net/smc/af_smc.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  *  AF_SMC protocol family socket handler keeping the AF_INET sock address type
   6  *  applies to SOCK_STREAM sockets only
   7  *  offers an alternative communication option for TCP-protocol sockets
   8  *  applicable with RoCE-cards only
   9  *
  10  *  Initial restrictions:
  11  *    - support for alternate links postponed
  12  *
  13  *  Copyright IBM Corp. 2016, 2018
  14  *
  15  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  16  *              based on prototype from Frank Blaschka
  17  */
  18
  19 #define KMSG_COMPONENT "smc"
  20 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  21
  22 #include <linux/module.h>
  23 #include <linux/socket.h>
  24 #include <linux/workqueue.h>
  25 #include <linux/in.h>
  26 #include <linux/sched/signal.h>
  27 #include <linux/if_vlan.h>
  28 #include <linux/rcupdate_wait.h>
  29 #include <linux/ctype.h>
  30
  31 #include <net/sock.h>
  32 #include <net/tcp.h>
  33 #include <net/smc.h>
  34 #include <asm/ioctls.h>
  35
  36 #include <net/net_namespace.h>
  37 #include <net/netns/generic.h>
  38 #include "smc_netns.h"
  39
  40 #include "smc.h"
  41 #include "smc_clc.h"
  42 #include "smc_llc.h"
  43 #include "smc_cdc.h"
  44 #include "smc_core.h"
  45 #include "smc_ib.h"
  46 #include "smc_ism.h"
  47 #include "smc_pnet.h"
  48 #include "smc_netlink.h"
  49 #include "smc_tx.h"
  50 #include "smc_rx.h"
  51 #include "smc_close.h"
  52 #include "smc_stats.h"
  53 #include "smc_tracepoint.h"
  54 #include "smc_sysctl.h"
  55
  56 static DEFINE_MUTEX(smc_server_lgr_pending);    /* serialize link group
  57                                                  * creation on server
  58                                                  */
  59 static DEFINE_MUTEX(smc_client_lgr_pending);    /* serialize link group
  60                                                  * creation on client
  61                                                  */
  62
  63 static struct workqueue_struct  *smc_tcp_ls_wq; /* wq for tcp listen work */
  64 struct workqueue_struct *smc_hs_wq;     /* wq for handshake work */
  65 struct workqueue_struct *smc_close_wq;  /* wq for close work */
  66
  67 static void smc_tcp_listen_work(struct work_struct *);
  68 static void smc_connect_work(struct work_struct *);
  69
  70 int smc_nl_dump_hs_limitation(struct sk_buff *skb, struct netlink_callback *cb)
  71 {
  72         struct smc_nl_dmp_ctx *cb_ctx = smc_nl_dmp_ctx(cb);
  73         void *hdr;
  74
  75         if (cb_ctx->pos[0])
  76                 goto out;
  77
  78         hdr = genlmsg_put(skb, NETLINK_CB(cb->skb).portid, cb->nlh->nlmsg_seq,
  79                           &smc_gen_nl_family, NLM_F_MULTI,
  80                           SMC_NETLINK_DUMP_HS_LIMITATION);
  81         if (!hdr)
  82                 return -ENOMEM;
  83
  84         if (nla_put_u8(skb, SMC_NLA_HS_LIMITATION_ENABLED,
  85                        sock_net(skb->sk)->smc.limit_smc_hs))
  86                 goto err;
  87
  88         genlmsg_end(skb, hdr);
  89         cb_ctx->pos[0] = 1;
  90 out:
  91         return skb->len;
  92 err:
  93         genlmsg_cancel(skb, hdr);
  94         return -EMSGSIZE;
  95 }
  96
  97 int smc_nl_enable_hs_limitation(struct sk_buff *skb, struct genl_info *info)
  98 {
  99         sock_net(skb->sk)->smc.limit_smc_hs = true;
 100         return 0;
 101 }
 102
 103 int smc_nl_disable_hs_limitation(struct sk_buff *skb, struct genl_info *info)
 104 {
 105         sock_net(skb->sk)->smc.limit_smc_hs = false;
 106         return 0;
 107 }
 108
 109 static void smc_set_keepalive(struct sock *sk, int val)
 110 {
 111         struct smc_sock *smc = smc_sk(sk);
 112
 113         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
 114 }
 115
 116 static struct sock *smc_tcp_syn_recv_sock(const struct sock *sk,
 117                                           struct sk_buff *skb,
 118                                           struct request_sock *req,
 119                                           struct dst_entry *dst,
 120                                           struct request_sock *req_unhash,
 121                                           bool *own_req)
 122 {
 123         struct smc_sock *smc;
 124         struct sock *child;
 125
 126         smc = smc_clcsock_user_data(sk);
 127
 128         if (READ_ONCE(sk->sk_ack_backlog) + atomic_read(&smc->queued_smc_hs) >
 129                                 sk->sk_max_ack_backlog)
 130                 goto drop;
 131
 132         if (sk_acceptq_is_full(&smc->sk)) {
 133                 NET_INC_STATS(sock_net(sk), LINUX_MIB_LISTENOVERFLOWS);
 134                 goto drop;
 135         }
 136
 137         /* passthrough to original syn recv sock fct */
 138         child = smc->ori_af_ops->syn_recv_sock(sk, skb, req, dst, req_unhash,
 139                                                own_req);
 140         /* child must not inherit smc or its ops */
 141         if (child) {
 142                 rcu_assign_sk_user_data(child, NULL);
 143
 144                 /* v4-mapped sockets don't inherit parent ops. Don't restore. */
 145                 if (inet_csk(child)->icsk_af_ops == inet_csk(sk)->icsk_af_ops)
 146                         inet_csk(child)->icsk_af_ops = smc->ori_af_ops;
 147         }
 148         return child;
 149
 150 drop:
 151         dst_release(dst);
 152         tcp_listendrop(sk);
 153         return NULL;
 154 }
 155
 156 static bool smc_hs_congested(const struct sock *sk)
 157 {
 158         const struct smc_sock *smc;
 159
 160         smc = smc_clcsock_user_data(sk);
 161
 162         if (!smc)
 163                 return true;
 164
 165         if (workqueue_congested(WORK_CPU_UNBOUND, smc_hs_wq))
 166                 return true;
 167
 168         return false;
 169 }
 170
 171 static struct smc_hashinfo smc_v4_hashinfo = {
 172         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
 173 };
 174
 175 static struct smc_hashinfo smc_v6_hashinfo = {
 176         .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
 177 };
 178
 179 int smc_hash_sk(struct sock *sk)
 180 {
 181         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
 182         struct hlist_head *head;
 183
 184         head = &h->ht;
 185
 186         write_lock_bh(&h->lock);
 187         sk_add_node(sk, head);
 188         write_unlock_bh(&h->lock);
 189         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
 190
 191         return 0;
 192 }
 193 EXPORT_SYMBOL_GPL(smc_hash_sk);
 194
 195 void smc_unhash_sk(struct sock *sk)
 196 {
 197         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
 198
 199         write_lock_bh(&h->lock);
 200         if (sk_del_node_init(sk))
 201                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
 202         write_unlock_bh(&h->lock);
 203 }
 204 EXPORT_SYMBOL_GPL(smc_unhash_sk);
 205
 206 /* This will be called before user really release sock_lock. So do the
 207  * work which we didn't do because of user hold the sock_lock in the
 208  * BH context
 209  */
 210 static void smc_release_cb(struct sock *sk)
 211 {
 212         struct smc_sock *smc = smc_sk(sk);
 213
 214         if (smc->conn.tx_in_release_sock) {
 215                 smc_tx_pending(&smc->conn);
 216                 smc->conn.tx_in_release_sock = false;
 217         }
 218 }
 219
 220 struct proto smc_proto = {
 221         .name           = "SMC",
 222         .owner          = THIS_MODULE,
 223         .keepalive      = smc_set_keepalive,
 224         .hash           = smc_hash_sk,
 225         .unhash         = smc_unhash_sk,
 226         .release_cb     = smc_release_cb,
 227         .obj_size       = sizeof(struct smc_sock),
 228         .h.smc_hash     = &smc_v4_hashinfo,
 229         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 230 };
 231 EXPORT_SYMBOL_GPL(smc_proto);
 232
 233 struct proto smc_proto6 = {
 234         .name           = "SMC6",
 235         .owner          = THIS_MODULE,
 236         .keepalive      = smc_set_keepalive,
 237         .hash           = smc_hash_sk,
 238         .unhash         = smc_unhash_sk,
 239         .release_cb     = smc_release_cb,
 240         .obj_size       = sizeof(struct smc_sock),
 241         .h.smc_hash     = &smc_v6_hashinfo,
 242         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 243 };
 244 EXPORT_SYMBOL_GPL(smc_proto6);
 245
 246 static void smc_fback_restore_callbacks(struct smc_sock *smc)
 247 {
 248         struct sock *clcsk = smc->clcsock->sk;
 249
 250         write_lock_bh(&clcsk->sk_callback_lock);
 251         clcsk->sk_user_data = NULL;
 252
 253         smc_clcsock_restore_cb(&clcsk->sk_state_change, &smc->clcsk_state_change);
 254         smc_clcsock_restore_cb(&clcsk->sk_data_ready, &smc->clcsk_data_ready);
 255         smc_clcsock_restore_cb(&clcsk->sk_write_space, &smc->clcsk_write_space);
 256         smc_clcsock_restore_cb(&clcsk->sk_error_report, &smc->clcsk_error_report);
 257
 258         write_unlock_bh(&clcsk->sk_callback_lock);
 259 }
 260
 261 static void smc_restore_fallback_changes(struct smc_sock *smc)
 262 {
 263         if (smc->clcsock->file) { /* non-accepted sockets have no file yet */
 264                 smc->clcsock->file->private_data = smc->sk.sk_socket;
 265                 smc->clcsock->file = NULL;
 266                 smc_fback_restore_callbacks(smc);
 267         }
 268 }
 269
 270 static int __smc_release(struct smc_sock *smc)
 271 {
 272         struct sock *sk = &smc->sk;
 273         int rc = 0;
 274
 275         if (!smc->use_fallback) {
 276                 rc = smc_close_active(smc);
 277                 smc_sock_set_flag(sk, SOCK_DEAD);
 278                 sk->sk_shutdown |= SHUTDOWN_MASK;
 279         } else {
 280                 if (sk->sk_state != SMC_CLOSED) {
 281                         if (sk->sk_state != SMC_LISTEN &&
 282                             sk->sk_state != SMC_INIT)
 283                                 sock_put(sk); /* passive closing */
 284                         if (sk->sk_state == SMC_LISTEN) {
 285                                 /* wake up clcsock accept */
 286                                 rc = kernel_sock_shutdown(smc->clcsock,
 287                                                           SHUT_RDWR);
 288                         }
 289                         sk->sk_state = SMC_CLOSED;
 290                         sk->sk_state_change(sk);
 291                 }
 292                 smc_restore_fallback_changes(smc);
 293         }
 294
 295         sk->sk_prot->unhash(sk);
 296
 297         if (sk->sk_state == SMC_CLOSED) {
 298                 if (smc->clcsock) {
 299                         release_sock(sk);
 300                         smc_clcsock_release(smc);
 301                         lock_sock(sk);
 302                 }
 303                 if (!smc->use_fallback)
 304                         smc_conn_free(&smc->conn);
 305         }
 306
 307         return rc;
 308 }
 309
 310 static int smc_release(struct socket *sock)
 311 {
 312         struct sock *sk = sock->sk;
 313         struct smc_sock *smc;
 314         int old_state, rc = 0;
 315
 316         if (!sk)
 317                 goto out;
 318
 319         sock_hold(sk); /* sock_put below */
 320         smc = smc_sk(sk);
 321
 322         old_state = sk->sk_state;
 323
 324         /* cleanup for a dangling non-blocking connect */
 325         if (smc->connect_nonblock && old_state == SMC_INIT)
 326                 tcp_abort(smc->clcsock->sk, ECONNABORTED);
 327
 328         if (cancel_work_sync(&smc->connect_work))
 329                 sock_put(&smc->sk); /* sock_hold in smc_connect for passive closing */
 330
 331         if (sk->sk_state == SMC_LISTEN)
 332                 /* smc_close_non_accepted() is called and acquires
 333                  * sock lock for child sockets again
 334                  */
 335                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 336         else
 337                 lock_sock(sk);
 338
 339         if (old_state == SMC_INIT && sk->sk_state == SMC_ACTIVE &&
 340             !smc->use_fallback)
 341                 smc_close_active_abort(smc);
 342
 343         rc = __smc_release(smc);
 344
 345         /* detach socket */
 346         sock_orphan(sk);
 347         sock->sk = NULL;
 348         release_sock(sk);
 349
 350         sock_put(sk); /* sock_hold above */
 351         sock_put(sk); /* final sock_put */
 352 out:
 353         return rc;
 354 }
 355
 356 static void smc_destruct(struct sock *sk)
 357 {
 358         if (sk->sk_state != SMC_CLOSED)
 359                 return;
 360         if (!sock_flag(sk, SOCK_DEAD))
 361                 return;
 362
 363         sk_refcnt_debug_dec(sk);
 364 }
 365
 366 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 367                                    int protocol)
 368 {
 369         struct smc_sock *smc;
 370         struct proto *prot;
 371         struct sock *sk;
 372
 373         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
 374         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
 375         if (!sk)
 376                 return NULL;
 377
 378         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 379         sk->sk_state = SMC_INIT;
 380         sk->sk_destruct = smc_destruct;
 381         sk->sk_protocol = protocol;
 382         WRITE_ONCE(sk->sk_sndbuf, 2 * READ_ONCE(net->smc.sysctl_wmem));
 383         WRITE_ONCE(sk->sk_rcvbuf, 2 * READ_ONCE(net->smc.sysctl_rmem));
 384         smc = smc_sk(sk);
 385         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 386         INIT_WORK(&smc->connect_work, smc_connect_work);
 387         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 388         INIT_LIST_HEAD(&smc->accept_q);
 389         spin_lock_init(&smc->accept_q_lock);
 390         spin_lock_init(&smc->conn.send_lock);
 391         sk->sk_prot->hash(sk);
 392         sk_refcnt_debug_inc(sk);
 393         mutex_init(&smc->clcsock_release_lock);
 394         smc_init_saved_callbacks(smc);
 395
 396         return sk;
 397 }
 398
 399 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 400                     int addr_len)
 401 {
 402         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 403         struct sock *sk = sock->sk;
 404         struct smc_sock *smc;
 405         int rc;
 406
 407         smc = smc_sk(sk);
 408
 409         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 410         rc = -EINVAL;
 411         if (addr_len < sizeof(struct sockaddr_in))
 412                 goto out;
 413
 414         rc = -EAFNOSUPPORT;
 415         if (addr->sin_family != AF_INET &&
 416             addr->sin_family != AF_INET6 &&
 417             addr->sin_family != AF_UNSPEC)
 418                 goto out;
 419         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 420         if (addr->sin_family == AF_UNSPEC &&
 421             addr->sin_addr.s_addr != htonl(INADDR_ANY))
 422                 goto out;
 423
 424         lock_sock(sk);
 425
 426         /* Check if socket is already active */
 427         rc = -EINVAL;
 428         if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
 429                 goto out_rel;
 430
 431         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 432         smc->clcsock->sk->sk_reuseport = sk->sk_reuseport;
 433         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 434
 435 out_rel:
 436         release_sock(sk);
 437 out:
 438         return rc;
 439 }
 440
 441 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 442  * clc socket (since smc is not called for these options from net/core)
 443  */
 444
 445 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 446                              (1UL << SOCK_KEEPOPEN) | \
 447                              (1UL << SOCK_LINGER) | \
 448                              (1UL << SOCK_BROADCAST) | \
 449                              (1UL << SOCK_TIMESTAMP) | \
 450                              (1UL << SOCK_DBG) | \
 451                              (1UL << SOCK_RCVTSTAMP) | \
 452                              (1UL << SOCK_RCVTSTAMPNS) | \
 453                              (1UL << SOCK_LOCALROUTE) | \
 454                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 455                              (1UL << SOCK_RXQ_OVFL) | \
 456                              (1UL << SOCK_WIFI_STATUS) | \
 457                              (1UL << SOCK_NOFCS) | \
 458                              (1UL << SOCK_FILTER_LOCKED) | \
 459                              (1UL << SOCK_TSTAMP_NEW))
 460
 461 /* if set, use value set by setsockopt() - else use IPv4 or SMC sysctl value */
 462 static void smc_adjust_sock_bufsizes(struct sock *nsk, struct sock *osk,
 463                                      unsigned long mask)
 464 {
 465         struct net *nnet = sock_net(nsk);
 466
 467         nsk->sk_userlocks = osk->sk_userlocks;
 468         if (osk->sk_userlocks & SOCK_SNDBUF_LOCK) {
 469                 nsk->sk_sndbuf = osk->sk_sndbuf;
 470         } else {
 471                 if (mask == SK_FLAGS_SMC_TO_CLC)
 472                         WRITE_ONCE(nsk->sk_sndbuf,
 473                                    READ_ONCE(nnet->ipv4.sysctl_tcp_wmem[1]));
 474                 else
 475                         WRITE_ONCE(nsk->sk_sndbuf,
 476                                    2 * READ_ONCE(nnet->smc.sysctl_wmem));
 477         }
 478         if (osk->sk_userlocks & SOCK_RCVBUF_LOCK) {
 479                 nsk->sk_rcvbuf = osk->sk_rcvbuf;
 480         } else {
 481                 if (mask == SK_FLAGS_SMC_TO_CLC)
 482                         WRITE_ONCE(nsk->sk_rcvbuf,
 483                                    READ_ONCE(nnet->ipv4.sysctl_tcp_rmem[1]));
 484                 else
 485                         WRITE_ONCE(nsk->sk_rcvbuf,
 486                                    2 * READ_ONCE(nnet->smc.sysctl_rmem));
 487         }
 488 }
 489
 490 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 491                                    unsigned long mask)
 492 {
 493         /* options we don't get control via setsockopt for */
 494         nsk->sk_type = osk->sk_type;
 495         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 496         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 497         nsk->sk_mark = READ_ONCE(osk->sk_mark);
 498         nsk->sk_priority = osk->sk_priority;
 499         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 500         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 501         nsk->sk_err = osk->sk_err;
 502
 503         nsk->sk_flags &= ~mask;
 504         nsk->sk_flags |= osk->sk_flags & mask;
 505
 506         smc_adjust_sock_bufsizes(nsk, osk, mask);
 507 }
 508
 509 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 510 {
 511         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 512 }
 513
 514 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 515                              (1UL << SOCK_KEEPOPEN) | \
 516                              (1UL << SOCK_LINGER) | \
 517                              (1UL << SOCK_DBG))
 518 /* copy only settings and flags relevant for smc from clc to smc socket */
 519 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 520 {
 521         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 522 }
 523
 524 /* register the new vzalloced sndbuf on all links */
 525 static int smcr_lgr_reg_sndbufs(struct smc_link *link,
 526                                 struct smc_buf_desc *snd_desc)
 527 {
 528         struct smc_link_group *lgr = link->lgr;
 529         int i, rc = 0;
 530
 531         if (!snd_desc->is_vm)
 532                 return -EINVAL;
 533
 534         /* protect against parallel smcr_link_reg_buf() */
 535         mutex_lock(&lgr->llc_conf_mutex);
 536         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 537                 if (!smc_link_active(&lgr->lnk[i]))
 538                         continue;
 539                 rc = smcr_link_reg_buf(&lgr->lnk[i], snd_desc);
 540                 if (rc)
 541                         break;
 542         }
 543         mutex_unlock(&lgr->llc_conf_mutex);
 544         return rc;
 545 }
 546
 547 /* register the new rmb on all links */
 548 static int smcr_lgr_reg_rmbs(struct smc_link *link,
 549                              struct smc_buf_desc *rmb_desc)
 550 {
 551         struct smc_link_group *lgr = link->lgr;
 552         int i, rc = 0;
 553
 554         rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
 555         if (rc)
 556                 return rc;
 557         /* protect against parallel smc_llc_cli_rkey_exchange() and
 558          * parallel smcr_link_reg_buf()
 559          */
 560         mutex_lock(&lgr->llc_conf_mutex);
 561         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 562                 if (!smc_link_active(&lgr->lnk[i]))
 563                         continue;
 564                 rc = smcr_link_reg_buf(&lgr->lnk[i], rmb_desc);
 565                 if (rc)
 566                         goto out;
 567         }
 568
 569         /* exchange confirm_rkey msg with peer */
 570         rc = smc_llc_do_confirm_rkey(link, rmb_desc);
 571         if (rc) {
 572                 rc = -EFAULT;
 573                 goto out;
 574         }
 575         rmb_desc->is_conf_rkey = true;
 576 out:
 577         mutex_unlock(&lgr->llc_conf_mutex);
 578         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
 579         return rc;
 580 }
 581
 582 static int smcr_clnt_conf_first_link(struct smc_sock *smc)
 583 {
 584         struct smc_link *link = smc->conn.lnk;
 585         struct smc_llc_qentry *qentry;
 586         int rc;
 587
 588         /* Receive CONFIRM LINK request from server over RoCE fabric.
