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_tx.h"
  49 #include "smc_rx.h"
  50 #include "smc_close.h"
  51
  52 static DEFINE_MUTEX(smc_server_lgr_pending);    /* serialize link group
  53                                                  * creation on server
  54                                                  */
  55 static DEFINE_MUTEX(smc_client_lgr_pending);    /* serialize link group
  56                                                  * creation on client
  57                                                  */
  58
  59 struct workqueue_struct *smc_hs_wq;     /* wq for handshake work */
  60 struct workqueue_struct *smc_close_wq;  /* wq for close work */
  61
  62 static void smc_tcp_listen_work(struct work_struct *);
  63 static void smc_connect_work(struct work_struct *);
  64
  65 static void smc_set_keepalive(struct sock *sk, int val)
  66 {
  67         struct smc_sock *smc = smc_sk(sk);
  68
  69         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
  70 }
  71
  72 static struct smc_hashinfo smc_v4_hashinfo = {
  73         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
  74 };
  75
  76 static struct smc_hashinfo smc_v6_hashinfo = {
  77         .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
  78 };
  79
  80 int smc_hash_sk(struct sock *sk)
  81 {
  82         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  83         struct hlist_head *head;
  84
  85         head = &h->ht;
  86
  87         write_lock_bh(&h->lock);
  88         sk_add_node(sk, head);
  89         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
  90         write_unlock_bh(&h->lock);
  91
  92         return 0;
  93 }
  94 EXPORT_SYMBOL_GPL(smc_hash_sk);
  95
  96 void smc_unhash_sk(struct sock *sk)
  97 {
  98         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  99
 100         write_lock_bh(&h->lock);
 101         if (sk_del_node_init(sk))
 102                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 103         write_unlock_bh(&h->lock);
 104 }
 105 EXPORT_SYMBOL_GPL(smc_unhash_sk);
 106
 107 struct proto smc_proto = {
 108         .name           = "SMC",
 109         .owner          = THIS_MODULE,
 110         .keepalive      = smc_set_keepalive,
 111         .hash           = smc_hash_sk,
 112         .unhash         = smc_unhash_sk,
 113         .obj_size       = sizeof(struct smc_sock),
 114         .h.smc_hash     = &smc_v4_hashinfo,
 115         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 116 };
 117 EXPORT_SYMBOL_GPL(smc_proto);
 118
 119 struct proto smc_proto6 = {
 120         .name           = "SMC6",
 121         .owner          = THIS_MODULE,
 122         .keepalive      = smc_set_keepalive,
 123         .hash           = smc_hash_sk,
 124         .unhash         = smc_unhash_sk,
 125         .obj_size       = sizeof(struct smc_sock),
 126         .h.smc_hash     = &smc_v6_hashinfo,
 127         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 128 };
 129 EXPORT_SYMBOL_GPL(smc_proto6);
 130
 131 static void smc_restore_fallback_changes(struct smc_sock *smc)
 132 {
 133         if (smc->clcsock->file) { /* non-accepted sockets have no file yet */
 134                 smc->clcsock->file->private_data = smc->sk.sk_socket;
 135                 smc->clcsock->file = NULL;
 136         }
 137 }
 138
 139 static int __smc_release(struct smc_sock *smc)
 140 {
 141         struct sock *sk = &smc->sk;
 142         int rc = 0;
 143
 144         if (!smc->use_fallback) {
 145                 rc = smc_close_active(smc);
 146                 sock_set_flag(sk, SOCK_DEAD);
 147                 sk->sk_shutdown |= SHUTDOWN_MASK;
 148         } else {
 149                 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
 150                         sock_put(sk); /* passive closing */
 151                 if (sk->sk_state == SMC_LISTEN) {
 152                         /* wake up clcsock accept */
 153                         rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
 154                 }
 155                 sk->sk_state = SMC_CLOSED;
 156                 sk->sk_state_change(sk);
 157                 smc_restore_fallback_changes(smc);
 158         }
 159
 160         sk->sk_prot->unhash(sk);
 161
 162         if (sk->sk_state == SMC_CLOSED) {
 163                 if (smc->clcsock) {
 164                         release_sock(sk);
 165                         smc_clcsock_release(smc);
 166                         lock_sock(sk);
 167                 }
 168                 if (!smc->use_fallback)
 169                         smc_conn_free(&smc->conn);
 170         }
 171
 172         return rc;
 173 }
 174
 175 static int smc_release(struct socket *sock)
 176 {
 177         struct sock *sk = sock->sk;
 178         struct smc_sock *smc;
 179         int rc = 0;
 180
 181         if (!sk)
 182                 goto out;
 183
 184         sock_hold(sk); /* sock_put below */
 185         smc = smc_sk(sk);
 186
 187         /* cleanup for a dangling non-blocking connect */
 188         if (smc->connect_nonblock && sk->sk_state == SMC_INIT)
 189                 tcp_abort(smc->clcsock->sk, ECONNABORTED);
 190         flush_work(&smc->connect_work);
 191
 192         if (sk->sk_state == SMC_LISTEN)
 193                 /* smc_close_non_accepted() is called and acquires
 194                  * sock lock for child sockets again
 195                  */
 196                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 197         else
 198                 lock_sock(sk);
 199
 200         rc = __smc_release(smc);
 201
 202         /* detach socket */
 203         sock_orphan(sk);
 204         sock->sk = NULL;
 205         release_sock(sk);
 206
 207         sock_put(sk); /* sock_hold above */
 208         sock_put(sk); /* final sock_put */
 209 out:
 210         return rc;
 211 }
 212
 213 static void smc_destruct(struct sock *sk)
 214 {
 215         if (sk->sk_state != SMC_CLOSED)
 216                 return;
 217         if (!sock_flag(sk, SOCK_DEAD))
 218                 return;
 219
 220         sk_refcnt_debug_dec(sk);
 221 }
 222
 223 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 224                                    int protocol)
 225 {
 226         struct smc_sock *smc;
 227         struct proto *prot;
 228         struct sock *sk;
 229
 230         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
 231         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
 232         if (!sk)
 233                 return NULL;
 234
 235         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 236         sk->sk_state = SMC_INIT;
 237         sk->sk_destruct = smc_destruct;
 238         sk->sk_protocol = protocol;
 239         smc = smc_sk(sk);
 240         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 241         INIT_WORK(&smc->connect_work, smc_connect_work);
 242         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 243         INIT_LIST_HEAD(&smc->accept_q);
 244         spin_lock_init(&smc->accept_q_lock);
 245         spin_lock_init(&smc->conn.send_lock);
 246         sk->sk_prot->hash(sk);
 247         sk_refcnt_debug_inc(sk);
 248         mutex_init(&smc->clcsock_release_lock);
 249
 250         return sk;
 251 }
 252
 253 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 254                     int addr_len)
 255 {
 256         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 257         struct sock *sk = sock->sk;
 258         struct smc_sock *smc;
 259         int rc;
 260
 261         smc = smc_sk(sk);
 262
 263         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 264         rc = -EINVAL;
 265         if (addr_len < sizeof(struct sockaddr_in))
 266                 goto out;
 267
 268         rc = -EAFNOSUPPORT;
 269         if (addr->sin_family != AF_INET &&
 270             addr->sin_family != AF_INET6 &&
 271             addr->sin_family != AF_UNSPEC)
 272                 goto out;
 273         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 274         if (addr->sin_family == AF_UNSPEC &&
 275             addr->sin_addr.s_addr != htonl(INADDR_ANY))
 276                 goto out;
 277
 278         lock_sock(sk);
 279
 280         /* Check if socket is already active */
 281         rc = -EINVAL;
 282         if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
 283                 goto out_rel;
 284
 285         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 286         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 287
 288 out_rel:
 289         release_sock(sk);
 290 out:
 291         return rc;
 292 }
 293
