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 <[email protected]>
  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 #include <linux/splice.h>
  31
  32 #include <net/sock.h>
  33 #include <net/tcp.h>
  34 #include <net/smc.h>
  35 #include <asm/ioctls.h>
  36
  37 #include <net/net_namespace.h>
  38 #include <net/netns/generic.h>
  39 #include "smc_netns.h"
  40
  41 #include "smc.h"
  42 #include "smc_clc.h"
  43 #include "smc_llc.h"
  44 #include "smc_cdc.h"
  45 #include "smc_core.h"
  46 #include "smc_ib.h"
  47 #include "smc_ism.h"
  48 #include "smc_pnet.h"
  49 #include "smc_netlink.h"
  50 #include "smc_tx.h"
  51 #include "smc_rx.h"
  52 #include "smc_close.h"
  53 #include "smc_stats.h"
  54 #include "smc_tracepoint.h"
  55 #include "smc_sysctl.h"
  56 #include "smc_loopback.h"
  57 #include "smc_inet.h"
  58
  59 static DEFINE_MUTEX(smc_server_lgr_pending);    /* serialize link group
  60                                                  * creation on server
  61                                                  */
  62 static DEFINE_MUTEX(smc_client_lgr_pending);    /* serialize link group
  63                                                  * creation on client
  64                                                  */
  65
  66 static struct workqueue_struct  *smc_tcp_ls_wq; /* wq for tcp listen work */
  67 struct workqueue_struct *smc_hs_wq;     /* wq for handshake work */
  68 struct workqueue_struct *smc_close_wq;  /* wq for close work */
  69
  70 static void smc_tcp_listen_work(struct work_struct *);
  71 static void smc_connect_work(struct work_struct *);
  72
  73 int smc_nl_dump_hs_limitation(struct sk_buff *skb, struct netlink_callback *cb)
  74 {
  75         struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
  76         void *hdr;
  77
  78         if (cb_ctx->pos[0])
  79                 goto out;
  80
  81         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
  82                           &smc_gen_nl_family, NLM_F_MULTI,
  83                           SMC_NETLINK_DUMP_HS_LIMITATION);
  84         if (!hdr)
  85                 return -ENOMEM;
  86
  87         if (nla_put_u8(skb, SMC_NLA_HS_LIMITATION_ENABLED,
  88                        sock_net(skb->sk)->smc.limit_smc_hs))
  89                 goto err;
  90
  91         genlmsg_end(skb, hdr);
  92         cb_ctx->pos[0] = 1;
  93 out:
  94         return skb->len;
  95 err:
  96         genlmsg_cancel(skb, hdr);
  97         return -EMSGSIZE;
  98 }
  99
 100 int smc_nl_enable_hs_limitation(struct sk_buff *skb, struct genl_info *info)
 101 {
 102         sock_net(skb->sk)->smc.limit_smc_hs = true;
 103         return 0;
 104 }
 105
 106 int smc_nl_disable_hs_limitation(struct sk_buff *skb, struct genl_info *info)
 107 {
 108         sock_net(skb->sk)->smc.limit_smc_hs = false;
 109         return 0;
 110 }
 111
 112 static void smc_set_keepalive(struct sock *sk, int val)
 113 {
 114         struct smc_sock *smc = smc_sk(sk);
 115
 116         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
 117 }
 118
 119 static struct sock *smc_tcp_syn_recv_sock(const struct sock *sk,
 120                                           struct sk_buff *skb,
 121                                           struct request_sock *req,
 122                                           struct dst_entry *dst,
 123                                           struct request_sock *req_unhash,
 124                                           bool *own_req)
 125 {
 126         struct smc_sock *smc;
 127         struct sock *child;
 128
 129         smc = smc_clcsock_user_data(sk);
 130
 131         if (READ_ONCE(sk->sk_ack_backlog) + atomic_read(&smc->queued_smc_hs) >
 132                                 sk->sk_max_ack_backlog)
 133                 goto drop;
 134
 135         if (sk_acceptq_is_full(&smc->sk)) {
 136                 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
 137                 goto drop;
 138         }
 139
 140         /* passthrough to original syn recv sock fct */
 141         child = smc->ori_af_ops->syn_recv_sock(sk, skb, req, dst, req_unhash,
 142                                                own_req);
 143         /* child must not inherit smc or its ops */
 144         if (child) {
 145                 rcu_assign_sk_user_data(child, NULL);
 146
 147                 /* v4-mapped sockets don't inherit parent ops. Don't restore. */
 148                 if (inet_csk(child)->icsk_af_ops == inet_csk(sk)->icsk_af_ops)
 149                         inet_csk(child)->icsk_af_ops = smc->ori_af_ops;
 150         }
 151         return child;
 152
 153 drop:
 154         dst_release(dst);
 155         tcp_listendrop(sk);
 156         return NULL;
 157 }
 158
 159 static bool smc_hs_congested(const struct sock *sk)
 160 {
 161         const struct smc_sock *smc;
 162
 163         smc = smc_clcsock_user_data(sk);
 164
 165         if (!smc)
 166                 return true;
 167
 168         if (workqueue_congested(WORK_CPU_UNBOUND, smc_hs_wq))
 169                 return true;
 170
 171         return false;
 172 }
 173
 174 struct smc_hashinfo smc_v4_hashinfo = {
 175         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
 176 };
 177
 178 struct smc_hashinfo smc_v6_hashinfo = {
 179         .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
 180 };
 181
 182 int smc_hash_sk(struct sock *sk)
 183 {
 184         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
 185         struct hlist_head *head;
 186
 187         head = &h->ht;
 188
 189         write_lock_bh(&h->lock);
 190         sk_add_node(sk, head);
 191         write_unlock_bh(&h->lock);
 192         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 193
 194         return 0;
 195 }
 196
 197 void smc_unhash_sk(struct sock *sk)
 198 {
 199         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
 200
 201         write_lock_bh(&h->lock);
 202         if (sk_del_node_init(sk))
 203                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 204         write_unlock_bh(&h->lock);
 205 }
 206
 207 /* This will be called before user really release sock_lock. So do the
 208  * work which we didn't do because of user hold the sock_lock in the
 209  * BH context
 210  */
 211 void smc_release_cb(struct sock *sk)
 212 {
 213         struct smc_sock *smc = smc_sk(sk);
 214
 215         if (smc->conn.tx_in_release_sock) {
 216                 smc_tx_pending(&smc->conn);
 217                 smc->conn.tx_in_release_sock = false;
 218         }
 219 }
 220
 221 struct proto smc_proto = {
 222         .name           = "SMC",
 223         .owner          = THIS_MODULE,
 224         .keepalive      = smc_set_keepalive,
 225         .hash           = smc_hash_sk,
 226         .unhash         = smc_unhash_sk,
 227         .release_cb     = smc_release_cb,
 228         .obj_size       = sizeof(struct smc_sock),
 229         .h.smc_hash     = &smc_v4_hashinfo,
 230         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 231 };
 232 EXPORT_SYMBOL_GPL(smc_proto);
 233
 234 struct proto smc_proto6 = {
 235         .name           = "SMC6",
 236         .owner          = THIS_MODULE,
 237         .keepalive      = smc_set_keepalive,
 238         .hash           = smc_hash_sk,
 239         .unhash         = smc_unhash_sk,
 240         .release_cb     = smc_release_cb,
 241         .obj_size       = sizeof(struct smc_sock),
 242         .h.smc_hash     = &smc_v6_hashinfo,
 243         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 244 };
 245 EXPORT_SYMBOL_GPL(smc_proto6);
 246
 247 static void smc_fback_restore_callbacks(struct smc_sock *smc)
 248 {
 249         struct sock *clcsk = smc->clcsock->sk;
 250
 251         write_lock_bh(&clcsk->sk_callback_lock);
 252         clcsk->sk_user_data = NULL;
 253
 254         smc_clcsock_restore_cb(&clcsk->sk_state_change, &smc->clcsk_state_change);
 255         smc_clcsock_restore_cb(&clcsk->sk_data_ready, &smc->clcsk_data_ready);
 256         smc_clcsock_restore_cb(&clcsk->sk_write_space, &smc->clcsk_write_space);
 257         smc_clcsock_restore_cb(&clcsk->sk_error_report, &smc->clcsk_error_report);
 258
 259         write_unlock_bh(&clcsk->sk_callback_lock);
 260 }
 261
 262 static void smc_restore_fallback_changes(struct smc_sock *smc)
 263 {
 264         if (smc->clcsock->file) { /* non-accepted sockets have no file yet */
 265                 smc->clcsock->file->private_data = smc->sk.sk_socket;
 266                 smc->clcsock->file = NULL;
 267                 smc_fback_restore_callbacks(smc);
 268         }
 269 }
 270
 271 static int __smc_release(struct smc_sock *smc)
 272 {
 273         struct sock *sk = &smc->sk;
 274         int rc = 0;
 275
 276         if (!smc->use_fallback) {
 277                 rc = smc_close_active(smc);
 278                 smc_sock_set_flag(sk, SOCK_DEAD);
 279                 sk->sk_shutdown |= SHUTDOWN_MASK;
 280         } else {
 281                 if (sk->sk_state != SMC_CLOSED) {
 282                         if (sk->sk_state != SMC_LISTEN &&
 283                             sk->sk_state != SMC_INIT)
 284                                 sock_put(sk); /* passive closing */
 285                         if (sk->sk_state == SMC_LISTEN) {
 286                                 /* wake up clcsock accept */
 287                                 rc = kernel_sock_shutdown(smc->clcsock,
 288                                                           SHUT_RDWR);
 289                         }
 290                         sk->sk_state = SMC_CLOSED;
 291                         sk->sk_state_change(sk);
 292                 }
 293                 smc_restore_fallback_changes(smc);
 294         }
 295
 296         sk->sk_prot->unhash(sk);
 297
 298         if (sk->sk_state == SMC_CLOSED) {
 299                 if (smc->clcsock) {
 300                         release_sock(sk);
 301                         smc_clcsock_release(smc);
 302                         lock_sock(sk);
 303                 }
 304                 if (!smc->use_fallback)
 305                         smc_conn_free(&smc->conn);
 306         }
 307
 308         return rc;
 309 }
 310
 311 int smc_release(struct socket *sock)
 312 {
 313         struct sock *sk = sock->sk;
 314         struct smc_sock *smc;
 315         int old_state, rc = 0;
 316
 317         if (!sk)
 318                 goto out;
 319
 320         sock_hold(sk); /* sock_put below */
 321         smc = smc_sk(sk);
 322
 323         old_state = sk->sk_state;
 324
 325         /* cleanup for a dangling non-blocking connect */
 326         if (smc->connect_nonblock && old_state == SMC_INIT)
 327                 tcp_abort(smc->clcsock->sk, ECONNABORTED);
 328
 329         if (cancel_work_sync(&smc->connect_work))
 330                 sock_put(&smc->sk); /* sock_hold in smc_connect for passive closing */
 331
 332         if (sk->sk_state == SMC_LISTEN)
 333                 /* smc_close_non_accepted() is called and acquires
 334                  * sock lock for child sockets again
 335                  */
 336                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 337         else
 338                 lock_sock(sk);
 339
 340         if (old_state == SMC_INIT && sk->sk_state == SMC_ACTIVE &&
 341             !smc->use_fallback)
 342                 smc_close_active_abort(smc);
 343
 344         rc = __smc_release(smc);
 345
 346         /* detach socket */
 347         sock_orphan(sk);
 348         sock->sk = NULL;
 349         release_sock(sk);
 350
 351         sock_put(sk); /* sock_hold above */
 352         sock_put(sk); /* final sock_put */
 353 out:
 354         return rc;
 355 }
 356
 357 static void smc_destruct(struct sock *sk)
 358 {
 359         if (sk->sk_state != SMC_CLOSED)
 360                 return;
 361         if (!sock_flag(sk, SOCK_DEAD))
 362                 return;
 363 }
 364
 365 void smc_sk_init(struct net *net, struct sock *sk, int protocol)
 366 {
 367         struct smc_sock *smc = smc_sk(sk);
 368
 369         sk->sk_state = SMC_INIT;
 370         sk->sk_destruct = smc_destruct;
 371         sk->sk_protocol = protocol;
 372         WRITE_ONCE(sk->sk_sndbuf, 2 * READ_ONCE(net->smc.sysctl_wmem));
 373         WRITE_ONCE(sk->sk_rcvbuf, 2 * READ_ONCE(net->smc.sysctl_rmem));
 374         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 375         INIT_WORK(&smc->connect_work, smc_connect_work);
 376         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 377         INIT_LIST_HEAD(&smc->accept_q);
 378         spin_lock_init(&smc->accept_q_lock);
 379         spin_lock_init(&smc->conn.send_lock);
 380         sk->sk_prot->hash(sk);
 381         mutex_init(&smc->clcsock_release_lock);
 382         smc_init_saved_callbacks(smc);
 383         smc->limit_smc_hs = net->smc.limit_smc_hs;
 384         smc->use_fallback = false; /* assume rdma capability first */
 385         smc->fallback_rsn = 0;
 386 }
 387
 388 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 389                                    int protocol)
 390 {
 391         struct proto *prot;
 392         struct sock *sk;
 393
 394         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
 395         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
 396         if (!sk)
 397                 return NULL;
 398
 399         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 400         smc_sk_init(net, sk, protocol);
 401
 402         return sk;
 403 }
 404
 405 int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 406              int addr_len)
 407 {
 408         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 409         struct sock *sk = sock->sk;
 410         struct smc_sock *smc;
 411         int rc;
 412
 413         smc = smc_sk(sk);
 414
 415         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 416         rc = -EINVAL;
 417         if (addr_len < sizeof(struct sockaddr_in))
 418                 goto out;
 419
 420         rc = -EAFNOSUPPORT;
 421         if (addr->sin_family != AF_INET &&
 422             addr->sin_family != AF_INET6 &&
 423             addr->sin_family != AF_UNSPEC)
 424                 goto out;
 425         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 426         if (addr->sin_family == AF_UNSPEC &&
 427             addr->sin_addr.s_addr != htonl(INADDR_ANY))
 428                 goto out;
 429
 430         lock_sock(sk);
 431
 432         /* Check if socket is already active */
 433         rc = -EINVAL;
 434         if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
 435                 goto out_rel;
 436
 437         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 438         smc->clcsock->sk->sk_reuseport = sk->sk_reuseport;
 439         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 440
 441 out_rel:
 442         release_sock(sk);
 443 out:
 444         return rc;
 445 }
 446
 447 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 448  * clc socket (since smc is not called for these options from net/core)
 449  */
 450
 451 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 452                              (1UL << SOCK_KEEPOPEN) | \
 453                              (1UL << SOCK_LINGER) | \
 454                              (1UL << SOCK_BROADCAST) | \
 455                              (1UL << SOCK_TIMESTAMP) | \
 456                              (1UL << SOCK_DBG) | \
 457                              (1UL << SOCK_RCVTSTAMP) | \
 458                              (1UL << SOCK_RCVTSTAMPNS) | \
 459                              (1UL << SOCK_LOCALROUTE) | \
 460                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 461                              (1UL << SOCK_RXQ_OVFL) | \
 462                              (1UL << SOCK_WIFI_STATUS) | \
 463                              (1UL << SOCK_NOFCS) | \
 464                              (1UL << SOCK_FILTER_LOCKED) | \
 465                              (1UL << SOCK_TSTAMP_NEW))
 466
 467 /* if set, use value set by setsockopt() - else use IPv4 or SMC sysctl value */
 468 static void smc_adjust_sock_bufsizes(struct sock *nsk, struct sock *osk,
 469                                      unsigned long mask)
 470 {
 471         nsk->sk_userlocks = osk->sk_userlocks;
 472         if (osk->sk_userlocks & SOCK_SNDBUF_LOCK)
 473                 nsk->sk_sndbuf = osk->sk_sndbuf;
 474         if (osk->sk_userlocks & SOCK_RCVBUF_LOCK)
 475                 nsk->sk_rcvbuf = osk->sk_rcvbuf;
 476 }
 477
 478 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 479                                    unsigned long mask)
 480 {
 481         /* options we don't get control via setsockopt for */
 482         nsk->sk_type = osk->sk_type;
 483         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 484         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 485         nsk->sk_mark = READ_ONCE(osk->sk_mark);
 486         nsk->sk_priority = READ_ONCE(osk->sk_priority);
 487         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 488         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 489         nsk->sk_err = osk->sk_err;
 490
 491         nsk->sk_flags &= ~mask;
 492         nsk->sk_flags |= osk->sk_flags & mask;
 493
 494         smc_adjust_sock_bufsizes(nsk, osk, mask);
 495 }
 496
 497 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 498 {
 499         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 500 }
 501
 502 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 503                              (1UL << SOCK_KEEPOPEN) | \
 504                              (1UL << SOCK_LINGER) | \
 505                              (1UL << SOCK_DBG))
 506 /* copy only settings and flags relevant for smc from clc to smc socket */
 507 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 508 {
 509         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 510 }
 511
 512 /* register the new vzalloced sndbuf on all links */
 513 static int smcr_lgr_reg_sndbufs(struct smc_link *link,
 514                                 struct smc_buf_desc *snd_desc)
 515 {
 516         struct smc_link_group *lgr = link->lgr;
 517         int i, rc = 0;
 518
 519         if (!snd_desc->is_vm)
 520                 return -EINVAL;
 521
 522         /* protect against parallel smcr_link_reg_buf() */
 523         down_write(&lgr->llc_conf_mutex);
 524         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 525                 if (!smc_link_active(&lgr->lnk[i]))
 526                         continue;
 527                 rc = smcr_link_reg_buf(&lgr->lnk[i], snd_desc);
 528                 if (rc)
 529                         break;
 530         }
 531         up_write(&lgr->llc_conf_mutex);
 532         return rc;
 533 }
 534
 535 /* register the new rmb on all links */
 536 static int smcr_lgr_reg_rmbs(struct smc_link *link,
 537                              struct smc_buf_desc *rmb_desc)
 538 {
 539         struct smc_link_group *lgr = link->lgr;
 540         bool do_slow = false;
 541         int i, rc = 0;
 542
 543         rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
 544         if (rc)
 545                 return rc;
 546
 547         down_read(&lgr->llc_conf_mutex);
 548         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 549                 if (!smc_link_active(&lgr->lnk[i]))
 550                         continue;
 551                 if (!rmb_desc->is_reg_mr[link->link_idx]) {
 552                         up_read(&lgr->llc_conf_mutex);
 553                         goto slow_path;
 554                 }
 555         }
 556         /* mr register already */
 557         goto fast_path;
 558 slow_path:
 559         do_slow = true;
 560         /* protect against parallel smc_llc_cli_rkey_exchange() and
 561          * parallel smcr_link_reg_buf()
 562          */
 563         down_write(&lgr->llc_conf_mutex);
 564         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 565                 if (!smc_link_active(&lgr->lnk[i]))
 566                         continue;
 567                 rc = smcr_link_reg_buf(&lgr->lnk[i], rmb_desc);
 568                 if (rc)
 569                         goto out;
 570         }
 571 fast_path:
 572         /* exchange confirm_rkey msg with peer */
 573         rc = smc_llc_do_confirm_rkey(link, rmb_desc);
 574         if (rc) {
 575                 rc = -EFAULT;
 576                 goto out;
 577         }
 578         rmb_desc->is_conf_rkey = true;
 579 out:
 580         do_slow ? up_write(&lgr->llc_conf_mutex) : up_read(&lgr->llc_conf_mutex);
 581         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
 582         return rc;
 583 }
 584
 585 static int smcr_clnt_conf_first_link(struct smc_sock *smc)
 586 {
 587         struct smc_link *link = smc->conn.lnk;
 588         struct smc_llc_qentry *qentry;
 589         int rc;
 590
 591         /* Receive CONFIRM LINK request from server over RoCE fabric.
