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         if (smc->connect_info && sk->sk_state == SMC_INIT)
 131                 tcp_abort(smc->clcsock->sk, ECONNABORTED);
 132         flush_work(&smc->connect_work);
 133         kfree(smc->connect_info);
 134         smc->connect_info = NULL;
 135
 136         if (sk->sk_state == SMC_LISTEN)
 137                 /* smc_close_non_accepted() is called and acquires
 138                  * sock lock for child sockets again
 139                  */
 140                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 141         else
 142                 lock_sock(sk);
 143
 144         if (!smc->use_fallback) {
 145                 rc = smc_close_active(smc);
 146                 sock_set_flag(sk, SOCK_DEAD);
 147                 sk->sk_shutdown |= SHUTDOWN_MASK;
 148         }
 149         if (smc->clcsock) {
 150                 sock_release(smc->clcsock);
 151                 smc->clcsock = NULL;
 152         }
 153         if (smc->use_fallback) {
 154                 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
 155                         sock_put(sk); /* passive closing */
 156                 sk->sk_state = SMC_CLOSED;
 157                 sk->sk_state_change(sk);
 158         }
 159
 160         /* detach socket */
 161         sock_orphan(sk);
 162         sock->sk = NULL;
 163         if (!smc->use_fallback && sk->sk_state == SMC_CLOSED)
 164                 smc_conn_free(&smc->conn);
 165         release_sock(sk);
 166
 167         sk->sk_prot->unhash(sk);
 168         sock_put(sk); /* final sock_put */
 169 out:
 170         return rc;
 171 }
 172
 173 static void smc_destruct(struct sock *sk)
 174 {
 175         if (sk->sk_state != SMC_CLOSED)
 176                 return;
 177         if (!sock_flag(sk, SOCK_DEAD))
 178                 return;
 179
 180         sk_refcnt_debug_dec(sk);
 181 }
 182
 183 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 184                                    int protocol)
 185 {
 186         struct smc_sock *smc;
 187         struct proto *prot;
 188         struct sock *sk;
 189
 190         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
 191         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
 192         if (!sk)
 193                 return NULL;
 194
 195         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 196         sk->sk_state = SMC_INIT;
 197         sk->sk_destruct = smc_destruct;
 198         sk->sk_protocol = protocol;
 199         smc = smc_sk(sk);
 200         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 201         INIT_WORK(&smc->connect_work, smc_connect_work);
 202         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 203         INIT_LIST_HEAD(&smc->accept_q);
 204         spin_lock_init(&smc->accept_q_lock);
 205         spin_lock_init(&smc->conn.send_lock);
 206         sk->sk_prot->hash(sk);
 207         sk_refcnt_debug_inc(sk);
 208
 209         return sk;
 210 }
 211
 212 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 213                     int addr_len)
 214 {
 215         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 216         struct sock *sk = sock->sk;
 217         struct smc_sock *smc;
 218         int rc;
 219
 220         smc = smc_sk(sk);
 221
 222         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 223         rc = -EINVAL;
 224         if (addr_len < sizeof(struct sockaddr_in))
 225                 goto out;
 226
 227         rc = -EAFNOSUPPORT;
 228         if (addr->sin_family != AF_INET &&
 229             addr->sin_family != AF_INET6 &&
 230             addr->sin_family != AF_UNSPEC)
 231                 goto out;
 232         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 233         if (addr->sin_family == AF_UNSPEC &&
 234             addr->sin_addr.s_addr != htonl(INADDR_ANY))
 235                 goto out;
 236
 237         lock_sock(sk);
 238
 239         /* Check if socket is already active */
 240         rc = -EINVAL;
 241         if (sk->sk_state != SMC_INIT)
 242                 goto out_rel;
 243
 244         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 245         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 246
 247 out_rel:
 248         release_sock(sk);
 249 out:
 250         return rc;
 251 }
 252
 253 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 254                                    unsigned long mask)
 255 {
 256         /* options we don't get control via setsockopt for */
 257         nsk->sk_type = osk->sk_type;
 258         nsk->sk_sndbuf = osk->sk_sndbuf;
 259         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 260         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 261         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 262         nsk->sk_mark = osk->sk_mark;
 263         nsk->sk_priority = osk->sk_priority;
 264         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 265         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 266         nsk->sk_err = osk->sk_err;
 267
 268         nsk->sk_flags &= ~mask;
 269         nsk->sk_flags |= osk->sk_flags & mask;
 270 }
 271
 272 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 273                              (1UL << SOCK_KEEPOPEN) | \
 274                              (1UL << SOCK_LINGER) | \
 275                              (1UL << SOCK_BROADCAST) | \
 276                              (1UL << SOCK_TIMESTAMP) | \
 277                              (1UL << SOCK_DBG) | \
 278                              (1UL << SOCK_RCVTSTAMP) | \
 279                              (1UL << SOCK_RCVTSTAMPNS) | \
 280                              (1UL << SOCK_LOCALROUTE) | \
 281                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 282                              (1UL << SOCK_RXQ_OVFL) | \
 283                              (1UL << SOCK_WIFI_STATUS) | \
 284                              (1UL << SOCK_NOFCS) | \
 285                              (1UL << SOCK_FILTER_LOCKED))
 286 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 287  * clc socket (since smc is not called for these options from net/core)
 288  */
 289 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 290 {
 291         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 292 }
 293
 294 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 295                              (1UL << SOCK_KEEPOPEN) | \
 296                              (1UL << SOCK_LINGER) | \
 297                              (1UL << SOCK_DBG))
 298 /* copy only settings and flags relevant for smc from clc to smc socket */
 299 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 300 {
 301         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 302 }
 303
 304 /* register a new rmb, optionally send confirm_rkey msg to register with peer */
 305 static int smc_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc,
 306                        bool conf_rkey)
 307 {
 308         /* register memory region for new rmb */
 309         if (smc_wr_reg_send(link, rmb_desc->mr_rx[SMC_SINGLE_LINK])) {
 310                 rmb_desc->regerr = 1;
 311                 return -EFAULT;
 312         }
 313         if (!conf_rkey)
 314                 return 0;
 315         /* exchange confirm_rkey msg with peer */
 316         if (smc_llc_do_confirm_rkey(link, rmb_desc)) {
 317                 rmb_desc->regerr = 1;
 318                 return -EFAULT;
 319         }
 320         return 0;
 321 }
 322
 323 static int smc_clnt_conf_first_link(struct smc_sock *smc)
 324 {
 325         struct net *net = sock_net(smc->clcsock->sk);
 326         struct smc_link_group *lgr = smc->conn.lgr;
 327         struct smc_link *link;
 328         int rest;
 329         int rc;
 330
 331         link = &lgr->lnk[SMC_SINGLE_LINK];
 332         /* receive CONFIRM LINK request from server over RoCE fabric */
 333         rest = wait_for_completion_interruptible_timeout(
 334                 &link->llc_confirm,
 335                 SMC_LLC_WAIT_FIRST_TIME);
 336         if (rest <= 0) {
 337                 struct smc_clc_msg_decline dclc;
 338
 339                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 340                                       SMC_CLC_DECLINE);
 341                 return rc;
 342         }
 343
 344         if (link->llc_confirm_rc)
 345                 return SMC_CLC_DECL_RMBE_EC;
 346
 347         rc = smc_ib_modify_qp_rts(link);
 348         if (rc)
 349                 return SMC_CLC_DECL_ERR_RDYLNK;
 350
 351         smc_wr_remember_qp_attr(link);
 352
 353         if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
 354                 return SMC_CLC_DECL_ERR_REGRMB;
 355
 356         /* send CONFIRM LINK response over RoCE fabric */
 357         rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
 358         if (rc < 0)
 359                 return SMC_CLC_DECL_TIMEOUT_CL;
 360
 361         /* receive ADD LINK request from server over RoCE fabric */
 362         rest = wait_for_completion_interruptible_timeout(&link->llc_add,
