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