 592          * Increasing the client's timeout by twice as much as the server's
 593          * timeout by default can temporarily avoid decline messages of
 594          * both sides crossing or colliding
 595          */
 596         qentry = smc_llc_wait(link->lgr, NULL, 2 * SMC_LLC_WAIT_TIME,
 597                               SMC_LLC_CONFIRM_LINK);
 598         if (!qentry) {
 599                 struct smc_clc_msg_decline dclc;
 600
 601                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 602                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 603                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
 604         }
 605         smc_llc_save_peer_uid(qentry);
 606         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
 607         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
 608         if (rc)
 609                 return SMC_CLC_DECL_RMBE_EC;
 610
 611         rc = smc_ib_modify_qp_rts(link);
 612         if (rc)
 613                 return SMC_CLC_DECL_ERR_RDYLNK;
 614
 615         smc_wr_remember_qp_attr(link);
 616
 617         /* reg the sndbuf if it was vzalloced */
 618         if (smc->conn.sndbuf_desc->is_vm) {
 619                 if (smcr_link_reg_buf(link, smc->conn.sndbuf_desc))
 620                         return SMC_CLC_DECL_ERR_REGBUF;
 621         }
 622
 623         /* reg the rmb */
 624         if (smcr_link_reg_buf(link, smc->conn.rmb_desc))
 625                 return SMC_CLC_DECL_ERR_REGBUF;
 626
 627         /* confirm_rkey is implicit on 1st contact */
 628         smc->conn.rmb_desc->is_conf_rkey = true;
 629
 630         /* send CONFIRM LINK response over RoCE fabric */
 631         rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
 632         if (rc < 0)
 633                 return SMC_CLC_DECL_TIMEOUT_CL;
 634
 635         smc_llc_link_active(link);
 636         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
 637
 638         if (link->lgr->max_links > 1) {
 639                 /* optional 2nd link, receive ADD LINK request from server */
 640                 qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
 641                                       SMC_LLC_ADD_LINK);
 642                 if (!qentry) {
 643                         struct smc_clc_msg_decline dclc;
 644
 645                         rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 646                                               SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 647                         if (rc == -EAGAIN)
 648                                 rc = 0; /* no DECLINE received, go with one link */
 649                         return rc;
 650                 }
 651                 smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
 652                 smc_llc_cli_add_link(link, qentry);
 653         }
 654         return 0;
 655 }
 656
 657 static bool smc_isascii(char *hostname)
 658 {
 659         int i;
 660
 661         for (i = 0; i < SMC_MAX_HOSTNAME_LEN; i++)
 662                 if (!isascii(hostname[i]))
 663                         return false;
 664         return true;
 665 }
 666
 667 static void smc_conn_save_peer_info_fce(struct smc_sock *smc,
 668                                         struct smc_clc_msg_accept_confirm *clc)
 669 {
 670         struct smc_clc_first_contact_ext *fce;
 671         int clc_v2_len;
 672
 673         if (clc->hdr.version == SMC_V1 ||
 674             !(clc->hdr.typev2 & SMC_FIRST_CONTACT_MASK))
 675                 return;
 676
 677         if (smc->conn.lgr->is_smcd) {
 678                 memcpy(smc->conn.lgr->negotiated_eid, clc->d1.eid,
 679                        SMC_MAX_EID_LEN);
 680                 clc_v2_len = offsetofend(struct smc_clc_msg_accept_confirm, d1);
 681         } else {
 682                 memcpy(smc->conn.lgr->negotiated_eid, clc->r1.eid,
 683                        SMC_MAX_EID_LEN);
 684                 clc_v2_len = offsetofend(struct smc_clc_msg_accept_confirm, r1);
 685         }
 686         fce = (struct smc_clc_first_contact_ext *)(((u8 *)clc) + clc_v2_len);
 687         smc->conn.lgr->peer_os = fce->os_type;
 688         smc->conn.lgr->peer_smc_release = fce->release;
 689         if (smc_isascii(fce->hostname))
 690                 memcpy(smc->conn.lgr->peer_hostname, fce->hostname,
 691                        SMC_MAX_HOSTNAME_LEN);
 692 }
 693
 694 static void smcr_conn_save_peer_info(struct smc_sock *smc,
 695                                      struct smc_clc_msg_accept_confirm *clc)
 696 {
 697         int bufsize = smc_uncompress_bufsize(clc->r0.rmbe_size);
 698
 699         smc->conn.peer_rmbe_idx = clc->r0.rmbe_idx;
 700         smc->conn.local_tx_ctrl.token = ntohl(clc->r0.rmbe_alert_token);
 701         smc->conn.peer_rmbe_size = bufsize;
 702         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 703         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
 704 }
 705
 706 static void smcd_conn_save_peer_info(struct smc_sock *smc,
 707                                      struct smc_clc_msg_accept_confirm *clc)
 708 {
 709         int bufsize = smc_uncompress_bufsize(clc->d0.dmbe_size);
 710
 711         smc->conn.peer_rmbe_idx = clc->d0.dmbe_idx;
 712         smc->conn.peer_token = ntohll(clc->d0.token);
 713         /* msg header takes up space in the buffer */
 714         smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
 715         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 716         smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
 717 }
 718
 719 static void smc_conn_save_peer_info(struct smc_sock *smc,
 720                                     struct smc_clc_msg_accept_confirm *clc)
 721 {
 722         if (smc->conn.lgr->is_smcd)
 723                 smcd_conn_save_peer_info(smc, clc);
 724         else
 725                 smcr_conn_save_peer_info(smc, clc);
 726         smc_conn_save_peer_info_fce(smc, clc);
 727 }
 728
 729 static void smc_link_save_peer_info(struct smc_link *link,
 730                                     struct smc_clc_msg_accept_confirm *clc,
 731                                     struct smc_init_info *ini)
 732 {
 733         link->peer_qpn = ntoh24(clc->r0.qpn);
 734         memcpy(link->peer_gid, ini->peer_gid, SMC_GID_SIZE);
 735         memcpy(link->peer_mac, ini->peer_mac, sizeof(link->peer_mac));
 736         link->peer_psn = ntoh24(clc->r0.psn);
 737         link->peer_mtu = clc->r0.qp_mtu;
 738 }
 739
 740 static void smc_stat_inc_fback_rsn_cnt(struct smc_sock *smc,
 741                                        struct smc_stats_fback *fback_arr)
 742 {
 743         int cnt;
 744
 745         for (cnt = 0; cnt < SMC_MAX_FBACK_RSN_CNT; cnt++) {
 746                 if (fback_arr[cnt].fback_code == smc->fallback_rsn) {
 747                         fback_arr[cnt].count++;
 748                         break;
 749                 }
 750                 if (!fback_arr[cnt].fback_code) {
 751                         fback_arr[cnt].fback_code = smc->fallback_rsn;
 752                         fback_arr[cnt].count++;
 753                         break;
 754                 }
 755         }
 756 }
 757
 758 static void smc_stat_fallback(struct smc_sock *smc)
 759 {
 760         struct net *net = sock_net(&smc->sk);
 761
 762         mutex_lock(&net->smc.mutex_fback_rsn);
 763         if (smc->listen_smc) {
 764                 smc_stat_inc_fback_rsn_cnt(smc, net->smc.fback_rsn->srv);
 765                 net->smc.fback_rsn->srv_fback_cnt++;
 766         } else {
 767                 smc_stat_inc_fback_rsn_cnt(smc, net->smc.fback_rsn->clnt);
 768                 net->smc.fback_rsn->clnt_fback_cnt++;
 769         }
 770         mutex_unlock(&net->smc.mutex_fback_rsn);
 771 }
 772
 773 /* must be called under rcu read lock */
 774 static void smc_fback_wakeup_waitqueue(struct smc_sock *smc, void *key)
 775 {
 776         struct socket_wq *wq;
 777         __poll_t flags;
 778
 779         wq = rcu_dereference(smc->sk.sk_wq);
 780         if (!skwq_has_sleeper(wq))
 781                 return;
 782
 783         /* wake up smc sk->sk_wq */
 784         if (!key) {
 785                 /* sk_state_change */
 786                 wake_up_interruptible_all(&wq->wait);
 787         } else {
 788                 flags = key_to_poll(key);
 789                 if (flags & (EPOLLIN | EPOLLOUT))
 790                         /* sk_data_ready or sk_write_space */
 791                         wake_up_interruptible_sync_poll(&wq->wait, flags);
 792                 else if (flags & EPOLLERR)
 793                         /* sk_error_report */
 794                         wake_up_interruptible_poll(&wq->wait, flags);
 795         }
 796 }
 797
 798 static int smc_fback_mark_woken(wait_queue_entry_t *wait,
 799                                 unsigned int mode, int sync, void *key)
 800 {
 801         struct smc_mark_woken *mark =
 802                 container_of(wait, struct smc_mark_woken, wait_entry);
 803
 804         mark->woken = true;
 805         mark->key = key;
 806         return 0;
 807 }
 808
 809 static void smc_fback_forward_wakeup(struct smc_sock *smc, struct sock *clcsk,
 810                                      void (*clcsock_callback)(struct sock *sk))
 811 {
 812         struct smc_mark_woken mark = { .woken = false };
 813         struct socket_wq *wq;
 814
 815         init_waitqueue_func_entry(&mark.wait_entry,
 816                                   smc_fback_mark_woken);
 817         rcu_read_lock();
 818         wq = rcu_dereference(clcsk->sk_wq);
 819         if (!wq)
 820                 goto out;
 821         add_wait_queue(sk_sleep(clcsk), &mark.wait_entry);
 822         clcsock_callback(clcsk);
 823         remove_wait_queue(sk_sleep(clcsk), &mark.wait_entry);
 824
 825         if (mark.woken)
 826                 smc_fback_wakeup_waitqueue(smc, mark.key);
 827 out:
 828         rcu_read_unlock();
 829 }
 830
 831 static void smc_fback_state_change(struct sock *clcsk)
 832 {
 833         struct smc_sock *smc;
 834
 835         read_lock_bh(&clcsk->sk_callback_lock);
 836         smc = smc_clcsock_user_data(clcsk);
 837         if (smc)
 838                 smc_fback_forward_wakeup(smc, clcsk,
 839                                          smc->clcsk_state_change);
 840         read_unlock_bh(&clcsk->sk_callback_lock);
 841 }
 842
 843 static void smc_fback_data_ready(struct sock *clcsk)
 844 {
 845         struct smc_sock *smc;
 846
 847         read_lock_bh(&clcsk->sk_callback_lock);
 848         smc = smc_clcsock_user_data(clcsk);
 849         if (smc)
 850                 smc_fback_forward_wakeup(smc, clcsk,
 851                                          smc->clcsk_data_ready);
 852         read_unlock_bh(&clcsk->sk_callback_lock);
 853 }
 854
 855 static void smc_fback_write_space(struct sock *clcsk)
 856 {
 857         struct smc_sock *smc;
 858
 859         read_lock_bh(&clcsk->sk_callback_lock);
 860         smc = smc_clcsock_user_data(clcsk);
 861         if (smc)
 862                 smc_fback_forward_wakeup(smc, clcsk,
 863                                          smc->clcsk_write_space);
 864         read_unlock_bh(&clcsk->sk_callback_lock);
 865 }
 866
 867 static void smc_fback_error_report(struct sock *clcsk)
 868 {
 869         struct smc_sock *smc;
 870
 871         read_lock_bh(&clcsk->sk_callback_lock);
 872         smc = smc_clcsock_user_data(clcsk);
 873         if (smc)
 874                 smc_fback_forward_wakeup(smc, clcsk,
 875                                          smc->clcsk_error_report);
 876         read_unlock_bh(&clcsk->sk_callback_lock);
 877 }
 878
 879 static void smc_fback_replace_callbacks(struct smc_sock *smc)
 880 {
 881         struct sock *clcsk = smc->clcsock->sk;
 882
 883         write_lock_bh(&clcsk->sk_callback_lock);
 884         clcsk->sk_user_data = (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
 885
 886         smc_clcsock_replace_cb(&clcsk->sk_state_change, smc_fback_state_change,
 887                                &smc->clcsk_state_change);
 888         smc_clcsock_replace_cb(&clcsk->sk_data_ready, smc_fback_data_ready,
 889                                &smc->clcsk_data_ready);
 890         smc_clcsock_replace_cb(&clcsk->sk_write_space, smc_fback_write_space,
 891                                &smc->clcsk_write_space);
 892         smc_clcsock_replace_cb(&clcsk->sk_error_report, smc_fback_error_report,
 893                                &smc->clcsk_error_report);
 894
 895         write_unlock_bh(&clcsk->sk_callback_lock);
 896 }
 897
 898 static int smc_switch_to_fallback(struct smc_sock *smc, int reason_code)
 899 {
 900         int rc = 0;
 901
 902         mutex_lock(&smc->clcsock_release_lock);
 903         if (!smc->clcsock) {
 904                 rc = -EBADF;
 905                 goto out;
 906         }
 907
 908         smc->use_fallback = true;
 909         smc->fallback_rsn = reason_code;
 910         smc_stat_fallback(smc);
 911         trace_smc_switch_to_fallback(smc, reason_code);
 912         if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
 913                 smc->clcsock->file = smc->sk.sk_socket->file;
 914                 smc->clcsock->file->private_data = smc->clcsock;
 915                 smc->clcsock->wq.fasync_list =
 916                         smc->sk.sk_socket->wq.fasync_list;
 917                 smc->sk.sk_socket->wq.fasync_list = NULL;
 918
 919                 /* There might be some wait entries remaining
 920                  * in smc sk->sk_wq and they should be woken up
 921                  * as clcsock's wait queue is woken up.
