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  *    - support for alternate links postponed
  11  *
  12  *  Copyright IBM Corp. 2016, 2018
  13  *
  14  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  15  *              based on prototype from Frank Blaschka
  16  */
  17
  18 #define KMSG_COMPONENT "smc"
  19 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  20
  21 #include <linux/module.h>
  22 #include <linux/socket.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/in.h>
  25 #include <linux/sched/signal.h>
  26 #include <linux/if_vlan.h>
  27
  28 #include <net/sock.h>
  29 #include <net/tcp.h>
  30 #include <net/smc.h>
  31 #include <asm/ioctls.h>
  32
  33 #include "smc.h"
  34 #include "smc_clc.h"
  35 #include "smc_llc.h"
  36 #include "smc_cdc.h"
  37 #include "smc_core.h"
  38 #include "smc_ib.h"
  39 #include "smc_ism.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 static void smc_tcp_listen_work(struct work_struct *);
  50 static void smc_connect_work(struct work_struct *);
  51
  52 static void smc_set_keepalive(struct sock *sk, int val)
  53 {
  54         struct smc_sock *smc = smc_sk(sk);
  55
  56         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
  57 }
  58
  59 static struct smc_hashinfo smc_v4_hashinfo = {
  60         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
  61 };
  62
  63 static struct smc_hashinfo smc_v6_hashinfo = {
  64         .lock = __RW_LOCK_UNLOCKED(smc_v6_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 struct proto smc_proto6 = {
 107         .name           = "SMC6",
 108         .owner          = THIS_MODULE,
 109         .keepalive      = smc_set_keepalive,
 110         .hash           = smc_hash_sk,
 111         .unhash         = smc_unhash_sk,
 112         .obj_size       = sizeof(struct smc_sock),
 113         .h.smc_hash     = &smc_v6_hashinfo,
 114         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 115 };
 116 EXPORT_SYMBOL_GPL(smc_proto6);
 117
 118 static int smc_release(struct socket *sock)
 119 {
 120         struct sock *sk = sock->sk;
 121         struct smc_sock *smc;
 122         int rc = 0;
 123
 124         if (!sk)
 125                 goto out;
 126
 127         smc = smc_sk(sk);
 128
 129         /* cleanup for a dangling non-blocking connect */
 130         flush_work(&smc->connect_work);
 131         kfree(smc->connect_info);
 132         smc->connect_info = NULL;
 133
 134         if (sk->sk_state == SMC_LISTEN)
 135                 /* smc_close_non_accepted() is called and acquires
 136                  * sock lock for child sockets again
 137                  */
 138                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 139         else
 140                 lock_sock(sk);
 141
 142         if (!smc->use_fallback) {
 143                 rc = smc_close_active(smc);
 144                 sock_set_flag(sk, SOCK_DEAD);
 145                 sk->sk_shutdown |= SHUTDOWN_MASK;
 146         }
 147         if (smc->clcsock) {
 148                 sock_release(smc->clcsock);
 149                 smc->clcsock = NULL;
 150         }
 151         if (smc->use_fallback) {
 152                 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
 153                         sock_put(sk); /* passive closing */
 154                 sk->sk_state = SMC_CLOSED;
 155                 sk->sk_state_change(sk);
 156         }
 157
 158         /* detach socket */
 159         sock_orphan(sk);
 160         sock->sk = NULL;
 161         if (!smc->use_fallback && sk->sk_state == SMC_CLOSED)
 162                 smc_conn_free(&smc->conn);
 163         release_sock(sk);
 164
 165         sk->sk_prot->unhash(sk);
 166         sock_put(sk); /* final sock_put */
 167 out:
 168         return rc;
 169 }
 170
 171 static void smc_destruct(struct sock *sk)
 172 {
 173         if (sk->sk_state != SMC_CLOSED)
 174                 return;
 175         if (!sock_flag(sk, SOCK_DEAD))
 176                 return;
 177
 178         sk_refcnt_debug_dec(sk);
 179 }
 180
 181 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 182                                    int protocol)
 183 {
 184         struct smc_sock *smc;
 185         struct proto *prot;
 186         struct sock *sk;
 187
 188         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
 189         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
 190         if (!sk)
 191                 return NULL;
 192
 193         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 194         sk->sk_state = SMC_INIT;
 195         sk->sk_destruct = smc_destruct;
 196         sk->sk_protocol = protocol;
 197         smc = smc_sk(sk);
 198         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 199         INIT_WORK(&smc->connect_work, smc_connect_work);
 200         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 201         INIT_LIST_HEAD(&smc->accept_q);
 202         spin_lock_init(&smc->accept_q_lock);
 203         spin_lock_init(&smc->conn.send_lock);
 204         sk->sk_prot->hash(sk);
 205         sk_refcnt_debug_inc(sk);
 206
 207         return sk;
 208 }
 209
 210 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 211                     int addr_len)
 212 {
 213         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 214         struct sock *sk = sock->sk;
 215         struct smc_sock *smc;
 216         int rc;
 217
 218         smc = smc_sk(sk);
 219
 220         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 221         rc = -EINVAL;
 222         if (addr_len < sizeof(struct sockaddr_in))
 223                 goto out;
 224
 225         rc = -EAFNOSUPPORT;
 226         if (addr->sin_family != AF_INET &&
 227             addr->sin_family != AF_INET6 &&
 228             addr->sin_family != AF_UNSPEC)
 229                 goto out;
 230         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 231         if (addr->sin_family == AF_UNSPEC &&
 232             addr->sin_addr.s_addr != htonl(INADDR_ANY))
 233                 goto out;
 234
 235         lock_sock(sk);
 236
 237         /* Check if socket is already active */
 238         rc = -EINVAL;
 239         if (sk->sk_state != SMC_INIT)
 240                 goto out_rel;
 241
 242         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 243         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 244
 245 out_rel:
 246         release_sock(sk);
 247 out:
 248         return rc;
 249 }
 250
 251 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 252                                    unsigned long mask)
 253 {
 254         /* options we don't get control via setsockopt for */
 255         nsk->sk_type = osk->sk_type;
 256         nsk->sk_sndbuf = osk->sk_sndbuf;
 257         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 258         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 259         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 260         nsk->sk_mark = osk->sk_mark;
 261         nsk->sk_priority = osk->sk_priority;
 262         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 263         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 264         nsk->sk_err = osk->sk_err;
 265
 266         nsk->sk_flags &= ~mask;
 267         nsk->sk_flags |= osk->sk_flags & mask;
 268 }
 269
 270 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 271                              (1UL << SOCK_KEEPOPEN) | \
 272                              (1UL << SOCK_LINGER) | \
 273                              (1UL << SOCK_BROADCAST) | \
 274                              (1UL << SOCK_TIMESTAMP) | \
 275                              (1UL << SOCK_DBG) | \
 276                              (1UL << SOCK_RCVTSTAMP) | \
 277                              (1UL << SOCK_RCVTSTAMPNS) | \
 278                              (1UL << SOCK_LOCALROUTE) | \
 279                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 280                              (1UL << SOCK_RXQ_OVFL) | \
 281                              (1UL << SOCK_WIFI_STATUS) | \
 282                              (1UL << SOCK_NOFCS) | \
 283                              (1UL << SOCK_FILTER_LOCKED))
 284 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 285  * clc socket (since smc is not called for these options from net/core)
 286  */
 287 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 288 {
 289         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 290 }
 291
 292 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 293                              (1UL << SOCK_KEEPOPEN) | \
 294                              (1UL << SOCK_LINGER) | \
 295                              (1UL << SOCK_DBG))