 363                                                          SMC_LLC_WAIT_TIME);
 364         if (rest <= 0) {
 365                 struct smc_clc_msg_decline dclc;
 366
 367                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 368                                       SMC_CLC_DECLINE);
 369                 return rc;
 370         }
 371
 372         /* send add link reject message, only one link supported for now */
 373         rc = smc_llc_send_add_link(link,
 374                                    link->smcibdev->mac[link->ibport - 1],
 375                                    link->gid, SMC_LLC_RESP);
 376         if (rc < 0)
 377                 return SMC_CLC_DECL_TIMEOUT_AL;
 378
 379         smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
 380
 381         return 0;
 382 }
 383
 384 static void smcr_conn_save_peer_info(struct smc_sock *smc,
 385                                      struct smc_clc_msg_accept_confirm *clc)
 386 {
 387         int bufsize = smc_uncompress_bufsize(clc->rmbe_size);
 388
 389         smc->conn.peer_rmbe_idx = clc->rmbe_idx;
 390         smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
 391         smc->conn.peer_rmbe_size = bufsize;
 392         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 393         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
 394 }
 395
 396 static void smcd_conn_save_peer_info(struct smc_sock *smc,
 397                                      struct smc_clc_msg_accept_confirm *clc)
 398 {
 399         int bufsize = smc_uncompress_bufsize(clc->dmbe_size);
 400
 401         smc->conn.peer_rmbe_idx = clc->dmbe_idx;
 402         smc->conn.peer_token = clc->token;
 403         /* msg header takes up space in the buffer */
 404         smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
 405         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 406         smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
 407 }
 408
 409 static void smc_conn_save_peer_info(struct smc_sock *smc,
 410                                     struct smc_clc_msg_accept_confirm *clc)
 411 {
 412         if (smc->conn.lgr->is_smcd)
 413                 smcd_conn_save_peer_info(smc, clc);
 414         else
 415                 smcr_conn_save_peer_info(smc, clc);
 416 }
 417
 418 static void smc_link_save_peer_info(struct smc_link *link,
 419                                     struct smc_clc_msg_accept_confirm *clc)
 420 {
 421         link->peer_qpn = ntoh24(clc->qpn);
 422         memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
 423         memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
 424         link->peer_psn = ntoh24(clc->psn);
 425         link->peer_mtu = clc->qp_mtu;
 426 }
 427
 428 /* fall back during connect */
 429 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
 430 {
 431         smc->use_fallback = true;
 432         smc->fallback_rsn = reason_code;
 433         smc_copy_sock_settings_to_clc(smc);
 434         if (smc->sk.sk_state == SMC_INIT)
 435                 smc->sk.sk_state = SMC_ACTIVE;
 436         return 0;
 437 }
 438
 439 /* decline and fall back during connect */
 440 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
 441 {
 442         int rc;
 443
 444         if (reason_code < 0) { /* error, fallback is not possible */
 445                 if (smc->sk.sk_state == SMC_INIT)
 446                         sock_put(&smc->sk); /* passive closing */
 447                 return reason_code;
 448         }
 449         if (reason_code != SMC_CLC_DECL_PEERDECL) {
 450                 rc = smc_clc_send_decline(smc, reason_code);
 451                 if (rc < 0) {
 452                         if (smc->sk.sk_state == SMC_INIT)
 453                                 sock_put(&smc->sk); /* passive closing */
 454                         return rc;
 455                 }
 456         }
 457         return smc_connect_fallback(smc, reason_code);
 458 }
 459
 460 /* abort connecting */
 461 static int smc_connect_abort(struct smc_sock *smc, int reason_code,
 462                              int local_contact)
 463 {
 464         if (local_contact == SMC_FIRST_CONTACT)
 465                 smc_lgr_forget(smc->conn.lgr);
 466         mutex_unlock(&smc_create_lgr_pending);
 467         smc_conn_free(&smc->conn);
 468         return reason_code;
 469 }
 470
 471 /* check if there is a rdma device available for this connection. */
 472 /* called for connect and listen */
 473 static int smc_check_rdma(struct smc_sock *smc, struct smc_ib_device **ibdev,
 474                           u8 *ibport, unsigned short vlan_id, u8 gid[])
 475 {
 476         int reason_code = 0;
 477
 478         /* PNET table look up: search active ib_device and port
 479          * within same PNETID that also contains the ethernet device
 480          * used for the internal TCP socket
 481          */
 482         smc_pnet_find_roce_resource(smc->clcsock->sk, ibdev, ibport, vlan_id,
 483                                     gid);
 484         if (!(*ibdev))
 485                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 486
 487         return reason_code;
 488 }
 489
 490 /* check if there is an ISM device available for this connection. */
 491 /* called for connect and listen */
 492 static int smc_check_ism(struct smc_sock *smc, struct smcd_dev **ismdev)
 493 {
 494         /* Find ISM device with same PNETID as connecting interface  */
 495         smc_pnet_find_ism_resource(smc->clcsock->sk, ismdev);
 496         if (!(*ismdev))
 497                 return SMC_CLC_DECL_CNFERR; /* configuration error */
 498         return 0;
 499 }
 500
 501 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
 502 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
 503                                       struct smcd_dev *ismdev,
 504                                       unsigned short vlan_id)
 505 {
 506         if (vlan_id && smc_ism_get_vlan(ismdev, vlan_id))
 507                 return SMC_CLC_DECL_CNFERR;
 508         return 0;
 509 }
 510
 511 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
 512  * used, the VLAN ID will be registered again during the connection setup.
 513  */
 514 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc, bool is_smcd,
 515                                         struct smcd_dev *ismdev,
 516                                         unsigned short vlan_id)
 517 {
 518         if (!is_smcd)
 519                 return 0;
 520         if (vlan_id && smc_ism_put_vlan(ismdev, vlan_id))
 521                 return SMC_CLC_DECL_CNFERR;
 522         return 0;
 523 }
 524
 525 /* CLC handshake during connect */
 526 static int smc_connect_clc(struct smc_sock *smc, int smc_type,
 527                            struct smc_clc_msg_accept_confirm *aclc,
 528                            struct smc_ib_device *ibdev, u8 ibport,
 529                            u8 gid[], struct smcd_dev *ismdev)
 530 {
 531         int rc = 0;
 532
 533         /* do inband token exchange */
 534         rc = smc_clc_send_proposal(smc, smc_type, ibdev, ibport, gid, ismdev);
 535         if (rc)
 536                 return rc;
 537         /* receive SMC Accept CLC message */
 538         return smc_clc_wait_msg(smc, aclc, sizeof(*aclc), SMC_CLC_ACCEPT);
 539 }
 540
 541 /* setup for RDMA connection of client */
 542 static int smc_connect_rdma(struct smc_sock *smc,
 543                             struct smc_clc_msg_accept_confirm *aclc,
 544                             struct smc_ib_device *ibdev, u8 ibport)
 545 {
 546         int local_contact = SMC_FIRST_CONTACT;
 547         struct smc_link *link;
 548         int reason_code = 0;
 549
 550         mutex_lock(&smc_create_lgr_pending);
 551         local_contact = smc_conn_create(smc, false, aclc->hdr.flag, ibdev,
 552                                         ibport, &aclc->lcl, NULL, 0);
 553         if (local_contact < 0) {
 554                 if (local_contact == -ENOMEM)
 555                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
 556                 else if (local_contact == -ENOLINK)
 557                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
 558                 else
 559                         reason_code = SMC_CLC_DECL_INTERR; /* other error */
 560                 return smc_connect_abort(smc, reason_code, 0);
 561         }
 562         link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
 563
 564         smc_conn_save_peer_info(smc, aclc);
 565
 566         /* create send buffer and rmb */
 567         if (smc_buf_create(smc, false))
 568                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, local_contact);
 569
 570         if (local_contact == SMC_FIRST_CONTACT)
 571                 smc_link_save_peer_info(link, aclc);
 572
 573         if (smc_rmb_rtoken_handling(&smc->conn, aclc))
 574                 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RTOK,
 575                                          local_contact);
 576
 577         smc_close_init(smc);
