net/smc/af_smc.c

   1 /*
   2  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   3  *
   4  *  AF_SMC protocol family socket handler keeping the AF_INET sock address type
   5  *  applies to SOCK_STREAM sockets only
   6  *  offers an alternative communication option for TCP-protocol sockets
   7  *  applicable with RoCE-cards only
   8  *
   9  *  Initial restrictions:
  10  *    - support for alternate links postponed
  11  *
  12  *  Copyright IBM Corp. 2016, 2018
  13  *
  14  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  15  *              based on prototype from Frank Blaschka
  16  */
  17
  18 #define KMSG_COMPONENT "smc"
  19 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  20
  21 #include <linux/module.h>
  22 #include <linux/socket.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/in.h>
  25 #include <linux/sched/signal.h>
  26 #include <linux/if_vlan.h>
  27
  28 #include <net/sock.h>
  29 #include <net/tcp.h>
  30 #include <net/smc.h>
  31 #include <asm/ioctls.h>
  32
  33 #include <net/net_namespace.h>
  34 #include <net/netns/generic.h>
  35 #include "smc_netns.h"
  36
  37 #include "smc.h"
  38 #include "smc_clc.h"
  39 #include "smc_llc.h"
  40 #include "smc_cdc.h"
  41 #include "smc_core.h"
  42 #include "smc_ib.h"
  43 #include "smc_ism.h"
  44 #include "smc_pnet.h"
  45 #include "smc_tx.h"
  46 #include "smc_rx.h"
  47 #include "smc_close.h"
  48
  49 static DEFINE_MUTEX(smc_server_lgr_pending);    /* serialize link group
  50                                                  * creation on server
  51                                                  */
  52 static DEFINE_MUTEX(smc_client_lgr_pending);    /* serialize link group
  53                                                  * creation on client
  54                                                  */
  55
  56 static void smc_tcp_listen_work(struct work_struct *);
  57 static void smc_connect_work(struct work_struct *);
  58
  59 static void smc_set_keepalive(struct sock *sk, int val)
  60 {
  61         struct smc_sock *smc = smc_sk(sk);
  62
  63         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
  64 }
  65
  66 static struct smc_hashinfo smc_v4_hashinfo = {
  67         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
  68 };
  69
  70 static struct smc_hashinfo smc_v6_hashinfo = {
  71         .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
  72 };
  73
  74 int smc_hash_sk(struct sock *sk)
  75 {
  76         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  77         struct hlist_head *head;
  78
  79         head = &h->ht;
  80
  81         write_lock_bh(&h->lock);
  82         sk_add_node(sk, head);
  83         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
  84         write_unlock_bh(&h->lock);
  85
  86         return 0;
  87 }
  88 EXPORT_SYMBOL_GPL(smc_hash_sk);
  89
  90 void smc_unhash_sk(struct sock *sk)
  91 {
  92         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  93
  94         write_lock_bh(&h->lock);
  95         if (sk_del_node_init(sk))
  96                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
  97         write_unlock_bh(&h->lock);
  98 }
  99 EXPORT_SYMBOL_GPL(smc_unhash_sk);
 100
 101 struct proto smc_proto = {
 102         .name           = "SMC",
 103         .owner          = THIS_MODULE,
 104         .keepalive      = smc_set_keepalive,
 105         .hash           = smc_hash_sk,
 106         .unhash         = smc_unhash_sk,
 107         .obj_size       = sizeof(struct smc_sock),
 108         .h.smc_hash     = &smc_v4_hashinfo,
 109         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 110 };
 111 EXPORT_SYMBOL_GPL(smc_proto);
 112
 113 struct proto smc_proto6 = {
 114         .name           = "SMC6",
 115         .owner          = THIS_MODULE,
 116         .keepalive      = smc_set_keepalive,
 117         .hash           = smc_hash_sk,
 118         .unhash         = smc_unhash_sk,
 119         .obj_size       = sizeof(struct smc_sock),
 120         .h.smc_hash     = &smc_v6_hashinfo,
 121         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 122 };
 123 EXPORT_SYMBOL_GPL(smc_proto6);
 124
 125 static int smc_release(struct socket *sock)
 126 {
 127         struct sock *sk = sock->sk;
 128         struct smc_sock *smc;
 129         int rc = 0;
 130
 131         if (!sk)
 132                 goto out;
 133
 134         smc = smc_sk(sk);
 135
 136         /* cleanup for a dangling non-blocking connect */
 137         if (smc->connect_nonblock && sk->sk_state == SMC_INIT)
 138                 tcp_abort(smc->clcsock->sk, ECONNABORTED);
 139         flush_work(&smc->connect_work);
 140
 141         if (sk->sk_state == SMC_LISTEN)
 142                 /* smc_close_non_accepted() is called and acquires
 143                  * sock lock for child sockets again
 144                  */
 145                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 146         else
 147                 lock_sock(sk);
 148
 149         if (!smc->use_fallback) {
 150                 rc = smc_close_active(smc);
 151                 sock_set_flag(sk, SOCK_DEAD);
 152                 sk->sk_shutdown |= SHUTDOWN_MASK;
 153         } else {
 154                 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
 155                         sock_put(sk); /* passive closing */
 156                 if (sk->sk_state == SMC_LISTEN) {
 157                         /* wake up clcsock accept */
 158                         rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
 159                 }
 160                 sk->sk_state = SMC_CLOSED;
 161                 sk->sk_state_change(sk);
 162         }
 163
 164         sk->sk_prot->unhash(sk);
 165
 166         if (sk->sk_state == SMC_CLOSED) {
 167                 if (smc->clcsock) {
 168                         mutex_lock(&smc->clcsock_release_lock);
 169                         sock_release(smc->clcsock);
 170                         smc->clcsock = NULL;
 171                         mutex_unlock(&smc->clcsock_release_lock);
 172                 }
 173                 if (!smc->use_fallback)
 174                         smc_conn_free(&smc->conn);
 175         }
 176
 177         /* detach socket */
 178         sock_orphan(sk);
 179         sock->sk = NULL;
 180         release_sock(sk);
 181
 182         sock_put(sk); /* final sock_put */
 183 out:
 184         return rc;
 185 }
 186
 187 static void smc_destruct(struct sock *sk)
 188 {
 189         if (sk->sk_state != SMC_CLOSED)
 190                 return;
 191         if (!sock_flag(sk, SOCK_DEAD))
 192                 return;
 193
 194         sk_refcnt_debug_dec(sk);
 195 }
 196
 197 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 198                                    int protocol)
 199 {
 200         struct smc_sock *smc;
 201         struct proto *prot;
 202         struct sock *sk;
 203
 204         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
 205         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
 206         if (!sk)
 207                 return NULL;
 208
 209         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 210         sk->sk_state = SMC_INIT;
 211         sk->sk_destruct = smc_destruct;
 212         sk->sk_protocol = protocol;
 213         smc = smc_sk(sk);
 214         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 215         INIT_WORK(&smc->connect_work, smc_connect_work);
 216         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 217         INIT_LIST_HEAD(&smc->accept_q);
 218         spin_lock_init(&smc->accept_q_lock);
 219         spin_lock_init(&smc->conn.send_lock);
 220         sk->sk_prot->hash(sk);
 221         sk_refcnt_debug_inc(sk);
 222         mutex_init(&smc->clcsock_release_lock);
 223
 224         return sk;
 225 }
 226
 227 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 228                     int addr_len)
 229 {
 230         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 231         struct sock *sk = sock->sk;
 232         struct smc_sock *smc;
 233         int rc;
 234
 235         smc = smc_sk(sk);
 236
 237         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 238         rc = -EINVAL;
 239         if (addr_len < sizeof(struct sockaddr_in))
 240                 goto out;
 241
 242         rc = -EAFNOSUPPORT;
 243         if (addr->sin_family != AF_INET &&
 244             addr->sin_family != AF_INET6 &&
 245             addr->sin_family != AF_UNSPEC)
 246                 goto out;
 247         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 248         if (addr->sin_family == AF_UNSPEC &&
 249             addr->sin_addr.s_addr != htonl(INADDR_ANY))
 250                 goto out;
 251
 252         lock_sock(sk);
 253
 254         /* Check if socket is already active */
 255         rc = -EINVAL;
 256         if (sk->sk_state != SMC_INIT)
 257                 goto out_rel;
 258
 259         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 260         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 261
 262 out_rel:
 263         release_sock(sk);
 264 out:
 265         return rc;
 266 }
 267
 268 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 269                                    unsigned long mask)
 270 {
 271         /* options we don't get control via setsockopt for */
 272         nsk->sk_type = osk->sk_type;
 273         nsk->sk_sndbuf = osk->sk_sndbuf;
 274         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 275         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 276         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 277         nsk->sk_mark = osk->sk_mark;
 278         nsk->sk_priority = osk->sk_priority;
 279         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 280         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 281         nsk->sk_err = osk->sk_err;
 282
 283         nsk->sk_flags &= ~mask;
 284         nsk->sk_flags |= osk->sk_flags & mask;
 285 }
 286
 287 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 288                              (1UL << SOCK_KEEPOPEN) | \
 289                              (1UL << SOCK_LINGER) | \
 290                              (1UL << SOCK_BROADCAST) | \
 291                              (1UL << SOCK_TIMESTAMP) | \
 292                              (1UL << SOCK_DBG) | \