 296 /* copy only settings and flags relevant for smc from clc to smc socket */
 297 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 298 {
 299         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 300 }
 301
 302 /* register a new rmb, optionally send confirm_rkey msg to register with peer */
 303 static int smc_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc,
 304                        bool conf_rkey)
 305 {
 306         /* register memory region for new rmb */
 307         if (smc_wr_reg_send(link, rmb_desc->mr_rx[SMC_SINGLE_LINK])) {
 308                 rmb_desc->regerr = 1;
 309                 return -EFAULT;
 310         }
 311         if (!conf_rkey)
 312                 return 0;
 313         /* exchange confirm_rkey msg with peer */
 314         if (smc_llc_do_confirm_rkey(link, rmb_desc)) {
 315                 rmb_desc->regerr = 1;
 316                 return -EFAULT;
 317         }
 318         return 0;
 319 }
 320
 321 static int smc_clnt_conf_first_link(struct smc_sock *smc)
 322 {
 323         struct net *net = sock_net(smc->clcsock->sk);
 324         struct smc_link_group *lgr = smc->conn.lgr;
 325         struct smc_link *link;
 326         int rest;
 327         int rc;
 328
 329         link = &lgr->lnk[SMC_SINGLE_LINK];
 330         /* receive CONFIRM LINK request from server over RoCE fabric */
 331         rest = wait_for_completion_interruptible_timeout(
 332                 &link->llc_confirm,
 333                 SMC_LLC_WAIT_FIRST_TIME);
 334         if (rest <= 0) {
 335                 struct smc_clc_msg_decline dclc;
 336
 337                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 338                                       SMC_CLC_DECLINE);
 339                 return rc;
 340         }
 341
 342         if (link->llc_confirm_rc)
 343                 return SMC_CLC_DECL_RMBE_EC;
 344
 345         rc = smc_ib_modify_qp_rts(link);
 346         if (rc)
 347                 return SMC_CLC_DECL_ERR_RDYLNK;
 348
 349         smc_wr_remember_qp_attr(link);
 350
 351         if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
 352                 return SMC_CLC_DECL_ERR_REGRMB;
 353
 354         /* send CONFIRM LINK response over RoCE fabric */
 355         rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
 356         if (rc < 0)
 357                 return SMC_CLC_DECL_TIMEOUT_CL;
 358
 359         /* receive ADD LINK request from server over RoCE fabric */
 360         rest = wait_for_completion_interruptible_timeout(&link->llc_add,
 361                                                          SMC_LLC_WAIT_TIME);
 362         if (rest <= 0) {
 363                 struct smc_clc_msg_decline dclc;
 364
 365                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 366                                       SMC_CLC_DECLINE);
 367                 return rc;
 368         }
 369
 370         /* send add link reject message, only one link supported for now */
 371         rc = smc_llc_send_add_link(link,
 372                                    link->smcibdev->mac[link->ibport - 1],
 373                                    link->gid, SMC_LLC_RESP);
 374         if (rc < 0)
 375                 return SMC_CLC_DECL_TIMEOUT_AL;
 376
 377         smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
 378
 379         return 0;
 380 }
 381
 382 static void smcr_conn_save_peer_info(struct smc_sock *smc,
 383                                      struct smc_clc_msg_accept_confirm *clc)
 384 {
 385         int bufsize = smc_uncompress_bufsize(clc->rmbe_size);
 386
 387         smc->conn.peer_rmbe_idx = clc->rmbe_idx;
 388         smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
 389         smc->conn.peer_rmbe_size = bufsize;
 390         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 391         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
 392 }
 393
 394 static void smcd_conn_save_peer_info(struct smc_sock *smc,
 395                                      struct smc_clc_msg_accept_confirm *clc)
 396 {
 397         int bufsize = smc_uncompress_bufsize(clc->dmbe_size);
 398
 399         smc->conn.peer_rmbe_idx = clc->dmbe_idx;
 400         smc->conn.peer_token = clc->token;
 401         /* msg header takes up space in the buffer */
 402         smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
 403         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 404         smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
 405 }
 406
 407 static void smc_conn_save_peer_info(struct smc_sock *smc,
 408                                     struct smc_clc_msg_accept_confirm *clc)
 409 {
 410         if (smc->conn.lgr->is_smcd)
 411                 smcd_conn_save_peer_info(smc, clc);
 412         else
 413                 smcr_conn_save_peer_info(smc, clc);
 414 }
 415
 416 static void smc_link_save_peer_info(struct smc_link *link,
 417                                     struct smc_clc_msg_accept_confirm *clc)
 418 {
 419         link->peer_qpn = ntoh24(clc->qpn);
 420         memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
 421         memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
 422         link->peer_psn = ntoh24(clc->psn);
 423         link->peer_mtu = clc->qp_mtu;
 424 }
 425
 426 /* fall back during connect */
 427 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
 428 {
 429         smc->use_fallback = true;
 430         smc->fallback_rsn = reason_code;
 431         smc_copy_sock_settings_to_clc(smc);
 432         if (smc->sk.sk_state == SMC_INIT)
 433                 smc->sk.sk_state = SMC_ACTIVE;
 434         return 0;
 435 }
 436
 437 /* decline and fall back during connect */
 438 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
 439 {
 440         int rc;
 441
 442         if (reason_code < 0) { /* error, fallback is not possible */
 443                 if (smc->sk.sk_state == SMC_INIT)
 444                         sock_put(&smc->sk); /* passive closing */
 445                 return reason_code;
 446         }
 447         if (reason_code != SMC_CLC_DECL_PEERDECL) {
 448                 rc = smc_clc_send_decline(smc, reason_code);
 449                 if (rc < 0) {
 450                         if (smc->sk.sk_state == SMC_INIT)
 451                                 sock_put(&smc->sk); /* passive closing */
 452                         return rc;
 453                 }
 454         }
 455         return smc_connect_fallback(smc, reason_code);
 456 }
 457
 458 /* abort connecting */
 459 static int smc_connect_abort(struct smc_sock *smc, int reason_code,
 460                              int local_contact)
 461 {
 462         if (local_contact == SMC_FIRST_CONTACT)
 463                 smc_lgr_forget(smc->conn.lgr);
 464         mutex_unlock(&smc_create_lgr_pending);
 465         smc_conn_free(&smc->conn);
 466         return reason_code;
 467 }
 468
 469 /* check if there is a rdma device available for this connection. */
 470 /* called for connect and listen */
 471 static int smc_check_rdma(struct smc_sock *smc, struct smc_ib_device **ibdev,
 472                           u8 *ibport, unsigned short vlan_id, u8 gid[])
 473 {
 474         int reason_code = 0;
 475
 476         /* PNET table look up: search active ib_device and port
 477          * within same PNETID that also contains the ethernet device
 478          * used for the internal TCP socket
 479          */
 480         smc_pnet_find_roce_resource(smc->clcsock->sk, ibdev, ibport, vlan_id,
 481                                     gid);
 482         if (!(*ibdev))
 483                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 484
 485         return reason_code;
 486 }
 487
 488 /* check if there is an ISM device available for this connection. */
 489 /* called for connect and listen */
 490 static int smc_check_ism(struct smc_sock *smc, struct smcd_dev **ismdev)
 491 {
 492         /* Find ISM device with same PNETID as connecting interface  */
 493         smc_pnet_find_ism_resource(smc->clcsock->sk, ismdev);
 494         if (!(*ismdev))
 495                 return SMC_CLC_DECL_CNFERR; /* configuration error */
 496         return 0;
 497 }
 498
 499 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
 500 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
 501                                       struct smcd_dev *ismdev,
 502                                       unsigned short vlan_id)
 503 {
 504         if (vlan_id && smc_ism_get_vlan(ismdev, vlan_id))
 505                 return SMC_CLC_DECL_CNFERR;
 506         return 0;
 507 }
 508
 509 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
 510  * used, the VLAN ID will be registered again during the connection setup.