 589          * Increasing the client's timeout by twice as much as the server's
 590          * timeout by default can temporarily avoid decline messages of
 591          * both sides crossing or colliding
 592          */
 593         qentry = smc_llc_wait(link->lgr, NULL, 2 * SMC_LLC_WAIT_TIME,
 594                               SMC_LLC_CONFIRM_LINK);
 595         if (!qentry) {
 596                 struct smc_clc_msg_decline dclc;
 597
 598                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 599                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 600                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
 601         }
 602         smc_llc_save_peer_uid(qentry);
 603         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
 604         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
 605         if (rc)
 606                 return SMC_CLC_DECL_RMBE_EC;
 607
 608         rc = smc_ib_modify_qp_rts(link);
 609         if (rc)
 610                 return SMC_CLC_DECL_ERR_RDYLNK;
 611
 612         smc_wr_remember_qp_attr(link);
 613
 614         /* reg the sndbuf if it was vzalloced */
 615         if (smc->conn.sndbuf_desc->is_vm) {
 616                 if (smcr_link_reg_buf(link, smc->conn.sndbuf_desc))
 617                         return SMC_CLC_DECL_ERR_REGBUF;
 618         }
 619
 620         /* reg the rmb */
 621         if (smcr_link_reg_buf(link, smc->conn.rmb_desc))
 622                 return SMC_CLC_DECL_ERR_REGBUF;
 623
 624         /* confirm_rkey is implicit on 1st contact */
 625         smc->conn.rmb_desc->is_conf_rkey = true;
 626
 627         /* send CONFIRM LINK response over RoCE fabric */
 628         rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
 629         if (rc < 0)
 630                 return SMC_CLC_DECL_TIMEOUT_CL;
 631
 632         smc_llc_link_active(link);
 633         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
 634
 635         /* optional 2nd link, receive ADD LINK request from server */
 636         qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
 637                               SMC_LLC_ADD_LINK);
 638         if (!qentry) {
 639                 struct smc_clc_msg_decline dclc;
 640
 641                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 642                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 643                 if (rc == -EAGAIN)
 644                         rc = 0; /* no DECLINE received, go with one link */
 645                 return rc;
 646         }
 647         smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
 648         smc_llc_cli_add_link(link, qentry);
 649         return 0;
 650 }
 651
 652 static bool smc_isascii(char *hostname)
 653 {
 654         int i;
 655
 656         for (i = 0; i < SMC_MAX_HOSTNAME_LEN; i++)
 657                 if (!isascii(hostname[i]))
 658                         return false;
 659         return true;
 660 }
 661
 662 static void smc_conn_save_peer_info_fce(struct smc_sock *smc,
 663                                         struct smc_clc_msg_accept_confirm *clc)
 664 {
 665         struct smc_clc_msg_accept_confirm_v2 *clc_v2 =
 666                 (struct smc_clc_msg_accept_confirm_v2 *)clc;
 667         struct smc_clc_first_contact_ext *fce;
 668         int clc_v2_len;
 669
 670         if (clc->hdr.version == SMC_V1 ||
 671             !(clc->hdr.typev2 & SMC_FIRST_CONTACT_MASK))
 672                 return;
 673
 674         if (smc->conn.lgr->is_smcd) {
 675                 memcpy(smc->conn.lgr->negotiated_eid, clc_v2->d1.eid,
 676                        SMC_MAX_EID_LEN);
 677                 clc_v2_len = offsetofend(struct smc_clc_msg_accept_confirm_v2,
 678                                          d1);
 679         } else {
 680                 memcpy(smc->conn.lgr->negotiated_eid, clc_v2->r1.eid,
 681                        SMC_MAX_EID_LEN);
 682                 clc_v2_len = offsetofend(struct smc_clc_msg_accept_confirm_v2,
 683                                          r1);
 684         }
 685         fce = (struct smc_clc_first_contact_ext *)(((u8 *)clc_v2) + clc_v2_len);
 686         smc->conn.lgr->peer_os = fce->os_type;
 687         smc->conn.lgr->peer_smc_release = fce->release;
 688         if (smc_isascii(fce->hostname))
 689                 memcpy(smc->conn.lgr->peer_hostname, fce->hostname,
 690                        SMC_MAX_HOSTNAME_LEN);
 691 }
 692
 693 static void smcr_conn_save_peer_info(struct smc_sock *smc,
 694                                      struct smc_clc_msg_accept_confirm *clc)
 695 {
 696         int bufsize = smc_uncompress_bufsize(clc->r0.rmbe_size);
 697
 698         smc->conn.peer_rmbe_idx = clc->r0.rmbe_idx;
 699         smc->conn.local_tx_ctrl.token = ntohl(clc->r0.rmbe_alert_token);
 700         smc->conn.peer_rmbe_size = bufsize;
 701         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 702         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
 703 }
 704
 705 static void smcd_conn_save_peer_info(struct smc_sock *smc,
 706                                      struct smc_clc_msg_accept_confirm *clc)
 707 {
 708         int bufsize = smc_uncompress_bufsize(clc->d0.dmbe_size);
 709
 710         smc->conn.peer_rmbe_idx = clc->d0.dmbe_idx;
 711         smc->conn.peer_token = clc->d0.token;
 712         /* msg header takes up space in the buffer */
 713         smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
 714         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 715         smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
 716 }
 717
 718 static void smc_conn_save_peer_info(struct smc_sock *smc,
 719                                     struct smc_clc_msg_accept_confirm *clc)
 720 {
 721         if (smc->conn.lgr->is_smcd)
 722                 smcd_conn_save_peer_info(smc, clc);
 723         else
 724                 smcr_conn_save_peer_info(smc, clc);
 725         smc_conn_save_peer_info_fce(smc, clc);
 726 }
 727
 728 static void smc_link_save_peer_info(struct smc_link *link,
 729                                     struct smc_clc_msg_accept_confirm *clc,
 730                                     struct smc_init_info *ini)
 731 {
 732         link->peer_qpn = ntoh24(clc->r0.qpn);
 733         memcpy(link->peer_gid, ini->peer_gid, SMC_GID_SIZE);
 734         memcpy(link->peer_mac, ini->peer_mac, sizeof(link->peer_mac));
 735         link->peer_psn = ntoh24(clc->r0.psn);
 736         link->peer_mtu = clc->r0.qp_mtu;
 737 }
 738
 739 static void smc_stat_inc_fback_rsn_cnt(struct smc_sock *smc,
 740                                        struct smc_stats_fback *fback_arr)
 741 {
 742         int cnt;
 743
 744         for (cnt = 0; cnt < SMC_MAX_FBACK_RSN_CNT; cnt++) {
 745                 if (fback_arr[cnt].fback_code == smc->fallback_rsn) {
 746                         fback_arr[cnt].count++;
 747                         break;
 748                 }
 749                 if (!fback_arr[cnt].fback_code) {
 750                         fback_arr[cnt].fback_code = smc->fallback_rsn;
 751                         fback_arr[cnt].count++;
 752                         break;
 753                 }
 754         }
 755 }
 756
 757 static void smc_stat_fallback(struct smc_sock *smc)
 758 {
 759         struct net *net = sock_net(&smc->sk);
 760
 761         mutex_lock(&net->smc.mutex_fback_rsn);
 762         if (smc->listen_smc) {
 763                 smc_stat_inc_fback_rsn_cnt(smc, net->smc.fback_rsn->srv);
 764                 net->smc.fback_rsn->srv_fback_cnt++;
 765         } else {
 766                 smc_stat_inc_fback_rsn_cnt(smc, net->smc.fback_rsn->clnt);
 767                 net->smc.fback_rsn->clnt_fback_cnt++;
 768         }
 769         mutex_unlock(&net->smc.mutex_fback_rsn);
 770 }
 771
 772 /* must be called under rcu read lock */
 773 static void smc_fback_wakeup_waitqueue(struct smc_sock *smc, void *key)
 774 {
 775         struct socket_wq *wq;
 776         __poll_t flags;
 777
 778         wq = rcu_dereference(smc->sk.sk_wq);
 779         if (!skwq_has_sleeper(wq))
 780                 return;
 781
 782         /* wake up smc sk->sk_wq */
 783         if (!key) {
 784                 /* sk_state_change */
 785                 wake_up_interruptible_all(&wq->wait);
 786         } else {
 787                 flags = key_to_poll(key);
 788                 if (flags & (EPOLLIN | EPOLLOUT))
 789                         /* sk_data_ready or sk_write_space */
 790                         wake_up_interruptible_sync_poll(&wq->wait, flags);
 791                 else if (flags & EPOLLERR)
 792                         /* sk_error_report */
 793                         wake_up_interruptible_poll(&wq->wait, flags);
 794         }
 795 }
 796
 797 static int smc_fback_mark_woken(wait_queue_entry_t *wait,
 798                                 unsigned int mode, int sync, void *key)
 799 {
 800         struct smc_mark_woken *mark =
 801                 container_of(wait, struct smc_mark_woken, wait_entry);
 802
 803         mark->woken = true;
 804         mark->key = key;
 805         return 0;
 806 }
 807
 808 static void smc_fback_forward_wakeup(struct smc_sock *smc, struct sock *clcsk,
 809                                      void (*clcsock_callback)(struct sock *sk))
 810 {
 811         struct smc_mark_woken mark = { .woken = false };
 812         struct socket_wq *wq;
 813
 814         init_waitqueue_func_entry(&mark.wait_entry,
 815                                   smc_fback_mark_woken);
 816         rcu_read_lock();
 817         wq = rcu_dereference(clcsk->sk_wq);
 818         if (!wq)
 819                 goto out;
 820         add_wait_queue(sk_sleep(clcsk), &mark.wait_entry);
 821         clcsock_callback(clcsk);
 822         remove_wait_queue(sk_sleep(clcsk), &mark.wait_entry);
 823
 824         if (mark.woken)
 825                 smc_fback_wakeup_waitqueue(smc, mark.key);
 826 out:
 827         rcu_read_unlock();
 828 }
 829
 830 static void smc_fback_state_change(struct sock *clcsk)
 831 {
 832         struct smc_sock *smc;
 833
 834         read_lock_bh(&clcsk->sk_callback_lock);
 835         smc = smc_clcsock_user_data(clcsk);
 836         if (smc)
 837                 smc_fback_forward_wakeup(smc, clcsk,
 838                                          smc->clcsk_state_change);
 839         read_unlock_bh(&clcsk->sk_callback_lock);
 840 }
 841
 842 static void smc_fback_data_ready(struct sock *clcsk)
 843 {
 844         struct smc_sock *smc;
 845
 846         read_lock_bh(&clcsk->sk_callback_lock);
 847         smc = smc_clcsock_user_data(clcsk);
 848         if (smc)
 849                 smc_fback_forward_wakeup(smc, clcsk,
 850                                          smc->clcsk_data_ready);
 851         read_unlock_bh(&clcsk->sk_callback_lock);
 852 }
 853
 854 static void smc_fback_write_space(struct sock *clcsk)
 855 {
 856         struct smc_sock *smc;
 857
 858         read_lock_bh(&clcsk->sk_callback_lock);
 859         smc = smc_clcsock_user_data(clcsk);
 860         if (smc)
 861                 smc_fback_forward_wakeup(smc, clcsk,
 862                                          smc->clcsk_write_space);
 863         read_unlock_bh(&clcsk->sk_callback_lock);
 864 }
 865
 866 static void smc_fback_error_report(struct sock *clcsk)
 867 {
 868         struct smc_sock *smc;
 869
 870         read_lock_bh(&clcsk->sk_callback_lock);
 871         smc = smc_clcsock_user_data(clcsk);
 872         if (smc)
 873                 smc_fback_forward_wakeup(smc, clcsk,
 874                                          smc->clcsk_error_report);
 875         read_unlock_bh(&clcsk->sk_callback_lock);
 876 }
 877
 878 static void smc_fback_replace_callbacks(struct smc_sock *smc)
 879 {
 880         struct sock *clcsk = smc->clcsock->sk;
 881
 882         write_lock_bh(&clcsk->sk_callback_lock);
 883         clcsk->sk_user_data = (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
 884
 885         smc_clcsock_replace_cb(&clcsk->sk_state_change, smc_fback_state_change,
 886                                &smc->clcsk_state_change);
 887         smc_clcsock_replace_cb(&clcsk->sk_data_ready, smc_fback_data_ready,
 888                                &smc->clcsk_data_ready);
 889         smc_clcsock_replace_cb(&clcsk->sk_write_space, smc_fback_write_space,
 890                                &smc->clcsk_write_space);
 891         smc_clcsock_replace_cb(&clcsk->sk_error_report, smc_fback_error_report,
 892                                &smc->clcsk_error_report);
 893
 894         write_unlock_bh(&clcsk->sk_callback_lock);
 895 }
 896
 897 static int smc_switch_to_fallback(struct smc_sock *smc, int reason_code)
 898 {
 899         int rc = 0;
 900
 901         mutex_lock(&smc->clcsock_release_lock);
 902         if (!smc->clcsock) {
 903                 rc = -EBADF;
 904                 goto out;
 905         }
 906
 907         smc->use_fallback = true;
 908         smc->fallback_rsn = reason_code;
 909         smc_stat_fallback(smc);
 910         trace_smc_switch_to_fallback(smc, reason_code);
 911         if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
 912                 smc->clcsock->file = smc->sk.sk_socket->file;
 913                 smc->clcsock->file->private_data = smc->clcsock;
 914                 smc->clcsock->wq.fasync_list =
 915                         smc->sk.sk_socket->wq.fasync_list;
 916
 917                 /* There might be some wait entries remaining
 918                  * in smc sk->sk_wq and they should be woken up
 919                  * as clcsock's wait queue is woken up.