 294 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 295                                    unsigned long mask)
 296 {
 297         /* options we don't get control via setsockopt for */
 298         nsk->sk_type = osk->sk_type;
 299         nsk->sk_sndbuf = osk->sk_sndbuf;
 300         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 301         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 302         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 303         nsk->sk_mark = osk->sk_mark;
 304         nsk->sk_priority = osk->sk_priority;
 305         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 306         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 307         nsk->sk_err = osk->sk_err;
 308
 309         nsk->sk_flags &= ~mask;
 310         nsk->sk_flags |= osk->sk_flags & mask;
 311 }
 312
 313 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 314                              (1UL << SOCK_KEEPOPEN) | \
 315                              (1UL << SOCK_LINGER) | \
 316                              (1UL << SOCK_BROADCAST) | \
 317                              (1UL << SOCK_TIMESTAMP) | \
 318                              (1UL << SOCK_DBG) | \
 319                              (1UL << SOCK_RCVTSTAMP) | \
 320                              (1UL << SOCK_RCVTSTAMPNS) | \
 321                              (1UL << SOCK_LOCALROUTE) | \
 322                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 323                              (1UL << SOCK_RXQ_OVFL) | \
 324                              (1UL << SOCK_WIFI_STATUS) | \
 325                              (1UL << SOCK_NOFCS) | \
 326                              (1UL << SOCK_FILTER_LOCKED) | \
 327                              (1UL << SOCK_TSTAMP_NEW))
 328 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 329  * clc socket (since smc is not called for these options from net/core)
 330  */
 331 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 332 {
 333         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 334 }
 335
 336 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 337                              (1UL << SOCK_KEEPOPEN) | \
 338                              (1UL << SOCK_LINGER) | \
 339                              (1UL << SOCK_DBG))
 340 /* copy only settings and flags relevant for smc from clc to smc socket */
 341 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 342 {
 343         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 344 }
 345
 346 /* register the new rmb on all links */
 347 static int smcr_lgr_reg_rmbs(struct smc_link *link,
 348                              struct smc_buf_desc *rmb_desc)
 349 {
 350         struct smc_link_group *lgr = link->lgr;
 351         int i, rc = 0;
 352
 353         rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
 354         if (rc)
 355                 return rc;
 356         /* protect against parallel smc_llc_cli_rkey_exchange() and
 357          * parallel smcr_link_reg_rmb()
 358          */
 359         mutex_lock(&lgr->llc_conf_mutex);
 360         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 361                 if (!smc_link_active(&lgr->lnk[i]))
 362                         continue;
 363                 rc = smcr_link_reg_rmb(&lgr->lnk[i], rmb_desc);
 364                 if (rc)
 365                         goto out;
 366         }
 367
 368         /* exchange confirm_rkey msg with peer */
 369         rc = smc_llc_do_confirm_rkey(link, rmb_desc);
 370         if (rc) {
 371                 rc = -EFAULT;
 372                 goto out;
 373         }
 374         rmb_desc->is_conf_rkey = true;
 375 out:
 376         mutex_unlock(&lgr->llc_conf_mutex);
 377         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
 378         return rc;
 379 }
 380
 381 static int smcr_clnt_conf_first_link(struct smc_sock *smc)
 382 {
 383         struct smc_link *link = smc->conn.lnk;
 384         struct smc_llc_qentry *qentry;
 385         int rc;
 386
 387         /* receive CONFIRM LINK request from server over RoCE fabric */
 388         qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
 389                               SMC_LLC_CONFIRM_LINK);
 390         if (!qentry) {
 391                 struct smc_clc_msg_decline dclc;
 392
 393                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 394                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 395                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
 396         }
 397         smc_llc_save_peer_uid(qentry);
 398         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
 399         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
 400         if (rc)
 401                 return SMC_CLC_DECL_RMBE_EC;
 402
 403         rc = smc_ib_modify_qp_rts(link);
 404         if (rc)
 405                 return SMC_CLC_DECL_ERR_RDYLNK;
 406
 407         smc_wr_remember_qp_attr(link);
 408
 409         if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
 410                 return SMC_CLC_DECL_ERR_REGRMB;
 411
 412         /* confirm_rkey is implicit on 1st contact */
 413         smc->conn.rmb_desc->is_conf_rkey = true;
 414
 415         /* send CONFIRM LINK response over RoCE fabric */
 416         rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
 417         if (rc < 0)
 418                 return SMC_CLC_DECL_TIMEOUT_CL;
 419
 420         smc_llc_link_active(link);
 421         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
 422
 423         /* optional 2nd link, receive ADD LINK request from server */
 424         qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
 425                               SMC_LLC_ADD_LINK);
 426         if (!qentry) {
 427                 struct smc_clc_msg_decline dclc;
 428
 429                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 430                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 431                 if (rc == -EAGAIN)
 432                         rc = 0; /* no DECLINE received, go with one link */
 433                 return rc;
 434         }
 435         smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
 436         smc_llc_cli_add_link(link, qentry);
 437         return 0;
 438 }
 439
 440 static void smcr_conn_save_peer_info(struct smc_sock *smc,
 441                                      struct smc_clc_msg_accept_confirm *clc)
 442 {
 443         int bufsize = smc_uncompress_bufsize(clc->r0.rmbe_size);
 444
 445         smc->conn.peer_rmbe_idx = clc->r0.rmbe_idx;
 446         smc->conn.local_tx_ctrl.token = ntohl(clc->r0.rmbe_alert_token);
 447         smc->conn.peer_rmbe_size = bufsize;
 448         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 449         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
 450 }
 451
 452 static bool smc_isascii(char *hostname)
 453 {
 454         int i;
 455
 456         for (i = 0; i < SMC_MAX_HOSTNAME_LEN; i++)
 457                 if (!isascii(hostname[i]))
 458                         return false;
 459         return true;
 460 }
 461
 462 static void smcd_conn_save_peer_info(struct smc_sock *smc,
 463                                      struct smc_clc_msg_accept_confirm *clc)
 464 {
 465         int bufsize = smc_uncompress_bufsize(clc->d0.dmbe_size);
 466
 467         smc->conn.peer_rmbe_idx = clc->d0.dmbe_idx;
 468         smc->conn.peer_token = clc->d0.token;
 469         /* msg header takes up space in the buffer */
 470         smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
 471         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 472         smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
 473         if (clc->hdr.version > SMC_V1 &&
 474             (clc->hdr.typev2 & SMC_FIRST_CONTACT_MASK)) {
 475                 struct smc_clc_msg_accept_confirm_v2 *clc_v2 =
 476                         (struct smc_clc_msg_accept_confirm_v2 *)clc;
 477                 struct smc_clc_first_contact_ext *fce =
 478                         (struct smc_clc_first_contact_ext *)
 479                                 (((u8 *)clc_v2) + sizeof(*clc_v2));
 480
 481                 memcpy(smc->conn.lgr->negotiated_eid, clc_v2->eid,
 482                        SMC_MAX_EID_LEN);
 483                 smc->conn.lgr->peer_os = fce->os_type;
 484                 smc->conn.lgr->peer_smc_release = fce->release;
 485                 if (smc_isascii(fce->hostname))
 486                         memcpy(smc->conn.lgr->peer_hostname, fce->hostname,
 487                                SMC_MAX_HOSTNAME_LEN);
 488         }
 489 }
 490