 293                              (1UL << SOCK_RCVTSTAMP) | \
 294                              (1UL << SOCK_RCVTSTAMPNS) | \
 295                              (1UL << SOCK_LOCALROUTE) | \
 296                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 297                              (1UL << SOCK_RXQ_OVFL) | \
 298                              (1UL << SOCK_WIFI_STATUS) | \
 299                              (1UL << SOCK_NOFCS) | \
 300                              (1UL << SOCK_FILTER_LOCKED) | \
 301                              (1UL << SOCK_TSTAMP_NEW))
 302 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 303  * clc socket (since smc is not called for these options from net/core)
 304  */
 305 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 306 {
 307         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 308 }
 309
 310 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 311                              (1UL << SOCK_KEEPOPEN) | \
 312                              (1UL << SOCK_LINGER) | \
 313                              (1UL << SOCK_DBG))
 314 /* copy only settings and flags relevant for smc from clc to smc socket */
 315 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 316 {
 317         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 318 }
 319
 320 /* register a new rmb, send confirm_rkey msg to register with peer */
 321 static int smc_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc,
 322                        bool conf_rkey)
 323 {
 324         if (!rmb_desc->wr_reg) {
 325                 /* register memory region for new rmb */
 326                 if (smc_wr_reg_send(link, rmb_desc->mr_rx[SMC_SINGLE_LINK])) {
 327                         rmb_desc->regerr = 1;
 328                         return -EFAULT;
 329                 }
 330                 rmb_desc->wr_reg = 1;
 331         }
 332         if (!conf_rkey)
 333                 return 0;
 334         /* exchange confirm_rkey msg with peer */
 335         if (smc_llc_do_confirm_rkey(link, rmb_desc)) {
 336                 rmb_desc->regerr = 1;
 337                 return -EFAULT;
 338         }
 339         return 0;
 340 }
 341
 342 static int smc_clnt_conf_first_link(struct smc_sock *smc)
 343 {
 344         struct net *net = sock_net(smc->clcsock->sk);
 345         struct smc_link_group *lgr = smc->conn.lgr;
 346         struct smc_link *link;
 347         int rest;
 348         int rc;
 349
 350         link = &lgr->lnk[SMC_SINGLE_LINK];
 351         /* receive CONFIRM LINK request from server over RoCE fabric */
 352         rest = wait_for_completion_interruptible_timeout(
 353                 &link->llc_confirm,
 354                 SMC_LLC_WAIT_FIRST_TIME);
 355         if (rest <= 0) {
 356                 struct smc_clc_msg_decline dclc;
 357
 358                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 359                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 360                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
 361         }
 362
 363         if (link->llc_confirm_rc)
 364                 return SMC_CLC_DECL_RMBE_EC;
 365
 366         rc = smc_ib_modify_qp_rts(link);
 367         if (rc)
 368                 return SMC_CLC_DECL_ERR_RDYLNK;
 369
 370         smc_wr_remember_qp_attr(link);
 371
 372         if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
 373                 return SMC_CLC_DECL_ERR_REGRMB;
 374
 375         /* send CONFIRM LINK response over RoCE fabric */
 376         rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
 377         if (rc < 0)
 378                 return SMC_CLC_DECL_TIMEOUT_CL;
 379
 380         /* receive ADD LINK request from server over RoCE fabric */
 381         rest = wait_for_completion_interruptible_timeout(&link->llc_add,
 382                                                          SMC_LLC_WAIT_TIME);
 383         if (rest <= 0) {
 384                 struct smc_clc_msg_decline dclc;
 385
 386                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 387                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 388                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_AL : rc;
 389         }
 390
 391         /* send add link reject message, only one link supported for now */
 392         rc = smc_llc_send_add_link(link,
 393                                    link->smcibdev->mac[link->ibport - 1],
 394                                    link->gid, SMC_LLC_RESP);
 395         if (rc < 0)
 396                 return SMC_CLC_DECL_TIMEOUT_AL;
 397
 398         smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
 399
 400         return 0;
 401 }
 402
 403 static void smcr_conn_save_peer_info(struct smc_sock *smc,
 404                                      struct smc_clc_msg_accept_confirm *clc)
 405 {
 406         int bufsize = smc_uncompress_bufsize(clc->rmbe_size);
 407
 408         smc->conn.peer_rmbe_idx = clc->rmbe_idx;
 409         smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
 410         smc->conn.peer_rmbe_size = bufsize;
 411         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 412         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
 413 }
 414
 415 static void smcd_conn_save_peer_info(struct smc_sock *smc,
 416                                      struct smc_clc_msg_accept_confirm *clc)
 417 {
 418         int bufsize = smc_uncompress_bufsize(clc->dmbe_size);
 419
 420         smc->conn.peer_rmbe_idx = clc->dmbe_idx;
 421         smc->conn.peer_token = clc->token;
 422         /* msg header takes up space in the buffer */
 423         smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
 424         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 425         smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
 426 }
 427
 428 static void smc_conn_save_peer_info(struct smc_sock *smc,
 429                                     struct smc_clc_msg_accept_confirm *clc)
 430 {
 431         if (smc->conn.lgr->is_smcd)
 432                 smcd_conn_save_peer_info(smc, clc);
 433         else
 434                 smcr_conn_save_peer_info(smc, clc);
 435 }
 436
 437 static void smc_link_save_peer_info(struct smc_link *link,
 438                                     struct smc_clc_msg_accept_confirm *clc)
 439 {
 440         link->peer_qpn = ntoh24(clc->qpn);
 441         memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
 442         memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
 443         link->peer_psn = ntoh24(clc->psn);
 444         link->peer_mtu = clc->qp_mtu;
 445 }
 446
 447 /* fall back during connect */
 448 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
 449 {
 450         smc->use_fallback = true;
 451         smc->fallback_rsn = reason_code;
 452         smc_copy_sock_settings_to_clc(smc);
 453         smc->connect_nonblock = 0;
 454         if (smc->sk.sk_state == SMC_INIT)
 455                 smc->sk.sk_state = SMC_ACTIVE;
 456         return 0;
 457 }
 458
 459 /* decline and fall back during connect */
 460 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
 461 {
 462         int rc;
 463
 464         if (reason_code < 0) { /* error, fallback is not possible */
 465                 if (smc->sk.sk_state == SMC_INIT)
 466                         sock_put(&smc->sk); /* passive closing */
 467                 return reason_code;
 468         }
 469         if (reason_code != SMC_CLC_DECL_PEERDECL) {
 470                 rc = smc_clc_send_decline(smc, reason_code);
 471                 if (rc < 0) {
 472                         if (smc->sk.sk_state == SMC_INIT)
 473                                 sock_put(&smc->sk); /* passive closing */
 474                         return rc;
 475                 }
 476         }
 477         return smc_connect_fallback(smc, reason_code);
 478 }
 479
 480 /* abort connecting */
 481 static int smc_connect_abort(struct smc_sock *smc, int reason_code,
 482                              int local_contact)
 483 {
 484         if (local_contact == SMC_FIRST_CONTACT)
 485                 smc_lgr_forget(smc->conn.lgr);
 486         if (smc->conn.lgr->is_smcd)
 487                 /* there is only one lgr role for SMC-D; use server lock */
 488                 mutex_unlock(&smc_server_lgr_pending);
 489         else
 490                 mutex_unlock(&smc_client_lgr_pending);
 491
 492         smc_conn_free(&smc->conn);
 493         smc->connect_nonblock = 0;
 494         return reason_code;
 495 }
 496
 497 /* check if there is a rdma device available for this connection. */
 498 /* called for connect and listen */
 499 static int smc_check_rdma(struct smc_sock *smc, struct smc_ib_device **ibdev,
 500                           u8 *ibport, unsigned short vlan_id, u8 gid[])
 501 {
 502         int reason_code = 0;
 503
 504         /* PNET table look up: search active ib_device and port
 505          * within same PNETID that also contains the ethernet device
 506          * used for the internal TCP socket
 507          */
 508         smc_pnet_find_roce_resource(smc->clcsock->sk, ibdev, ibport, vlan_id,
 509                                     gid);
 510         if (!(*ibdev))
 511                 reason_code = SMC_CLC_DECL_CNFERR; /* configuration error */
 512
 513         return reason_code;
 514 }
 515
 516 /* check if there is an ISM device available for this connection. */
 517 /* called for connect and listen */
 518 static int smc_check_ism(struct smc_sock *smc, struct smcd_dev **ismdev)
 519 {
 520         /* Find ISM device with same PNETID as connecting interface  */
 521         smc_pnet_find_ism_resource(smc->clcsock->sk, ismdev);
 522         if (!(*ismdev))
 523                 return SMC_CLC_DECL_CNFERR; /* configuration error */
 524         return 0;
 525 }
 526
 527 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
 528 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
 529                                       struct smcd_dev *ismdev,
 530                                       unsigned short vlan_id)
 531 {
 532         if (vlan_id && smc_ism_get_vlan(ismdev, vlan_id))
 533                 return SMC_CLC_DECL_CNFERR;
 534         return 0;
 535 }
 536
 537 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
 538  * used, the VLAN ID will be registered again during the connection setup.
