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 : version;
 983         if (rc)
 984                 goto vlan_cleanup;
 985
 986         /* depending on previous steps, connect using rdma or ism */
 987         if (aclc->hdr.typev1 == SMC_TYPE_R)
 988                 rc = smc_connect_rdma(smc, aclc, ini);
 989         else if (aclc->hdr.typev1 == SMC_TYPE_D)
 990                 rc = smc_connect_ism(smc, aclc, ini);
 991         if (rc)
 992                 goto vlan_cleanup;
 993
 994         smc_connect_ism_vlan_cleanup(smc, ini);
 995         kfree(buf);
 996         kfree(ini);
 997         return 0;
 998
 999 vlan_cleanup:
1000         smc_connect_ism_vlan_cleanup(smc, ini);
1001         kfree(buf);
1002 fallback:
1003         kfree(ini);
1004         return smc_connect_decline_fallback(smc, rc, version);
1005 }
1006
1007 static void smc_connect_work(struct work_struct *work)
1008 {
1009         struct smc_sock *smc = container_of(work, struct smc_sock,
1010                                             connect_work);
1011         long timeo = smc->sk.sk_sndtimeo;
1012         int rc = 0;
1013
1014         if (!timeo)
1015                 timeo = MAX_SCHEDULE_TIMEOUT;
1016         lock_sock(smc->clcsock->sk);
1017         if (smc->clcsock->sk->sk_err) {
1018                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
1019         } else if ((1 << smc->clcsock->sk->sk_state) &
1020                                         (TCPF_SYN_SENT | TCP_SYN_RECV)) {
1021                 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
1022                 if ((rc == -EPIPE) &&
1023                     ((1 << smc->clcsock->sk->sk_state) &
1024                                         (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
1025                         rc = 0;
1026         }
1027         release_sock(smc->clcsock->sk);
1028         lock_sock(&smc->sk);
1029         if (rc != 0 || smc->sk.sk_err) {
1030                 smc->sk.sk_state = SMC_CLOSED;
1031                 if (rc == -EPIPE || rc == -EAGAIN)
1032                         smc->sk.sk_err = EPIPE;
1033                 else if (signal_pending(current))
1034                         smc->sk.sk_err = -sock_intr_errno(timeo);
1035                 sock_put(&smc->sk); /* passive closing */
1036                 goto out;
1037         }
1038
1039         rc = __smc_connect(smc);
1040         if (rc < 0)
1041                 smc->sk.sk_err = -rc;
1042
1043 out:
1044         if (!sock_flag(&smc->sk, SOCK_DEAD)) {
1045                 if (smc->sk.sk_err) {
1046                         smc->sk.sk_state_change(&smc->sk);
1047                 } else { /* allow polling before and after fallback decision */
1048                         smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
1049                         smc->sk.sk_write_space(&smc->sk);
1050                 }
1051         }
1052         release_sock(&smc->sk);
1053 }
1054
1055 static int smc_connect(struct socket *sock, struct sockaddr *addr,
1056                        int alen, int flags)
1057 {
1058         struct sock *sk = sock->sk;
1059         struct smc_sock *smc;
1060         int rc = -EINVAL;
1061
1062         smc = smc_sk(sk);
1063
1064         /* separate smc parameter checking to be safe */
1065         if (alen < sizeof(addr->sa_family))
1066                 goto out_err;
1067         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
1068                 goto out_err;
1069
1070         lock_sock(sk);
1071         switch (sk->sk_state) {
1072         default:
1073                 goto out;
1074         case SMC_ACTIVE:
1075                 rc = -EISCONN;
1076                 goto out;
1077         case SMC_INIT:
1078                 rc = 0;
1079                 break;
1080         }
1081
1082         smc_copy_sock_settings_to_clc(smc);
1083         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1084         if (smc->connect_nonblock) {
1085                 rc = -EALREADY;
1086                 goto out;
1087         }
1088         rc = kernel_connect(smc->clcsock, addr, alen, flags);
1089         if (rc && rc != -EINPROGRESS)
1090                 goto out;
1091
1092         sock_hold(&smc->sk); /* sock put in passive closing */
1093         if (smc->use_fallback)
1094                 goto out;
1095         if (flags & O_NONBLOCK) {
1096                 if (queue_work(smc_hs_wq, &smc->connect_work))
1097                         smc->connect_nonblock = 1;
1098                 rc = -EINPROGRESS;
1099         } else {
1100                 rc = __smc_connect(smc);
1101                 if (rc < 0)
1102                         goto out;
1103                 else
1104                         rc = 0; /* success cases including fallback */
1105         }
1106
1107 out:
1108         release_sock(sk);
1109 out_err:
1110         return rc;
1111 }
1112
1113 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
1114 {
1115         struct socket *new_clcsock = NULL;
1116         struct sock *lsk = &lsmc->sk;
1117         struct sock *new_sk;
1118         int rc = -EINVAL;
1119
1120         release_sock(lsk);
1121         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
1122         if (!new_sk) {
1123                 rc = -ENOMEM;
1124                 lsk->sk_err = ENOMEM;
1125                 *new_smc = NULL;
1126                 lock_sock(lsk);
1127                 goto out;
1128         }
1129         *new_smc = smc_sk(new_sk);
1130
1131         mutex_lock(&lsmc->clcsock_release_lock);
1132         if (lsmc->clcsock)
1133                 rc = kernel_accept(lsmc->clcsock, &new_clcsock, SOCK_NONBLOCK);
1134         mutex_unlock(&lsmc->clcsock_release_lock);
1135         lock_sock(lsk);
1136         if  (rc < 0 && rc != -EAGAIN)
1137                 lsk->sk_err = -rc;
1138         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
1139                 new_sk->sk_prot->unhash(new_sk);
1140                 if (new_clcsock)
1141                         sock_release(new_clcsock);
1142                 new_sk->sk_state = SMC_CLOSED;
1143                 sock_set_flag(new_sk, SOCK_DEAD);
1144                 sock_put(new_sk); /* final */
1145                 *new_smc = NULL;
1146                 goto out;
1147         }
1148
1149         /* new clcsock has inherited the smc listen-specific sk_data_ready
1150          * function; switch it back to the original sk_data_ready function
1151          */
1152         new_clcsock->sk->sk_data_ready = lsmc->clcsk_data_ready;
1153         (*new_smc)->clcsock = new_clcsock;
1154 out:
1155         return rc;
1156 }
1157
1158 /* add a just created sock to the accept queue of the listen sock as
1159  * candidate for a following socket accept call from user space
1160  */
1161 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
1162 {
1163         struct smc_sock *par = smc_sk(parent);
1164
1165         sock_hold(sk); /* sock_put in smc_accept_unlink () */
1166         spin_lock(&par->accept_q_lock);
1167         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
1168         spin_unlock(&par->accept_q_lock);
1169         sk_acceptq_added(parent);
1170 }
1171
1172 /* remove a socket from the accept queue of its parental listening socket */
1173 static void smc_accept_unlink(struct sock *sk)
1174 {
1175         struct smc_sock *par = smc_sk(sk)->listen_smc;
1176
1177         spin_lock(&par->accept_q_lock);
1178         list_del_init(&smc_sk(sk)->accept_q);
1179         spin_unlock(&par->accept_q_lock);
1180         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
1181         sock_put(sk); /* sock_hold in smc_accept_enqueue */
