         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 174         sk->sk_state = SMC_INIT;
 175         sk->sk_destruct = smc_destruct;
 176         sk->sk_protocol = SMCPROTO_SMC;
 177         smc = smc_sk(sk);
 178         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 179         INIT_LIST_HEAD(&smc->accept_q);
 180         spin_lock_init(&smc->accept_q_lock);
 181         sk->sk_prot->hash(sk);
 182         sk_refcnt_debug_inc(sk);
 183
 184         return sk;
 185 }
 186
 187 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 188                     int addr_len)
 189 {
 190         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 191         struct sock *sk = sock->sk;
 192         struct smc_sock *smc;
 193         int rc;
 194
 195         smc = smc_sk(sk);
 196
 197         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 198         rc = -EINVAL;
 199         if (addr_len < sizeof(struct sockaddr_in))
 200                 goto out;
 201
 202         rc = -EAFNOSUPPORT;
 203         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 204         if ((addr->sin_family != AF_INET) &&
 205             ((addr->sin_family != AF_UNSPEC) ||
 206              (addr->sin_addr.s_addr != htonl(INADDR_ANY))))
 207                 goto out;
 208
 209         lock_sock(sk);
 210
 211         /* Check if socket is already active */
 212         rc = -EINVAL;
 213         if (sk->sk_state != SMC_INIT)
 214                 goto out_rel;
 215
 216         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 217         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 218
 219 out_rel:
 220         release_sock(sk);
 221 out:
 222         return rc;
 223 }
 224
 225 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 226                                    unsigned long mask)
 227 {
 228         /* options we don't get control via setsockopt for */
 229         nsk->sk_type = osk->sk_type;
 230         nsk->sk_sndbuf = osk->sk_sndbuf;
 231         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 232         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 233         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 234         nsk->sk_mark = osk->sk_mark;
 235         nsk->sk_priority = osk->sk_priority;
 236         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 237         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 238         nsk->sk_err = osk->sk_err;
 239
 240         nsk->sk_flags &= ~mask;
 241         nsk->sk_flags |= osk->sk_flags & mask;
 242 }
 243
 244 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 245                              (1UL << SOCK_KEEPOPEN) | \
 246                              (1UL << SOCK_LINGER) | \
 247                              (1UL << SOCK_BROADCAST) | \
 248                              (1UL << SOCK_TIMESTAMP) | \
 249                              (1UL << SOCK_DBG) | \
 250                              (1UL << SOCK_RCVTSTAMP) | \
 251                              (1UL << SOCK_RCVTSTAMPNS) | \
 252                              (1UL << SOCK_LOCALROUTE) | \
 253                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 254                              (1UL << SOCK_RXQ_OVFL) | \
 255                              (1UL << SOCK_WIFI_STATUS) | \
 256                              (1UL << SOCK_NOFCS) | \
 257                              (1UL << SOCK_FILTER_LOCKED))
 258 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 259  * clc socket (since smc is not called for these options from net/core)
 260  */
 261 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 262 {
 263         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 264 }
 265
 266 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 267                              (1UL << SOCK_KEEPOPEN) | \
 268                              (1UL << SOCK_LINGER) | \
 269                              (1UL << SOCK_DBG))
 270 /* copy only settings and flags relevant for smc from clc to smc socket */
 271 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 272 {
 273         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 274 }
 275
 276 /* determine subnet and mask of internal TCP socket */
 277 int smc_netinfo_by_tcpsk(struct socket *clcsock,
 278                          __be32 *subnet, u8 *prefix_len)
 279 {
 280         struct dst_entry *dst = sk_dst_get(clcsock->sk);
 281         struct in_device *in_dev;
 282         struct sockaddr_in addr;
 283         int rc = -ENOENT;
 284         int len;
 285
 286         if (!dst) {
 287                 rc = -ENOTCONN;
 288                 goto out;
 289         }
 290         if (!dst->dev) {
 291                 rc = -ENODEV;
 292                 goto out_rel;
 293         }
 294
 295         /* get address to which the internal TCP socket is bound */
 296         kernel_getsockname(clcsock, (struct sockaddr *)&addr, &len);
 297         /* analyze IPv4 specific data of net_device belonging to TCP socket */
 298         rcu_read_lock();
 299         in_dev = __in_dev_get_rcu(dst->dev);
 300         for_ifa(in_dev) {
 301                 if (!inet_ifa_match(addr.sin_addr.s_addr, ifa))
 302                         continue;
 303                 *prefix_len = inet_mask_len(ifa->ifa_mask);
 304                 *subnet = ifa->ifa_address & ifa->ifa_mask;
 305                 rc = 0;
 306                 break;
 307         } endfor_ifa(in_dev);
 308         rcu_read_unlock();
 309
 310 out_rel:
 311         dst_release(dst);
 312 out:
 313         return rc;
 314 }
 315
 316 static int smc_clnt_conf_first_link(struct smc_sock *smc, union ib_gid *gid)
 317 {
 318         struct smc_link_group *lgr = smc->conn.lgr;
 319         struct smc_link *link;
 320         int rest;
 321         int rc;
 322
 323         link = &lgr->lnk[SMC_SINGLE_LINK];
 324         /* receive CONFIRM LINK request from server over RoCE fabric */
 325         rest = wait_for_completion_interruptible_timeout(
 326                 &link->llc_confirm,
 327                 SMC_LLC_WAIT_FIRST_TIME);
 328         if (rest <= 0) {
 329                 struct smc_clc_msg_decline dclc;
 330
 331                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 332                                       SMC_CLC_DECLINE);
