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 (rc == -ECONNREFUSED)
1227                         smc->sk.sk_err = ECONNREFUSED;
1228                 else if (signal_pending(current))
1229                         smc->sk.sk_err = -sock_intr_errno(timeo);
1230                 sock_put(&smc->sk); /* passive closing */
1231                 goto out;
1232         }
1233
1234         rc = __smc_connect(smc);
1235         if (rc < 0)
1236                 smc->sk.sk_err = -rc;
1237
1238 out:
1239         if (!sock_flag(&smc->sk, SOCK_DEAD)) {
1240                 if (smc->sk.sk_err) {
1241                         smc->sk.sk_state_change(&smc->sk);
1242                 } else { /* allow polling before and after fallback decision */
1243                         smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
1244                         smc->sk.sk_write_space(&smc->sk);
1245                 }
1246         }
1247         release_sock(&smc->sk);
1248 }
1249
1250 static int smc_connect(struct socket *sock, struct sockaddr *addr,
1251                        int alen, int flags)
1252 {
1253         struct sock *sk = sock->sk;
1254         struct smc_sock *smc;
1255         int rc = -EINVAL;
1256
1257         smc = smc_sk(sk);
1258
1259         /* separate smc parameter checking to be safe */
1260         if (alen < sizeof(addr->sa_family))
1261                 goto out_err;
1262         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
1263                 goto out_err;
1264
1265         lock_sock(sk);
1266         switch (sk->sk_state) {
1267         default:
1268                 goto out;
1269         case SMC_ACTIVE:
1270                 rc = -EISCONN;
1271                 goto out;
1272         case SMC_INIT:
1273                 break;
1274         }
1275
1276         smc_copy_sock_settings_to_clc(smc);
1277         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1278         if (smc->connect_nonblock) {
1279                 rc = -EALREADY;
1280                 goto out;
1281         }
1282         rc = kernel_connect(smc->clcsock, addr, alen, flags);
1283         if (rc && rc != -EINPROGRESS)
1284                 goto out;
1285
1286         sock_hold(&smc->sk); /* sock put in passive closing */
1287         if (smc->use_fallback)
1288                 goto out;
1289         if (flags & O_NONBLOCK) {
1290                 if (queue_work(smc_hs_wq, &smc->connect_work))
1291                         smc->connect_nonblock = 1;
1292                 rc = -EINPROGRESS;
1293         } else {
1294                 rc = __smc_connect(smc);
1295                 if (rc < 0)
1296                         goto out;
1297                 else
1298                         rc = 0; /* success cases including fallback */
1299         }
1300
1301 out:
1302         release_sock(sk);
1303 out_err:
1304         return rc;
1305 }
1306
1307 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
1308 {
1309         struct socket *new_clcsock = NULL;
1310         struct sock *lsk = &lsmc->sk;
1311         struct sock *new_sk;
1312         int rc = -EINVAL;
1313
1314         release_sock(lsk);
1315         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
1316         if (!new_sk) {
1317                 rc = -ENOMEM;
1318                 lsk->sk_err = ENOMEM;
1319                 *new_smc = NULL;
1320                 lock_sock(lsk);
1321                 goto out;
1322         }
1323         *new_smc = smc_sk(new_sk);
1324
1325         mutex_lock(&lsmc->clcsock_release_lock);
1326         if (lsmc->clcsock)
1327                 rc = kernel_accept(lsmc->clcsock, &new_clcsock, SOCK_NONBLOCK);
1328         mutex_unlock(&lsmc->clcsock_release_lock);
1329         lock_sock(lsk);
1330         if  (rc < 0 && rc != -EAGAIN)
1331                 lsk->sk_err = -rc;
1332         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
1333                 new_sk->sk_prot->unhash(new_sk);
1334                 if (new_clcsock)
1335                         sock_release(new_clcsock);
1336                 new_sk->sk_state = SMC_CLOSED;
1337                 sock_set_flag(new_sk, SOCK_DEAD);
1338                 sock_put(new_sk); /* final */
1339                 *new_smc = NULL;
1340                 goto out;
1341         }
1342
1343         /* new clcsock has inherited the smc listen-specific sk_data_ready
1344          * function; switch it back to the original sk_data_ready function
1345          */
1346         new_clcsock->sk->sk_data_ready = lsmc->clcsk_data_ready;
1347         (*new_smc)->clcsock = new_clcsock;
1348 out:
1349         return rc;
1350 }
1351
1352 /* add a just created sock to the accept queue of the listen sock as
1353  * candidate for a following socket accept call from user space
1354  */
1355 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
1356 {
1357         struct smc_sock *par = smc_sk(parent);
1358
1359         sock_hold(sk); /* sock_put in smc_accept_unlink () */
1360         spin_lock(&par->accept_q_lock);
1361         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
1362         spin_unlock(&par->accept_q_lock);
1363         sk_acceptq_added(parent);
1364 }
1365
1366 /* remove a socket from the accept queue of its parental listening socket */
1367 static void smc_accept_unlink(struct sock *sk)
1368 {
1369         struct smc_sock *par = smc_sk(sk)->listen_smc;
1370
1371         spin_lock(&par->accept_q_lock);
1372         list_del_init(&smc_sk(sk)->accept_q);
1373         spin_unlock(&par->accept_q_lock);
1374         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
1375         sock_put(sk); /* sock_hold in smc_accept_enqueue */
1376 }
1377
1378 /* remove a sock from the accept queue to bind it to a new socket created
1379  * for a socket accept call from user space
1380  */
1381 struct sock *smc_accept_dequeue(struct sock *parent,
1382                                 struct socket *new_sock)
1383 {
1384         struct smc_sock *isk, *n;
1385         struct sock *new_sk;
1386
1387         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1388                 new_sk = (struct sock *)isk;
1389
1390                 smc_accept_unlink(new_sk);
1391                 if (new_sk->sk_state == SMC_CLOSED) {
1392                         new_sk->sk_prot->unhash(new_sk);
1393                         if (isk->clcsock) {
1394                                 sock_release(isk->clcsock);
1395                                 isk->clcsock = NULL;
1396                         }
1397                         sock_put(new_sk); /* final */
1398                         continue;
1399                 }
1400                 if (new_sock) {
1401                         sock_graft(new_sk, new_sock);
1402                         if (isk->use_fallback) {
1403                                 smc_sk(new_sk)->clcsock->file = new_sock->file;
1404                                 isk->clcsock->file->private_data = isk->clcsock;
1405                         }
1406                 }
1407                 return new_sk;
1408         }
1409         return NULL;
1410 }
1411
1412 /* clean up for a created but never accepted sock */
1413 void smc_close_non_accepted(struct sock *sk)
1414 {
1415         struct smc_sock *smc = smc_sk(sk);
1416
1417         sock_hold(sk); /* sock_put below */
1418         lock_sock(sk);
1419         if (!sk->sk_lingertime)
1420                 /* wait for peer closing */
1421                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
1422         __smc_release(smc);
1423         release_sock(sk);
1424         sock_put(sk); /* sock_hold above */
1425         sock_put(sk); /* final sock_put */