1182 }
1183
1184 /* remove a sock from the accept queue to bind it to a new socket created
1185  * for a socket accept call from user space
1186  */
1187 struct sock *smc_accept_dequeue(struct sock *parent,
1188                                 struct socket *new_sock)
1189 {
1190         struct smc_sock *isk, *n;
1191         struct sock *new_sk;
1192
1193         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1194                 new_sk = (struct sock *)isk;
1195
1196                 smc_accept_unlink(new_sk);
1197                 if (new_sk->sk_state == SMC_CLOSED) {
1198                         new_sk->sk_prot->unhash(new_sk);
1199                         if (isk->clcsock) {
1200                                 sock_release(isk->clcsock);
1201                                 isk->clcsock = NULL;
1202                         }
1203                         sock_put(new_sk); /* final */
1204                         continue;
1205                 }
1206                 if (new_sock) {
1207                         sock_graft(new_sk, new_sock);
1208                         if (isk->use_fallback) {
1209                                 smc_sk(new_sk)->clcsock->file = new_sock->file;
1210                                 isk->clcsock->file->private_data = isk->clcsock;
1211                         }
1212                 }
1213                 return new_sk;
1214         }
1215         return NULL;
1216 }
1217
1218 /* clean up for a created but never accepted sock */
1219 void smc_close_non_accepted(struct sock *sk)
1220 {
1221         struct smc_sock *smc = smc_sk(sk);
1222
1223         sock_hold(sk); /* sock_put below */
1224         lock_sock(sk);
1225         if (!sk->sk_lingertime)
1226                 /* wait for peer closing */
1227                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
1228         __smc_release(smc);
1229         release_sock(sk);
1230         sock_put(sk); /* sock_hold above */
1231         sock_put(sk); /* final sock_put */
1232 }
1233
1234 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1235 {
1236         struct smc_link *link = smc->conn.lnk;
1237         struct smc_llc_qentry *qentry;
1238         int rc;
1239
1240         if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
1241                 return SMC_CLC_DECL_ERR_REGRMB;
1242
1243         /* send CONFIRM LINK request to client over the RoCE fabric */
1244         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1245         if (rc < 0)
1246                 return SMC_CLC_DECL_TIMEOUT_CL;
1247
1248         /* receive CONFIRM LINK response from client over the RoCE fabric */
1249         qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1250                               SMC_LLC_CONFIRM_LINK);
1251         if (!qentry) {
1252                 struct smc_clc_msg_decline dclc;
1253
1254                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1255                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1256                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1257         }
1258         smc_llc_save_peer_uid(qentry);
1259         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1260         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1261         if (rc)
1262                 return SMC_CLC_DECL_RMBE_EC;
1263
1264         /* confirm_rkey is implicit on 1st contact */
1265         smc->conn.rmb_desc->is_conf_rkey = true;
1266
1267         smc_llc_link_active(link);
1268         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1269
1270         /* initial contact - try to establish second link */
1271         smc_llc_srv_add_link(link);
1272         return 0;
1273 }
1274
1275 /* listen worker: finish */
1276 static void smc_listen_out(struct smc_sock *new_smc)
1277 {
1278         struct smc_sock *lsmc = new_smc->listen_smc;
1279         struct sock *newsmcsk = &new_smc->sk;
1280
1281         if (lsmc->sk.sk_state == SMC_LISTEN) {
1282                 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1283                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1284                 release_sock(&lsmc->sk);
1285         } else { /* no longer listening */
1286                 smc_close_non_accepted(newsmcsk);
1287         }
1288
1289         /* Wake up accept */
1290         lsmc->sk.sk_data_ready(&lsmc->sk);
1291         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1292 }
1293
1294 /* listen worker: finish in state connected */
1295 static void smc_listen_out_connected(struct smc_sock *new_smc)
1296 {
1297         struct sock *newsmcsk = &new_smc->sk;
1298
1299         sk_refcnt_debug_inc(newsmcsk);
1300         if (newsmcsk->sk_state == SMC_INIT)
1301                 newsmcsk->sk_state = SMC_ACTIVE;
1302
1303         smc_listen_out(new_smc);
1304 }
1305
1306 /* listen worker: finish in error state */
1307 static void smc_listen_out_err(struct smc_sock *new_smc)
1308 {
1309         struct sock *newsmcsk = &new_smc->sk;
1310
1311         if (newsmcsk->sk_state == SMC_INIT)
1312                 sock_put(&new_smc->sk); /* passive closing */
1313         newsmcsk->sk_state = SMC_CLOSED;
1314
1315         smc_listen_out(new_smc);
1316 }
1317
1318 /* listen worker: decline and fall back if possible */
1319 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1320                                int local_first, u8 version)
1321 {
1322         /* RDMA setup failed, switch back to TCP */
1323         if (local_first)
1324                 smc_lgr_cleanup_early(&new_smc->conn);
1325         else
1326                 smc_conn_free(&new_smc->conn);
1327         if (reason_code < 0) { /* error, no fallback possible */
1328                 smc_listen_out_err(new_smc);
1329                 return;
1330         }
1331         smc_switch_to_fallback(new_smc);
1332         new_smc->fallback_rsn = reason_code;
1333         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1334                 if (smc_clc_send_decline(new_smc, reason_code, version) < 0) {
1335                         smc_listen_out_err(new_smc);
1336                         return;
1337                 }
1338         }
1339         smc_listen_out_connected(new_smc);
1340 }
1341
1342 /* listen worker: version checking */
1343 static int smc_listen_v2_check(struct smc_sock *new_smc,
1344                                struct smc_clc_msg_proposal *pclc,
1345                                struct smc_init_info *ini)
1346 {
1347         struct smc_clc_smcd_v2_extension *pclc_smcd_v2_ext;
1348         struct smc_clc_v2_extension *pclc_v2_ext;
1349
1350         ini->smc_type_v1 = pclc->hdr.typev1;
1351         ini->smc_type_v2 = pclc->hdr.typev2;
1352         ini->smcd_version = ini->smc_type_v1 != SMC_TYPE_N ? SMC_V1 : 0;
1353         if (pclc->hdr.version > SMC_V1)
1354                 ini->smcd_version |=
1355                                 ini->smc_type_v2 != SMC_TYPE_N ? SMC_V2 : 0;
1356         if (!smc_ism_v2_capable) {
1357                 ini->smcd_version &= ~SMC_V2;
1358                 goto out;
1359         }
1360         pclc_v2_ext = smc_get_clc_v2_ext(pclc);
1361         if (!pclc_v2_ext) {
1362                 ini->smcd_version &= ~SMC_V2;
1363                 goto out;
1364         }
1365         pclc_smcd_v2_ext = smc_get_clc_smcd_v2_ext(pclc_v2_ext);
1366         if (!pclc_smcd_v2_ext)
1367                 ini->smcd_version &= ~SMC_V2;
1368
1369 out:
1370         if (!ini->smcd_version) {
1371                 if (pclc->hdr.typev1 == SMC_TYPE_B ||
1372                     pclc->hdr.typev2 == SMC_TYPE_B)