 578         smc_rx_init(smc);
 579
 580         if (local_contact == SMC_FIRST_CONTACT) {
 581                 if (smc_ib_ready_link(link))
 582                         return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RDYLNK,
 583                                                  local_contact);
 584         } else {
 585                 if (!smc->conn.rmb_desc->reused &&
 586                     smc_reg_rmb(link, smc->conn.rmb_desc, true))
 587                         return smc_connect_abort(smc, SMC_CLC_DECL_ERR_REGRMB,
 588                                                  local_contact);
 589         }
 590         smc_rmb_sync_sg_for_device(&smc->conn);
 591
 592         reason_code = smc_clc_send_confirm(smc);
 593         if (reason_code)
 594                 return smc_connect_abort(smc, reason_code, local_contact);
 595
 596         smc_tx_init(smc);
 597
 598         if (local_contact == SMC_FIRST_CONTACT) {
 599                 /* QP confirmation over RoCE fabric */
 600                 reason_code = smc_clnt_conf_first_link(smc);
 601                 if (reason_code)
 602                         return smc_connect_abort(smc, reason_code,
 603                                                  local_contact);
 604         }
 605         mutex_unlock(&smc_create_lgr_pending);
 606
 607         smc_copy_sock_settings_to_clc(smc);
 608         if (smc->sk.sk_state == SMC_INIT)
 609                 smc->sk.sk_state = SMC_ACTIVE;
 610
 611         return 0;
 612 }
 613
 614 /* setup for ISM connection of client */
 615 static int smc_connect_ism(struct smc_sock *smc,
 616                            struct smc_clc_msg_accept_confirm *aclc,
 617                            struct smcd_dev *ismdev)
 618 {
 619         int local_contact = SMC_FIRST_CONTACT;
 620         int rc = 0;
 621
 622         mutex_lock(&smc_create_lgr_pending);
 623         local_contact = smc_conn_create(smc, true, aclc->hdr.flag, NULL, 0,
 624                                         NULL, ismdev, aclc->gid);
 625         if (local_contact < 0)
 626                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, 0);
 627
 628         /* Create send and receive buffers */
 629         if (smc_buf_create(smc, true))
 630                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, local_contact);
 631
 632         smc_conn_save_peer_info(smc, aclc);
 633         smc_close_init(smc);
 634         smc_rx_init(smc);
 635         smc_tx_init(smc);
 636
 637         rc = smc_clc_send_confirm(smc);
 638         if (rc)
 639                 return smc_connect_abort(smc, rc, local_contact);
 640         mutex_unlock(&smc_create_lgr_pending);
 641
 642         smc_copy_sock_settings_to_clc(smc);
 643         if (smc->sk.sk_state == SMC_INIT)
 644                 smc->sk.sk_state = SMC_ACTIVE;
 645
 646         return 0;
 647 }
 648
 649 /* perform steps before actually connecting */
 650 static int __smc_connect(struct smc_sock *smc)
 651 {
 652         bool ism_supported = false, rdma_supported = false;
 653         struct smc_clc_msg_accept_confirm aclc;
 654         struct smc_ib_device *ibdev;
 655         struct smcd_dev *ismdev;
 656         u8 gid[SMC_GID_SIZE];
 657         unsigned short vlan;
 658         int smc_type;
 659         int rc = 0;
 660         u8 ibport;
 661
 662         sock_hold(&smc->sk); /* sock put in passive closing */
 663
 664         if (smc->use_fallback)
 665                 return smc_connect_fallback(smc, smc->fallback_rsn);
 666
 667         /* if peer has not signalled SMC-capability, fall back */
 668         if (!tcp_sk(smc->clcsock->sk)->syn_smc)
 669                 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
 670
 671         /* IPSec connections opt out of SMC-R optimizations */
 672         if (using_ipsec(smc))
 673                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
 674
 675         /* check for VLAN ID */
 676         if (smc_vlan_by_tcpsk(smc->clcsock, &vlan))
 677                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_CNFERR);
 678
 679         /* check if there is an ism device available */
 680         if (!smc_check_ism(smc, &ismdev) &&
 681             !smc_connect_ism_vlan_setup(smc, ismdev, vlan)) {
 682                 /* ISM is supported for this connection */
 683                 ism_supported = true;
 684                 smc_type = SMC_TYPE_D;
 685         }
 686
 687         /* check if there is a rdma device available */
 688         if (!smc_check_rdma(smc, &ibdev, &ibport, vlan, gid)) {
 689                 /* RDMA is supported for this connection */
 690                 rdma_supported = true;
 691                 if (ism_supported)
 692                         smc_type = SMC_TYPE_B; /* both */
 693                 else
 694                         smc_type = SMC_TYPE_R; /* only RDMA */
 695         }
 696
 697         /* if neither ISM nor RDMA are supported, fallback */
 698         if (!rdma_supported && !ism_supported)
 699                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_NOSMCDEV);
 700
 701         /* perform CLC handshake */
 702         rc = smc_connect_clc(smc, smc_type, &aclc, ibdev, ibport, gid, ismdev);
 703         if (rc) {
 704                 smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
 705                 return smc_connect_decline_fallback(smc, rc);
 706         }
 707
 708         /* depending on previous steps, connect using rdma or ism */
 709         if (rdma_supported && aclc.hdr.path == SMC_TYPE_R)
 710                 rc = smc_connect_rdma(smc, &aclc, ibdev, ibport);
 711         else if (ism_supported && aclc.hdr.path == SMC_TYPE_D)
 712                 rc = smc_connect_ism(smc, &aclc, ismdev);
 713         else
 714                 rc = SMC_CLC_DECL_MODEUNSUPP;
 715         if (rc) {
 716                 smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
 717                 return smc_connect_decline_fallback(smc, rc);
 718         }
 719
 720         smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
 721         return 0;
 722 }
 723
 724 static void smc_connect_work(struct work_struct *work)
 725 {
 726         struct smc_sock *smc = container_of(work, struct smc_sock,
 727                                             connect_work);
 728         int rc;
 729
 730         lock_sock(&smc->sk);
 731         rc = kernel_connect(smc->clcsock, &smc->connect_info->addr,
 732                             smc->connect_info->alen, smc->connect_info->flags);
 733         if (smc->clcsock->sk->sk_err) {
 734                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
 735                 goto out;
 736         }
 737         if (rc < 0) {
 738                 smc->sk.sk_err = -rc;
 739                 goto out;
 740         }
 741
 742         rc = __smc_connect(smc);
 743         if (rc < 0)
 744                 smc->sk.sk_err = -rc;
 745
 746 out:
 747         if (smc->sk.sk_err)
 748                 smc->sk.sk_state_change(&smc->sk);
 749         else
 750                 smc->sk.sk_write_space(&smc->sk);
 751         kfree(smc->connect_info);
 752         smc->connect_info = NULL;
 753         release_sock(&smc->sk);
 754 }
 755
 756 static int smc_connect(struct socket *sock, struct sockaddr *addr,
 757                        int alen, int flags)
 758 {
 759         struct sock *sk = sock->sk;
 760         struct smc_sock *smc;
 761         int rc = -EINVAL;
 762
 763         smc = smc_sk(sk);
 764
 765         /* separate smc parameter checking to be safe */
 766         if (alen < sizeof(addr->sa_family))
 767                 goto out_err;
 768         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
 769                 goto out_err;
 770
 771         lock_sock(sk);
 772         switch (sk->sk_state) {
 773         default:
 774                 goto out;
 775         case SMC_ACTIVE:
 776                 rc = -EISCONN;
 777                 goto out;
 778         case SMC_INIT:
 779                 rc = 0;
 780                 break;
 781         }
 782
 783         smc_copy_sock_settings_to_clc(smc);
 784         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
 785         if (flags & O_NONBLOCK) {
 786                 if (smc->connect_info) {
 787                         rc = -EALREADY;
 788                         goto out;
 789                 }
 790                 smc->connect_info = kzalloc(alen + 2 * sizeof(int), GFP_KERNEL);
 791                 if (!smc->connect_info) {
 792                         rc = -ENOMEM;
 793                         goto out;
 794                 }
 795                 smc->connect_info->alen = alen;
 796                 smc->connect_info->flags = flags ^ O_NONBLOCK;
 797                 memcpy(&smc->connect_info->addr, addr, alen);
 798                 schedule_work(&smc->connect_work);
 799                 rc = -EINPROGRESS;
 800         } else {
 801                 rc = kernel_connect(smc->clcsock, addr, alen, flags);
 802                 if (rc)
 803                         goto out;
 804