 539  */
 540 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc, bool is_smcd,
 541                                         struct smcd_dev *ismdev,
 542                                         unsigned short vlan_id)
 543 {
 544         if (!is_smcd)
 545                 return 0;
 546         if (vlan_id && smc_ism_put_vlan(ismdev, vlan_id))
 547                 return SMC_CLC_DECL_CNFERR;
 548         return 0;
 549 }
 550
 551 /* CLC handshake during connect */
 552 static int smc_connect_clc(struct smc_sock *smc, int smc_type,
 553                            struct smc_clc_msg_accept_confirm *aclc,
 554                            struct smc_ib_device *ibdev, u8 ibport,
 555                            u8 gid[], struct smcd_dev *ismdev)
 556 {
 557         int rc = 0;
 558
 559         /* do inband token exchange */
 560         rc = smc_clc_send_proposal(smc, smc_type, ibdev, ibport, gid, ismdev);
 561         if (rc)
 562                 return rc;
 563         /* receive SMC Accept CLC message */
 564         return smc_clc_wait_msg(smc, aclc, sizeof(*aclc), SMC_CLC_ACCEPT,
 565                                 CLC_WAIT_TIME);
 566 }
 567
 568 /* setup for RDMA connection of client */
 569 static int smc_connect_rdma(struct smc_sock *smc,
 570                             struct smc_clc_msg_accept_confirm *aclc,
 571                             struct smc_ib_device *ibdev, u8 ibport)
 572 {
 573         int local_contact = SMC_FIRST_CONTACT;
 574         struct smc_link *link;
 575         int reason_code = 0;
 576
 577         mutex_lock(&smc_client_lgr_pending);
 578         local_contact = smc_conn_create(smc, false, aclc->hdr.flag, ibdev,
 579                                         ibport, ntoh24(aclc->qpn), &aclc->lcl,
 580                                         NULL, 0);
 581         if (local_contact < 0) {
 582                 if (local_contact == -ENOMEM)
 583                         reason_code = SMC_CLC_DECL_MEM;/* insufficient memory*/
 584                 else if (local_contact == -ENOLINK)
 585                         reason_code = SMC_CLC_DECL_SYNCERR; /* synchr. error */
 586                 else
 587                         reason_code = SMC_CLC_DECL_INTERR; /* other error */
 588                 mutex_unlock(&smc_client_lgr_pending);
 589                 return reason_code;
 590         }
 591         link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
 592
 593         smc_conn_save_peer_info(smc, aclc);
 594
 595         /* create send buffer and rmb */
 596         if (smc_buf_create(smc, false))
 597                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, local_contact);
 598
 599         if (local_contact == SMC_FIRST_CONTACT)
 600                 smc_link_save_peer_info(link, aclc);
 601
 602         if (smc_rmb_rtoken_handling(&smc->conn, aclc))
 603                 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RTOK,
 604                                          local_contact);
 605
 606         smc_close_init(smc);
 607         smc_rx_init(smc);
 608
 609         if (local_contact == SMC_FIRST_CONTACT) {
 610                 if (smc_ib_ready_link(link))
 611                         return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RDYLNK,
 612                                                  local_contact);
 613         } else {
 614                 if (smc_reg_rmb(link, smc->conn.rmb_desc, true))
 615                         return smc_connect_abort(smc, SMC_CLC_DECL_ERR_REGRMB,
 616                                                  local_contact);
 617         }
 618         smc_rmb_sync_sg_for_device(&smc->conn);
 619
 620         reason_code = smc_clc_send_confirm(smc);
 621         if (reason_code)
 622                 return smc_connect_abort(smc, reason_code, local_contact);
 623
 624         smc_tx_init(smc);
 625
 626         if (local_contact == SMC_FIRST_CONTACT) {
 627                 /* QP confirmation over RoCE fabric */
 628                 reason_code = smc_clnt_conf_first_link(smc);
 629                 if (reason_code)
 630                         return smc_connect_abort(smc, reason_code,
 631                                                  local_contact);
 632         }
 633         mutex_unlock(&smc_client_lgr_pending);
 634
 635         smc_copy_sock_settings_to_clc(smc);
 636         smc->connect_nonblock = 0;
 637         if (smc->sk.sk_state == SMC_INIT)
 638                 smc->sk.sk_state = SMC_ACTIVE;
 639
 640         return 0;
 641 }
 642
 643 /* setup for ISM connection of client */
 644 static int smc_connect_ism(struct smc_sock *smc,
 645                            struct smc_clc_msg_accept_confirm *aclc,
 646                            struct smcd_dev *ismdev)
 647 {
 648         int local_contact = SMC_FIRST_CONTACT;
 649         int rc = 0;
 650
 651         /* there is only one lgr role for SMC-D; use server lock */
 652         mutex_lock(&smc_server_lgr_pending);
 653         local_contact = smc_conn_create(smc, true, aclc->hdr.flag, NULL, 0, 0,
 654                                         NULL, ismdev, aclc->gid);
 655         if (local_contact < 0) {
 656                 mutex_unlock(&smc_server_lgr_pending);
 657                 return SMC_CLC_DECL_MEM;
 658         }
 659
 660         /* Create send and receive buffers */
 661         if (smc_buf_create(smc, true))
 662                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM, local_contact);
 663
 664         smc_conn_save_peer_info(smc, aclc);
 665         smc_close_init(smc);
 666         smc_rx_init(smc);
 667         smc_tx_init(smc);
 668
 669         rc = smc_clc_send_confirm(smc);
 670         if (rc)
 671                 return smc_connect_abort(smc, rc, local_contact);
 672         mutex_unlock(&smc_server_lgr_pending);
 673
 674         smc_copy_sock_settings_to_clc(smc);
 675         smc->connect_nonblock = 0;
 676         if (smc->sk.sk_state == SMC_INIT)
 677                 smc->sk.sk_state = SMC_ACTIVE;
 678
 679         return 0;
 680 }
 681
 682 /* perform steps before actually connecting */
 683 static int __smc_connect(struct smc_sock *smc)
 684 {
 685         bool ism_supported = false, rdma_supported = false;
 686         struct smc_clc_msg_accept_confirm aclc;
 687         struct smc_ib_device *ibdev;
 688         struct smcd_dev *ismdev;
 689         u8 gid[SMC_GID_SIZE];
 690         unsigned short vlan;
 691         int smc_type;
 692         int rc = 0;
 693         u8 ibport;
 694
 695         sock_hold(&smc->sk); /* sock put in passive closing */
 696
 697         if (smc->use_fallback)
 698                 return smc_connect_fallback(smc, smc->fallback_rsn);
 699
 700         /* if peer has not signalled SMC-capability, fall back */
 701         if (!tcp_sk(smc->clcsock->sk)->syn_smc)
 702                 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
 703
 704         /* IPSec connections opt out of SMC-R optimizations */
 705         if (using_ipsec(smc))
 706                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
 707
 708         /* check for VLAN ID */
 709         if (smc_vlan_by_tcpsk(smc->clcsock, &vlan))
 710                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_CNFERR);
 711
 712         /* check if there is an ism device available */
 713         if (!smc_check_ism(smc, &ismdev) &&
 714             !smc_connect_ism_vlan_setup(smc, ismdev, vlan)) {
 715                 /* ISM is supported for this connection */
 716                 ism_supported = true;
 717                 smc_type = SMC_TYPE_D;
 718         }
 719
 720         /* check if there is a rdma device available */
 721         if (!smc_check_rdma(smc, &ibdev, &ibport, vlan, gid)) {
 722                 /* RDMA is supported for this connection */
 723                 rdma_supported = true;
 724                 if (ism_supported)
 725                         smc_type = SMC_TYPE_B; /* both */
 726                 else
 727                         smc_type = SMC_TYPE_R; /* only RDMA */
 728         }
 729
 730         /* if neither ISM nor RDMA are supported, fallback */
 731         if (!rdma_supported && !ism_supported)
 732                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_NOSMCDEV);
 733
 734         /* perform CLC handshake */
 735         rc = smc_connect_clc(smc, smc_type, &aclc, ibdev, ibport, gid, ismdev);
 736         if (rc) {
 737                 smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
 738                 return smc_connect_decline_fallback(smc, rc);
 739         }
 740
 741         /* depending on previous steps, connect using rdma or ism */
 742         if (rdma_supported && aclc.hdr.path == SMC_TYPE_R)
 743                 rc = smc_connect_rdma(smc, &aclc, ibdev, ibport);
 744         else if (ism_supported && aclc.hdr.path == SMC_TYPE_D)
 745                 rc = smc_connect_ism(smc, &aclc, ismdev);
 746         else
 747                 rc = SMC_CLC_DECL_MODEUNSUPP;
 748         if (rc) {
 749                 smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
 750                 return smc_connect_decline_fallback(smc, rc);
 751         }
 752
 753         smc_connect_ism_vlan_cleanup(smc, ism_supported, ismdev, vlan);
 754         return 0;
 755 }
 756
 757 static void smc_connect_work(struct work_struct *work)
 758 {
 759         struct smc_sock *smc = container_of(work, struct smc_sock,
 760                                             connect_work);
 761         long timeo = smc->sk.sk_sndtimeo;
 762         int rc = 0;
 763
 764         if (!timeo)
 765                 timeo = MAX_SCHEDULE_TIMEOUT;
 766         lock_sock(smc->clcsock->sk);
 767         if (smc->clcsock->sk->sk_err) {
 768                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
 769         } else if ((1 << smc->clcsock->sk->sk_state) &
 770                                         (TCPF_SYN_SENT | TCP_SYN_RECV)) {
 771                 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