1426 }
1427
1428 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1429 {
1430         struct smc_link *link = smc->conn.lnk;
1431         struct smc_llc_qentry *qentry;
1432         int rc;
1433
1434         if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
1435                 return SMC_CLC_DECL_ERR_REGRMB;
1436
1437         /* send CONFIRM LINK request to client over the RoCE fabric */
1438         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1439         if (rc < 0)
1440                 return SMC_CLC_DECL_TIMEOUT_CL;
1441
1442         /* receive CONFIRM LINK response from client over the RoCE fabric */
1443         qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1444                               SMC_LLC_CONFIRM_LINK);
1445         if (!qentry) {
1446                 struct smc_clc_msg_decline dclc;
1447
1448                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1449                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1450                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1451         }
1452         smc_llc_save_peer_uid(qentry);
1453         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1454         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1455         if (rc)
1456                 return SMC_CLC_DECL_RMBE_EC;
1457
1458         /* confirm_rkey is implicit on 1st contact */
1459         smc->conn.rmb_desc->is_conf_rkey = true;
1460
1461         smc_llc_link_active(link);
1462         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1463
1464         /* initial contact - try to establish second link */
1465         smc_llc_srv_add_link(link);
1466         return 0;
1467 }
1468
1469 /* listen worker: finish */
1470 static void smc_listen_out(struct smc_sock *new_smc)
1471 {
1472         struct smc_sock *lsmc = new_smc->listen_smc;
1473         struct sock *newsmcsk = &new_smc->sk;
1474
1475         if (lsmc->sk.sk_state == SMC_LISTEN) {
1476                 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1477                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1478                 release_sock(&lsmc->sk);
1479         } else { /* no longer listening */
1480                 smc_close_non_accepted(newsmcsk);
1481         }
1482
1483         /* Wake up accept */
1484         lsmc->sk.sk_data_ready(&lsmc->sk);
1485         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1486 }
1487
1488 /* listen worker: finish in state connected */
1489 static void smc_listen_out_connected(struct smc_sock *new_smc)
1490 {
1491         struct sock *newsmcsk = &new_smc->sk;
1492
1493         sk_refcnt_debug_inc(newsmcsk);
1494         if (newsmcsk->sk_state == SMC_INIT)
1495                 newsmcsk->sk_state = SMC_ACTIVE;
1496
1497         smc_listen_out(new_smc);
1498 }
1499
1500 /* listen worker: finish in error state */
1501 static void smc_listen_out_err(struct smc_sock *new_smc)
1502 {
1503         struct sock *newsmcsk = &new_smc->sk;
1504         struct net *net = sock_net(newsmcsk);
1505
1506         this_cpu_inc(net->smc.smc_stats->srv_hshake_err_cnt);
1507         if (newsmcsk->sk_state == SMC_INIT)
1508                 sock_put(&new_smc->sk); /* passive closing */
1509         newsmcsk->sk_state = SMC_CLOSED;
1510
1511         smc_listen_out(new_smc);
1512 }
1513
1514 /* listen worker: decline and fall back if possible */
1515 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1516                                int local_first, u8 version)
1517 {
1518         /* RDMA setup failed, switch back to TCP */
1519         smc_conn_abort(new_smc, local_first);
1520         if (reason_code < 0 ||
1521             smc_switch_to_fallback(new_smc, reason_code)) {
1522                 /* error, no fallback possible */
1523                 smc_listen_out_err(new_smc);
1524                 return;
1525         }
1526         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1527                 if (smc_clc_send_decline(new_smc, reason_code, version) < 0) {
1528                         smc_listen_out_err(new_smc);
1529                         return;
1530                 }
1531         }
1532         smc_listen_out_connected(new_smc);
1533 }
1534
1535 /* listen worker: version checking */
1536 static int smc_listen_v2_check(struct smc_sock *new_smc,
1537                                struct smc_clc_msg_proposal *pclc,
1538                                struct smc_init_info *ini)
1539 {
1540         struct smc_clc_smcd_v2_extension *pclc_smcd_v2_ext;
1541         struct smc_clc_v2_extension *pclc_v2_ext;
1542         int rc = SMC_CLC_DECL_PEERNOSMC;
1543
1544         ini->smc_type_v1 = pclc->hdr.typev1;
1545         ini->smc_type_v2 = pclc->hdr.typev2;
1546         ini->smcd_version = ini->smc_type_v1 != SMC_TYPE_N ? SMC_V1 : 0;
1547         if (pclc->hdr.version > SMC_V1)
1548                 ini->smcd_version |=
1549                                 ini->smc_type_v2 != SMC_TYPE_N ? SMC_V2 : 0;
1550         if (!(ini->smcd_version & SMC_V2)) {
1551                 rc = SMC_CLC_DECL_PEERNOSMC;
1552                 goto out;
1553         }
1554         if (!smc_ism_is_v2_capable()) {
1555                 ini->smcd_version &= ~SMC_V2;
1556                 rc = SMC_CLC_DECL_NOISM2SUPP;
1557                 goto out;
1558         }
1559         pclc_v2_ext = smc_get_clc_v2_ext(pclc);
1560         if (!pclc_v2_ext) {
1561                 ini->smcd_version &= ~SMC_V2;
1562                 rc = SMC_CLC_DECL_NOV2EXT;
1563                 goto out;
1564         }
1565         pclc_smcd_v2_ext = smc_get_clc_smcd_v2_ext(pclc_v2_ext);
1566         if (!pclc_smcd_v2_ext) {
1567                 ini->smcd_version &= ~SMC_V2;
1568                 rc = SMC_CLC_DECL_NOV2DEXT;
1569         }
1570
1571 out:
1572         if (!ini->smcd_version)
1573                 return rc;
1574
1575         return 0;
1576 }
1577
1578 /* listen worker: check prefixes */
1579 static int smc_listen_prfx_check(struct smc_sock *new_smc,
1580                                  struct smc_clc_msg_proposal *pclc)
1581 {
1582         struct smc_clc_msg_proposal_prefix *pclc_prfx;
1583         struct socket *newclcsock = new_smc->clcsock;
1584
1585         if (pclc->hdr.typev1 == SMC_TYPE_N)
1586                 return 0;
1587         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1588         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1589                 return SMC_CLC_DECL_DIFFPREFIX;
1590
1591         return 0;
1592 }
1593
1594 /* listen worker: initialize connection and buffers */
1595 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1596                                 struct smc_init_info *ini)
1597 {
1598         int rc;
1599
1600         /* allocate connection / link group */
1601         rc = smc_conn_create(new_smc, ini);
1602         if (rc)
1603                 return rc;
1604
1605         /* create send buffer and rmb */
1606         if (smc_buf_create(new_smc, false))
1607                 return SMC_CLC_DECL_MEM;
1608
1609         return 0;
1610 }
1611
1612 /* listen worker: initialize connection and buffers for SMC-D */
1613 static int smc_listen_ism_init(struct smc_sock *new_smc,
1614                                struct smc_init_info *ini)
1615 {
1616         int rc;
1617
1618         rc = smc_conn_create(new_smc, ini);
1619         if (rc)
1620                 return rc;
1621
1622         /* Create send and receive buffers */
1623         rc = smc_buf_create(new_smc, true);
1624         if (rc) {
1625                 smc_conn_abort(new_smc, ini->first_contact_local);
1626                 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
