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