 511  */
 512 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc, bool is_smcd,
 513                                         struct smcd_dev *ismdev,
 514                                         unsigned short vlan_id)
 515 {
 516         if (!is_smcd)
 517                 return 0;
 518         if (vlan_id && smc_ism_put_vlan(ismdev, vlan_id))
 519                 return SMC_CLC_DECL_CNFERR;
 520         return 0;
 521 }
 522
 523 /* CLC handshake during connect */
 524 static int smc_connect_clc(struct smc_sock *smc, int smc_type,
 525                            struct smc_clc_msg_accept_confirm *aclc,
 526                            struct smc_ib_device *ibdev, u8 ibport,
 527                            u8 gid[], struct smcd_dev *ismdev)
 528 {
 529         int rc = 0;
 530
 531         /* do inband token exchange */
 532         rc = smc_clc_send_proposal(smc, smc_type, ibdev, ibport, gid, ismdev);
 533         if (rc)
 534                 return rc;
 535         /* receive SMC Accept CLC message */
 536         return smc_clc_wait_msg(smc, aclc, sizeof(*aclc), SMC_CLC_ACCEPT);
 537 }
 538
 539 /* setup for RDMA connection of client */
 540 static int smc_connect_rdma(struct smc_sock *smc,
 541                             struct smc_clc_msg_accept_confirm *aclc,
 542                             struct smc_ib_device *ibdev, u8 ibport)
 543 {
 544         int local_contact = SMC_FIRST_CONTACT;
 545         struct smc_link *link;
 546         int reason_code = 0;
 547
 548         mutex_lock(&smc_create_lgr_pending);
 549         local_contact = smc_conn_create(smc, false, aclc->hdr.flag, ibdev,
 550                                         ibport, &aclc->lcl, NULL, 0);
 551         if (local_contact < 0) {
 552                 if (local_contact == -ENOMEM)
 553                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
 554                 else if (local_contact == -ENOLINK)
 555                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
 556                 else
 557                         reason_code = SMC_CLC_DECL_INTERR; /* other error */
 558                 return smc_connect_abort(smc, reason_code, 0);
 559         }
 560         link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
 561
 562         smc_conn_save_peer_info(smc, aclc);
 563
 564         /* create send buffer and rmb */
 565         if (smc_buf_create(smc, false))
 566                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, local_contact);
 567
 568         if (local_contact == SMC_FIRST_CONTACT)
 569                 smc_link_save_peer_info(link, aclc);
 570
 571         if (smc_rmb_rtoken_handling(&smc->conn, aclc))
 572                 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RTOK,
 573                                          local_contact);
 574
 575         smc_close_init(smc);
 576         smc_rx_init(smc);
 577
 578         if (local_contact == SMC_FIRST_CONTACT) {
 579                 if (smc_ib_ready_link(link))
 580                         return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RDYLNK,
 581                                                  local_contact);
 582         } else {
 583                 if (!smc->conn.rmb_desc->reused &&
 584                     smc_reg_rmb(link, smc->conn.rmb_desc, true))
 585                         return smc_connect_abort(smc, SMC_CLC_DECL_ERR_REGRMB,
 586                                                  local_contact);
 587         }
 588         smc_rmb_sync_sg_for_device(&smc->conn);
 589
 590         reason_code = smc_clc_send_confirm(smc);
 591         if (reason_code)
 592                 return smc_connect_abort(smc, reason_code, local_contact);
 593
 594         smc_tx_init(smc);
 595
 596         if (local_contact == SMC_FIRST_CONTACT) {
 597                 /* QP confirmation over RoCE fabric */
 598                 reason_code = smc_clnt_conf_first_link(smc);
 599                 if (reason_code)
 600                         return smc_connect_abort(smc, reason_code,
 601                                                  local_contact);
 602         }
 603         mutex_unlock(&smc_create_lgr_pending);
 604
 605         smc_copy_sock_settings_to_clc(smc);
 606         if (smc->sk.sk_state == SMC_INIT)
 607                 smc->sk.sk_state = SMC_ACTIVE;
 608
 609         return 0;
 610 }
 611
 612 /* setup for ISM connection of client */
 613 static int smc_connect_ism(struct smc_sock *smc,
 614                            struct smc_clc_msg_accept_confirm *aclc,
 615                            struct smcd_dev *ismdev)
 616 {
 617         int local_contact = SMC_FIRST_CONTACT;
 618         int rc = 0;
 619
 620         mutex_lock(&smc_create_lgr_pending);
 621         local_contact = smc_conn_create(smc, true, aclc->hdr.flag, NULL, 0,
 622                                         NULL, ismdev, aclc->gid);
 623         if (local_contact < 0)
 624                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, 0);
 625
 626         /* Create send and receive buffers */
 627         if (smc_buf_create(smc, true))
 628                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, local_contact);
 629
 630         smc_conn_save_peer_info(smc, aclc);
 631         smc_close_init(smc);
 632         smc_rx_init(smc);
 633         smc_tx_init(smc);
 634
 635         rc = smc_clc_send_confirm(smc);
 636         if (rc)
 637                 return smc_connect_abort(smc, rc, local_contact);
 638         mutex_unlock(&smc_create_lgr_pending);
 639
 640         smc_copy_sock_settings_to_clc(smc);
 641         if (smc->sk.sk_state == SMC_INIT)
 642                 smc->sk.sk_state = SMC_ACTIVE;
 643
 644         return 0;
 645 }
 646
 647 /* perform steps before actually connecting */
 648 static int __smc_connect(struct smc_sock *smc)
 649 {
 650         bool ism_supported = false, rdma_supported = false;
 651         struct smc_clc_msg_accept_confirm aclc;
 652         struct smc_ib_device *ibdev;
 653         struct smcd_dev *ismdev;
 654         u8 gid[SMC_GID_SIZE];
 655         unsigned short vlan;
 656         int smc_type;
 657         int rc = 0;
 658         u8 ibport;
 659
 660         sock_hold(&smc->sk); /* sock put in passive closing */
 661
 662         if (smc->use_fallback)
 663                 return smc_connect_fallback(smc, smc->fallback_rsn);
 664
 665         /* if peer has not signalled SMC-capability, fall back */
 666         if (!tcp_sk(smc->clcsock->sk)->syn_smc)
 667                 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
 668
 669         /* IPSec connections opt out of SMC-R optimizations */
 670         if (using_ipsec(smc))
 671                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
 672
 673         /* check for VLAN ID */
 674         if (smc_vlan_by_tcpsk(smc->clcsock, &vlan))
 675                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_CNFERR);
 676
 677         /* check if there is an ism device available */
 678         if (!smc_check_ism(smc, &ismdev) &&
 679             !smc_connect_ism_vlan_setup(smc, ismdev, vlan)) {
 680                 /* ISM is supported for this connection */
 681                 ism_supported = true;
 682                 smc_type = SMC_TYPE_D;
 683         }
 684
 685         /* check if there is a rdma device available */
 686         if (!smc_check_rdma(smc, &ibdev, &ibport, vlan, gid)) {
 687                 /* RDMA is supported for this connection */
 688                 rdma_supported = true;
 689                 if (ism_supported)
 690                         smc_type = SMC_TYPE_B; /* both */
 691                 else
 692                         smc_type = SMC_TYPE_R; /* only RDMA */
 693         }
 694
 695         /* if neither ISM nor RDMA are supported, fallback */
 696         if (!rdma_supported && !ism_supported)
 697                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_NOSMCDEV);
 698
 699         /* perform CLC handshake */
 700         rc = smc_connect_clc(smc, smc_type, &aclc, ibdev, ibport, gid, ismdev);
 701         if (rc) {
 702                 smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
 703                 return smc_connect_decline_fallback(smc, rc);
 704         }
 705
 706         /* depending on previous steps, connect using rdma or ism */
 707         if (rdma_supported && aclc.hdr.path == SMC_TYPE_R)
 708                 rc = smc_connect_rdma(smc, &aclc, ibdev, ibport);
 709         else if (ism_supported && aclc.hdr.path == SMC_TYPE_D)
 710                 rc = smc_connect_ism(smc, &aclc, ismdev);
 711         else
 712                 rc = SMC_CLC_DECL_MODEUNSUPP;
 713         if (rc) {
 714                 smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
 715                 return smc_connect_decline_fallback(smc, rc);
 716         }
 717
 718         smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
 719         return 0;
 720 }
 721
 722 static void smc_connect_work(struct work_struct *work)
 723 {
 724         struct smc_sock *smc = container_of(work, struct smc_sock,
 725                                             connect_work);
 726         int rc;
 727
 728         lock_sock(&smc->sk);
 729         rc = kernel_connect(smc->clcsock, &smc->connect_info->addr,
 730                             smc->connect_info->alen, smc->connect_info->flags);
 731         if (smc->clcsock->sk->sk_err) {
 732                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
 733                 goto out;
 734         }