1373                         return SMC_CLC_DECL_NOSMCDEV;
1374                 if (pclc->hdr.typev1 == SMC_TYPE_D ||
1375                     pclc->hdr.typev2 == SMC_TYPE_D)
1376                         return SMC_CLC_DECL_NOSMCDDEV;
1377                 return SMC_CLC_DECL_NOSMCRDEV;
1378         }
1379
1380         return 0;
1381 }
1382
1383 /* listen worker: check prefixes */
1384 static int smc_listen_prfx_check(struct smc_sock *new_smc,
1385                                  struct smc_clc_msg_proposal *pclc)
1386 {
1387         struct smc_clc_msg_proposal_prefix *pclc_prfx;
1388         struct socket *newclcsock = new_smc->clcsock;
1389
1390         if (pclc->hdr.typev1 == SMC_TYPE_N)
1391                 return 0;
1392         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1393         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1394                 return SMC_CLC_DECL_DIFFPREFIX;
1395
1396         return 0;
1397 }
1398
1399 /* listen worker: initialize connection and buffers */
1400 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1401                                 struct smc_init_info *ini)
1402 {
1403         int rc;
1404
1405         /* allocate connection / link group */
1406         rc = smc_conn_create(new_smc, ini);
1407         if (rc)
1408                 return rc;
1409
1410         /* create send buffer and rmb */
1411         if (smc_buf_create(new_smc, false))
1412                 return SMC_CLC_DECL_MEM;
1413
1414         return 0;
1415 }
1416
1417 /* listen worker: initialize connection and buffers for SMC-D */
1418 static int smc_listen_ism_init(struct smc_sock *new_smc,
1419                                struct smc_init_info *ini)
1420 {
1421         int rc;
1422
1423         rc = smc_conn_create(new_smc, ini);
1424         if (rc)
1425                 return rc;
1426
1427         /* Create send and receive buffers */
1428         rc = smc_buf_create(new_smc, true);
1429         if (rc) {
1430                 if (ini->first_contact_local)
1431                         smc_lgr_cleanup_early(&new_smc->conn);
1432                 else
1433                         smc_conn_free(&new_smc->conn);
1434                 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
1435                                          SMC_CLC_DECL_MEM;
1436         }
1437
1438         return 0;
1439 }
1440
1441 static bool smc_is_already_selected(struct smcd_dev *smcd,
1442                                     struct smc_init_info *ini,
1443                                     int matches)
1444 {
1445         int i;
1446
1447         for (i = 0; i < matches; i++)
1448                 if (smcd == ini->ism_dev[i])
1449                         return true;
1450
1451         return false;
1452 }
1453
1454 /* check for ISM devices matching proposed ISM devices */
1455 static void smc_check_ism_v2_match(struct smc_init_info *ini,
1456                                    u16 proposed_chid, u64 proposed_gid,
1457                                    unsigned int *matches)
1458 {
1459         struct smcd_dev *smcd;
1460
1461         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1462                 if (smcd->going_away)
1463                         continue;
1464                 if (smc_is_already_selected(smcd, ini, *matches))
1465                         continue;
1466                 if (smc_ism_get_chid(smcd) == proposed_chid &&
1467                     !smc_ism_cantalk(proposed_gid, ISM_RESERVED_VLANID, smcd)) {
1468                         ini->ism_peer_gid[*matches] = proposed_gid;
1469                         ini->ism_dev[*matches] = smcd;
1470                         (*matches)++;
1471                         break;
1472                 }
1473         }
1474 }
1475
1476 static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
1477                                         struct smc_clc_msg_proposal *pclc,
1478                                         struct smc_init_info *ini)
1479 {
1480         struct smc_clc_smcd_v2_extension *smcd_v2_ext;
1481         struct smc_clc_v2_extension *smc_v2_ext;
1482         struct smc_clc_msg_smcd *pclc_smcd;
1483         unsigned int matches = 0;
1484         u8 smcd_version;
1485         u8 *eid = NULL;
1486         int i;
1487
1488         if (!(ini->smcd_version & SMC_V2) || !smcd_indicated(ini->smc_type_v2))
1489                 goto not_found;
1490
1491         pclc_smcd = smc_get_clc_msg_smcd(pclc);
1492         smc_v2_ext = smc_get_clc_v2_ext(pclc);
1493         smcd_v2_ext = smc_get_clc_smcd_v2_ext(smc_v2_ext);
1494         if (!smcd_v2_ext ||
1495             !smc_v2_ext->hdr.flag.seid) /* no system EID support for SMCD */
1496                 goto not_found;
1497
1498         mutex_lock(&smcd_dev_list.mutex);
1499         if (pclc_smcd->ism.chid)
1500                 /* check for ISM device matching proposed native ISM device */
1501                 smc_check_ism_v2_match(ini, ntohs(pclc_smcd->ism.chid),
1502                                        ntohll(pclc_smcd->ism.gid), &matches);
1503         for (i = 1; i <= smc_v2_ext->hdr.ism_gid_cnt; i++) {
1504                 /* check for ISM devices matching proposed non-native ISM
1505                  * devices
1506                  */
1507                 smc_check_ism_v2_match(ini,
1508                                        ntohs(smcd_v2_ext->gidchid[i - 1].chid),
1509                                        ntohll(smcd_v2_ext->gidchid[i - 1].gid),
1510                                        &matches);
1511         }
1512         mutex_unlock(&smcd_dev_list.mutex);
1513
1514         if (ini->ism_dev[0]) {
1515                 smc_ism_get_system_eid(ini->ism_dev[0], &eid);
1516                 if (memcmp(eid, smcd_v2_ext->system_eid, SMC_MAX_EID_LEN))
1517                         goto not_found;
1518         } else {
1519                 goto not_found;
1520         }
1521
1522         /* separate - outside the smcd_dev_list.lock */
1523         smcd_version = ini->smcd_version;
1524         for (i = 0; i < matches; i++) {
1525                 ini->smcd_version = SMC_V2;
1526                 ini->is_smcd = true;
1527                 ini->ism_selected = i;
1528                 if (smc_listen_ism_init(new_smc, ini))
1529                         /* try next active ISM device */
1530                         continue;
1531                 return; /* matching and usable V2 ISM device found */
1532         }
1533         /* no V2 ISM device could be initialized */
1534         ini->smcd_version = smcd_version;       /* restore original value */
1535
1536 not_found:
1537         ini->smcd_version &= ~SMC_V2;
1538         ini->ism_dev[0] = NULL;
1539         ini->is_smcd = false;
1540 }
1541
1542 static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
1543                                         struct smc_clc_msg_proposal *pclc,
1544                                         struct smc_init_info *ini)
1545 {
1546         struct smc_clc_msg_smcd *pclc_smcd = smc_get_clc_msg_smcd(pclc);
1547
1548         /* check if ISM V1 is available */
1549         if (!(ini->smcd_version & SMC_V1) || !smcd_indicated(ini->smc_type_v1))
1550                 goto not_found;
1551         ini->is_smcd = true; /* prepare ISM check */
1552         ini->ism_peer_gid[0] = ntohll(pclc_smcd->ism.gid);
1553         if (smc_find_ism_device(new_smc, ini))
1554                 goto not_found;
1555         ini->ism_selected = 0;
1556         if (!smc_listen_ism_init(new_smc, ini))
1557                 return;         /* V1 ISM device found */
1558
1559 not_found:
1560         ini->ism_dev[0] = NULL;