 922                  */
 923                 smc_fback_replace_callbacks(smc);
 924         }
 925 out:
 926         mutex_unlock(&smc->clcsock_release_lock);
 927         return rc;
 928 }
 929
 930 /* fall back during connect */
 931 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
 932 {
 933         struct net *net = sock_net(&smc->sk);
 934         int rc = 0;
 935
 936         rc = smc_switch_to_fallback(smc, reason_code);
 937         if (rc) { /* fallback fails */
 938                 this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
 939                 if (smc->sk.sk_state == SMC_INIT)
 940                         sock_put(&smc->sk); /* passive closing */
 941                 return rc;
 942         }
 943         smc_copy_sock_settings_to_clc(smc);
 944         smc->connect_nonblock = 0;
 945         if (smc->sk.sk_state == SMC_INIT)
 946                 smc->sk.sk_state = SMC_ACTIVE;
 947         return 0;
 948 }
 949
 950 /* decline and fall back during connect */
 951 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code,
 952                                         u8 version)
 953 {
 954         struct net *net = sock_net(&smc->sk);
 955         int rc;
 956
 957         if (reason_code < 0) { /* error, fallback is not possible */
 958                 this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
 959                 if (smc->sk.sk_state == SMC_INIT)
 960                         sock_put(&smc->sk); /* passive closing */
 961                 return reason_code;
 962         }
 963         if (reason_code != SMC_CLC_DECL_PEERDECL) {
 964                 rc = smc_clc_send_decline(smc, reason_code, version);
 965                 if (rc < 0) {
 966                         this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
 967                         if (smc->sk.sk_state == SMC_INIT)
 968                                 sock_put(&smc->sk); /* passive closing */
 969                         return rc;
 970                 }
 971         }
 972         return smc_connect_fallback(smc, reason_code);
 973 }
 974
 975 static void smc_conn_abort(struct smc_sock *smc, int local_first)
 976 {
 977         struct smc_connection *conn = &smc->conn;
 978         struct smc_link_group *lgr = conn->lgr;
 979         bool lgr_valid = false;
 980
 981         if (smc_conn_lgr_valid(conn))
 982                 lgr_valid = true;
 983
 984         smc_conn_free(conn);
 985         if (local_first && lgr_valid)
 986                 smc_lgr_cleanup_early(lgr);
 987 }
 988
 989 /* check if there is a rdma device available for this connection. */
 990 /* called for connect and listen */
 991 static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
 992 {
 993         /* PNET table look up: search active ib_device and port
 994          * within same PNETID that also contains the ethernet device
 995          * used for the internal TCP socket
 996          */
 997         smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
 998         if (!ini->check_smcrv2 && !ini->ib_dev)
 999                 return SMC_CLC_DECL_NOSMCRDEV;
1000         if (ini->check_smcrv2 && !ini->smcrv2.ib_dev_v2)
1001                 return SMC_CLC_DECL_NOSMCRDEV;
1002         return 0;
1003 }
1004
1005 /* check if there is an ISM device available for this connection. */
1006 /* called for connect and listen */
1007 static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
1008 {
1009         /* Find ISM device with same PNETID as connecting interface  */
1010         smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
1011         if (!ini->ism_dev[0])
1012                 return SMC_CLC_DECL_NOSMCDDEV;
1013         else
1014                 ini->ism_chid[0] = smc_ism_get_chid(ini->ism_dev[0]);
1015         return 0;
1016 }
1017
1018 /* is chid unique for the ism devices that are already determined? */
1019 static bool smc_find_ism_v2_is_unique_chid(u16 chid, struct smc_init_info *ini,
1020                                            int cnt)
1021 {
1022         int i = (!ini->ism_dev[0]) ? 1 : 0;
1023
1024         for (; i < cnt; i++)
1025                 if (ini->ism_chid[i] == chid)
1026                         return false;
1027         return true;
1028 }
1029
1030 /* determine possible V2 ISM devices (either without PNETID or with PNETID plus
1031  * PNETID matching net_device)
1032  */
1033 static int smc_find_ism_v2_device_clnt(struct smc_sock *smc,
1034                                        struct smc_init_info *ini)
1035 {
1036         int rc = SMC_CLC_DECL_NOSMCDDEV;
1037         struct smcd_dev *smcd;
1038         int i = 1, entry = 1;
1039         bool is_emulated;
1040         u16 chid;
1041
1042         if (smcd_indicated(ini->smc_type_v1))
1043                 rc = 0;         /* already initialized for V1 */
1044         mutex_lock(&smcd_dev_list.mutex);
1045         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1046                 if (smcd->going_away || smcd == ini->ism_dev[0])
1047                         continue;
1048                 chid = smc_ism_get_chid(smcd);
1049                 if (!smc_find_ism_v2_is_unique_chid(chid, ini, i))
1050                         continue;
1051                 is_emulated = __smc_ism_is_emulated(chid);
1052                 if (!smc_pnet_is_pnetid_set(smcd->pnetid) ||
1053                     smc_pnet_is_ndev_pnetid(sock_net(&smc->sk), smcd->pnetid)) {
1054                         if (is_emulated && entry == SMCD_CLC_MAX_V2_GID_ENTRIES)
1055                                 /* It's the last GID-CHID entry left in CLC
1056                                  * Proposal SMC-Dv2 extension, but an Emulated-
1057                                  * ISM device will take two entries. So give
1058                                  * up it and try the next potential ISM device.
1059                                  */
1060                                 continue;
1061                         ini->ism_dev[i] = smcd;
1062                         ini->ism_chid[i] = chid;
1063                         ini->is_smcd = true;
1064                         rc = 0;
1065                         i++;
1066                         entry = is_emulated ? entry + 2 : entry + 1;
1067                         if (entry > SMCD_CLC_MAX_V2_GID_ENTRIES)
1068                                 break;
1069                 }
1070         }
1071         mutex_unlock(&smcd_dev_list.mutex);
1072         ini->ism_offered_cnt = i - 1;
1073         if (!ini->ism_dev[0] && !ini->ism_dev[1])
1074                 ini->smcd_version = 0;
1075
1076         return rc;
1077 }
1078
1079 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
1080 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
1081                                       struct smc_init_info *ini)
1082 {
1083         if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev[0], ini->vlan_id))
1084                 return SMC_CLC_DECL_ISMVLANERR;
1085         return 0;
1086 }
1087
1088 static int smc_find_proposal_devices(struct smc_sock *smc,
1089                                      struct smc_init_info *ini)
1090 {
1091         int rc = 0;
1092
1093         /* check if there is an ism device available */
1094         if (!(ini->smcd_version & SMC_V1) ||
1095             smc_find_ism_device(smc, ini) ||
1096             smc_connect_ism_vlan_setup(smc, ini))
1097                 ini->smcd_version &= ~SMC_V1;
1098         /* else ISM V1 is supported for this connection */
1099
1100         /* check if there is an rdma device available */
1101         if (!(ini->smcr_version & SMC_V1) ||
1102             smc_find_rdma_device(smc, ini))
1103                 ini->smcr_version &= ~SMC_V1;
1104         /* else RDMA is supported for this connection */
1105
1106         ini->smc_type_v1 = smc_indicated_type(ini->smcd_version & SMC_V1,
1107                                               ini->smcr_version & SMC_V1);
1108
1109         /* check if there is an ism v2 device available */
1110         if (!(ini->smcd_version & SMC_V2) ||
1111             !smc_ism_is_v2_capable() ||
1112             smc_find_ism_v2_device_clnt(smc, ini))
1113                 ini->smcd_version &= ~SMC_V2;
1114
1115         /* check if there is an rdma v2 device available */
1116         ini->check_smcrv2 = true;
1117         ini->smcrv2.saddr = smc->clcsock->sk->sk_rcv_saddr;
1118         if (!(ini->smcr_version & SMC_V2) ||
1119             smc->clcsock->sk->sk_family != AF_INET ||
1120             !smc_clc_ueid_count() ||
1121             smc_find_rdma_device(smc, ini))
1122                 ini->smcr_version &= ~SMC_V2;
1123         ini->check_smcrv2 = false;
1124
1125         ini->smc_type_v2 = smc_indicated_type(ini->smcd_version & SMC_V2,
1126                                               ini->smcr_version & SMC_V2);
1127
1128         /* if neither ISM nor RDMA are supported, fallback */
1129         if (ini->smc_type_v1 == SMC_TYPE_N && ini->smc_type_v2 == SMC_TYPE_N)
1130                 rc = SMC_CLC_DECL_NOSMCDEV;
1131
1132         return rc;
1133 }
1134
1135 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
1136  * used, the VLAN ID will be registered again during the connection setup.
1137  */
1138 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc,
1139                                         struct smc_init_info *ini)
1140 {
1141         if (!smcd_indicated(ini->smc_type_v1))
1142                 return 0;
1143         if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev[0], ini->vlan_id))
1144                 return SMC_CLC_DECL_CNFERR;
1145         return 0;
1146 }
1147
1148 #define SMC_CLC_MAX_ACCEPT_LEN \
1149         (sizeof(struct smc_clc_msg_accept_confirm) + \
1150          sizeof(struct smc_clc_first_contact_ext_v2x) + \
1151          sizeof(struct smc_clc_msg_trail))
1152
1153 /* CLC handshake during connect */
1154 static int smc_connect_clc(struct smc_sock *smc,
1155                            struct smc_clc_msg_accept_confirm *aclc,
1156                            struct smc_init_info *ini)
1157 {
1158         int rc = 0;
1159
1160         /* do inband token exchange */
1161         rc = smc_clc_send_proposal(smc, ini);
1162         if (rc)
1163                 return rc;
1164         /* receive SMC Accept CLC message */
1165         return smc_clc_wait_msg(smc, aclc, SMC_CLC_MAX_ACCEPT_LEN,
1166                                 SMC_CLC_ACCEPT, CLC_WAIT_TIME);
1167 }
1168
1169 void smc_fill_gid_list(struct smc_link_group *lgr,
1170                        struct smc_gidlist *gidlist,
1171                        struct smc_ib_device *known_dev, u8 *known_gid)
1172 {
1173         struct smc_init_info *alt_ini = NULL;
1174
1175         memset(gidlist, 0, sizeof(*gidlist));
1176         memcpy(gidlist->list[gidlist->len++], known_gid, SMC_GID_SIZE);
1177
1178         alt_ini = kzalloc(sizeof(*alt_ini), GFP_KERNEL);
1179         if (!alt_ini)
1180                 goto out;
1181
1182         alt_ini->vlan_id = lgr->vlan_id;
1183         alt_ini->check_smcrv2 = true;
1184         alt_ini->smcrv2.saddr = lgr->saddr;
1185         smc_pnet_find_alt_roce(lgr, alt_ini, known_dev);
1186
1187         if (!alt_ini->smcrv2.ib_dev_v2)
1188                 goto out;
1189
1190         memcpy(gidlist->list[gidlist->len++], alt_ini->smcrv2.ib_gid_v2,
1191                SMC_GID_SIZE);
1192
1193 out:
1194         kfree(alt_ini);
1195 }
1196
1197 static int smc_connect_rdma_v2_prepare(struct smc_sock *smc,
1198                                        struct smc_clc_msg_accept_confirm *aclc,
1199                                        struct smc_init_info *ini)
1200 {
1201         struct smc_clc_first_contact_ext *fce =
1202                 smc_get_clc_first_contact_ext(aclc, false);
1203         struct net *net = sock_net(&smc->sk);
1204         int rc;
1205
1206         if (!ini->first_contact_peer || aclc->hdr.version == SMC_V1)
1207                 return 0;
1208
1209         if (fce->v2_direct) {
1210                 memcpy(ini->smcrv2.nexthop_mac, &aclc->r0.lcl.mac, ETH_ALEN);
1211                 ini->smcrv2.uses_gateway = false;
1212         } else {
1213                 if (smc_ib_find_route(net, smc->clcsock->sk->sk_rcv_saddr,
1214                                       smc_ib_gid_to_ipv4(aclc->r0.lcl.gid),
1215                                       ini->smcrv2.nexthop_mac,
1216                                       &ini->smcrv2.uses_gateway))
1217                         return SMC_CLC_DECL_NOROUTE;
1218                 if (!ini->smcrv2.uses_gateway) {
1219                         /* mismatch: peer claims indirect, but its direct */
1220                         return SMC_CLC_DECL_NOINDIRECT;
1221                 }
1222         }
1223
1224         ini->release_nr = fce->release;
1225         rc = smc_clc_clnt_v2x_features_validate(fce, ini);
1226         if (rc)
1227                 return rc;
1228
1229         return 0;
1230 }
1231
1232 /* setup for RDMA connection of client */
1233 static int smc_connect_rdma(struct smc_sock *smc,
1234                             struct smc_clc_msg_accept_confirm *aclc,
1235                             struct smc_init_info *ini)
1236 {
1237         int i, reason_code = 0;
1238         struct smc_link *link;
1239         u8 *eid = NULL;
1240
1241         ini->is_smcd = false;
1242         ini->ib_clcqpn = ntoh24(aclc->r0.qpn);
1243         ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
1244         memcpy(ini->peer_systemid, aclc->r0.lcl.id_for_peer, SMC_SYSTEMID_LEN);
1245         memcpy(ini->peer_gid, aclc->r0.lcl.gid, SMC_GID_SIZE);
1246         memcpy(ini->peer_mac, aclc->r0.lcl.mac, ETH_ALEN);
1247         ini->max_conns = SMC_CONN_PER_LGR_MAX;
1248         ini->max_links = SMC_LINKS_ADD_LNK_MAX;
1249
1250         reason_code = smc_connect_rdma_v2_prepare(smc, aclc, ini);
1251         if (reason_code)
1252                 return reason_code;
1253
1254         mutex_lock(&smc_client_lgr_pending);
1255         reason_code = smc_conn_create(smc, ini);
1256         if (reason_code) {
1257                 mutex_unlock(&smc_client_lgr_pending);
1258                 return reason_code;
1259         }
1260
1261         smc_conn_save_peer_info(smc, aclc);
1262
1263         if (ini->first_contact_local) {
1264                 link = smc->conn.lnk;
1265         } else {
1266                 /* set link that was assigned by server */
1267                 link = NULL;
1268                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1269                         struct smc_link *l = &smc->conn.lgr->lnk[i];
1270
1271                         if (l->peer_qpn == ntoh24(aclc->r0.qpn) &&
1272                             !memcmp(l->peer_gid, &aclc->r0.lcl.gid,
1273                                     SMC_GID_SIZE) &&
1274                             (aclc->hdr.version > SMC_V1 ||
1275                              !memcmp(l->peer_mac, &aclc->r0.lcl.mac,
1276                                      sizeof(l->peer_mac)))) {
1277                                 link = l;
1278                                 break;
1279                         }
1280                 }
1281                 if (!link) {
1282                         reason_code = SMC_CLC_DECL_NOSRVLINK;
1283                         goto connect_abort;
1284                 }
1285                 smc_switch_link_and_count(&smc->conn, link);
1286         }
1287
1288         /* create send buffer and rmb */
1289         if (smc_buf_create(smc, false)) {
1290                 reason_code = SMC_CLC_DECL_MEM;
1291                 goto connect_abort;
1292         }
1293
1294         if (ini->first_contact_local)
1295                 smc_link_save_peer_info(link, aclc, ini);
1296
1297         if (smc_rmb_rtoken_handling(&smc->conn, link, aclc)) {
1298                 reason_code = SMC_CLC_DECL_ERR_RTOK;
1299                 goto connect_abort;
1300         }
1301
1302         smc_close_init(smc);
1303         smc_rx_init(smc);
1304
1305         if (ini->first_contact_local) {
1306                 if (smc_ib_ready_link(link)) {
1307                         reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1308                         goto connect_abort;
1309                 }
1310         } else {
1311                 /* reg sendbufs if they were vzalloced */
1312                 if (smc->conn.sndbuf_desc->is_vm) {
1313                         if (smcr_lgr_reg_sndbufs(link, smc->conn.sndbuf_desc)) {
1314                                 reason_code = SMC_CLC_DECL_ERR_REGBUF;
1315                                 goto connect_abort;
1316                         }
1317                 }
1318                 if (smcr_lgr_reg_rmbs(link, smc->conn.rmb_desc)) {
1319                         reason_code = SMC_CLC_DECL_ERR_REGBUF;
1320                         goto connect_abort;
1321                 }
1322         }
1323
1324         if (aclc->hdr.version > SMC_V1) {
1325                 eid = aclc->r1.eid;
1326                 if (ini->first_contact_local)
1327                         smc_fill_gid_list(link->lgr, &ini->smcrv2.gidlist,
1328                                           link->smcibdev, link->gid);
1329         }
1330
1331         reason_code = smc_clc_send_confirm(smc, ini->first_contact_local,
1332                                            aclc->hdr.version, eid, ini);
1333         if (reason_code)
1334                 goto connect_abort;
1335
1336         smc_tx_init(smc);
1337
1338         if (ini->first_contact_local) {
1339                 /* QP confirmation over RoCE fabric */
1340                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1341                 reason_code = smcr_clnt_conf_first_link(smc);
1342                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1343                 if (reason_code)
1344                         goto connect_abort;
1345         }
1346         mutex_unlock(&smc_client_lgr_pending);
1347
1348         smc_copy_sock_settings_to_clc(smc);
1349         smc->connect_nonblock = 0;
1350         if (smc->sk.sk_state == SMC_INIT)
1351                 smc->sk.sk_state = SMC_ACTIVE;
1352
1353         return 0;
1354 connect_abort:
1355         smc_conn_abort(smc, ini->first_contact_local);
1356         mutex_unlock(&smc_client_lgr_pending);
1357         smc->connect_nonblock = 0;
1358
1359         return reason_code;
1360 }
1361
1362 /* The server has chosen one of the proposed ISM devices for the communication.
