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