1561         ini->is_smcd = false;
1562 }
1563
1564 /* listen worker: register buffers */
1565 static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
1566 {
1567         struct smc_connection *conn = &new_smc->conn;
1568
1569         if (!local_first) {
1570                 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
1571                         return SMC_CLC_DECL_ERR_REGRMB;
1572         }
1573         smc_rmb_sync_sg_for_device(&new_smc->conn);
1574
1575         return 0;
1576 }
1577
1578 static int smc_find_rdma_v1_device_serv(struct smc_sock *new_smc,
1579                                         struct smc_clc_msg_proposal *pclc,
1580                                         struct smc_init_info *ini)
1581 {
1582         int rc;
1583
1584         if (!smcr_indicated(ini->smc_type_v1))
1585                 return SMC_CLC_DECL_NOSMCDEV;
1586
1587         /* prepare RDMA check */
1588         ini->ib_lcl = &pclc->lcl;
1589         rc = smc_find_rdma_device(new_smc, ini);
1590         if (rc) {
1591                 /* no RDMA device found */
1592                 if (ini->smc_type_v1 == SMC_TYPE_B)
1593                         /* neither ISM nor RDMA device found */
1594                         rc = SMC_CLC_DECL_NOSMCDEV;
1595                 return rc;
1596         }
1597         rc = smc_listen_rdma_init(new_smc, ini);
1598         if (rc)
1599                 return rc;
1600         return smc_listen_rdma_reg(new_smc, ini->first_contact_local);
1601 }
1602
1603 /* determine the local device matching to proposal */
1604 static int smc_listen_find_device(struct smc_sock *new_smc,
1605                                   struct smc_clc_msg_proposal *pclc,
1606                                   struct smc_init_info *ini)
1607 {
1608         int rc;
1609
1610         /* check for ISM device matching V2 proposed device */
1611         smc_find_ism_v2_device_serv(new_smc, pclc, ini);
1612         if (ini->ism_dev[0])
1613                 return 0;
1614
1615         if (!(ini->smcd_version & SMC_V1))
1616                 return SMC_CLC_DECL_NOSMCDEV;
1617
1618         /* check for matching IP prefix and subnet length */
1619         rc = smc_listen_prfx_check(new_smc, pclc);
1620         if (rc)
1621                 return rc;
1622
1623         /* get vlan id from IP device */
1624         if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
1625                 return SMC_CLC_DECL_GETVLANERR;
1626
1627         /* check for ISM device matching V1 proposed device */
1628         smc_find_ism_v1_device_serv(new_smc, pclc, ini);
1629         if (ini->ism_dev[0])
1630                 return 0;
1631
1632         if (pclc->hdr.typev1 == SMC_TYPE_D)
1633                 return SMC_CLC_DECL_NOSMCDDEV; /* skip RDMA and decline */
1634
1635         /* check if RDMA is available */
1636         return smc_find_rdma_v1_device_serv(new_smc, pclc, ini);
1637 }
1638
1639 /* listen worker: finish RDMA setup */
1640 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1641                                   struct smc_clc_msg_accept_confirm *cclc,
1642                                   bool local_first)
1643 {
1644         struct smc_link *link = new_smc->conn.lnk;
1645         int reason_code = 0;
1646
1647         if (local_first)
1648                 smc_link_save_peer_info(link, cclc);
1649
1650         if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc))
1651                 return SMC_CLC_DECL_ERR_RTOK;
1652
1653         if (local_first) {
1654                 if (smc_ib_ready_link(link))
1655                         return SMC_CLC_DECL_ERR_RDYLNK;
1656                 /* QP confirmation over RoCE fabric */
1657                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1658                 reason_code = smcr_serv_conf_first_link(new_smc);
1659                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1660         }
1661         return reason_code;
1662 }
1663
1664 /* setup for connection of server */
1665 static void smc_listen_work(struct work_struct *work)
1666 {
1667         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1668                                                 smc_listen_work);
1669         u8 version = smc_ism_v2_capable ? SMC_V2 : SMC_V1;
1670         struct socket *newclcsock = new_smc->clcsock;
1671         struct smc_clc_msg_accept_confirm *cclc;
1672         struct smc_clc_msg_proposal_area *buf;
1673         struct smc_clc_msg_proposal *pclc;
1674         struct smc_init_info *ini = NULL;
1675         int rc = 0;
1676
1677         if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1678                 return smc_listen_out_err(new_smc);
1679
1680         if (new_smc->use_fallback) {
1681                 smc_listen_out_connected(new_smc);
1682                 return;
1683         }
1684
1685         /* check if peer is smc capable */
1686         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1687                 smc_switch_to_fallback(new_smc);
1688                 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1689                 smc_listen_out_connected(new_smc);
1690                 return;
1691         }
1692
1693         /* do inband token exchange -
1694          * wait for and receive SMC Proposal CLC message
1695          */
1696         buf = kzalloc(sizeof(*buf), GFP_KERNEL);
1697         if (!buf) {
1698                 rc = SMC_CLC_DECL_MEM;
1699                 goto out_decl;
1700         }
1701         pclc = (struct smc_clc_msg_proposal *)buf;
1702         rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
1703                               SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1704         if (rc)
1705                 goto out_decl;
1706         version = pclc->hdr.version == SMC_V1 ? SMC_V1 : version;
1707
1708         /* IPSec connections opt out of SMC optimizations */
1709         if (using_ipsec(new_smc)) {
1710                 rc = SMC_CLC_DECL_IPSEC;
1711                 goto out_decl;
1712         }
1713
1714         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1715         if (!ini) {
1716                 rc = SMC_CLC_DECL_MEM;
1717                 goto out_decl;
1718         }
1719
1720         /* initial version checking */
1721         rc = smc_listen_v2_check(new_smc, pclc, ini);
1722         if (rc)
1723                 goto out_decl;
1724
1725         mutex_lock(&smc_server_lgr_pending);
1726         smc_close_init(new_smc);
1727         smc_rx_init(new_smc);
1728         smc_tx_init(new_smc);
1729
1730         /* determine ISM or RoCE device used for connection */
1731         rc = smc_listen_find_device(new_smc, pclc, ini);
1732         if (rc)
1733                 goto out_unlock;
1734
1735         /* send SMC Accept CLC message */
1736         rc = smc_clc_send_accept(new_smc, ini->first_contact_local,
1737                                  ini->smcd_version == SMC_V2 ? SMC_V2 : SMC_V1);
1738         if (rc)
1739                 goto out_unlock;
1740
1741         /* SMC-D does not need this lock any more */
1742         if (ini->is_smcd)
1743                 mutex_unlock(&smc_server_lgr_pending);
1744
1745         /* receive SMC Confirm CLC message */
1746         memset(buf, 0, sizeof(*buf));
1747         cclc = (struct smc_clc_msg_accept_confirm *)buf;
1748         rc = smc_clc_wait_msg(new_smc, cclc, sizeof(*buf),
1749                               SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1750         if (rc) {
1751                 if (!ini->is_smcd)
1752                         goto out_unlock;
1753                 goto out_decl;
1754         }
1755