 333                 return rc;
 334         }
 335
 336         rc = smc_ib_modify_qp_rts(link);
 337         if (rc)
 338                 return SMC_CLC_DECL_INTERR;
 339
 340         smc_wr_remember_qp_attr(link);
 341
 342         rc = smc_wr_reg_send(link,
 343                              smc->conn.rmb_desc->mr_rx[SMC_SINGLE_LINK]);
 344         if (rc)
 345                 return SMC_CLC_DECL_INTERR;
 346
 347         /* send CONFIRM LINK response over RoCE fabric */
 348         rc = smc_llc_send_confirm_link(link,
 349                                        link->smcibdev->mac[link->ibport - 1],
 350                                        gid, SMC_LLC_RESP);
 351         if (rc < 0)
 352                 return SMC_CLC_DECL_TCL;
 353
 354         return rc;
 355 }
 356
 357 static void smc_conn_save_peer_info(struct smc_sock *smc,
 358                                     struct smc_clc_msg_accept_confirm *clc)
 359 {
 360         smc->conn.peer_conn_idx = clc->conn_idx;
 361         smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
 362         smc->conn.peer_rmbe_size = smc_uncompress_bufsize(clc->rmbe_size);
 363         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 364 }
 365
 366 static void smc_link_save_peer_info(struct smc_link *link,
 367                                     struct smc_clc_msg_accept_confirm *clc)
 368 {
 369         link->peer_qpn = ntoh24(clc->qpn);
 370         memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
 371         memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
 372         link->peer_psn = ntoh24(clc->psn);
 373         link->peer_mtu = clc->qp_mtu;
 374 }
 375
 376 static void smc_lgr_forget(struct smc_link_group *lgr)
 377 {
 378         spin_lock_bh(&smc_lgr_list.lock);
 379         /* do not use this link group for new connections */
 380         if (!list_empty(&lgr->list))
 381                 list_del_init(&lgr->list);
 382         spin_unlock_bh(&smc_lgr_list.lock);
 383 }
 384
 385 /* setup for RDMA connection of client */
 386 static int smc_connect_rdma(struct smc_sock *smc)
 387 {
 388         struct sockaddr_in *inaddr = (struct sockaddr_in *)smc->addr;
 389         struct smc_clc_msg_accept_confirm aclc;
 390         int local_contact = SMC_FIRST_CONTACT;
 391         struct smc_ib_device *smcibdev;
 392         struct smc_link *link;
 393         u8 srv_first_contact;
 394         int reason_code = 0;
 395         int rc = 0;
 396         u8 ibport;
 397
 398         sock_hold(&smc->sk); /* sock put in passive closing */
 399
 400         if (!tcp_sk(smc->clcsock->sk)->syn_smc) {
 401                 /* peer has not signalled SMC-capability */
 402                 smc->use_fallback = true;
 403                 goto out_connected;
 404         }
 405
 406         /* IPSec connections opt out of SMC-R optimizations */
 407         if (using_ipsec(smc)) {
 408                 reason_code = SMC_CLC_DECL_IPSEC;
 409                 goto decline_rdma;
 410         }
 411
 412         /* PNET table look up: search active ib_device and port
 413          * within same PNETID that also contains the ethernet device
 414          * used for the internal TCP socket
 415          */
 416         smc_pnet_find_roce_resource(smc->clcsock->sk, &smcibdev, &ibport);
 417         if (!smcibdev) {
 418                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 419                 goto decline_rdma;
 420         }
 421
 422         /* do inband token exchange */
 423         reason_code = smc_clc_send_proposal(smc, smcibdev, ibport);
 424         if (reason_code < 0) {
 425                 rc = reason_code;
 426                 goto out_err;
 427         }
 428         if (reason_code > 0) /* configuration error */
 429                 goto decline_rdma;
 430         /* receive SMC Accept CLC message */
 431         reason_code = smc_clc_wait_msg(smc, &aclc, sizeof(aclc),
 432                                        SMC_CLC_ACCEPT);
 433         if (reason_code < 0) {
 434                 rc = reason_code;
 435                 goto out_err;
 436         }
 437         if (reason_code > 0)
 438                 goto decline_rdma;
 439
 440         srv_first_contact = aclc.hdr.flag;
 441         mutex_lock(&smc_create_lgr_pending);
 442         local_contact = smc_conn_create(smc, inaddr->sin_addr.s_addr, smcibdev,
 443                                         ibport, &aclc.lcl, srv_first_contact);
 444         if (local_contact < 0) {
 445                 rc = local_contact;
 446                 if (rc == -ENOMEM)
 447                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
 448                 else if (rc == -ENOLINK)
 449                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
 450                 goto decline_rdma_unlock;
 451         }
 452         link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
 453
 454         smc_conn_save_peer_info(smc, &aclc);
 455
 456         /* create send buffer and rmb */
 457         rc = smc_buf_create(smc);
 458         if (rc) {
 459                 reason_code = SMC_CLC_DECL_MEM;
 460                 goto decline_rdma_unlock;
 461         }
 462
 463         if (local_contact == SMC_FIRST_CONTACT)
 464                 smc_link_save_peer_info(link, &aclc);
 465
 466         rc = smc_rmb_rtoken_handling(&smc->conn, &aclc);
 467         if (rc) {
 468                 reason_code = SMC_CLC_DECL_INTERR;
 469                 goto decline_rdma_unlock;
 470         }
 471
 472         smc_close_init(smc);
 473         smc_rx_init(smc);
 474
 475         if (local_contact == SMC_FIRST_CONTACT) {
 476                 rc = smc_ib_ready_link(link);
 477                 if (rc) {
 478                         reason_code = SMC_CLC_DECL_INTERR;
 479                         goto decline_rdma_unlock;
 480                 }
 481         } else {
 482                 struct smc_buf_desc *buf_desc = smc->conn.rmb_desc;
 483
 484                 if (!buf_desc->reused) {
 485                         /* register memory region for new rmb */
 486                         rc = smc_wr_reg_send(link,
 487                                              buf_desc->mr_rx[SMC_SINGLE_LINK]);
 488                         if (rc) {
 489                                 reason_code = SMC_CLC_DECL_INTERR;