 772                 if ((rc == -EPIPE) &&
 773                     ((1 << smc->clcsock->sk->sk_state) &
 774                                         (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
 775                         rc = 0;
 776         }
 777         release_sock(smc->clcsock->sk);
 778         lock_sock(&smc->sk);
 779         if (rc != 0 || smc->sk.sk_err) {
 780                 smc->sk.sk_state = SMC_CLOSED;
 781                 if (rc == -EPIPE || rc == -EAGAIN)
 782                         smc->sk.sk_err = EPIPE;
 783                 else if (signal_pending(current))
 784                         smc->sk.sk_err = -sock_intr_errno(timeo);
 785                 goto out;
 786         }
 787
 788         rc = __smc_connect(smc);
 789         if (rc < 0)
 790                 smc->sk.sk_err = -rc;
 791
 792 out:
 793         if (smc->sk.sk_err)
 794                 smc->sk.sk_state_change(&smc->sk);
 795         else
 796                 smc->sk.sk_write_space(&smc->sk);
 797         release_sock(&smc->sk);
 798 }
 799
 800 static int smc_connect(struct socket *sock, struct sockaddr *addr,
 801                        int alen, int flags)
 802 {
 803         struct sock *sk = sock->sk;
 804         struct smc_sock *smc;
 805         int rc = -EINVAL;
 806
 807         smc = smc_sk(sk);
 808
 809         /* separate smc parameter checking to be safe */
 810         if (alen < sizeof(addr->sa_family))
 811                 goto out_err;
 812         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
 813                 goto out_err;
 814
 815         lock_sock(sk);
 816         switch (sk->sk_state) {
 817         default:
 818                 goto out;
 819         case SMC_ACTIVE:
 820                 rc = -EISCONN;
 821                 goto out;
 822         case SMC_INIT:
 823                 rc = 0;
 824                 break;
 825         }
 826
 827         smc_copy_sock_settings_to_clc(smc);
 828         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
 829         if (smc->connect_nonblock) {
 830                 rc = -EALREADY;
 831                 goto out;
 832         }
 833         rc = kernel_connect(smc->clcsock, addr, alen, flags);
 834         if (rc && rc != -EINPROGRESS)
 835                 goto out;
 836         if (flags & O_NONBLOCK) {
 837                 if (schedule_work(&smc->connect_work))
 838                         smc->connect_nonblock = 1;
 839                 rc = -EINPROGRESS;
 840         } else {
 841                 rc = __smc_connect(smc);
 842                 if (rc < 0)
 843                         goto out;
 844                 else
 845                         rc = 0; /* success cases including fallback */
 846         }
 847
 848 out:
 849         release_sock(sk);
 850 out_err:
 851         return rc;
 852 }
 853
 854 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 855 {
 856         struct socket *new_clcsock = NULL;
 857         struct sock *lsk = &lsmc->sk;
 858         struct sock *new_sk;
 859         int rc = -EINVAL;
 860
 861         release_sock(lsk);
 862         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
 863         if (!new_sk) {
 864                 rc = -ENOMEM;
 865                 lsk->sk_err = ENOMEM;
 866                 *new_smc = NULL;
 867                 lock_sock(lsk);
 868                 goto out;
 869         }
 870         *new_smc = smc_sk(new_sk);
 871
 872         mutex_lock(&lsmc->clcsock_release_lock);
 873         if (lsmc->clcsock)
 874                 rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
 875         mutex_unlock(&lsmc->clcsock_release_lock);
 876         lock_sock(lsk);
 877         if  (rc < 0)
 878                 lsk->sk_err = -rc;
 879         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
 880                 if (new_clcsock)
 881                         sock_release(new_clcsock);
 882                 new_sk->sk_state = SMC_CLOSED;
 883                 sock_set_flag(new_sk, SOCK_DEAD);
 884                 new_sk->sk_prot->unhash(new_sk);
 885                 sock_put(new_sk); /* final */
 886                 *new_smc = NULL;
 887                 goto out;
 888         }
 889
 890         (*new_smc)->clcsock = new_clcsock;
 891 out:
 892         return rc;
 893 }
 894
 895 /* add a just created sock to the accept queue of the listen sock as
 896  * candidate for a following socket accept call from user space
 897  */
 898 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
 899 {
 900         struct smc_sock *par = smc_sk(parent);
 901
 902         sock_hold(sk); /* sock_put in smc_accept_unlink () */
 903         spin_lock(&par->accept_q_lock);
 904         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
 905         spin_unlock(&par->accept_q_lock);
 906         sk_acceptq_added(parent);
 907 }
 908
 909 /* remove a socket from the accept queue of its parental listening socket */
 910 static void smc_accept_unlink(struct sock *sk)
 911 {
 912         struct smc_sock *par = smc_sk(sk)->listen_smc;
 913
 914         spin_lock(&par->accept_q_lock);
 915         list_del_init(&smc_sk(sk)->accept_q);
 916         spin_unlock(&par->accept_q_lock);
 917         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
 918         sock_put(sk); /* sock_hold in smc_accept_enqueue */
 919 }
 920
 921 /* remove a sock from the accept queue to bind it to a new socket created
 922  * for a socket accept call from user space
 923  */
 924 struct sock *smc_accept_dequeue(struct sock *parent,
 925                                 struct socket *new_sock)
 926 {
 927         struct smc_sock *isk, *n;
 928         struct sock *new_sk;
 929
 930         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
 931                 new_sk = (struct sock *)isk;
 932
 933                 smc_accept_unlink(new_sk);
 934                 if (new_sk->sk_state == SMC_CLOSED) {
 935                         if (isk->clcsock) {
 936                                 sock_release(isk->clcsock);
 937                                 isk->clcsock = NULL;
 938                         }
 939                         new_sk->sk_prot->unhash(new_sk);
 940                         sock_put(new_sk); /* final */
 941                         continue;
 942                 }
 943                 if (new_sock)
 944                         sock_graft(new_sk, new_sock);
 945                 return new_sk;
 946         }
 947         return NULL;
 948 }
 949
 950 /* clean up for a created but never accepted sock */
 951 void smc_close_non_accepted(struct sock *sk)
 952 {
 953         struct smc_sock *smc = smc_sk(sk);
 954
 955         lock_sock(sk);
 956         if (!sk->sk_lingertime)
 957                 /* wait for peer closing */
 958                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
 959         if (!smc->use_fallback) {
 960                 smc_close_active(smc);
 961                 sock_set_flag(sk, SOCK_DEAD);
 962                 sk->sk_shutdown |= SHUTDOWN_MASK;
 963         }
 964         if (smc->clcsock) {
 965                 struct socket *tcp;
 966
 967                 tcp = smc->clcsock;
 968                 smc->clcsock = NULL;
 969                 sock_release(tcp);
 970         }
 971         if (smc->use_fallback) {
 972                 sock_put(sk); /* passive closing */
 973                 sk->sk_state = SMC_CLOSED;
 974         } else {
 975                 if (sk->sk_state == SMC_CLOSED)
 976                         smc_conn_free(&smc->conn);
 977         }
 978         release_sock(sk);
 979         sk->sk_prot->unhash(sk);
 980         sock_put(sk); /* final sock_put */
 981 }
 982
 983 static int smc_serv_conf_first_link(struct smc_sock *smc)
 984 {
 985         struct net *net = sock_net(smc->clcsock->sk);
 986         struct smc_link_group *lgr = smc->conn.lgr;
 987         struct smc_link *link;
 988         int rest;
 989         int rc;
 990
 991         link = &lgr->lnk[SMC_SINGLE_LINK];
 992
 993         if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
 994                 return SMC_CLC_DECL_ERR_REGRMB;
 995
 996         /* send CONFIRM LINK request to client over the RoCE fabric */
 997         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
 998         if (rc < 0)
 999                 return SMC_CLC_DECL_TIMEOUT_CL;
1000
1001         /* receive CONFIRM LINK response from client over the RoCE fabric */
1002         rest = wait_for_completion_interruptible_timeout(
1003                 &link->llc_confirm_resp,
1004                 SMC_LLC_WAIT_FIRST_TIME);
1005         if (rest <= 0) {
1006                 struct smc_clc_msg_decline dclc;
1007
1008                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1009                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1010                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1011         }
1012
1013         if (link->llc_confirm_resp_rc)
1014                 return SMC_CLC_DECL_RMBE_EC;
1015
1016         /* send ADD LINK request to client over the RoCE fabric */
1017         rc = smc_llc_send_add_link(link,
1018                                    link->smcibdev->mac[link->ibport - 1],
1019                                    link->gid, SMC_LLC_REQ);
1020         if (rc < 0)
1021                 return SMC_CLC_DECL_TIMEOUT_AL;
1022
1023         /* receive ADD LINK response from client over the RoCE fabric */
1024         rest = wait_for_completion_interruptible_timeout(&link->llc_add_resp,
1025                                                          SMC_LLC_WAIT_TIME);
1026         if (rest <= 0) {
1027                 struct smc_clc_msg_decline dclc;
1028
1029                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1030                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1031                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_AL : rc;
1032         }
1033
1034         smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