1627                                          SMC_CLC_DECL_MEM;
1628         }
1629
1630         return 0;
1631 }
1632
1633 static bool smc_is_already_selected(struct smcd_dev *smcd,
1634                                     struct smc_init_info *ini,
1635                                     int matches)
1636 {
1637         int i;
1638
1639         for (i = 0; i < matches; i++)
1640                 if (smcd == ini->ism_dev[i])
1641                         return true;
1642
1643         return false;
1644 }
1645
1646 /* check for ISM devices matching proposed ISM devices */
1647 static void smc_check_ism_v2_match(struct smc_init_info *ini,
1648                                    u16 proposed_chid, u64 proposed_gid,
1649                                    unsigned int *matches)
1650 {
1651         struct smcd_dev *smcd;
1652
1653         list_for_each_entry(smcd, &smcd_dev_list.list, list) {
1654                 if (smcd->going_away)
1655                         continue;
1656                 if (smc_is_already_selected(smcd, ini, *matches))
1657                         continue;
1658                 if (smc_ism_get_chid(smcd) == proposed_chid &&
1659                     !smc_ism_cantalk(proposed_gid, ISM_RESERVED_VLANID, smcd)) {
1660                         ini->ism_peer_gid[*matches] = proposed_gid;
1661                         ini->ism_dev[*matches] = smcd;
1662                         (*matches)++;
1663                         break;
1664                 }
1665         }
1666 }
1667
1668 static void smc_find_ism_store_rc(u32 rc, struct smc_init_info *ini)
1669 {
1670         if (!ini->rc)
1671                 ini->rc = rc;
1672 }
1673
1674 static void smc_find_ism_v2_device_serv(struct smc_sock *new_smc,
1675                                         struct smc_clc_msg_proposal *pclc,
1676                                         struct smc_init_info *ini)
1677 {
1678         struct smc_clc_smcd_v2_extension *smcd_v2_ext;
1679         struct smc_clc_v2_extension *smc_v2_ext;
1680         struct smc_clc_msg_smcd *pclc_smcd;
1681         unsigned int matches = 0;
1682         u8 smcd_version;
1683         u8 *eid = NULL;
1684         int i, rc;
1685
1686         if (!(ini->smcd_version & SMC_V2) || !smcd_indicated(ini->smc_type_v2))
1687                 goto not_found;
1688
1689         pclc_smcd = smc_get_clc_msg_smcd(pclc);
1690         smc_v2_ext = smc_get_clc_v2_ext(pclc);
1691         smcd_v2_ext = smc_get_clc_smcd_v2_ext(smc_v2_ext);
1692         if (!smcd_v2_ext ||
1693             !smc_v2_ext->hdr.flag.seid) { /* no system EID support for SMCD */
1694                 smc_find_ism_store_rc(SMC_CLC_DECL_NOSEID, ini);
1695                 goto not_found;
1696         }
1697
1698         mutex_lock(&smcd_dev_list.mutex);
1699         if (pclc_smcd->ism.chid)
1700                 /* check for ISM device matching proposed native ISM device */
1701                 smc_check_ism_v2_match(ini, ntohs(pclc_smcd->ism.chid),
1702                                        ntohll(pclc_smcd->ism.gid), &matches);
1703         for (i = 1; i <= smc_v2_ext->hdr.ism_gid_cnt; i++) {
1704                 /* check for ISM devices matching proposed non-native ISM
1705                  * devices
1706                  */
1707                 smc_check_ism_v2_match(ini,
1708                                        ntohs(smcd_v2_ext->gidchid[i - 1].chid),
1709                                        ntohll(smcd_v2_ext->gidchid[i - 1].gid),
1710                                        &matches);
1711         }
1712         mutex_unlock(&smcd_dev_list.mutex);
1713
1714         if (ini->ism_dev[0]) {
1715                 smc_ism_get_system_eid(ini->ism_dev[0], &eid);
1716                 if (memcmp(eid, smcd_v2_ext->system_eid, SMC_MAX_EID_LEN))
1717                         goto not_found;
1718         } else {
1719                 goto not_found;
1720         }
1721
1722         /* separate - outside the smcd_dev_list.lock */
1723         smcd_version = ini->smcd_version;
1724         for (i = 0; i < matches; i++) {
1725                 ini->smcd_version = SMC_V2;
1726                 ini->is_smcd = true;
1727                 ini->ism_selected = i;
1728                 rc = smc_listen_ism_init(new_smc, ini);
1729                 if (rc) {
1730                         smc_find_ism_store_rc(rc, ini);
1731                         /* try next active ISM device */
1732                         continue;
1733                 }
1734                 return; /* matching and usable V2 ISM device found */
1735         }
1736         /* no V2 ISM device could be initialized */
1737         ini->smcd_version = smcd_version;       /* restore original value */
1738
1739 not_found:
1740         ini->smcd_version &= ~SMC_V2;
1741         ini->ism_dev[0] = NULL;
1742         ini->is_smcd = false;
1743 }
1744
1745 static void smc_find_ism_v1_device_serv(struct smc_sock *new_smc,
1746                                         struct smc_clc_msg_proposal *pclc,
1747                                         struct smc_init_info *ini)
1748 {
1749         struct smc_clc_msg_smcd *pclc_smcd = smc_get_clc_msg_smcd(pclc);
1750         int rc = 0;
1751
1752         /* check if ISM V1 is available */
1753         if (!(ini->smcd_version & SMC_V1) || !smcd_indicated(ini->smc_type_v1))
1754                 goto not_found;
1755         ini->is_smcd = true; /* prepare ISM check */
1756         ini->ism_peer_gid[0] = ntohll(pclc_smcd->ism.gid);
1757         rc = smc_find_ism_device(new_smc, ini);
1758         if (rc)
1759                 goto not_found;
1760         ini->ism_selected = 0;
1761         rc = smc_listen_ism_init(new_smc, ini);
1762         if (!rc)
1763                 return;         /* V1 ISM device found */
1764
1765 not_found:
1766         smc_find_ism_store_rc(rc, ini);
1767         ini->ism_dev[0] = NULL;
1768         ini->is_smcd = false;
1769 }
1770
1771 /* listen worker: register buffers */
1772 static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
1773 {
1774         struct smc_connection *conn = &new_smc->conn;
1775
1776         if (!local_first) {
1777                 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
1778                         return SMC_CLC_DECL_ERR_REGRMB;
1779         }
1780         smc_rmb_sync_sg_for_device(&new_smc->conn);
1781
1782         return 0;
1783 }
1784
1785 static int smc_find_rdma_v1_device_serv(struct smc_sock *new_smc,
1786                                         struct smc_clc_msg_proposal *pclc,
1787                                         struct smc_init_info *ini)
1788 {
1789         int rc;
1790
1791         if (!smcr_indicated(ini->smc_type_v1))
1792                 return SMC_CLC_DECL_NOSMCDEV;
1793
1794         /* prepare RDMA check */
1795         ini->ib_lcl = &pclc->lcl;
1796         rc = smc_find_rdma_device(new_smc, ini);
1797         if (rc) {
1798                 /* no RDMA device found */
1799                 if (ini->smc_type_v1 == SMC_TYPE_B)
1800                         /* neither ISM nor RDMA device found */
1801                         rc = SMC_CLC_DECL_NOSMCDEV;
1802                 return rc;
1803         }
1804         rc = smc_listen_rdma_init(new_smc, ini);
1805         if (rc)
1806                 return rc;
1807         return smc_listen_rdma_reg(new_smc, ini->first_contact_local);
1808 }
1809
1810 /* determine the local device matching to proposal */
1811 static int smc_listen_find_device(struct smc_sock *new_smc,
1812                                   struct smc_clc_msg_proposal *pclc,
1813                                   struct smc_init_info *ini)