 735         if (rc < 0) {
 736                 smc->sk.sk_err = -rc;
 737                 goto out;
 738         }
 739
 740         rc = __smc_connect(smc);
 741         if (rc < 0)
 742                 smc->sk.sk_err = -rc;
 743
 744 out:
 745         smc->sk.sk_state_change(&smc->sk);
 746         kfree(smc->connect_info);
 747         smc->connect_info = NULL;
 748         release_sock(&smc->sk);
 749 }
 750
 751 static int smc_connect(struct socket *sock, struct sockaddr *addr,
 752                        int alen, int flags)
 753 {
 754         struct sock *sk = sock->sk;
 755         struct smc_sock *smc;
 756         int rc = -EINVAL;
 757
 758         smc = smc_sk(sk);
 759
 760         /* separate smc parameter checking to be safe */
 761         if (alen < sizeof(addr->sa_family))
 762                 goto out_err;
 763         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
 764                 goto out_err;
 765
 766         lock_sock(sk);
 767         switch (sk->sk_state) {
 768         default:
 769                 goto out;
 770         case SMC_ACTIVE:
 771                 rc = -EISCONN;
 772                 goto out;
 773         case SMC_INIT:
 774                 rc = 0;
 775                 break;
 776         }
 777
 778         smc_copy_sock_settings_to_clc(smc);
 779         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
 780         if (flags & O_NONBLOCK) {
 781                 if (smc->connect_info) {
 782                         rc = -EALREADY;
 783                         goto out;
 784                 }
 785                 smc->connect_info = kzalloc(alen + 2 * sizeof(int), GFP_KERNEL);
 786                 if (!smc->connect_info) {
 787                         rc = -ENOMEM;
 788                         goto out;
 789                 }
 790                 smc->connect_info->alen = alen;
 791                 smc->connect_info->flags = flags ^ O_NONBLOCK;
 792                 memcpy(&smc->connect_info->addr, addr, alen);
 793                 schedule_work(&smc->connect_work);
 794                 rc = -EINPROGRESS;
 795         } else {
 796                 rc = kernel_connect(smc->clcsock, addr, alen, flags);
 797                 if (rc)
 798                         goto out;
 799
 800                 rc = __smc_connect(smc);
 801                 if (rc < 0)
 802                         goto out;
 803                 else
 804                         rc = 0; /* success cases including fallback */
 805         }
 806
 807 out:
 808         release_sock(sk);
 809 out_err:
 810         return rc;
 811 }
 812
 813 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 814 {
 815         struct socket *new_clcsock = NULL;
 816         struct sock *lsk = &lsmc->sk;
 817         struct sock *new_sk;
 818         int rc;
 819
 820         release_sock(lsk);
 821         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
 822         if (!new_sk) {
 823                 rc = -ENOMEM;
 824                 lsk->sk_err = ENOMEM;
 825                 *new_smc = NULL;
 826                 lock_sock(lsk);
 827                 goto out;
 828         }
 829         *new_smc = smc_sk(new_sk);
 830
 831         rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
 832         lock_sock(lsk);
 833         if  (rc < 0)
 834                 lsk->sk_err = -rc;
 835         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
 836                 if (new_clcsock)
 837                         sock_release(new_clcsock);
 838                 new_sk->sk_state = SMC_CLOSED;
 839                 sock_set_flag(new_sk, SOCK_DEAD);
 840                 new_sk->sk_prot->unhash(new_sk);
 841                 sock_put(new_sk); /* final */
 842                 *new_smc = NULL;
 843                 goto out;
 844         }
 845
 846         (*new_smc)->clcsock = new_clcsock;
 847 out:
 848         return rc;
 849 }
 850
 851 /* add a just created sock to the accept queue of the listen sock as
 852  * candidate for a following socket accept call from user space
 853  */
 854 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
 855 {
 856         struct smc_sock *par = smc_sk(parent);
 857
 858         sock_hold(sk); /* sock_put in smc_accept_unlink () */
 859         spin_lock(&par->accept_q_lock);
 860         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
 861         spin_unlock(&par->accept_q_lock);
 862         sk_acceptq_added(parent);
 863 }
 864
 865 /* remove a socket from the accept queue of its parental listening socket */
 866 static void smc_accept_unlink(struct sock *sk)
 867 {
 868         struct smc_sock *par = smc_sk(sk)->listen_smc;
 869
 870         spin_lock(&par->accept_q_lock);
 871         list_del_init(&smc_sk(sk)->accept_q);
 872         spin_unlock(&par->accept_q_lock);
 873         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
 874         sock_put(sk); /* sock_hold in smc_accept_enqueue */
 875 }
 876
 877 /* remove a sock from the accept queue to bind it to a new socket created
 878  * for a socket accept call from user space
 879  */
 880 struct sock *smc_accept_dequeue(struct sock *parent,
 881                                 struct socket *new_sock)
 882 {
 883         struct smc_sock *isk, *n;
 884         struct sock *new_sk;
 885
 886         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
 887                 new_sk = (struct sock *)isk;
 888
 889                 smc_accept_unlink(new_sk);
 890                 if (new_sk->sk_state == SMC_CLOSED) {
 891                         if (isk->clcsock) {
 892                                 sock_release(isk->clcsock);
 893                                 isk->clcsock = NULL;
 894                         }
 895                         new_sk->sk_prot->unhash(new_sk);
 896                         sock_put(new_sk); /* final */
 897                         continue;
 898                 }
 899                 if (new_sock)
 900                         sock_graft(new_sk, new_sock);
 901                 return new_sk;
 902         }
 903         return NULL;
 904 }
 905
 906 /* clean up for a created but never accepted sock */
 907 void smc_close_non_accepted(struct sock *sk)
 908 {
 909         struct smc_sock *smc = smc_sk(sk);
 910
 911         lock_sock(sk);
 912         if (!sk->sk_lingertime)
 913                 /* wait for peer closing */
 914                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
 915         if (!smc->use_fallback) {
 916                 smc_close_active(smc);
 917                 sock_set_flag(sk, SOCK_DEAD);
 918                 sk->sk_shutdown |= SHUTDOWN_MASK;
 919         }
 920         if (smc->clcsock) {
 921                 struct socket *tcp;
 922
 923                 tcp = smc->clcsock;
 924                 smc->clcsock = NULL;
 925                 sock_release(tcp);
 926         }
 927         if (smc->use_fallback) {
 928                 sock_put(sk); /* passive closing */
 929                 sk->sk_state = SMC_CLOSED;
 930         } else {
 931                 if (sk->sk_state == SMC_CLOSED)
 932                         smc_conn_free(&smc->conn);
 933         }
 934         release_sock(sk);
 935         sk->sk_prot->unhash(sk);
 936         sock_put(sk); /* final sock_put */
 937 }
 938
 939 static int smc_serv_conf_first_link(struct smc_sock *smc)
 940 {
 941         struct net *net = sock_net(smc->clcsock->sk);
 942         struct smc_link_group *lgr = smc->conn.lgr;
 943         struct smc_link *link;
 944         int rest;
 945         int rc;
 946
 947         link = &lgr->lnk[SMC_SINGLE_LINK];
 948
 949         if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
 950                 return SMC_CLC_DECL_ERR_REGRMB;
 951
 952         /* send CONFIRM LINK request to client over the RoCE fabric */
 953         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
 954         if (rc < 0)
 955                 return SMC_CLC_DECL_TIMEOUT_CL;
 956
 957         /* receive CONFIRM LINK response from client over the RoCE fabric */
 958         rest = wait_for_completion_interruptible_timeout(
 959                 &link->llc_confirm_resp,
 960                 SMC_LLC_WAIT_FIRST_TIME);
 961         if (rest <= 0) {
 962                 struct smc_clc_msg_decline dclc;
 963
 964                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 965                                       SMC_CLC_DECLINE);
 966                 return rc;
 967         }
 968
 969         if (link->llc_confirm_resp_rc)
 970                 return SMC_CLC_DECL_RMBE_EC;
 971
 972         /* send ADD LINK request to client over the RoCE fabric */
 973         rc = smc_llc_send_add_link(link,
 974                                    link->smcibdev->mac[link->ibport - 1],
 975                                    link->gid, SMC_LLC_REQ);
 976         if (rc < 0)
 977                 return SMC_CLC_DECL_TIMEOUT_AL;
 978
 979         /* receive ADD LINK response from client over the RoCE fabric */
 980         rest = wait_for_completion_interruptible_timeout(&link->llc_add_resp,
 981                                                          SMC_LLC_WAIT_TIME);
 982         if (rest <= 0) {
 983                 struct smc_clc_msg_decline dclc;
 984
 985                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 986                                       SMC_CLC_DECLINE);
 987                 return rc;
 988         }
 989
 990         smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
 991
 992         return 0;
 993 }
 994