 490                                 goto decline_rdma_unlock;
 491                         }
 492                 }
 493         }
 494         smc_rmb_sync_sg_for_device(&smc->conn);
 495
 496         rc = smc_clc_send_confirm(smc);
 497         if (rc)
 498                 goto out_err_unlock;
 499
 500         if (local_contact == SMC_FIRST_CONTACT) {
 501                 /* QP confirmation over RoCE fabric */
 502                 reason_code = smc_clnt_conf_first_link(
 503                         smc, &smcibdev->gid[ibport - 1]);
 504                 if (reason_code < 0) {
 505                         rc = reason_code;
 506                         goto out_err_unlock;
 507                 }
 508                 if (reason_code > 0)
 509                         goto decline_rdma_unlock;
 510         }
 511
 512         mutex_unlock(&smc_create_lgr_pending);
 513         smc_tx_init(smc);
 514
 515 out_connected:
 516         smc_copy_sock_settings_to_clc(smc);
 517         if (smc->sk.sk_state == SMC_INIT)
 518                 smc->sk.sk_state = SMC_ACTIVE;
 519
 520         return rc ? rc : local_contact;
 521
 522 decline_rdma_unlock:
 523         if (local_contact == SMC_FIRST_CONTACT)
 524                 smc_lgr_forget(smc->conn.lgr);
 525         mutex_unlock(&smc_create_lgr_pending);
 526         smc_conn_free(&smc->conn);
 527 decline_rdma:
 528         /* RDMA setup failed, switch back to TCP */
 529         smc->use_fallback = true;
 530         if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
 531                 rc = smc_clc_send_decline(smc, reason_code);
 532                 if (rc < 0)
 533                         goto out_err;
 534         }
 535         goto out_connected;
 536
 537 out_err_unlock:
 538         if (local_contact == SMC_FIRST_CONTACT)
 539                 smc_lgr_forget(smc->conn.lgr);
 540         mutex_unlock(&smc_create_lgr_pending);
 541         smc_conn_free(&smc->conn);
 542 out_err:
 543         if (smc->sk.sk_state == SMC_INIT)
 544                 sock_put(&smc->sk); /* passive closing */
 545         return rc;
 546 }
 547
 548 static int smc_connect(struct socket *sock, struct sockaddr *addr,
 549                        int alen, int flags)
 550 {
 551         struct sock *sk = sock->sk;
 552         struct smc_sock *smc;
 553         int rc = -EINVAL;
 554
 555         smc = smc_sk(sk);
 556
 557         /* separate smc parameter checking to be safe */
 558         if (alen < sizeof(addr->sa_family))
 559                 goto out_err;
 560         if (addr->sa_family != AF_INET)
 561                 goto out_err;
 562         smc->addr = addr;       /* needed for nonblocking connect */
 563
 564         lock_sock(sk);
 565         switch (sk->sk_state) {
 566         default:
 567                 goto out;
 568         case SMC_ACTIVE:
 569                 rc = -EISCONN;
 570                 goto out;
 571         case SMC_INIT:
 572                 rc = 0;
 573                 break;
 574         }
 575
 576         smc_copy_sock_settings_to_clc(smc);
 577         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
 578         rc = kernel_connect(smc->clcsock, addr, alen, flags);
 579         if (rc)
 580                 goto out;
 581
 582         /* setup RDMA connection */
 583         rc = smc_connect_rdma(smc);
 584         if (rc < 0)
 585                 goto out;
 586         else
 587                 rc = 0; /* success cases including fallback */
 588
 589 out:
 590         release_sock(sk);
 591 out_err:
 592         return rc;
 593 }
 594
 595 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 596 {
 597         struct socket *new_clcsock = NULL;
 598         struct sock *lsk = &lsmc->sk;
 599         struct sock *new_sk;
 600         int rc;
 601
 602         release_sock(lsk);
 603         new_sk = smc_sock_alloc(sock_net(lsk), NULL);
 604         if (!new_sk) {
 605                 rc = -ENOMEM;
 606                 lsk->sk_err = ENOMEM;
 607                 *new_smc = NULL;
 608                 lock_sock(lsk);
 609                 goto out;
 610         }
 611         *new_smc = smc_sk(new_sk);
 612
 613         rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
 614         lock_sock(lsk);
 615         if  (rc < 0)
 616                 lsk->sk_err = -rc;
 617         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
 618                 if (new_clcsock)
 619                         sock_release(new_clcsock);
 620                 new_sk->sk_state = SMC_CLOSED;
 621                 sock_set_flag(new_sk, SOCK_DEAD);
 622                 new_sk->sk_prot->unhash(new_sk);
 623                 sock_put(new_sk); /* final */
 624                 *new_smc = NULL;
 625                 goto out;
 626         }
 627
 628         (*new_smc)->clcsock = new_clcsock;
 629 out:
 630         return rc;
 631 }
 632
 633 /* add a just created sock to the accept queue of the listen sock as
 634  * candidate for a following socket accept call from user space
 635  */
 636 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
 637 {
 638         struct smc_sock *par = smc_sk(parent);
 639
 640         sock_hold(sk); /* sock_put in smc_accept_unlink () */
 641         spin_lock(&par->accept_q_lock);
 642         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
 643         spin_unlock(&par->accept_q_lock);
 644         sk_acceptq_added(parent);
 645 }
 646
 647 /* remove a socket from the accept queue of its parental listening socket */
 648 static void smc_accept_unlink(struct sock *sk)
 649 {
 650         struct smc_sock *par = smc_sk(sk)->listen_smc;
 651
 652         spin_lock(&par->accept_q_lock);
 653         list_del_init(&smc_sk(sk)->accept_q);
 654         spin_unlock(&par->accept_q_lock);
 655         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
 656         sock_put(sk); /* sock_hold in smc_accept_enqueue */
 657 }
 658
 659 /* remove a sock from the accept queue to bind it to a new socket created
 660  * for a socket accept call from user space
 661  */
 662 struct sock *smc_accept_dequeue(struct sock *parent,
 663                                 struct socket *new_sock)
 664 {
 665         struct smc_sock *isk, *n;
 666         struct sock *new_sk;
 667
 668         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