1814 {
1815         int rc;
1816
1817         /* check for ISM device matching V2 proposed device */
1818         smc_find_ism_v2_device_serv(new_smc, pclc, ini);
1819         if (ini->ism_dev[0])
1820                 return 0;
1821
1822         if (!(ini->smcd_version & SMC_V1))
1823                 return ini->rc ?: SMC_CLC_DECL_NOSMCD2DEV;
1824
1825         /* check for matching IP prefix and subnet length */
1826         rc = smc_listen_prfx_check(new_smc, pclc);
1827         if (rc)
1828                 return ini->rc ?: rc;
1829
1830         /* get vlan id from IP device */
1831         if (smc_vlan_by_tcpsk(new_smc->clcsock, ini))
1832                 return ini->rc ?: SMC_CLC_DECL_GETVLANERR;
1833
1834         /* check for ISM device matching V1 proposed device */
1835         smc_find_ism_v1_device_serv(new_smc, pclc, ini);
1836         if (ini->ism_dev[0])
1837                 return 0;
1838
1839         if (pclc->hdr.typev1 == SMC_TYPE_D)
1840                 /* skip RDMA and decline */
1841                 return ini->rc ?: SMC_CLC_DECL_NOSMCDDEV;
1842
1843         /* check if RDMA is available */
1844         rc = smc_find_rdma_v1_device_serv(new_smc, pclc, ini);
1845         smc_find_ism_store_rc(rc, ini);
1846
1847         return (!rc) ? 0 : ini->rc;
1848 }
1849
1850 /* listen worker: finish RDMA setup */
1851 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1852                                   struct smc_clc_msg_accept_confirm *cclc,
1853                                   bool local_first)
1854 {
1855         struct smc_link *link = new_smc->conn.lnk;
1856         int reason_code = 0;
1857
1858         if (local_first)
1859                 smc_link_save_peer_info(link, cclc);
1860
1861         if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc))
1862                 return SMC_CLC_DECL_ERR_RTOK;
1863
1864         if (local_first) {
1865                 if (smc_ib_ready_link(link))
1866                         return SMC_CLC_DECL_ERR_RDYLNK;
1867                 /* QP confirmation over RoCE fabric */
1868                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1869                 reason_code = smcr_serv_conf_first_link(new_smc);
1870                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1871         }
1872         return reason_code;
1873 }
1874
1875 /* setup for connection of server */
1876 static void smc_listen_work(struct work_struct *work)
1877 {
1878         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1879                                                 smc_listen_work);
1880         u8 version = smc_ism_is_v2_capable() ? SMC_V2 : SMC_V1;
1881         struct socket *newclcsock = new_smc->clcsock;
1882         struct smc_clc_msg_accept_confirm *cclc;
1883         struct smc_clc_msg_proposal_area *buf;
1884         struct smc_clc_msg_proposal *pclc;
1885         struct smc_init_info *ini = NULL;
1886         int rc = 0;
1887
1888         if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1889                 return smc_listen_out_err(new_smc);
1890
1891         if (new_smc->use_fallback) {
1892                 smc_listen_out_connected(new_smc);
1893                 return;
1894         }
1895
1896         /* check if peer is smc capable */
1897         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1898                 rc = smc_switch_to_fallback(new_smc, SMC_CLC_DECL_PEERNOSMC);
1899                 if (rc)
1900                         smc_listen_out_err(new_smc);
1901                 else
1902                         smc_listen_out_connected(new_smc);
1903                 return;
1904         }
1905
1906         /* do inband token exchange -
1907          * wait for and receive SMC Proposal CLC message
1908          */
1909         buf = kzalloc(sizeof(*buf), GFP_KERNEL);
1910         if (!buf) {
1911                 rc = SMC_CLC_DECL_MEM;
1912                 goto out_decl;
1913         }
1914         pclc = (struct smc_clc_msg_proposal *)buf;
1915         rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
1916                               SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1917         if (rc)
1918                 goto out_decl;
1919         version = pclc->hdr.version == SMC_V1 ? SMC_V1 : version;
1920
1921         /* IPSec connections opt out of SMC optimizations */
1922         if (using_ipsec(new_smc)) {
1923                 rc = SMC_CLC_DECL_IPSEC;
1924                 goto out_decl;
1925         }
1926
1927         ini = kzalloc(sizeof(*ini), GFP_KERNEL);
1928         if (!ini) {
1929                 rc = SMC_CLC_DECL_MEM;
1930                 goto out_decl;
1931         }
1932
1933         /* initial version checking */
1934         rc = smc_listen_v2_check(new_smc, pclc, ini);
1935         if (rc)
1936                 goto out_decl;
1937
1938         mutex_lock(&smc_server_lgr_pending);
1939         smc_close_init(new_smc);
1940         smc_rx_init(new_smc);
1941         smc_tx_init(new_smc);
1942
1943         /* determine ISM or RoCE device used for connection */
1944         rc = smc_listen_find_device(new_smc, pclc, ini);
1945         if (rc)
1946                 goto out_unlock;
1947
1948         /* send SMC Accept CLC message */
1949         rc = smc_clc_send_accept(new_smc, ini->first_contact_local,
1950                                  ini->smcd_version == SMC_V2 ? SMC_V2 : SMC_V1);
1951         if (rc)
1952                 goto out_unlock;
1953
1954         /* SMC-D does not need this lock any more */
1955         if (ini->is_smcd)
1956                 mutex_unlock(&smc_server_lgr_pending);
1957
1958         /* receive SMC Confirm CLC message */
1959         memset(buf, 0, sizeof(*buf));
1960         cclc = (struct smc_clc_msg_accept_confirm *)buf;
1961         rc = smc_clc_wait_msg(new_smc, cclc, sizeof(*buf),
1962                               SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1963         if (rc) {
1964                 if (!ini->is_smcd)
1965                         goto out_unlock;
1966                 goto out_decl;
1967         }
1968
1969         /* finish worker */
1970         if (!ini->is_smcd) {
1971                 rc = smc_listen_rdma_finish(new_smc, cclc,
1972                                             ini->first_contact_local);
1973                 if (rc)
1974                         goto out_unlock;
1975                 mutex_unlock(&smc_server_lgr_pending);
1976         }
1977         smc_conn_save_peer_info(new_smc, cclc);
1978         smc_listen_out_connected(new_smc);
1979         SMC_STAT_SERV_SUCC_INC(sock_net(newclcsock->sk), ini);
1980         goto out_free;
1981
1982 out_unlock:
1983         mutex_unlock(&smc_server_lgr_pending);
1984 out_decl:
1985         smc_listen_decline(new_smc, rc, ini ? ini->first_contact_local : 0,
1986                            version);
1987 out_free:
1988         kfree(ini);
1989         kfree(buf);
1990 }
1991
1992 static void smc_tcp_listen_work(struct work_struct *work)
1993 {
1994         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1995                                              tcp_listen_work);
1996         struct sock *lsk = &lsmc->sk;
1997         struct smc_sock *new_smc;
1998         int rc = 0;
1999
2000         lock_sock(lsk);
2001         while (lsk->sk_state == SMC_LISTEN) {
2002                 rc = smc_clcsock_accept(lsmc, &new_smc);
2003                 if (rc) /* clcsock accept queue empty or error */
2004                         goto out;
2005                 if (!new_smc)
2006                         continue;
2007
2008                 new_smc->listen_smc = lsmc;