1035
1036         return 0;
1037 }
1038
1039 /* listen worker: finish */
1040 static void smc_listen_out(struct smc_sock *new_smc)
1041 {
1042         struct smc_sock *lsmc = new_smc->listen_smc;
1043         struct sock *newsmcsk = &new_smc->sk;
1044
1045         lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1046         if (lsmc->sk.sk_state == SMC_LISTEN) {
1047                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1048         } else { /* no longer listening */
1049                 smc_close_non_accepted(newsmcsk);
1050         }
1051         release_sock(&lsmc->sk);
1052
1053         /* Wake up accept */
1054         lsmc->sk.sk_data_ready(&lsmc->sk);
1055         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1056 }
1057
1058 /* listen worker: finish in state connected */
1059 static void smc_listen_out_connected(struct smc_sock *new_smc)
1060 {
1061         struct sock *newsmcsk = &new_smc->sk;
1062
1063         sk_refcnt_debug_inc(newsmcsk);
1064         if (newsmcsk->sk_state == SMC_INIT)
1065                 newsmcsk->sk_state = SMC_ACTIVE;
1066
1067         smc_listen_out(new_smc);
1068 }
1069
1070 /* listen worker: finish in error state */
1071 static void smc_listen_out_err(struct smc_sock *new_smc)
1072 {
1073         struct sock *newsmcsk = &new_smc->sk;
1074
1075         if (newsmcsk->sk_state == SMC_INIT)
1076                 sock_put(&new_smc->sk); /* passive closing */
1077         newsmcsk->sk_state = SMC_CLOSED;
1078         smc_conn_free(&new_smc->conn);
1079
1080         smc_listen_out(new_smc);
1081 }
1082
1083 /* listen worker: decline and fall back if possible */
1084 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1085                                int local_contact)
1086 {
1087         /* RDMA setup failed, switch back to TCP */
1088         if (local_contact == SMC_FIRST_CONTACT)
1089                 smc_lgr_forget(new_smc->conn.lgr);
1090         if (reason_code < 0) { /* error, no fallback possible */
1091                 smc_listen_out_err(new_smc);
1092                 return;
1093         }
1094         smc_conn_free(&new_smc->conn);
1095         new_smc->use_fallback = true;
1096         new_smc->fallback_rsn = reason_code;
1097         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1098                 if (smc_clc_send_decline(new_smc, reason_code) < 0) {
1099                         smc_listen_out_err(new_smc);
1100                         return;
1101                 }
1102         }
1103         smc_listen_out_connected(new_smc);
1104 }
1105
1106 /* listen worker: check prefixes */
1107 static int smc_listen_rdma_check(struct smc_sock *new_smc,
1108                                  struct smc_clc_msg_proposal *pclc)
1109 {
1110         struct smc_clc_msg_proposal_prefix *pclc_prfx;
1111         struct socket *newclcsock = new_smc->clcsock;
1112
1113         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1114         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1115                 return SMC_CLC_DECL_CNFERR;
1116
1117         return 0;
1118 }
1119
1120 /* listen worker: initialize connection and buffers */
1121 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1122                                 struct smc_clc_msg_proposal *pclc,
1123                                 struct smc_ib_device *ibdev, u8 ibport,
1124                                 int *local_contact)
1125 {
1126         /* allocate connection / link group */
1127         *local_contact = smc_conn_create(new_smc, false, 0, ibdev, ibport, 0,
1128                                          &pclc->lcl, NULL, 0);
1129         if (*local_contact < 0) {
1130                 if (*local_contact == -ENOMEM)
1131                         return SMC_CLC_DECL_MEM;/* insufficient memory*/
1132                 return SMC_CLC_DECL_INTERR; /* other error */
1133         }
1134
1135         /* create send buffer and rmb */
1136         if (smc_buf_create(new_smc, false))
1137                 return SMC_CLC_DECL_MEM;
1138
1139         return 0;
1140 }
1141
1142 /* listen worker: initialize connection and buffers for SMC-D */
1143 static int smc_listen_ism_init(struct smc_sock *new_smc,
1144                                struct smc_clc_msg_proposal *pclc,
1145                                struct smcd_dev *ismdev,
1146                                int *local_contact)
1147 {
1148         struct smc_clc_msg_smcd *pclc_smcd;
1149
1150         pclc_smcd = smc_get_clc_msg_smcd(pclc);
1151         *local_contact = smc_conn_create(new_smc, true, 0, NULL, 0, 0, NULL,
1152                                          ismdev, pclc_smcd->gid);
1153         if (*local_contact < 0) {
1154                 if (*local_contact == -ENOMEM)
1155                         return SMC_CLC_DECL_MEM;/* insufficient memory*/
1156                 return SMC_CLC_DECL_INTERR; /* other error */
1157         }
1158
1159         /* Check if peer can be reached via ISM device */
1160         if (smc_ism_cantalk(new_smc->conn.lgr->peer_gid,
1161                             new_smc->conn.lgr->vlan_id,
1162                             new_smc->conn.lgr->smcd)) {
1163                 if (*local_contact == SMC_FIRST_CONTACT)
1164                         smc_lgr_forget(new_smc->conn.lgr);
1165                 smc_conn_free(&new_smc->conn);
1166                 return SMC_CLC_DECL_CNFERR;
1167         }
1168
1169         /* Create send and receive buffers */
1170         if (smc_buf_create(new_smc, true)) {
1171                 if (*local_contact == SMC_FIRST_CONTACT)
1172                         smc_lgr_forget(new_smc->conn.lgr);
1173                 smc_conn_free(&new_smc->conn);
1174                 return SMC_CLC_DECL_MEM;
1175         }
1176
1177         return 0;
1178 }
1179
1180 /* listen worker: register buffers */
1181 static int smc_listen_rdma_reg(struct smc_sock *new_smc, int local_contact)
1182 {
1183         struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
1184
1185         if (local_contact != SMC_FIRST_CONTACT) {
1186                 if (smc_reg_rmb(link, new_smc->conn.rmb_desc, true))
1187                         return SMC_CLC_DECL_ERR_REGRMB;
1188         }
1189         smc_rmb_sync_sg_for_device(&new_smc->conn);
1190
1191         return 0;
1192 }
1193
1194 /* listen worker: finish RDMA setup */
1195 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1196                                   struct smc_clc_msg_accept_confirm *cclc,
1197                                   int local_contact)
1198 {
1199         struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
1200         int reason_code = 0;
1201
1202         if (local_contact == SMC_FIRST_CONTACT)
1203                 smc_link_save_peer_info(link, cclc);
1204
1205         if (smc_rmb_rtoken_handling(&new_smc->conn, cclc)) {
1206                 reason_code = SMC_CLC_DECL_ERR_RTOK;
1207                 goto decline;
1208         }
1209
1210         if (local_contact == SMC_FIRST_CONTACT) {
1211                 if (smc_ib_ready_link(link)) {
1212                         reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1213                         goto decline;
1214                 }
1215                 /* QP confirmation over RoCE fabric */
1216                 reason_code = smc_serv_conf_first_link(new_smc);
1217                 if (reason_code)
1218                         goto decline;
1219         }
1220         return 0;
1221
1222 decline:
1223         smc_listen_decline(new_smc, reason_code, local_contact);
1224         return reason_code;
1225 }
1226
1227 /* setup for RDMA connection of server */
1228 static void smc_listen_work(struct work_struct *work)
1229 {
1230         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1231                                                 smc_listen_work);
1232         struct socket *newclcsock = new_smc->clcsock;
1233         struct smc_clc_msg_accept_confirm cclc;
1234         struct smc_clc_msg_proposal *pclc;
1235         struct smc_ib_device *ibdev;
1236         bool ism_supported = false;
1237         struct smcd_dev *ismdev;
1238         u8 buf[SMC_CLC_MAX_LEN];
1239         int local_contact = 0;
1240         unsigned short vlan;
1241         int reason_code = 0;
1242         int rc = 0;
1243         u8 ibport;
1244
1245         if (new_smc->use_fallback) {
1246                 smc_listen_out_connected(new_smc);
1247                 return;
1248         }
1249
1250         /* check if peer is smc capable */
1251         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1252                 new_smc->use_fallback = true;
1253                 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1254                 smc_listen_out_connected(new_smc);
1255                 return;
1256         }
1257
1258         /* do inband token exchange -
1259          * wait for and receive SMC Proposal CLC message
1260          */
1261         pclc = (struct smc_clc_msg_proposal *)&buf;
1262         reason_code = smc_clc_wait_msg(new_smc, pclc, SMC_CLC_MAX_LEN,
1263                                        SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1264         if (reason_code) {
1265                 smc_listen_decline(new_smc, reason_code, 0);
1266                 return;
1267         }
1268
1269         /* IPSec connections opt out of SMC-R optimizations */
1270         if (using_ipsec(new_smc)) {
1271                 smc_listen_decline(new_smc, SMC_CLC_DECL_IPSEC, 0);
1272                 return;
1273         }
1274
1275         mutex_lock(&smc_server_lgr_pending);
1276         smc_close_init(new_smc);
1277         smc_rx_init(new_smc);
1278         smc_tx_init(new_smc);
1279
1280         /* check if ISM is available */
1281         if ((pclc->hdr.path == SMC_TYPE_D || pclc->hdr.path == SMC_TYPE_B) &&
1282             !smc_check_ism(new_smc, &ismdev) &&
1283             !smc_listen_ism_init(new_smc, pclc, ismdev, &local_contact)) {
1284                 ism_supported = true;
1285         }