 805                 rc = __smc_connect(smc);
 806                 if (rc < 0)
 807                         goto out;
 808                 else
 809                         rc = 0; /* success cases including fallback */
 810         }
 811
 812 out:
 813         release_sock(sk);
 814 out_err:
 815         return rc;
 816 }
 817
 818 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 819 {
 820         struct socket *new_clcsock = NULL;
 821         struct sock *lsk = &lsmc->sk;
 822         struct sock *new_sk;
 823         int rc;
 824
 825         release_sock(lsk);
 826         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
 827         if (!new_sk) {
 828                 rc = -ENOMEM;
 829                 lsk->sk_err = ENOMEM;
 830                 *new_smc = NULL;
 831                 lock_sock(lsk);
 832                 goto out;
 833         }
 834         *new_smc = smc_sk(new_sk);
 835
 836         rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
 837         lock_sock(lsk);
 838         if  (rc < 0)
 839                 lsk->sk_err = -rc;
 840         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
 841                 if (new_clcsock)
 842                         sock_release(new_clcsock);
 843                 new_sk->sk_state = SMC_CLOSED;
 844                 sock_set_flag(new_sk, SOCK_DEAD);
 845                 new_sk->sk_prot->unhash(new_sk);
 846                 sock_put(new_sk); /* final */
 847                 *new_smc = NULL;
 848                 goto out;
 849         }
 850
 851         (*new_smc)->clcsock = new_clcsock;
 852 out:
 853         return rc;
 854 }
 855
 856 /* add a just created sock to the accept queue of the listen sock as
 857  * candidate for a following socket accept call from user space
 858  */
 859 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
 860 {
 861         struct smc_sock *par = smc_sk(parent);
 862
 863         sock_hold(sk); /* sock_put in smc_accept_unlink () */
 864         spin_lock(&par->accept_q_lock);
 865         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
 866         spin_unlock(&par->accept_q_lock);
 867         sk_acceptq_added(parent);
 868 }
 869
 870 /* remove a socket from the accept queue of its parental listening socket */
 871 static void smc_accept_unlink(struct sock *sk)
 872 {
 873         struct smc_sock *par = smc_sk(sk)->listen_smc;
 874
 875         spin_lock(&par->accept_q_lock);
 876         list_del_init(&smc_sk(sk)->accept_q);
 877         spin_unlock(&par->accept_q_lock);
 878         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
 879         sock_put(sk); /* sock_hold in smc_accept_enqueue */
 880 }
 881
 882 /* remove a sock from the accept queue to bind it to a new socket created
 883  * for a socket accept call from user space
 884  */
 885 struct sock *smc_accept_dequeue(struct sock *parent,
 886                                 struct socket *new_sock)
 887 {
 888         struct smc_sock *isk, *n;
 889         struct sock *new_sk;
 890
 891         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
 892                 new_sk = (struct sock *)isk;
 893
 894                 smc_accept_unlink(new_sk);
 895                 if (new_sk->sk_state == SMC_CLOSED) {
 896                         if (isk->clcsock) {
 897                                 sock_release(isk->clcsock);
 898                                 isk->clcsock = NULL;
 899                         }
 900                         new_sk->sk_prot->unhash(new_sk);
 901                         sock_put(new_sk); /* final */
 902                         continue;
 903                 }
 904                 if (new_sock)
 905                         sock_graft(new_sk, new_sock);
 906                 return new_sk;
 907         }
 908         return NULL;
 909 }
 910
 911 /* clean up for a created but never accepted sock */
 912 void smc_close_non_accepted(struct sock *sk)
 913 {
 914         struct smc_sock *smc = smc_sk(sk);
 915
 916         lock_sock(sk);
 917         if (!sk->sk_lingertime)
 918                 /* wait for peer closing */
 919                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
 920         if (!smc->use_fallback) {
 921                 smc_close_active(smc);
 922                 sock_set_flag(sk, SOCK_DEAD);
 923                 sk->sk_shutdown |= SHUTDOWN_MASK;
 924         }
 925         if (smc->clcsock) {
 926                 struct socket *tcp;
 927
 928                 tcp = smc->clcsock;
 929                 smc->clcsock = NULL;
 930                 sock_release(tcp);
 931         }
 932         if (smc->use_fallback) {
 933                 sock_put(sk); /* passive closing */
 934                 sk->sk_state = SMC_CLOSED;
 935         } else {
 936                 if (sk->sk_state == SMC_CLOSED)
 937                         smc_conn_free(&smc->conn);
 938         }
 939         release_sock(sk);
 940         sk->sk_prot->unhash(sk);
 941         sock_put(sk); /* final sock_put */
 942 }
 943
 944 static int smc_serv_conf_first_link(struct smc_sock *smc)
 945 {
 946         struct net *net = sock_net(smc->clcsock->sk);
 947         struct smc_link_group *lgr = smc->conn.lgr;
 948         struct smc_link *link;
 949         int rest;
 950         int rc;
 951
 952         link = &lgr->lnk[SMC_SINGLE_LINK];
 953
 954         if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
 955                 return SMC_CLC_DECL_ERR_REGRMB;
 956
 957         /* send CONFIRM LINK request to client over the RoCE fabric */
 958         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
 959         if (rc < 0)
 960                 return SMC_CLC_DECL_TIMEOUT_CL;
 961
 962         /* receive CONFIRM LINK response from client over the RoCE fabric */
 963         rest = wait_for_completion_interruptible_timeout(
 964                 &link->llc_confirm_resp,
 965                 SMC_LLC_WAIT_FIRST_TIME);
 966         if (rest <= 0) {
 967                 struct smc_clc_msg_decline dclc;
 968
 969                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 970                                       SMC_CLC_DECLINE);
 971                 return rc;
 972         }
 973
 974         if (link->llc_confirm_resp_rc)
 975                 return SMC_CLC_DECL_RMBE_EC;
 976
 977         /* send ADD LINK request to client over the RoCE fabric */
 978         rc = smc_llc_send_add_link(link,
 979                                    link->smcibdev->mac[link->ibport - 1],
 980                                    link->gid, SMC_LLC_REQ);
 981         if (rc < 0)
 982                 return SMC_CLC_DECL_TIMEOUT_AL;
 983
 984         /* receive ADD LINK response from client over the RoCE fabric */
 985         rest = wait_for_completion_interruptible_timeout(&link->llc_add_resp,
 986                                                          SMC_LLC_WAIT_TIME);
 987         if (rest <= 0) {
 988                 struct smc_clc_msg_decline dclc;
 989
 990                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 991                                       SMC_CLC_DECLINE);
 992                 return rc;
 993         }
 994
 995         smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
 996
 997         return 0;
 998 }
 999
1000 /* listen worker: finish */
1001 static void smc_listen_out(struct smc_sock *new_smc)
1002 {
1003         struct smc_sock *lsmc = new_smc->listen_smc;
1004         struct sock *newsmcsk = &new_smc->sk;
1005
1006         lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1007         if (lsmc->sk.sk_state == SMC_LISTEN) {
1008                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1009         } else { /* no longer listening */
1010                 smc_close_non_accepted(newsmcsk);
1011         }
1012         release_sock(&lsmc->sk);
1013
1014         /* Wake up accept */
1015         lsmc->sk.sk_data_ready(&lsmc->sk);
1016         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1017 }
1018
1019 /* listen worker: finish in state connected */
1020 static void smc_listen_out_connected(struct smc_sock *new_smc)
1021 {
1022         struct sock *newsmcsk = &new_smc->sk;
1023
1024         sk_refcnt_debug_inc(newsmcsk);
1025         if (newsmcsk->sk_state == SMC_INIT)
1026                 newsmcsk->sk_state = SMC_ACTIVE;
1027
1028         smc_listen_out(new_smc);
1029 }
1030
1031 /* listen worker: finish in error state */
1032 static void smc_listen_out_err(struct smc_sock *new_smc)
1033 {
1034         struct sock *newsmcsk = &new_smc->sk;
1035
1036         if (newsmcsk->sk_state == SMC_INIT)
1037                 sock_put(&new_smc->sk); /* passive closing */
1038         newsmcsk->sk_state = SMC_CLOSED;
1039         smc_conn_free(&new_smc->conn);
1040
1041         smc_listen_out(new_smc);
1042 }
1043
1044 /* listen worker: decline and fall back if possible */
1045 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1046                                int local_contact)
1047 {
1048         /* RDMA setup failed, switch back to TCP */
1049         if (local_contact == SMC_FIRST_CONTACT)