2009                 new_smc->use_fallback = lsmc->use_fallback;
2010                 new_smc->fallback_rsn = lsmc->fallback_rsn;
2011                 sock_hold(lsk); /* sock_put in smc_listen_work */
2012                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
2013                 smc_copy_sock_settings_to_smc(new_smc);
2014                 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
2015                 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
2016                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
2017                 if (!queue_work(smc_hs_wq, &new_smc->smc_listen_work))
2018                         sock_put(&new_smc->sk);
2019         }
2020
2021 out:
2022         release_sock(lsk);
2023         sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
2024 }
2025
2026 static void smc_clcsock_data_ready(struct sock *listen_clcsock)
2027 {
2028         struct smc_sock *lsmc =
2029                 smc_clcsock_user_data(listen_clcsock);
2030
2031         if (!lsmc)
2032                 return;
2033         lsmc->clcsk_data_ready(listen_clcsock);
2034         if (lsmc->sk.sk_state == SMC_LISTEN) {
2035                 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
2036                 if (!queue_work(smc_hs_wq, &lsmc->tcp_listen_work))
2037                         sock_put(&lsmc->sk);
2038         }
2039 }
2040
2041 static int smc_listen(struct socket *sock, int backlog)
2042 {
2043         struct sock *sk = sock->sk;
2044         struct smc_sock *smc;
2045         int rc;
2046
2047         smc = smc_sk(sk);
2048         lock_sock(sk);
2049
2050         rc = -EINVAL;
2051         if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
2052             smc->connect_nonblock)
2053                 goto out;
2054
2055         rc = 0;
2056         if (sk->sk_state == SMC_LISTEN) {
2057                 sk->sk_max_ack_backlog = backlog;
2058                 goto out;
2059         }
2060         /* some socket options are handled in core, so we could not apply
2061          * them to the clc socket -- copy smc socket options to clc socket
2062          */
2063         smc_copy_sock_settings_to_clc(smc);
2064         if (!smc->use_fallback)
2065                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
2066
2067         /* save original sk_data_ready function and establish
2068          * smc-specific sk_data_ready function
2069          */
2070         smc->clcsk_data_ready = smc->clcsock->sk->sk_data_ready;
2071         smc->clcsock->sk->sk_data_ready = smc_clcsock_data_ready;
2072         smc->clcsock->sk->sk_user_data =
2073                 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
2074         rc = kernel_listen(smc->clcsock, backlog);
2075         if (rc) {
2076                 smc->clcsock->sk->sk_data_ready = smc->clcsk_data_ready;
2077                 goto out;
2078         }
2079         sk->sk_max_ack_backlog = backlog;
2080         sk->sk_ack_backlog = 0;
2081         sk->sk_state = SMC_LISTEN;
2082
2083 out:
2084         release_sock(sk);
2085         return rc;
2086 }
2087
2088 static int smc_accept(struct socket *sock, struct socket *new_sock,
2089                       int flags, bool kern)
2090 {
2091         struct sock *sk = sock->sk, *nsk;
2092         DECLARE_WAITQUEUE(wait, current);
2093         struct smc_sock *lsmc;
2094         long timeo;
2095         int rc = 0;
2096
2097         lsmc = smc_sk(sk);
2098         sock_hold(sk); /* sock_put below */
2099         lock_sock(sk);
2100
2101         if (lsmc->sk.sk_state != SMC_LISTEN) {
2102                 rc = -EINVAL;
2103                 release_sock(sk);
2104                 goto out;
2105         }
2106
2107         /* Wait for an incoming connection */
2108         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
2109         add_wait_queue_exclusive(sk_sleep(sk), &wait);
2110         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
2111                 set_current_state(TASK_INTERRUPTIBLE);
2112                 if (!timeo) {
2113                         rc = -EAGAIN;
2114                         break;
2115                 }
2116                 release_sock(sk);
2117                 timeo = schedule_timeout(timeo);
2118                 /* wakeup by sk_data_ready in smc_listen_work() */
2119                 sched_annotate_sleep();
2120                 lock_sock(sk);
2121                 if (signal_pending(current)) {
2122                         rc = sock_intr_errno(timeo);
2123                         break;
2124                 }
2125         }
2126         set_current_state(TASK_RUNNING);
2127         remove_wait_queue(sk_sleep(sk), &wait);
2128
2129         if (!rc)
2130                 rc = sock_error(nsk);
2131         release_sock(sk);
2132         if (rc)
2133                 goto out;
2134
2135         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
2136                 /* wait till data arrives on the socket */
2137                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
2138                                                                 MSEC_PER_SEC);
2139                 if (smc_sk(nsk)->use_fallback) {
2140                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
2141
2142                         lock_sock(clcsk);
2143                         if (skb_queue_empty(&clcsk->sk_receive_queue))
2144                                 sk_wait_data(clcsk, &timeo, NULL);
2145                         release_sock(clcsk);
2146                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
2147                         lock_sock(nsk);
2148                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
2149                         release_sock(nsk);
2150                 }
2151         }
2152
2153 out:
2154         sock_put(sk); /* sock_hold above */
2155         return rc;
2156 }
2157
2158 static int smc_getname(struct socket *sock, struct sockaddr *addr,
2159                        int peer)
2160 {
2161         struct smc_sock *smc;
2162
2163         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
2164             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
2165                 return -ENOTCONN;
2166
2167         smc = smc_sk(sock->sk);
2168
2169         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
2170 }
2171
2172 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
2173 {
2174         struct sock *sk = sock->sk;
2175         struct smc_sock *smc;
2176         int rc = -EPIPE;
2177
2178         smc = smc_sk(sk);
2179         lock_sock(sk);
2180         if ((sk->sk_state != SMC_ACTIVE) &&
2181             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2182             (sk->sk_state != SMC_INIT))
2183                 goto out;
2184
2185         if (msg->msg_flags & MSG_FASTOPEN) {
2186                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2187                         rc = smc_switch_to_fallback(smc, SMC_CLC_DECL_OPTUNSUPP);
2188                         if (rc)
2189                                 goto out;
2190                 } else {
2191                         rc = -EINVAL;
2192                         goto out;
2193                 }
2194         }
2195
2196         if (smc->use_fallback) {
2197                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
2198         } else {
2199                 rc = smc_tx_sendmsg(smc, msg, len);
2200                 SMC_STAT_TX_PAYLOAD(smc, len, rc);
2201         }
2202 out:
2203         release_sock(sk);
2204         return rc;
2205 }
2206
2207 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