1286
1287         /* check if RDMA is available */
1288         if (!ism_supported &&
1289             ((pclc->hdr.path != SMC_TYPE_R && pclc->hdr.path != SMC_TYPE_B) ||
1290              smc_vlan_by_tcpsk(new_smc->clcsock, &vlan) ||
1291              smc_check_rdma(new_smc, &ibdev, &ibport, vlan, NULL) ||
1292              smc_listen_rdma_check(new_smc, pclc) ||
1293              smc_listen_rdma_init(new_smc, pclc, ibdev, ibport,
1294                                   &local_contact) ||
1295              smc_listen_rdma_reg(new_smc, local_contact))) {
1296                 /* SMC not supported, decline */
1297                 mutex_unlock(&smc_server_lgr_pending);
1298                 smc_listen_decline(new_smc, SMC_CLC_DECL_MODEUNSUPP,
1299                                    local_contact);
1300                 return;
1301         }
1302
1303         /* send SMC Accept CLC message */
1304         rc = smc_clc_send_accept(new_smc, local_contact);
1305         if (rc) {
1306                 mutex_unlock(&smc_server_lgr_pending);
1307                 smc_listen_decline(new_smc, rc, local_contact);
1308                 return;
1309         }
1310
1311         /* SMC-D does not need this lock any more */
1312         if (ism_supported)
1313                 mutex_unlock(&smc_server_lgr_pending);
1314
1315         /* receive SMC Confirm CLC message */
1316         reason_code = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
1317                                        SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1318         if (reason_code) {
1319                 if (!ism_supported)
1320                         mutex_unlock(&smc_server_lgr_pending);
1321                 smc_listen_decline(new_smc, reason_code, local_contact);
1322                 return;
1323         }
1324
1325         /* finish worker */
1326         if (!ism_supported) {
1327                 rc = smc_listen_rdma_finish(new_smc, &cclc, local_contact);
1328                 mutex_unlock(&smc_server_lgr_pending);
1329                 if (rc)
1330                         return;
1331         }
1332         smc_conn_save_peer_info(new_smc, &cclc);
1333         smc_listen_out_connected(new_smc);
1334 }
1335
1336 static void smc_tcp_listen_work(struct work_struct *work)
1337 {
1338         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1339                                              tcp_listen_work);
1340         struct sock *lsk = &lsmc->sk;
1341         struct smc_sock *new_smc;
1342         int rc = 0;
1343
1344         lock_sock(lsk);
1345         while (lsk->sk_state == SMC_LISTEN) {
1346                 rc = smc_clcsock_accept(lsmc, &new_smc);
1347                 if (rc)
1348                         goto out;
1349                 if (!new_smc)
1350                         continue;
1351
1352                 new_smc->listen_smc = lsmc;
1353                 new_smc->use_fallback = lsmc->use_fallback;
1354                 new_smc->fallback_rsn = lsmc->fallback_rsn;
1355                 sock_hold(lsk); /* sock_put in smc_listen_work */
1356                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1357                 smc_copy_sock_settings_to_smc(new_smc);
1358                 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1359                 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1360                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1361                 if (!schedule_work(&new_smc->smc_listen_work))
1362                         sock_put(&new_smc->sk);
1363         }
1364
1365 out:
1366         release_sock(lsk);
1367         sock_put(&lsmc->sk); /* sock_hold in smc_listen */
1368 }
1369
1370 static int smc_listen(struct socket *sock, int backlog)
1371 {
1372         struct sock *sk = sock->sk;
1373         struct smc_sock *smc;
1374         int rc;
1375
1376         smc = smc_sk(sk);
1377         lock_sock(sk);
1378
1379         rc = -EINVAL;
1380         if ((sk->sk_state != SMC_INIT) && (sk->sk_state != SMC_LISTEN))
1381                 goto out;
1382
1383         rc = 0;
1384         if (sk->sk_state == SMC_LISTEN) {
1385                 sk->sk_max_ack_backlog = backlog;
1386                 goto out;
1387         }
1388         /* some socket options are handled in core, so we could not apply
1389          * them to the clc socket -- copy smc socket options to clc socket
1390          */
1391         smc_copy_sock_settings_to_clc(smc);
1392         if (!smc->use_fallback)
1393                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1394
1395         rc = kernel_listen(smc->clcsock, backlog);
1396         if (rc)
1397                 goto out;
1398         sk->sk_max_ack_backlog = backlog;
1399         sk->sk_ack_backlog = 0;
1400         sk->sk_state = SMC_LISTEN;
1401         sock_hold(sk); /* sock_hold in tcp_listen_worker */
1402         if (!schedule_work(&smc->tcp_listen_work))
1403                 sock_put(sk);
1404
1405 out:
1406         release_sock(sk);
1407         return rc;
1408 }
1409
1410 static int smc_accept(struct socket *sock, struct socket *new_sock,
1411                       int flags, bool kern)
1412 {
1413         struct sock *sk = sock->sk, *nsk;
1414         DECLARE_WAITQUEUE(wait, current);
1415         struct smc_sock *lsmc;
1416         long timeo;
1417         int rc = 0;
1418
1419         lsmc = smc_sk(sk);
1420         sock_hold(sk); /* sock_put below */
1421         lock_sock(sk);
1422
1423         if (lsmc->sk.sk_state != SMC_LISTEN) {
1424                 rc = -EINVAL;
1425                 release_sock(sk);
1426                 goto out;
1427         }
1428
1429         /* Wait for an incoming connection */
1430         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1431         add_wait_queue_exclusive(sk_sleep(sk), &wait);
1432         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1433                 set_current_state(TASK_INTERRUPTIBLE);
1434                 if (!timeo) {
1435                         rc = -EAGAIN;
1436                         break;
1437                 }
1438                 release_sock(sk);
1439                 timeo = schedule_timeout(timeo);
1440                 /* wakeup by sk_data_ready in smc_listen_work() */
1441                 sched_annotate_sleep();
1442                 lock_sock(sk);
1443                 if (signal_pending(current)) {
1444                         rc = sock_intr_errno(timeo);
1445                         break;
1446                 }
1447         }
1448         set_current_state(TASK_RUNNING);
1449         remove_wait_queue(sk_sleep(sk), &wait);
1450
1451         if (!rc)
1452                 rc = sock_error(nsk);
1453         release_sock(sk);
1454         if (rc)
1455                 goto out;
1456
1457         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1458                 /* wait till data arrives on the socket */
1459                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1460                                                                 MSEC_PER_SEC);
1461                 if (smc_sk(nsk)->use_fallback) {
1462                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1463
1464                         lock_sock(clcsk);
1465                         if (skb_queue_empty(&clcsk->sk_receive_queue))
1466                                 sk_wait_data(clcsk, &timeo, NULL);
1467                         release_sock(clcsk);
1468                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1469                         lock_sock(nsk);
1470                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1471                         release_sock(nsk);
1472                 }
1473         }
1474
1475 out:
1476         sock_put(sk); /* sock_hold above */
1477         return rc;
1478 }
1479
1480 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1481                        int peer)
1482 {
1483         struct smc_sock *smc;
1484
1485         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1486             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1487                 return -ENOTCONN;
1488
1489         smc = smc_sk(sock->sk);
1490
1491         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1492 }
1493
1494 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1495 {
1496         struct sock *sk = sock->sk;
1497         struct smc_sock *smc;
1498         int rc = -EPIPE;
1499
1500         smc = smc_sk(sk);
1501         lock_sock(sk);
1502         if ((sk->sk_state != SMC_ACTIVE) &&
1503             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1504             (sk->sk_state != SMC_INIT))
1505                 goto out;
1506
1507         if (msg->msg_flags & MSG_FASTOPEN) {
1508                 if (sk->sk_state == SMC_INIT) {
1509                         smc->use_fallback = true;
1510                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1511                 } else {
1512                         rc = -EINVAL;
1513                         goto out;
1514                 }
1515         }
1516
1517         if (smc->use_fallback)
1518                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1519         else
1520                 rc = smc_tx_sendmsg(smc, msg, len);
1521 out:
1522         release_sock(sk);
1523         return rc;
1524 }
1525
1526 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1527                        int flags)
1528 {
1529         struct sock *sk = sock->sk;
1530         struct smc_sock *smc;
1531         int rc = -ENOTCONN;
1532
1533         smc = smc_sk(sk);
1534         lock_sock(sk);
1535         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1536                 /* socket was connected before, no more data to read */
1537                 rc = 0;
1538                 goto out;
1539         }
1540         if ((sk->sk_state == SMC_INIT) ||
1541             (sk->sk_state == SMC_LISTEN) ||
1542             (sk->sk_state == SMC_CLOSED))
1543                 goto out;
1544