1050                 smc_lgr_forget(new_smc->conn.lgr);
1051         if (reason_code < 0) { /* error, no fallback possible */
1052                 smc_listen_out_err(new_smc);
1053                 return;
1054         }
1055         smc_conn_free(&new_smc->conn);
1056         new_smc->use_fallback = true;
1057         new_smc->fallback_rsn = reason_code;
1058         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1059                 if (smc_clc_send_decline(new_smc, reason_code) < 0) {
1060                         smc_listen_out_err(new_smc);
1061                         return;
1062                 }
1063         }
1064         smc_listen_out_connected(new_smc);
1065 }
1066
1067 /* listen worker: check prefixes */
1068 static int smc_listen_rdma_check(struct smc_sock *new_smc,
1069                                  struct smc_clc_msg_proposal *pclc)
1070 {
1071         struct smc_clc_msg_proposal_prefix *pclc_prfx;
1072         struct socket *newclcsock = new_smc->clcsock;
1073
1074         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1075         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1076                 return SMC_CLC_DECL_CNFERR;
1077
1078         return 0;
1079 }
1080
1081 /* listen worker: initialize connection and buffers */
1082 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1083                                 struct smc_clc_msg_proposal *pclc,
1084                                 struct smc_ib_device *ibdev, u8 ibport,
1085                                 int *local_contact)
1086 {
1087         /* allocate connection / link group */
1088         *local_contact = smc_conn_create(new_smc, false, 0, ibdev, ibport,
1089                                          &pclc->lcl, NULL, 0);
1090         if (*local_contact < 0) {
1091                 if (*local_contact == -ENOMEM)
1092                         return SMC_CLC_DECL_MEM;/* insufficient memory*/
1093                 return SMC_CLC_DECL_INTERR; /* other error */
1094         }
1095
1096         /* create send buffer and rmb */
1097         if (smc_buf_create(new_smc, false))
1098                 return SMC_CLC_DECL_MEM;
1099
1100         return 0;
1101 }
1102
1103 /* listen worker: initialize connection and buffers for SMC-D */
1104 static int smc_listen_ism_init(struct smc_sock *new_smc,
1105                                struct smc_clc_msg_proposal *pclc,
1106                                struct smcd_dev *ismdev,
1107                                int *local_contact)
1108 {
1109         struct smc_clc_msg_smcd *pclc_smcd;
1110
1111         pclc_smcd = smc_get_clc_msg_smcd(pclc);
1112         *local_contact = smc_conn_create(new_smc, true, 0, NULL, 0, NULL,
1113                                          ismdev, pclc_smcd->gid);
1114         if (*local_contact < 0) {
1115                 if (*local_contact == -ENOMEM)
1116                         return SMC_CLC_DECL_MEM;/* insufficient memory*/
1117                 return SMC_CLC_DECL_INTERR; /* other error */
1118         }
1119
1120         /* Check if peer can be reached via ISM device */
1121         if (smc_ism_cantalk(new_smc->conn.lgr->peer_gid,
1122                             new_smc->conn.lgr->vlan_id,
1123                             new_smc->conn.lgr->smcd)) {
1124                 if (*local_contact == SMC_FIRST_CONTACT)
1125                         smc_lgr_forget(new_smc->conn.lgr);
1126                 smc_conn_free(&new_smc->conn);
1127                 return SMC_CLC_DECL_CNFERR;
1128         }
1129
1130         /* Create send and receive buffers */
1131         if (smc_buf_create(new_smc, true)) {
1132                 if (*local_contact == SMC_FIRST_CONTACT)
1133                         smc_lgr_forget(new_smc->conn.lgr);
1134                 smc_conn_free(&new_smc->conn);
1135                 return SMC_CLC_DECL_MEM;
1136         }
1137
1138         return 0;
1139 }
1140
1141 /* listen worker: register buffers */
1142 static int smc_listen_rdma_reg(struct smc_sock *new_smc, int local_contact)
1143 {
1144         struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
1145
1146         if (local_contact != SMC_FIRST_CONTACT) {
1147                 if (!new_smc->conn.rmb_desc->reused) {
1148                         if (smc_reg_rmb(link, new_smc->conn.rmb_desc, true))
1149                                 return SMC_CLC_DECL_ERR_REGRMB;
1150                 }
1151         }
1152         smc_rmb_sync_sg_for_device(&new_smc->conn);
1153
1154         return 0;
1155 }
1156
1157 /* listen worker: finish RDMA setup */
1158 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1159                                   struct smc_clc_msg_accept_confirm *cclc,
1160                                   int local_contact)
1161 {
1162         struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
1163         int reason_code = 0;
1164
1165         if (local_contact == SMC_FIRST_CONTACT)
1166                 smc_link_save_peer_info(link, cclc);
1167
1168         if (smc_rmb_rtoken_handling(&new_smc->conn, cclc)) {
1169                 reason_code = SMC_CLC_DECL_ERR_RTOK;
1170                 goto decline;
1171         }
1172
1173         if (local_contact == SMC_FIRST_CONTACT) {
1174                 if (smc_ib_ready_link(link)) {
1175                         reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1176                         goto decline;
1177                 }
1178                 /* QP confirmation over RoCE fabric */
1179                 reason_code = smc_serv_conf_first_link(new_smc);
1180                 if (reason_code)
1181                         goto decline;
1182         }
1183         return 0;
1184
1185 decline:
1186         mutex_unlock(&smc_create_lgr_pending);
1187         smc_listen_decline(new_smc, reason_code, local_contact);
1188         return reason_code;
1189 }
1190
1191 /* setup for RDMA connection of server */
1192 static void smc_listen_work(struct work_struct *work)
1193 {
1194         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1195                                                 smc_listen_work);
1196         struct socket *newclcsock = new_smc->clcsock;
1197         struct smc_clc_msg_accept_confirm cclc;
1198         struct smc_clc_msg_proposal *pclc;
1199         struct smc_ib_device *ibdev;
1200         bool ism_supported = false;
1201         struct smcd_dev *ismdev;
1202         u8 buf[SMC_CLC_MAX_LEN];
1203         int local_contact = 0;
1204         unsigned short vlan;
1205         int reason_code = 0;
1206         int rc = 0;
1207         u8 ibport;
1208
1209         if (new_smc->use_fallback) {
1210                 smc_listen_out_connected(new_smc);
1211                 return;
1212         }
1213
1214         /* check if peer is smc capable */
1215         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1216                 new_smc->use_fallback = true;
1217                 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1218                 smc_listen_out_connected(new_smc);
1219                 return;
1220         }
1221
1222         /* do inband token exchange -
1223          * wait for and receive SMC Proposal CLC message
1224          */
1225         pclc = (struct smc_clc_msg_proposal *)&buf;
1226         reason_code = smc_clc_wait_msg(new_smc, pclc, SMC_CLC_MAX_LEN,
1227                                        SMC_CLC_PROPOSAL);
1228         if (reason_code) {
1229                 smc_listen_decline(new_smc, reason_code, 0);
1230                 return;
1231         }
1232
1233         /* IPSec connections opt out of SMC-R optimizations */
1234         if (using_ipsec(new_smc)) {
1235                 smc_listen_decline(new_smc, SMC_CLC_DECL_IPSEC, 0);
1236                 return;
1237         }
1238
1239         mutex_lock(&smc_create_lgr_pending);
1240         smc_close_init(new_smc);
1241         smc_rx_init(new_smc);
1242         smc_tx_init(new_smc);
1243
1244         /* check if ISM is available */
1245         if ((pclc->hdr.path == SMC_TYPE_D || pclc->hdr.path == SMC_TYPE_B) &&
1246             !smc_check_ism(new_smc, &ismdev) &&
1247             !smc_listen_ism_init(new_smc, pclc, ismdev, &local_contact)) {
1248                 ism_supported = true;
1249         }
1250
1251         /* check if RDMA is available */
1252         if (!ism_supported &&
1253             ((pclc->hdr.path != SMC_TYPE_R && pclc->hdr.path != SMC_TYPE_B) ||
1254              smc_vlan_by_tcpsk(new_smc->clcsock, &vlan) ||
1255              smc_check_rdma(new_smc, &ibdev, &ibport, vlan, NULL) ||
1256              smc_listen_rdma_check(new_smc, pclc) ||
1257              smc_listen_rdma_init(new_smc, pclc, ibdev, ibport,
1258                                   &local_contact) ||
1259              smc_listen_rdma_reg(new_smc, local_contact))) {
1260                 /* SMC not supported, decline */
1261                 mutex_unlock(&smc_create_lgr_pending);
1262                 smc_listen_decline(new_smc, SMC_CLC_DECL_MODEUNSUPP,
1263                                    local_contact);
1264                 return;