 491 static void smc_conn_save_peer_info(struct smc_sock *smc,
 492                                     struct smc_clc_msg_accept_confirm *clc)
 493 {
 494         if (smc->conn.lgr->is_smcd)
 495                 smcd_conn_save_peer_info(smc, clc);
 496         else
 497                 smcr_conn_save_peer_info(smc, clc);
 498 }
 499
 500 static void smc_link_save_peer_info(struct smc_link *link,
 501                                     struct smc_clc_msg_accept_confirm *clc)
 502 {
 503         link->peer_qpn = ntoh24(clc->r0.qpn);
 504         memcpy(link->peer_gid, clc->r0.lcl.gid, SMC_GID_SIZE);
 505         memcpy(link->peer_mac, clc->r0.lcl.mac, sizeof(link->peer_mac));
 506         link->peer_psn = ntoh24(clc->r0.psn);
 507         link->peer_mtu = clc->r0.qp_mtu;
 508 }
 509
 510 static void smc_switch_to_fallback(struct smc_sock *smc)
 511 {
 512         smc->use_fallback = true;
 513         if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
 514                 smc->clcsock->file = smc->sk.sk_socket->file;
 515                 smc->clcsock->file->private_data = smc->clcsock;
 516                 smc->clcsock->wq.fasync_list =
 517                         smc->sk.sk_socket->wq.fasync_list;
 518         }
 519 }
 520
 521 /* fall back during connect */
 522 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
 523 {
 524         smc_switch_to_fallback(smc);
 525         smc->fallback_rsn = reason_code;
 526         smc_copy_sock_settings_to_clc(smc);
 527         smc->connect_nonblock = 0;
 528         if (smc->sk.sk_state == SMC_INIT)
 529                 smc->sk.sk_state = SMC_ACTIVE;
 530         return 0;
 531 }
 532
 533 /* decline and fall back during connect */
 534 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code,
 535                                         u8 version)
 536 {
 537         int rc;
 538
 539         if (reason_code < 0) { /* error, fallback is not possible */
 540                 if (smc->sk.sk_state == SMC_INIT)
 541                         sock_put(&smc->sk); /* passive closing */
 542                 return reason_code;
 543         }
 544         if (reason_code != SMC_CLC_DECL_PEERDECL) {
 545                 rc = smc_clc_send_decline(smc, reason_code, version);
 546                 if (rc < 0) {
 547                         if (smc->sk.sk_state == SMC_INIT)
 548                                 sock_put(&smc->sk); /* passive closing */
 549                         return rc;
 550                 }
 551         }
 552         return smc_connect_fallback(smc, reason_code);
 553 }
 554
 555 /* abort connecting */
 556 static void smc_connect_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_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_connect_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_connect_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_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_v2_capable ? SMC_V2 : 0;
 950         ini->smc_type_v1 = SMC_TYPE_B;
 951         ini->smc_type_v2 = smc_ism_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         if (local_first)
1325                 smc_lgr_cleanup_early(&new_smc->conn);
1326         else
1327                 smc_conn_free(&new_smc->conn);
1328         if (reason_code < 0) { /* error, no fallback possible */
1329                 smc_listen_out_err(new_smc);
1330                 return;
1331         }
1332         smc_switch_to_fallback(new_smc);
1333         new_smc->fallback_rsn = reason_code;
1334         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1335                 if (smc_clc_send_decline(new_smc, reason_code, version) < 0) {
1336                         smc_listen_out_err(new_smc);
1337                         return;
1338                 }
1339         }
1340         smc_listen_out_connected(new_smc);
1341 }
1342
1343 /* listen worker: version checking */
1344 static int smc_listen_v2_check(struct smc_sock *new_smc,
1345                                struct smc_clc_msg_proposal *pclc,
1346                                struct smc_init_info *ini)
1347 {
1348         struct smc_clc_smcd_v2_extension *pclc_smcd_v2_ext;
1349         struct smc_clc_v2_extension *pclc_v2_ext;
1350
1351         ini->smc_type_v1 = pclc->hdr.typev1;
1352         ini->smc_type_v2 = pclc->hdr.typev2;
1353         ini->smcd_version = ini->smc_type_v1 != SMC_TYPE_N ? SMC_V1 : 0;
1354         if (pclc->hdr.version > SMC_V1)
1355                 ini->smcd_version |=
1356                                 ini->smc_type_v2 != SMC_TYPE_N ? SMC_V2 : 0;
1357         if (!smc_ism_v2_capable) {
1358                 ini->smcd_version &= ~SMC_V2;
1359                 goto out;
1360         }
1361         pclc_v2_ext = smc_get_clc_v2_ext(pclc);
1362         if (!pclc_v2_ext) {
1363                 ini->smcd_version &= ~SMC_V2;
1364                 goto out;
1365         }
1366         pclc_smcd_v2_ext = smc_get_clc_smcd_v2_ext(pclc_v2_ext);
1367         if (!pclc_smcd_v2_ext)
1368                 ini->smcd_version &= ~SMC_V2;
1369
1370 out:
1371         if (!ini->smcd_version) {
1372                 if (pclc->hdr.typev1 == SMC_TYPE_B ||
1373                     pclc->hdr.typev2 == SMC_TYPE_B)
1374                         return SMC_CLC_DECL_NOSMCDEV;
1375                 if (pclc->hdr.typev1 == SMC_TYPE_D ||
1376                     pclc->hdr.typev2 == SMC_TYPE_D)
1377                         return SMC_CLC_DECL_NOSMCDDEV;
1378                 return SMC_CLC_DECL_NOSMCRDEV;
1379         }
1380
1381         return 0;
1382 }
1383
1384 /* listen worker: check prefixes */
1385 static int smc_listen_prfx_check(struct smc_sock *new_smc,
1386                                  struct smc_clc_msg_proposal *pclc)
1387 {
1388         struct smc_clc_msg_proposal_prefix *pclc_prfx;
1389         struct socket *newclcsock = new_smc->clcsock;
1390
1391         if (pclc->hdr.typev1 == SMC_TYPE_N)
1392                 return 0;
1393         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1394         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1395                 return SMC_CLC_DECL_DIFFPREFIX;
1396
1397         return 0;
1398 }
1399
1400 /* listen worker: initialize connection and buffers */
1401 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1402                                 struct smc_init_info *ini)
1403 {
1404         int rc;
1405
1406         /* allocate connection / link group */
1407         rc = smc_conn_create(new_smc, ini);
1408         if (rc)
1409                 return rc;
1410
1411         /* create send buffer and rmb */
1412         if (smc_buf_create(new_smc, false))
1413                 return SMC_CLC_DECL_MEM;
1414
1415         return 0;
1416 }
1417
1418 /* listen worker: initialize connection and buffers for SMC-D */
1419 static int smc_listen_ism_init(struct smc_sock *new_smc,
1420                                struct smc_init_info *ini)
1421 {
1422         int rc;
1423
1424         rc = smc_conn_create(new_smc, ini);
1425         if (rc)
1426                 return rc;
1427
1428         /* Create send and receive buffers */
1429         rc = smc_buf_create(new_smc, true);
1430         if (rc) {
1431                 if (ini->first_contact_local)
1432                         smc_lgr_cleanup_early(&new_smc->conn);
1433                 else
1434                         smc_conn_free(&new_smc->conn);
1435                 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
1436                                          SMC_CLC_DECL_MEM;
1437         }
1438
1439         return 0;
1440 }
1441
1442 static bool smc_is_already_selected(struct smcd_dev *smcd,
1443                                     struct smc_init_info *ini,
1444                                     int matches)
1445 {
1446         int i;
1447
1448         for (i = 0; i < matches; i++)
1449                 if (smcd == ini->ism_dev[i])
1450                         return true;
1451
1452         return false;
1453 }
1454
1455 /* check for ISM devices matching proposed ISM devices */
1456 static void smc_check_ism_v2_match(struct smc_init_info *ini,
1457                                    u16 proposed_chid, u64 proposed_gid,
1458                                    unsigned int *matches)
1459 {
1460         struct smcd_dev *smcd;
1461
1462         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1463                 if (smcd->going_away)
1464                         continue;
1465                 if (smc_is_already_selected(smcd, ini, *matches))
1466                         continue;
1467                 if (smc_ism_get_chid(smcd) == proposed_chid &&
1468                     !smc_ism_cantalk(proposed_gid, ISM_RESERVED_VLANID, smcd)) {