2208                        int flags)
2209 {
2210         struct sock *sk = sock->sk;
2211         struct smc_sock *smc;
2212         int rc = -ENOTCONN;
2213
2214         smc = smc_sk(sk);
2215         lock_sock(sk);
2216         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2217                 /* socket was connected before, no more data to read */
2218                 rc = 0;
2219                 goto out;
2220         }
2221         if ((sk->sk_state == SMC_INIT) ||
2222             (sk->sk_state == SMC_LISTEN) ||
2223             (sk->sk_state == SMC_CLOSED))
2224                 goto out;
2225
2226         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2227                 rc = 0;
2228                 goto out;
2229         }
2230
2231         if (smc->use_fallback) {
2232                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
2233         } else {
2234                 msg->msg_namelen = 0;
2235                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
2236                 SMC_STAT_RX_PAYLOAD(smc, rc, rc);
2237         }
2238
2239 out:
2240         release_sock(sk);
2241         return rc;
2242 }
2243
2244 static __poll_t smc_accept_poll(struct sock *parent)
2245 {
2246         struct smc_sock *isk = smc_sk(parent);
2247         __poll_t mask = 0;
2248
2249         spin_lock(&isk->accept_q_lock);
2250         if (!list_empty(&isk->accept_q))
2251                 mask = EPOLLIN | EPOLLRDNORM;
2252         spin_unlock(&isk->accept_q_lock);
2253
2254         return mask;
2255 }
2256
2257 static __poll_t smc_poll(struct file *file, struct socket *sock,
2258                              poll_table *wait)
2259 {
2260         struct sock *sk = sock->sk;
2261         struct smc_sock *smc;
2262         __poll_t mask = 0;
2263
2264         if (!sk)
2265                 return EPOLLNVAL;
2266
2267         smc = smc_sk(sock->sk);
2268         if (smc->use_fallback) {
2269                 /* delegate to CLC child sock */
2270                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
2271                 sk->sk_err = smc->clcsock->sk->sk_err;
2272         } else {
2273                 if (sk->sk_state != SMC_CLOSED)
2274                         sock_poll_wait(file, sock, wait);
2275                 if (sk->sk_err)
2276                         mask |= EPOLLERR;
2277                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
2278                     (sk->sk_state == SMC_CLOSED))
2279                         mask |= EPOLLHUP;
2280                 if (sk->sk_state == SMC_LISTEN) {
2281                         /* woken up by sk_data_ready in smc_listen_work() */
2282                         mask |= smc_accept_poll(sk);
2283                 } else if (smc->use_fallback) { /* as result of connect_work()*/
2284                         mask |= smc->clcsock->ops->poll(file, smc->clcsock,
2285                                                            wait);
2286                         sk->sk_err = smc->clcsock->sk->sk_err;
2287                 } else {
2288                         if ((sk->sk_state != SMC_INIT &&
2289                              atomic_read(&smc->conn.sndbuf_space)) ||
2290                             sk->sk_shutdown & SEND_SHUTDOWN) {
2291                                 mask |= EPOLLOUT | EPOLLWRNORM;
2292                         } else {
2293                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
2294                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
2295                         }
2296                         if (atomic_read(&smc->conn.bytes_to_rcv))
2297                                 mask |= EPOLLIN | EPOLLRDNORM;
2298                         if (sk->sk_shutdown & RCV_SHUTDOWN)
2299                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
2300                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
2301                                 mask |= EPOLLIN;
2302                         if (smc->conn.urg_state == SMC_URG_VALID)
2303                                 mask |= EPOLLPRI;
2304                 }
2305         }
2306
2307         return mask;
2308 }
2309
2310 static int smc_shutdown(struct socket *sock, int how)
2311 {
2312         struct sock *sk = sock->sk;
2313         bool do_shutdown = true;
2314         struct smc_sock *smc;
2315         int rc = -EINVAL;
2316         int old_state;
2317         int rc1 = 0;
2318
2319         smc = smc_sk(sk);
2320
2321         if ((how < SHUT_RD) || (how > SHUT_RDWR))
2322                 return rc;
2323
2324         lock_sock(sk);
2325
2326         rc = -ENOTCONN;
2327         if ((sk->sk_state != SMC_ACTIVE) &&
2328             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
2329             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
2330             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
2331             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
2332             (sk->sk_state != SMC_APPFINCLOSEWAIT))
2333                 goto out;
2334         if (smc->use_fallback) {
2335                 rc = kernel_sock_shutdown(smc->clcsock, how);
2336                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
2337                 if (sk->sk_shutdown == SHUTDOWN_MASK) {
2338                         sk->sk_state = SMC_CLOSED;
2339                         sock_put(sk);
2340                 }
2341                 goto out;
2342         }
2343         switch (how) {
2344         case SHUT_RDWR:         /* shutdown in both directions */
2345                 old_state = sk->sk_state;
2346                 rc = smc_close_active(smc);
2347                 if (old_state == SMC_ACTIVE &&
2348                     sk->sk_state == SMC_PEERCLOSEWAIT1)
2349                         do_shutdown = false;
2350                 break;
2351         case SHUT_WR:
2352                 rc = smc_close_shutdown_write(smc);
2353                 break;
2354         case SHUT_RD:
2355                 rc = 0;
2356                 /* nothing more to do because peer is not involved */
2357                 break;
2358         }
2359         if (do_shutdown && smc->clcsock)
2360                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
2361         /* map sock_shutdown_cmd constants to sk_shutdown value range */
2362         sk->sk_shutdown |= how + 1;
2363
2364 out:
2365         release_sock(sk);
2366         return rc ? rc : rc1;
2367 }
2368
2369 static int smc_setsockopt(struct socket *sock, int level, int optname,
2370                           sockptr_t optval, unsigned int optlen)
2371 {
2372         struct sock *sk = sock->sk;
2373         struct smc_sock *smc;
2374         int val, rc;
2375
2376         if (level == SOL_TCP && optname == TCP_ULP)
2377                 return -EOPNOTSUPP;
2378
2379         smc = smc_sk(sk);
2380
2381         /* generic setsockopts reaching us here always apply to the
2382          * CLC socket
2383          */
2384         mutex_lock(&smc->clcsock_release_lock);
2385         if (!smc->clcsock) {
2386                 mutex_unlock(&smc->clcsock_release_lock);
2387                 return -EBADF;
2388         }
2389         if (unlikely(!smc->clcsock->ops->setsockopt))
2390                 rc = -EOPNOTSUPP;
2391         else
2392                 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
2393                                                    optval, optlen);
2394         if (smc->clcsock->sk->sk_err) {
2395                 sk->sk_err = smc->clcsock->sk->sk_err;
2396                 sk_error_report(sk);
2397         }
2398         mutex_unlock(&smc->clcsock_release_lock);