1265         }
1266
1267         /* send SMC Accept CLC message */
1268         rc = smc_clc_send_accept(new_smc, local_contact);
1269         if (rc) {
1270                 mutex_unlock(&smc_create_lgr_pending);
1271                 smc_listen_decline(new_smc, rc, local_contact);
1272                 return;
1273         }
1274
1275         /* receive SMC Confirm CLC message */
1276         reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
1277                                        SMC_CLC_CONFIRM);
1278         if (reason_code) {
1279                 mutex_unlock(&smc_create_lgr_pending);
1280                 smc_listen_decline(new_smc, reason_code, local_contact);
1281                 return;
1282         }
1283
1284         /* finish worker */
1285         if (!ism_supported) {
1286                 if (smc_listen_rdma_finish(new_smc, &cclc, local_contact))
1287                         return;
1288         }
1289         smc_conn_save_peer_info(new_smc, &cclc);
1290         mutex_unlock(&smc_create_lgr_pending);
1291         smc_listen_out_connected(new_smc);
1292 }
1293
1294 static void smc_tcp_listen_work(struct work_struct *work)
1295 {
1296         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1297                                              tcp_listen_work);
1298         struct sock *lsk = &lsmc->sk;
1299         struct smc_sock *new_smc;
1300         int rc = 0;
1301
1302         lock_sock(lsk);
1303         while (lsk->sk_state == SMC_LISTEN) {
1304                 rc = smc_clcsock_accept(lsmc, &new_smc);
1305                 if (rc)
1306                         goto out;
1307                 if (!new_smc)
1308                         continue;
1309
1310                 new_smc->listen_smc = lsmc;
1311                 new_smc->use_fallback = lsmc->use_fallback;
1312                 new_smc->fallback_rsn = lsmc->fallback_rsn;
1313                 sock_hold(lsk); /* sock_put in smc_listen_work */
1314                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1315                 smc_copy_sock_settings_to_smc(new_smc);
1316                 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1317                 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1318                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1319                 if (!schedule_work(&new_smc->smc_listen_work))
1320                         sock_put(&new_smc->sk);
1321         }
1322
1323 out:
1324         release_sock(lsk);
1325         sock_put(&lsmc->sk); /* sock_hold in smc_listen */
1326 }
1327
1328 static int smc_listen(struct socket *sock, int backlog)
1329 {
1330         struct sock *sk = sock->sk;
1331         struct smc_sock *smc;
1332         int rc;
1333
1334         smc = smc_sk(sk);
1335         lock_sock(sk);
1336
1337         rc = -EINVAL;
1338         if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
1339                 goto out;
1340
1341         rc = 0;
1342         if (sk->sk_state == SMC_LISTEN) {
1343                 sk->sk_max_ack_backlog = backlog;
1344                 goto out;
1345         }
1346         /* some socket options are handled in core, so we could not apply
1347          * them to the clc socket -- copy smc socket options to clc socket
1348          */
1349         smc_copy_sock_settings_to_clc(smc);
1350         if (!smc->use_fallback)
1351                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1352
1353         rc = kernel_listen(smc->clcsock, backlog);
1354         if (rc)
1355                 goto out;
1356         sk->sk_max_ack_backlog = backlog;
1357         sk->sk_ack_backlog = 0;
1358         sk->sk_state = SMC_LISTEN;
1359         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
1360         sock_hold(sk); /* sock_hold in tcp_listen_worker */
1361         if (!schedule_work(&smc->tcp_listen_work))
1362                 sock_put(sk);
1363
1364 out:
1365         release_sock(sk);
1366         return rc;
1367 }
1368
1369 static int smc_accept(struct socket *sock, struct socket *new_sock,
1370                       int flags, bool kern)
1371 {
1372         struct sock *sk = sock->sk, *nsk;
1373         DECLARE_WAITQUEUE(wait, current);
1374         struct smc_sock *lsmc;
1375         long timeo;
1376         int rc = 0;
1377
1378         lsmc = smc_sk(sk);
1379         sock_hold(sk); /* sock_put below */
1380         lock_sock(sk);
1381
1382         if (lsmc->sk.sk_state != SMC_LISTEN) {
1383                 rc = -EINVAL;
1384                 release_sock(sk);
1385                 goto out;
1386         }
1387
1388         /* Wait for an incoming connection */
1389         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1390         add_wait_queue_exclusive(sk_sleep(sk), &wait);
1391         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1392                 set_current_state(TASK_INTERRUPTIBLE);
1393                 if (!timeo) {
1394                         rc = -EAGAIN;
1395                         break;
1396                 }
1397                 release_sock(sk);
1398                 timeo = schedule_timeout(timeo);
1399                 /* wakeup by sk_data_ready in smc_listen_work() */
1400                 sched_annotate_sleep();
1401                 lock_sock(sk);
1402                 if (signal_pending(current)) {
1403                         rc = sock_intr_errno(timeo);
1404                         break;
1405                 }
1406         }
1407         set_current_state(TASK_RUNNING);
1408         remove_wait_queue(sk_sleep(sk), &wait);
1409
1410         if (!rc)
1411                 rc = sock_error(nsk);
1412         release_sock(sk);
1413         if (rc)
1414                 goto out;
1415
1416         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1417                 /* wait till data arrives on the socket */
1418                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1419                                                                 MSEC_PER_SEC);
1420                 if (smc_sk(nsk)->use_fallback) {
1421                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1422
1423                         lock_sock(clcsk);
1424                         if (skb_queue_empty(&clcsk->sk_receive_queue))
1425                                 sk_wait_data(clcsk, &timeo, NULL);
1426                         release_sock(clcsk);
1427                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1428                         lock_sock(nsk);
1429                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1430                         release_sock(nsk);
1431                 }
1432         }
1433
1434 out:
1435         sock_put(sk); /* sock_hold above */
1436         return rc;
1437 }
1438
1439 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1440                        int peer)
1441 {
1442         struct smc_sock *smc;
1443
1444         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1445             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1446                 return -ENOTCONN;
1447
1448         smc = smc_sk(sock->sk);
1449
1450         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1451 }
1452
1453 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1454 {
1455         struct sock *sk = sock->sk;
1456         struct smc_sock *smc;
1457         int rc = -EPIPE;
1458
1459         smc = smc_sk(sk);
1460         lock_sock(sk);
1461         if ((sk->sk_state != SMC_ACTIVE) &&
1462             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1463             (sk->sk_state != SMC_INIT))
1464                 goto out;
1465
1466         if (msg->msg_flags & MSG_FASTOPEN) {
1467                 if (sk->sk_state == SMC_INIT) {
1468                         smc->use_fallback = true;
1469                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1470                 } else {
1471                         rc = -EINVAL;
1472                         goto out;
1473                 }
1474         }
1475
1476         if (smc->use_fallback)
1477                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1478         else
1479                 rc = smc_tx_sendmsg(smc, msg, len);
1480 out:
1481         release_sock(sk);
1482         return rc;
1483 }
1484
1485 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1486                        int flags)
1487 {
1488         struct sock *sk = sock->sk;
1489         struct smc_sock *smc;
1490         int rc = -ENOTCONN;
1491
1492         smc = smc_sk(sk);
1493         lock_sock(sk);
1494         if ((sk->sk_state == SMC_INIT) ||
1495             (sk->sk_state == SMC_LISTEN) ||
1496             (sk->sk_state == SMC_CLOSED))
1497                 goto out;
1498
1499         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1500                 rc = 0;
1501                 goto out;
1502         }
1503
1504         if (smc->use_fallback) {
1505                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1506         } else {
1507                 msg->msg_namelen = 0;