1363  * Determine from the CHID of the received CLC ACCEPT the ISM device chosen.
1364  */
1365 static int
1366 smc_v2_determine_accepted_chid(struct smc_clc_msg_accept_confirm *aclc,
1367                                struct smc_init_info *ini)
1368 {
1369         int i;
1370
1371         for (i = 0; i < ini->ism_offered_cnt + 1; i++) {
1372                 if (ini->ism_chid[i] == ntohs(aclc->d1.chid)) {
1373                         ini->ism_selected = i;
1374                         return 0;
1375                 }
1376         }
1377
1378         return -EPROTO;
1379 }
1380
1381 /* setup for ISM connection of client */
1382 static int smc_connect_ism(struct smc_sock *smc,
1383                            struct smc_clc_msg_accept_confirm *aclc,
1384                            struct smc_init_info *ini)
1385 {
1386         u8 *eid = NULL;
1387         int rc = 0;
1388
1389         ini->is_smcd = true;
1390         ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
1391
1392         if (aclc->hdr.version == SMC_V2) {
1393                 if (ini->first_contact_peer) {
1394                         struct smc_clc_first_contact_ext *fce =
1395                                 smc_get_clc_first_contact_ext(aclc, true);
1396
1397                         ini->release_nr = fce->release;
1398                         rc = smc_clc_clnt_v2x_features_validate(fce, ini);
1399                         if (rc)
1400                                 return rc;
1401                 }
1402
1403                 rc = smc_v2_determine_accepted_chid(aclc, ini);
1404                 if (rc)
1405                         return rc;
1406
1407                 if (__smc_ism_is_emulated(ini->ism_chid[ini->ism_selected]))
1408                         ini->ism_peer_gid[ini->ism_selected].gid_ext =
1409                                                 ntohll(aclc->d1.gid_ext);
1410                 /* for non-Emulated-ISM devices, peer gid_ext remains 0. */
1411         }
1412         ini->ism_peer_gid[ini->ism_selected].gid = ntohll(aclc->d0.gid);
1413
1414         /* there is only one lgr role for SMC-D; use server lock */
1415         mutex_lock(&smc_server_lgr_pending);
1416         rc = smc_conn_create(smc, ini);
1417         if (rc) {
1418                 mutex_unlock(&smc_server_lgr_pending);
1419                 return rc;
1420         }
1421
1422         /* Create send and receive buffers */
1423         rc = smc_buf_create(smc, true);
1424         if (rc) {
1425                 rc = (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB : SMC_CLC_DECL_MEM;
1426                 goto connect_abort;
1427         }
1428
1429         smc_conn_save_peer_info(smc, aclc);
1430
1431         if (smc_ism_support_dmb_nocopy(smc->conn.lgr->smcd)) {
1432                 rc = smcd_buf_attach(smc);
1433                 if (rc) {
1434                         rc = SMC_CLC_DECL_MEM;  /* try to fallback */
1435                         goto connect_abort;
1436                 }
1437         }
1438         smc_close_init(smc);
1439         smc_rx_init(smc);
1440         smc_tx_init(smc);
1441
1442         if (aclc->hdr.version > SMC_V1)
1443                 eid = aclc->d1.eid;
1444
1445         rc = smc_clc_send_confirm(smc, ini->first_contact_local,
1446                                   aclc->hdr.version, eid, ini);
1447         if (rc)
1448                 goto connect_abort;
1449         mutex_unlock(&smc_server_lgr_pending);
1450
1451         smc_copy_sock_settings_to_clc(smc);
1452         smc->connect_nonblock = 0;
1453         if (smc->sk.sk_state == SMC_INIT)
1454                 smc->sk.sk_state = SMC_ACTIVE;
1455
1456         return 0;
1457 connect_abort:
1458         smc_conn_abort(smc, ini->first_contact_local);
1459         mutex_unlock(&smc_server_lgr_pending);
1460         smc->connect_nonblock = 0;
1461
1462         return rc;
1463 }
1464
1465 /* check if received accept type and version matches a proposed one */
1466 static int smc_connect_check_aclc(struct smc_init_info *ini,
1467                                   struct smc_clc_msg_accept_confirm *aclc)
1468 {
1469         if (aclc->hdr.version >= SMC_V2) {
1470                 if ((aclc->hdr.typev1 == SMC_TYPE_R &&
1471                      !smcr_indicated(ini->smc_type_v2)) ||
1472                     (aclc->hdr.typev1 == SMC_TYPE_D &&
1473                      !smcd_indicated(ini->smc_type_v2)))
1474                         return SMC_CLC_DECL_MODEUNSUPP;
1475         } else {
1476                 if ((aclc->hdr.typev1 == SMC_TYPE_R &&
1477                      !smcr_indicated(ini->smc_type_v1)) ||
1478                     (aclc->hdr.typev1 == SMC_TYPE_D &&
1479                      !smcd_indicated(ini->smc_type_v1)))
1480                         return SMC_CLC_DECL_MODEUNSUPP;
1481         }
1482
1483         return 0;
1484 }
1485
1486 /* perform steps before actually connecting */
1487 static int __smc_connect(struct smc_sock *smc)
1488 {
1489         u8 version = smc_ism_is_v2_capable() ? SMC_V2 : SMC_V1;
1490         struct smc_clc_msg_accept_confirm *aclc;
1491         struct smc_init_info *ini = NULL;
1492         u8 *buf = NULL;
1493         int rc = 0;
1494
1495         if (smc->use_fallback)
1496                 return smc_connect_fallback(smc, smc->fallback_rsn);
1497
1498         /* if peer has not signalled SMC-capability, fall back */
1499         if (!tcp_sk(smc->clcsock->sk)->syn_smc)
1500                 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
1501
1502         /* IPSec connections opt out of SMC optimizations */
1503         if (using_ipsec(smc))
1504                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC,
1505                                                     version);
1506
1507         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1508         if (!ini)
1509                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_MEM,
1510                                                     version);
1511
1512         ini->smcd_version = SMC_V1 | SMC_V2;
1513         ini->smcr_version = SMC_V1 | SMC_V2;
1514         ini->smc_type_v1 = SMC_TYPE_B;
1515         ini->smc_type_v2 = SMC_TYPE_B;
1516
1517         /* get vlan id from IP device */
1518         if (smc_vlan_by_tcpsk(smc->clcsock, ini)) {
1519                 ini->smcd_version &= ~SMC_V1;
1520                 ini->smcr_version = 0;
1521                 ini->smc_type_v1 = SMC_TYPE_N;
1522         }
1523
1524         rc = smc_find_proposal_devices(smc, ini);
1525         if (rc)
1526                 goto fallback;
1527
1528         buf = kzalloc(SMC_CLC_MAX_ACCEPT_LEN, GFP_KERNEL);
1529         if (!buf) {
1530                 rc = SMC_CLC_DECL_MEM;
1531                 goto fallback;
1532         }
1533         aclc = (struct smc_clc_msg_accept_confirm *)buf;
1534
1535         /* perform CLC handshake */
1536         rc = smc_connect_clc(smc, aclc, ini);
1537         if (rc) {
1538                 /* -EAGAIN on timeout, see tcp_recvmsg() */
1539                 if (rc == -EAGAIN) {
1540                         rc = -ETIMEDOUT;
1541                         smc->sk.sk_err = ETIMEDOUT;
1542                 }
1543                 goto vlan_cleanup;
1544         }
1545
1546         /* check if smc modes and versions of CLC proposal and accept match */
1547         rc = smc_connect_check_aclc(ini, aclc);
1548         version = aclc->hdr.version == SMC_V1 ? SMC_V1 : SMC_V2;
1549         if (rc)
1550                 goto vlan_cleanup;
1551
1552         /* depending on previous steps, connect using rdma or ism */
1553         if (aclc->hdr.typev1 == SMC_TYPE_R) {
1554                 ini->smcr_version = version;
1555                 rc = smc_connect_rdma(smc, aclc, ini);
1556         } else if (aclc->hdr.typev1 == SMC_TYPE_D) {
1557                 ini->smcd_version = version;
1558                 rc = smc_connect_ism(smc, aclc, ini);
1559         }
1560         if (rc)
1561                 goto vlan_cleanup;
1562
1563         SMC_STAT_CLNT_SUCC_INC(sock_net(smc->clcsock->sk), aclc);
1564         smc_connect_ism_vlan_cleanup(smc, ini);
1565         kfree(buf);
1566         kfree(ini);
1567         return 0;
1568
1569 vlan_cleanup:
1570         smc_connect_ism_vlan_cleanup(smc, ini);
1571         kfree(buf);
1572 fallback:
1573         kfree(ini);
1574         return smc_connect_decline_fallback(smc, rc, version);
1575 }
1576
1577 static void smc_connect_work(struct work_struct *work)
1578 {
1579         struct smc_sock *smc = container_of(work, struct smc_sock,
1580                                             connect_work);
1581         long timeo = smc->sk.sk_sndtimeo;
1582         int rc = 0;
1583
1584         if (!timeo)
1585                 timeo = MAX_SCHEDULE_TIMEOUT;
1586         lock_sock(smc->clcsock->sk);
1587         if (smc->clcsock->sk->sk_err) {
1588                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
1589         } else if ((1 << smc->clcsock->sk->sk_state) &
1590                                         (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
1591                 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
1592                 if ((rc == -EPIPE) &&
1593                     ((1 << smc->clcsock->sk->sk_state) &
1594                                         (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
1595                         rc = 0;
1596         }
1597         release_sock(smc->clcsock->sk);
1598         lock_sock(&smc->sk);
1599         if (rc != 0 || smc->sk.sk_err) {
1600                 smc->sk.sk_state = SMC_CLOSED;
1601                 if (rc == -EPIPE || rc == -EAGAIN)
1602                         smc->sk.sk_err = EPIPE;
1603                 else if (rc == -ECONNREFUSED)
1604                         smc->sk.sk_err = ECONNREFUSED;
1605                 else if (signal_pending(current))
1606                         smc->sk.sk_err = -sock_intr_errno(timeo);
1607                 sock_put(&smc->sk); /* passive closing */
1608                 goto out;
1609         }
1610
1611         rc = __smc_connect(smc);
1612         if (rc < 0)
1613                 smc->sk.sk_err = -rc;
1614
1615 out:
1616         if (!sock_flag(&smc->sk, SOCK_DEAD)) {
1617                 if (smc->sk.sk_err) {
1618                         smc->sk.sk_state_change(&smc->sk);
1619                 } else { /* allow polling before and after fallback decision */
1620                         smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
1621                         smc->sk.sk_write_space(&smc->sk);
1622                 }
1623         }
1624         release_sock(&smc->sk);
1625 }
1626
1627 int smc_connect(struct socket *sock, struct sockaddr *addr,
1628                 int alen, int flags)
1629 {
1630         struct sock *sk = sock->sk;
1631         struct smc_sock *smc;
1632         int rc = -EINVAL;
1633
1634         smc = smc_sk(sk);
1635
1636         /* separate smc parameter checking to be safe */
1637         if (alen < sizeof(addr->sa_family))
1638                 goto out_err;
1639         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
1640                 goto out_err;
1641
1642         lock_sock(sk);
1643         switch (sock->state) {
1644         default:
1645                 rc = -EINVAL;
1646                 goto out;
1647         case SS_CONNECTED:
1648                 rc = sk->sk_state == SMC_ACTIVE ? -EISCONN : -EINVAL;
1649                 goto out;
1650         case SS_CONNECTING:
1651                 if (sk->sk_state == SMC_ACTIVE)
1652                         goto connected;
1653                 break;
1654         case SS_UNCONNECTED:
1655                 sock->state = SS_CONNECTING;
1656                 break;
1657         }
1658
1659         switch (sk->sk_state) {
1660         default:
1661                 goto out;
1662         case SMC_CLOSED:
1663                 rc = sock_error(sk) ? : -ECONNABORTED;
1664                 sock->state = SS_UNCONNECTED;
1665                 goto out;
1666         case SMC_ACTIVE:
1667                 rc = -EISCONN;
1668                 goto out;
1669         case SMC_INIT:
1670                 break;
1671         }
1672
1673         smc_copy_sock_settings_to_clc(smc);
1674         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1675         if (smc->connect_nonblock) {
1676                 rc = -EALREADY;
1677                 goto out;
1678         }
1679         rc = kernel_connect(smc->clcsock, addr, alen, flags);
1680         if (rc && rc != -EINPROGRESS)
1681                 goto out;
1682
1683         if (smc->use_fallback) {
1684                 sock->state = rc ? SS_CONNECTING : SS_CONNECTED;
1685                 goto out;
1686         }
1687         sock_hold(&smc->sk); /* sock put in passive closing */
1688         if (flags & O_NONBLOCK) {
1689                 if (queue_work(smc_hs_wq, &smc->connect_work))
1690                         smc->connect_nonblock = 1;
1691                 rc = -EINPROGRESS;
1692                 goto out;
1693         } else {
1694                 rc = __smc_connect(smc);
1695                 if (rc < 0)
1696                         goto out;
1697         }
1698
1699 connected:
1700         rc = 0;
1701         sock->state = SS_CONNECTED;
1702 out:
1703         release_sock(sk);
1704 out_err:
1705         return rc;
1706 }
1707
1708 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
1709 {
1710         struct socket *new_clcsock = NULL;
1711         struct sock *lsk = &lsmc->sk;
1712         struct sock *new_sk;
1713         int rc = -EINVAL;
1714
1715         release_sock(lsk);
1716         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
1717         if (!new_sk) {
1718                 rc = -ENOMEM;
1719                 lsk->sk_err = ENOMEM;
1720                 *new_smc = NULL;
1721                 lock_sock(lsk);
1722                 goto out;
1723         }
1724         *new_smc = smc_sk(new_sk);
1725
1726         mutex_lock(&lsmc->clcsock_release_lock);
1727         if (lsmc->clcsock)
1728                 rc = kernel_accept(lsmc->clcsock, &new_clcsock, SOCK_NONBLOCK);
1729         mutex_unlock(&lsmc->clcsock_release_lock);
1730         lock_sock(lsk);
1731         if  (rc < 0 && rc != -EAGAIN)
1732                 lsk->sk_err = -rc;
1733         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
1734                 new_sk->sk_prot->unhash(new_sk);
1735                 if (new_clcsock)
1736                         sock_release(new_clcsock);
1737                 new_sk->sk_state = SMC_CLOSED;
1738                 smc_sock_set_flag(new_sk, SOCK_DEAD);
1739                 sock_put(new_sk); /* final */
1740                 *new_smc = NULL;
1741                 goto out;
1742         }
1743
1744         /* new clcsock has inherited the smc listen-specific sk_data_ready
1745          * function; switch it back to the original sk_data_ready function
1746          */
1747         new_clcsock->sk->sk_data_ready = lsmc->clcsk_data_ready;
1748
1749         /* if new clcsock has also inherited the fallback-specific callback
1750          * functions, switch them back to the original ones.