 669                 new_sk = (struct sock *)isk;
 670
 671                 smc_accept_unlink(new_sk);
 672                 if (new_sk->sk_state == SMC_CLOSED) {
 673                         if (isk->clcsock) {
 674                                 sock_release(isk->clcsock);
 675                                 isk->clcsock = NULL;
 676                         }
 677                         new_sk->sk_prot->unhash(new_sk);
 678                         sock_put(new_sk); /* final */
 679                         continue;
 680                 }
 681                 if (new_sock)
 682                         sock_graft(new_sk, new_sock);
 683                 return new_sk;
 684         }
 685         return NULL;
 686 }
 687
 688 /* clean up for a created but never accepted sock */
 689 void smc_close_non_accepted(struct sock *sk)
 690 {
 691         struct smc_sock *smc = smc_sk(sk);
 692
 693         lock_sock(sk);
 694         if (!sk->sk_lingertime)
 695                 /* wait for peer closing */
 696                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
 697         if (!smc->use_fallback) {
 698                 smc_close_active(smc);
 699                 sock_set_flag(sk, SOCK_DEAD);
 700                 sk->sk_shutdown |= SHUTDOWN_MASK;
 701         }
 702         if (smc->clcsock) {
 703                 struct socket *tcp;
 704
 705                 tcp = smc->clcsock;
 706                 smc->clcsock = NULL;
 707                 sock_release(tcp);
 708         }
 709         if (smc->use_fallback) {
 710                 sock_put(sk); /* passive closing */
 711                 sk->sk_state = SMC_CLOSED;
 712         } else {
 713                 if (sk->sk_state == SMC_CLOSED)
 714                         smc_conn_free(&smc->conn);
 715         }
 716         release_sock(sk);
 717         sk->sk_prot->unhash(sk);
 718         sock_put(sk); /* final sock_put */
 719 }
 720
 721 static int smc_serv_conf_first_link(struct smc_sock *smc)
 722 {
 723         struct smc_link_group *lgr = smc->conn.lgr;
 724         struct smc_link *link;
 725         int rest;
 726         int rc;
 727
 728         link = &lgr->lnk[SMC_SINGLE_LINK];
 729
 730         rc = smc_wr_reg_send(link,
 731                              smc->conn.rmb_desc->mr_rx[SMC_SINGLE_LINK]);
 732         if (rc)
 733                 return SMC_CLC_DECL_INTERR;
 734
 735         /* send CONFIRM LINK request to client over the RoCE fabric */
 736         rc = smc_llc_send_confirm_link(link,
 737                                        link->smcibdev->mac[link->ibport - 1],
 738                                        &link->smcibdev->gid[link->ibport - 1],
 739                                        SMC_LLC_REQ);
 740         if (rc < 0)
 741                 return SMC_CLC_DECL_TCL;
 742
 743         /* receive CONFIRM LINK response from client over the RoCE fabric */
 744         rest = wait_for_completion_interruptible_timeout(
 745                 &link->llc_confirm_resp,
 746                 SMC_LLC_WAIT_FIRST_TIME);
 747         if (rest <= 0) {
 748                 struct smc_clc_msg_decline dclc;
 749
 750                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 751                                       SMC_CLC_DECLINE);
 752         }
 753
 754         return rc;
 755 }
 756
 757 /* setup for RDMA connection of server */
 758 static void smc_listen_work(struct work_struct *work)
 759 {
 760         struct smc_sock *new_smc = container_of(work, struct smc_sock,
 761                                                 smc_listen_work);
 762         struct smc_clc_msg_proposal_prefix *pclc_prfx;
 763         struct socket *newclcsock = new_smc->clcsock;
 764         struct smc_sock *lsmc = new_smc->listen_smc;
 765         struct smc_clc_msg_accept_confirm cclc;
 766         int local_contact = SMC_REUSE_CONTACT;
 767         struct sock *newsmcsk = &new_smc->sk;
 768         struct smc_clc_msg_proposal *pclc;
 769         struct smc_ib_device *smcibdev;
 770         struct sockaddr_in peeraddr;
 771         u8 buf[SMC_CLC_MAX_LEN];
 772         struct smc_link *link;
 773         int reason_code = 0;
 774         int rc = 0, len;
 775         __be32 subnet;
 776         u8 prefix_len;
 777         u8 ibport;
 778
 779         /* check if peer is smc capable */
 780         if (!tcp_sk(newclcsock->sk)->syn_smc) {
 781                 new_smc->use_fallback = true;
 782                 goto out_connected;
 783         }
 784
 785         /* do inband token exchange -
 786          *wait for and receive SMC Proposal CLC message
 787          */
 788         reason_code = smc_clc_wait_msg(new_smc, &buf, sizeof(buf),
 789                                        SMC_CLC_PROPOSAL);
 790         if (reason_code < 0)
 791                 goto out_err;
 792         if (reason_code > 0)
 793                 goto decline_rdma;
 794
 795         /* IPSec connections opt out of SMC-R optimizations */
 796         if (using_ipsec(new_smc)) {
 797                 reason_code = SMC_CLC_DECL_IPSEC;
 798                 goto decline_rdma;
 799         }
 800
 801         /* PNET table look up: search active ib_device and port
 802          * within same PNETID that also contains the ethernet device
 803          * used for the internal TCP socket
 804          */
 805         smc_pnet_find_roce_resource(newclcsock->sk, &smcibdev, &ibport);
 806         if (!smcibdev) {
 807                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 808                 goto decline_rdma;
 809         }
 810
 811         /* determine subnet and mask from internal TCP socket */
 812         rc = smc_netinfo_by_tcpsk(newclcsock, &subnet, &prefix_len);
 813         if (rc) {
 814                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 815                 goto decline_rdma;
 816         }
 817
 818         pclc = (struct smc_clc_msg_proposal *)&buf;
 819         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
 820         if (pclc_prfx->outgoing_subnet != subnet ||
 821             pclc_prfx->prefix_len != prefix_len) {
 822                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 823                 goto decline_rdma;
 824         }
 825
 826         /* get address of the peer connected to the internal TCP socket */