 995 /* listen worker: finish */
 996 static void smc_listen_out(struct smc_sock *new_smc)
 997 {
 998         struct smc_sock *lsmc = new_smc->listen_smc;
 999         struct sock *newsmcsk = &new_smc->sk;
1000
1001         lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1002         if (lsmc->sk.sk_state == SMC_LISTEN) {
1003                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1004         } else { /* no longer listening */
1005                 smc_close_non_accepted(newsmcsk);
1006         }
1007         release_sock(&lsmc->sk);
1008
1009         /* Wake up accept */
1010         lsmc->sk.sk_data_ready(&lsmc->sk);
1011         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1012 }
1013
1014 /* listen worker: finish in state connected */
1015 static void smc_listen_out_connected(struct smc_sock *new_smc)
1016 {
1017         struct sock *newsmcsk = &new_smc->sk;
1018
1019         sk_refcnt_debug_inc(newsmcsk);
1020         if (newsmcsk->sk_state == SMC_INIT)
1021                 newsmcsk->sk_state = SMC_ACTIVE;
1022
1023         smc_listen_out(new_smc);
1024 }
1025
1026 /* listen worker: finish in error state */
1027 static void smc_listen_out_err(struct smc_sock *new_smc)
1028 {
1029         struct sock *newsmcsk = &new_smc->sk;
1030
1031         if (newsmcsk->sk_state == SMC_INIT)
1032                 sock_put(&new_smc->sk); /* passive closing */
1033         newsmcsk->sk_state = SMC_CLOSED;
1034         smc_conn_free(&new_smc->conn);
1035
1036         smc_listen_out(new_smc);
1037 }
1038
1039 /* listen worker: decline and fall back if possible */
1040 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1041                                int local_contact)
1042 {
1043         /* RDMA setup failed, switch back to TCP */
1044         if (local_contact == SMC_FIRST_CONTACT)
1045                 smc_lgr_forget(new_smc->conn.lgr);
1046         if (reason_code < 0) { /* error, no fallback possible */
1047                 smc_listen_out_err(new_smc);
1048                 return;
1049         }
1050         smc_conn_free(&new_smc->conn);
1051         new_smc->use_fallback = true;
1052         new_smc->fallback_rsn = reason_code;
1053         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1054                 if (smc_clc_send_decline(new_smc, reason_code) < 0) {
1055                         smc_listen_out_err(new_smc);
1056                         return;
1057                 }
1058         }
1059         smc_listen_out_connected(new_smc);
1060 }
1061
1062 /* listen worker: check prefixes */
1063 static int smc_listen_rdma_check(struct smc_sock *new_smc,
1064                                  struct smc_clc_msg_proposal *pclc)
1065 {
1066         struct smc_clc_msg_proposal_prefix *pclc_prfx;
1067         struct socket *newclcsock = new_smc->clcsock;
1068
1069         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1070         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1071                 return SMC_CLC_DECL_CNFERR;
1072
1073         return 0;
1074 }
1075
1076 /* listen worker: initialize connection and buffers */
1077 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1078                                 struct smc_clc_msg_proposal *pclc,
1079                                 struct smc_ib_device *ibdev, u8 ibport,
1080                                 int *local_contact)
1081 {
1082         /* allocate connection / link group */
1083         *local_contact = smc_conn_create(new_smc, false, 0, ibdev, ibport,
1084                                          &pclc->lcl, NULL, 0);
1085         if (*local_contact < 0) {
1086                 if (*local_contact == -ENOMEM)
1087                         return SMC_CLC_DECL_MEM;/* insufficient memory*/
1088                 return SMC_CLC_DECL_INTERR; /* other error */
1089         }
1090
1091         /* create send buffer and rmb */
1092         if (smc_buf_create(new_smc, false))
1093                 return SMC_CLC_DECL_MEM;
1094
1095         return 0;
1096 }
1097
1098 /* listen worker: initialize connection and buffers for SMC-D */
1099 static int smc_listen_ism_init(struct smc_sock *new_smc,
1100                                struct smc_clc_msg_proposal *pclc,
1101                                struct smcd_dev *ismdev,
1102                                int *local_contact)
1103 {
1104         struct smc_clc_msg_smcd *pclc_smcd;
1105
1106         pclc_smcd = smc_get_clc_msg_smcd(pclc);
1107         *local_contact = smc_conn_create(new_smc, true, 0, NULL, 0, NULL,
1108                                          ismdev, pclc_smcd->gid);
1109         if (*local_contact < 0) {
1110                 if (*local_contact == -ENOMEM)
1111                         return SMC_CLC_DECL_MEM;/* insufficient memory*/
1112                 return SMC_CLC_DECL_INTERR; /* other error */
1113         }
1114
1115         /* Check if peer can be reached via ISM device */
1116         if (smc_ism_cantalk(new_smc->conn.lgr->peer_gid,
1117                             new_smc->conn.lgr->vlan_id,
1118                             new_smc->conn.lgr->smcd)) {
1119                 if (*local_contact == SMC_FIRST_CONTACT)
1120                         smc_lgr_forget(new_smc->conn.lgr);
1121                 smc_conn_free(&new_smc->conn);
1122                 return SMC_CLC_DECL_CNFERR;
1123         }
1124
1125         /* Create send and receive buffers */
1126         if (smc_buf_create(new_smc, true)) {
1127                 if (*local_contact == SMC_FIRST_CONTACT)
1128                         smc_lgr_forget(new_smc->conn.lgr);
1129                 smc_conn_free(&new_smc->conn);
1130                 return SMC_CLC_DECL_MEM;
1131         }
1132
1133         return 0;
1134 }
1135
1136 /* listen worker: register buffers */
1137 static int smc_listen_rdma_reg(struct smc_sock *new_smc, int local_contact)
1138 {
1139         struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
1140
1141         if (local_contact != SMC_FIRST_CONTACT) {
1142                 if (!new_smc->conn.rmb_desc->reused) {
1143                         if (smc_reg_rmb(link, new_smc->conn.rmb_desc, true))
1144                                 return SMC_CLC_DECL_ERR_REGRMB;
1145                 }
1146         }
1147         smc_rmb_sync_sg_for_device(&new_smc->conn);
1148
1149         return 0;
1150 }
1151
1152 /* listen worker: finish RDMA setup */
1153 static void smc_listen_rdma_finish(struct smc_sock *new_smc,
1154                                    struct smc_clc_msg_accept_confirm *cclc,
1155                                    int local_contact)
1156 {
1157         struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
1158         int reason_code = 0;
1159
1160         if (local_contact == SMC_FIRST_CONTACT)
1161                 smc_link_save_peer_info(link, cclc);
1162
1163         if (smc_rmb_rtoken_handling(&new_smc->conn, cclc)) {
1164                 reason_code = SMC_CLC_DECL_ERR_RTOK;
1165                 goto decline;
1166         }
1167
1168         if (local_contact == SMC_FIRST_CONTACT) {
1169                 if (smc_ib_ready_link(link)) {
1170                         reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1171                         goto decline;
1172                 }
1173                 /* QP confirmation over RoCE fabric */
1174                 reason_code = smc_serv_conf_first_link(new_smc);
1175                 if (reason_code)
1176                         goto decline;
1177         }
1178         return;
1179
1180 decline:
1181         mutex_unlock(&smc_create_lgr_pending);
1182         smc_listen_decline(new_smc, reason_code, local_contact);
1183 }
1184
1185 /* setup for RDMA connection of server */
1186 static void smc_listen_work(struct work_struct *work)
1187 {
1188         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1189                                                 smc_listen_work);
1190         struct socket *newclcsock = new_smc->clcsock;
1191         struct smc_clc_msg_accept_confirm cclc;
1192         struct smc_clc_msg_proposal *pclc;
1193         struct smc_ib_device *ibdev;
1194         bool ism_supported = false;
1195         struct smcd_dev *ismdev;
1196         u8 buf[SMC_CLC_MAX_LEN];
1197         int local_contact = 0;
1198         unsigned short vlan;
1199         int reason_code = 0;
1200         int rc = 0;
1201         u8 ibport;
1202
1203         if (new_smc->use_fallback) {
1204                 smc_listen_out_connected(new_smc);
1205                 return;
1206         }
1207
1208         /* check if peer is smc capable */
1209         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1210                 new_smc->use_fallback = true;
1211                 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1212                 smc_listen_out_connected(new_smc);
1213                 return;
1214         }
1215
1216         /* do inband token exchange -
1217          * wait for and receive SMC Proposal CLC message
1218          */
1219         pclc = (struct smc_clc_msg_proposal *)&buf;
1220         reason_code = smc_clc_wait_msg(new_smc, pclc, SMC_CLC_MAX_LEN,
1221                                        SMC_CLC_PROPOSAL);
1222         if (reason_code) {
1223                 smc_listen_decline(new_smc, reason_code, 0);
1224                 return;
1225         }
1226
1227         /* IPSec connections opt out of SMC-R optimizations */
1228         if (using_ipsec(new_smc)) {
1229                 smc_listen_decline(new_smc, SMC_CLC_DECL_IPSEC, 0);
1230                 return;