1751          */
1752         if (lsmc->use_fallback) {
1753                 if (lsmc->clcsk_state_change)
1754                         new_clcsock->sk->sk_state_change = lsmc->clcsk_state_change;
1755                 if (lsmc->clcsk_write_space)
1756                         new_clcsock->sk->sk_write_space = lsmc->clcsk_write_space;
1757                 if (lsmc->clcsk_error_report)
1758                         new_clcsock->sk->sk_error_report = lsmc->clcsk_error_report;
1759         }
1760
1761         (*new_smc)->clcsock = new_clcsock;
1762 out:
1763         return rc;
1764 }
1765
1766 /* add a just created sock to the accept queue of the listen sock as
1767  * candidate for a following socket accept call from user space
1768  */
1769 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
1770 {
1771         struct smc_sock *par = smc_sk(parent);
1772
1773         sock_hold(sk); /* sock_put in smc_accept_unlink () */
1774         spin_lock(&par->accept_q_lock);
1775         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
1776         spin_unlock(&par->accept_q_lock);
1777         sk_acceptq_added(parent);
1778 }
1779
1780 /* remove a socket from the accept queue of its parental listening socket */
1781 static void smc_accept_unlink(struct sock *sk)
1782 {
1783         struct smc_sock *par = smc_sk(sk)->listen_smc;
1784
1785         spin_lock(&par->accept_q_lock);
1786         list_del_init(&smc_sk(sk)->accept_q);
1787         spin_unlock(&par->accept_q_lock);
1788         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
1789         sock_put(sk); /* sock_hold in smc_accept_enqueue */
1790 }
1791
1792 /* remove a sock from the accept queue to bind it to a new socket created
1793  * for a socket accept call from user space
1794  */
1795 struct sock *smc_accept_dequeue(struct sock *parent,
1796                                 struct socket *new_sock)
1797 {
1798         struct smc_sock *isk, *n;
1799         struct sock *new_sk;
1800
1801         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1802                 new_sk = (struct sock *)isk;
1803
1804                 smc_accept_unlink(new_sk);
1805                 if (new_sk->sk_state == SMC_CLOSED) {
1806                         new_sk->sk_prot->unhash(new_sk);
1807                         if (isk->clcsock) {
1808                                 sock_release(isk->clcsock);
1809                                 isk->clcsock = NULL;
1810                         }
1811                         sock_put(new_sk); /* final */
1812                         continue;
1813                 }
1814                 if (new_sock) {
1815                         sock_graft(new_sk, new_sock);
1816                         new_sock->state = SS_CONNECTED;
1817                         if (isk->use_fallback) {
1818                                 smc_sk(new_sk)->clcsock->file = new_sock->file;
1819                                 isk->clcsock->file->private_data = isk->clcsock;
1820                         }
1821                 }
1822                 return new_sk;
1823         }
1824         return NULL;
1825 }
1826
1827 /* clean up for a created but never accepted sock */
1828 void smc_close_non_accepted(struct sock *sk)
1829 {
1830         struct smc_sock *smc = smc_sk(sk);
1831
1832         sock_hold(sk); /* sock_put below */
1833         lock_sock(sk);
1834         if (!sk->sk_lingertime)
1835                 /* wait for peer closing */
1836                 WRITE_ONCE(sk->sk_lingertime, SMC_MAX_STREAM_WAIT_TIMEOUT);
1837         __smc_release(smc);
1838         release_sock(sk);
1839         sock_put(sk); /* sock_hold above */
1840         sock_put(sk); /* final sock_put */
1841 }
1842
1843 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1844 {
1845         struct smc_link *link = smc->conn.lnk;
1846         struct smc_llc_qentry *qentry;
1847         int rc;
1848
1849         /* reg the sndbuf if it was vzalloced*/
1850         if (smc->conn.sndbuf_desc->is_vm) {
1851                 if (smcr_link_reg_buf(link, smc->conn.sndbuf_desc))
1852                         return SMC_CLC_DECL_ERR_REGBUF;
1853         }
1854
1855         /* reg the rmb */
1856         if (smcr_link_reg_buf(link, smc->conn.rmb_desc))
1857                 return SMC_CLC_DECL_ERR_REGBUF;
1858
1859         /* send CONFIRM LINK request to client over the RoCE fabric */
1860         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1861         if (rc < 0)
1862                 return SMC_CLC_DECL_TIMEOUT_CL;
1863
1864         /* receive CONFIRM LINK response from client over the RoCE fabric */
1865         qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1866                               SMC_LLC_CONFIRM_LINK);
1867         if (!qentry) {
1868                 struct smc_clc_msg_decline dclc;
1869
1870                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1871                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1872                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1873         }
1874         smc_llc_save_peer_uid(qentry);
1875         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1876         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1877         if (rc)
1878                 return SMC_CLC_DECL_RMBE_EC;
1879
1880         /* confirm_rkey is implicit on 1st contact */
1881         smc->conn.rmb_desc->is_conf_rkey = true;
1882
1883         smc_llc_link_active(link);
1884         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1885
1886         if (link->lgr->max_links > 1) {
1887                 down_write(&link->lgr->llc_conf_mutex);
1888                 /* initial contact - try to establish second link */
1889                 smc_llc_srv_add_link(link, NULL);
1890                 up_write(&link->lgr->llc_conf_mutex);
1891         }
1892         return 0;
1893 }
1894
1895 /* listen worker: finish */
1896 static void smc_listen_out(struct smc_sock *new_smc)
1897 {
1898         struct smc_sock *lsmc = new_smc->listen_smc;
1899         struct sock *newsmcsk = &new_smc->sk;
1900
1901         if (tcp_sk(new_smc->clcsock->sk)->syn_smc)
1902                 atomic_dec(&lsmc->queued_smc_hs);
1903
1904         if (lsmc->sk.sk_state == SMC_LISTEN) {
1905                 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1906                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1907                 release_sock(&lsmc->sk);
1908         } else { /* no longer listening */
1909                 smc_close_non_accepted(newsmcsk);
1910         }
1911
1912         /* Wake up accept */
1913         lsmc->sk.sk_data_ready(&lsmc->sk);
1914         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1915 }
1916
1917 /* listen worker: finish in state connected */
1918 static void smc_listen_out_connected(struct smc_sock *new_smc)
1919 {
1920         struct sock *newsmcsk = &new_smc->sk;
1921
1922         if (newsmcsk->sk_state == SMC_INIT)
1923                 newsmcsk->sk_state = SMC_ACTIVE;
1924
1925         smc_listen_out(new_smc);
1926 }
1927
1928 /* listen worker: finish in error state */
1929 static void smc_listen_out_err(struct smc_sock *new_smc)
1930 {
1931         struct sock *newsmcsk = &new_smc->sk;
1932         struct net *net = sock_net(newsmcsk);
1933
1934         this_cpu_inc(net->smc.smc_stats->srv_hshake_err_cnt);
1935         if (newsmcsk->sk_state == SMC_INIT)
1936                 sock_put(&new_smc->sk); /* passive closing */
1937         newsmcsk->sk_state = SMC_CLOSED;
1938
1939         smc_listen_out(new_smc);
1940 }
1941
1942 /* listen worker: decline and fall back if possible */
1943 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1944                                int local_first, u8 version)
1945 {
1946         /* RDMA setup failed, switch back to TCP */
1947         smc_conn_abort(new_smc, local_first);
1948         if (reason_code < 0 ||
1949             smc_switch_to_fallback(new_smc, reason_code)) {
1950                 /* error, no fallback possible */
1951                 smc_listen_out_err(new_smc);
1952                 return;
1953         }
1954         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1955                 if (smc_clc_send_decline(new_smc, reason_code, version) < 0) {
1956                         smc_listen_out_err(new_smc);
1957                         return;
1958                 }
1959         }
1960         smc_listen_out_connected(new_smc);
1961 }
1962
1963 /* listen worker: version checking */
1964 static int smc_listen_v2_check(struct smc_sock *new_smc,
1965                                struct smc_clc_msg_proposal *pclc,
1966                                struct smc_init_info *ini)
1967 {
1968         struct smc_clc_smcd_v2_extension *pclc_smcd_v2_ext;
1969         struct smc_clc_v2_extension *pclc_v2_ext;
1970         int rc = SMC_CLC_DECL_PEERNOSMC;
1971
1972         ini->smc_type_v1 = pclc->hdr.typev1;
1973         ini->smc_type_v2 = pclc->hdr.typev2;
1974         ini->smcd_version = smcd_indicated(ini->smc_type_v1) ? SMC_V1 : 0;
1975         ini->smcr_version = smcr_indicated(ini->smc_type_v1) ? SMC_V1 : 0;
1976         if (pclc->hdr.version > SMC_V1) {
1977                 if (smcd_indicated(ini->smc_type_v2))
1978                         ini->smcd_version |= SMC_V2;
1979                 if (smcr_indicated(ini->smc_type_v2))
1980                         ini->smcr_version |= SMC_V2;
1981         }
1982         if (!(ini->smcd_version & SMC_V2) && !(ini->smcr_version & SMC_V2)) {
1983                 rc = SMC_CLC_DECL_PEERNOSMC;
1984                 goto out;
1985         }
1986         pclc_v2_ext = smc_get_clc_v2_ext(pclc);
1987         if (!pclc_v2_ext) {
1988                 ini->smcd_version &= ~SMC_V2;
1989                 ini->smcr_version &= ~SMC_V2;
1990                 rc = SMC_CLC_DECL_NOV2EXT;
1991                 goto out;
1992         }
1993         pclc_smcd_v2_ext = smc_get_clc_smcd_v2_ext(pclc_v2_ext);
1994         if (ini->smcd_version & SMC_V2) {
1995                 if (!smc_ism_is_v2_capable()) {
1996                         ini->smcd_version &= ~SMC_V2;
1997                         rc = SMC_CLC_DECL_NOISM2SUPP;
1998                 } else if (!pclc_smcd_v2_ext) {
1999                         ini->smcd_version &= ~SMC_V2;
2000                         rc = SMC_CLC_DECL_NOV2DEXT;
2001                 } else if (!pclc_v2_ext->hdr.eid_cnt &&
2002                            !pclc_v2_ext->hdr.flag.seid) {
2003                         ini->smcd_version &= ~SMC_V2;
2004                         rc = SMC_CLC_DECL_NOUEID;
2005                 }
2006         }
2007         if (ini->smcr_version & SMC_V2) {
2008                 if (!pclc_v2_ext->hdr.eid_cnt) {
2009                         ini->smcr_version &= ~SMC_V2;
2010                         rc = SMC_CLC_DECL_NOUEID;
2011                 }
2012         }
2013
2014         ini->release_nr = pclc_v2_ext->hdr.flag.release;
2015         if (pclc_v2_ext->hdr.flag.release > SMC_RELEASE)
2016                 ini->release_nr = SMC_RELEASE;
2017
2018 out:
2019         if (!ini->smcd_version && !ini->smcr_version)
2020                 return rc;
2021
2022         return 0;
2023 }
2024
2025 /* listen worker: check prefixes */
2026 static int smc_listen_prfx_check(struct smc_sock *new_smc,
2027                                  struct smc_clc_msg_proposal *pclc)
2028 {
2029         struct smc_clc_msg_proposal_prefix *pclc_prfx;
2030         struct socket *newclcsock = new_smc->clcsock;
2031
2032         if (pclc->hdr.typev1 == SMC_TYPE_N)
2033                 return 0;
2034         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
2035         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
2036                 return SMC_CLC_DECL_DIFFPREFIX;
2037
2038         return 0;
2039 }
2040
2041 /* listen worker: initialize connection and buffers */
2042 static int smc_listen_rdma_init(struct smc_sock *new_smc,
2043                                 struct smc_init_info *ini)
2044 {
2045         int rc;
2046
2047         /* allocate connection / link group */
2048         rc = smc_conn_create(new_smc, ini);
2049         if (rc)
2050                 return rc;
2051
2052         /* create send buffer and rmb */
2053         if (smc_buf_create(new_smc, false)) {
2054                 smc_conn_abort(new_smc, ini->first_contact_local);
2055                 return SMC_CLC_DECL_MEM;
2056         }
2057
2058         return 0;
2059 }
2060
2061 /* listen worker: initialize connection and buffers for SMC-D */
2062 static int smc_listen_ism_init(struct smc_sock *new_smc,
2063                                struct smc_init_info *ini)
2064 {
2065         int rc;
2066
2067         rc = smc_conn_create(new_smc, ini);
2068         if (rc)
2069                 return rc;
2070
2071         /* Create send and receive buffers */
2072         rc = smc_buf_create(new_smc, true);
2073         if (rc) {
2074                 smc_conn_abort(new_smc, ini->first_contact_local);
2075                 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
2076                                          SMC_CLC_DECL_MEM;
2077         }
2078
2079         return 0;
2080 }
2081
2082 static bool smc_is_already_selected(struct smcd_dev *smcd,
2083                                     struct smc_init_info *ini,
2084                                     int matches)
2085 {
2086         int i;
2087
2088         for (i = 0; i < matches; i++)
2089                 if (smcd == ini->ism_dev[i])
2090                         return true;
2091
2092         return false;
2093 }
2094
2095 /* check for ISM devices matching proposed ISM devices */
2096 static void smc_check_ism_v2_match(struct smc_init_info *ini,
2097                                    u16 proposed_chid,
2098                                    struct smcd_gid *proposed_gid,
2099                                    unsigned int *matches)
2100 {
2101         struct smcd_dev *smcd;
2102
2103         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2104                 if (smcd->going_away)
2105                         continue;
2106                 if (smc_is_already_selected(smcd, ini, *matches))
2107                         continue;
2108                 if (smc_ism_get_chid(smcd) == proposed_chid &&
2109                     !smc_ism_cantalk(proposed_gid, ISM_RESERVED_VLANID, smcd)) {
2110                         ini->ism_peer_gid[*matches].gid = proposed_gid->gid;
2111                         if (__smc_ism_is_emulated(proposed_chid))
2112                                 ini->ism_peer_gid[*matches].gid_ext =
2113                                                         proposed_gid->gid_ext;
2114                                 /* non-Emulated-ISM's peer gid_ext remains 0. */
2115                         ini->ism_dev[*matches] = smcd;
2116                         (*matches)++;
2117                         break;
2118                 }
2119         }
2120 }
2121
2122 static void smc_find_ism_store_rc(u32 rc, struct smc_init_info *ini)
2123 {
2124         if (!ini->rc)
2125                 ini->rc = rc;
2126 }
2127
2128 static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
2129                                         struct smc_clc_msg_proposal *pclc,
2130                                         struct smc_init_info *ini)
2131 {
2132         struct smc_clc_smcd_v2_extension *smcd_v2_ext;
2133         struct smc_clc_v2_extension *smc_v2_ext;
2134         struct smc_clc_msg_smcd *pclc_smcd;
2135         unsigned int matches = 0;
2136         struct smcd_gid smcd_gid;
2137         u8 smcd_version;
2138         u8 *eid = NULL;
2139         int i, rc;
2140         u16 chid;
2141
2142         if (!(ini->smcd_version & SMC_V2) || !smcd_indicated(ini->smc_type_v2))
2143                 goto not_found;
2144
2145         pclc_smcd = smc_get_clc_msg_smcd(pclc);
2146         smc_v2_ext = smc_get_clc_v2_ext(pclc);
2147         smcd_v2_ext = smc_get_clc_smcd_v2_ext(smc_v2_ext);
2148
2149         mutex_lock(&smcd_dev_list.mutex);
2150         if (pclc_smcd->ism.chid) {
2151                 /* check for ISM device matching proposed native ISM device */
2152                 smcd_gid.gid = ntohll(pclc_smcd->ism.gid);
2153                 smcd_gid.gid_ext = 0;
2154                 smc_check_ism_v2_match(ini, ntohs(pclc_smcd->ism.chid),
2155                                        &smcd_gid, &matches);
2156         }
2157         for (i = 0; i < smc_v2_ext->hdr.ism_gid_cnt; i++) {
2158                 /* check for ISM devices matching proposed non-native ISM
2159                  * devices
2160                  */
2161                 smcd_gid.gid = ntohll(smcd_v2_ext->gidchid[i].gid);
2162                 smcd_gid.gid_ext = 0;
2163                 chid = ntohs(smcd_v2_ext->gidchid[i].chid);
2164                 if (__smc_ism_is_emulated(chid)) {
2165                         if ((i + 1) == smc_v2_ext->hdr.ism_gid_cnt ||
2166                             chid != ntohs(smcd_v2_ext->gidchid[i + 1].chid))
2167                                 /* each Emulated-ISM device takes two GID-CHID
2168                                  * entries and CHID of the second entry repeats
2169                                  * that of the first entry.
2170                                  *
2171                                  * So check if the next GID-CHID entry exists
2172                                  * and both two entries' CHIDs are the same.