1469                         ini->ism_peer_gid[*matches] = proposed_gid;
1470                         ini->ism_dev[*matches] = smcd;
1471                         (*matches)++;
1472                         break;
1473                 }
1474         }
1475 }
1476
1477 static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
1478                                         struct smc_clc_msg_proposal *pclc,
1479                                         struct smc_init_info *ini)
1480 {
1481         struct smc_clc_smcd_v2_extension *smcd_v2_ext;
1482         struct smc_clc_v2_extension *smc_v2_ext;
1483         struct smc_clc_msg_smcd *pclc_smcd;
1484         unsigned int matches = 0;
1485         u8 smcd_version;
1486         u8 *eid = NULL;
1487         int i;
1488
1489         if (!(ini->smcd_version & SMC_V2) || !smcd_indicated(ini->smc_type_v2))
1490                 goto not_found;
1491
1492         pclc_smcd = smc_get_clc_msg_smcd(pclc);
1493         smc_v2_ext = smc_get_clc_v2_ext(pclc);
1494         smcd_v2_ext = smc_get_clc_smcd_v2_ext(smc_v2_ext);
1495         if (!smcd_v2_ext ||
1496             !smc_v2_ext->hdr.flag.seid) /* no system EID support for SMCD */
1497                 goto not_found;
1498
1499         mutex_lock(&smcd_dev_list.mutex);
1500         if (pclc_smcd->ism.chid)
1501                 /* check for ISM device matching proposed native ISM device */
1502                 smc_check_ism_v2_match(ini, ntohs(pclc_smcd->ism.chid),
1503                                        ntohll(pclc_smcd->ism.gid), &matches);
1504         for (i = 1; i <= smc_v2_ext->hdr.ism_gid_cnt; i++) {
1505                 /* check for ISM devices matching proposed non-native ISM
1506                  * devices
1507                  */
1508                 smc_check_ism_v2_match(ini,
1509                                        ntohs(smcd_v2_ext->gidchid[i - 1].chid),
1510                                        ntohll(smcd_v2_ext->gidchid[i - 1].gid),
1511                                        &matches);
1512         }
1513         mutex_unlock(&smcd_dev_list.mutex);
1514
1515         if (ini->ism_dev[0]) {
1516                 smc_ism_get_system_eid(ini->ism_dev[0], &eid);
1517                 if (memcmp(eid, smcd_v2_ext->system_eid, SMC_MAX_EID_LEN))
1518                         goto not_found;
1519         } else {
1520                 goto not_found;
1521         }
1522
1523         /* separate - outside the smcd_dev_list.lock */
1524         smcd_version = ini->smcd_version;
1525         for (i = 0; i < matches; i++) {
1526                 ini->smcd_version = SMC_V2;
1527                 ini->is_smcd = true;
1528                 ini->ism_selected = i;
1529                 if (smc_listen_ism_init(new_smc, ini))
1530                         /* try next active ISM device */
1531                         continue;
1532                 return; /* matching and usable V2 ISM device found */
1533         }
1534         /* no V2 ISM device could be initialized */
1535         ini->smcd_version = smcd_version;       /* restore original value */
1536
1537 not_found:
1538         ini->smcd_version &= ~SMC_V2;
1539         ini->ism_dev[0] = NULL;
1540         ini->is_smcd = false;
1541 }
1542
1543 static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
1544                                         struct smc_clc_msg_proposal *pclc,
1545                                         struct smc_init_info *ini)
1546 {
1547         struct smc_clc_msg_smcd *pclc_smcd = smc_get_clc_msg_smcd(pclc);
1548
1549         /* check if ISM V1 is available */
1550         if (!(ini->smcd_version & SMC_V1) || !smcd_indicated(ini->smc_type_v1))
1551                 goto not_found;
1552         ini->is_smcd = true; /* prepare ISM check */
1553         ini->ism_peer_gid[0] = ntohll(pclc_smcd->ism.gid);
1554         if (smc_find_ism_device(new_smc, ini))
1555                 goto not_found;
1556         ini->ism_selected = 0;
1557         if (!smc_listen_ism_init(new_smc, ini))
1558                 return;         /* V1 ISM device found */
1559
1560 not_found:
1561         ini->ism_dev[0] = NULL;
1562         ini->is_smcd = false;
1563 }
1564
1565 /* listen worker: register buffers */
1566 static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
1567 {
1568         struct smc_connection *conn = &new_smc->conn;
1569
1570         if (!local_first) {
1571                 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
1572                         return SMC_CLC_DECL_ERR_REGRMB;
1573         }
1574         smc_rmb_sync_sg_for_device(&new_smc->conn);
1575
1576         return 0;
1577 }
1578
1579 static int smc_find_rdma_v1_device_serv(struct smc_sock *new_smc,
1580                                         struct smc_clc_msg_proposal *pclc,
1581                                         struct smc_init_info *ini)
1582 {
1583         int rc;
1584
1585         if (!smcr_indicated(ini->smc_type_v1))
1586                 return SMC_CLC_DECL_NOSMCDEV;
1587
1588         /* prepare RDMA check */
1589         ini->ib_lcl = &pclc->lcl;
1590         rc = smc_find_rdma_device(new_smc, ini);
1591         if (rc) {
1592                 /* no RDMA device found */
1593                 if (ini->smc_type_v1 == SMC_TYPE_B)
1594                         /* neither ISM nor RDMA device found */
1595                         rc = SMC_CLC_DECL_NOSMCDEV;
1596                 return rc;
1597         }
1598         rc = smc_listen_rdma_init(new_smc, ini);
1599         if (rc)
1600                 return rc;
1601         return smc_listen_rdma_reg(new_smc, ini->first_contact_local);
1602 }
1603
1604 /* determine the local device matching to proposal */
1605 static int smc_listen_find_device(struct smc_sock *new_smc,
1606                                   struct smc_clc_msg_proposal *pclc,
1607                                   struct smc_init_info *ini)
1608 {
1609         int rc;
1610
1611         /* check for ISM device matching V2 proposed device */
1612         smc_find_ism_v2_device_serv(new_smc, pclc, ini);
1613         if (ini->ism_dev[0])
1614                 return 0;
1615
1616         if (!(ini->smcd_version & SMC_V1))
1617                 return SMC_CLC_DECL_NOSMCDEV;
1618
1619         /* check for matching IP prefix and subnet length */
1620         rc = smc_listen_prfx_check(new_smc, pclc);
1621         if (rc)
1622                 return rc;
1623
1624         /* get vlan id from IP device */
1625         if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
1626                 return SMC_CLC_DECL_GETVLANERR;
1627
1628         /* check for ISM device matching V1 proposed device */
1629         smc_find_ism_v1_device_serv(new_smc, pclc, ini);
1630         if (ini->ism_dev[0])
1631                 return 0;
1632
1633         if (pclc->hdr.typev1 == SMC_TYPE_D)
1634                 return SMC_CLC_DECL_NOSMCDDEV; /* skip RDMA and decline */
1635
1636         /* check if RDMA is available */
1637         return smc_find_rdma_v1_device_serv(new_smc, pclc, ini);
1638 }
1639
1640 /* listen worker: finish RDMA setup */
1641 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1642                                   struct smc_clc_msg_accept_confirm *cclc,
1643                                   bool local_first)
1644 {
1645         struct smc_link *link = new_smc->conn.lnk;
1646         int reason_code = 0;
1647
1648         if (local_first)
1649                 smc_link_save_peer_info(link, cclc);
1650
1651         if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc))
1652                 return SMC_CLC_DECL_ERR_RTOK;
1653
1654         if (local_first) {
1655                 if (smc_ib_ready_link(link))
1656                         return SMC_CLC_DECL_ERR_RDYLNK;
1657                 /* QP confirmation over RoCE fabric */
1658                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1659                 reason_code = smcr_serv_conf_first_link(new_smc);
1660                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1661         }
1662         return reason_code;
1663 }
1664
1665 /* setup for connection of server */
1666 static void smc_listen_work(struct work_struct *work)
1667 {
1668         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1669                                                 smc_listen_work);
1670         u8 version = smc_ism_v2_capable ? SMC_V2 : SMC_V1;
1671         struct socket *newclcsock = new_smc->clcsock;
1672         struct smc_clc_msg_accept_confirm *cclc;
1673         struct smc_clc_msg_proposal_area *buf;
1674         struct smc_clc_msg_proposal *pclc;