1756         /* finish worker */
1757         if (!ini->is_smcd) {
1758                 rc = smc_listen_rdma_finish(new_smc, cclc,
1759                                             ini->first_contact_local);
1760                 if (rc)
1761                         goto out_unlock;
1762                 mutex_unlock(&smc_server_lgr_pending);
1763         }
1764         smc_conn_save_peer_info(new_smc, cclc);
1765         smc_listen_out_connected(new_smc);
1766         goto out_free;
1767
1768 out_unlock:
1769         mutex_unlock(&smc_server_lgr_pending);
1770 out_decl:
1771         smc_listen_decline(new_smc, rc, ini ? ini->first_contact_local : 0,
1772                            version);
1773 out_free:
1774         kfree(ini);
1775         kfree(buf);
1776 }
1777
1778 static void smc_tcp_listen_work(struct work_struct *work)
1779 {
1780         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1781                                              tcp_listen_work);
1782         struct sock *lsk = &lsmc->sk;
1783         struct smc_sock *new_smc;
1784         int rc = 0;
1785
1786         lock_sock(lsk);
1787         while (lsk->sk_state == SMC_LISTEN) {
1788                 rc = smc_clcsock_accept(lsmc, &new_smc);
1789                 if (rc) /* clcsock accept queue empty or error */
1790                         goto out;
1791                 if (!new_smc)
1792                         continue;
1793
1794                 new_smc->listen_smc = lsmc;
1795                 new_smc->use_fallback = lsmc->use_fallback;
1796                 new_smc->fallback_rsn = lsmc->fallback_rsn;
1797                 sock_hold(lsk); /* sock_put in smc_listen_work */
1798                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1799                 smc_copy_sock_settings_to_smc(new_smc);
1800                 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1801                 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1802                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1803                 if (!queue_work(smc_hs_wq, &new_smc->smc_listen_work))
1804                         sock_put(&new_smc->sk);
1805         }
1806
1807 out:
1808         release_sock(lsk);
1809         sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
1810 }
1811
1812 static void smc_clcsock_data_ready(struct sock *listen_clcsock)
1813 {
1814         struct smc_sock *lsmc;
1815
1816         lsmc = (struct smc_sock *)
1817                ((uintptr_t)listen_clcsock->sk_user_data & ~SK_USER_DATA_NOCOPY);
1818         if (!lsmc)
1819                 return;
1820         lsmc->clcsk_data_ready(listen_clcsock);
1821         if (lsmc->sk.sk_state == SMC_LISTEN) {
1822                 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
1823                 if (!queue_work(smc_hs_wq, &lsmc->tcp_listen_work))
1824                         sock_put(&lsmc->sk);
1825         }
1826 }
1827
1828 static int smc_listen(struct socket *sock, int backlog)
1829 {
1830         struct sock *sk = sock->sk;
1831         struct smc_sock *smc;
1832         int rc;
1833
1834         smc = smc_sk(sk);
1835         lock_sock(sk);
1836
1837         rc = -EINVAL;
1838         if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
1839             smc->connect_nonblock)
1840                 goto out;
1841
1842         rc = 0;
1843         if (sk->sk_state == SMC_LISTEN) {
1844                 sk->sk_max_ack_backlog = backlog;
1845                 goto out;
1846         }
1847         /* some socket options are handled in core, so we could not apply
1848          * them to the clc socket -- copy smc socket options to clc socket
1849          */
1850         smc_copy_sock_settings_to_clc(smc);
1851         if (!smc->use_fallback)
1852                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1853
1854         /* save original sk_data_ready function and establish
1855          * smc-specific sk_data_ready function
1856          */
1857         smc->clcsk_data_ready = smc->clcsock->sk->sk_data_ready;
1858         smc->clcsock->sk->sk_data_ready = smc_clcsock_data_ready;
1859         smc->clcsock->sk->sk_user_data =
1860                 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
1861         rc = kernel_listen(smc->clcsock, backlog);
1862         if (rc)
1863                 goto out;
1864         sk->sk_max_ack_backlog = backlog;
1865         sk->sk_ack_backlog = 0;
1866         sk->sk_state = SMC_LISTEN;
1867
1868 out:
1869         release_sock(sk);
1870         return rc;
1871 }
1872
1873 static int smc_accept(struct socket *sock, struct socket *new_sock,
1874                       int flags, bool kern)
1875 {
1876         struct sock *sk = sock->sk, *nsk;
1877         DECLARE_WAITQUEUE(wait, current);
1878         struct smc_sock *lsmc;
1879         long timeo;
1880         int rc = 0;
1881
1882         lsmc = smc_sk(sk);
1883         sock_hold(sk); /* sock_put below */
1884         lock_sock(sk);
1885
1886         if (lsmc->sk.sk_state != SMC_LISTEN) {
1887                 rc = -EINVAL;
1888                 release_sock(sk);
1889                 goto out;
1890         }
1891
1892         /* Wait for an incoming connection */
1893         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1894         add_wait_queue_exclusive(sk_sleep(sk), &wait);
1895         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1896                 set_current_state(TASK_INTERRUPTIBLE);
1897                 if (!timeo) {
1898                         rc = -EAGAIN;
1899                         break;
1900                 }
1901                 release_sock(sk);
1902                 timeo = schedule_timeout(timeo);
1903                 /* wakeup by sk_data_ready in smc_listen_work() */
1904                 sched_annotate_sleep();
1905                 lock_sock(sk);
1906                 if (signal_pending(current)) {
1907                         rc = sock_intr_errno(timeo);
1908                         break;
1909                 }
1910         }
1911         set_current_state(TASK_RUNNING);
1912         remove_wait_queue(sk_sleep(sk), &wait);
1913
1914         if (!rc)
1915                 rc = sock_error(nsk);
1916         release_sock(sk);
1917         if (rc)
1918                 goto out;
1919
1920         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1921                 /* wait till data arrives on the socket */
1922                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1923                                                                 MSEC_PER_SEC);
1924                 if (smc_sk(nsk)->use_fallback) {
1925                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1926
1927                         lock_sock(clcsk);
1928                         if (skb_queue_empty(&clcsk->sk_receive_queue))
1929                                 sk_wait_data(clcsk, &timeo, NULL);
1930                         release_sock(clcsk);
1931                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1932                         lock_sock(nsk);
1933                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1934                         release_sock(nsk);
1935                 }
1936         }
1937
1938 out:
1939         sock_put(sk); /* sock_hold above */
1940         return rc;
1941 }
1942
1943 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1944                        int peer)
1945 {
1946         struct smc_sock *smc;
1947
1948         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1949             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1950                 return -ENOTCONN;
