net/smc/af_smc.c

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