1675         struct smc_init_info *ini = NULL;
1676         int rc = 0;
1677
1678         if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1679                 return smc_listen_out_err(new_smc);
1680
1681         if (new_smc->use_fallback) {
1682                 smc_listen_out_connected(new_smc);
1683                 return;
1684         }
1685
1686         /* check if peer is smc capable */
1687         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1688                 smc_switch_to_fallback(new_smc);
1689                 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1690                 smc_listen_out_connected(new_smc);
1691                 return;
1692         }
1693
1694         /* do inband token exchange -
1695          * wait for and receive SMC Proposal CLC message
1696          */
1697         buf = kzalloc(sizeof(*buf), GFP_KERNEL);
1698         if (!buf) {
1699                 rc = SMC_CLC_DECL_MEM;
1700                 goto out_decl;
1701         }
1702         pclc = (struct smc_clc_msg_proposal *)buf;
1703         rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
1704                               SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1705         if (rc)
1706                 goto out_decl;
1707         version = pclc->hdr.version == SMC_V1 ? SMC_V1 : version;
1708
1709         /* IPSec connections opt out of SMC optimizations */
1710         if (using_ipsec(new_smc)) {
1711                 rc = SMC_CLC_DECL_IPSEC;
1712                 goto out_decl;
1713         }
1714
1715         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1716         if (!ini) {
1717                 rc = SMC_CLC_DECL_MEM;
1718                 goto out_decl;
1719         }
1720
1721         /* initial version checking */
1722         rc = smc_listen_v2_check(new_smc, pclc, ini);
1723         if (rc)
1724                 goto out_decl;
1725
1726         mutex_lock(&smc_server_lgr_pending);
1727         smc_close_init(new_smc);
1728         smc_rx_init(new_smc);
1729         smc_tx_init(new_smc);
1730
1731         /* determine ISM or RoCE device used for connection */
1732         rc = smc_listen_find_device(new_smc, pclc, ini);
1733         if (rc)
1734                 goto out_unlock;
1735
1736         /* send SMC Accept CLC message */
1737         rc = smc_clc_send_accept(new_smc, ini->first_contact_local,
1738                                  ini->smcd_version == SMC_V2 ? SMC_V2 : SMC_V1);
1739         if (rc)
1740                 goto out_unlock;
1741
1742         /* SMC-D does not need this lock any more */
1743         if (ini->is_smcd)
1744                 mutex_unlock(&smc_server_lgr_pending);
1745
1746         /* receive SMC Confirm CLC message */
1747         memset(buf, 0, sizeof(*buf));
1748         cclc = (struct smc_clc_msg_accept_confirm *)buf;
1749         rc = smc_clc_wait_msg(new_smc, cclc, sizeof(*buf),
1750                               SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1751         if (rc) {
1752                 if (!ini->is_smcd)
1753                         goto out_unlock;
1754                 goto out_decl;
1755         }
1756
1757         /* finish worker */
1758         if (!ini->is_smcd) {
1759                 rc = smc_listen_rdma_finish(new_smc, cclc,
1760                                             ini->first_contact_local);
1761                 if (rc)
1762                         goto out_unlock;
1763                 mutex_unlock(&smc_server_lgr_pending);
1764         }
1765         smc_conn_save_peer_info(new_smc, cclc);
1766         smc_listen_out_connected(new_smc);
1767         goto out_free;
1768
1769 out_unlock:
1770         mutex_unlock(&smc_server_lgr_pending);
1771 out_decl:
1772         smc_listen_decline(new_smc, rc, ini ? ini->first_contact_local : 0,
1773                            version);
1774 out_free:
1775         kfree(ini);
1776         kfree(buf);
1777 }
1778
1779 static void smc_tcp_listen_work(struct work_struct *work)
1780 {
1781         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1782                                              tcp_listen_work);
1783         struct sock *lsk = &lsmc->sk;
1784         struct smc_sock *new_smc;
1785         int rc = 0;
1786
1787         lock_sock(lsk);
1788         while (lsk->sk_state == SMC_LISTEN) {
1789                 rc = smc_clcsock_accept(lsmc, &new_smc);
1790                 if (rc) /* clcsock accept queue empty or error */
1791                         goto out;
1792                 if (!new_smc)
1793                         continue;
1794
1795                 new_smc->listen_smc = lsmc;
1796                 new_smc->use_fallback = lsmc->use_fallback;
1797                 new_smc->fallback_rsn = lsmc->fallback_rsn;
1798                 sock_hold(lsk); /* sock_put in smc_listen_work */
1799                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1800                 smc_copy_sock_settings_to_smc(new_smc);
1801                 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1802                 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1803                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1804                 if (!queue_work(smc_hs_wq, &new_smc->smc_listen_work))
1805                         sock_put(&new_smc->sk);
1806         }
1807
1808 out:
1809         release_sock(lsk);
1810         sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
1811 }
1812
1813 static void smc_clcsock_data_ready(struct sock *listen_clcsock)
1814 {
1815         struct smc_sock *lsmc;
1816
1817         lsmc = (struct smc_sock *)
1818                ((uintptr_t)listen_clcsock->sk_user_data & ~SK_USER_DATA_NOCOPY);
1819         if (!lsmc)
1820                 return;
1821         lsmc->clcsk_data_ready(listen_clcsock);
1822         if (lsmc->sk.sk_state == SMC_LISTEN) {
1823                 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
1824                 if (!queue_work(smc_hs_wq, &lsmc->tcp_listen_work))
1825                         sock_put(&lsmc->sk);
1826         }
1827 }
1828
1829 static int smc_listen(struct socket *sock, int backlog)
1830 {
1831         struct sock *sk = sock->sk;
1832         struct smc_sock *smc;
1833         int rc;
1834
1835         smc = smc_sk(sk);
1836         lock_sock(sk);
1837
1838         rc = -EINVAL;
1839         if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
1840             smc->connect_nonblock)
1841                 goto out;
1842
1843         rc = 0;
1844         if (sk->sk_state == SMC_LISTEN) {
1845                 sk->sk_max_ack_backlog = backlog;
1846                 goto out;
1847         }
1848         /* some socket options are handled in core, so we could not apply
1849          * them to the clc socket -- copy smc socket options to clc socket
1850          */
1851         smc_copy_sock_settings_to_clc(smc);
1852         if (!smc->use_fallback)
1853                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1854
1855         /* save original sk_data_ready function and establish
1856          * smc-specific sk_data_ready function
1857          */
1858         smc->clcsk_data_ready = smc->clcsock->sk->sk_data_ready;
1859         smc->clcsock->sk->sk_data_ready = smc_clcsock_data_ready;
1860         smc->clcsock->sk->sk_user_data =
1861                 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
1862         rc = kernel_listen(smc->clcsock, backlog);
1863         if (rc)
1864                 goto out;
1865         sk->sk_max_ack_backlog = backlog;
1866         sk->sk_ack_backlog = 0;
1867         sk->sk_state = SMC_LISTEN;
1868
1869 out:
1870         release_sock(sk);
1871         return rc;
1872 }
1873
1874 static int smc_accept(struct socket *sock, struct socket *new_sock,
1875                       int flags, bool kern)
1876 {
1877         struct sock *sk = sock->sk, *nsk;
1878         DECLARE_WAITQUEUE(wait, current);
1879         struct smc_sock *lsmc;
1880         long timeo;
1881         int rc = 0;
1882
1883         lsmc = smc_sk(sk);
1884         sock_hold(sk); /* sock_put below */
1885         lock_sock(sk);
1886
1887         if (lsmc->sk.sk_state != SMC_LISTEN) {
1888                 rc = -EINVAL;
1889                 release_sock(sk);
1890                 goto out;
1891         }
1892
1893         /* Wait for an incoming connection */
1894         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1895         add_wait_queue_exclusive(sk_sleep(sk), &wait);
1896         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1897                 set_current_state(TASK_INTERRUPTIBLE);
1898                 if (!timeo) {