 920                  */
 921                 smc_fback_replace_callbacks(smc);
 922         }
 923 out:
 924         mutex_unlock(&smc->clcsock_release_lock);
 925         return rc;
 926 }
 927
 928 /* fall back during connect */
 929 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
 930 {
 931         struct net *net = sock_net(&smc->sk);
 932         int rc = 0;
 933
 934         rc = smc_switch_to_fallback(smc, reason_code);
 935         if (rc) { /* fallback fails */
 936                 this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
 937                 if (smc->sk.sk_state == SMC_INIT)
 938                         sock_put(&smc->sk); /* passive closing */
 939                 return rc;
 940         }
 941         smc_copy_sock_settings_to_clc(smc);
 942         smc->connect_nonblock = 0;
 943         if (smc->sk.sk_state == SMC_INIT)
 944                 smc->sk.sk_state = SMC_ACTIVE;
 945         return 0;
 946 }
 947
 948 /* decline and fall back during connect */
 949 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code,
 950                                         u8 version)
 951 {
 952         struct net *net = sock_net(&smc->sk);
 953         int rc;
 954
 955         if (reason_code < 0) { /* error, fallback is not possible */
 956                 this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
 957                 if (smc->sk.sk_state == SMC_INIT)
 958                         sock_put(&smc->sk); /* passive closing */
 959                 return reason_code;
 960         }
 961         if (reason_code != SMC_CLC_DECL_PEERDECL) {
 962                 rc = smc_clc_send_decline(smc, reason_code, version);
 963                 if (rc < 0) {
 964                         this_cpu_inc(net->smc.smc_stats->clnt_hshake_err_cnt);
 965                         if (smc->sk.sk_state == SMC_INIT)
 966                                 sock_put(&smc->sk); /* passive closing */
 967                         return rc;
 968                 }
 969         }
 970         return smc_connect_fallback(smc, reason_code);
 971 }
 972
 973 static void smc_conn_abort(struct smc_sock *smc, int local_first)
 974 {
 975         struct smc_connection *conn = &smc->conn;
 976         struct smc_link_group *lgr = conn->lgr;
 977         bool lgr_valid = false;
 978
 979         if (smc_conn_lgr_valid(conn))
 980                 lgr_valid = true;
 981
 982         smc_conn_free(conn);
 983         if (local_first && lgr_valid)
 984                 smc_lgr_cleanup_early(lgr);
 985 }
 986
 987 /* check if there is a rdma device available for this connection. */
 988 /* called for connect and listen */
 989 static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
 990 {
 991         /* PNET table look up: search active ib_device and port
 992          * within same PNETID that also contains the ethernet device
 993          * used for the internal TCP socket
 994          */
 995         smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
 996         if (!ini->check_smcrv2 && !ini->ib_dev)
 997                 return SMC_CLC_DECL_NOSMCRDEV;
 998         if (ini->check_smcrv2 && !ini->smcrv2.ib_dev_v2)
 999                 return SMC_CLC_DECL_NOSMCRDEV;
1000         return 0;
1001 }
1002
1003 /* check if there is an ISM device available for this connection. */
1004 /* called for connect and listen */
1005 static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
1006 {
1007         /* Find ISM device with same PNETID as connecting interface  */
1008         smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
1009         if (!ini->ism_dev[0])
1010                 return SMC_CLC_DECL_NOSMCDDEV;
1011         else
1012                 ini->ism_chid[0] = smc_ism_get_chid(ini->ism_dev[0]);
1013         return 0;
1014 }
1015
1016 /* is chid unique for the ism devices that are already determined? */
1017 static bool smc_find_ism_v2_is_unique_chid(u16 chid, struct smc_init_info *ini,
1018                                            int cnt)
1019 {
1020         int i = (!ini->ism_dev[0]) ? 1 : 0;
1021
1022         for (; i < cnt; i++)
1023                 if (ini->ism_chid[i] == chid)
1024                         return false;
1025         return true;
1026 }
1027
1028 /* determine possible V2 ISM devices (either without PNETID or with PNETID plus
1029  * PNETID matching net_device)
1030  */
1031 static int smc_find_ism_v2_device_clnt(struct smc_sock *smc,
1032                                        struct smc_init_info *ini)
1033 {
1034         int rc = SMC_CLC_DECL_NOSMCDDEV;
1035         struct smcd_dev *smcd;
1036         int i = 1;
1037         u16 chid;
1038
1039         if (smcd_indicated(ini->smc_type_v1))
1040                 rc = 0;         /* already initialized for V1 */
1041         mutex_lock(&smcd_dev_list.mutex);
1042         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1043                 if (smcd->going_away || smcd == ini->ism_dev[0])
1044                         continue;
1045                 chid = smc_ism_get_chid(smcd);
1046                 if (!smc_find_ism_v2_is_unique_chid(chid, ini, i))
1047                         continue;
1048                 if (!smc_pnet_is_pnetid_set(smcd->pnetid) ||
1049                     smc_pnet_is_ndev_pnetid(sock_net(&smc->sk), smcd->pnetid)) {
1050                         ini->ism_dev[i] = smcd;
1051                         ini->ism_chid[i] = chid;
1052                         ini->is_smcd = true;
1053                         rc = 0;
1054                         i++;
1055                         if (i > SMC_MAX_ISM_DEVS)
1056                                 break;
1057                 }
1058         }
1059         mutex_unlock(&smcd_dev_list.mutex);
1060         ini->ism_offered_cnt = i - 1;
1061         if (!ini->ism_dev[0] && !ini->ism_dev[1])
1062                 ini->smcd_version = 0;
1063
1064         return rc;
1065 }
1066
1067 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
1068 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
1069                                       struct smc_init_info *ini)
1070 {
1071         if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev[0], ini->vlan_id))
1072                 return SMC_CLC_DECL_ISMVLANERR;
1073         return 0;
1074 }
1075
1076 static int smc_find_proposal_devices(struct smc_sock *smc,
1077                                      struct smc_init_info *ini)
1078 {
1079         int rc = 0;
1080
1081         /* check if there is an ism device available */
1082         if (!(ini->smcd_version & SMC_V1) ||
1083             smc_find_ism_device(smc, ini) ||
1084             smc_connect_ism_vlan_setup(smc, ini))
1085                 ini->smcd_version &= ~SMC_V1;
1086         /* else ISM V1 is supported for this connection */
1087
1088         /* check if there is an rdma device available */
1089         if (!(ini->smcr_version & SMC_V1) ||
1090             smc_find_rdma_device(smc, ini))
1091                 ini->smcr_version &= ~SMC_V1;
1092         /* else RDMA is supported for this connection */
1093
1094         ini->smc_type_v1 = smc_indicated_type(ini->smcd_version & SMC_V1,
1095                                               ini->smcr_version & SMC_V1);
1096
1097         /* check if there is an ism v2 device available */
1098         if (!(ini->smcd_version & SMC_V2) ||
1099             !smc_ism_is_v2_capable() ||
1100             smc_find_ism_v2_device_clnt(smc, ini))
1101                 ini->smcd_version &= ~SMC_V2;
1102
1103         /* check if there is an rdma v2 device available */
1104         ini->check_smcrv2 = true;
1105         ini->smcrv2.saddr = smc->clcsock->sk->sk_rcv_saddr;
1106         if (!(ini->smcr_version & SMC_V2) ||
1107             smc->clcsock->sk->sk_family != AF_INET ||
1108             !smc_clc_ueid_count() ||
1109             smc_find_rdma_device(smc, ini))
1110                 ini->smcr_version &= ~SMC_V2;
1111         ini->check_smcrv2 = false;
1112
1113         ini->smc_type_v2 = smc_indicated_type(ini->smcd_version & SMC_V2,
1114                                               ini->smcr_version & SMC_V2);
1115
1116         /* if neither ISM nor RDMA are supported, fallback */
1117         if (ini->smc_type_v1 == SMC_TYPE_N && ini->smc_type_v2 == SMC_TYPE_N)
1118                 rc = SMC_CLC_DECL_NOSMCDEV;
1119
1120         return rc;
1121 }
1122
1123 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
1124  * used, the VLAN ID will be registered again during the connection setup.
1125  */
1126 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc,
1127                                         struct smc_init_info *ini)
1128 {
1129         if (!smcd_indicated(ini->smc_type_v1))
1130                 return 0;
1131         if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev[0], ini->vlan_id))
1132                 return SMC_CLC_DECL_CNFERR;
1133         return 0;
1134 }
1135
1136 #define SMC_CLC_MAX_ACCEPT_LEN \
1137         (sizeof(struct smc_clc_msg_accept_confirm_v2) + \
1138          sizeof(struct smc_clc_first_contact_ext) + \
1139          sizeof(struct smc_clc_msg_trail))
1140
1141 /* CLC handshake during connect */
1142 static int smc_connect_clc(struct smc_sock *smc,
1143                            struct smc_clc_msg_accept_confirm_v2 *aclc2,
1144                            struct smc_init_info *ini)
1145 {
1146         int rc = 0;
1147
1148         /* do inband token exchange */
1149         rc = smc_clc_send_proposal(smc, ini);
1150         if (rc)
1151                 return rc;
1152         /* receive SMC Accept CLC message */
1153         return smc_clc_wait_msg(smc, aclc2, SMC_CLC_MAX_ACCEPT_LEN,
1154                                 SMC_CLC_ACCEPT, CLC_WAIT_TIME);
1155 }
1156
1157 void smc_fill_gid_list(struct smc_link_group *lgr,
1158                        struct smc_gidlist *gidlist,
1159                        struct smc_ib_device *known_dev, u8 *known_gid)
1160 {
1161         struct smc_init_info *alt_ini = NULL;
1162
1163         memset(gidlist, 0, sizeof(*gidlist));
1164         memcpy(gidlist->list[gidlist->len++], known_gid, SMC_GID_SIZE);
1165
1166         alt_ini = kzalloc(sizeof(*alt_ini), GFP_KERNEL);
1167         if (!alt_ini)
1168                 goto out;
1169
1170         alt_ini->vlan_id = lgr->vlan_id;
1171         alt_ini->check_smcrv2 = true;
1172         alt_ini->smcrv2.saddr = lgr->saddr;
1173         smc_pnet_find_alt_roce(lgr, alt_ini, known_dev);
1174
1175         if (!alt_ini->smcrv2.ib_dev_v2)
1176                 goto out;
1177
1178         memcpy(gidlist->list[gidlist->len++], alt_ini->smcrv2.ib_gid_v2,
1179                SMC_GID_SIZE);
1180
1181 out:
1182         kfree(alt_ini);
1183 }
1184
1185 static int smc_connect_rdma_v2_prepare(struct smc_sock *smc,
1186                                        struct smc_clc_msg_accept_confirm *aclc,
1187                                        struct smc_init_info *ini)
1188 {
1189         struct smc_clc_msg_accept_confirm_v2 *clc_v2 =
1190                 (struct smc_clc_msg_accept_confirm_v2 *)aclc;
1191         struct smc_clc_first_contact_ext *fce =
1192                 (struct smc_clc_first_contact_ext *)
1193                         (((u8 *)clc_v2) + sizeof(*clc_v2));
1194         struct net *net = sock_net(&smc->sk);
1195
1196         if (!ini->first_contact_peer || aclc->hdr.version == SMC_V1)
1197                 return 0;
1198
1199         if (fce->v2_direct) {
1200                 memcpy(ini->smcrv2.nexthop_mac, &aclc->r0.lcl.mac, ETH_ALEN);
1201                 ini->smcrv2.uses_gateway = false;
1202         } else {
1203                 if (smc_ib_find_route(net, smc->clcsock->sk->sk_rcv_saddr,
1204                                       smc_ib_gid_to_ipv4(aclc->r0.lcl.gid),
1205                                       ini->smcrv2.nexthop_mac,
1206                                       &ini->smcrv2.uses_gateway))
1207                         return SMC_CLC_DECL_NOROUTE;
1208                 if (!ini->smcrv2.uses_gateway) {
1209                         /* mismatch: peer claims indirect, but its direct */
1210                         return SMC_CLC_DECL_NOINDIRECT;
1211                 }
1212         }
1213         return 0;
1214 }
1215
1216 /* setup for RDMA connection of client */
1217 static int smc_connect_rdma(struct smc_sock *smc,
1218                             struct smc_clc_msg_accept_confirm *aclc,
1219                             struct smc_init_info *ini)
1220 {
1221         int i, reason_code = 0;
1222         struct smc_link *link;
1223         u8 *eid = NULL;
1224
1225         ini->is_smcd = false;
1226         ini->ib_clcqpn = ntoh24(aclc->r0.qpn);
1227         ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
1228         memcpy(ini->peer_systemid, aclc->r0.lcl.id_for_peer, SMC_SYSTEMID_LEN);
1229         memcpy(ini->peer_gid, aclc->r0.lcl.gid, SMC_GID_SIZE);
1230         memcpy(ini->peer_mac, aclc->r0.lcl.mac, ETH_ALEN);
1231
1232         reason_code = smc_connect_rdma_v2_prepare(smc, aclc, ini);
1233         if (reason_code)
1234                 return reason_code;
1235
1236         mutex_lock(&smc_client_lgr_pending);
1237         reason_code = smc_conn_create(smc, ini);
1238         if (reason_code) {
1239                 mutex_unlock(&smc_client_lgr_pending);
1240                 return reason_code;
1241         }
1242
1243         smc_conn_save_peer_info(smc, aclc);
1244
1245         if (ini->first_contact_local) {
1246                 link = smc->conn.lnk;
1247         } else {
1248                 /* set link that was assigned by server */
1249                 link = NULL;
1250                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
1251                         struct smc_link *l = &smc->conn.lgr->lnk[i];
1252
1253                         if (l->peer_qpn == ntoh24(aclc->r0.qpn) &&
1254                             !memcmp(l->peer_gid, &aclc->r0.lcl.gid,
1255                                     SMC_GID_SIZE) &&
1256                             (aclc->hdr.version > SMC_V1 ||
1257                              !memcmp(l->peer_mac, &aclc->r0.lcl.mac,
1258                                      sizeof(l->peer_mac)))) {
1259                                 link = l;
1260                                 break;
1261                         }
1262                 }
1263                 if (!link) {
1264                         reason_code = SMC_CLC_DECL_NOSRVLINK;
1265                         goto connect_abort;
1266                 }
1267                 smc_switch_link_and_count(&smc->conn, link);
1268         }
1269
1270         /* create send buffer and rmb */
1271         if (smc_buf_create(smc, false)) {
1272                 reason_code = SMC_CLC_DECL_MEM;
1273                 goto connect_abort;
1274         }
1275
1276         if (ini->first_contact_local)
1277                 smc_link_save_peer_info(link, aclc, ini);
1278
1279         if (smc_rmb_rtoken_handling(&smc->conn, link, aclc)) {
1280                 reason_code = SMC_CLC_DECL_ERR_RTOK;
1281                 goto connect_abort;
1282         }
1283
1284         smc_close_init(smc);
1285         smc_rx_init(smc);
1286
1287         if (ini->first_contact_local) {
1288                 if (smc_ib_ready_link(link)) {
1289                         reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1290                         goto connect_abort;
1291                 }
1292         } else {
1293                 /* reg sendbufs if they were vzalloced */
1294                 if (smc->conn.sndbuf_desc->is_vm) {
1295                         if (smcr_lgr_reg_sndbufs(link, smc->conn.sndbuf_desc)) {
1296                                 reason_code = SMC_CLC_DECL_ERR_REGBUF;
1297                                 goto connect_abort;
1298                         }
1299                 }
1300                 if (smcr_lgr_reg_rmbs(link, smc->conn.rmb_desc)) {
1301                         reason_code = SMC_CLC_DECL_ERR_REGBUF;
1302                         goto connect_abort;
1303                 }
1304         }
1305
1306         if (aclc->hdr.version > SMC_V1) {
1307                 struct smc_clc_msg_accept_confirm_v2 *clc_v2 =
1308                         (struct smc_clc_msg_accept_confirm_v2 *)aclc;
1309
1310                 eid = clc_v2->r1.eid;
1311                 if (ini->first_contact_local)
1312                         smc_fill_gid_list(link->lgr, &ini->smcrv2.gidlist,
1313                                           link->smcibdev, link->gid);
1314         }
1315
1316         reason_code = smc_clc_send_confirm(smc, ini->first_contact_local,
1317                                            aclc->hdr.version, eid, ini);
1318         if (reason_code)
1319                 goto connect_abort;
1320
1321         smc_tx_init(smc);
1322
1323         if (ini->first_contact_local) {
1324                 /* QP confirmation over RoCE fabric */
1325                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1326                 reason_code = smcr_clnt_conf_first_link(smc);
1327                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1328                 if (reason_code)
1329                         goto connect_abort;
1330         }
1331         mutex_unlock(&smc_client_lgr_pending);
1332
1333         smc_copy_sock_settings_to_clc(smc);
1334         smc->connect_nonblock = 0;
1335         if (smc->sk.sk_state == SMC_INIT)
1336                 smc->sk.sk_state = SMC_ACTIVE;
1337
1338         return 0;
1339 connect_abort:
1340         smc_conn_abort(smc, ini->first_contact_local);
1341         mutex_unlock(&smc_client_lgr_pending);
1342         smc->connect_nonblock = 0;
1343
1344         return reason_code;
1345 }
1346
1347 /* The server has chosen one of the proposed ISM devices for the communication.