 827         kernel_getpeername(newclcsock, (struct sockaddr *)&peeraddr, &len);
 828
 829         /* allocate connection / link group */
 830         mutex_lock(&smc_create_lgr_pending);
 831         local_contact = smc_conn_create(new_smc, peeraddr.sin_addr.s_addr,
 832                                         smcibdev, ibport, &pclc->lcl, 0);
 833         if (local_contact < 0) {
 834                 rc = local_contact;
 835                 if (rc == -ENOMEM)
 836                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
 837                 goto decline_rdma_unlock;
 838         }
 839         link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
 840
 841         /* create send buffer and rmb */
 842         rc = smc_buf_create(new_smc);
 843         if (rc) {
 844                 reason_code = SMC_CLC_DECL_MEM;
 845                 goto decline_rdma_unlock;
 846         }
 847
 848         smc_close_init(new_smc);
 849         smc_rx_init(new_smc);
 850
 851         if (local_contact != SMC_FIRST_CONTACT) {
 852                 struct smc_buf_desc *buf_desc = new_smc->conn.rmb_desc;
 853
 854                 if (!buf_desc->reused) {
 855                         /* register memory region for new rmb */
 856                         rc = smc_wr_reg_send(link,
 857                                              buf_desc->mr_rx[SMC_SINGLE_LINK]);
 858                         if (rc) {
 859                                 reason_code = SMC_CLC_DECL_INTERR;
 860                                 goto decline_rdma_unlock;
 861                         }
 862                 }
 863         }
 864         smc_rmb_sync_sg_for_device(&new_smc->conn);
 865
 866         rc = smc_clc_send_accept(new_smc, local_contact);
 867         if (rc)
 868                 goto out_err_unlock;
 869
 870         /* receive SMC Confirm CLC message */
 871         reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
 872                                        SMC_CLC_CONFIRM);
 873         if (reason_code < 0)
 874                 goto out_err_unlock;
 875         if (reason_code > 0)
 876                 goto decline_rdma_unlock;
 877         smc_conn_save_peer_info(new_smc, &cclc);
 878         if (local_contact == SMC_FIRST_CONTACT)
 879                 smc_link_save_peer_info(link, &cclc);
 880
 881         rc = smc_rmb_rtoken_handling(&new_smc->conn, &cclc);
 882         if (rc) {
 883                 reason_code = SMC_CLC_DECL_INTERR;
 884                 goto decline_rdma_unlock;
 885         }
 886
 887         if (local_contact == SMC_FIRST_CONTACT) {
 888                 rc = smc_ib_ready_link(link);
 889                 if (rc) {
 890                         reason_code = SMC_CLC_DECL_INTERR;
 891                         goto decline_rdma_unlock;
 892                 }
 893                 /* QP confirmation over RoCE fabric */
 894                 reason_code = smc_serv_conf_first_link(new_smc);
 895                 if (reason_code < 0)
 896                         /* peer is not aware of a problem */
 897                         goto out_err_unlock;
 898                 if (reason_code > 0)
 899                         goto decline_rdma_unlock;
 900         }
 901
 902         smc_tx_init(new_smc);
 903         mutex_unlock(&smc_create_lgr_pending);
 904
 905 out_connected:
 906         sk_refcnt_debug_inc(newsmcsk);
 907         if (newsmcsk->sk_state == SMC_INIT)
 908                 newsmcsk->sk_state = SMC_ACTIVE;
 909 enqueue:
 910         lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
 911         if (lsmc->sk.sk_state == SMC_LISTEN) {
 912                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
 913         } else { /* no longer listening */
 914                 smc_close_non_accepted(newsmcsk);
 915         }
 916         release_sock(&lsmc->sk);
 917
 918         /* Wake up accept */
 919         lsmc->sk.sk_data_ready(&lsmc->sk);
 920         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
 921         return;
 922
 923 decline_rdma_unlock:
 924         if (local_contact == SMC_FIRST_CONTACT)
 925                 smc_lgr_forget(new_smc->conn.lgr);
 926         mutex_unlock(&smc_create_lgr_pending);
 927 decline_rdma:
 928         /* RDMA setup failed, switch back to TCP */
 929         smc_conn_free(&new_smc->conn);
 930         new_smc->use_fallback = true;
 931         if (reason_code && (reason_code != SMC_CLC_DECL_REPLY)) {
 932                 if (smc_clc_send_decline(new_smc, reason_code) < 0)
 933                         goto out_err;
 934         }
 935         goto out_connected;
 936
 937 out_err_unlock:
 938         if (local_contact == SMC_FIRST_CONTACT)
 939                 smc_lgr_forget(new_smc->conn.lgr);
 940         mutex_unlock(&smc_create_lgr_pending);
 941 out_err:
 942         if (newsmcsk->sk_state == SMC_INIT)
 943                 sock_put(&new_smc->sk); /* passive closing */
 944         newsmcsk->sk_state = SMC_CLOSED;
 945         smc_conn_free(&new_smc->conn);
 946         goto enqueue; /* queue new sock with sk_err set */
 947 }
 948
 949 static void smc_tcp_listen_work(struct work_struct *work)
 950 {
 951         struct smc_sock *lsmc = container_of(work, struct smc_sock,
 952                                              tcp_listen_work);
 953         struct sock *lsk = &lsmc->sk;
 954         struct smc_sock *new_smc;
 955         int rc = 0;
 956
 957         lock_sock(lsk);
 958         while (lsk->sk_state == SMC_LISTEN) {
 959                 rc = smc_clcsock_accept(lsmc, &new_smc);
 960                 if (rc)
 961                         goto out;
 962                 if (!new_smc)
 963                         continue;
 964
 965                 new_smc->listen_smc = lsmc;
 966                 new_smc->use_fallback = false; /* assume rdma capability first*/
 967                 sock_hold(lsk); /* sock_put in smc_listen_work */
 968                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
 969                 smc_copy_sock_settings_to_smc(new_smc);
 970                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
 971                 if (!schedule_work(&new_smc->smc_listen_work))
 972                         sock_put(&new_smc->sk);
 973         }
 974
 975 out:
 976         if (lsmc->clcsock) {
 977                 sock_release(lsmc->clcsock);
 978                 lsmc->clcsock = NULL;
 979         }
 980         release_sock(lsk);
 981         /* no more listening, wake up smc_close_wait_listen_clcsock and
 982          * accept
 983          */
 984         lsk->sk_state_change(lsk);
 985         sock_put(&lsmc->sk); /* sock_hold in smc_listen */
 986 }
 987
 988 static int smc_listen(struct socket *sock, int backlog)
 989 {
 990         struct sock *sk = sock->sk;
 991         struct smc_sock *smc;
 992         int rc;
 993
 994         smc = smc_sk(sk);
 995         lock_sock(sk);
 996
 997         rc = -EINVAL;
 998         if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
 999                 goto out;
1000
1001         rc = 0;
1002         if (sk->sk_state == SMC_LISTEN) {
1003                 sk->sk_max_ack_backlog = backlog;
1004                 goto out;
1005         }
1006         /* some socket options are handled in core, so we could not apply
1007          * them to the clc socket -- copy smc socket options to clc socket
1008          */
1009         smc_copy_sock_settings_to_clc(smc);
1010         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1011
1012         rc = kernel_listen(smc->clcsock, backlog);
1013         if (rc)
1014                 goto out;
1015         sk->sk_max_ack_backlog = backlog;
1016         sk->sk_ack_backlog = 0;
1017         sk->sk_state = SMC_LISTEN;
1018         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
1019         sock_hold(sk); /* sock_hold in tcp_listen_worker */
1020         if (!schedule_work(&smc->tcp_listen_work))
1021                 sock_put(sk);
1022
1023 out:
1024         release_sock(sk);
1025         return rc;
1026 }
1027
1028 static int smc_accept(struct socket *sock, struct socket *new_sock,
1029                       int flags, bool kern)
1030 {
1031         struct sock *sk = sock->sk, *nsk;
1032         DECLARE_WAITQUEUE(wait, current);
1033         struct smc_sock *lsmc;
1034         long timeo;
1035         int rc = 0;
1036
1037         lsmc = smc_sk(sk);
1038         sock_hold(sk); /* sock_put below */
1039         lock_sock(sk);
1040
1041         if (lsmc->sk.sk_state != SMC_LISTEN) {
1042                 rc = -EINVAL;
1043                 goto out;
1044         }
1045
1046         /* Wait for an incoming connection */
1047         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1048         add_wait_queue_exclusive(sk_sleep(sk), &wait);
1049         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1050                 set_current_state(TASK_INTERRUPTIBLE);
1051                 if (!timeo) {
1052                         rc = -EAGAIN;
1053                         break;
1054                 }
1055                 release_sock(sk);
1056                 timeo = schedule_timeout(timeo);
1057                 /* wakeup by sk_data_ready in smc_listen_work() */
1058                 sched_annotate_sleep();
1059                 lock_sock(sk);
1060                 if (signal_pending(current)) {
1061                         rc = sock_intr_errno(timeo);
1062                         break;
1063                 }
1064         }
1065         set_current_state(TASK_RUNNING);
1066         remove_wait_queue(sk_sleep(sk), &wait);
1067
1068         if (!rc)
1069                 rc = sock_error(nsk);
1070
1071 out:
1072         release_sock(sk);
1073         sock_put(sk); /* sock_hold above */
1074         return rc;
1075 }
1076
1077 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1078                        int *len, int peer)
1079 {
1080         struct smc_sock *smc;
1081
1082         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1083             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1084                 return -ENOTCONN;
1085
1086         smc = smc_sk(sock->sk);
1087
1088         return smc->clcsock->ops->getname(smc->clcsock, addr, len, peer);
1089 }
1090
1091 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1092 {
1093         struct sock *sk = sock->sk;
1094         struct smc_sock *smc;
1095         int rc = -EPIPE;
1096
1097         smc = smc_sk(sk);
1098         lock_sock(sk);
1099         if ((sk->sk_state != SMC_ACTIVE) &&
1100             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1101             (sk->sk_state != SMC_INIT))
1102                 goto out;
1103         if (smc->use_fallback)
1104                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1105         else
1106                 rc = smc_tx_sendmsg(smc, msg, len);
1107 out:
1108         release_sock(sk);
1109         return rc;
1110 }
1111
1112 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1113                        int flags)
1114 {
1115         struct sock *sk = sock->sk;
1116         struct smc_sock *smc;
1117         int rc = -ENOTCONN;
1118
1119         smc = smc_sk(sk);
1120         lock_sock(sk);
1121         if ((sk->sk_state == SMC_INIT) ||
1122             (sk->sk_state == SMC_LISTEN) ||
1123             (sk->sk_state == SMC_CLOSED))
1124                 goto out;
1125
1126         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1127                 rc = 0;
1128                 goto out;
1129         }
1130
1131         if (smc->use_fallback)
1132                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1133         else
1134                 rc = smc_rx_recvmsg(smc, msg, len, flags);
1135
1136 out:
1137         release_sock(sk);
1138         return rc;
1139 }
1140
1141 static __poll_t smc_accept_poll(struct sock *parent)
1142 {
1143         struct smc_sock *isk = smc_sk(parent);
1144         __poll_t mask = 0;
1145
1146         spin_lock(&isk->accept_q_lock);
1147         if (!list_empty(&isk->accept_q))
1148                 mask = EPOLLIN | EPOLLRDNORM;
1149         spin_unlock(&isk->accept_q_lock);
1150
1151         return mask;
1152 }
1153
1154 static __poll_t smc_poll(struct file *file, struct socket *sock,
1155                              poll_table *wait)
1156 {
1157         struct sock *sk = sock->sk;
1158         __poll_t mask = 0;
1159         struct smc_sock *smc;
1160         int rc;
1161
1162         if (!sk)
1163                 return EPOLLNVAL;
1164
1165         smc = smc_sk(sock->sk);
1166         sock_hold(sk);