2399
2400         if (optlen < sizeof(int))
2401                 return -EINVAL;
2402         if (copy_from_sockptr(&val, optval, sizeof(int)))
2403                 return -EFAULT;
2404
2405         lock_sock(sk);
2406         if (rc || smc->use_fallback)
2407                 goto out;
2408         switch (optname) {
2409         case TCP_FASTOPEN:
2410         case TCP_FASTOPEN_CONNECT:
2411         case TCP_FASTOPEN_KEY:
2412         case TCP_FASTOPEN_NO_COOKIE:
2413                 /* option not supported by SMC */
2414                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
2415                         rc = smc_switch_to_fallback(smc, SMC_CLC_DECL_OPTUNSUPP);
2416                 } else {
2417                         rc = -EINVAL;
2418                 }
2419                 break;
2420         case TCP_NODELAY:
2421                 if (sk->sk_state != SMC_INIT &&
2422                     sk->sk_state != SMC_LISTEN &&
2423                     sk->sk_state != SMC_CLOSED) {
2424                         if (val) {
2425                                 SMC_STAT_INC(smc, ndly_cnt);
2426                                 smc_tx_pending(&smc->conn);
2427                                 cancel_delayed_work(&smc->conn.tx_work);
2428                         }
2429                 }
2430                 break;
2431         case TCP_CORK:
2432                 if (sk->sk_state != SMC_INIT &&
2433                     sk->sk_state != SMC_LISTEN &&
2434                     sk->sk_state != SMC_CLOSED) {
2435                         if (!val) {
2436                                 SMC_STAT_INC(smc, cork_cnt);
2437                                 smc_tx_pending(&smc->conn);
2438                                 cancel_delayed_work(&smc->conn.tx_work);
2439                         }
2440                 }
2441                 break;
2442         case TCP_DEFER_ACCEPT:
2443                 smc->sockopt_defer_accept = val;
2444                 break;
2445         default:
2446                 break;
2447         }
2448 out:
2449         release_sock(sk);
2450
2451         return rc;
2452 }
2453
2454 static int smc_getsockopt(struct socket *sock, int level, int optname,
2455                           char __user *optval, int __user *optlen)
2456 {
2457         struct smc_sock *smc;
2458         int rc;
2459
2460         smc = smc_sk(sock->sk);
2461         mutex_lock(&smc->clcsock_release_lock);
2462         if (!smc->clcsock) {
2463                 mutex_unlock(&smc->clcsock_release_lock);
2464                 return -EBADF;
2465         }
2466         /* socket options apply to the CLC socket */
2467         if (unlikely(!smc->clcsock->ops->getsockopt)) {
2468                 mutex_unlock(&smc->clcsock_release_lock);
2469                 return -EOPNOTSUPP;
2470         }
2471         rc = smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
2472                                            optval, optlen);
2473         mutex_unlock(&smc->clcsock_release_lock);
2474         return rc;
2475 }
2476
2477 static int smc_ioctl(struct socket *sock, unsigned int cmd,
2478                      unsigned long arg)
2479 {
2480         union smc_host_cursor cons, urg;
2481         struct smc_connection *conn;
2482         struct smc_sock *smc;
2483         int answ;
2484
2485         smc = smc_sk(sock->sk);
2486         conn = &smc->conn;
2487         lock_sock(&smc->sk);
2488         if (smc->use_fallback) {
2489                 if (!smc->clcsock) {
2490                         release_sock(&smc->sk);
2491                         return -EBADF;
2492                 }
2493                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
2494                 release_sock(&smc->sk);
2495                 return answ;
2496         }
2497         switch (cmd) {
2498         case SIOCINQ: /* same as FIONREAD */
2499                 if (smc->sk.sk_state == SMC_LISTEN) {
2500                         release_sock(&smc->sk);
2501                         return -EINVAL;
2502                 }
2503                 if (smc->sk.sk_state == SMC_INIT ||
2504                     smc->sk.sk_state == SMC_CLOSED)
2505                         answ = 0;
2506                 else
2507                         answ = atomic_read(&smc->conn.bytes_to_rcv);
2508                 break;
2509         case SIOCOUTQ:
2510                 /* output queue size (not send + not acked) */
2511                 if (smc->sk.sk_state == SMC_LISTEN) {
2512                         release_sock(&smc->sk);
2513                         return -EINVAL;
2514                 }
2515                 if (smc->sk.sk_state == SMC_INIT ||
2516                     smc->sk.sk_state == SMC_CLOSED)
2517                         answ = 0;
2518                 else
2519                         answ = smc->conn.sndbuf_desc->len -
2520                                         atomic_read(&smc->conn.sndbuf_space);
2521                 break;
2522         case SIOCOUTQNSD:
2523                 /* output queue size (not send only) */
2524                 if (smc->sk.sk_state == SMC_LISTEN) {
2525                         release_sock(&smc->sk);
2526                         return -EINVAL;
2527                 }
2528                 if (smc->sk.sk_state == SMC_INIT ||
2529                     smc->sk.sk_state == SMC_CLOSED)
2530                         answ = 0;
2531                 else
2532                         answ = smc_tx_prepared_sends(&smc->conn);
2533                 break;
2534         case SIOCATMARK:
2535                 if (smc->sk.sk_state == SMC_LISTEN) {
2536                         release_sock(&smc->sk);
2537                         return -EINVAL;
2538                 }
2539                 if (smc->sk.sk_state == SMC_INIT ||
2540                     smc->sk.sk_state == SMC_CLOSED) {
2541                         answ = 0;
2542                 } else {
2543                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
2544                         smc_curs_copy(&urg, &conn->urg_curs, conn);
2545                         answ = smc_curs_diff(conn->rmb_desc->len,
2546                                              &cons, &urg) == 1;
2547                 }
2548                 break;
2549         default:
2550                 release_sock(&smc->sk);
2551                 return -ENOIOCTLCMD;
2552         }
2553         release_sock(&smc->sk);
2554
2555         return put_user(answ, (int __user *)arg);
2556 }
2557
2558 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
2559                             int offset, size_t size, int flags)
2560 {
2561         struct sock *sk = sock->sk;
2562         struct smc_sock *smc;
2563         int rc = -EPIPE;
2564
2565         smc = smc_sk(sk);
2566         lock_sock(sk);
2567         if (sk->sk_state != SMC_ACTIVE) {
2568                 release_sock(sk);
2569                 goto out;
2570         }
2571         release_sock(sk);
2572         if (smc->use_fallback) {
2573                 rc = kernel_sendpage(smc->clcsock, page, offset,
2574                                      size, flags);
2575         } else {
2576                 SMC_STAT_INC(smc, sendpage_cnt);
2577                 rc = sock_no_sendpage(sock, page, offset, size, flags);
2578         }
2579
2580 out:
2581         return rc;
2582 }
2583
2584 /* Map the affected portions of the rmbe into an spd, note the number of bytes
2585  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
2586  * updates till whenever a respective page has been fully processed.
2587  * Note that subsequent recv() calls have to wait till all splice() processing
2588  * completed.