2173                                  */
2174                                 continue;
2175                         smcd_gid.gid_ext =
2176                                 ntohll(smcd_v2_ext->gidchid[++i].gid);
2177                 }
2178                 smc_check_ism_v2_match(ini, chid, &smcd_gid, &matches);
2179         }
2180         mutex_unlock(&smcd_dev_list.mutex);
2181
2182         if (!ini->ism_dev[0]) {
2183                 smc_find_ism_store_rc(SMC_CLC_DECL_NOSMCD2DEV, ini);
2184                 goto not_found;
2185         }
2186
2187         smc_ism_get_system_eid(&eid);
2188         if (!smc_clc_match_eid(ini->negotiated_eid, smc_v2_ext,
2189                                smcd_v2_ext->system_eid, eid))
2190                 goto not_found;
2191
2192         /* separate - outside the smcd_dev_list.lock */
2193         smcd_version = ini->smcd_version;
2194         for (i = 0; i < matches; i++) {
2195                 ini->smcd_version = SMC_V2;
2196                 ini->is_smcd = true;
2197                 ini->ism_selected = i;
2198                 rc = smc_listen_ism_init(new_smc, ini);
2199                 if (rc) {
2200                         smc_find_ism_store_rc(rc, ini);
2201                         /* try next active ISM device */
2202                         continue;
2203                 }
2204                 return; /* matching and usable V2 ISM device found */
2205         }
2206         /* no V2 ISM device could be initialized */
2207         ini->smcd_version = smcd_version;       /* restore original value */
2208         ini->negotiated_eid[0] = 0;
2209
2210 not_found:
2211         ini->smcd_version &= ~SMC_V2;
2212         ini->ism_dev[0] = NULL;
2213         ini->is_smcd = false;
2214 }
2215
2216 static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
2217                                         struct smc_clc_msg_proposal *pclc,
2218                                         struct smc_init_info *ini)
2219 {
2220         struct smc_clc_msg_smcd *pclc_smcd = smc_get_clc_msg_smcd(pclc);
2221         int rc = 0;
2222
2223         /* check if ISM V1 is available */
2224         if (!(ini->smcd_version & SMC_V1) || !smcd_indicated(ini->smc_type_v1))
2225                 goto not_found;
2226         ini->is_smcd = true; /* prepare ISM check */
2227         ini->ism_peer_gid[0].gid = ntohll(pclc_smcd->ism.gid);
2228         ini->ism_peer_gid[0].gid_ext = 0;
2229         rc = smc_find_ism_device(new_smc, ini);
2230         if (rc)
2231                 goto not_found;
2232         ini->ism_selected = 0;
2233         rc = smc_listen_ism_init(new_smc, ini);
2234         if (!rc)
2235                 return;         /* V1 ISM device found */
2236
2237 not_found:
2238         smc_find_ism_store_rc(rc, ini);
2239         ini->smcd_version &= ~SMC_V1;
2240         ini->ism_dev[0] = NULL;
2241         ini->is_smcd = false;
2242 }
2243
2244 /* listen worker: register buffers */
2245 static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
2246 {
2247         struct smc_connection *conn = &new_smc->conn;
2248
2249         if (!local_first) {
2250                 /* reg sendbufs if they were vzalloced */
2251                 if (conn->sndbuf_desc->is_vm) {
2252                         if (smcr_lgr_reg_sndbufs(conn->lnk,
2253                                                  conn->sndbuf_desc))
2254                                 return SMC_CLC_DECL_ERR_REGBUF;
2255                 }
2256                 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
2257                         return SMC_CLC_DECL_ERR_REGBUF;
2258         }
2259
2260         return 0;
2261 }
2262
2263 static void smc_find_rdma_v2_device_serv(struct smc_sock *new_smc,
2264                                          struct smc_clc_msg_proposal *pclc,
2265                                          struct smc_init_info *ini)
2266 {
2267         struct smc_clc_v2_extension *smc_v2_ext;
2268         u8 smcr_version;
2269         int rc;
2270
2271         if (!(ini->smcr_version & SMC_V2) || !smcr_indicated(ini->smc_type_v2))
2272                 goto not_found;
2273
2274         smc_v2_ext = smc_get_clc_v2_ext(pclc);
2275         if (!smc_clc_match_eid(ini->negotiated_eid, smc_v2_ext, NULL, NULL))
2276                 goto not_found;
2277
2278         /* prepare RDMA check */
2279         memcpy(ini->peer_systemid, pclc->lcl.id_for_peer, SMC_SYSTEMID_LEN);
2280         memcpy(ini->peer_gid, smc_v2_ext->roce, SMC_GID_SIZE);
2281         memcpy(ini->peer_mac, pclc->lcl.mac, ETH_ALEN);
2282         ini->check_smcrv2 = true;
2283         ini->smcrv2.clc_sk = new_smc->clcsock->sk;
2284         ini->smcrv2.saddr = new_smc->clcsock->sk->sk_rcv_saddr;
2285         ini->smcrv2.daddr = smc_ib_gid_to_ipv4(smc_v2_ext->roce);
2286         rc = smc_find_rdma_device(new_smc, ini);
2287         if (rc) {
2288                 smc_find_ism_store_rc(rc, ini);
2289                 goto not_found;
2290         }
2291         if (!ini->smcrv2.uses_gateway)
2292                 memcpy(ini->smcrv2.nexthop_mac, pclc->lcl.mac, ETH_ALEN);
2293
2294         smcr_version = ini->smcr_version;
2295         ini->smcr_version = SMC_V2;
2296         rc = smc_listen_rdma_init(new_smc, ini);
2297         if (!rc) {
2298                 rc = smc_listen_rdma_reg(new_smc, ini->first_contact_local);
2299                 if (rc)
2300                         smc_conn_abort(new_smc, ini->first_contact_local);
2301         }
2302         if (!rc)
2303                 return;
2304         ini->smcr_version = smcr_version;
2305         smc_find_ism_store_rc(rc, ini);
2306
2307 not_found:
2308         ini->smcr_version &= ~SMC_V2;
2309         ini->smcrv2.ib_dev_v2 = NULL;
2310         ini->check_smcrv2 = false;
2311 }
2312
2313 static int smc_find_rdma_v1_device_serv(struct smc_sock *new_smc,
2314                                         struct smc_clc_msg_proposal *pclc,
2315                                         struct smc_init_info *ini)
2316 {
2317         int rc;
2318
2319         if (!(ini->smcr_version & SMC_V1) || !smcr_indicated(ini->smc_type_v1))
2320                 return SMC_CLC_DECL_NOSMCDEV;
2321
2322         /* prepare RDMA check */
2323         memcpy(ini->peer_systemid, pclc->lcl.id_for_peer, SMC_SYSTEMID_LEN);
2324         memcpy(ini->peer_gid, pclc->lcl.gid, SMC_GID_SIZE);
2325         memcpy(ini->peer_mac, pclc->lcl.mac, ETH_ALEN);
2326         rc = smc_find_rdma_device(new_smc, ini);
2327         if (rc) {
2328                 /* no RDMA device found */
2329                 return SMC_CLC_DECL_NOSMCDEV;
2330         }
2331         rc = smc_listen_rdma_init(new_smc, ini);
2332         if (rc)
2333                 return rc;
2334         return smc_listen_rdma_reg(new_smc, ini->first_contact_local);
2335 }
2336
2337 /* determine the local device matching to proposal */
2338 static int smc_listen_find_device(struct smc_sock *new_smc,
2339                                   struct smc_clc_msg_proposal *pclc,
2340                                   struct smc_init_info *ini)
2341 {
2342         int prfx_rc;
2343
2344         /* check for ISM device matching V2 proposed device */
2345         smc_find_ism_v2_device_serv(new_smc, pclc, ini);
2346         if (ini->ism_dev[0])
2347                 return 0;
2348
2349         /* check for matching IP prefix and subnet length (V1) */
2350         prfx_rc = smc_listen_prfx_check(new_smc, pclc);
2351         if (prfx_rc)
2352                 smc_find_ism_store_rc(prfx_rc, ini);
2353
2354         /* get vlan id from IP device */
2355         if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
2356                 return ini->rc ?: SMC_CLC_DECL_GETVLANERR;
2357
2358         /* check for ISM device matching V1 proposed device */
2359         if (!prfx_rc)
2360                 smc_find_ism_v1_device_serv(new_smc, pclc, ini);
2361         if (ini->ism_dev[0])
2362                 return 0;
2363
2364         if (!smcr_indicated(pclc->hdr.typev1) &&
2365             !smcr_indicated(pclc->hdr.typev2))
2366                 /* skip RDMA and decline */
2367                 return ini->rc ?: SMC_CLC_DECL_NOSMCDDEV;
2368
2369         /* check if RDMA V2 is available */
2370         smc_find_rdma_v2_device_serv(new_smc, pclc, ini);
2371         if (ini->smcrv2.ib_dev_v2)
2372                 return 0;
2373
2374         /* check if RDMA V1 is available */
2375         if (!prfx_rc) {
2376                 int rc;
2377
2378                 rc = smc_find_rdma_v1_device_serv(new_smc, pclc, ini);
2379                 smc_find_ism_store_rc(rc, ini);
2380                 return (!rc) ? 0 : ini->rc;
2381         }
2382         return prfx_rc;
2383 }
2384
2385 /* listen worker: finish RDMA setup */
2386 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
2387                                   struct smc_clc_msg_accept_confirm *cclc,
2388                                   bool local_first,
2389                                   struct smc_init_info *ini)
2390 {
2391         struct smc_link *link = new_smc->conn.lnk;
2392         int reason_code = 0;
2393
2394         if (local_first)
2395                 smc_link_save_peer_info(link, cclc, ini);
2396
2397         if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc))
2398                 return SMC_CLC_DECL_ERR_RTOK;
2399
2400         if (local_first) {
2401                 if (smc_ib_ready_link(link))
2402                         return SMC_CLC_DECL_ERR_RDYLNK;
2403                 /* QP confirmation over RoCE fabric */
2404                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
2405                 reason_code = smcr_serv_conf_first_link(new_smc);
2406                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
2407         }
2408         return reason_code;
2409 }
2410
2411 /* setup for connection of server */
2412 static void smc_listen_work(struct work_struct *work)
2413 {
2414         struct smc_sock *new_smc = container_of(work, struct smc_sock,
2415                                                 smc_listen_work);
2416         struct socket *newclcsock = new_smc->clcsock;
2417         struct smc_clc_msg_accept_confirm *cclc;
2418         struct smc_clc_msg_proposal_area *buf;
2419         struct smc_clc_msg_proposal *pclc;
2420         struct smc_init_info *ini = NULL;
2421         u8 proposal_version = SMC_V1;
2422         u8 accept_version;
2423         int rc = 0;
2424
2425         if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
2426                 return smc_listen_out_err(new_smc);
2427
2428         if (new_smc->use_fallback) {
2429                 smc_listen_out_connected(new_smc);
2430                 return;
2431         }
2432
2433         /* check if peer is smc capable */
2434         if (!tcp_sk(newclcsock->sk)->syn_smc) {
2435                 rc = smc_switch_to_fallback(new_smc, SMC_CLC_DECL_PEERNOSMC);
2436                 if (rc)
2437                         smc_listen_out_err(new_smc);
2438                 else
2439                         smc_listen_out_connected(new_smc);
2440                 return;
2441         }
2442
2443         /* do inband token exchange -
2444          * wait for and receive SMC Proposal CLC message
2445          */
2446         buf = kzalloc(sizeof(*buf), GFP_KERNEL);
2447         if (!buf) {
2448                 rc = SMC_CLC_DECL_MEM;
2449                 goto out_decl;
2450         }
2451         pclc = (struct smc_clc_msg_proposal *)buf;
2452         rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
2453                               SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
2454         if (rc)
2455                 goto out_decl;
2456
2457         if (pclc->hdr.version > SMC_V1)
2458                 proposal_version = SMC_V2;
2459
2460         /* IPSec connections opt out of SMC optimizations */
2461         if (using_ipsec(new_smc)) {
2462                 rc = SMC_CLC_DECL_IPSEC;
2463                 goto out_decl;
2464         }
2465
2466         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
2467         if (!ini) {
2468                 rc = SMC_CLC_DECL_MEM;
2469                 goto out_decl;
2470         }
2471
2472         /* initial version checking */
2473         rc = smc_listen_v2_check(new_smc, pclc, ini);
2474         if (rc)
2475                 goto out_decl;
2476
2477         rc = smc_clc_srv_v2x_features_validate(new_smc, pclc, ini);
2478         if (rc)
2479                 goto out_decl;
2480
2481         mutex_lock(&smc_server_lgr_pending);
2482         smc_close_init(new_smc);
2483         smc_rx_init(new_smc);
2484         smc_tx_init(new_smc);
2485
2486         /* determine ISM or RoCE device used for connection */
2487         rc = smc_listen_find_device(new_smc, pclc, ini);
2488         if (rc)
2489                 goto out_unlock;
2490
2491         /* send SMC Accept CLC message */
2492         accept_version = ini->is_smcd ? ini->smcd_version : ini->smcr_version;
2493         rc = smc_clc_send_accept(new_smc, ini->first_contact_local,
2494                                  accept_version, ini->negotiated_eid, ini);
2495         if (rc)
2496                 goto out_unlock;
2497
2498         /* SMC-D does not need this lock any more */
2499         if (ini->is_smcd)
2500                 mutex_unlock(&smc_server_lgr_pending);
2501
2502         /* receive SMC Confirm CLC message */
2503         memset(buf, 0, sizeof(*buf));
2504         cclc = (struct smc_clc_msg_accept_confirm *)buf;
2505         rc = smc_clc_wait_msg(new_smc, cclc, sizeof(*buf),
2506                               SMC_CLC_CONFIRM, CLC_WAIT_TIME);
2507         if (rc) {
2508                 if (!ini->is_smcd)
2509                         goto out_unlock;
2510                 goto out_decl;
2511         }
2512
2513         rc = smc_clc_v2x_features_confirm_check(cclc, ini);
2514         if (rc) {
2515                 if (!ini->is_smcd)
2516                         goto out_unlock;
2517                 goto out_decl;
2518         }
2519
2520         /* fce smc release version is needed in smc_listen_rdma_finish,
2521          * so save fce info here.
2522          */
2523         smc_conn_save_peer_info_fce(new_smc, cclc);
2524
2525         /* finish worker */
2526         if (!ini->is_smcd) {
2527                 rc = smc_listen_rdma_finish(new_smc, cclc,
2528                                             ini->first_contact_local, ini);
2529                 if (rc)
2530                         goto out_unlock;
2531                 mutex_unlock(&smc_server_lgr_pending);
2532         }
2533         smc_conn_save_peer_info(new_smc, cclc);
2534
2535         if (ini->is_smcd &&
2536             smc_ism_support_dmb_nocopy(new_smc->conn.lgr->smcd)) {
2537                 rc = smcd_buf_attach(new_smc);
2538                 if (rc)
2539                         goto out_decl;
2540         }
2541
2542         smc_listen_out_connected(new_smc);
2543         SMC_STAT_SERV_SUCC_INC(sock_net(newclcsock->sk), ini);
2544         goto out_free;
2545
2546 out_unlock:
2547         mutex_unlock(&smc_server_lgr_pending);
2548 out_decl:
2549         smc_listen_decline(new_smc, rc, ini ? ini->first_contact_local : 0,
2550                            proposal_version);
2551 out_free:
2552         kfree(ini);
2553         kfree(buf);
2554 }
2555
2556 static void smc_tcp_listen_work(struct work_struct *work)
2557 {
2558         struct smc_sock *lsmc = container_of(work, struct smc_sock,
2559                                              tcp_listen_work);
2560         struct sock *lsk = &lsmc->sk;
2561         struct smc_sock *new_smc;
2562         int rc = 0;
2563
2564         lock_sock(lsk);
2565         while (lsk->sk_state == SMC_LISTEN) {
2566                 rc = smc_clcsock_accept(lsmc, &new_smc);
2567                 if (rc) /* clcsock accept queue empty or error */
2568                         goto out;
2569                 if (!new_smc)
2570                         continue;
2571
2572                 if (tcp_sk(new_smc->clcsock->sk)->syn_smc)
2573                         atomic_inc(&lsmc->queued_smc_hs);
2574
2575                 new_smc->listen_smc = lsmc;
2576                 new_smc->use_fallback = lsmc->use_fallback;
2577                 new_smc->fallback_rsn = lsmc->fallback_rsn;
2578                 sock_hold(lsk); /* sock_put in smc_listen_work */
2579                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
2580                 smc_copy_sock_settings_to_smc(new_smc);
2581                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
2582                 if (!queue_work(smc_hs_wq, &new_smc->smc_listen_work))
2583                         sock_put(&new_smc->sk);
2584         }
2585
2586 out:
2587         release_sock(lsk);
2588         sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
2589 }
2590
2591 static void smc_clcsock_data_ready(struct sock *listen_clcsock)
2592 {
2593         struct smc_sock *lsmc;
2594
2595         read_lock_bh(&listen_clcsock->sk_callback_lock);
2596         lsmc = smc_clcsock_user_data(listen_clcsock);
2597         if (!lsmc)
2598                 goto out;
2599         lsmc->clcsk_data_ready(listen_clcsock);
2600         if (lsmc->sk.sk_state == SMC_LISTEN) {
2601                 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
2602                 if (!queue_work(smc_tcp_ls_wq, &lsmc->tcp_listen_work))
2603                         sock_put(&lsmc->sk);
2604         }
2605 out:
2606         read_unlock_bh(&listen_clcsock->sk_callback_lock);
2607 }
2608
2609 int smc_listen(struct socket *sock, int backlog)
2610 {
2611         struct sock *sk = sock->sk;
2612         struct smc_sock *smc;
2613         int rc;
2614
2615         smc = smc_sk(sk);
2616         lock_sock(sk);
2617
2618         rc = -EINVAL;
2619         if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
2620             smc->connect_nonblock || sock->state != SS_UNCONNECTED)
2621                 goto out;
2622
2623         rc = 0;
2624         if (sk->sk_state == SMC_LISTEN) {
2625                 sk->sk_max_ack_backlog = backlog;
2626                 goto out;
2627         }
2628         /* some socket options are handled in core, so we could not apply
2629          * them to the clc socket -- copy smc socket options to clc socket
2630          */
2631         smc_copy_sock_settings_to_clc(smc);
2632         if (!smc->use_fallback)
2633                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
2634
2635         /* save original sk_data_ready function and establish
2636          * smc-specific sk_data_ready function
2637          */
2638         write_lock_bh(&smc->clcsock->sk->sk_callback_lock);
2639         smc->clcsock->sk->sk_user_data =
2640                 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
2641         smc_clcsock_replace_cb(&smc->clcsock->sk->sk_data_ready,
2642                                smc_clcsock_data_ready, &smc->clcsk_data_ready);
2643         write_unlock_bh(&smc->clcsock->sk->sk_callback_lock);
2644
2645         /* save original ops */
2646         smc->ori_af_ops = inet_csk(smc->clcsock->sk)->icsk_af_ops;
2647
2648         smc->af_ops = *smc->ori_af_ops;
2649         smc->af_ops.syn_recv_sock = smc_tcp_syn_recv_sock;
2650
2651         inet_csk(smc->clcsock->sk)->icsk_af_ops = &smc->af_ops;
2652
2653         if (smc->limit_smc_hs)
2654                 tcp_sk(smc->clcsock->sk)->smc_hs_congested = smc_hs_congested;
2655
2656         rc = kernel_listen(smc->clcsock, backlog);
2657         if (rc) {
2658                 write_lock_bh(&smc->clcsock->sk->sk_callback_lock);
2659                 smc_clcsock_restore_cb(&smc->clcsock->sk->sk_data_ready,
2660                                        &smc->clcsk_data_ready);
2661                 smc->clcsock->sk->sk_user_data = NULL;
2662                 write_unlock_bh(&smc->clcsock->sk->sk_callback_lock);
2663                 goto out;
2664         }
2665         sk->sk_max_ack_backlog = backlog;
2666         sk->sk_ack_backlog = 0;
2667         sk->sk_state = SMC_LISTEN;
2668
2669 out:
2670         release_sock(sk);
2671         return rc;
2672 }
2673
2674 int smc_accept(struct socket *sock, struct socket *new_sock,
2675                struct proto_accept_arg *arg)
2676 {
2677         struct sock *sk = sock->sk, *nsk;
2678         DECLARE_WAITQUEUE(wait, current);
2679         struct smc_sock *lsmc;
2680         long timeo;
2681         int rc = 0;
2682
2683         lsmc = smc_sk(sk);
2684         sock_hold(sk); /* sock_put below */
2685         lock_sock(sk);
2686
2687         if (lsmc->sk.sk_state != SMC_LISTEN) {
2688                 rc = -EINVAL;
2689                 release_sock(sk);
2690                 goto out;
2691         }
2692
2693         /* Wait for an incoming connection */
2694         timeo = sock_rcvtimeo(sk, arg->flags & O_NONBLOCK);
2695         add_wait_queue_exclusive(sk_sleep(sk), &wait);
2696         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