1231         }
1232
1233         mutex_lock(&smc_create_lgr_pending);
1234         smc_close_init(new_smc);
1235         smc_rx_init(new_smc);
1236         smc_tx_init(new_smc);
1237
1238         /* check if ISM is available */
1239         if ((pclc->hdr.path == SMC_TYPE_D || pclc->hdr.path == SMC_TYPE_B) &&
1240             !smc_check_ism(new_smc, &ismdev) &&
1241             !smc_listen_ism_init(new_smc, pclc, ismdev, &local_contact)) {
1242                 ism_supported = true;
1243         }
1244
1245         /* check if RDMA is available */
1246         if (!ism_supported &&
1247             ((pclc->hdr.path != SMC_TYPE_R && pclc->hdr.path != SMC_TYPE_B) ||
1248              smc_vlan_by_tcpsk(new_smc->clcsock, &vlan) ||
1249              smc_check_rdma(new_smc, &ibdev, &ibport, vlan, NULL) ||
1250              smc_listen_rdma_check(new_smc, pclc) ||
1251              smc_listen_rdma_init(new_smc, pclc, ibdev, ibport,
1252                                   &local_contact) ||
1253              smc_listen_rdma_reg(new_smc, local_contact))) {
1254                 /* SMC not supported, decline */
1255                 mutex_unlock(&smc_create_lgr_pending);
1256                 smc_listen_decline(new_smc, SMC_CLC_DECL_MODEUNSUPP,
1257                                    local_contact);
1258                 return;
1259         }
1260
1261         /* send SMC Accept CLC message */
1262         rc = smc_clc_send_accept(new_smc, local_contact);
1263         if (rc) {
1264                 mutex_unlock(&smc_create_lgr_pending);
1265                 smc_listen_decline(new_smc, rc, local_contact);
1266                 return;
1267         }
1268
1269         /* receive SMC Confirm CLC message */
1270         reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
1271                                        SMC_CLC_CONFIRM);
1272         if (reason_code) {
1273                 mutex_unlock(&smc_create_lgr_pending);
1274                 smc_listen_decline(new_smc, reason_code, local_contact);
1275                 return;
1276         }
1277
1278         /* finish worker */
1279         if (!ism_supported)
1280                 smc_listen_rdma_finish(new_smc, &cclc, local_contact);
1281         smc_conn_save_peer_info(new_smc, &cclc);
1282         mutex_unlock(&smc_create_lgr_pending);
1283         smc_listen_out_connected(new_smc);
1284 }
1285
1286 static void smc_tcp_listen_work(struct work_struct *work)
1287 {
1288         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1289                                              tcp_listen_work);
1290         struct sock *lsk = &lsmc->sk;
1291         struct smc_sock *new_smc;
1292         int rc = 0;
1293
1294         lock_sock(lsk);
1295         while (lsk->sk_state == SMC_LISTEN) {
1296                 rc = smc_clcsock_accept(lsmc, &new_smc);
1297                 if (rc)
1298                         goto out;
1299                 if (!new_smc)
1300                         continue;
1301
1302                 new_smc->listen_smc = lsmc;
1303                 new_smc->use_fallback = lsmc->use_fallback;
1304                 new_smc->fallback_rsn = lsmc->fallback_rsn;
1305                 sock_hold(lsk); /* sock_put in smc_listen_work */
1306                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1307                 smc_copy_sock_settings_to_smc(new_smc);
1308                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1309                 if (!schedule_work(&new_smc->smc_listen_work))
1310                         sock_put(&new_smc->sk);
1311         }
1312
1313 out:
1314         release_sock(lsk);
1315         sock_put(&lsmc->sk); /* sock_hold in smc_listen */
1316 }
1317
1318 static int smc_listen(struct socket *sock, int backlog)
1319 {
1320         struct sock *sk = sock->sk;
1321         struct smc_sock *smc;
1322         int rc;
1323
1324         smc = smc_sk(sk);
1325         lock_sock(sk);
1326
1327         rc = -EINVAL;
1328         if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
1329                 goto out;
1330
1331         rc = 0;
1332         if (sk->sk_state == SMC_LISTEN) {
1333                 sk->sk_max_ack_backlog = backlog;
1334                 goto out;
1335         }
1336         /* some socket options are handled in core, so we could not apply
1337          * them to the clc socket -- copy smc socket options to clc socket
1338          */
1339         smc_copy_sock_settings_to_clc(smc);
1340         if (!smc->use_fallback)
1341                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1342
1343         rc = kernel_listen(smc->clcsock, backlog);
1344         if (rc)
1345                 goto out;
1346         sk->sk_max_ack_backlog = backlog;
1347         sk->sk_ack_backlog = 0;
1348         sk->sk_state = SMC_LISTEN;
1349         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
1350         sock_hold(sk); /* sock_hold in tcp_listen_worker */
1351         if (!schedule_work(&smc->tcp_listen_work))
1352                 sock_put(sk);
1353
1354 out:
1355         release_sock(sk);
1356         return rc;
1357 }
1358
1359 static int smc_accept(struct socket *sock, struct socket *new_sock,
1360                       int flags, bool kern)
1361 {
1362         struct sock *sk = sock->sk, *nsk;
1363         DECLARE_WAITQUEUE(wait, current);
1364         struct smc_sock *lsmc;
1365         long timeo;
1366         int rc = 0;
1367
1368         lsmc = smc_sk(sk);
1369         sock_hold(sk); /* sock_put below */
1370         lock_sock(sk);
1371
1372         if (lsmc->sk.sk_state != SMC_LISTEN) {
1373                 rc = -EINVAL;
1374                 release_sock(sk);
1375                 goto out;
1376         }
1377
1378         /* Wait for an incoming connection */
1379         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1380         add_wait_queue_exclusive(sk_sleep(sk), &wait);
1381         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1382                 set_current_state(TASK_INTERRUPTIBLE);
1383                 if (!timeo) {
1384                         rc = -EAGAIN;
1385                         break;
1386                 }
1387                 release_sock(sk);
1388                 timeo = schedule_timeout(timeo);
1389                 /* wakeup by sk_data_ready in smc_listen_work() */
1390                 sched_annotate_sleep();
1391                 lock_sock(sk);
1392                 if (signal_pending(current)) {
1393                         rc = sock_intr_errno(timeo);
1394                         break;
1395                 }
1396         }
1397         set_current_state(TASK_RUNNING);
1398         remove_wait_queue(sk_sleep(sk), &wait);
1399
1400         if (!rc)
1401                 rc = sock_error(nsk);
1402         release_sock(sk);
1403         if (rc)
1404                 goto out;
1405
1406         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1407                 /* wait till data arrives on the socket */
1408                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1409                                                                 MSEC_PER_SEC);
1410                 if (smc_sk(nsk)->use_fallback) {
1411                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1412
1413                         lock_sock(clcsk);
1414                         if (skb_queue_empty(&clcsk->sk_receive_queue))
1415                                 sk_wait_data(clcsk, &timeo, NULL);
1416                         release_sock(clcsk);
1417                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1418                         lock_sock(nsk);
1419                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1420                         release_sock(nsk);
1421                 }
1422         }
1423
1424 out:
1425         sock_put(sk); /* sock_hold above */
1426         return rc;
1427 }
1428
1429 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1430                        int peer)
1431 {
1432         struct smc_sock *smc;
1433
1434         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1435             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1436                 return -ENOTCONN;
1437
1438         smc = smc_sk(sock->sk);
1439
1440         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1441 }
1442
1443 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1444 {
1445         struct sock *sk = sock->sk;
1446         struct smc_sock *smc;
1447         int rc = -EPIPE;
1448
1449         smc = smc_sk(sk);
1450         lock_sock(sk);
1451         if ((sk->sk_state != SMC_ACTIVE) &&
1452             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1453             (sk->sk_state != SMC_INIT))
1454                 goto out;
1455
1456         if (msg->msg_flags & MSG_FASTOPEN) {
1457                 if (sk->sk_state == SMC_INIT) {
1458                         smc->use_fallback = true;
1459                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1460                 } else {
1461                         rc = -EINVAL;
1462                         goto out;
1463                 }
1464         }
1465
1466         if (smc->use_fallback)
1467                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1468         else
1469                 rc = smc_tx_sendmsg(smc, msg, len);
1470 out:
1471         release_sock(sk);
1472         return rc;
1473 }
1474
1475 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1476                        int flags)
1477 {
1478         struct sock *sk = sock->sk;
