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
  29 #include <net/sock.h>
  30 #include <net/tcp.h>
  31 #include <net/smc.h>
  32 #include <asm/ioctls.h>
  33
  34 #include <net/net_namespace.h>
  35 #include <net/netns/generic.h>
  36 #include "smc_netns.h"
  37
  38 #include "smc.h"
  39 #include "smc_clc.h"
  40 #include "smc_llc.h"
  41 #include "smc_cdc.h"
  42 #include "smc_core.h"
  43 #include "smc_ib.h"
  44 #include "smc_ism.h"
  45 #include "smc_pnet.h"
  46 #include "smc_tx.h"
  47 #include "smc_rx.h"
  48 #include "smc_close.h"
  49
  50 static DEFINE_MUTEX(smc_server_lgr_pending);    /* serialize link group
  51                                                  * creation on server
  52                                                  */
  53 static DEFINE_MUTEX(smc_client_lgr_pending);    /* serialize link group
  54                                                  * creation on client
  55                                                  */
  56
  57 static void smc_tcp_listen_work(struct work_struct *);
  58 static void smc_connect_work(struct work_struct *);
  59
  60 static void smc_set_keepalive(struct sock *sk, int val)
  61 {
  62         struct smc_sock *smc = smc_sk(sk);
  63
  64         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
  65 }
  66
  67 static struct smc_hashinfo smc_v4_hashinfo = {
  68         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
  69 };
  70
  71 static struct smc_hashinfo smc_v6_hashinfo = {
  72         .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
  73 };
  74
  75 int smc_hash_sk(struct sock *sk)
  76 {
  77         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  78         struct hlist_head *head;
  79
  80         head = &h->ht;
  81
  82         write_lock_bh(&h->lock);
  83         sk_add_node(sk, head);
  84         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
  85         write_unlock_bh(&h->lock);
  86
  87         return 0;
  88 }
  89 EXPORT_SYMBOL_GPL(smc_hash_sk);
  90
  91 void smc_unhash_sk(struct sock *sk)
  92 {
  93         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  94
  95         write_lock_bh(&h->lock);
  96         if (sk_del_node_init(sk))
  97                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
  98         write_unlock_bh(&h->lock);
  99 }
 100 EXPORT_SYMBOL_GPL(smc_unhash_sk);
 101
 102 struct proto smc_proto = {
 103         .name           = "SMC",
 104         .owner          = THIS_MODULE,
 105         .keepalive      = smc_set_keepalive,
 106         .hash           = smc_hash_sk,
 107         .unhash         = smc_unhash_sk,
 108         .obj_size       = sizeof(struct smc_sock),
 109         .h.smc_hash     = &smc_v4_hashinfo,
 110         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 111 };
 112 EXPORT_SYMBOL_GPL(smc_proto);
 113
 114 struct proto smc_proto6 = {
 115         .name           = "SMC6",
 116         .owner          = THIS_MODULE,
 117         .keepalive      = smc_set_keepalive,
 118         .hash           = smc_hash_sk,
 119         .unhash         = smc_unhash_sk,
 120         .obj_size       = sizeof(struct smc_sock),
 121         .h.smc_hash     = &smc_v6_hashinfo,
 122         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 123 };
 124 EXPORT_SYMBOL_GPL(smc_proto6);
 125
 126 static void smc_restore_fallback_changes(struct smc_sock *smc)
 127 {
 128         smc->clcsock->file->private_data = smc->sk.sk_socket;
 129         smc->clcsock->file = NULL;
 130 }
 131
 132 static int __smc_release(struct smc_sock *smc)
 133 {
 134         struct sock *sk = &smc->sk;
 135         int rc = 0;
 136
 137         if (!smc->use_fallback) {
 138                 rc = smc_close_active(smc);
 139                 sock_set_flag(sk, SOCK_DEAD);
 140                 sk->sk_shutdown |= SHUTDOWN_MASK;
 141         } else {
 142                 if (sk->sk_state != SMC_CLOSED) {
 143                         if (sk->sk_state != SMC_LISTEN &&
 144                             sk->sk_state != SMC_INIT)
 145                                 sock_put(sk); /* passive closing */
 146                         if (sk->sk_state == SMC_LISTEN) {
 147                                 /* wake up clcsock accept */
 148                                 rc = kernel_sock_shutdown(smc->clcsock,