1899                         rc = -EAGAIN;
1900                         break;
1901                 }
1902                 release_sock(sk);
1903                 timeo = schedule_timeout(timeo);
1904                 /* wakeup by sk_data_ready in smc_listen_work() */
1905                 sched_annotate_sleep();
1906                 lock_sock(sk);
1907                 if (signal_pending(current)) {
1908                         rc = sock_intr_errno(timeo);
1909                         break;
1910                 }
1911         }
1912         set_current_state(TASK_RUNNING);
1913         remove_wait_queue(sk_sleep(sk), &wait);
1914
1915         if (!rc)
1916                 rc = sock_error(nsk);
1917         release_sock(sk);
1918         if (rc)
1919                 goto out;
1920
1921         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1922                 /* wait till data arrives on the socket */
1923                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1924                                                                 MSEC_PER_SEC);
1925                 if (smc_sk(nsk)->use_fallback) {
1926                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1927
1928                         lock_sock(clcsk);
1929                         if (skb_queue_empty(&clcsk->sk_receive_queue))
1930                                 sk_wait_data(clcsk, &timeo, NULL);
1931                         release_sock(clcsk);
1932                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1933                         lock_sock(nsk);
1934                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1935                         release_sock(nsk);
1936                 }
1937         }
1938
1939 out:
1940         sock_put(sk); /* sock_hold above */
1941         return rc;
1942 }
1943
1944 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1945                        int peer)
1946 {
1947         struct smc_sock *smc;
1948
1949         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1950             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1951                 return -ENOTCONN;
1952
1953         smc = smc_sk(sock->sk);
1954
1955         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1956 }
1957
1958 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1959 {
1960         struct sock *sk = sock->sk;
1961         struct smc_sock *smc;
1962         int rc = -EPIPE;
1963
1964         smc = smc_sk(sk);
1965         lock_sock(sk);
1966         if ((sk->sk_state != SMC_ACTIVE) &&
1967             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1968             (sk->sk_state != SMC_INIT))
1969                 goto out;
1970
1971         if (msg->msg_flags & MSG_FASTOPEN) {
1972                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1973                         smc_switch_to_fallback(smc);
1974                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1975                 } else {
1976                         rc = -EINVAL;
1977                         goto out;
1978                 }
1979         }
1980
1981         if (smc->use_fallback)
1982                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1983         else
1984                 rc = smc_tx_sendmsg(smc, msg, len);
1985 out:
1986         release_sock(sk);
1987         return rc;
1988 }
1989
1990 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1991                        int flags)
1992 {
1993         struct sock *sk = sock->sk;
1994         struct smc_sock *smc;
1995         int rc = -ENOTCONN;
1996
1997         smc = smc_sk(sk);
1998         lock_sock(sk);
1999         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2000                 /* socket was connected before, no more data to read */
2001                 rc = 0;
2002                 goto out;
2003         }
2004         if ((sk->sk_state == SMC_INIT) ||
2005             (sk->sk_state == SMC_LISTEN) ||
2006             (sk->sk_state == SMC_CLOSED))
2007                 goto out;
2008
2009         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2010                 rc = 0;
2011                 goto out;
2012         }
2013
2014         if (smc->use_fallback) {
2015                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
2016         } else {
2017                 msg->msg_namelen = 0;
2018                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
2019         }
2020
2021 out:
2022         release_sock(sk);
2023         return rc;
2024 }
2025
2026 static __poll_t smc_accept_poll(struct sock *parent)
2027 {
2028         struct smc_sock *isk = smc_sk(parent);
2029         __poll_t mask = 0;
2030
2031         spin_lock(&isk->accept_q_lock);
2032         if (!list_empty(&isk->accept_q))
2033                 mask = EPOLLIN | EPOLLRDNORM;
2034         spin_unlock(&isk->accept_q_lock);
2035
2036         return mask;
2037 }
2038
2039 static __poll_t smc_poll(struct file *file, struct socket *sock,
2040                              poll_table *wait)
2041 {
2042         struct sock *sk = sock->sk;
2043         struct smc_sock *smc;
2044         __poll_t mask = 0;
2045
2046         if (!sk)
2047                 return EPOLLNVAL;
2048
2049         smc = smc_sk(sock->sk);
2050         if (smc->use_fallback) {
2051                 /* delegate to CLC child sock */
2052                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
2053                 sk->sk_err = smc->clcsock->sk->sk_err;
2054         } else {
2055                 if (sk->sk_state != SMC_CLOSED)
2056                         sock_poll_wait(file, sock, wait);
2057                 if (sk->sk_err)
2058                         mask |= EPOLLERR;
2059                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
2060                     (sk->sk_state == SMC_CLOSED))
2061                         mask |= EPOLLHUP;
2062                 if (sk->sk_state == SMC_LISTEN) {
2063                         /* woken up by sk_data_ready in smc_listen_work() */
2064                         mask |= smc_accept_poll(sk);
2065                 } else if (smc->use_fallback) { /* as result of connect_work()*/
2066                         mask |= smc->clcsock->ops->poll(file, smc->clcsock,
2067                                                            wait);
2068                         sk->sk_err = smc->clcsock->sk->sk_err;
2069                 } else {
2070                         if ((sk->sk_state != SMC_INIT &&
2071                              atomic_read(&smc->conn.sndbuf_space)) ||
2072                             sk->sk_shutdown & SEND_SHUTDOWN) {
2073                                 mask |= EPOLLOUT | EPOLLWRNORM;
2074                         } else {
2075                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
2076                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2077                         }
2078                         if (atomic_read(&smc->conn.bytes_to_rcv))
2079                                 mask |= EPOLLIN | EPOLLRDNORM;
2080                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2081                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2082                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
2083                                 mask |= EPOLLIN;
2084                         if (smc->conn.urg_state == SMC_URG_VALID)
2085                                 mask |= EPOLLPRI;
2086                 }
2087         }
2088
2089         return mask;
2090 }
2091
2092 static int smc_shutdown(struct socket *sock, int how)
2093 {
2094         struct sock *sk = sock->sk;
2095         struct smc_sock *smc;
2096         int rc = -EINVAL;
2097         int rc1 = 0;
2098
2099         smc = smc_sk(sk);
2100
2101         if ((how < SHUT_RD) || (how > SHUT_RDWR))
2102                 return rc;
2103
2104         lock_sock(sk);
2105
2106         rc = -ENOTCONN;
2107         if ((sk->sk_state != SMC_ACTIVE) &&
2108             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
2109             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
2110             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2111             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
2112             (sk->sk_state != SMC_APPFINCLOSEWAIT))
2113                 goto out;
2114         if (smc->use_fallback) {
2115                 rc = kernel_sock_shutdown(smc->clcsock, how);
2116                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
2117                 if (sk->sk_shutdown == SHUTDOWN_MASK)
2118                         sk->sk_state = SMC_CLOSED;