1348  * Determine from the CHID of the received CLC ACCEPT the ISM device chosen.
1349  */
1350 static int
1351 smc_v2_determine_accepted_chid(struct smc_clc_msg_accept_confirm_v2 *aclc,
1352                                struct smc_init_info *ini)
1353 {
1354         int i;
1355
1356         for (i = 0; i < ini->ism_offered_cnt + 1; i++) {
1357                 if (ini->ism_chid[i] == ntohs(aclc->d1.chid)) {
1358                         ini->ism_selected = i;
1359                         return 0;
1360                 }
1361         }
1362
1363         return -EPROTO;
1364 }
1365
1366 /* setup for ISM connection of client */
1367 static int smc_connect_ism(struct smc_sock *smc,
1368                            struct smc_clc_msg_accept_confirm *aclc,
1369                            struct smc_init_info *ini)
1370 {
1371         u8 *eid = NULL;
1372         int rc = 0;
1373
1374         ini->is_smcd = true;
1375         ini->first_contact_peer = aclc->hdr.typev2 & SMC_FIRST_CONTACT_MASK;
1376
1377         if (aclc->hdr.version == SMC_V2) {
1378                 struct smc_clc_msg_accept_confirm_v2 *aclc_v2 =
1379                         (struct smc_clc_msg_accept_confirm_v2 *)aclc;
1380
1381                 rc = smc_v2_determine_accepted_chid(aclc_v2, ini);
1382                 if (rc)
1383                         return rc;
1384         }
1385         ini->ism_peer_gid[ini->ism_selected] = aclc->d0.gid;
1386
1387         /* there is only one lgr role for SMC-D; use server lock */
1388         mutex_lock(&smc_server_lgr_pending);
1389         rc = smc_conn_create(smc, ini);
1390         if (rc) {
1391                 mutex_unlock(&smc_server_lgr_pending);
1392                 return rc;
1393         }
1394
1395         /* Create send and receive buffers */
1396         rc = smc_buf_create(smc, true);
1397         if (rc) {
1398                 rc = (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB : SMC_CLC_DECL_MEM;
1399                 goto connect_abort;
1400         }
1401
1402         smc_conn_save_peer_info(smc, aclc);
1403         smc_close_init(smc);
1404         smc_rx_init(smc);
1405         smc_tx_init(smc);
1406
1407         if (aclc->hdr.version > SMC_V1) {
1408                 struct smc_clc_msg_accept_confirm_v2 *clc_v2 =
1409                         (struct smc_clc_msg_accept_confirm_v2 *)aclc;
1410
1411                 eid = clc_v2->d1.eid;
1412         }
1413
1414         rc = smc_clc_send_confirm(smc, ini->first_contact_local,
1415                                   aclc->hdr.version, eid, NULL);
1416         if (rc)
1417                 goto connect_abort;
1418         mutex_unlock(&smc_server_lgr_pending);
1419
1420         smc_copy_sock_settings_to_clc(smc);
1421         smc->connect_nonblock = 0;
1422         if (smc->sk.sk_state == SMC_INIT)
1423                 smc->sk.sk_state = SMC_ACTIVE;
1424
1425         return 0;
1426 connect_abort:
1427         smc_conn_abort(smc, ini->first_contact_local);
1428         mutex_unlock(&smc_server_lgr_pending);
1429         smc->connect_nonblock = 0;
1430
1431         return rc;
1432 }
1433
1434 /* check if received accept type and version matches a proposed one */
1435 static int smc_connect_check_aclc(struct smc_init_info *ini,
1436                                   struct smc_clc_msg_accept_confirm *aclc)
1437 {
1438         if (aclc->hdr.typev1 != SMC_TYPE_R &&
1439             aclc->hdr.typev1 != SMC_TYPE_D)
1440                 return SMC_CLC_DECL_MODEUNSUPP;
1441
1442         if (aclc->hdr.version >= SMC_V2) {
1443                 if ((aclc->hdr.typev1 == SMC_TYPE_R &&
1444                      !smcr_indicated(ini->smc_type_v2)) ||
1445                     (aclc->hdr.typev1 == SMC_TYPE_D &&
1446                      !smcd_indicated(ini->smc_type_v2)))
1447                         return SMC_CLC_DECL_MODEUNSUPP;
1448         } else {
1449                 if ((aclc->hdr.typev1 == SMC_TYPE_R &&
1450                      !smcr_indicated(ini->smc_type_v1)) ||
1451                     (aclc->hdr.typev1 == SMC_TYPE_D &&
1452                      !smcd_indicated(ini->smc_type_v1)))
1453                         return SMC_CLC_DECL_MODEUNSUPP;
1454         }
1455
1456         return 0;
1457 }
1458
1459 /* perform steps before actually connecting */
1460 static int __smc_connect(struct smc_sock *smc)
1461 {
1462         u8 version = smc_ism_is_v2_capable() ? SMC_V2 : SMC_V1;
1463         struct smc_clc_msg_accept_confirm_v2 *aclc2;
1464         struct smc_clc_msg_accept_confirm *aclc;
1465         struct smc_init_info *ini = NULL;
1466         u8 *buf = NULL;
1467         int rc = 0;
1468
1469         if (smc->use_fallback)
1470                 return smc_connect_fallback(smc, smc->fallback_rsn);
1471
1472         /* if peer has not signalled SMC-capability, fall back */
1473         if (!tcp_sk(smc->clcsock->sk)->syn_smc)
1474                 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
1475
1476         /* IPSec connections opt out of SMC optimizations */
1477         if (using_ipsec(smc))
1478                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC,
1479                                                     version);
1480
1481         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1482         if (!ini)
1483                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_MEM,
1484                                                     version);
1485
1486         ini->smcd_version = SMC_V1 | SMC_V2;
1487         ini->smcr_version = SMC_V1 | SMC_V2;
1488         ini->smc_type_v1 = SMC_TYPE_B;
1489         ini->smc_type_v2 = SMC_TYPE_B;
1490
1491         /* get vlan id from IP device */
1492         if (smc_vlan_by_tcpsk(smc->clcsock, ini)) {
1493                 ini->smcd_version &= ~SMC_V1;
1494                 ini->smcr_version = 0;
1495                 ini->smc_type_v1 = SMC_TYPE_N;
1496                 if (!ini->smcd_version) {
1497                         rc = SMC_CLC_DECL_GETVLANERR;
1498                         goto fallback;
1499                 }
1500         }
1501
1502         rc = smc_find_proposal_devices(smc, ini);
1503         if (rc)
1504                 goto fallback;
1505
1506         buf = kzalloc(SMC_CLC_MAX_ACCEPT_LEN, GFP_KERNEL);
1507         if (!buf) {
1508                 rc = SMC_CLC_DECL_MEM;
1509                 goto fallback;
1510         }
1511         aclc2 = (struct smc_clc_msg_accept_confirm_v2 *)buf;
1512         aclc = (struct smc_clc_msg_accept_confirm *)aclc2;
1513
1514         /* perform CLC handshake */
1515         rc = smc_connect_clc(smc, aclc2, ini);
1516         if (rc) {
1517                 /* -EAGAIN on timeout, see tcp_recvmsg() */
1518                 if (rc == -EAGAIN) {
1519                         rc = -ETIMEDOUT;
1520                         smc->sk.sk_err = ETIMEDOUT;
1521                 }
1522                 goto vlan_cleanup;
1523         }
1524
1525         /* check if smc modes and versions of CLC proposal and accept match */
1526         rc = smc_connect_check_aclc(ini, aclc);
1527         version = aclc->hdr.version == SMC_V1 ? SMC_V1 : SMC_V2;
1528         if (rc)
1529                 goto vlan_cleanup;
1530
1531         /* depending on previous steps, connect using rdma or ism */
1532         if (aclc->hdr.typev1 == SMC_TYPE_R) {
1533                 ini->smcr_version = version;
1534                 rc = smc_connect_rdma(smc, aclc, ini);
1535         } else if (aclc->hdr.typev1 == SMC_TYPE_D) {
1536                 ini->smcd_version = version;
1537                 rc = smc_connect_ism(smc, aclc, ini);
1538         }
1539         if (rc)
1540                 goto vlan_cleanup;
1541
1542         SMC_STAT_CLNT_SUCC_INC(sock_net(smc->clcsock->sk), aclc);
1543         smc_connect_ism_vlan_cleanup(smc, ini);
1544         kfree(buf);
1545         kfree(ini);
1546         return 0;
1547
1548 vlan_cleanup:
1549         smc_connect_ism_vlan_cleanup(smc, ini);
1550         kfree(buf);
1551 fallback:
1552         kfree(ini);
1553         return smc_connect_decline_fallback(smc, rc, version);
1554 }
1555
1556 static void smc_connect_work(struct work_struct *work)
1557 {
1558         struct smc_sock *smc = container_of(work, struct smc_sock,
1559                                             connect_work);
1560         long timeo = smc->sk.sk_sndtimeo;
1561         int rc = 0;
1562
1563         if (!timeo)
1564                 timeo = MAX_SCHEDULE_TIMEOUT;
1565         lock_sock(smc->clcsock->sk);
1566         if (smc->clcsock->sk->sk_err) {
1567                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
1568         } else if ((1 << smc->clcsock->sk->sk_state) &
1569                                         (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
1570                 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
1571                 if ((rc == -EPIPE) &&
1572                     ((1 << smc->clcsock->sk->sk_state) &
1573                                         (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
1574                         rc = 0;
1575         }
1576         release_sock(smc->clcsock->sk);
1577         lock_sock(&smc->sk);
1578         if (rc != 0 || smc->sk.sk_err) {
1579                 smc->sk.sk_state = SMC_CLOSED;
1580                 if (rc == -EPIPE || rc == -EAGAIN)
1581                         smc->sk.sk_err = EPIPE;
1582                 else if (rc == -ECONNREFUSED)
1583                         smc->sk.sk_err = ECONNREFUSED;
1584                 else if (signal_pending(current))
1585                         smc->sk.sk_err = -sock_intr_errno(timeo);
1586                 sock_put(&smc->sk); /* passive closing */
1587                 goto out;
1588         }
1589
1590         rc = __smc_connect(smc);
1591         if (rc < 0)
1592                 smc->sk.sk_err = -rc;
1593
1594 out:
1595         if (!sock_flag(&smc->sk, SOCK_DEAD)) {
1596                 if (smc->sk.sk_err) {
1597                         smc->sk.sk_state_change(&smc->sk);
1598                 } else { /* allow polling before and after fallback decision */
1599                         smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
1600                         smc->sk.sk_write_space(&smc->sk);
1601                 }
1602         }
1603         release_sock(&smc->sk);
1604 }
1605
1606 static int smc_connect(struct socket *sock, struct sockaddr *addr,
1607                        int alen, int flags)
1608 {
1609         struct sock *sk = sock->sk;
1610         struct smc_sock *smc;
1611         int rc = -EINVAL;
1612
1613         smc = smc_sk(sk);
1614
1615         /* separate smc parameter checking to be safe */
1616         if (alen < sizeof(addr->sa_family))
1617                 goto out_err;
1618         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
1619                 goto out_err;
1620
1621         lock_sock(sk);
1622         switch (sock->state) {
1623         default:
1624                 rc = -EINVAL;
1625                 goto out;
1626         case SS_CONNECTED:
1627                 rc = sk->sk_state == SMC_ACTIVE ? -EISCONN : -EINVAL;
1628                 goto out;
1629         case SS_CONNECTING:
1630                 if (sk->sk_state == SMC_ACTIVE)
1631                         goto connected;
1632                 break;
1633         case SS_UNCONNECTED:
1634                 sock->state = SS_CONNECTING;
1635                 break;
1636         }
1637
1638         switch (sk->sk_state) {
1639         default:
1640                 goto out;
1641         case SMC_CLOSED:
1642                 rc = sock_error(sk) ? : -ECONNABORTED;
1643                 sock->state = SS_UNCONNECTED;
1644                 goto out;
1645         case SMC_ACTIVE:
1646                 rc = -EISCONN;
1647                 goto out;
1648         case SMC_INIT:
1649                 break;
1650         }
1651
1652         smc_copy_sock_settings_to_clc(smc);
1653         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1654         if (smc->connect_nonblock) {
1655                 rc = -EALREADY;
1656                 goto out;
1657         }
1658         rc = kernel_connect(smc->clcsock, addr, alen, flags);
1659         if (rc && rc != -EINPROGRESS)
1660                 goto out;
1661
1662         if (smc->use_fallback) {
1663                 sock->state = rc ? SS_CONNECTING : SS_CONNECTED;
1664                 goto out;
1665         }
1666         sock_hold(&smc->sk); /* sock put in passive closing */
1667         if (flags & O_NONBLOCK) {
1668                 if (queue_work(smc_hs_wq, &smc->connect_work))
1669                         smc->connect_nonblock = 1;
1670                 rc = -EINPROGRESS;
1671                 goto out;
1672         } else {
1673                 rc = __smc_connect(smc);
1674                 if (rc < 0)
1675                         goto out;
1676         }
1677
1678 connected:
1679         rc = 0;
1680         sock->state = SS_CONNECTED;
1681 out:
1682         release_sock(sk);
1683 out_err:
1684         return rc;
1685 }
1686
1687 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
1688 {
1689         struct socket *new_clcsock = NULL;
1690         struct sock *lsk = &lsmc->sk;
1691         struct sock *new_sk;
1692         int rc = -EINVAL;
1693
1694         release_sock(lsk);
1695         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
1696         if (!new_sk) {
1697                 rc = -ENOMEM;
1698                 lsk->sk_err = ENOMEM;
1699                 *new_smc = NULL;
1700                 lock_sock(lsk);
1701                 goto out;
1702         }
1703         *new_smc = smc_sk(new_sk);
1704
1705         mutex_lock(&lsmc->clcsock_release_lock);
1706         if (lsmc->clcsock)
1707                 rc = kernel_accept(lsmc->clcsock, &new_clcsock, SOCK_NONBLOCK);
1708         mutex_unlock(&lsmc->clcsock_release_lock);
1709         lock_sock(lsk);
1710         if  (rc < 0 && rc != -EAGAIN)
1711                 lsk->sk_err = -rc;
1712         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
1713                 new_sk->sk_prot->unhash(new_sk);
1714                 if (new_clcsock)
1715                         sock_release(new_clcsock);
1716                 new_sk->sk_state = SMC_CLOSED;
1717                 smc_sock_set_flag(new_sk, SOCK_DEAD);
1718                 sock_put(new_sk); /* final */
1719                 *new_smc = NULL;
1720                 goto out;
1721         }
1722
1723         /* new clcsock has inherited the smc listen-specific sk_data_ready
1724          * function; switch it back to the original sk_data_ready function
1725          */
1726         new_clcsock->sk->sk_data_ready = lsmc->clcsk_data_ready;
1727
1728         /* if new clcsock has also inherited the fallback-specific callback
1729          * functions, switch them back to the original ones.