1479         struct smc_sock *smc;
1480         int rc = -ENOTCONN;
1481
1482         smc = smc_sk(sk);
1483         lock_sock(sk);
1484         if ((sk->sk_state == SMC_INIT) ||
1485             (sk->sk_state == SMC_LISTEN) ||
1486             (sk->sk_state == SMC_CLOSED))
1487                 goto out;
1488
1489         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1490                 rc = 0;
1491                 goto out;
1492         }
1493
1494         if (smc->use_fallback) {
1495                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1496         } else {
1497                 msg->msg_namelen = 0;
1498                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1499         }
1500
1501 out:
1502         release_sock(sk);
1503         return rc;
1504 }
1505
1506 static __poll_t smc_accept_poll(struct sock *parent)
1507 {
1508         struct smc_sock *isk = smc_sk(parent);
1509         __poll_t mask = 0;
1510
1511         spin_lock(&isk->accept_q_lock);
1512         if (!list_empty(&isk->accept_q))
1513                 mask = EPOLLIN | EPOLLRDNORM;
1514         spin_unlock(&isk->accept_q_lock);
1515
1516         return mask;
1517 }
1518
1519 static __poll_t smc_poll(struct file *file, struct socket *sock,
1520                              poll_table *wait)
1521 {
1522         struct sock *sk = sock->sk;
1523         __poll_t mask = 0;
1524         struct smc_sock *smc;
1525
1526         if (!sk)
1527                 return EPOLLNVAL;
1528
1529         smc = smc_sk(sock->sk);
1530         if ((sk->sk_state == SMC_INIT) || smc->use_fallback) {
1531                 /* delegate to CLC child sock */
1532                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1533                 sk->sk_err = smc->clcsock->sk->sk_err;
1534                 if (sk->sk_err)
1535                         mask |= EPOLLERR;
1536         } else {
1537                 if (sk->sk_state != SMC_CLOSED)
1538                         sock_poll_wait(file, sk_sleep(sk), wait);
1539                 if (sk->sk_err)
1540                         mask |= EPOLLERR;
1541                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1542                     (sk->sk_state == SMC_CLOSED))
1543                         mask |= EPOLLHUP;
1544                 if (sk->sk_state == SMC_LISTEN) {
1545                         /* woken up by sk_data_ready in smc_listen_work() */
1546                         mask = smc_accept_poll(sk);
1547                 } else {
1548                         if (atomic_read(&smc->conn.sndbuf_space) ||
1549                             sk->sk_shutdown & SEND_SHUTDOWN) {
1550                                 mask |= EPOLLOUT | EPOLLWRNORM;
1551                         } else {
1552                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1553                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1554                         }
1555                         if (atomic_read(&smc->conn.bytes_to_rcv))
1556                                 mask |= EPOLLIN | EPOLLRDNORM;
1557                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1558                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1559                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
1560                                 mask |= EPOLLIN;
1561                 }
1562                 if (smc->conn.urg_state == SMC_URG_VALID)
1563                         mask |= EPOLLPRI;
1564         }
1565
1566         return mask;
1567 }
1568
1569 static int smc_shutdown(struct socket *sock, int how)
1570 {
1571         struct sock *sk = sock->sk;
1572         struct smc_sock *smc;
1573         int rc = -EINVAL;
1574         int rc1 = 0;
1575
1576         smc = smc_sk(sk);
1577
1578         if ((how < SHUT_RD) || (how > SHUT_RDWR))
1579                 return rc;
1580
1581         lock_sock(sk);
1582
1583         rc = -ENOTCONN;
1584         if ((sk->sk_state != SMC_LISTEN) &&
1585             (sk->sk_state != SMC_ACTIVE) &&
1586             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1587             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1588             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1589             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1590             (sk->sk_state != SMC_APPFINCLOSEWAIT))
1591                 goto out;
1592         if (smc->use_fallback) {
1593                 rc = kernel_sock_shutdown(smc->clcsock, how);
1594                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1595                 if (sk->sk_shutdown == SHUTDOWN_MASK)
1596                         sk->sk_state = SMC_CLOSED;
1597                 goto out;
1598         }
1599         switch (how) {
1600         case SHUT_RDWR:         /* shutdown in both directions */
1601                 rc = smc_close_active(smc);
1602                 break;
1603         case SHUT_WR:
1604                 rc = smc_close_shutdown_write(smc);
1605                 break;
1606         case SHUT_RD:
1607                 rc = 0;
1608                 /* nothing more to do because peer is not involved */
1609                 break;
1610         }
1611         if (smc->clcsock)
1612                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1613         /* map sock_shutdown_cmd constants to sk_shutdown value range */
1614         sk->sk_shutdown |= how + 1;
1615
1616 out:
1617         release_sock(sk);
1618         return rc ? rc : rc1;
1619 }
1620
1621 static int smc_setsockopt(struct socket *sock, int level, int optname,
1622                           char __user *optval, unsigned int optlen)
1623 {
1624         struct sock *sk = sock->sk;
1625         struct smc_sock *smc;
1626         int val, rc;
1627
1628         smc = smc_sk(sk);
1629
1630         /* generic setsockopts reaching us here always apply to the
1631          * CLC socket
1632          */
1633         rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1634                                            optval, optlen);
1635         if (smc->clcsock->sk->sk_err) {
1636                 sk->sk_err = smc->clcsock->sk->sk_err;
1637                 sk->sk_error_report(sk);
1638         }
1639         if (rc)
1640                 return rc;
1641
1642         if (optlen < sizeof(int))
1643                 return -EINVAL;
1644         if (get_user(val, (int __user *)optval))
1645                 return -EFAULT;
1646
1647         lock_sock(sk);
1648         switch (optname) {
1649         case TCP_ULP:
1650         case TCP_FASTOPEN:
1651         case TCP_FASTOPEN_CONNECT:
1652         case TCP_FASTOPEN_KEY:
1653         case TCP_FASTOPEN_NO_COOKIE:
1654                 /* option not supported by SMC */
1655                 if (sk->sk_state == SMC_INIT) {
1656                         smc->use_fallback = true;
1657                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1658                 } else {
1659                         if (!smc->use_fallback)
1660                                 rc = -EINVAL;
1661                 }
1662                 break;
1663         case TCP_NODELAY:
1664                 if (sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) {
1665                         if (val && !smc->use_fallback)
1666                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1667                                                  0);
1668                 }
1669                 break;
1670         case TCP_CORK:
1671                 if (sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) {
1672                         if (!val && !smc->use_fallback)
1673                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1674                                                  0);
1675                 }
1676                 break;
1677         case TCP_DEFER_ACCEPT:
1678                 smc->sockopt_defer_accept = val;
1679                 break;
1680         default:
1681                 break;
1682         }
1683         release_sock(sk);
1684
1685         return rc;
1686 }
1687
1688 static int smc_getsockopt(struct socket *sock, int level, int optname,
1689                           char __user *optval, int __user *optlen)
1690 {
1691         struct smc_sock *smc;
1692
1693         smc = smc_sk(sock->sk);
1694         /* socket options apply to the CLC socket */
1695         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1696                                              optval, optlen);
1697 }
1698
1699 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1700                      unsigned long arg)
1701 {
1702         union smc_host_cursor cons, urg;
1703         struct smc_connection *conn;
1704         struct smc_sock *smc;
1705         int answ;
1706
1707         smc = smc_sk(sock->sk);
1708         conn = &smc->conn;
1709         if (smc->use_fallback) {
1710                 if (!smc->clcsock)
1711                         return -EBADF;
1712                 return smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1713         }
1714         lock_sock(&smc->sk);
1715         switch (cmd) {
1716         case SIOCINQ: /* same as FIONREAD */
1717                 if (smc->sk.sk_state == SMC_LISTEN) {
1718                         release_sock(&smc->sk);
1719                         return -EINVAL;
1720                 }
1721                 if (smc->sk.sk_state == SMC_INIT ||
1722                     smc->sk.sk_state == SMC_CLOSED)
1723                         answ = 0;
1724                 else