1167         lock_sock(sk);
1168         if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
1169                 /* delegate to CLC child sock */
1170                 release_sock(sk);
1171                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1172                 /* if non-blocking connect finished ... */
1173                 lock_sock(sk);
1174                 if ((sk->sk_state == SMC_INIT) && (mask & EPOLLOUT)) {
1175                         sk->sk_err = smc->clcsock->sk->sk_err;
1176                         if (sk->sk_err) {
1177                                 mask |= EPOLLERR;
1178                         } else {
1179                                 rc = smc_connect_rdma(smc);
1180                                 if (rc < 0)
1181                                         mask |= EPOLLERR;
1182                                 /* success cases including fallback */
1183                                 mask |= EPOLLOUT | EPOLLWRNORM;
1184                         }
1185                 }
1186         } else {
1187                 if (sk->sk_state != SMC_CLOSED) {
1188                         release_sock(sk);
1189                         sock_poll_wait(file, sk_sleep(sk), wait);
1190                         lock_sock(sk);
1191                 }
1192                 if (sk->sk_err)
1193                         mask |= EPOLLERR;
1194                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1195                     (sk->sk_state == SMC_CLOSED))
1196                         mask |= EPOLLHUP;
1197                 if (sk->sk_state == SMC_LISTEN) {
1198                         /* woken up by sk_data_ready in smc_listen_work() */
1199                         mask = smc_accept_poll(sk);
1200                 } else {
1201                         if (atomic_read(&smc->conn.sndbuf_space) ||
1202                             sk->sk_shutdown & SEND_SHUTDOWN) {
1203                                 mask |= EPOLLOUT | EPOLLWRNORM;
1204                         } else {
1205                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1206                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1207                         }
1208                         if (atomic_read(&smc->conn.bytes_to_rcv))
1209                                 mask |= EPOLLIN | EPOLLRDNORM;
1210                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1211                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1212                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
1213                                 mask |= EPOLLIN;
1214                 }
1215
1216         }
1217         release_sock(sk);
1218         sock_put(sk);
1219
1220         return mask;
1221 }
1222
1223 static int smc_shutdown(struct socket *sock, int how)
1224 {
1225         struct sock *sk = sock->sk;
1226         struct smc_sock *smc;
1227         int rc = -EINVAL;
1228         int rc1 = 0;
1229
1230         smc = smc_sk(sk);
1231
1232         if ((how < SHUT_RD) || (how > SHUT_RDWR))
1233                 return rc;
1234
1235         lock_sock(sk);
1236
1237         rc = -ENOTCONN;
1238         if ((sk->sk_state != SMC_LISTEN) &&
1239             (sk->sk_state != SMC_ACTIVE) &&
1240             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1241             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1242             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1243             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1244             (sk->sk_state != SMC_APPFINCLOSEWAIT))
1245                 goto out;
1246         if (smc->use_fallback) {
1247                 rc = kernel_sock_shutdown(smc->clcsock, how);
1248                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1249                 if (sk->sk_shutdown == SHUTDOWN_MASK)
1250                         sk->sk_state = SMC_CLOSED;
1251                 goto out;
1252         }
1253         switch (how) {
1254         case SHUT_RDWR:         /* shutdown in both directions */
1255                 rc = smc_close_active(smc);
1256                 break;
1257         case SHUT_WR:
1258                 rc = smc_close_shutdown_write(smc);
1259                 break;
1260         case SHUT_RD:
1261                 if (sk->sk_state == SMC_LISTEN)
1262                         rc = smc_close_active(smc);
1263                 else
1264                         rc = 0;
1265                         /* nothing more to do because peer is not involved */
1266                 break;
1267         }
1268         rc1 = kernel_sock_shutdown(smc->clcsock, how);
1269         /* map sock_shutdown_cmd constants to sk_shutdown value range */
1270         sk->sk_shutdown |= how + 1;
1271
1272 out:
1273         release_sock(sk);
1274         return rc ? rc : rc1;
1275 }
1276
1277 static int smc_setsockopt(struct socket *sock, int level, int optname,
1278                           char __user *optval, unsigned int optlen)
1279 {
1280         struct sock *sk = sock->sk;
1281         struct smc_sock *smc;
1282
1283         smc = smc_sk(sk);
1284
1285         /* generic setsockopts reaching us here always apply to the
1286          * CLC socket
1287          */
1288         return smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1289                                              optval, optlen);
1290 }
1291
1292 static int smc_getsockopt(struct socket *sock, int level, int optname,
1293                           char __user *optval, int __user *optlen)
1294 {
1295         struct smc_sock *smc;
1296
1297         smc = smc_sk(sock->sk);
1298         /* socket options apply to the CLC socket */
1299         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1300                                              optval, optlen);
1301 }
1302
1303 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1304                      unsigned long arg)
1305 {
1306         struct smc_sock *smc;
1307
1308         smc = smc_sk(sock->sk);
1309         if (smc->use_fallback)
1310                 return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1311         else
1312                 return sock_no_ioctl(sock, cmd, arg);
1313 }
1314
1315 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1316                             int offset, size_t size, int flags)