1730          */
1731         if (lsmc->use_fallback) {
1732                 if (lsmc->clcsk_state_change)
1733                         new_clcsock->sk->sk_state_change = lsmc->clcsk_state_change;
1734                 if (lsmc->clcsk_write_space)
1735                         new_clcsock->sk->sk_write_space = lsmc->clcsk_write_space;
1736                 if (lsmc->clcsk_error_report)
1737                         new_clcsock->sk->sk_error_report = lsmc->clcsk_error_report;
1738         }
1739
1740         (*new_smc)->clcsock = new_clcsock;
1741 out:
1742         return rc;
1743 }
1744
1745 /* add a just created sock to the accept queue of the listen sock as
1746  * candidate for a following socket accept call from user space
1747  */
1748 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
1749 {
1750         struct smc_sock *par = smc_sk(parent);
1751
1752         sock_hold(sk); /* sock_put in smc_accept_unlink () */
1753         spin_lock(&par->accept_q_lock);
1754         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
1755         spin_unlock(&par->accept_q_lock);
1756         sk_acceptq_added(parent);
1757 }
1758
1759 /* remove a socket from the accept queue of its parental listening socket */
1760 static void smc_accept_unlink(struct sock *sk)
1761 {
1762         struct smc_sock *par = smc_sk(sk)->listen_smc;
1763
1764         spin_lock(&par->accept_q_lock);
1765         list_del_init(&smc_sk(sk)->accept_q);
1766         spin_unlock(&par->accept_q_lock);
1767         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
1768         sock_put(sk); /* sock_hold in smc_accept_enqueue */
1769 }
1770
1771 /* remove a sock from the accept queue to bind it to a new socket created
1772  * for a socket accept call from user space
1773  */
1774 struct sock *smc_accept_dequeue(struct sock *parent,
1775                                 struct socket *new_sock)
1776 {
1777         struct smc_sock *isk, *n;
1778         struct sock *new_sk;
1779
1780         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1781                 new_sk = (struct sock *)isk;
1782
1783                 smc_accept_unlink(new_sk);
1784                 if (new_sk->sk_state == SMC_CLOSED) {
1785                         new_sk->sk_prot->unhash(new_sk);
1786                         if (isk->clcsock) {
1787                                 sock_release(isk->clcsock);
1788                                 isk->clcsock = NULL;
1789                         }
1790                         sock_put(new_sk); /* final */
1791                         continue;
1792                 }
1793                 if (new_sock) {
1794                         sock_graft(new_sk, new_sock);
1795                         new_sock->state = SS_CONNECTED;
1796                         if (isk->use_fallback) {
1797                                 smc_sk(new_sk)->clcsock->file = new_sock->file;
1798                                 isk->clcsock->file->private_data = isk->clcsock;
1799                         }
1800                 }
1801                 return new_sk;
1802         }
1803         return NULL;
1804 }
1805
1806 /* clean up for a created but never accepted sock */
1807 void smc_close_non_accepted(struct sock *sk)
1808 {
1809         struct smc_sock *smc = smc_sk(sk);
1810
1811         sock_hold(sk); /* sock_put below */
1812         lock_sock(sk);
1813         if (!sk->sk_lingertime)
1814                 /* wait for peer closing */
1815                 WRITE_ONCE(sk->sk_lingertime, SMC_MAX_STREAM_WAIT_TIMEOUT);
1816         __smc_release(smc);
1817         release_sock(sk);
1818         sock_put(sk); /* sock_hold above */
1819         sock_put(sk); /* final sock_put */
1820 }
1821
1822 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1823 {
1824         struct smc_link *link = smc->conn.lnk;
1825         struct smc_llc_qentry *qentry;
1826         int rc;
1827
1828         /* reg the sndbuf if it was vzalloced*/
1829         if (smc->conn.sndbuf_desc->is_vm) {
1830                 if (smcr_link_reg_buf(link, smc->conn.sndbuf_desc))
1831                         return SMC_CLC_DECL_ERR_REGBUF;
1832         }
1833
1834         /* reg the rmb */
1835         if (smcr_link_reg_buf(link, smc->conn.rmb_desc))
1836                 return SMC_CLC_DECL_ERR_REGBUF;
1837
1838         /* send CONFIRM LINK request to client over the RoCE fabric */
1839         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1840         if (rc < 0)
1841                 return SMC_CLC_DECL_TIMEOUT_CL;
1842
1843         /* receive CONFIRM LINK response from client over the RoCE fabric */
1844         qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1845                               SMC_LLC_CONFIRM_LINK);
1846         if (!qentry) {
1847                 struct smc_clc_msg_decline dclc;
1848
1849                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1850                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1851                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1852         }
1853         smc_llc_save_peer_uid(qentry);
1854         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1855         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1856         if (rc)
1857                 return SMC_CLC_DECL_RMBE_EC;
1858
1859         /* confirm_rkey is implicit on 1st contact */
1860         smc->conn.rmb_desc->is_conf_rkey = true;
1861
1862         smc_llc_link_active(link);
1863         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1864
1865         mutex_lock(&link->lgr->llc_conf_mutex);
1866         /* initial contact - try to establish second link */
1867         smc_llc_srv_add_link(link, NULL);
1868         mutex_unlock(&link->lgr->llc_conf_mutex);
1869         return 0;
1870 }
1871
1872 /* listen worker: finish */
1873 static void smc_listen_out(struct smc_sock *new_smc)
1874 {
1875         struct smc_sock *lsmc = new_smc->listen_smc;
1876         struct sock *newsmcsk = &new_smc->sk;
1877
1878         if (tcp_sk(new_smc->clcsock->sk)->syn_smc)
1879                 atomic_dec(&lsmc->queued_smc_hs);
1880
1881         if (lsmc->sk.sk_state == SMC_LISTEN) {
1882                 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1883                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1884                 release_sock(&lsmc->sk);
1885         } else { /* no longer listening */
1886                 smc_close_non_accepted(newsmcsk);
1887         }
1888
1889         /* Wake up accept */
1890         lsmc->sk.sk_data_ready(&lsmc->sk);
1891         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1892 }
1893
1894 /* listen worker: finish in state connected */
1895 static void smc_listen_out_connected(struct smc_sock *new_smc)
1896 {
1897         struct sock *newsmcsk = &new_smc->sk;
1898
1899         if (newsmcsk->sk_state == SMC_INIT)
1900                 newsmcsk->sk_state = SMC_ACTIVE;
1901
1902         smc_listen_out(new_smc);
1903 }
1904
1905 /* listen worker: finish in error state */
1906 static void smc_listen_out_err(struct smc_sock *new_smc)
1907 {
1908         struct sock *newsmcsk = &new_smc->sk;
1909         struct net *net = sock_net(newsmcsk);
1910
1911         this_cpu_inc(net->smc.smc_stats->srv_hshake_err_cnt);
1912         if (newsmcsk->sk_state == SMC_INIT)
1913                 sock_put(&new_smc->sk); /* passive closing */
1914         newsmcsk->sk_state = SMC_CLOSED;
1915
1916         smc_listen_out(new_smc);
1917 }
1918
1919 /* listen worker: decline and fall back if possible */
1920 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1921                                int local_first, u8 version)
1922 {
1923         /* RDMA setup failed, switch back to TCP */
1924         smc_conn_abort(new_smc, local_first);
1925         if (reason_code < 0 ||
1926             smc_switch_to_fallback(new_smc, reason_code)) {
1927                 /* error, no fallback possible */
1928                 smc_listen_out_err(new_smc);
1929                 return;
1930         }
1931         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1932                 if (smc_clc_send_decline(new_smc, reason_code, version) < 0) {
1933                         smc_listen_out_err(new_smc);
1934                         return;
1935                 }
1936         }
1937         smc_listen_out_connected(new_smc);
1938 }
1939
1940 /* listen worker: version checking */
1941 static int smc_listen_v2_check(struct smc_sock *new_smc,
1942                                struct smc_clc_msg_proposal *pclc,
1943                                struct smc_init_info *ini)
1944 {
1945         struct smc_clc_smcd_v2_extension *pclc_smcd_v2_ext;
1946         struct smc_clc_v2_extension *pclc_v2_ext;
1947         int rc = SMC_CLC_DECL_PEERNOSMC;
1948
1949         ini->smc_type_v1 = pclc->hdr.typev1;
1950         ini->smc_type_v2 = pclc->hdr.typev2;
1951         ini->smcd_version = smcd_indicated(ini->smc_type_v1) ? SMC_V1 : 0;
1952         ini->smcr_version = smcr_indicated(ini->smc_type_v1) ? SMC_V1 : 0;
1953         if (pclc->hdr.version > SMC_V1) {
1954                 if (smcd_indicated(ini->smc_type_v2))
1955                         ini->smcd_version |= SMC_V2;
1956                 if (smcr_indicated(ini->smc_type_v2))
1957                         ini->smcr_version |= SMC_V2;
1958         }
1959         if (!(ini->smcd_version & SMC_V2) && !(ini->smcr_version & SMC_V2)) {
1960                 rc = SMC_CLC_DECL_PEERNOSMC;
1961                 goto out;
1962         }
1963         pclc_v2_ext = smc_get_clc_v2_ext(pclc);
1964         if (!pclc_v2_ext) {
1965                 ini->smcd_version &= ~SMC_V2;
1966                 ini->smcr_version &= ~SMC_V2;
1967                 rc = SMC_CLC_DECL_NOV2EXT;
1968                 goto out;
1969         }
1970         pclc_smcd_v2_ext = smc_get_clc_smcd_v2_ext(pclc_v2_ext);
1971         if (ini->smcd_version & SMC_V2) {
1972                 if (!smc_ism_is_v2_capable()) {
1973                         ini->smcd_version &= ~SMC_V2;
1974                         rc = SMC_CLC_DECL_NOISM2SUPP;
1975                 } else if (!pclc_smcd_v2_ext) {
1976                         ini->smcd_version &= ~SMC_V2;
1977                         rc = SMC_CLC_DECL_NOV2DEXT;
1978                 } else if (!pclc_v2_ext->hdr.eid_cnt &&
1979                            !pclc_v2_ext->hdr.flag.seid) {
1980                         ini->smcd_version &= ~SMC_V2;
1981                         rc = SMC_CLC_DECL_NOUEID;
1982                 }
1983         }
1984         if (ini->smcr_version & SMC_V2) {
1985                 if (!pclc_v2_ext->hdr.eid_cnt) {
1986                         ini->smcr_version &= ~SMC_V2;
1987                         rc = SMC_CLC_DECL_NOUEID;
1988                 }
1989         }
1990
1991 out:
1992         if (!ini->smcd_version && !ini->smcr_version)
1993                 return rc;
1994
1995         return 0;
1996 }
1997
1998 /* listen worker: check prefixes */
1999 static int smc_listen_prfx_check(struct smc_sock *new_smc,
2000                                  struct smc_clc_msg_proposal *pclc)
2001 {
2002         struct smc_clc_msg_proposal_prefix *pclc_prfx;
2003         struct socket *newclcsock = new_smc->clcsock;
2004
2005         if (pclc->hdr.typev1 == SMC_TYPE_N)
2006                 return 0;
2007         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
2008         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
2009                 return SMC_CLC_DECL_DIFFPREFIX;
2010
2011         return 0;
2012 }
2013
2014 /* listen worker: initialize connection and buffers */
2015 static int smc_listen_rdma_init(struct smc_sock *new_smc,
2016                                 struct smc_init_info *ini)
2017 {
2018         int rc;
2019
2020         /* allocate connection / link group */
2021         rc = smc_conn_create(new_smc, ini);
2022         if (rc)
2023                 return rc;
2024
2025         /* create send buffer and rmb */
2026         if (smc_buf_create(new_smc, false)) {
2027                 smc_conn_abort(new_smc, ini->first_contact_local);
2028                 return SMC_CLC_DECL_MEM;
2029         }
2030
2031         return 0;
2032 }
2033
2034 /* listen worker: initialize connection and buffers for SMC-D */
2035 static int smc_listen_ism_init(struct smc_sock *new_smc,
2036                                struct smc_init_info *ini)
2037 {
2038         int rc;
2039
2040         rc = smc_conn_create(new_smc, ini);
2041         if (rc)
2042                 return rc;
2043
2044         /* Create send and receive buffers */
2045         rc = smc_buf_create(new_smc, true);
2046         if (rc) {
2047                 smc_conn_abort(new_smc, ini->first_contact_local);
2048                 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
2049                                          SMC_CLC_DECL_MEM;
2050         }
2051
2052         return 0;
2053 }
2054
2055 static bool smc_is_already_selected(struct smcd_dev *smcd,
2056                                     struct smc_init_info *ini,
2057                                     int matches)
2058 {
2059         int i;
2060
2061         for (i = 0; i < matches; i++)
2062                 if (smcd == ini->ism_dev[i])
2063                         return true;
2064
2065         return false;
2066 }
2067
2068 /* check for ISM devices matching proposed ISM devices */
2069 static void smc_check_ism_v2_match(struct smc_init_info *ini,
2070                                    u16 proposed_chid, u64 proposed_gid,
2071                                    unsigned int *matches)
2072 {
2073         struct smcd_dev *smcd;
2074
2075         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
2076                 if (smcd->going_away)
2077                         continue;
2078                 if (smc_is_already_selected(smcd, ini, *matches))
2079                         continue;
2080                 if (smc_ism_get_chid(smcd) == proposed_chid &&
2081                     !smc_ism_cantalk(proposed_gid, ISM_RESERVED_VLANID, smcd)) {
2082                         ini->ism_peer_gid[*matches] = proposed_gid;
2083                         ini->ism_dev[*matches] = smcd;
2084                         (*matches)++;
2085                         break;
2086                 }
2087         }
2088 }
2089
2090 static void smc_find_ism_store_rc(u32 rc, struct smc_init_info *ini)
2091 {
2092         if (!ini->rc)
2093                 ini->rc = rc;
2094 }
2095
2096 static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
2097                                         struct smc_clc_msg_proposal *pclc,
2098                                         struct smc_init_info *ini)
2099 {
2100         struct smc_clc_smcd_v2_extension *smcd_v2_ext;
2101         struct smc_clc_v2_extension *smc_v2_ext;
2102         struct smc_clc_msg_smcd *pclc_smcd;
2103         unsigned int matches = 0;
2104         u8 smcd_version;
2105         u8 *eid = NULL;
2106         int i, rc;
2107
2108         if (!(ini->smcd_version & SMC_V2) || !smcd_indicated(ini->smc_type_v2))
2109                 goto not_found;
2110
2111         pclc_smcd = smc_get_clc_msg_smcd(pclc);
2112         smc_v2_ext = smc_get_clc_v2_ext(pclc);
2113         smcd_v2_ext = smc_get_clc_smcd_v2_ext(smc_v2_ext);
2114
2115         mutex_lock(&smcd_dev_list.mutex);
2116         if (pclc_smcd->ism.chid)
2117                 /* check for ISM device matching proposed native ISM device */
2118                 smc_check_ism_v2_match(ini, ntohs(pclc_smcd->ism.chid),
2119                                        ntohll(pclc_smcd->ism.gid), &matches);
2120         for (i = 1; i <= smc_v2_ext->hdr.ism_gid_cnt; i++) {
2121                 /* check for ISM devices matching proposed non-native ISM
2122                  * devices
2123                  */
2124                 smc_check_ism_v2_match(ini,
2125                                        ntohs(smcd_v2_ext->gidchid[i - 1].chid),
2126                                        ntohll(smcd_v2_ext->gidchid[i - 1].gid),
2127                                        &matches);
2128         }
2129         mutex_unlock(&smcd_dev_list.mutex);
2130
2131         if (!ini->ism_dev[0]) {
2132                 smc_find_ism_store_rc(SMC_CLC_DECL_NOSMCD2DEV, ini);
2133                 goto not_found;
2134         }
2135
2136         smc_ism_get_system_eid(&eid);
2137         if (!smc_clc_match_eid(ini->negotiated_eid, smc_v2_ext,
2138                                smcd_v2_ext->system_eid, eid))
2139                 goto not_found;
2140
2141         /* separate - outside the smcd_dev_list.lock */
2142         smcd_version = ini->smcd_version;
2143         for (i = 0; i < matches; i++) {
2144                 ini->smcd_version = SMC_V2;
2145                 ini->is_smcd = true;
2146                 ini->ism_selected = i;
2147                 rc = smc_listen_ism_init(new_smc, ini);
2148                 if (rc) {
2149                         smc_find_ism_store_rc(rc, ini);
2150                         /* try next active ISM device */
2151                         continue;
2152                 }
2153                 return; /* matching and usable V2 ISM device found */
2154         }
2155         /* no V2 ISM device could be initialized */
2156         ini->smcd_version = smcd_version;       /* restore original value */
2157         ini->negotiated_eid[0] = 0;
2158
2159 not_found:
2160         ini->smcd_version &= ~SMC_V2;
2161         ini->ism_dev[0] = NULL;
2162         ini->is_smcd = false;
2163 }
2164
2165 static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
2166                                         struct smc_clc_msg_proposal *pclc,
2167                                         struct smc_init_info *ini)
2168 {
2169         struct smc_clc_msg_smcd *pclc_smcd = smc_get_clc_msg_smcd(pclc);
2170         int rc = 0;
2171
2172         /* check if ISM V1 is available */
2173         if (!(ini->smcd_version & SMC_V1) || !smcd_indicated(ini->smc_type_v1))
2174                 goto not_found;
2175         ini->is_smcd = true; /* prepare ISM check */
2176         ini->ism_peer_gid[0] = ntohll(pclc_smcd->ism.gid);
2177         rc = smc_find_ism_device(new_smc, ini);
2178         if (rc)
2179                 goto not_found;
2180         ini->ism_selected = 0;
2181         rc = smc_listen_ism_init(new_smc, ini);
2182         if (!rc)
2183                 return;         /* V1 ISM device found */
2184
2185 not_found:
2186         smc_find_ism_store_rc(rc, ini);
2187         ini->smcd_version &= ~SMC_V1;
2188         ini->ism_dev[0] = NULL;
2189         ini->is_smcd = false;
2190 }
2191
2192 /* listen worker: register buffers */
2193 static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
2194 {
2195         struct smc_connection *conn = &new_smc->conn;
2196
2197         if (!local_first) {
2198                 /* reg sendbufs if they were vzalloced */
2199                 if (conn->sndbuf_desc->is_vm) {
2200                         if (smcr_lgr_reg_sndbufs(conn->lnk,
2201                                                  conn->sndbuf_desc))
2202                                 return SMC_CLC_DECL_ERR_REGBUF;
2203                 }
2204                 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
2205                         return SMC_CLC_DECL_ERR_REGBUF;