1951
1952         smc = smc_sk(sock->sk);
1953
1954         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1955 }
1956
1957 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1958 {
1959         struct sock *sk = sock->sk;
1960         struct smc_sock *smc;
1961         int rc = -EPIPE;
1962
1963         smc = smc_sk(sk);
1964         lock_sock(sk);
1965         if ((sk->sk_state != SMC_ACTIVE) &&
1966             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1967             (sk->sk_state != SMC_INIT))
1968                 goto out;
1969
1970         if (msg->msg_flags & MSG_FASTOPEN) {
1971                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1972                         smc_switch_to_fallback(smc);
1973                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1974                 } else {
1975                         rc = -EINVAL;
1976                         goto out;
1977                 }
1978         }
1979
1980         if (smc->use_fallback)
1981                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1982         else
1983                 rc = smc_tx_sendmsg(smc, msg, len);
1984 out:
1985         release_sock(sk);
1986         return rc;
1987 }
1988
1989 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1990                        int flags)
1991 {
1992         struct sock *sk = sock->sk;
1993         struct smc_sock *smc;
1994         int rc = -ENOTCONN;
1995
1996         smc = smc_sk(sk);
1997         lock_sock(sk);
1998         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1999                 /* socket was connected before, no more data to read */
2000                 rc = 0;
2001                 goto out;
2002         }
2003         if ((sk->sk_state == SMC_INIT) ||
2004             (sk->sk_state == SMC_LISTEN) ||
2005             (sk->sk_state == SMC_CLOSED))
2006                 goto out;
2007
2008         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2009                 rc = 0;
2010                 goto out;
2011         }
2012
2013         if (smc->use_fallback) {
2014                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
2015         } else {
2016                 msg->msg_namelen = 0;
2017                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
2018         }
2019
2020 out:
2021         release_sock(sk);
2022         return rc;
2023 }
2024
2025 static __poll_t smc_accept_poll(struct sock *parent)
2026 {
2027         struct smc_sock *isk = smc_sk(parent);
2028         __poll_t mask = 0;
2029
2030         spin_lock(&isk->accept_q_lock);
2031         if (!list_empty(&isk->accept_q))
2032                 mask = EPOLLIN | EPOLLRDNORM;
2033         spin_unlock(&isk->accept_q_lock);
2034
2035         return mask;
2036 }
2037
2038 static __poll_t smc_poll(struct file *file, struct socket *sock,
2039                              poll_table *wait)
2040 {
2041         struct sock *sk = sock->sk;
2042         struct smc_sock *smc;
2043         __poll_t mask = 0;
2044
2045         if (!sk)
2046                 return EPOLLNVAL;
2047
2048         smc = smc_sk(sock->sk);
2049         if (smc->use_fallback) {
2050                 /* delegate to CLC child sock */
2051                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
2052                 sk->sk_err = smc->clcsock->sk->sk_err;
2053         } else {
2054                 if (sk->sk_state != SMC_CLOSED)
2055                         sock_poll_wait(file, sock, wait);
2056                 if (sk->sk_err)
2057                         mask |= EPOLLERR;
2058                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
2059                     (sk->sk_state == SMC_CLOSED))
2060                         mask |= EPOLLHUP;
2061                 if (sk->sk_state == SMC_LISTEN) {
2062                         /* woken up by sk_data_ready in smc_listen_work() */
2063                         mask |= smc_accept_poll(sk);
2064                 } else if (smc->use_fallback) { /* as result of connect_work()*/
2065                         mask |= smc->clcsock->ops->poll(file, smc->clcsock,
2066                                                            wait);
2067                         sk->sk_err = smc->clcsock->sk->sk_err;
2068                 } else {
2069                         if ((sk->sk_state != SMC_INIT &&
2070                              atomic_read(&smc->conn.sndbuf_space)) ||
2071                             sk->sk_shutdown & SEND_SHUTDOWN) {
2072                                 mask |= EPOLLOUT | EPOLLWRNORM;
2073                         } else {
2074                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
2075                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2076                         }
2077                         if (atomic_read(&smc->conn.bytes_to_rcv))
2078                                 mask |= EPOLLIN | EPOLLRDNORM;
2079                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2080                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2081                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
2082                                 mask |= EPOLLIN;
2083                         if (smc->conn.urg_state == SMC_URG_VALID)
2084                                 mask |= EPOLLPRI;
2085                 }
2086         }
2087
2088         return mask;
2089 }
2090
2091 static int smc_shutdown(struct socket *sock, int how)
2092 {
2093         struct sock *sk = sock->sk;
2094         struct smc_sock *smc;
2095         int rc = -EINVAL;
2096         int rc1 = 0;
2097
2098         smc = smc_sk(sk);
2099
2100         if ((how < SHUT_RD) || (how > SHUT_RDWR))
2101                 return rc;
2102
2103         lock_sock(sk);
2104
2105         rc = -ENOTCONN;
2106         if ((sk->sk_state != SMC_ACTIVE) &&
2107             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
2108             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
2109             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2110             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
2111             (sk->sk_state != SMC_APPFINCLOSEWAIT))
2112                 goto out;
2113         if (smc->use_fallback) {
2114                 rc = kernel_sock_shutdown(smc->clcsock, how);
2115                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
2116                 if (sk->sk_shutdown == SHUTDOWN_MASK)
2117                         sk->sk_state = SMC_CLOSED;
2118                 goto out;
2119         }
2120         switch (how) {
2121         case SHUT_RDWR:         /* shutdown in both directions */
2122                 rc = smc_close_active(smc);
2123                 break;
2124         case SHUT_WR:
2125                 rc = smc_close_shutdown_write(smc);
2126                 break;
2127         case SHUT_RD:
2128                 rc = 0;
2129                 /* nothing more to do because peer is not involved */
2130                 break;
2131         }
2132         if (smc->clcsock)
2133                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
2134         /* map sock_shutdown_cmd constants to sk_shutdown value range */
2135         sk->sk_shutdown |= how + 1;
2136
2137 out:
2138         release_sock(sk);
2139         return rc ? rc : rc1;
2140 }
2141
2142 static int smc_setsockopt(struct socket *sock, int level, int optname,
2143                           sockptr_t optval, unsigned int optlen)
2144 {
2145         struct sock *sk = sock->sk;
2146         struct smc_sock *smc;
2147         int val, rc;
2148
2149         smc = smc_sk(sk);