2697                 set_current_state(TASK_INTERRUPTIBLE);
2698                 if (!timeo) {
2699                         rc = -EAGAIN;
2700                         break;
2701                 }
2702                 release_sock(sk);
2703                 timeo = schedule_timeout(timeo);
2704                 /* wakeup by sk_data_ready in smc_listen_work() */
2705                 sched_annotate_sleep();
2706                 lock_sock(sk);
2707                 if (signal_pending(current)) {
2708                         rc = sock_intr_errno(timeo);
2709                         break;
2710                 }
2711         }
2712         set_current_state(TASK_RUNNING);
2713         remove_wait_queue(sk_sleep(sk), &wait);
2714
2715         if (!rc)
2716                 rc = sock_error(nsk);
2717         release_sock(sk);
2718         if (rc)
2719                 goto out;
2720
2721         if (lsmc->sockopt_defer_accept && !(arg->flags & O_NONBLOCK)) {
2722                 /* wait till data arrives on the socket */
2723                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
2724                                                                 MSEC_PER_SEC);
2725                 if (smc_sk(nsk)->use_fallback) {
2726                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
2727
2728                         lock_sock(clcsk);
2729                         if (skb_queue_empty(&clcsk->sk_receive_queue))
2730                                 sk_wait_data(clcsk, &timeo, NULL);
2731                         release_sock(clcsk);
2732                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
2733                         lock_sock(nsk);
2734                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
2735                         release_sock(nsk);
2736                 }
2737         }
2738
2739 out:
2740         sock_put(sk); /* sock_hold above */
2741         return rc;
2742 }
2743
2744 int smc_getname(struct socket *sock, struct sockaddr *addr,
2745                 int peer)
2746 {
2747         struct smc_sock *smc;
2748
2749         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
2750             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
2751                 return -ENOTCONN;
2752
2753         smc = smc_sk(sock->sk);
2754
2755         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
2756 }
2757
2758 int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
2759 {
2760         struct sock *sk = sock->sk;
2761         struct smc_sock *smc;
2762         int rc;
2763
2764         smc = smc_sk(sk);
2765         lock_sock(sk);
2766
2767         /* SMC does not support connect with fastopen */
2768         if (msg->msg_flags & MSG_FASTOPEN) {
2769                 /* not connected yet, fallback */
2770                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2771                         rc = smc_switch_to_fallback(smc, SMC_CLC_DECL_OPTUNSUPP);
2772                         if (rc)
2773                                 goto out;
2774                 } else {
2775                         rc = -EINVAL;
2776                         goto out;
2777                 }
2778         } else if ((sk->sk_state != SMC_ACTIVE) &&
2779                    (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2780                    (sk->sk_state != SMC_INIT)) {
2781                 rc = -EPIPE;
2782                 goto out;
2783         }
2784
2785         if (smc->use_fallback) {
2786                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
2787         } else {
2788                 rc = smc_tx_sendmsg(smc, msg, len);
2789                 SMC_STAT_TX_PAYLOAD(smc, len, rc);
2790         }
2791 out:
2792         release_sock(sk);
2793         return rc;
2794 }
2795
2796 int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
2797                 int flags)
2798 {
2799         struct sock *sk = sock->sk;
2800         struct smc_sock *smc;
2801         int rc = -ENOTCONN;
2802
2803         smc = smc_sk(sk);
2804         lock_sock(sk);
2805         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2806                 /* socket was connected before, no more data to read */
2807                 rc = 0;
2808                 goto out;
2809         }
2810         if ((sk->sk_state == SMC_INIT) ||
2811             (sk->sk_state == SMC_LISTEN) ||
2812             (sk->sk_state == SMC_CLOSED))
2813                 goto out;
2814
2815         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2816                 rc = 0;
2817                 goto out;
2818         }
2819
2820         if (smc->use_fallback) {
2821                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
2822         } else {
2823                 msg->msg_namelen = 0;
2824                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
2825                 SMC_STAT_RX_PAYLOAD(smc, rc, rc);
2826         }
2827
2828 out:
2829         release_sock(sk);
2830         return rc;
2831 }
2832
2833 static __poll_t smc_accept_poll(struct sock *parent)
2834 {
2835         struct smc_sock *isk = smc_sk(parent);
2836         __poll_t mask = 0;
2837
2838         spin_lock(&isk->accept_q_lock);
2839         if (!list_empty(&isk->accept_q))
2840                 mask = EPOLLIN | EPOLLRDNORM;
2841         spin_unlock(&isk->accept_q_lock);
2842
2843         return mask;
2844 }
2845
2846 __poll_t smc_poll(struct file *file, struct socket *sock,
2847                   poll_table *wait)
2848 {
2849         struct sock *sk = sock->sk;
2850         struct smc_sock *smc;
2851         __poll_t mask = 0;
2852
2853         if (!sk)
2854                 return EPOLLNVAL;
2855
2856         smc = smc_sk(sock->sk);
2857         if (smc->use_fallback) {
2858                 /* delegate to CLC child sock */
2859                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
2860                 sk->sk_err = smc->clcsock->sk->sk_err;
2861         } else {
2862                 if (sk->sk_state != SMC_CLOSED)
2863                         sock_poll_wait(file, sock, wait);
2864                 if (sk->sk_err)
2865                         mask |= EPOLLERR;
2866                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
2867                     (sk->sk_state == SMC_CLOSED))
2868                         mask |= EPOLLHUP;
2869                 if (sk->sk_state == SMC_LISTEN) {
2870                         /* woken up by sk_data_ready in smc_listen_work() */
2871                         mask |= smc_accept_poll(sk);
2872                 } else if (smc->use_fallback) { /* as result of connect_work()*/
2873                         mask |= smc->clcsock->ops->poll(file, smc->clcsock,
2874                                                            wait);
2875                         sk->sk_err = smc->clcsock->sk->sk_err;
2876                 } else {
2877                         if ((sk->sk_state != SMC_INIT &&
2878                              atomic_read(&smc->conn.sndbuf_space)) ||
2879                             sk->sk_shutdown & SEND_SHUTDOWN) {
2880                                 mask |= EPOLLOUT | EPOLLWRNORM;
2881                         } else {
2882                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
2883                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2884                         }
2885                         if (atomic_read(&smc->conn.bytes_to_rcv))
2886                                 mask |= EPOLLIN | EPOLLRDNORM;
2887                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2888                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2889                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
2890                                 mask |= EPOLLIN;
2891                         if (smc->conn.urg_state == SMC_URG_VALID)
2892                                 mask |= EPOLLPRI;
2893                 }
2894         }
2895
2896         return mask;
2897 }
2898
2899 int smc_shutdown(struct socket *sock, int how)
2900 {
2901         struct sock *sk = sock->sk;
2902         bool do_shutdown = true;
2903         struct smc_sock *smc;
2904         int rc = -EINVAL;
2905         int old_state;
2906         int rc1 = 0;
2907
2908         smc = smc_sk(sk);
2909
2910         if ((how < SHUT_RD) || (how > SHUT_RDWR))
2911                 return rc;
2912
2913         lock_sock(sk);
2914
2915         if (sock->state == SS_CONNECTING) {
2916                 if (sk->sk_state == SMC_ACTIVE)
2917                         sock->state = SS_CONNECTED;
2918                 else if (sk->sk_state == SMC_PEERCLOSEWAIT1 ||
2919                          sk->sk_state == SMC_PEERCLOSEWAIT2 ||
2920                          sk->sk_state == SMC_APPCLOSEWAIT1 ||
2921                          sk->sk_state == SMC_APPCLOSEWAIT2 ||
2922                          sk->sk_state == SMC_APPFINCLOSEWAIT)
2923                         sock->state = SS_DISCONNECTING;
2924         }
2925
2926         rc = -ENOTCONN;
2927         if ((sk->sk_state != SMC_ACTIVE) &&
2928             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
2929             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
2930             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2931             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
2932             (sk->sk_state != SMC_APPFINCLOSEWAIT))
2933                 goto out;
2934         if (smc->use_fallback) {
2935                 rc = kernel_sock_shutdown(smc->clcsock, how);
2936                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
2937                 if (sk->sk_shutdown == SHUTDOWN_MASK) {
2938                         sk->sk_state = SMC_CLOSED;
2939                         sk->sk_socket->state = SS_UNCONNECTED;
2940                         sock_put(sk);
2941                 }
2942                 goto out;
2943         }
2944         switch (how) {
2945         case SHUT_RDWR:         /* shutdown in both directions */
2946                 old_state = sk->sk_state;
2947                 rc = smc_close_active(smc);
2948                 if (old_state == SMC_ACTIVE &&
2949                     sk->sk_state == SMC_PEERCLOSEWAIT1)
2950                         do_shutdown = false;
2951                 break;
2952         case SHUT_WR:
2953                 rc = smc_close_shutdown_write(smc);
2954                 break;
2955         case SHUT_RD:
2956                 rc = 0;
2957                 /* nothing more to do because peer is not involved */
2958                 break;
2959         }
2960         if (do_shutdown && smc->clcsock)
2961                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
2962         /* map sock_shutdown_cmd constants to sk_shutdown value range */
2963         sk->sk_shutdown |= how + 1;
2964
2965         if (sk->sk_state == SMC_CLOSED)
2966                 sock->state = SS_UNCONNECTED;
2967         else
2968                 sock->state = SS_DISCONNECTING;
2969 out:
2970         release_sock(sk);
2971         return rc ? rc : rc1;
2972 }
2973
2974 static int __smc_getsockopt(struct socket *sock, int level, int optname,
2975                             char __user *optval, int __user *optlen)
2976 {
2977         struct smc_sock *smc;
2978         int val, len;
2979
2980         smc = smc_sk(sock->sk);
2981
2982         if (get_user(len, optlen))
2983                 return -EFAULT;
2984
2985         len = min_t(int, len, sizeof(int));
2986
2987         if (len < 0)
2988                 return -EINVAL;
2989
2990         switch (optname) {
2991         case SMC_LIMIT_HS:
2992                 val = smc->limit_smc_hs;
2993                 break;
2994         default:
2995                 return -EOPNOTSUPP;
2996         }
2997
2998         if (put_user(len, optlen))
2999                 return -EFAULT;
3000         if (copy_to_user(optval, &val, len))
3001                 return -EFAULT;
3002
3003         return 0;
3004 }
3005
3006 static int __smc_setsockopt(struct socket *sock, int level, int optname,
3007                             sockptr_t optval, unsigned int optlen)
3008 {
3009         struct sock *sk = sock->sk;
3010         struct smc_sock *smc;
3011         int val, rc;
3012
3013         smc = smc_sk(sk);
3014
3015         lock_sock(sk);
3016         switch (optname) {
3017         case SMC_LIMIT_HS:
3018                 if (optlen < sizeof(int)) {
3019                         rc = -EINVAL;
3020                         break;
3021                 }
3022                 if (copy_from_sockptr(&val, optval, sizeof(int))) {
3023                         rc = -EFAULT;
3024                         break;
3025                 }
3026
3027                 smc->limit_smc_hs = !!val;
3028                 rc = 0;
3029                 break;
3030         default:
3031                 rc = -EOPNOTSUPP;
3032                 break;
3033         }
3034         release_sock(sk);
3035
3036         return rc;
3037 }
3038
3039 int smc_setsockopt(struct socket *sock, int level, int optname,
3040                    sockptr_t optval, unsigned int optlen)
3041 {
3042         struct sock *sk = sock->sk;
3043         struct smc_sock *smc;
3044         int val, rc;
3045
3046         if (level == SOL_TCP && optname == TCP_ULP)
3047                 return -EOPNOTSUPP;
3048         else if (level == SOL_SMC)
3049                 return __smc_setsockopt(sock, level, optname, optval, optlen);
3050
3051         smc = smc_sk(sk);
3052
3053         /* generic setsockopts reaching us here always apply to the
3054          * CLC socket
3055          */
3056         mutex_lock(&smc->clcsock_release_lock);
3057         if (!smc->clcsock) {
3058                 mutex_unlock(&smc->clcsock_release_lock);
3059                 return -EBADF;
3060         }
3061         if (unlikely(!smc->clcsock->ops->setsockopt))
3062                 rc = -EOPNOTSUPP;
3063         else
3064                 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
3065                                                    optval, optlen);
3066         if (smc->clcsock->sk->sk_err) {
3067                 sk->sk_err = smc->clcsock->sk->sk_err;
3068                 sk_error_report(sk);
3069         }
3070         mutex_unlock(&smc->clcsock_release_lock);
3071
3072         if (optlen < sizeof(int))
3073                 return -EINVAL;
3074         if (copy_from_sockptr(&val, optval, sizeof(int)))
3075                 return -EFAULT;
3076
3077         lock_sock(sk);
3078         if (rc || smc->use_fallback)
3079                 goto out;
3080         switch (optname) {
3081         case TCP_FASTOPEN:
3082         case TCP_FASTOPEN_CONNECT:
3083         case TCP_FASTOPEN_KEY:
3084         case TCP_FASTOPEN_NO_COOKIE:
3085                 /* option not supported by SMC */
3086                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
3087                         rc = smc_switch_to_fallback(smc, SMC_CLC_DECL_OPTUNSUPP);
3088                 } else {
3089                         rc = -EINVAL;
3090                 }
3091                 break;
3092         case TCP_NODELAY:
3093                 if (sk->sk_state != SMC_INIT &&
3094                     sk->sk_state != SMC_LISTEN &&
3095                     sk->sk_state != SMC_CLOSED) {
3096                         if (val) {
3097                                 SMC_STAT_INC(smc, ndly_cnt);
3098                                 smc_tx_pending(&smc->conn);
3099                                 cancel_delayed_work(&smc->conn.tx_work);
3100                         }
3101                 }
3102                 break;
3103         case TCP_CORK:
3104                 if (sk->sk_state != SMC_INIT &&
3105                     sk->sk_state != SMC_LISTEN &&
3106                     sk->sk_state != SMC_CLOSED) {
3107                         if (!val) {
3108                                 SMC_STAT_INC(smc, cork_cnt);
3109                                 smc_tx_pending(&smc->conn);
3110                                 cancel_delayed_work(&smc->conn.tx_work);
3111                         }
3112                 }
3113                 break;
3114         case TCP_DEFER_ACCEPT:
3115                 smc->sockopt_defer_accept = val;
3116                 break;
3117         default:
3118                 break;
3119         }
3120 out:
3121         release_sock(sk);
3122
3123         return rc;
3124 }
3125
3126 int smc_getsockopt(struct socket *sock, int level, int optname,
3127                    char __user *optval, int __user *optlen)
3128 {
3129         struct smc_sock *smc;
3130         int rc;
3131
3132         if (level == SOL_SMC)
3133                 return __smc_getsockopt(sock, level, optname, optval, optlen);
3134
3135         smc = smc_sk(sock->sk);
3136         mutex_lock(&smc->clcsock_release_lock);
3137         if (!smc->clcsock) {
3138                 mutex_unlock(&smc->clcsock_release_lock);
3139                 return -EBADF;
3140         }
3141         /* socket options apply to the CLC socket */
3142         if (unlikely(!smc->clcsock->ops->getsockopt)) {
3143                 mutex_unlock(&smc->clcsock_release_lock);
3144                 return -EOPNOTSUPP;
3145         }
3146         rc = smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
3147                                            optval, optlen);
3148         mutex_unlock(&smc->clcsock_release_lock);
3149         return rc;
3150 }
3151
3152 int smc_ioctl(struct socket *sock, unsigned int cmd,
3153               unsigned long arg)
3154 {
3155         union smc_host_cursor cons, urg;
3156         struct smc_connection *conn;
3157         struct smc_sock *smc;
3158         int answ;
3159
3160         smc = smc_sk(sock->sk);
3161         conn = &smc->conn;
3162         lock_sock(&smc->sk);
3163         if (smc->use_fallback) {
3164                 if (!smc->clcsock) {
3165                         release_sock(&smc->sk);
3166                         return -EBADF;
3167                 }
3168                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
3169                 release_sock(&smc->sk);
3170                 return answ;
3171         }
3172         switch (cmd) {
3173         case SIOCINQ: /* same as FIONREAD */
3174                 if (smc->sk.sk_state == SMC_LISTEN) {
3175                         release_sock(&smc->sk);
3176                         return -EINVAL;
3177                 }
3178                 if (smc->sk.sk_state == SMC_INIT ||
3179                     smc->sk.sk_state == SMC_CLOSED)
3180                         answ = 0;
3181                 else
3182                         answ = atomic_read(&smc->conn.bytes_to_rcv);
3183                 break;
3184         case SIOCOUTQ:
3185                 /* output queue size (not send + not acked) */
3186                 if (smc->sk.sk_state == SMC_LISTEN) {
3187                         release_sock(&smc->sk);
3188                         return -EINVAL;
3189                 }
3190                 if (smc->sk.sk_state == SMC_INIT ||
3191                     smc->sk.sk_state == SMC_CLOSED)
3192                         answ = 0;
3193                 else
3194                         answ = smc->conn.sndbuf_desc->len -
3195                                         atomic_read(&smc->conn.sndbuf_space);
3196                 break;
3197         case SIOCOUTQNSD:
3198                 /* output queue size (not send only) */
3199                 if (smc->sk.sk_state == SMC_LISTEN) {
3200                         release_sock(&smc->sk);
3201                         return -EINVAL;
3202                 }
3203                 if (smc->sk.sk_state == SMC_INIT ||
3204                     smc->sk.sk_state == SMC_CLOSED)
3205                         answ = 0;
3206                 else
3207                         answ = smc_tx_prepared_sends(&smc->conn);
3208                 break;
3209         case SIOCATMARK:
3210                 if (smc->sk.sk_state == SMC_LISTEN) {
3211                         release_sock(&smc->sk);
3212                         return -EINVAL;
3213                 }
3214                 if (smc->sk.sk_state == SMC_INIT ||
3215                     smc->sk.sk_state == SMC_CLOSED) {
3216                         answ = 0;
3217                 } else {
3218                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
3219                         smc_curs_copy(&urg, &conn->urg_curs, conn);
3220                         answ = smc_curs_diff(conn->rmb_desc->len,
3221                                              &cons, &urg) == 1;
3222                 }
3223                 break;
3224         default:
3225                 release_sock(&smc->sk);
3226                 return -ENOIOCTLCMD;
3227         }
3228         release_sock(&smc->sk);
3229
3230         return put_user(answ, (int __user *)arg);
3231 }
3232
3233 /* Map the affected portions of the rmbe into an spd, note the number of bytes
3234  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
3235  * updates till whenever a respective page has been fully processed.
