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