2150
2151         /* generic setsockopts reaching us here always apply to the
2152          * CLC socket
2153          */
2154         if (unlikely(!smc->clcsock->ops->setsockopt))
2155                 rc = -EOPNOTSUPP;
2156         else
2157                 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
2158                                                    optval, optlen);
2159         if (smc->clcsock->sk->sk_err) {
2160                 sk->sk_err = smc->clcsock->sk->sk_err;
2161                 sk->sk_error_report(sk);
2162         }
2163
2164         if (optlen < sizeof(int))
2165                 return -EINVAL;
2166         if (copy_from_sockptr(&val, optval, sizeof(int)))
2167                 return -EFAULT;
2168
2169         lock_sock(sk);
2170         if (rc || smc->use_fallback)
2171                 goto out;
2172         switch (optname) {
2173         case TCP_ULP:
2174         case TCP_FASTOPEN:
2175         case TCP_FASTOPEN_CONNECT:
2176         case TCP_FASTOPEN_KEY:
2177         case TCP_FASTOPEN_NO_COOKIE:
2178                 /* option not supported by SMC */
2179                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2180                         smc_switch_to_fallback(smc);
2181                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
2182                 } else {
2183                         rc = -EINVAL;
2184                 }
2185                 break;
2186         case TCP_NODELAY:
2187                 if (sk->sk_state != SMC_INIT &&
2188                     sk->sk_state != SMC_LISTEN &&
2189                     sk->sk_state != SMC_CLOSED) {
2190                         if (val)
2191                                 mod_delayed_work(smc->conn.lgr->tx_wq,
2192                                                  &smc->conn.tx_work, 0);
2193                 }
2194                 break;
2195         case TCP_CORK:
2196                 if (sk->sk_state != SMC_INIT &&
2197                     sk->sk_state != SMC_LISTEN &&
2198                     sk->sk_state != SMC_CLOSED) {
2199                         if (!val)
2200                                 mod_delayed_work(smc->conn.lgr->tx_wq,
2201                                                  &smc->conn.tx_work, 0);
2202                 }
2203                 break;
2204         case TCP_DEFER_ACCEPT:
2205                 smc->sockopt_defer_accept = val;
2206                 break;
2207         default:
2208                 break;
2209         }
2210 out:
2211         release_sock(sk);
2212
2213         return rc;
2214 }
2215
2216 static int smc_getsockopt(struct socket *sock, int level, int optname,
2217                           char __user *optval, int __user *optlen)
2218 {
2219         struct smc_sock *smc;
2220
2221         smc = smc_sk(sock->sk);
2222         /* socket options apply to the CLC socket */
2223         if (unlikely(!smc->clcsock->ops->getsockopt))
2224                 return -EOPNOTSUPP;
2225         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
2226                                              optval, optlen);
2227 }
2228
2229 static int smc_ioctl(struct socket *sock, unsigned int cmd,
2230                      unsigned long arg)
2231 {
2232         union smc_host_cursor cons, urg;
2233         struct smc_connection *conn;
2234         struct smc_sock *smc;
2235         int answ;
2236
2237         smc = smc_sk(sock->sk);
2238         conn = &smc->conn;
2239         lock_sock(&smc->sk);
2240         if (smc->use_fallback) {
2241                 if (!smc->clcsock) {
2242                         release_sock(&smc->sk);
2243                         return -EBADF;
2244                 }
2245                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
2246                 release_sock(&smc->sk);
2247                 return answ;
2248         }
2249         switch (cmd) {
2250         case SIOCINQ: /* same as FIONREAD */
2251                 if (smc->sk.sk_state == SMC_LISTEN) {
2252                         release_sock(&smc->sk);
2253                         return -EINVAL;
2254                 }
2255                 if (smc->sk.sk_state == SMC_INIT ||
2256                     smc->sk.sk_state == SMC_CLOSED)
2257                         answ = 0;
2258                 else
2259                         answ = atomic_read(&smc->conn.bytes_to_rcv);
2260                 break;
2261         case SIOCOUTQ:
2262                 /* output queue size (not send + not acked) */
2263                 if (smc->sk.sk_state == SMC_LISTEN) {
2264                         release_sock(&smc->sk);
2265                         return -EINVAL;
2266                 }
2267                 if (smc->sk.sk_state == SMC_INIT ||
2268                     smc->sk.sk_state == SMC_CLOSED)
2269                         answ = 0;
2270                 else
2271                         answ = smc->conn.sndbuf_desc->len -
2272                                         atomic_read(&smc->conn.sndbuf_space);
2273                 break;
2274         case SIOCOUTQNSD:
2275                 /* output queue size (not send only) */
2276                 if (smc->sk.sk_state == SMC_LISTEN) {
2277                         release_sock(&smc->sk);
2278                         return -EINVAL;
2279                 }
2280                 if (smc->sk.sk_state == SMC_INIT ||
2281                     smc->sk.sk_state == SMC_CLOSED)
2282                         answ = 0;
2283                 else
2284                         answ = smc_tx_prepared_sends(&smc->conn);
2285                 break;
2286         case SIOCATMARK:
2287                 if (smc->sk.sk_state == SMC_LISTEN) {
2288                         release_sock(&smc->sk);
2289                         return -EINVAL;
2290                 }
2291                 if (smc->sk.sk_state == SMC_INIT ||
2292                     smc->sk.sk_state == SMC_CLOSED) {
2293                         answ = 0;
2294                 } else {
2295                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
2296                         smc_curs_copy(&urg, &conn->urg_curs, conn);
2297                         answ = smc_curs_diff(conn->rmb_desc->len,
2298                                              &cons, &urg) == 1;
2299                 }
2300                 break;
2301         default:
2302                 release_sock(&smc->sk);
2303                 return -ENOIOCTLCMD;
2304         }
2305         release_sock(&smc->sk);
2306
2307         return put_user(answ, (int __user *)arg);
2308 }
2309
2310 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
2311                             int offset, size_t size, int flags)
2312 {
2313         struct sock *sk = sock->sk;
2314         struct smc_sock *smc;
2315         int rc = -EPIPE;
2316
2317         smc = smc_sk(sk);
2318         lock_sock(sk);
2319         if (sk->sk_state != SMC_ACTIVE) {
2320                 release_sock(sk);
2321                 goto out;
2322         }
2323         release_sock(sk);
2324         if (smc->use_fallback)
2325                 rc = kernel_sendpage(smc->clcsock, page, offset,
2326                                      size, flags);
2327         else
2328                 rc = sock_no_sendpage(sock, page, offset, size, flags);
2329
2330 out:
2331         return rc;
2332 }
2333
2334 /* Map the affected portions of the rmbe into an spd, note the number of bytes
2335  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
2336  * updates till whenever a respective page has been fully processed.
2337  * Note that subsequent recv() calls have to wait till all splice() processing
2338  * completed.