1545         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1546                 rc = 0;
1547                 goto out;
1548         }
1549
1550         if (smc->use_fallback) {
1551                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1552         } else {
1553                 msg->msg_namelen = 0;
1554                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1555         }
1556
1557 out:
1558         release_sock(sk);
1559         return rc;
1560 }
1561
1562 static __poll_t smc_accept_poll(struct sock *parent)
1563 {
1564         struct smc_sock *isk = smc_sk(parent);
1565         __poll_t mask = 0;
1566
1567         spin_lock(&isk->accept_q_lock);
1568         if (!list_empty(&isk->accept_q))
1569                 mask = EPOLLIN | EPOLLRDNORM;
1570         spin_unlock(&isk->accept_q_lock);
1571
1572         return mask;
1573 }
1574
1575 static __poll_t smc_poll(struct file *file, struct socket *sock,
1576                              poll_table *wait)
1577 {
1578         struct sock *sk = sock->sk;
1579         struct smc_sock *smc;
1580         __poll_t mask = 0;
1581
1582         if (!sk)
1583                 return EPOLLNVAL;
1584
1585         smc = smc_sk(sock->sk);
1586         if (smc->use_fallback) {
1587                 /* delegate to CLC child sock */
1588                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1589                 sk->sk_err = smc->clcsock->sk->sk_err;
1590         } else {
1591                 if (sk->sk_state != SMC_CLOSED)
1592                         sock_poll_wait(file, sock, wait);
1593                 if (sk->sk_err)
1594                         mask |= EPOLLERR;
1595                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1596                     (sk->sk_state == SMC_CLOSED))
1597                         mask |= EPOLLHUP;
1598                 if (sk->sk_state == SMC_LISTEN) {
1599                         /* woken up by sk_data_ready in smc_listen_work() */
1600                         mask |= smc_accept_poll(sk);
1601                 } else if (smc->use_fallback) { /* as result of connect_work()*/
1602                         mask |= smc->clcsock->ops->poll(file, smc->clcsock,
1603                                                            wait);
1604                         sk->sk_err = smc->clcsock->sk->sk_err;
1605                 } else {
1606                         if ((sk->sk_state != SMC_INIT &&
1607                              atomic_read(&smc->conn.sndbuf_space)) ||
1608                             sk->sk_shutdown & SEND_SHUTDOWN) {
1609                                 mask |= EPOLLOUT | EPOLLWRNORM;
1610                         } else {
1611                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1612                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1613                         }
1614                         if (atomic_read(&smc->conn.bytes_to_rcv))
1615                                 mask |= EPOLLIN | EPOLLRDNORM;
1616                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1617                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1618                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
1619                                 mask |= EPOLLIN;
1620                         if (smc->conn.urg_state == SMC_URG_VALID)
1621                                 mask |= EPOLLPRI;
1622                 }
1623         }
1624
1625         return mask;
1626 }
1627
1628 static int smc_shutdown(struct socket *sock, int how)
1629 {
1630         struct sock *sk = sock->sk;
1631         struct smc_sock *smc;
1632         int rc = -EINVAL;
1633         int rc1 = 0;
1634
1635         smc = smc_sk(sk);
1636
1637         if ((how < SHUT_RD) || (how > SHUT_RDWR))
1638                 return rc;
1639
1640         lock_sock(sk);
1641
1642         rc = -ENOTCONN;
1643         if ((sk->sk_state != SMC_ACTIVE) &&
1644             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1645             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1646             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1647             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1648             (sk->sk_state != SMC_APPFINCLOSEWAIT))
1649                 goto out;
1650         if (smc->use_fallback) {
1651                 rc = kernel_sock_shutdown(smc->clcsock, how);
1652                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1653                 if (sk->sk_shutdown == SHUTDOWN_MASK)
1654                         sk->sk_state = SMC_CLOSED;
1655                 goto out;
1656         }
1657         switch (how) {
1658         case SHUT_RDWR:         /* shutdown in both directions */
1659                 rc = smc_close_active(smc);
1660                 break;
1661         case SHUT_WR:
1662                 rc = smc_close_shutdown_write(smc);
1663                 break;
1664         case SHUT_RD:
1665                 rc = 0;
1666                 /* nothing more to do because peer is not involved */
1667                 break;
1668         }
1669         if (smc->clcsock)
1670                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1671         /* map sock_shutdown_cmd constants to sk_shutdown value range */
1672         sk->sk_shutdown |= how + 1;
1673
1674 out:
1675         release_sock(sk);
1676         return rc ? rc : rc1;
1677 }
1678
1679 static int smc_setsockopt(struct socket *sock, int level, int optname,
1680                           char __user *optval, unsigned int optlen)
1681 {
1682         struct sock *sk = sock->sk;
1683         struct smc_sock *smc;
1684         int val, rc;
1685
1686         smc = smc_sk(sk);
1687
1688         /* generic setsockopts reaching us here always apply to the
1689          * CLC socket
1690          */
1691         rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1692                                            optval, optlen);
1693         if (smc->clcsock->sk->sk_err) {
1694                 sk->sk_err = smc->clcsock->sk->sk_err;
1695                 sk->sk_error_report(sk);
1696         }
1697         if (rc)
1698                 return rc;
1699
1700         if (optlen < sizeof(int))
1701                 return -EINVAL;
1702         if (get_user(val, (int __user *)optval))
1703                 return -EFAULT;
1704
1705         lock_sock(sk);
1706         switch (optname) {
1707         case TCP_ULP:
1708         case TCP_FASTOPEN:
1709         case TCP_FASTOPEN_CONNECT:
1710         case TCP_FASTOPEN_KEY:
1711         case TCP_FASTOPEN_NO_COOKIE:
1712                 /* option not supported by SMC */
1713                 if (sk->sk_state == SMC_INIT) {
1714                         smc->use_fallback = true;
1715                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1716                 } else {
1717                         if (!smc->use_fallback)
1718                                 rc = -EINVAL;
1719                 }
1720                 break;
1721         case TCP_NODELAY:
1722                 if (sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) {
1723                         if (val && !smc->use_fallback)
1724                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1725                                                  0);
1726                 }
1727                 break;
1728         case TCP_CORK:
1729                 if (sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) {
1730                         if (!val && !smc->use_fallback)
1731                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1732                                                  0);
1733                 }
1734                 break;
1735         case TCP_DEFER_ACCEPT:
1736                 smc->sockopt_defer_accept = val;
1737                 break;
1738         default:
1739                 break;
1740         }
1741         release_sock(sk);
1742
1743         return rc;
1744 }
1745
1746 static int smc_getsockopt(struct socket *sock, int level, int optname,
1747                           char __user *optval, int __user *optlen)
1748 {
1749         struct smc_sock *smc;
1750
1751         smc = smc_sk(sock->sk);
1752         /* socket options apply to the CLC socket */
1753         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1754                                              optval, optlen);
1755 }
1756
1757 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1758                      unsigned long arg)
1759 {
1760         union smc_host_cursor cons, urg;
1761         struct smc_connection *conn;
1762         struct smc_sock *smc;
1763         int answ;
1764
1765         smc = smc_sk(sock->sk);
1766         conn = &smc->conn;
1767         lock_sock(&smc->sk);
1768         if (smc->use_fallback) {
1769                 if (!smc->clcsock) {
1770                         release_sock(&smc->sk);
1771                         return -EBADF;
1772                 }
1773                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1774                 release_sock(&smc->sk);
1775                 return answ;
1776         }
1777         switch (cmd) {
1778         case SIOCINQ: /* same as FIONREAD */
1779                 if (smc->sk.sk_state == SMC_LISTEN) {
1780                         release_sock(&smc->sk);
1781                         return -EINVAL;
1782                 }
1783                 if (smc->sk.sk_state == SMC_INIT ||
1784                     smc->sk.sk_state == SMC_CLOSED)
1785                         answ = 0;
1786                 else
1787                         answ = atomic_read(&smc->conn.bytes_to_rcv);
1788                 break;
1789         case SIOCOUTQ:
1790                 /* output queue size (not send + not acked) */
1791                 if (smc->sk.sk_state == SMC_LISTEN) {
1792                         release_sock(&smc->sk);
1793                         return -EINVAL;
1794                 }
1795                 if (smc->sk.sk_state == SMC_INIT ||