1317 {
1318         struct sock *sk = sock->sk;
1319         struct smc_sock *smc;
1320         int rc = -EPIPE;
1321
1322         smc = smc_sk(sk);
1323         lock_sock(sk);
1324         if (sk->sk_state != SMC_ACTIVE)
1325                 goto out;
1326         if (smc->use_fallback)
1327                 rc = kernel_sendpage(smc->clcsock, page, offset,
1328                                      size, flags);
1329         else
1330                 rc = sock_no_sendpage(sock, page, offset, size, flags);
1331
1332 out:
1333         release_sock(sk);
1334         return rc;
1335 }
1336
1337 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1338                                struct pipe_inode_info *pipe, size_t len,
1339                                     unsigned int flags)
1340 {
1341         struct sock *sk = sock->sk;
1342         struct smc_sock *smc;
1343         int rc = -ENOTCONN;
1344
1345         smc = smc_sk(sk);
1346         lock_sock(sk);
1347         if ((sk->sk_state != SMC_ACTIVE) && (sk->sk_state != SMC_CLOSED))
1348                 goto out;
1349         if (smc->use_fallback) {
1350                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1351                                                     pipe, len, flags);
1352         } else {
1353                 rc = -EOPNOTSUPP;
1354         }
1355 out:
1356         release_sock(sk);
1357         return rc;
1358 }
1359
1360 /* must look like tcp */
1361 static const struct proto_ops smc_sock_ops = {
1362         .family         = PF_SMC,
1363         .owner          = THIS_MODULE,
1364         .release        = smc_release,
1365         .bind           = smc_bind,
1366         .connect        = smc_connect,
1367         .socketpair     = sock_no_socketpair,
1368         .accept         = smc_accept,
1369         .getname        = smc_getname,
1370         .poll           = smc_poll,
1371         .ioctl          = smc_ioctl,
1372         .listen         = smc_listen,
1373         .shutdown       = smc_shutdown,
1374         .setsockopt     = smc_setsockopt,
1375         .getsockopt     = smc_getsockopt,
1376         .sendmsg        = smc_sendmsg,
1377         .recvmsg        = smc_recvmsg,
1378         .mmap           = sock_no_mmap,
1379         .sendpage       = smc_sendpage,
1380         .splice_read    = smc_splice_read,
1381 };
1382
1383 static int smc_create(struct net *net, struct socket *sock, int protocol,
1384                       int kern)
1385 {
1386         struct smc_sock *smc;
1387         struct sock *sk;
1388         int rc;
1389
1390         rc = -ESOCKTNOSUPPORT;
1391         if (sock->type != SOCK_STREAM)
1392                 goto out;
1393
1394         rc = -EPROTONOSUPPORT;
1395         if ((protocol != IPPROTO_IP) && (protocol != IPPROTO_TCP))
1396                 goto out;
1397
1398         rc = -ENOBUFS;
1399         sock->ops = &smc_sock_ops;
1400         sk = smc_sock_alloc(net, sock);
1401         if (!sk)
1402                 goto out;
1403
1404         /* create internal TCP socket for CLC handshake and fallback */
1405         smc = smc_sk(sk);
1406         smc->use_fallback = false; /* assume rdma capability first */
1407         rc = sock_create_kern(net, PF_INET, SOCK_STREAM,
1408                               IPPROTO_TCP, &smc->clcsock);
1409         if (rc) {
1410                 sk_common_release(sk);
1411                 goto out;
1412         }
1413         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1414         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1415
1416 out:
1417         return rc;
1418 }
1419
1420 static const struct net_proto_family smc_sock_family_ops = {
1421         .family = PF_SMC,
1422         .owner  = THIS_MODULE,
1423         .create = smc_create,
1424 };
1425
1426 static int __init smc_init(void)
1427 {
1428         int rc;
1429
1430         rc = smc_pnet_init();
1431         if (rc)
1432                 return rc;
1433
1434         rc = smc_llc_init();
1435         if (rc) {
1436                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
1437                 goto out_pnet;
1438         }
1439
1440         rc = smc_cdc_init();
1441         if (rc) {
1442                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
1443                 goto out_pnet;
1444         }
1445
1446         rc = proto_register(&smc_proto, 1);
1447         if (rc) {
1448                 pr_err("%s: proto_register fails with %d\n", __func__, rc);
1449                 goto out_pnet;
1450         }
1451
1452         rc = sock_register(&smc_sock_family_ops);
1453         if (rc) {
1454                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1455                 goto out_proto;
1456         }
1457         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
1458
1459         rc = smc_ib_register_client();
1460         if (rc) {
1461                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1462                 goto out_sock;
1463         }
1464
1465         static_branch_enable(&tcp_have_smc);
1466         return 0;
1467
1468 out_sock:
1469         sock_unregister(PF_SMC);
1470 out_proto:
1471         proto_unregister(&smc_proto);
1472 out_pnet:
1473         smc_pnet_exit();
1474         return rc;
1475 }
1476
1477 static void __exit smc_exit(void)
1478 {
1479         struct smc_link_group *lgr, *lg;
1480         LIST_HEAD(lgr_freeing_list);
1481
1482         spin_lock_bh(&smc_lgr_list.lock);
1483         if (!list_empty(&smc_lgr_list.list))
1484                 list_splice_init(&smc_lgr_list.list, &lgr_freeing_list);
1485         spin_unlock_bh(&smc_lgr_list.lock);
1486         list_for_each_entry_safe(lgr, lg, &lgr_freeing_list, list) {
1487                 list_del_init(&lgr->list);
1488                 smc_lgr_free(lgr); /* free link group */
1489         }
1490         static_branch_disable(&tcp_have_smc);
1491         smc_ib_unregister_client();
1492         sock_unregister(PF_SMC);
1493         proto_unregister(&smc_proto);
1494         smc_pnet_exit();
1495 }
1496
1497 module_init(smc_init);
1498 module_exit(smc_exit);
1499
1500 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1501 MODULE_DESCRIPTION("smc socket address family");
1502 MODULE_LICENSE("GPL");
1503 MODULE_ALIAS_NETPROTO(PF_SMC);