2206         }
2207
2208         return 0;
2209 }
2210
2211 static void smc_find_rdma_v2_device_serv(struct smc_sock *new_smc,
2212                                          struct smc_clc_msg_proposal *pclc,
2213                                          struct smc_init_info *ini)
2214 {
2215         struct smc_clc_v2_extension *smc_v2_ext;
2216         u8 smcr_version;
2217         int rc;
2218
2219         if (!(ini->smcr_version & SMC_V2) || !smcr_indicated(ini->smc_type_v2))
2220                 goto not_found;
2221
2222         smc_v2_ext = smc_get_clc_v2_ext(pclc);
2223         if (!smc_clc_match_eid(ini->negotiated_eid, smc_v2_ext, NULL, NULL))
2224                 goto not_found;
2225
2226         /* prepare RDMA check */
2227         memcpy(ini->peer_systemid, pclc->lcl.id_for_peer, SMC_SYSTEMID_LEN);
2228         memcpy(ini->peer_gid, smc_v2_ext->roce, SMC_GID_SIZE);
2229         memcpy(ini->peer_mac, pclc->lcl.mac, ETH_ALEN);
2230         ini->check_smcrv2 = true;
2231         ini->smcrv2.clc_sk = new_smc->clcsock->sk;
2232         ini->smcrv2.saddr = new_smc->clcsock->sk->sk_rcv_saddr;
2233         ini->smcrv2.daddr = smc_ib_gid_to_ipv4(smc_v2_ext->roce);
2234         rc = smc_find_rdma_device(new_smc, ini);
2235         if (rc) {
2236                 smc_find_ism_store_rc(rc, ini);
2237                 goto not_found;
2238         }
2239         if (!ini->smcrv2.uses_gateway)
2240                 memcpy(ini->smcrv2.nexthop_mac, pclc->lcl.mac, ETH_ALEN);
2241
2242         smcr_version = ini->smcr_version;
2243         ini->smcr_version = SMC_V2;
2244         rc = smc_listen_rdma_init(new_smc, ini);
2245         if (!rc) {
2246                 rc = smc_listen_rdma_reg(new_smc, ini->first_contact_local);
2247                 if (rc)
2248                         smc_conn_abort(new_smc, ini->first_contact_local);
2249         }
2250         if (!rc)
2251                 return;
2252         ini->smcr_version = smcr_version;
2253         smc_find_ism_store_rc(rc, ini);
2254
2255 not_found:
2256         ini->smcr_version &= ~SMC_V2;
2257         ini->smcrv2.ib_dev_v2 = NULL;
2258         ini->check_smcrv2 = false;
2259 }
2260
2261 static int smc_find_rdma_v1_device_serv(struct smc_sock *new_smc,
2262                                         struct smc_clc_msg_proposal *pclc,
2263                                         struct smc_init_info *ini)
2264 {
2265         int rc;
2266
2267         if (!(ini->smcr_version & SMC_V1) || !smcr_indicated(ini->smc_type_v1))
2268                 return SMC_CLC_DECL_NOSMCDEV;
2269
2270         /* prepare RDMA check */
2271         memcpy(ini->peer_systemid, pclc->lcl.id_for_peer, SMC_SYSTEMID_LEN);
2272         memcpy(ini->peer_gid, pclc->lcl.gid, SMC_GID_SIZE);
2273         memcpy(ini->peer_mac, pclc->lcl.mac, ETH_ALEN);
2274         rc = smc_find_rdma_device(new_smc, ini);
2275         if (rc) {
2276                 /* no RDMA device found */
2277                 return SMC_CLC_DECL_NOSMCDEV;
2278         }
2279         rc = smc_listen_rdma_init(new_smc, ini);
2280         if (rc)
2281                 return rc;
2282         return smc_listen_rdma_reg(new_smc, ini->first_contact_local);
2283 }
2284
2285 /* determine the local device matching to proposal */
2286 static int smc_listen_find_device(struct smc_sock *new_smc,
2287                                   struct smc_clc_msg_proposal *pclc,
2288                                   struct smc_init_info *ini)
2289 {
2290         int prfx_rc;
2291
2292         /* check for ISM device matching V2 proposed device */
2293         smc_find_ism_v2_device_serv(new_smc, pclc, ini);
2294         if (ini->ism_dev[0])
2295                 return 0;
2296
2297         /* check for matching IP prefix and subnet length (V1) */
2298         prfx_rc = smc_listen_prfx_check(new_smc, pclc);
2299         if (prfx_rc)
2300                 smc_find_ism_store_rc(prfx_rc, ini);
2301
2302         /* get vlan id from IP device */
2303         if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
2304                 return ini->rc ?: SMC_CLC_DECL_GETVLANERR;
2305
2306         /* check for ISM device matching V1 proposed device */
2307         if (!prfx_rc)
2308                 smc_find_ism_v1_device_serv(new_smc, pclc, ini);
2309         if (ini->ism_dev[0])
2310                 return 0;
2311
2312         if (!smcr_indicated(pclc->hdr.typev1) &&
2313             !smcr_indicated(pclc->hdr.typev2))
2314                 /* skip RDMA and decline */
2315                 return ini->rc ?: SMC_CLC_DECL_NOSMCDDEV;
2316
2317         /* check if RDMA V2 is available */
2318         smc_find_rdma_v2_device_serv(new_smc, pclc, ini);
2319         if (ini->smcrv2.ib_dev_v2)
2320                 return 0;
2321
2322         /* check if RDMA V1 is available */
2323         if (!prfx_rc) {
2324                 int rc;
2325
2326                 rc = smc_find_rdma_v1_device_serv(new_smc, pclc, ini);
2327                 smc_find_ism_store_rc(rc, ini);
2328                 return (!rc) ? 0 : ini->rc;
2329         }
2330         return prfx_rc;
2331 }
2332
2333 /* listen worker: finish RDMA setup */
2334 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
2335                                   struct smc_clc_msg_accept_confirm *cclc,
2336                                   bool local_first,
2337                                   struct smc_init_info *ini)
2338 {
2339         struct smc_link *link = new_smc->conn.lnk;
2340         int reason_code = 0;
2341
2342         if (local_first)
2343                 smc_link_save_peer_info(link, cclc, ini);
2344
2345         if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc))
2346                 return SMC_CLC_DECL_ERR_RTOK;
2347
2348         if (local_first) {
2349                 if (smc_ib_ready_link(link))
2350                         return SMC_CLC_DECL_ERR_RDYLNK;
2351                 /* QP confirmation over RoCE fabric */
2352                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
2353                 reason_code = smcr_serv_conf_first_link(new_smc);
2354                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
2355         }
2356         return reason_code;
2357 }
2358
2359 /* setup for connection of server */
2360 static void smc_listen_work(struct work_struct *work)
2361 {
2362         struct smc_sock *new_smc = container_of(work, struct smc_sock,
2363                                                 smc_listen_work);
2364         struct socket *newclcsock = new_smc->clcsock;
2365         struct smc_clc_msg_accept_confirm *cclc;
2366         struct smc_clc_msg_proposal_area *buf;
2367         struct smc_clc_msg_proposal *pclc;
2368         struct smc_init_info *ini = NULL;
2369         u8 proposal_version = SMC_V1;
2370         u8 accept_version;
2371         int rc = 0;
2372
2373         if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
2374                 return smc_listen_out_err(new_smc);
2375
2376         if (new_smc->use_fallback) {
2377                 smc_listen_out_connected(new_smc);
2378                 return;
2379         }
2380
2381         /* check if peer is smc capable */
2382         if (!tcp_sk(newclcsock->sk)->syn_smc) {
2383                 rc = smc_switch_to_fallback(new_smc, SMC_CLC_DECL_PEERNOSMC);
2384                 if (rc)
2385                         smc_listen_out_err(new_smc);
2386                 else
2387                         smc_listen_out_connected(new_smc);
2388                 return;
2389         }
2390
2391         /* do inband token exchange -
2392          * wait for and receive SMC Proposal CLC message
2393          */
2394         buf = kzalloc(sizeof(*buf), GFP_KERNEL);
2395         if (!buf) {
2396                 rc = SMC_CLC_DECL_MEM;
2397                 goto out_decl;
2398         }
2399         pclc = (struct smc_clc_msg_proposal *)buf;
2400         rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
2401                               SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
2402         if (rc)
2403                 goto out_decl;
2404
2405         if (pclc->hdr.version > SMC_V1)
2406                 proposal_version = SMC_V2;
2407
2408         /* IPSec connections opt out of SMC optimizations */
2409         if (using_ipsec(new_smc)) {
2410                 rc = SMC_CLC_DECL_IPSEC;
2411                 goto out_decl;
2412         }
2413
2414         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
2415         if (!ini) {
2416                 rc = SMC_CLC_DECL_MEM;
2417                 goto out_decl;
2418         }
2419
2420         /* initial version checking */
2421         rc = smc_listen_v2_check(new_smc, pclc, ini);
2422         if (rc)
2423                 goto out_decl;
2424
2425         mutex_lock(&smc_server_lgr_pending);
2426         smc_close_init(new_smc);
2427         smc_rx_init(new_smc);
2428         smc_tx_init(new_smc);
2429
2430         /* determine ISM or RoCE device used for connection */
2431         rc = smc_listen_find_device(new_smc, pclc, ini);
2432         if (rc)
2433                 goto out_unlock;
2434
2435         /* send SMC Accept CLC message */
2436         accept_version = ini->is_smcd ? ini->smcd_version : ini->smcr_version;
2437         rc = smc_clc_send_accept(new_smc, ini->first_contact_local,
2438                                  accept_version, ini->negotiated_eid);
2439         if (rc)
2440                 goto out_unlock;
2441
2442         /* SMC-D does not need this lock any more */
2443         if (ini->is_smcd)
2444                 mutex_unlock(&smc_server_lgr_pending);
2445
2446         /* receive SMC Confirm CLC message */
2447         memset(buf, 0, sizeof(*buf));
2448         cclc = (struct smc_clc_msg_accept_confirm *)buf;
2449         rc = smc_clc_wait_msg(new_smc, cclc, sizeof(*buf),
2450                               SMC_CLC_CONFIRM, CLC_WAIT_TIME);
2451         if (rc) {
2452                 if (!ini->is_smcd)
2453                         goto out_unlock;
2454                 goto out_decl;
2455         }
2456
2457         /* finish worker */
2458         if (!ini->is_smcd) {
2459                 rc = smc_listen_rdma_finish(new_smc, cclc,
2460                                             ini->first_contact_local, ini);
2461                 if (rc)
2462                         goto out_unlock;
2463                 mutex_unlock(&smc_server_lgr_pending);
2464         }
2465         smc_conn_save_peer_info(new_smc, cclc);
2466         smc_listen_out_connected(new_smc);
2467         SMC_STAT_SERV_SUCC_INC(sock_net(newclcsock->sk), ini);
2468         goto out_free;
2469
2470 out_unlock:
2471         mutex_unlock(&smc_server_lgr_pending);
2472 out_decl:
2473         smc_listen_decline(new_smc, rc, ini ? ini->first_contact_local : 0,
2474                            proposal_version);
2475 out_free:
2476         kfree(ini);
2477         kfree(buf);
2478 }
2479
2480 static void smc_tcp_listen_work(struct work_struct *work)
2481 {
2482         struct smc_sock *lsmc = container_of(work, struct smc_sock,
2483                                              tcp_listen_work);
2484         struct sock *lsk = &lsmc->sk;
2485         struct smc_sock *new_smc;
2486         int rc = 0;
2487
2488         lock_sock(lsk);
2489         while (lsk->sk_state == SMC_LISTEN) {
2490                 rc = smc_clcsock_accept(lsmc, &new_smc);
2491                 if (rc) /* clcsock accept queue empty or error */
2492                         goto out;
2493                 if (!new_smc)
2494                         continue;
2495
2496                 if (tcp_sk(new_smc->clcsock->sk)->syn_smc)
2497                         atomic_inc(&lsmc->queued_smc_hs);
2498
2499                 new_smc->listen_smc = lsmc;
2500                 new_smc->use_fallback = lsmc->use_fallback;
2501                 new_smc->fallback_rsn = lsmc->fallback_rsn;
2502                 sock_hold(lsk); /* sock_put in smc_listen_work */
2503                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
2504                 smc_copy_sock_settings_to_smc(new_smc);
2505                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
2506                 if (!queue_work(smc_hs_wq, &new_smc->smc_listen_work))
2507                         sock_put(&new_smc->sk);
2508         }
2509
2510 out:
2511         release_sock(lsk);
2512         sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
2513 }
2514
2515 static void smc_clcsock_data_ready(struct sock *listen_clcsock)
2516 {
2517         struct smc_sock *lsmc;
2518
2519         read_lock_bh(&listen_clcsock->sk_callback_lock);
2520         lsmc = smc_clcsock_user_data(listen_clcsock);
2521         if (!lsmc)
2522                 goto out;
2523         lsmc->clcsk_data_ready(listen_clcsock);
2524         if (lsmc->sk.sk_state == SMC_LISTEN) {
2525                 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
2526                 if (!queue_work(smc_tcp_ls_wq, &lsmc->tcp_listen_work))
2527                         sock_put(&lsmc->sk);
2528         }
2529 out:
2530         read_unlock_bh(&listen_clcsock->sk_callback_lock);
2531 }
2532
2533 static int smc_listen(struct socket *sock, int backlog)
2534 {
2535         struct sock *sk = sock->sk;
2536         struct smc_sock *smc;
2537         int rc;
2538
2539         smc = smc_sk(sk);
2540         lock_sock(sk);
2541
2542         rc = -EINVAL;
2543         if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
2544             smc->connect_nonblock || sock->state != SS_UNCONNECTED)
2545                 goto out;
2546
2547         rc = 0;
2548         if (sk->sk_state == SMC_LISTEN) {
2549                 sk->sk_max_ack_backlog = backlog;
2550                 goto out;
2551         }
2552         /* some socket options are handled in core, so we could not apply
2553          * them to the clc socket -- copy smc socket options to clc socket
2554          */
2555         smc_copy_sock_settings_to_clc(smc);
2556         if (!smc->use_fallback)
2557                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
2558
2559         /* save original sk_data_ready function and establish
2560          * smc-specific sk_data_ready function
2561          */
2562         write_lock_bh(&smc->clcsock->sk->sk_callback_lock);
2563         smc->clcsock->sk->sk_user_data =
2564                 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
2565         smc_clcsock_replace_cb(&smc->clcsock->sk->sk_data_ready,
2566                                smc_clcsock_data_ready, &smc->clcsk_data_ready);
2567         write_unlock_bh(&smc->clcsock->sk->sk_callback_lock);
2568
2569         /* save original ops */
2570         smc->ori_af_ops = inet_csk(smc->clcsock->sk)->icsk_af_ops;
2571
2572         smc->af_ops = *smc->ori_af_ops;
2573         smc->af_ops.syn_recv_sock = smc_tcp_syn_recv_sock;
2574
2575         inet_csk(smc->clcsock->sk)->icsk_af_ops = &smc->af_ops;
2576
2577         if (smc->limit_smc_hs)
2578                 tcp_sk(smc->clcsock->sk)->smc_hs_congested = smc_hs_congested;
2579
2580         rc = kernel_listen(smc->clcsock, backlog);
2581         if (rc) {
2582                 write_lock_bh(&smc->clcsock->sk->sk_callback_lock);
2583                 smc_clcsock_restore_cb(&smc->clcsock->sk->sk_data_ready,
2584                                        &smc->clcsk_data_ready);
2585                 smc->clcsock->sk->sk_user_data = NULL;
2586                 write_unlock_bh(&smc->clcsock->sk->sk_callback_lock);
2587                 goto out;
2588         }
2589         sk->sk_max_ack_backlog = backlog;
2590         sk->sk_ack_backlog = 0;
2591         sk->sk_state = SMC_LISTEN;
2592
2593 out:
2594         release_sock(sk);
2595         return rc;
2596 }
2597
2598 static int smc_accept(struct socket *sock, struct socket *new_sock,
2599                       int flags, bool kern)
2600 {
2601         struct sock *sk = sock->sk, *nsk;
2602         DECLARE_WAITQUEUE(wait, current);
2603         struct smc_sock *lsmc;
2604         long timeo;
2605         int rc = 0;
2606
2607         lsmc = smc_sk(sk);
2608         sock_hold(sk); /* sock_put below */
2609         lock_sock(sk);
2610
2611         if (lsmc->sk.sk_state != SMC_LISTEN) {
2612                 rc = -EINVAL;
2613                 release_sock(sk);
2614                 goto out;
2615         }
2616
2617         /* Wait for an incoming connection */
2618         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
2619         add_wait_queue_exclusive(sk_sleep(sk), &wait);
2620         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
2621                 set_current_state(TASK_INTERRUPTIBLE);
2622                 if (!timeo) {
2623                         rc = -EAGAIN;
2624                         break;
2625                 }
2626                 release_sock(sk);
2627                 timeo = schedule_timeout(timeo);
2628                 /* wakeup by sk_data_ready in smc_listen_work() */
2629                 sched_annotate_sleep();
2630                 lock_sock(sk);
2631                 if (signal_pending(current)) {
2632                         rc = sock_intr_errno(timeo);
2633                         break;
2634                 }
2635         }
2636         set_current_state(TASK_RUNNING);
2637         remove_wait_queue(sk_sleep(sk), &wait);
2638
2639         if (!rc)
2640                 rc = sock_error(nsk);
2641         release_sock(sk);
2642         if (rc)
2643                 goto out;
2644
2645         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
2646                 /* wait till data arrives on the socket */
2647                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
2648                                                                 MSEC_PER_SEC);
2649                 if (smc_sk(nsk)->use_fallback) {
2650                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
2651
2652                         lock_sock(clcsk);
2653                         if (skb_queue_empty(&clcsk->sk_receive_queue))
2654                                 sk_wait_data(clcsk, &timeo, NULL);
2655                         release_sock(clcsk);
2656                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
2657                         lock_sock(nsk);
2658                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
2659                         release_sock(nsk);
2660                 }
2661         }
2662
2663 out:
2664         sock_put(sk); /* sock_hold above */
2665         return rc;
2666 }
2667
2668 static int smc_getname(struct socket *sock, struct sockaddr *addr,
2669                        int peer)
2670 {
2671         struct smc_sock *smc;
2672
2673         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
2674             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
2675                 return -ENOTCONN;
2676
2677         smc = smc_sk(sock->sk);
2678
2679         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
2680 }
2681
2682 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
2683 {
2684         struct sock *sk = sock->sk;
2685         struct smc_sock *smc;
2686         int rc;
2687
2688         smc = smc_sk(sk);
2689         lock_sock(sk);
2690
2691         /* SMC does not support connect with fastopen */
2692         if (msg->msg_flags & MSG_FASTOPEN) {