2339  */
2340 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
2341                                struct pipe_inode_info *pipe, size_t len,
2342                                unsigned int flags)
2343 {
2344         struct sock *sk = sock->sk;
2345         struct smc_sock *smc;
2346         int rc = -ENOTCONN;
2347
2348         smc = smc_sk(sk);
2349         lock_sock(sk);
2350         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2351                 /* socket was connected before, no more data to read */
2352                 rc = 0;
2353                 goto out;
2354         }
2355         if (sk->sk_state == SMC_INIT ||
2356             sk->sk_state == SMC_LISTEN ||
2357             sk->sk_state == SMC_CLOSED)
2358                 goto out;
2359
2360         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2361                 rc = 0;
2362                 goto out;
2363         }
2364
2365         if (smc->use_fallback) {
2366                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
2367                                                     pipe, len, flags);
2368         } else {
2369                 if (*ppos) {
2370                         rc = -ESPIPE;
2371                         goto out;
2372                 }
2373                 if (flags & SPLICE_F_NONBLOCK)
2374                         flags = MSG_DONTWAIT;
2375                 else
2376                         flags = 0;
2377                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
2378         }
2379 out:
2380         release_sock(sk);
2381
2382         return rc;
2383 }
2384
2385 /* must look like tcp */
2386 static const struct proto_ops smc_sock_ops = {
2387         .family         = PF_SMC,
2388         .owner          = THIS_MODULE,
2389         .release        = smc_release,
2390         .bind           = smc_bind,
2391         .connect        = smc_connect,
2392         .socketpair     = sock_no_socketpair,
2393         .accept         = smc_accept,
2394         .getname        = smc_getname,
2395         .poll           = smc_poll,
2396         .ioctl          = smc_ioctl,
2397         .listen         = smc_listen,
2398         .shutdown       = smc_shutdown,
2399         .setsockopt     = smc_setsockopt,
2400         .getsockopt     = smc_getsockopt,
2401         .sendmsg        = smc_sendmsg,
2402         .recvmsg        = smc_recvmsg,
2403         .mmap           = sock_no_mmap,
2404         .sendpage       = smc_sendpage,
2405         .splice_read    = smc_splice_read,
2406 };
2407
2408 static int smc_create(struct net *net, struct socket *sock, int protocol,
2409                       int kern)
2410 {
2411         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
2412         struct smc_sock *smc;
2413         struct sock *sk;
2414         int rc;
2415
2416         rc = -ESOCKTNOSUPPORT;
2417         if (sock->type != SOCK_STREAM)
2418                 goto out;
2419
2420         rc = -EPROTONOSUPPORT;
2421         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
2422                 goto out;
2423
2424         rc = -ENOBUFS;
2425         sock->ops = &smc_sock_ops;
2426         sk = smc_sock_alloc(net, sock, protocol);
2427         if (!sk)
2428                 goto out;
2429
2430         /* create internal TCP socket for CLC handshake and fallback */
2431         smc = smc_sk(sk);
2432         smc->use_fallback = false; /* assume rdma capability first */
2433         smc->fallback_rsn = 0;
2434         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2435                               &smc->clcsock);
2436         if (rc) {
2437                 sk_common_release(sk);
2438                 goto out;
2439         }
2440         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2441         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
2442
2443 out:
2444         return rc;
2445 }
2446
2447 static const struct net_proto_family smc_sock_family_ops = {
2448         .family = PF_SMC,
2449         .owner  = THIS_MODULE,
2450         .create = smc_create,
2451 };
2452
2453 unsigned int smc_net_id;
2454
2455 static __net_init int smc_net_init(struct net *net)
2456 {
2457         return smc_pnet_net_init(net);
2458 }
2459
2460 static void __net_exit smc_net_exit(struct net *net)
2461 {
2462         smc_pnet_net_exit(net);
2463 }
2464
2465 static struct pernet_operations smc_net_ops = {
2466         .init = smc_net_init,
2467         .exit = smc_net_exit,
2468         .id   = &smc_net_id,
2469         .size = sizeof(struct smc_net),
2470 };
2471
2472 static int __init smc_init(void)
2473 {
2474         int rc;
2475
2476         rc = register_pernet_subsys(&smc_net_ops);
2477         if (rc)
2478                 return rc;
2479
2480         smc_ism_init();
2481         smc_clc_init();
2482
2483         rc = smc_pnet_init();
2484         if (rc)
2485                 goto out_pernet_subsys;
2486
2487         rc = -ENOMEM;
2488         smc_hs_wq = alloc_workqueue("smc_hs_wq", 0, 0);
2489         if (!smc_hs_wq)
2490                 goto out_pnet;
2491
2492         smc_close_wq = alloc_workqueue("smc_close_wq", 0, 0);
2493         if (!smc_close_wq)
2494                 goto out_alloc_hs_wq;
2495
2496         rc = smc_core_init();
2497         if (rc) {
2498                 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
2499                 goto out_alloc_wqs;
2500         }
2501
2502         rc = smc_llc_init();
2503         if (rc) {
2504                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2505                 goto out_core;
2506         }
2507
2508         rc = smc_cdc_init();
2509         if (rc) {
2510                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2511                 goto out_core;
2512         }
2513
2514         rc = proto_register(&smc_proto, 1);
2515         if (rc) {
2516                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2517                 goto out_core;
2518         }
2519
2520         rc = proto_register(&smc_proto6, 1);
2521         if (rc) {
2522                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2523                 goto out_proto;
2524         }
2525
2526         rc = sock_register(&smc_sock_family_ops);
2527         if (rc) {
2528                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2529                 goto out_proto6;
2530         }
2531         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2532         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2533
2534         rc = smc_ib_register_client();
2535         if (rc) {
2536                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2537                 goto out_sock;
2538         }
2539
2540         static_branch_enable(&tcp_have_smc);
2541         return 0;
2542
2543 out_sock:
2544         sock_unregister(PF_SMC);
2545 out_proto6:
2546         proto_unregister(&smc_proto6);
2547 out_proto:
2548         proto_unregister(&smc_proto);
2549 out_core:
2550         smc_core_exit();
2551 out_alloc_wqs:
2552         destroy_workqueue(smc_close_wq);
2553 out_alloc_hs_wq:
2554         destroy_workqueue(smc_hs_wq);
2555 out_pnet:
2556         smc_pnet_exit();
2557 out_pernet_subsys:
2558         unregister_pernet_subsys(&smc_net_ops);
2559
2560         return rc;
2561 }
2562
2563 static void __exit smc_exit(void)
2564 {
2565         static_branch_disable(&tcp_have_smc);
2566         sock_unregister(PF_SMC);
2567         smc_core_exit();
2568         smc_ib_unregister_client();
2569         destroy_workqueue(smc_close_wq);
2570         destroy_workqueue(smc_hs_wq);
2571         proto_unregister(&smc_proto6);
2572         proto_unregister(&smc_proto);
2573         smc_pnet_exit();
2574         unregister_pernet_subsys(&smc_net_ops);
2575         rcu_barrier();
2576 }
2577
2578 module_init(smc_init);
2579 module_exit(smc_exit);
2580
2581 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2582 MODULE_DESCRIPTION("smc socket address family");
2583 MODULE_LICENSE("GPL");
2584 MODULE_ALIAS_NETPROTO(PF_SMC);