2119                 goto out;
2120         }
2121         switch (how) {
2122         case SHUT_RDWR:         /* shutdown in both directions */
2123                 rc = smc_close_active(smc);
2124                 break;
2125         case SHUT_WR:
2126                 rc = smc_close_shutdown_write(smc);
2127                 break;
2128         case SHUT_RD:
2129                 rc = 0;
2130                 /* nothing more to do because peer is not involved */
2131                 break;
2132         }
2133         if (smc->clcsock)
2134                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
2135         /* map sock_shutdown_cmd constants to sk_shutdown value range */
2136         sk->sk_shutdown |= how + 1;
2137
2138 out:
2139         release_sock(sk);
2140         return rc ? rc : rc1;
2141 }
2142
2143 static int smc_setsockopt(struct socket *sock, int level, int optname,
2144                           sockptr_t optval, unsigned int optlen)
2145 {
2146         struct sock *sk = sock->sk;
2147         struct smc_sock *smc;
2148         int val, rc;
2149
2150         smc = smc_sk(sk);
2151
2152         /* generic setsockopts reaching us here always apply to the
2153          * CLC socket
2154          */
2155         if (unlikely(!smc->clcsock->ops->setsockopt))
2156                 rc = -EOPNOTSUPP;
2157         else
2158                 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
2159                                                    optval, optlen);
2160         if (smc->clcsock->sk->sk_err) {
2161                 sk->sk_err = smc->clcsock->sk->sk_err;
2162                 sk->sk_error_report(sk);
2163         }
2164
2165         if (optlen < sizeof(int))
2166                 return -EINVAL;
2167         if (copy_from_sockptr(&val, optval, sizeof(int)))
2168                 return -EFAULT;
2169
2170         lock_sock(sk);
2171         if (rc || smc->use_fallback)
2172                 goto out;
2173         switch (optname) {
2174         case TCP_ULP:
2175         case TCP_FASTOPEN:
2176         case TCP_FASTOPEN_CONNECT:
2177         case TCP_FASTOPEN_KEY:
2178         case TCP_FASTOPEN_NO_COOKIE:
2179                 /* option not supported by SMC */
2180                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2181                         smc_switch_to_fallback(smc);
2182                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
2183                 } else {
2184                         rc = -EINVAL;
2185                 }
2186                 break;
2187         case TCP_NODELAY:
2188                 if (sk->sk_state != SMC_INIT &&
2189                     sk->sk_state != SMC_LISTEN &&
2190                     sk->sk_state != SMC_CLOSED) {
2191                         if (val)
2192                                 mod_delayed_work(smc->conn.lgr->tx_wq,
2193                                                  &smc->conn.tx_work, 0);
2194                 }
2195                 break;
2196         case TCP_CORK:
2197                 if (sk->sk_state != SMC_INIT &&
2198                     sk->sk_state != SMC_LISTEN &&
2199                     sk->sk_state != SMC_CLOSED) {
2200                         if (!val)
2201                                 mod_delayed_work(smc->conn.lgr->tx_wq,
2202                                                  &smc->conn.tx_work, 0);
2203                 }
2204                 break;
2205         case TCP_DEFER_ACCEPT:
2206                 smc->sockopt_defer_accept = val;
2207                 break;
2208         default:
2209                 break;
2210         }
2211 out:
2212         release_sock(sk);
2213
2214         return rc;
2215 }
2216
2217 static int smc_getsockopt(struct socket *sock, int level, int optname,
2218                           char __user *optval, int __user *optlen)
2219 {
2220         struct smc_sock *smc;
2221
2222         smc = smc_sk(sock->sk);
2223         /* socket options apply to the CLC socket */
2224         if (unlikely(!smc->clcsock->ops->getsockopt))
2225                 return -EOPNOTSUPP;
2226         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
2227                                              optval, optlen);
2228 }
2229
2230 static int smc_ioctl(struct socket *sock, unsigned int cmd,
2231                      unsigned long arg)
2232 {
2233         union smc_host_cursor cons, urg;
2234         struct smc_connection *conn;
2235         struct smc_sock *smc;
2236         int answ;
2237
2238         smc = smc_sk(sock->sk);
2239         conn = &smc->conn;
2240         lock_sock(&smc->sk);
2241         if (smc->use_fallback) {
2242                 if (!smc->clcsock) {
2243                         release_sock(&smc->sk);
2244                         return -EBADF;
2245                 }
2246                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
2247                 release_sock(&smc->sk);
2248                 return answ;
2249         }
2250         switch (cmd) {
2251         case SIOCINQ: /* same as FIONREAD */
2252                 if (smc->sk.sk_state == SMC_LISTEN) {
2253                         release_sock(&smc->sk);
2254                         return -EINVAL;
2255                 }
2256                 if (smc->sk.sk_state == SMC_INIT ||
2257                     smc->sk.sk_state == SMC_CLOSED)
2258                         answ = 0;
2259                 else
2260                         answ = atomic_read(&smc->conn.bytes_to_rcv);
2261                 break;
2262         case SIOCOUTQ:
2263                 /* output queue size (not send + not acked) */
2264                 if (smc->sk.sk_state == SMC_LISTEN) {
2265                         release_sock(&smc->sk);
2266                         return -EINVAL;
2267                 }
2268                 if (smc->sk.sk_state == SMC_INIT ||
2269                     smc->sk.sk_state == SMC_CLOSED)
2270                         answ = 0;
2271                 else
2272                         answ = smc->conn.sndbuf_desc->len -
2273                                         atomic_read(&smc->conn.sndbuf_space);
2274                 break;
2275         case SIOCOUTQNSD:
2276                 /* output queue size (not send only) */
2277                 if (smc->sk.sk_state == SMC_LISTEN) {
2278                         release_sock(&smc->sk);
2279                         return -EINVAL;
2280                 }
2281                 if (smc->sk.sk_state == SMC_INIT ||
2282                     smc->sk.sk_state == SMC_CLOSED)
2283                         answ = 0;
2284                 else
2285                         answ = smc_tx_prepared_sends(&smc->conn);
2286                 break;
2287         case SIOCATMARK:
2288                 if (smc->sk.sk_state == SMC_LISTEN) {
2289                         release_sock(&smc->sk);
2290                         return -EINVAL;
2291                 }
2292                 if (smc->sk.sk_state == SMC_INIT ||
2293                     smc->sk.sk_state == SMC_CLOSED) {
2294                         answ = 0;
2295                 } else {
2296                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
2297                         smc_curs_copy(&urg, &conn->urg_curs, conn);
2298                         answ = smc_curs_diff(conn->rmb_desc->len,
2299                                              &cons, &urg) == 1;
2300                 }
2301                 break;
2302         default:
2303                 release_sock(&smc->sk);
2304                 return -ENOIOCTLCMD;
2305         }
2306         release_sock(&smc->sk);
2307
2308         return put_user(answ, (int __user *)arg);
2309 }
2310
2311 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
2312                             int offset, size_t size, int flags)
2313 {
2314         struct sock *sk = sock->sk;
2315         struct smc_sock *smc;
2316         int rc = -EPIPE;
2317
2318         smc = smc_sk(sk);
2319         lock_sock(sk);
2320         if (sk->sk_state != SMC_ACTIVE) {
2321                 release_sock(sk);
2322                 goto out;
2323         }
2324         release_sock(sk);
2325         if (smc->use_fallback)
2326                 rc = kernel_sendpage(smc->clcsock, page, offset,
2327                                      size, flags);
2328         else
2329                 rc = sock_no_sendpage(sock, page, offset, size, flags);
2330
2331 out:
2332         return rc;
2333 }
2334
2335 /* Map the affected portions of the rmbe into an spd, note the number of bytes
2336  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
2337  * updates till whenever a respective page has been fully processed.