1796                     smc->sk.sk_state == SMC_CLOSED)
1797                         answ = 0;
1798                 else
1799                         answ = smc->conn.sndbuf_desc->len -
1800                                         atomic_read(&smc->conn.sndbuf_space);
1801                 break;
1802         case SIOCOUTQNSD:
1803                 /* output queue size (not send only) */
1804                 if (smc->sk.sk_state == SMC_LISTEN) {
1805                         release_sock(&smc->sk);
1806                         return -EINVAL;
1807                 }
1808                 if (smc->sk.sk_state == SMC_INIT ||
1809                     smc->sk.sk_state == SMC_CLOSED)
1810                         answ = 0;
1811                 else
1812                         answ = smc_tx_prepared_sends(&smc->conn);
1813                 break;
1814         case SIOCATMARK:
1815                 if (smc->sk.sk_state == SMC_LISTEN) {
1816                         release_sock(&smc->sk);
1817                         return -EINVAL;
1818                 }
1819                 if (smc->sk.sk_state == SMC_INIT ||
1820                     smc->sk.sk_state == SMC_CLOSED) {
1821                         answ = 0;
1822                 } else {
1823                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
1824                         smc_curs_copy(&urg, &conn->urg_curs, conn);
1825                         answ = smc_curs_diff(conn->rmb_desc->len,
1826                                              &cons, &urg) == 1;
1827                 }
1828                 break;
1829         default:
1830                 release_sock(&smc->sk);
1831                 return -ENOIOCTLCMD;
1832         }
1833         release_sock(&smc->sk);
1834
1835         return put_user(answ, (int __user *)arg);
1836 }
1837
1838 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1839                             int offset, size_t size, int flags)
1840 {
1841         struct sock *sk = sock->sk;
1842         struct smc_sock *smc;
1843         int rc = -EPIPE;
1844
1845         smc = smc_sk(sk);
1846         lock_sock(sk);
1847         if (sk->sk_state != SMC_ACTIVE) {
1848                 release_sock(sk);
1849                 goto out;
1850         }
1851         release_sock(sk);
1852         if (smc->use_fallback)
1853                 rc = kernel_sendpage(smc->clcsock, page, offset,
1854                                      size, flags);
1855         else
1856                 rc = sock_no_sendpage(sock, page, offset, size, flags);
1857
1858 out:
1859         return rc;
1860 }
1861
1862 /* Map the affected portions of the rmbe into an spd, note the number of bytes
1863  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
1864  * updates till whenever a respective page has been fully processed.
1865  * Note that subsequent recv() calls have to wait till all splice() processing
1866  * completed.
1867  */
1868 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1869                                struct pipe_inode_info *pipe, size_t len,
1870                                unsigned int flags)
1871 {
1872         struct sock *sk = sock->sk;
1873         struct smc_sock *smc;
1874         int rc = -ENOTCONN;
1875
1876         smc = smc_sk(sk);
1877         lock_sock(sk);
1878         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1879                 /* socket was connected before, no more data to read */
1880                 rc = 0;
1881                 goto out;
1882         }
1883         if (sk->sk_state == SMC_INIT ||
1884             sk->sk_state == SMC_LISTEN ||
1885             sk->sk_state == SMC_CLOSED)
1886                 goto out;
1887
1888         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1889                 rc = 0;
1890                 goto out;
1891         }
1892
1893         if (smc->use_fallback) {
1894                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1895                                                     pipe, len, flags);
1896         } else {
1897                 if (*ppos) {
1898                         rc = -ESPIPE;
1899                         goto out;
1900                 }
1901                 if (flags & SPLICE_F_NONBLOCK)
1902                         flags = MSG_DONTWAIT;
1903                 else
1904                         flags = 0;
1905                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
1906         }
1907 out:
1908         release_sock(sk);
1909
1910         return rc;
1911 }
1912
1913 /* must look like tcp */
1914 static const struct proto_ops smc_sock_ops = {
1915         .family         = PF_SMC,
1916         .owner          = THIS_MODULE,
1917         .release        = smc_release,
1918         .bind           = smc_bind,
1919         .connect        = smc_connect,
1920         .socketpair     = sock_no_socketpair,
1921         .accept         = smc_accept,
1922         .getname        = smc_getname,
1923         .poll           = smc_poll,
1924         .ioctl          = smc_ioctl,
1925         .listen         = smc_listen,
1926         .shutdown       = smc_shutdown,
1927         .setsockopt     = smc_setsockopt,
1928         .getsockopt     = smc_getsockopt,
1929         .sendmsg        = smc_sendmsg,
1930         .recvmsg        = smc_recvmsg,
1931         .mmap           = sock_no_mmap,
1932         .sendpage       = smc_sendpage,
1933         .splice_read    = smc_splice_read,
1934 };
1935
1936 static int smc_create(struct net *net, struct socket *sock, int protocol,
1937                       int kern)
1938 {
1939         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
1940         struct smc_sock *smc;
1941         struct sock *sk;
1942         int rc;
1943
1944         rc = -ESOCKTNOSUPPORT;
1945         if (sock->type != SOCK_STREAM)
1946                 goto out;
1947
1948         rc = -EPROTONOSUPPORT;
1949         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
1950                 goto out;
1951
1952         rc = -ENOBUFS;
1953         sock->ops = &smc_sock_ops;
1954         sk = smc_sock_alloc(net, sock, protocol);
1955         if (!sk)
1956                 goto out;
1957
1958         /* create internal TCP socket for CLC handshake and fallback */
1959         smc = smc_sk(sk);
1960         smc->use_fallback = false; /* assume rdma capability first */
1961         smc->fallback_rsn = 0;
1962         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
1963                               &smc->clcsock);
1964         if (rc) {
1965                 sk_common_release(sk);
1966                 goto out;
1967         }
1968         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
1969         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
1970
1971 out:
1972         return rc;
1973 }
1974
1975 static const struct net_proto_family smc_sock_family_ops = {
1976         .family = PF_SMC,
1977         .owner  = THIS_MODULE,
1978         .create = smc_create,
1979 };
1980
1981 unsigned int smc_net_id;
1982
1983 static __net_init int smc_net_init(struct net *net)
1984 {
1985         return smc_pnet_net_init(net);
1986 }
1987
1988 static void __net_exit smc_net_exit(struct net *net)
1989 {
1990         smc_pnet_net_exit(net);
1991 }
1992
1993 static struct pernet_operations smc_net_ops = {
1994         .init = smc_net_init,
1995         .exit = smc_net_exit,
1996         .id   = &smc_net_id,
1997         .size = sizeof(struct smc_net),
1998 };
1999
2000 static int __init smc_init(void)
2001 {
2002         int rc;
2003
2004         rc = register_pernet_subsys(&smc_net_ops);
2005         if (rc)
2006                 return rc;
2007
2008         rc = smc_pnet_init();
2009         if (rc)
2010                 return rc;
2011
2012         rc = smc_llc_init();
2013         if (rc) {
2014                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2015                 goto out_pnet;
2016         }
2017
2018         rc = smc_cdc_init();
2019         if (rc) {
2020                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2021                 goto out_pnet;
2022         }
2023
2024         rc = proto_register(&smc_proto, 1);
2025         if (rc) {
2026                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2027                 goto out_pnet;
2028         }
2029
2030         rc = proto_register(&smc_proto6, 1);
2031         if (rc) {
2032                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2033                 goto out_proto;
2034         }
2035
2036         rc = sock_register(&smc_sock_family_ops);
2037         if (rc) {
2038                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2039                 goto out_proto6;
2040         }
2041         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2042         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2043
2044         rc = smc_ib_register_client();
2045         if (rc) {
2046                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2047                 goto out_sock;
2048         }
2049
2050         static_branch_enable(&tcp_have_smc);
2051         return 0;
2052
2053 out_sock:
2054         sock_unregister(PF_SMC);
2055 out_proto6:
2056         proto_unregister(&smc_proto6);
2057 out_proto:
2058         proto_unregister(&smc_proto);
2059 out_pnet:
2060         smc_pnet_exit();
2061         return rc;
2062 }
2063
2064 static void __exit smc_exit(void)
2065 {
2066         smc_core_exit();
2067         static_branch_disable(&tcp_have_smc);
2068         smc_ib_unregister_client();
2069         sock_unregister(PF_SMC);
2070         proto_unregister(&smc_proto6);
2071         proto_unregister(&smc_proto);
2072         smc_pnet_exit();
2073         unregister_pernet_subsys(&smc_net_ops);
2074 }
2075
2076 module_init(smc_init);
2077 module_exit(smc_exit);
2078
2079 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2080 MODULE_DESCRIPTION("smc socket address family");
2081 MODULE_LICENSE("GPL");
2082 MODULE_ALIAS_NETPROTO(PF_SMC);