2589  */
2590 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
2591                                struct pipe_inode_info *pipe, size_t len,
2592                                unsigned int flags)
2593 {
2594         struct sock *sk = sock->sk;
2595         struct smc_sock *smc;
2596         int rc = -ENOTCONN;
2597
2598         smc = smc_sk(sk);
2599         lock_sock(sk);
2600         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
2601                 /* socket was connected before, no more data to read */
2602                 rc = 0;
2603                 goto out;
2604         }
2605         if (sk->sk_state == SMC_INIT ||
2606             sk->sk_state == SMC_LISTEN ||
2607             sk->sk_state == SMC_CLOSED)
2608                 goto out;
2609
2610         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
2611                 rc = 0;
2612                 goto out;
2613         }
2614
2615         if (smc->use_fallback) {
2616                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
2617                                                     pipe, len, flags);
2618         } else {
2619                 if (*ppos) {
2620                         rc = -ESPIPE;
2621                         goto out;
2622                 }
2623                 if (flags & SPLICE_F_NONBLOCK)
2624                         flags = MSG_DONTWAIT;
2625                 else
2626                         flags = 0;
2627                 SMC_STAT_INC(smc, splice_cnt);
2628                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
2629         }
2630 out:
2631         release_sock(sk);
2632
2633         return rc;
2634 }
2635
2636 /* must look like tcp */
2637 static const struct proto_ops smc_sock_ops = {
2638         .family         = PF_SMC,
2639         .owner          = THIS_MODULE,
2640         .release        = smc_release,
2641         .bind           = smc_bind,
2642         .connect        = smc_connect,
2643         .socketpair     = sock_no_socketpair,
2644         .accept         = smc_accept,
2645         .getname        = smc_getname,
2646         .poll           = smc_poll,
2647         .ioctl          = smc_ioctl,
2648         .listen         = smc_listen,
2649         .shutdown       = smc_shutdown,
2650         .setsockopt     = smc_setsockopt,
2651         .getsockopt     = smc_getsockopt,
2652         .sendmsg        = smc_sendmsg,
2653         .recvmsg        = smc_recvmsg,
2654         .mmap           = sock_no_mmap,
2655         .sendpage       = smc_sendpage,
2656         .splice_read    = smc_splice_read,
2657 };
2658
2659 static int smc_create(struct net *net, struct socket *sock, int protocol,
2660                       int kern)
2661 {
2662         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
2663         struct smc_sock *smc;
2664         struct sock *sk;
2665         int rc;
2666
2667         rc = -ESOCKTNOSUPPORT;
2668         if (sock->type != SOCK_STREAM)
2669                 goto out;
2670
2671         rc = -EPROTONOSUPPORT;
2672         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
2673                 goto out;
2674
2675         rc = -ENOBUFS;
2676         sock->ops = &smc_sock_ops;
2677         sk = smc_sock_alloc(net, sock, protocol);
2678         if (!sk)
2679                 goto out;
2680
2681         /* create internal TCP socket for CLC handshake and fallback */
2682         smc = smc_sk(sk);
2683         smc->use_fallback = false; /* assume rdma capability first */
2684         smc->fallback_rsn = 0;
2685         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2686                               &smc->clcsock);
2687         if (rc) {
2688                 sk_common_release(sk);
2689                 goto out;
2690         }
2691         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2692         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
2693
2694 out:
2695         return rc;
2696 }
2697
2698 static const struct net_proto_family smc_sock_family_ops = {
2699         .family = PF_SMC,
2700         .owner  = THIS_MODULE,
2701         .create = smc_create,
2702 };
2703
2704 unsigned int smc_net_id;
2705
2706 static __net_init int smc_net_init(struct net *net)
2707 {
2708         return smc_pnet_net_init(net);
2709 }
2710
2711 static void __net_exit smc_net_exit(struct net *net)
2712 {
2713         smc_pnet_net_exit(net);
2714 }
2715
2716 static __net_init int smc_net_stat_init(struct net *net)
2717 {
2718         return smc_stats_init(net);
2719 }
2720
2721 static void __net_exit smc_net_stat_exit(struct net *net)
2722 {
2723         smc_stats_exit(net);
2724 }
2725
2726 static struct pernet_operations smc_net_ops = {
2727         .init = smc_net_init,
2728         .exit = smc_net_exit,
2729         .id   = &smc_net_id,
2730         .size = sizeof(struct smc_net),
2731 };
2732
2733 static struct pernet_operations smc_net_stat_ops = {
2734         .init = smc_net_stat_init,
2735         .exit = smc_net_stat_exit,
2736 };
2737
2738 static int __init smc_init(void)
2739 {
2740         int rc;
2741
2742         rc = register_pernet_subsys(&smc_net_ops);
2743         if (rc)
2744                 return rc;
2745
2746         rc = register_pernet_subsys(&smc_net_stat_ops);
2747         if (rc)
2748                 return rc;
2749
2750         smc_ism_init();
2751         smc_clc_init();
2752
2753         rc = smc_nl_init();
2754         if (rc)
2755                 goto out_pernet_subsys;
2756
2757         rc = smc_pnet_init();
2758         if (rc)
2759                 goto out_nl;
2760
2761         rc = -ENOMEM;
2762         smc_hs_wq = alloc_workqueue("smc_hs_wq", 0, 0);
2763         if (!smc_hs_wq)
2764                 goto out_pnet;
2765
2766         smc_close_wq = alloc_workqueue("smc_close_wq", 0, 0);
2767         if (!smc_close_wq)
2768                 goto out_alloc_hs_wq;
2769
2770         rc = smc_core_init();
2771         if (rc) {
2772                 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
2773                 goto out_alloc_wqs;
2774         }
2775
2776         rc = smc_llc_init();
2777         if (rc) {
2778                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2779                 goto out_core;
2780         }
2781
2782         rc = smc_cdc_init();
2783         if (rc) {
2784                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2785                 goto out_core;
2786         }
2787
2788         rc = proto_register(&smc_proto, 1);
2789         if (rc) {
2790                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2791                 goto out_core;
2792         }
2793
2794         rc = proto_register(&smc_proto6, 1);
2795         if (rc) {
2796                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2797                 goto out_proto;
2798         }
2799
2800         rc = sock_register(&smc_sock_family_ops);
2801         if (rc) {
2802                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2803                 goto out_proto6;
2804         }
2805         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2806         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2807
2808         rc = smc_ib_register_client();
2809         if (rc) {
2810                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2811                 goto out_sock;
2812         }
2813
2814         static_branch_enable(&tcp_have_smc);
2815         return 0;
2816
2817 out_sock:
2818         sock_unregister(PF_SMC);
2819 out_proto6:
2820         proto_unregister(&smc_proto6);
2821 out_proto:
2822         proto_unregister(&smc_proto);
2823 out_core:
2824         smc_core_exit();
2825 out_alloc_wqs:
2826         destroy_workqueue(smc_close_wq);
2827 out_alloc_hs_wq:
2828         destroy_workqueue(smc_hs_wq);
2829 out_pnet:
2830         smc_pnet_exit();
2831 out_nl:
2832         smc_nl_exit();
2833 out_pernet_subsys:
2834         unregister_pernet_subsys(&smc_net_ops);
2835
2836         return rc;
2837 }
2838
2839 static void __exit smc_exit(void)
2840 {
2841         static_branch_disable(&tcp_have_smc);
2842         sock_unregister(PF_SMC);
2843         smc_core_exit();
2844         smc_ib_unregister_client();
2845         destroy_workqueue(smc_close_wq);
2846         destroy_workqueue(smc_hs_wq);
2847         proto_unregister(&smc_proto6);
2848         proto_unregister(&smc_proto);
2849         smc_pnet_exit();
2850         smc_nl_exit();
2851         unregister_pernet_subsys(&smc_net_stat_ops);
2852         unregister_pernet_subsys(&smc_net_ops);
2853         rcu_barrier();
2854 }
2855
2856 module_init(smc_init);
2857 module_exit(smc_exit);
2858
2859 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2860 MODULE_DESCRIPTION("smc socket address family");
2861 MODULE_LICENSE("GPL");
2862 MODULE_ALIAS_NETPROTO(PF_SMC);