2693                 /* not connected yet, fallback */
2694                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2695                         rc = smc_switch_to_fallback(smc, SMC_CLC_DECL_OPTUNSUPP);
2696                         if (rc)
2697                                 goto out;
2698                 } else {
2699                         rc = -EINVAL;
2700                         goto out;
2701                 }
2702         } else if ((sk->sk_state != SMC_ACTIVE) &&
2703                    (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2704                    (sk->sk_state != SMC_INIT)) {
2705                 rc = -EPIPE;
2706                 goto out;
2707         }
2708
2709         if (smc->use_fallback) {
2710                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
2711         } else {
2712                 rc = smc_tx_sendmsg(smc, msg, len);
2713                 SMC_STAT_TX_PAYLOAD(smc, len, rc);
2714         }
2715 out:
2716         release_sock(sk);
2717         return rc;
2718 }
2719
2720 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
2721                        int flags)
2722 {
2723         struct sock *sk = sock->sk;
2724         struct smc_sock *smc;
2725         int rc = -ENOTCONN;
2726
2727         smc = smc_sk(sk);
2728         lock_sock(sk);
2729         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2730                 /* socket was connected before, no more data to read */
2731                 rc = 0;
2732                 goto out;
2733         }
2734         if ((sk->sk_state == SMC_INIT) ||
2735             (sk->sk_state == SMC_LISTEN) ||
2736             (sk->sk_state == SMC_CLOSED))
2737                 goto out;
2738
2739         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2740                 rc = 0;
2741                 goto out;
2742         }
2743
2744         if (smc->use_fallback) {
2745                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
2746         } else {
2747                 msg->msg_namelen = 0;
2748                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
2749                 SMC_STAT_RX_PAYLOAD(smc, rc, rc);
2750         }
2751
2752 out:
2753         release_sock(sk);
2754         return rc;
2755 }
2756
2757 static __poll_t smc_accept_poll(struct sock *parent)
2758 {
2759         struct smc_sock *isk = smc_sk(parent);
2760         __poll_t mask = 0;
2761
2762         spin_lock(&isk->accept_q_lock);
2763         if (!list_empty(&isk->accept_q))
2764                 mask = EPOLLIN | EPOLLRDNORM;
2765         spin_unlock(&isk->accept_q_lock);
2766
2767         return mask;
2768 }
2769
2770 static __poll_t smc_poll(struct file *file, struct socket *sock,
2771                              poll_table *wait)
2772 {
2773         struct sock *sk = sock->sk;
2774         struct smc_sock *smc;
2775         __poll_t mask = 0;
2776
2777         if (!sk)
2778                 return EPOLLNVAL;
2779
2780         smc = smc_sk(sock->sk);
2781         if (smc->use_fallback) {
2782                 /* delegate to CLC child sock */
2783                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
2784                 sk->sk_err = smc->clcsock->sk->sk_err;
2785         } else {
2786                 if (sk->sk_state != SMC_CLOSED)
2787                         sock_poll_wait(file, sock, wait);
2788                 if (sk->sk_err)
2789                         mask |= EPOLLERR;
2790                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
2791                     (sk->sk_state == SMC_CLOSED))
2792                         mask |= EPOLLHUP;
2793                 if (sk->sk_state == SMC_LISTEN) {
2794                         /* woken up by sk_data_ready in smc_listen_work() */
2795                         mask |= smc_accept_poll(sk);
2796                 } else if (smc->use_fallback) { /* as result of connect_work()*/
2797                         mask |= smc->clcsock->ops->poll(file, smc->clcsock,
2798                                                            wait);
2799                         sk->sk_err = smc->clcsock->sk->sk_err;
2800                 } else {
2801                         if ((sk->sk_state != SMC_INIT &&
2802                              atomic_read(&smc->conn.sndbuf_space)) ||
2803                             sk->sk_shutdown & SEND_SHUTDOWN) {
2804                                 mask |= EPOLLOUT | EPOLLWRNORM;
2805                         } else {
2806                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
2807                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2808                         }
2809                         if (atomic_read(&smc->conn.bytes_to_rcv))
2810                                 mask |= EPOLLIN | EPOLLRDNORM;
2811                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2812                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2813                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
2814                                 mask |= EPOLLIN;
2815                         if (smc->conn.urg_state == SMC_URG_VALID)
2816                                 mask |= EPOLLPRI;
2817                 }
2818         }
2819
2820         return mask;
2821 }
2822
2823 static int smc_shutdown(struct socket *sock, int how)
2824 {
2825         struct sock *sk = sock->sk;
2826         bool do_shutdown = true;
2827         struct smc_sock *smc;
2828         int rc = -EINVAL;
2829         int old_state;
2830         int rc1 = 0;
2831
2832         smc = smc_sk(sk);
2833
2834         if ((how < SHUT_RD) || (how > SHUT_RDWR))
2835                 return rc;
2836
2837         lock_sock(sk);
2838
2839         if (sock->state == SS_CONNECTING) {
2840                 if (sk->sk_state == SMC_ACTIVE)
2841                         sock->state = SS_CONNECTED;
2842                 else if (sk->sk_state == SMC_PEERCLOSEWAIT1 ||
2843                          sk->sk_state == SMC_PEERCLOSEWAIT2 ||
2844                          sk->sk_state == SMC_APPCLOSEWAIT1 ||
2845                          sk->sk_state == SMC_APPCLOSEWAIT2 ||
2846                          sk->sk_state == SMC_APPFINCLOSEWAIT)
2847                         sock->state = SS_DISCONNECTING;
2848         }
2849
2850         rc = -ENOTCONN;
2851         if ((sk->sk_state != SMC_ACTIVE) &&
2852             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
2853             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
2854             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2855             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
2856             (sk->sk_state != SMC_APPFINCLOSEWAIT))
2857                 goto out;
2858         if (smc->use_fallback) {
2859                 rc = kernel_sock_shutdown(smc->clcsock, how);
2860                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
2861                 if (sk->sk_shutdown == SHUTDOWN_MASK) {
2862                         sk->sk_state = SMC_CLOSED;
2863                         sk->sk_socket->state = SS_UNCONNECTED;
2864                         sock_put(sk);
2865                 }
2866                 goto out;
2867         }
2868         switch (how) {
2869         case SHUT_RDWR:         /* shutdown in both directions */
2870                 old_state = sk->sk_state;
2871                 rc = smc_close_active(smc);
2872                 if (old_state == SMC_ACTIVE &&
2873                     sk->sk_state == SMC_PEERCLOSEWAIT1)
2874                         do_shutdown = false;
2875                 break;
2876         case SHUT_WR:
2877                 rc = smc_close_shutdown_write(smc);
2878                 break;
2879         case SHUT_RD:
2880                 rc = 0;
2881                 /* nothing more to do because peer is not involved */
2882                 break;
2883         }
2884         if (do_shutdown && smc->clcsock)
2885                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
2886         /* map sock_shutdown_cmd constants to sk_shutdown value range */
2887         sk->sk_shutdown |= how + 1;
2888
2889         if (sk->sk_state == SMC_CLOSED)
2890                 sock->state = SS_UNCONNECTED;
2891         else
2892                 sock->state = SS_DISCONNECTING;
2893 out:
2894         release_sock(sk);
2895         return rc ? rc : rc1;
2896 }
2897
2898 static int __smc_getsockopt(struct socket *sock, int level, int optname,
2899                             char __user *optval, int __user *optlen)
2900 {
2901         struct smc_sock *smc;
2902         int val, len;
2903
2904         smc = smc_sk(sock->sk);
2905
2906         if (get_user(len, optlen))
2907                 return -EFAULT;
2908
2909         len = min_t(int, len, sizeof(int));
2910
2911         if (len < 0)
2912                 return -EINVAL;
2913
2914         switch (optname) {
2915         case SMC_LIMIT_HS:
2916                 val = smc->limit_smc_hs;
2917                 break;
2918         default:
2919                 return -EOPNOTSUPP;
2920         }
2921
2922         if (put_user(len, optlen))
2923                 return -EFAULT;
2924         if (copy_to_user(optval, &val, len))
2925                 return -EFAULT;
2926
2927         return 0;
2928 }
2929
2930 static int __smc_setsockopt(struct socket *sock, int level, int optname,
2931                             sockptr_t optval, unsigned int optlen)
2932 {
2933         struct sock *sk = sock->sk;
2934         struct smc_sock *smc;
2935         int val, rc;
2936
2937         smc = smc_sk(sk);
2938
2939         lock_sock(sk);
2940         switch (optname) {
2941         case SMC_LIMIT_HS:
2942                 if (optlen < sizeof(int)) {
2943                         rc = -EINVAL;
2944                         break;
2945                 }
2946                 if (copy_from_sockptr(&val, optval, sizeof(int))) {
2947                         rc = -EFAULT;
2948                         break;
2949                 }
2950
2951                 smc->limit_smc_hs = !!val;
2952                 rc = 0;
2953                 break;
2954         default:
2955                 rc = -EOPNOTSUPP;
2956                 break;
2957         }
2958         release_sock(sk);
2959
2960         return rc;
2961 }
2962
2963 static int smc_setsockopt(struct socket *sock, int level, int optname,
2964                           sockptr_t optval, unsigned int optlen)
2965 {
2966         struct sock *sk = sock->sk;
2967         struct smc_sock *smc;
2968         int val, rc;
2969
2970         if (level == SOL_TCP && optname == TCP_ULP)
2971                 return -EOPNOTSUPP;
2972         else if (level == SOL_SMC)
2973                 return __smc_setsockopt(sock, level, optname, optval, optlen);
2974
2975         smc = smc_sk(sk);
2976
2977         /* generic setsockopts reaching us here always apply to the
2978          * CLC socket
2979          */
2980         mutex_lock(&smc->clcsock_release_lock);
2981         if (!smc->clcsock) {
2982                 mutex_unlock(&smc->clcsock_release_lock);
2983                 return -EBADF;
2984         }
2985         if (unlikely(!smc->clcsock->ops->setsockopt))
2986                 rc = -EOPNOTSUPP;
2987         else
2988                 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
2989                                                    optval, optlen);
2990         if (smc->clcsock->sk->sk_err) {
2991                 sk->sk_err = smc->clcsock->sk->sk_err;
2992                 sk_error_report(sk);
2993         }
2994         mutex_unlock(&smc->clcsock_release_lock);
2995
2996         if (optlen < sizeof(int))
2997                 return -EINVAL;
2998         if (copy_from_sockptr(&val, optval, sizeof(int)))
2999                 return -EFAULT;
3000
3001         lock_sock(sk);
3002         if (rc || smc->use_fallback)
3003                 goto out;
3004         switch (optname) {
3005         case TCP_FASTOPEN:
3006         case TCP_FASTOPEN_CONNECT:
3007         case TCP_FASTOPEN_KEY:
3008         case TCP_FASTOPEN_NO_COOKIE:
3009                 /* option not supported by SMC */
3010                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
3011                         rc = smc_switch_to_fallback(smc, SMC_CLC_DECL_OPTUNSUPP);
3012                 } else {
3013                         rc = -EINVAL;
3014                 }
3015                 break;
3016         case TCP_NODELAY:
3017                 if (sk->sk_state != SMC_INIT &&
3018                     sk->sk_state != SMC_LISTEN &&
3019                     sk->sk_state != SMC_CLOSED) {
3020                         if (val) {
3021                                 SMC_STAT_INC(smc, ndly_cnt);
3022                                 smc_tx_pending(&smc->conn);
3023                                 cancel_delayed_work(&smc->conn.tx_work);
3024                         }
3025                 }
3026                 break;
3027         case TCP_CORK:
3028                 if (sk->sk_state != SMC_INIT &&
3029                     sk->sk_state != SMC_LISTEN &&
3030                     sk->sk_state != SMC_CLOSED) {
3031                         if (!val) {
3032                                 SMC_STAT_INC(smc, cork_cnt);
3033                                 smc_tx_pending(&smc->conn);
3034                                 cancel_delayed_work(&smc->conn.tx_work);
3035                         }
3036                 }
3037                 break;
3038         case TCP_DEFER_ACCEPT:
3039                 smc->sockopt_defer_accept = val;
3040                 break;
3041         default:
3042                 break;
3043         }
3044 out:
3045         release_sock(sk);
3046
3047         return rc;
3048 }
3049
3050 static int smc_getsockopt(struct socket *sock, int level, int optname,
3051                           char __user *optval, int __user *optlen)
3052 {
3053         struct smc_sock *smc;
3054         int rc;
3055
3056         if (level == SOL_SMC)
3057                 return __smc_getsockopt(sock, level, optname, optval, optlen);
3058
3059         smc = smc_sk(sock->sk);
3060         mutex_lock(&smc->clcsock_release_lock);
3061         if (!smc->clcsock) {
3062                 mutex_unlock(&smc->clcsock_release_lock);
3063                 return -EBADF;
3064         }
3065         /* socket options apply to the CLC socket */
3066         if (unlikely(!smc->clcsock->ops->getsockopt)) {
3067                 mutex_unlock(&smc->clcsock_release_lock);
3068                 return -EOPNOTSUPP;
3069         }
3070         rc = smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
3071                                            optval, optlen);
3072         mutex_unlock(&smc->clcsock_release_lock);
3073         return rc;
3074 }
3075
3076 static int smc_ioctl(struct socket *sock, unsigned int cmd,
3077                      unsigned long arg)
3078 {
3079         union smc_host_cursor cons, urg;
3080         struct smc_connection *conn;
3081         struct smc_sock *smc;
3082         int answ;
3083
3084         smc = smc_sk(sock->sk);
3085         conn = &smc->conn;
3086         lock_sock(&smc->sk);
3087         if (smc->use_fallback) {
3088                 if (!smc->clcsock) {
3089                         release_sock(&smc->sk);
3090                         return -EBADF;
3091                 }
3092                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
3093                 release_sock(&smc->sk);
3094                 return answ;
3095         }
3096         switch (cmd) {
3097         case SIOCINQ: /* same as FIONREAD */
3098                 if (smc->sk.sk_state == SMC_LISTEN) {
3099                         release_sock(&smc->sk);
3100                         return -EINVAL;
3101                 }
3102                 if (smc->sk.sk_state == SMC_INIT ||
3103                     smc->sk.sk_state == SMC_CLOSED)
3104                         answ = 0;
3105                 else
3106                         answ = atomic_read(&smc->conn.bytes_to_rcv);
3107                 break;
3108         case SIOCOUTQ:
3109                 /* output queue size (not send + not acked) */
3110                 if (smc->sk.sk_state == SMC_LISTEN) {
3111                         release_sock(&smc->sk);
3112                         return -EINVAL;
3113                 }
3114                 if (smc->sk.sk_state == SMC_INIT ||
3115                     smc->sk.sk_state == SMC_CLOSED)
3116                         answ = 0;
3117                 else
3118                         answ = smc->conn.sndbuf_desc->len -
3119                                         atomic_read(&smc->conn.sndbuf_space);
3120                 break;
3121         case SIOCOUTQNSD:
3122                 /* output queue size (not send only) */
3123                 if (smc->sk.sk_state == SMC_LISTEN) {
3124                         release_sock(&smc->sk);
3125                         return -EINVAL;
3126                 }
3127                 if (smc->sk.sk_state == SMC_INIT ||
3128                     smc->sk.sk_state == SMC_CLOSED)
3129                         answ = 0;
3130                 else
3131                         answ = smc_tx_prepared_sends(&smc->conn);
3132                 break;
3133         case SIOCATMARK:
3134                 if (smc->sk.sk_state == SMC_LISTEN) {
3135                         release_sock(&smc->sk);
3136                         return -EINVAL;
3137                 }
3138                 if (smc->sk.sk_state == SMC_INIT ||
3139                     smc->sk.sk_state == SMC_CLOSED) {
3140                         answ = 0;
3141                 } else {
3142                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
3143                         smc_curs_copy(&urg, &conn->urg_curs, conn);
3144                         answ = smc_curs_diff(conn->rmb_desc->len,
3145                                              &cons, &urg) == 1;
3146                 }
3147                 break;
3148         default:
3149                 release_sock(&smc->sk);
3150                 return -ENOIOCTLCMD;
3151         }
3152         release_sock(&smc->sk);
3153
3154         return put_user(answ, (int __user *)arg);
3155 }
3156
3157 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
3158                             int offset, size_t size, int flags)
3159 {
3160         struct sock *sk = sock->sk;
3161         struct smc_sock *smc;
3162         int rc = -EPIPE;
3163
3164         smc = smc_sk(sk);
3165         lock_sock(sk);
3166         if (sk->sk_state != SMC_ACTIVE) {
3167                 release_sock(sk);
3168                 goto out;
3169         }
3170         release_sock(sk);
3171         if (smc->use_fallback) {
3172                 rc = kernel_sendpage(smc->clcsock, page, offset,
3173                                      size, flags);
3174         } else {
3175                 lock_sock(sk);
3176                 rc = smc_tx_sendpage(smc, page, offset, size, flags);
3177                 release_sock(sk);
3178                 SMC_STAT_INC(smc, sendpage_cnt);
3179         }
3180
3181 out:
3182         return rc;
3183 }
3184
3185 /* Map the affected portions of the rmbe into an spd, note the number of bytes
3186  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
3187  * updates till whenever a respective page has been fully processed.