 149                                                           SHUT_RDWR);
 150                         }
 151                         sk->sk_state = SMC_CLOSED;
 152                         sk->sk_state_change(sk);
 153                 }
 154                 smc_restore_fallback_changes(smc);
 155         }
 156
 157         sk->sk_prot->unhash(sk);
 158
 159         if (sk->sk_state == SMC_CLOSED) {
 160                 if (smc->clcsock) {
 161                         release_sock(sk);
 162                         smc_clcsock_release(smc);
 163                         lock_sock(sk);
 164                 }
 165                 if (!smc->use_fallback)
 166                         smc_conn_free(&smc->conn);
 167         }
 168
 169         return rc;
 170 }
 171
 172 static int smc_release(struct socket *sock)
 173 {
 174         struct sock *sk = sock->sk;
 175         struct smc_sock *smc;
 176         int rc = 0;
 177
 178         if (!sk)
 179                 goto out;
 180
 181         smc = smc_sk(sk);
 182
 183         /* cleanup for a dangling non-blocking connect */
 184         if (smc->connect_nonblock && sk->sk_state == SMC_INIT)
 185                 tcp_abort(smc->clcsock->sk, ECONNABORTED);
 186
 187         if (cancel_work_sync(&smc->connect_work))
 188                 sock_put(&smc->sk); /* sock_hold in smc_connect for passive closing */
 189
 190         if (sk->sk_state == SMC_LISTEN)
 191                 /* smc_close_non_accepted() is called and acquires
 192                  * sock lock for child sockets again
 193                  */
 194                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 195         else
 196                 lock_sock(sk);
 197
 198         rc = __smc_release(smc);
 199
 200         /* detach socket */
 201         sock_orphan(sk);
 202         sock->sk = NULL;
 203         release_sock(sk);
 204
 205         sock_put(sk); /* final sock_put */
 206 out:
 207         return rc;
 208 }
 209
 210 static void smc_destruct(struct sock *sk)
 211 {
 212         if (sk->sk_state != SMC_CLOSED)
 213                 return;
 214         if (!sock_flag(sk, SOCK_DEAD))
 215                 return;
 216
 217         sk_refcnt_debug_dec(sk);
 218 }
 219
 220 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 221                                    int protocol)
 222 {
 223         struct smc_sock *smc;
 224         struct proto *prot;
 225         struct sock *sk;
 226
 227         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
 228         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
 229         if (!sk)
 230                 return NULL;
 231
 232         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 233         sk->sk_state = SMC_INIT;
 234         sk->sk_destruct = smc_destruct;
 235         sk->sk_protocol = protocol;
 236         smc = smc_sk(sk);
 237         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 238         INIT_WORK(&smc->connect_work, smc_connect_work);
 239         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 240         INIT_LIST_HEAD(&smc->accept_q);
 241         spin_lock_init(&smc->accept_q_lock);
 242         spin_lock_init(&smc->conn.send_lock);
 243         sk->sk_prot->hash(sk);
 244         sk_refcnt_debug_inc(sk);
 245         mutex_init(&smc->clcsock_release_lock);
 246
 247         return sk;
 248 }
 249
 250 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 251                     int addr_len)
 252 {
 253         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 254         struct sock *sk = sock->sk;
 255         struct smc_sock *smc;
 256         int rc;
 257
 258         smc = smc_sk(sk);
 259
 260         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 261         rc = -EINVAL;
 262         if (addr_len < sizeof(struct sockaddr_in))
 263                 goto out;
 264
 265         rc = -EAFNOSUPPORT;
 266         if (addr->sin_family != AF_INET &&
 267             addr->sin_family != AF_INET6 &&
 268             addr->sin_family != AF_UNSPEC)
 269                 goto out;
 270         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 271         if (addr->sin_family == AF_UNSPEC &&
 272             addr->sin_addr.s_addr != htonl(INADDR_ANY))
 273                 goto out;