2338  * Note that subsequent recv() calls have to wait till all splice() processing
2339  * completed.
2340  */
2341 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
2342                                struct pipe_inode_info *pipe, size_t len,
2343                                unsigned int flags)
2344 {
2345         struct sock *sk = sock->sk;
2346         struct smc_sock *smc;
2347         int rc = -ENOTCONN;
2348
2349         smc = smc_sk(sk);
2350         lock_sock(sk);
2351         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2352                 /* socket was connected before, no more data to read */
2353                 rc = 0;
2354                 goto out;
2355         }
2356         if (sk->sk_state == SMC_INIT ||
2357             sk->sk_state == SMC_LISTEN ||
2358             sk->sk_state == SMC_CLOSED)
2359                 goto out;
2360
2361         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2362                 rc = 0;
2363                 goto out;
2364         }
2365
2366         if (smc->use_fallback) {
2367                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
2368                                                     pipe, len, flags);
2369         } else {
2370                 if (*ppos) {
2371                         rc = -ESPIPE;
2372                         goto out;
2373                 }
2374                 if (flags & SPLICE_F_NONBLOCK)
2375                         flags = MSG_DONTWAIT;
2376                 else
2377                         flags = 0;
2378                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
2379         }
2380 out:
2381         release_sock(sk);
2382
2383         return rc;
2384 }
2385
2386 /* must look like tcp */
2387 static const struct proto_ops smc_sock_ops = {
2388         .family         = PF_SMC,
2389         .owner          = THIS_MODULE,
2390         .release        = smc_release,
2391         .bind           = smc_bind,
2392         .connect        = smc_connect,
2393         .socketpair     = sock_no_socketpair,
2394         .accept         = smc_accept,
2395         .getname        = smc_getname,
2396         .poll           = smc_poll,
2397         .ioctl          = smc_ioctl,
2398         .listen         = smc_listen,
2399         .shutdown       = smc_shutdown,
2400         .setsockopt     = smc_setsockopt,
2401         .getsockopt     = smc_getsockopt,
2402         .sendmsg        = smc_sendmsg,
2403         .recvmsg        = smc_recvmsg,
2404         .mmap           = sock_no_mmap,
2405         .sendpage       = smc_sendpage,
2406         .splice_read    = smc_splice_read,
2407 };
2408
2409 static int smc_create(struct net *net, struct socket *sock, int protocol,
2410                       int kern)
2411 {
2412         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
2413         struct smc_sock *smc;
2414         struct sock *sk;
2415         int rc;
2416
2417         rc = -ESOCKTNOSUPPORT;
2418         if (sock->type != SOCK_STREAM)
2419                 goto out;
2420
2421         rc = -EPROTONOSUPPORT;
2422         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
2423                 goto out;
2424
2425         rc = -ENOBUFS;
2426         sock->ops = &smc_sock_ops;
2427         sk = smc_sock_alloc(net, sock, protocol);
2428         if (!sk)
2429                 goto out;
2430
2431         /* create internal TCP socket for CLC handshake and fallback */
2432         smc = smc_sk(sk);
2433         smc->use_fallback = false; /* assume rdma capability first */
2434         smc->fallback_rsn = 0;
2435         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2436                               &smc->clcsock);
2437         if (rc) {
2438                 sk_common_release(sk);
2439                 goto out;
2440         }
2441         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2442         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
2443
2444 out:
2445         return rc;
2446 }
2447
2448 static const struct net_proto_family smc_sock_family_ops = {
2449         .family = PF_SMC,
2450         .owner  = THIS_MODULE,
2451         .create = smc_create,
2452 };
2453
2454 unsigned int smc_net_id;
2455
2456 static __net_init int smc_net_init(struct net *net)
2457 {
2458         return smc_pnet_net_init(net);
2459 }
2460
2461 static void __net_exit smc_net_exit(struct net *net)
2462 {
2463         smc_pnet_net_exit(net);
2464 }
2465
2466 static struct pernet_operations smc_net_ops = {
2467         .init = smc_net_init,
2468         .exit = smc_net_exit,
2469         .id   = &smc_net_id,
2470         .size = sizeof(struct smc_net),
2471 };
2472
2473 static int __init smc_init(void)
2474 {
2475         int rc;
2476
2477         rc = register_pernet_subsys(&smc_net_ops);
2478         if (rc)
2479                 return rc;
2480
2481         smc_ism_init();
2482         smc_clc_init();
2483
2484         rc = smc_pnet_init();
2485         if (rc)
2486                 goto out_pernet_subsys;
2487
2488         rc = -ENOMEM;
2489         smc_hs_wq = alloc_workqueue("smc_hs_wq", 0, 0);
2490         if (!smc_hs_wq)
2491                 goto out_pnet;
2492
2493         smc_close_wq = alloc_workqueue("smc_close_wq", 0, 0);
2494         if (!smc_close_wq)
2495                 goto out_alloc_hs_wq;
2496
2497         rc = smc_core_init();
2498         if (rc) {
2499                 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
2500                 goto out_alloc_wqs;
2501         }
2502
2503         rc = smc_llc_init();
2504         if (rc) {
2505                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2506                 goto out_core;
2507         }
2508
2509         rc = smc_cdc_init();
2510         if (rc) {
2511                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2512                 goto out_core;
2513         }
2514
2515         rc = proto_register(&smc_proto, 1);
2516         if (rc) {
2517                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2518                 goto out_core;
2519         }
2520
2521         rc = proto_register(&smc_proto6, 1);
2522         if (rc) {
2523                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2524                 goto out_proto;
2525         }
2526
2527         rc = sock_register(&smc_sock_family_ops);
2528         if (rc) {
2529                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2530                 goto out_proto6;
2531         }
2532         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2533         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2534
2535         rc = smc_ib_register_client();
2536         if (rc) {
2537                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2538                 goto out_sock;
2539         }
2540
2541         static_branch_enable(&tcp_have_smc);
2542         return 0;
2543
2544 out_sock:
2545         sock_unregister(PF_SMC);
2546 out_proto6:
2547         proto_unregister(&smc_proto6);
2548 out_proto:
2549         proto_unregister(&smc_proto);
2550 out_core:
2551         smc_core_exit();
2552 out_alloc_wqs:
2553         destroy_workqueue(smc_close_wq);
2554 out_alloc_hs_wq:
2555         destroy_workqueue(smc_hs_wq);
2556 out_pnet:
2557         smc_pnet_exit();
2558 out_pernet_subsys:
2559         unregister_pernet_subsys(&smc_net_ops);
2560
2561         return rc;
2562 }
2563
2564 static void __exit smc_exit(void)
2565 {
2566         static_branch_disable(&tcp_have_smc);
2567         sock_unregister(PF_SMC);
2568         smc_core_exit();
2569         smc_ib_unregister_client();
2570         destroy_workqueue(smc_close_wq);
2571         destroy_workqueue(smc_hs_wq);
2572         proto_unregister(&smc_proto6);
2573         proto_unregister(&smc_proto);
2574         smc_pnet_exit();
2575         unregister_pernet_subsys(&smc_net_ops);
2576         rcu_barrier();
2577 }
2578
2579 module_init(smc_init);
2580 module_exit(smc_exit);
2581
2582 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2583 MODULE_DESCRIPTION("smc socket address family");
2584 MODULE_LICENSE("GPL");
2585 MODULE_ALIAS_NETPROTO(PF_SMC);