1508                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1509         }
1510
1511 out:
1512         release_sock(sk);
1513         return rc;
1514 }
1515
1516 static __poll_t smc_accept_poll(struct sock *parent)
1517 {
1518         struct smc_sock *isk = smc_sk(parent);
1519         __poll_t mask = 0;
1520
1521         spin_lock(&isk->accept_q_lock);
1522         if (!list_empty(&isk->accept_q))
1523                 mask = EPOLLIN | EPOLLRDNORM;
1524         spin_unlock(&isk->accept_q_lock);
1525
1526         return mask;
1527 }
1528
1529 static __poll_t smc_poll(struct file *file, struct socket *sock,
1530                              poll_table *wait)
1531 {
1532         struct sock *sk = sock->sk;
1533         __poll_t mask = 0;
1534         struct smc_sock *smc;
1535
1536         if (!sk)
1537                 return EPOLLNVAL;
1538
1539         smc = smc_sk(sock->sk);
1540         if (smc->use_fallback) {
1541                 /* delegate to CLC child sock */
1542                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1543                 sk->sk_err = smc->clcsock->sk->sk_err;
1544                 if (sk->sk_err)
1545                         mask |= EPOLLERR;
1546         } else {
1547                 if (sk->sk_state != SMC_CLOSED)
1548                         sock_poll_wait(file, sock, wait);
1549                 if (sk->sk_err)
1550                         mask |= EPOLLERR;
1551                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1552                     (sk->sk_state == SMC_CLOSED))
1553                         mask |= EPOLLHUP;
1554                 if (sk->sk_state == SMC_LISTEN) {
1555                         /* woken up by sk_data_ready in smc_listen_work() */
1556                         mask = smc_accept_poll(sk);
1557                 } else {
1558                         if (atomic_read(&smc->conn.sndbuf_space) ||
1559                             sk->sk_shutdown & SEND_SHUTDOWN) {
1560                                 mask |= EPOLLOUT | EPOLLWRNORM;
1561                         } else {
1562                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1563                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1564                         }
1565                         if (atomic_read(&smc->conn.bytes_to_rcv))
1566                                 mask |= EPOLLIN | EPOLLRDNORM;
1567                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1568                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1569                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
1570                                 mask |= EPOLLIN;
1571                         if (smc->conn.urg_state == SMC_URG_VALID)
1572                                 mask |= EPOLLPRI;
1573                 }
1574         }
1575
1576         return mask;
1577 }
1578
1579 static int smc_shutdown(struct socket *sock, int how)
1580 {
1581         struct sock *sk = sock->sk;
1582         struct smc_sock *smc;
1583         int rc = -EINVAL;
1584         int rc1 = 0;
1585
1586         smc = smc_sk(sk);
1587
1588         if ((how < SHUT_RD) || (how > SHUT_RDWR))
1589                 return rc;
1590
1591         lock_sock(sk);
1592
1593         rc = -ENOTCONN;
1594         if ((sk->sk_state != SMC_ACTIVE) &&
1595             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1596             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1597             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1598             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1599             (sk->sk_state != SMC_APPFINCLOSEWAIT))
1600                 goto out;
1601         if (smc->use_fallback) {
1602                 rc = kernel_sock_shutdown(smc->clcsock, how);
1603                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1604                 if (sk->sk_shutdown == SHUTDOWN_MASK)
1605                         sk->sk_state = SMC_CLOSED;
1606                 goto out;
1607         }
1608         switch (how) {
1609         case SHUT_RDWR:         /* shutdown in both directions */
1610                 rc = smc_close_active(smc);
1611                 break;
1612         case SHUT_WR:
1613                 rc = smc_close_shutdown_write(smc);
1614                 break;
1615         case SHUT_RD:
1616                 rc = 0;
1617                 /* nothing more to do because peer is not involved */
1618                 break;
1619         }
1620         if (smc->clcsock)
1621                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1622         /* map sock_shutdown_cmd constants to sk_shutdown value range */
1623         sk->sk_shutdown |= how + 1;
1624
1625 out:
1626         release_sock(sk);
1627         return rc ? rc : rc1;
1628 }
1629
1630 static int smc_setsockopt(struct socket *sock, int level, int optname,
1631                           char __user *optval, unsigned int optlen)
1632 {
1633         struct sock *sk = sock->sk;
1634         struct smc_sock *smc;
1635         int val, rc;
1636
1637         smc = smc_sk(sk);
1638
1639         /* generic setsockopts reaching us here always apply to the
1640          * CLC socket
1641          */
1642         rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1643                                            optval, optlen);
1644         if (smc->clcsock->sk->sk_err) {
1645                 sk->sk_err = smc->clcsock->sk->sk_err;
1646                 sk->sk_error_report(sk);
1647         }
1648         if (rc)
1649                 return rc;
1650
1651         if (optlen < sizeof(int))
1652                 return -EINVAL;
1653         if (get_user(val, (int __user *)optval))
1654                 return -EFAULT;
1655
1656         lock_sock(sk);
1657         switch (optname) {
1658         case TCP_ULP:
1659         case TCP_FASTOPEN:
1660         case TCP_FASTOPEN_CONNECT:
1661         case TCP_FASTOPEN_KEY:
1662         case TCP_FASTOPEN_NO_COOKIE:
1663                 /* option not supported by SMC */
1664                 if (sk->sk_state == SMC_INIT) {
1665                         smc->use_fallback = true;
1666                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1667                 } else {
1668                         if (!smc->use_fallback)
1669                                 rc = -EINVAL;
1670                 }
1671                 break;
1672         case TCP_NODELAY:
1673                 if (sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) {
1674                         if (val && !smc->use_fallback)
1675                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1676                                                  0);
1677                 }
1678                 break;
1679         case TCP_CORK:
1680                 if (sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) {
1681                         if (!val && !smc->use_fallback)
1682                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1683                                                  0);
1684                 }
1685                 break;
1686         case TCP_DEFER_ACCEPT:
1687                 smc->sockopt_defer_accept = val;
1688                 break;
1689         default:
1690                 break;
1691         }
1692         release_sock(sk);
1693
1694         return rc;
1695 }
1696
1697 static int smc_getsockopt(struct socket *sock, int level, int optname,
1698                           char __user *optval, int __user *optlen)
1699 {
1700         struct smc_sock *smc;
1701
1702         smc = smc_sk(sock->sk);
1703         /* socket options apply to the CLC socket */
1704         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1705                                              optval, optlen);
1706 }
1707
1708 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1709                      unsigned long arg)
1710 {
1711         union smc_host_cursor cons, urg;
1712         struct smc_connection *conn;
1713         struct smc_sock *smc;
1714         int answ;
1715
1716         smc = smc_sk(sock->sk);
1717         conn = &smc->conn;
1718         lock_sock(&smc->sk);
1719         if (smc->use_fallback) {
1720                 if (!smc->clcsock) {
1721                         release_sock(&smc->sk);
1722                         return -EBADF;
1723                 }
1724                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1725                 release_sock(&smc->sk);
1726                 return answ;
1727         }
1728         switch (cmd) {
1729         case SIOCINQ: /* same as FIONREAD */
1730                 if (smc->sk.sk_state == SMC_LISTEN) {
1731                         release_sock(&smc->sk);
1732                         return -EINVAL;
1733                 }
1734                 if (smc->sk.sk_state == SMC_INIT ||
1735                     smc->sk.sk_state == SMC_CLOSED)
1736                         answ = 0;
1737                 else
1738                         answ = atomic_read(&smc->conn.bytes_to_rcv);
1739                 break;
1740         case SIOCOUTQ:
1741                 /* output queue size (not send + not acked) */
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->conn.sndbuf_desc->len -
1751                                         atomic_read(&smc->conn.sndbuf_space);
1752                 break;
1753         case SIOCOUTQNSD:
1754                 /* output queue size (not send only) */
1755                 if (smc->sk.sk_state == SMC_LISTEN) {
1756                         release_sock(&smc->sk);
1757                         return -EINVAL;
1758                 }
1759                 if (smc->sk.sk_state == SMC_INIT ||
1760                     smc->sk.sk_state == SMC_CLOSED)
1761                         answ = 0;
1762                 else
1763                         answ = smc_tx_prepared_sends(&smc->conn);
1764                 break;
1765         case SIOCATMARK:
1766                 if (smc->sk.sk_state == SMC_LISTEN) {
1767                         release_sock(&smc->sk);
1768                         return -EINVAL;
1769                 }
1770                 if (smc->sk.sk_state == SMC_INIT ||
1771                     smc->sk.sk_state == SMC_CLOSED) {
1772                         answ = 0;
1773                 } else {
1774                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
1775                         smc_curs_copy(&urg, &conn->urg_curs, conn);
1776                         answ = smc_curs_diff(conn->rmb_desc->len,
1777                                              &cons, &urg) == 1;
1778                 }
1779                 break;
1780         default:
1781                 release_sock(&smc->sk);
1782                 return -ENOIOCTLCMD;
1783         }
1784         release_sock(&smc->sk);
1785
1786         return put_user(answ, (int __user *)arg);
1787 }
1788
1789 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1790                             int offset, size_t size, int flags)
1791 {
1792         struct sock *sk = sock->sk;
1793         struct smc_sock *smc;
1794         int rc = -EPIPE;
1795
1796         smc = smc_sk(sk);
1797         lock_sock(sk);
1798         if (sk->sk_state != SMC_ACTIVE) {
1799                 release_sock(sk);
1800                 goto out;
1801         }
1802         release_sock(sk);
1803         if (smc->use_fallback)
1804                 rc = kernel_sendpage(smc->clcsock, page, offset,
1805                                      size, flags);
1806         else
1807                 rc = sock_no_sendpage(sock, page, offset, size, flags);
1808
1809 out:
1810         return rc;
1811 }
1812
1813 /* Map the affected portions of the rmbe into an spd, note the number of bytes
1814  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
1815  * updates till whenever a respective page has been fully processed.
1816  * Note that subsequent recv() calls have to wait till all splice() processing
1817  * completed.
1818  */
1819 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1820                                struct pipe_inode_info *pipe, size_t len,
1821                                unsigned int flags)
1822 {
1823         struct sock *sk = sock->sk;
1824         struct smc_sock *smc;
1825         int rc = -ENOTCONN;
1826
1827         smc = smc_sk(sk);
1828         lock_sock(sk);
1829
1830         if (sk->sk_state == SMC_INIT ||
1831             sk->sk_state == SMC_LISTEN ||
1832             sk->sk_state == SMC_CLOSED)
1833                 goto out;
1834
1835         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1836                 rc = 0;
1837                 goto out;
1838         }
1839
1840         if (smc->use_fallback) {
1841                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1842                                                     pipe, len, flags);
1843         } else {
1844                 if (*ppos) {
1845                         rc = -ESPIPE;
1846                         goto out;
1847                 }
1848                 if (flags & SPLICE_F_NONBLOCK)
1849                         flags = MSG_DONTWAIT;
1850                 else
1851                         flags = 0;
1852                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
1853         }
1854 out:
1855         release_sock(sk);
1856
1857         return rc;
1858 }
1859
1860 /* must look like tcp */
1861 static const struct proto_ops smc_sock_ops = {
1862         .family         = PF_SMC,
1863         .owner          = THIS_MODULE,
1864         .release        = smc_release,
1865         .bind           = smc_bind,
1866         .connect        = smc_connect,
1867         .socketpair     = sock_no_socketpair,
1868         .accept         = smc_accept,
1869         .getname        = smc_getname,
1870         .poll           = smc_poll,
1871         .ioctl          = smc_ioctl,
1872         .listen         = smc_listen,
1873         .shutdown       = smc_shutdown,
1874         .setsockopt     = smc_setsockopt,
1875         .getsockopt     = smc_getsockopt,
1876         .sendmsg        = smc_sendmsg,
1877         .recvmsg        = smc_recvmsg,
1878         .mmap           = sock_no_mmap,
1879         .sendpage       = smc_sendpage,
1880         .splice_read    = smc_splice_read,
1881 };
1882
1883 static int smc_create(struct net *net, struct socket *sock, int protocol,
1884                       int kern)
1885 {
1886         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
1887         struct smc_sock *smc;
1888         struct sock *sk;
1889         int rc;
1890
1891         rc = -ESOCKTNOSUPPORT;
1892         if (sock->type != SOCK_STREAM)
1893                 goto out;
1894
1895         rc = -EPROTONOSUPPORT;
1896         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
1897                 goto out;
1898
1899         rc = -ENOBUFS;
1900         sock->ops = &smc_sock_ops;
1901         sk = smc_sock_alloc(net, sock, protocol);
1902         if (!sk)
1903                 goto out;
1904
1905         /* create internal TCP socket for CLC handshake and fallback */
1906         smc = smc_sk(sk);
1907         smc->use_fallback = false; /* assume rdma capability first */
1908         smc->fallback_rsn = 0;
1909         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
1910                               &smc->clcsock);
1911         if (rc) {
1912                 sk_common_release(sk);
1913                 goto out;
1914         }
1915         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1916         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1917
1918 out:
1919         return rc;
1920 }
1921
1922 static const struct net_proto_family smc_sock_family_ops = {
1923         .family = PF_SMC,
1924         .owner  = THIS_MODULE,
1925         .create = smc_create,
1926 };
1927
1928 static int __init smc_init(void)
1929 {
1930         int rc;
1931
1932         rc = smc_pnet_init();
1933         if (rc)
1934                 return rc;
1935
1936         rc = smc_llc_init();
1937         if (rc) {
1938                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
1939                 goto out_pnet;
1940         }
1941
1942         rc = smc_cdc_init();
1943         if (rc) {
1944                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
1945                 goto out_pnet;
1946         }
1947
1948         rc = proto_register(&smc_proto, 1);
1949         if (rc) {
1950                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
1951                 goto out_pnet;
1952         }
1953
1954         rc = proto_register(&smc_proto6, 1);
1955         if (rc) {
1956                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
1957                 goto out_proto;
1958         }
1959
1960         rc = sock_register(&smc_sock_family_ops);
1961         if (rc) {
1962                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
1963                 goto out_proto6;
1964         }
1965         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
1966         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
1967
1968         rc = smc_ib_register_client();
1969         if (rc) {
1970                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
1971                 goto out_sock;
1972         }
1973
1974         static_branch_enable(&tcp_have_smc);
1975         return 0;
1976
1977 out_sock:
1978         sock_unregister(PF_SMC);
1979 out_proto6:
1980         proto_unregister(&smc_proto6);
1981 out_proto:
1982         proto_unregister(&smc_proto);
1983 out_pnet:
1984         smc_pnet_exit();
1985         return rc;
1986 }
1987
1988 static void __exit smc_exit(void)
1989 {
1990         smc_core_exit();
1991         static_branch_disable(&tcp_have_smc);
1992         smc_ib_unregister_client();
1993         sock_unregister(PF_SMC);
1994         proto_unregister(&smc_proto6);
1995         proto_unregister(&smc_proto);
1996         smc_pnet_exit();
1997 }
1998
1999 module_init(smc_init);
2000 module_exit(smc_exit);
2001
2002 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2003 MODULE_DESCRIPTION("smc socket address family");
2004 MODULE_LICENSE("GPL");
2005 MODULE_ALIAS_NETPROTO(PF_SMC);