 274
 275         lock_sock(sk);
 276
 277         /* Check if socket is already active */
 278         rc = -EINVAL;
 279         if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
 280                 goto out_rel;
 281
 282         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 283         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 284
 285 out_rel:
 286         release_sock(sk);
 287 out:
 288         return rc;
 289 }
 290
 291 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 292                                    unsigned long mask)
 293 {
 294         /* options we don't get control via setsockopt for */
 295         nsk->sk_type = osk->sk_type;
 296         nsk->sk_sndbuf = osk->sk_sndbuf;
 297         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 298         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 299         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 300         nsk->sk_mark = osk->sk_mark;
 301         nsk->sk_priority = osk->sk_priority;
 302         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 303         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 304         nsk->sk_err = osk->sk_err;
 305
 306         nsk->sk_flags &= ~mask;
 307         nsk->sk_flags |= osk->sk_flags & mask;
 308 }
 309
 310 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 311                              (1UL << SOCK_KEEPOPEN) | \
 312                              (1UL << SOCK_LINGER) | \
 313                              (1UL << SOCK_BROADCAST) | \
 314                              (1UL << SOCK_TIMESTAMP) | \
 315                              (1UL << SOCK_DBG) | \
 316                              (1UL << SOCK_RCVTSTAMP) | \
 317                              (1UL << SOCK_RCVTSTAMPNS) | \
 318                              (1UL << SOCK_LOCALROUTE) | \
 319                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 320                              (1UL << SOCK_RXQ_OVFL) | \
 321                              (1UL << SOCK_WIFI_STATUS) | \
 322                              (1UL << SOCK_NOFCS) | \
 323                              (1UL << SOCK_FILTER_LOCKED) | \
 324                              (1UL << SOCK_TSTAMP_NEW))
 325 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 326  * clc socket (since smc is not called for these options from net/core)
 327  */
 328 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 329 {
 330         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 331 }
 332
 333 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 334                              (1UL << SOCK_KEEPOPEN) | \
 335                              (1UL << SOCK_LINGER) | \
 336                              (1UL << SOCK_DBG))
 337 /* copy only settings and flags relevant for smc from clc to smc socket */
 338 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 339 {
 340         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 341 }
 342
 343 /* register a new rmb, send confirm_rkey msg to register with peer */
 344 static int smc_reg_rmb(struct smc_link *link, struct smc_buf_desc *rmb_desc,
 345                        bool conf_rkey)
 346 {
 347         if (!rmb_desc->wr_reg) {
 348                 /* register memory region for new rmb */
 349                 if (smc_wr_reg_send(link, rmb_desc->mr_rx[SMC_SINGLE_LINK])) {
 350                         rmb_desc->regerr = 1;
 351                         return -EFAULT;
 352                 }
 353                 rmb_desc->wr_reg = 1;
 354         }
 355         if (!conf_rkey)
 356                 return 0;
 357         /* exchange confirm_rkey msg with peer */
 358         if (smc_llc_do_confirm_rkey(link, rmb_desc)) {
 359                 rmb_desc->regerr = 1;
 360                 return -EFAULT;
 361         }
 362         return 0;
 363 }
 364
 365 static int smc_clnt_conf_first_link(struct smc_sock *smc)
 366 {
 367         struct net *net = sock_net(smc->clcsock->sk);
 368         struct smc_link_group *lgr = smc->conn.lgr;
 369         struct smc_link *link;
 370         int rest;
 371         int rc;
 372
 373         link = &lgr->lnk[SMC_SINGLE_LINK];
 374         /* receive CONFIRM LINK request from server over RoCE fabric */
 375         rest = wait_for_completion_interruptible_timeout(