3188  * Note that subsequent recv() calls have to wait till all splice() processing
3189  * completed.
3190  */
3191 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
3192                                struct pipe_inode_info *pipe, size_t len,
3193                                unsigned int flags)
3194 {
3195         struct sock *sk = sock->sk;
3196         struct smc_sock *smc;
3197         int rc = -ENOTCONN;
3198
3199         smc = smc_sk(sk);
3200         lock_sock(sk);
3201         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
3202                 /* socket was connected before, no more data to read */
3203                 rc = 0;
3204                 goto out;
3205         }
3206         if (sk->sk_state == SMC_INIT ||
3207             sk->sk_state == SMC_LISTEN ||
3208             sk->sk_state == SMC_CLOSED)
3209                 goto out;
3210
3211         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
3212                 rc = 0;
3213                 goto out;
3214         }
3215
3216         if (smc->use_fallback) {
3217                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
3218                                                     pipe, len, flags);
3219         } else {
3220                 if (*ppos) {
3221                         rc = -ESPIPE;
3222                         goto out;
3223                 }
3224                 if (flags & SPLICE_F_NONBLOCK)
3225                         flags = MSG_DONTWAIT;
3226                 else
3227                         flags = 0;
3228                 SMC_STAT_INC(smc, splice_cnt);
3229                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
3230         }
3231 out:
3232         release_sock(sk);
3233
3234         return rc;
3235 }
3236
3237 /* must look like tcp */
3238 static const struct proto_ops smc_sock_ops = {
3239         .family         = PF_SMC,
3240         .owner          = THIS_MODULE,
3241         .release        = smc_release,
3242         .bind           = smc_bind,
3243         .connect        = smc_connect,
3244         .socketpair     = sock_no_socketpair,
3245         .accept         = smc_accept,
3246         .getname        = smc_getname,
3247         .poll           = smc_poll,
3248         .ioctl          = smc_ioctl,
3249         .listen         = smc_listen,
3250         .shutdown       = smc_shutdown,
3251         .setsockopt     = smc_setsockopt,
3252         .getsockopt     = smc_getsockopt,
3253         .sendmsg        = smc_sendmsg,
3254         .recvmsg        = smc_recvmsg,
3255         .mmap           = sock_no_mmap,
3256         .sendpage       = smc_sendpage,
3257         .splice_read    = smc_splice_read,
3258 };
3259
3260 static int __smc_create(struct net *net, struct socket *sock, int protocol,
3261                         int kern, struct socket *clcsock)
3262 {
3263         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
3264         struct smc_sock *smc;
3265         struct sock *sk;
3266         int rc;
3267
3268         rc = -ESOCKTNOSUPPORT;
3269         if (sock->type != SOCK_STREAM)
3270                 goto out;
3271
3272         rc = -EPROTONOSUPPORT;
3273         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
3274                 goto out;
3275
3276         rc = -ENOBUFS;
3277         sock->ops = &smc_sock_ops;
3278         sock->state = SS_UNCONNECTED;
3279         sk = smc_sock_alloc(net, sock, protocol);
3280         if (!sk)
3281                 goto out;
3282
3283         /* create internal TCP socket for CLC handshake and fallback */
3284         smc = smc_sk(sk);
3285         smc->use_fallback = false; /* assume rdma capability first */
3286         smc->fallback_rsn = 0;
3287
3288         /* default behavior from limit_smc_hs in every net namespace */
3289         smc->limit_smc_hs = net->smc.limit_smc_hs;
3290
3291         rc = 0;
3292         if (!clcsock) {
3293                 rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
3294                                       &smc->clcsock);
3295                 if (rc) {
3296                         sk_common_release(sk);
3297                         goto out;
3298                 }
3299         } else {
3300                 smc->clcsock = clcsock;
3301         }
3302
3303 out:
3304         return rc;
3305 }
3306
3307 static int smc_create(struct net *net, struct socket *sock, int protocol,
3308                       int kern)
3309 {
3310         return __smc_create(net, sock, protocol, kern, NULL);
3311 }
3312
3313 static const struct net_proto_family smc_sock_family_ops = {
3314         .family = PF_SMC,
3315         .owner  = THIS_MODULE,
3316         .create = smc_create,
3317 };
3318
3319 static int smc_ulp_init(struct sock *sk)
3320 {
3321         struct socket *tcp = sk->sk_socket;
3322         struct net *net = sock_net(sk);
3323         struct socket *smcsock;
3324         int protocol, ret;
3325
3326         /* only TCP can be replaced */
3327         if (tcp->type != SOCK_STREAM || sk->sk_protocol != IPPROTO_TCP ||
3328             (sk->sk_family != AF_INET && sk->sk_family != AF_INET6))
3329                 return -ESOCKTNOSUPPORT;
3330         /* don't handle wq now */
3331         if (tcp->state != SS_UNCONNECTED || !tcp->file || tcp->wq.fasync_list)
3332                 return -ENOTCONN;
3333
3334         if (sk->sk_family == AF_INET)
3335                 protocol = SMCPROTO_SMC;
3336         else
3337                 protocol = SMCPROTO_SMC6;
3338
3339         smcsock = sock_alloc();
3340         if (!smcsock)
3341                 return -ENFILE;
3342
3343         smcsock->type = SOCK_STREAM;
3344         __module_get(THIS_MODULE); /* tried in __tcp_ulp_find_autoload */
3345         ret = __smc_create(net, smcsock, protocol, 1, tcp);
3346         if (ret) {
3347                 sock_release(smcsock); /* module_put() which ops won't be NULL */
3348                 return ret;
3349         }
3350
3351         /* replace tcp socket to smc */
3352         smcsock->file = tcp->file;
3353         smcsock->file->private_data = smcsock;
3354         smcsock->file->f_inode = SOCK_INODE(smcsock); /* replace inode when sock_close */
3355         smcsock->file->f_path.dentry->d_inode = SOCK_INODE(smcsock); /* dput() in __fput */
3356         tcp->file = NULL;
3357
3358         return ret;
3359 }
3360
3361 static void smc_ulp_clone(const struct request_sock *req, struct sock *newsk,
3362                           const gfp_t priority)
3363 {
3364         struct inet_connection_sock *icsk = inet_csk(newsk);
3365
3366         /* don't inherit ulp ops to child when listen */
3367         icsk->icsk_ulp_ops = NULL;
3368 }
3369
3370 static struct tcp_ulp_ops smc_ulp_ops __read_mostly = {
3371         .name           = "smc",
3372         .owner          = THIS_MODULE,
3373         .init           = smc_ulp_init,
3374         .clone          = smc_ulp_clone,
3375 };
3376
3377 unsigned int smc_net_id;
3378
3379 static __net_init int smc_net_init(struct net *net)
3380 {
3381         int rc;
3382
3383         rc = smc_sysctl_net_init(net);
3384         if (rc)
3385                 return rc;
3386         return smc_pnet_net_init(net);
3387 }
3388
3389 static void __net_exit smc_net_exit(struct net *net)
3390 {
3391         smc_sysctl_net_exit(net);
3392         smc_pnet_net_exit(net);
3393 }
3394
3395 static __net_init int smc_net_stat_init(struct net *net)
3396 {
3397         return smc_stats_init(net);
3398 }
3399
3400 static void __net_exit smc_net_stat_exit(struct net *net)
3401 {
3402         smc_stats_exit(net);
3403 }
3404
3405 static struct pernet_operations smc_net_ops = {
3406         .init = smc_net_init,
3407         .exit = smc_net_exit,
3408         .id   = &smc_net_id,
3409         .size = sizeof(struct smc_net),
3410 };
3411
3412 static struct pernet_operations smc_net_stat_ops = {
3413         .init = smc_net_stat_init,
3414         .exit = smc_net_stat_exit,
3415 };
3416
3417 static int __init smc_init(void)
3418 {
3419         int rc;
3420
3421         rc = register_pernet_subsys(&smc_net_ops);
3422         if (rc)
3423                 return rc;
3424
3425         rc = register_pernet_subsys(&smc_net_stat_ops);
3426         if (rc)
3427                 goto out_pernet_subsys;
3428
3429         smc_ism_init();
3430         smc_clc_init();
3431
3432         rc = smc_nl_init();
3433         if (rc)
3434                 goto out_pernet_subsys_stat;
3435
3436         rc = smc_pnet_init();
3437         if (rc)
3438                 goto out_nl;
3439
3440         rc = -ENOMEM;
3441
3442         smc_tcp_ls_wq = alloc_workqueue("smc_tcp_ls_wq", 0, 0);
3443         if (!smc_tcp_ls_wq)
3444                 goto out_pnet;
3445
3446         smc_hs_wq = alloc_workqueue("smc_hs_wq", 0, 0);
3447         if (!smc_hs_wq)
3448                 goto out_alloc_tcp_ls_wq;
3449
3450         smc_close_wq = alloc_workqueue("smc_close_wq", 0, 0);
3451         if (!smc_close_wq)
3452                 goto out_alloc_hs_wq;
3453
3454         rc = smc_core_init();
3455         if (rc) {
3456                 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
3457                 goto out_alloc_wqs;
3458         }
3459
3460         rc = smc_llc_init();
3461         if (rc) {
3462                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
3463                 goto out_core;
3464         }
3465
3466         rc = smc_cdc_init();
3467         if (rc) {
3468                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
3469                 goto out_core;
3470         }
3471
3472         rc = proto_register(&smc_proto, 1);
3473         if (rc) {
3474                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
3475                 goto out_core;
3476         }
3477
3478         rc = proto_register(&smc_proto6, 1);
3479         if (rc) {
3480                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
3481                 goto out_proto;
3482         }
3483
3484         rc = sock_register(&smc_sock_family_ops);
3485         if (rc) {
3486                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
3487                 goto out_proto6;
3488         }
3489         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
3490         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
3491
3492         rc = smc_ib_register_client();
3493         if (rc) {
3494                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
3495                 goto out_sock;
3496         }
3497
3498         rc = tcp_register_ulp(&smc_ulp_ops);
3499         if (rc) {
3500                 pr_err("%s: tcp_ulp_register fails with %d\n", __func__, rc);
3501                 goto out_ib;
3502         }
3503
3504         static_branch_enable(&tcp_have_smc);
3505         return 0;
3506
3507 out_ib:
3508         smc_ib_unregister_client();
3509 out_sock:
3510         sock_unregister(PF_SMC);
3511 out_proto6:
3512         proto_unregister(&smc_proto6);
3513 out_proto:
3514         proto_unregister(&smc_proto);
3515 out_core:
3516         smc_core_exit();
3517 out_alloc_wqs:
3518         destroy_workqueue(smc_close_wq);
3519 out_alloc_hs_wq:
3520         destroy_workqueue(smc_hs_wq);
3521 out_alloc_tcp_ls_wq:
3522         destroy_workqueue(smc_tcp_ls_wq);
3523 out_pnet:
3524         smc_pnet_exit();
3525 out_nl:
3526         smc_nl_exit();
3527 out_pernet_subsys_stat:
3528         unregister_pernet_subsys(&smc_net_stat_ops);
3529 out_pernet_subsys:
3530         unregister_pernet_subsys(&smc_net_ops);
3531
3532         return rc;
3533 }
3534
3535 static void __exit smc_exit(void)
3536 {
3537         static_branch_disable(&tcp_have_smc);
3538         tcp_unregister_ulp(&smc_ulp_ops);
3539         sock_unregister(PF_SMC);
3540         smc_core_exit();
3541         smc_ib_unregister_client();
3542         destroy_workqueue(smc_close_wq);
3543         destroy_workqueue(smc_tcp_ls_wq);
3544         destroy_workqueue(smc_hs_wq);
3545         proto_unregister(&smc_proto6);
3546         proto_unregister(&smc_proto);
3547         smc_pnet_exit();
3548         smc_nl_exit();
3549         smc_clc_exit();
3550         unregister_pernet_subsys(&smc_net_stat_ops);
3551         unregister_pernet_subsys(&smc_net_ops);
3552         rcu_barrier();
3553 }
3554
3555 module_init(smc_init);
3556 module_exit(smc_exit);
3557
3558 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
3559 MODULE_DESCRIPTION("smc socket address family");
3560 MODULE_LICENSE("GPL");
3561 MODULE_ALIAS_NETPROTO(PF_SMC);
3562 MODULE_ALIAS_TCP_ULP("smc");
3563 MODULE_ALIAS_GENL_FAMILY(SMC_GENL_FAMILY_NAME);