3236  * Note that subsequent recv() calls have to wait till all splice() processing
3237  * completed.
3238  */
3239 ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
3240                         struct pipe_inode_info *pipe, size_t len,
3241                         unsigned int flags)
3242 {
3243         struct sock *sk = sock->sk;
3244         struct smc_sock *smc;
3245         int rc = -ENOTCONN;
3246
3247         smc = smc_sk(sk);
3248         lock_sock(sk);
3249         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
3250                 /* socket was connected before, no more data to read */
3251                 rc = 0;
3252                 goto out;
3253         }
3254         if (sk->sk_state == SMC_INIT ||
3255             sk->sk_state == SMC_LISTEN ||
3256             sk->sk_state == SMC_CLOSED)
3257                 goto out;
3258
3259         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
3260                 rc = 0;
3261                 goto out;
3262         }
3263
3264         if (smc->use_fallback) {
3265                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
3266                                                     pipe, len, flags);
3267         } else {
3268                 if (*ppos) {
3269                         rc = -ESPIPE;
3270                         goto out;
3271                 }
3272                 if (flags & SPLICE_F_NONBLOCK)
3273                         flags = MSG_DONTWAIT;
3274                 else
3275                         flags = 0;
3276                 SMC_STAT_INC(smc, splice_cnt);
3277                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
3278         }
3279 out:
3280         release_sock(sk);
3281
3282         return rc;
3283 }
3284
3285 /* must look like tcp */
3286 static const struct proto_ops smc_sock_ops = {
3287         .family         = PF_SMC,
3288         .owner          = THIS_MODULE,
3289         .release        = smc_release,
3290         .bind           = smc_bind,
3291         .connect        = smc_connect,
3292         .socketpair     = sock_no_socketpair,
3293         .accept         = smc_accept,
3294         .getname        = smc_getname,
3295         .poll           = smc_poll,
3296         .ioctl          = smc_ioctl,
3297         .listen         = smc_listen,
3298         .shutdown       = smc_shutdown,
3299         .setsockopt     = smc_setsockopt,
3300         .getsockopt     = smc_getsockopt,
3301         .sendmsg        = smc_sendmsg,
3302         .recvmsg        = smc_recvmsg,
3303         .mmap           = sock_no_mmap,
3304         .splice_read    = smc_splice_read,
3305 };
3306
3307 int smc_create_clcsk(struct net *net, struct sock *sk, int family)
3308 {
3309         struct smc_sock *smc = smc_sk(sk);
3310         int rc;
3311
3312         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
3313                               &smc->clcsock);
3314         if (rc)
3315                 return rc;
3316
3317         /* smc_clcsock_release() does not wait smc->clcsock->sk's
3318          * destruction;  its sk_state might not be TCP_CLOSE after
3319          * smc->sk is close()d, and TCP timers can be fired later,
3320          * which need net ref.
3321          */
3322         sk = smc->clcsock->sk;
3323         __netns_tracker_free(net, &sk->ns_tracker, false);
3324         sk->sk_net_refcnt = 1;
3325         get_net_track(net, &sk->ns_tracker, GFP_KERNEL);
3326         sock_inuse_add(net, 1);
3327         return 0;
3328 }
3329
3330 static int __smc_create(struct net *net, struct socket *sock, int protocol,
3331                         int kern, struct socket *clcsock)
3332 {
3333         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
3334         struct smc_sock *smc;
3335         struct sock *sk;
3336         int rc;
3337
3338         rc = -ESOCKTNOSUPPORT;
3339         if (sock->type != SOCK_STREAM)
3340                 goto out;
3341
3342         rc = -EPROTONOSUPPORT;
3343         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
3344                 goto out;
3345
3346         rc = -ENOBUFS;
3347         sock->ops = &smc_sock_ops;
3348         sock->state = SS_UNCONNECTED;
3349         sk = smc_sock_alloc(net, sock, protocol);
3350         if (!sk)
3351                 goto out;
3352
3353         /* create internal TCP socket for CLC handshake and fallback */
3354         smc = smc_sk(sk);
3355
3356         rc = 0;
3357         if (clcsock)
3358                 smc->clcsock = clcsock;
3359         else
3360                 rc = smc_create_clcsk(net, sk, family);
3361
3362         if (rc)
3363                 sk_common_release(sk);
3364 out:
3365         return rc;
3366 }
3367
3368 static int smc_create(struct net *net, struct socket *sock, int protocol,
3369                       int kern)
3370 {
3371         return __smc_create(net, sock, protocol, kern, NULL);
3372 }
3373
3374 static const struct net_proto_family smc_sock_family_ops = {
3375         .family = PF_SMC,
3376         .owner  = THIS_MODULE,
3377         .create = smc_create,
3378 };
3379
3380 static int smc_ulp_init(struct sock *sk)
3381 {
3382         struct socket *tcp = sk->sk_socket;
3383         struct net *net = sock_net(sk);
3384         struct socket *smcsock;
3385         int protocol, ret;
3386
3387         /* only TCP can be replaced */
3388         if (tcp->type != SOCK_STREAM || sk->sk_protocol != IPPROTO_TCP ||
3389             (sk->sk_family != AF_INET && sk->sk_family != AF_INET6))
3390                 return -ESOCKTNOSUPPORT;
3391         /* don't handle wq now */
3392         if (tcp->state != SS_UNCONNECTED || !tcp->file || tcp->wq.fasync_list)
3393                 return -ENOTCONN;
3394
3395         if (sk->sk_family == AF_INET)
3396                 protocol = SMCPROTO_SMC;
3397         else
3398                 protocol = SMCPROTO_SMC6;
3399
3400         smcsock = sock_alloc();
3401         if (!smcsock)
3402                 return -ENFILE;
3403
3404         smcsock->type = SOCK_STREAM;
3405         __module_get(THIS_MODULE); /* tried in __tcp_ulp_find_autoload */
3406         ret = __smc_create(net, smcsock, protocol, 1, tcp);
3407         if (ret) {
3408                 sock_release(smcsock); /* module_put() which ops won't be NULL */
3409                 return ret;
3410         }
3411
3412         /* replace tcp socket to smc */
3413         smcsock->file = tcp->file;
3414         smcsock->file->private_data = smcsock;
3415         smcsock->file->f_inode = SOCK_INODE(smcsock); /* replace inode when sock_close */
3416         smcsock->file->f_path.dentry->d_inode = SOCK_INODE(smcsock); /* dput() in __fput */
3417         tcp->file = NULL;
3418
3419         return ret;
3420 }
3421
3422 static void smc_ulp_clone(const struct request_sock *req, struct sock *newsk,
3423                           const gfp_t priority)
3424 {
3425         struct inet_connection_sock *icsk = inet_csk(newsk);
3426
3427         /* don't inherit ulp ops to child when listen */
3428         icsk->icsk_ulp_ops = NULL;
3429 }
3430
3431 static struct tcp_ulp_ops smc_ulp_ops __read_mostly = {
3432         .name           = "smc",
3433         .owner          = THIS_MODULE,
3434         .init           = smc_ulp_init,
3435         .clone          = smc_ulp_clone,
3436 };
3437
3438 unsigned int smc_net_id;
3439
3440 static __net_init int smc_net_init(struct net *net)
3441 {
3442         int rc;
3443
3444         rc = smc_sysctl_net_init(net);
3445         if (rc)
3446                 return rc;
3447         return smc_pnet_net_init(net);
3448 }
3449
3450 static void __net_exit smc_net_exit(struct net *net)
3451 {
3452         smc_sysctl_net_exit(net);
3453         smc_pnet_net_exit(net);
3454 }
3455
3456 static __net_init int smc_net_stat_init(struct net *net)
3457 {
3458         return smc_stats_init(net);
3459 }
3460
3461 static void __net_exit smc_net_stat_exit(struct net *net)
3462 {
3463         smc_stats_exit(net);
3464 }
3465
3466 static struct pernet_operations smc_net_ops = {
3467         .init = smc_net_init,
3468         .exit = smc_net_exit,
3469         .id   = &smc_net_id,
3470         .size = sizeof(struct smc_net),
3471 };
3472
3473 static struct pernet_operations smc_net_stat_ops = {
3474         .init = smc_net_stat_init,
3475         .exit = smc_net_stat_exit,
3476 };
3477
3478 static int __init smc_init(void)
3479 {
3480         int rc;
3481
3482         rc = register_pernet_subsys(&smc_net_ops);
3483         if (rc)
3484                 return rc;
3485
3486         rc = register_pernet_subsys(&smc_net_stat_ops);
3487         if (rc)
3488                 goto out_pernet_subsys;
3489
3490         rc = smc_ism_init();
3491         if (rc)
3492                 goto out_pernet_subsys_stat;
3493         smc_clc_init();
3494
3495         rc = smc_nl_init();
3496         if (rc)
3497                 goto out_ism;
3498
3499         rc = smc_pnet_init();
3500         if (rc)
3501                 goto out_nl;
3502
3503         rc = -ENOMEM;
3504
3505         smc_tcp_ls_wq = alloc_workqueue("smc_tcp_ls_wq", 0, 0);
3506         if (!smc_tcp_ls_wq)
3507                 goto out_pnet;
3508
3509         smc_hs_wq = alloc_workqueue("smc_hs_wq", 0, 0);
3510         if (!smc_hs_wq)
3511                 goto out_alloc_tcp_ls_wq;
3512
3513         smc_close_wq = alloc_workqueue("smc_close_wq", 0, 0);
3514         if (!smc_close_wq)
3515                 goto out_alloc_hs_wq;
3516
3517         rc = smc_core_init();
3518         if (rc) {
3519                 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
3520                 goto out_alloc_wqs;
3521         }
3522
3523         rc = smc_llc_init();
3524         if (rc) {
3525                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
3526                 goto out_core;
3527         }
3528
3529         rc = smc_cdc_init();
3530         if (rc) {
3531                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
3532                 goto out_core;
3533         }
3534
3535         rc = proto_register(&smc_proto, 1);
3536         if (rc) {
3537                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
3538                 goto out_core;
3539         }
3540
3541         rc = proto_register(&smc_proto6, 1);
3542         if (rc) {
3543                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
3544                 goto out_proto;
3545         }
3546
3547         rc = sock_register(&smc_sock_family_ops);
3548         if (rc) {
3549                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
3550                 goto out_proto6;
3551         }
3552         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
3553         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
3554
3555         rc = smc_ib_register_client();
3556         if (rc) {
3557                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
3558                 goto out_sock;
3559         }
3560
3561         rc = smc_loopback_init();
3562         if (rc) {
3563                 pr_err("%s: smc_loopback_init fails with %d\n", __func__, rc);
3564                 goto out_ib;
3565         }
3566
3567         rc = tcp_register_ulp(&smc_ulp_ops);
3568         if (rc) {
3569                 pr_err("%s: tcp_ulp_register fails with %d\n", __func__, rc);
3570                 goto out_lo;
3571         }
3572         rc = smc_inet_init();
3573         if (rc) {
3574                 pr_err("%s: smc_inet_init fails with %d\n", __func__, rc);
3575                 goto out_ulp;
3576         }
3577         static_branch_enable(&tcp_have_smc);
3578         return 0;
3579 out_ulp:
3580         tcp_unregister_ulp(&smc_ulp_ops);
3581 out_lo:
3582         smc_loopback_exit();
3583 out_ib:
3584         smc_ib_unregister_client();
3585 out_sock:
3586         sock_unregister(PF_SMC);
3587 out_proto6:
3588         proto_unregister(&smc_proto6);
3589 out_proto:
3590         proto_unregister(&smc_proto);
3591 out_core:
3592         smc_core_exit();
3593 out_alloc_wqs:
3594         destroy_workqueue(smc_close_wq);
3595 out_alloc_hs_wq:
3596         destroy_workqueue(smc_hs_wq);
3597 out_alloc_tcp_ls_wq:
3598         destroy_workqueue(smc_tcp_ls_wq);
3599 out_pnet:
3600         smc_pnet_exit();
3601 out_nl:
3602         smc_nl_exit();
3603 out_ism:
3604         smc_clc_exit();
3605         smc_ism_exit();
3606 out_pernet_subsys_stat:
3607         unregister_pernet_subsys(&smc_net_stat_ops);
3608 out_pernet_subsys:
3609         unregister_pernet_subsys(&smc_net_ops);
3610
3611         return rc;
3612 }
3613
3614 static void __exit smc_exit(void)
3615 {
3616         static_branch_disable(&tcp_have_smc);
3617         smc_inet_exit();
3618         tcp_unregister_ulp(&smc_ulp_ops);
3619         sock_unregister(PF_SMC);
3620         smc_core_exit();
3621         smc_loopback_exit();
3622         smc_ib_unregister_client();
3623         smc_ism_exit();
3624         destroy_workqueue(smc_close_wq);
3625         destroy_workqueue(smc_tcp_ls_wq);
3626         destroy_workqueue(smc_hs_wq);
3627         proto_unregister(&smc_proto6);
3628         proto_unregister(&smc_proto);
3629         smc_pnet_exit();
3630         smc_nl_exit();
3631         smc_clc_exit();
3632         unregister_pernet_subsys(&smc_net_stat_ops);
3633         unregister_pernet_subsys(&smc_net_ops);
3634         rcu_barrier();
3635 }
3636
3637 module_init(smc_init);
3638 module_exit(smc_exit);
3639
3640 MODULE_AUTHOR("Ursula Braun <[email protected]>");
3641 MODULE_DESCRIPTION("smc socket address family");
3642 MODULE_LICENSE("GPL");
3643 MODULE_ALIAS_NETPROTO(PF_SMC);
3644 MODULE_ALIAS_TCP_ULP("smc");
3645 /* 256 for IPPROTO_SMC and 1 for SOCK_STREAM */
3646 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_INET, 256, 1);
3647 #if IS_ENABLED(CONFIG_IPV6)
3648 MODULE_ALIAS_NET_PF_PROTO_TYPE(PF_INET6, 256, 1);
3649 #endif /* CONFIG_IPV6 */
3650 MODULE_ALIAS_GENL_FAMILY(SMC_GENL_FAMILY_NAME);