1725                         answ = atomic_read(&smc->conn.bytes_to_rcv);
1726                 break;
1727         case SIOCOUTQ:
1728                 /* output queue size (not send + not acked) */
1729                 if (smc->sk.sk_state == SMC_LISTEN) {
1730                         release_sock(&smc->sk);
1731                         return -EINVAL;
1732                 }
1733                 if (smc->sk.sk_state == SMC_INIT ||
1734                     smc->sk.sk_state == SMC_CLOSED)
1735                         answ = 0;
1736                 else
1737                         answ = smc->conn.sndbuf_desc->len -
1738                                         atomic_read(&smc->conn.sndbuf_space);
1739                 break;
1740         case SIOCOUTQNSD:
1741                 /* output queue size (not send only) */
1742                 if (smc->sk.sk_state == SMC_LISTEN) {
1743                         release_sock(&smc->sk);
1744                         return -EINVAL;
1745                 }
1746                 if (smc->sk.sk_state == SMC_INIT ||
1747                     smc->sk.sk_state == SMC_CLOSED)
1748                         answ = 0;
1749                 else
1750                         answ = smc_tx_prepared_sends(&smc->conn);
1751                 break;
1752         case SIOCATMARK:
1753                 if (smc->sk.sk_state == SMC_LISTEN) {
1754                         release_sock(&smc->sk);
1755                         return -EINVAL;
1756                 }
1757                 if (smc->sk.sk_state == SMC_INIT ||
1758                     smc->sk.sk_state == SMC_CLOSED) {
1759                         answ = 0;
1760                 } else {
1761                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
1762                         smc_curs_copy(&urg, &conn->urg_curs, conn);
1763                         answ = smc_curs_diff(conn->rmb_desc->len,
1764                                              &cons, &urg) == 1;
1765                 }
1766                 break;
1767         default:
1768                 release_sock(&smc->sk);
1769                 return -ENOIOCTLCMD;
1770         }
1771         release_sock(&smc->sk);
1772
1773         return put_user(answ, (int __user *)arg);
1774 }
1775
1776 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1777                             int offset, size_t size, int flags)
1778 {
1779         struct sock *sk = sock->sk;
1780         struct smc_sock *smc;
1781         int rc = -EPIPE;
1782
1783         smc = smc_sk(sk);
1784         lock_sock(sk);
1785         if (sk->sk_state != SMC_ACTIVE) {
1786                 release_sock(sk);
1787                 goto out;
1788         }
1789         release_sock(sk);
1790         if (smc->use_fallback)
1791                 rc = kernel_sendpage(smc->clcsock, page, offset,
1792                                      size, flags);
1793         else
1794                 rc = sock_no_sendpage(sock, page, offset, size, flags);
1795
1796 out:
1797         return rc;
1798 }
1799
1800 /* Map the affected portions of the rmbe into an spd, note the number of bytes
1801  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
1802  * updates till whenever a respective page has been fully processed.
1803  * Note that subsequent recv() calls have to wait till all splice() processing
1804  * completed.
1805  */
1806 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1807                                struct pipe_inode_info *pipe, size_t len,
1808                                unsigned int flags)
1809 {
1810         struct sock *sk = sock->sk;
1811         struct smc_sock *smc;
1812         int rc = -ENOTCONN;
1813
1814         smc = smc_sk(sk);
1815         lock_sock(sk);
1816
1817         if (sk->sk_state == SMC_INIT ||
1818             sk->sk_state == SMC_LISTEN ||
1819             sk->sk_state == SMC_CLOSED)
1820                 goto out;
1821
1822         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1823                 rc = 0;
1824                 goto out;
1825         }
1826
1827         if (smc->use_fallback) {
1828                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1829                                                     pipe, len, flags);
1830         } else {
1831                 if (*ppos) {
1832                         rc = -ESPIPE;
1833                         goto out;
1834                 }
1835                 if (flags & SPLICE_F_NONBLOCK)
1836                         flags = MSG_DONTWAIT;
1837                 else
1838                         flags = 0;
1839                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
1840         }
1841 out:
1842         release_sock(sk);
1843
1844         return rc;
1845 }
1846
1847 /* must look like tcp */
1848 static const struct proto_ops smc_sock_ops = {
1849         .family         = PF_SMC,
1850         .owner          = THIS_MODULE,
1851         .release        = smc_release,
1852         .bind           = smc_bind,
1853         .connect        = smc_connect,
1854         .socketpair     = sock_no_socketpair,
1855         .accept         = smc_accept,
1856         .getname        = smc_getname,
1857         .poll           = smc_poll,
1858         .ioctl          = smc_ioctl,
1859         .listen         = smc_listen,
1860         .shutdown       = smc_shutdown,
1861         .setsockopt     = smc_setsockopt,
1862         .getsockopt     = smc_getsockopt,
1863         .sendmsg        = smc_sendmsg,
1864         .recvmsg        = smc_recvmsg,
1865         .mmap           = sock_no_mmap,
1866         .sendpage       = smc_sendpage,
1867         .splice_read    = smc_splice_read,
1868 };
1869
1870 static int smc_create(struct net *net, struct socket *sock, int protocol,
1871                       int kern)
1872 {
1873         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
1874         struct smc_sock *smc;
1875         struct sock *sk;
1876         int rc;
1877
1878         rc = -ESOCKTNOSUPPORT;
1879         if (sock->type != SOCK_STREAM)
1880                 goto out;
1881
1882         rc = -EPROTONOSUPPORT;
1883         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
1884                 goto out;
1885
1886         rc = -ENOBUFS;
1887         sock->ops = &smc_sock_ops;
1888         sk = smc_sock_alloc(net, sock, protocol);
1889         if (!sk)
1890                 goto out;
1891
1892         /* create internal TCP socket for CLC handshake and fallback */
1893         smc = smc_sk(sk);
1894         smc->use_fallback = false; /* assume rdma capability first */
1895         smc->fallback_rsn = 0;
1896         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
1897                               &smc->clcsock);
1898         if (rc) {
1899                 sk_common_release(sk);
1900                 goto out;
1901         }
1902         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1903         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1904
1905 out:
1906         return rc;
1907 }
1908
1909 static const struct net_proto_family smc_sock_family_ops = {
1910         .family = PF_SMC,
1911         .owner  = THIS_MODULE,
1912         .create = smc_create,
1913 };
1914
1915 static int __init smc_init(void)
1916 {
1917         int rc;
1918
1919         rc = smc_pnet_init();
1920         if (rc)
1921                 return rc;
1922
1923         rc = smc_llc_init();
1924         if (rc) {
1925                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
1926                 goto out_pnet;
1927         }
1928
1929         rc = smc_cdc_init();
1930         if (rc) {
1931                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
1932                 goto out_pnet;
1933         }
1934
1935         rc = proto_register(&smc_proto, 1);
1936         if (rc) {
1937                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
1938                 goto out_pnet;
1939         }
1940
1941         rc = proto_register(&smc_proto6, 1);
1942         if (rc) {
1943                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
1944                 goto out_proto;
1945         }
1946
1947         rc = sock_register(&smc_sock_family_ops);
1948         if (rc) {
1949                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1950                 goto out_proto6;
1951         }
1952         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
1953         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
1954
1955         rc = smc_ib_register_client();
1956         if (rc) {
1957                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1958                 goto out_sock;
1959         }
1960
1961         static_branch_enable(&tcp_have_smc);
1962         return 0;
1963
1964 out_sock:
1965         sock_unregister(PF_SMC);
1966 out_proto6:
1967         proto_unregister(&smc_proto6);
1968 out_proto:
1969         proto_unregister(&smc_proto);
1970 out_pnet:
1971         smc_pnet_exit();
1972         return rc;
1973 }
1974
1975 static void __exit smc_exit(void)
1976 {
1977         smc_core_exit();
1978         static_branch_disable(&tcp_have_smc);
1979         smc_ib_unregister_client();
1980         sock_unregister(PF_SMC);
1981         proto_unregister(&smc_proto6);
1982         proto_unregister(&smc_proto);
1983         smc_pnet_exit();
1984 }
1985
1986 module_init(smc_init);
1987 module_exit(smc_exit);
1988
1989 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
1990 MODULE_DESCRIPTION("smc socket address family");
1991 MODULE_LICENSE("GPL");
1992 MODULE_ALIAS_NETPROTO(PF_SMC);