 376                 &link->llc_confirm,
 377                 SMC_LLC_WAIT_FIRST_TIME);
 378         if (rest <= 0) {
 379                 struct smc_clc_msg_decline dclc;
 380
 381                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 382                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 383                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
 384         }
 385
 386         if (link->llc_confirm_rc)
 387                 return SMC_CLC_DECL_RMBE_EC;
 388
 389         rc = smc_ib_modify_qp_rts(link);
 390         if (rc)
 391                 return SMC_CLC_DECL_ERR_RDYLNK;
 392
 393         smc_wr_remember_qp_attr(link);
 394
 395         if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
 396                 return SMC_CLC_DECL_ERR_REGRMB;
 397
 398         /* send CONFIRM LINK response over RoCE fabric */
 399         rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
 400         if (rc < 0)
 401                 return SMC_CLC_DECL_TIMEOUT_CL;
 402
 403         /* receive ADD LINK request from server over RoCE fabric */
 404         rest = wait_for_completion_interruptible_timeout(&link->llc_add,
 405                                                          SMC_LLC_WAIT_TIME);
 406         if (rest <= 0) {
 407                 struct smc_clc_msg_decline dclc;
 408
 409                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 410                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 411                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_AL : rc;
 412         }
 413
 414         /* send add link reject message, only one link supported for now */
 415         rc = smc_llc_send_add_link(link,
 416                                    link->smcibdev->mac[link->ibport - 1],
 417                                    link->gid, SMC_LLC_RESP);
 418         if (rc < 0)
 419                 return SMC_CLC_DECL_TIMEOUT_AL;
 420
 421         smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
 422
 423         return 0;
 424 }
 425
 426 static void smcr_conn_save_peer_info(struct smc_sock *smc,
 427                                      struct smc_clc_msg_accept_confirm *clc)
 428 {
 429         int bufsize = smc_uncompress_bufsize(clc->rmbe_size);
 430
 431         smc->conn.peer_rmbe_idx = clc->rmbe_idx;
 432         smc->conn.local_tx_ctrl.token = ntohl(clc->rmbe_alert_token);
 433         smc->conn.peer_rmbe_size = bufsize;
 434         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 435         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
 436 }
 437
 438 static void smcd_conn_save_peer_info(struct smc_sock *smc,
 439                                      struct smc_clc_msg_accept_confirm *clc)
 440 {
 441         int bufsize = smc_uncompress_bufsize(clc->dmbe_size);
 442
 443         smc->conn.peer_rmbe_idx = clc->dmbe_idx;
 444         smc->conn.peer_token = clc->token;
 445         /* msg header takes up space in the buffer */
 446         smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
 447         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 448         smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
 449 }
 450
 451 static void smc_conn_save_peer_info(struct smc_sock *smc,
 452                                     struct smc_clc_msg_accept_confirm *clc)
 453 {
 454         if (smc->conn.lgr->is_smcd)
 455                 smcd_conn_save_peer_info(smc, clc);
 456         else
 457                 smcr_conn_save_peer_info(smc, clc);
 458 }
 459
 460 static void smc_link_save_peer_info(struct smc_link *link,
 461                                     struct smc_clc_msg_accept_confirm *clc)
 462 {
 463         link->peer_qpn = ntoh24(clc->qpn);
 464         memcpy(link->peer_gid, clc->lcl.gid, SMC_GID_SIZE);
 465         memcpy(link->peer_mac, clc->lcl.mac, sizeof(link->peer_mac));
 466         link->peer_psn = ntoh24(clc->psn);
 467         link->peer_mtu = clc->qp_mtu;
 468 }
 469
 470 static void smc_switch_to_fallback(struct smc_sock *smc)
 471 {
 472         wait_queue_head_t *smc_wait = sk_sleep(&smc->sk);
 473         wait_queue_head_t *clc_wait = sk_sleep(smc->clcsock->sk);
 474         unsigned long flags;
 475
 476         smc->use_fallback = true;
 477         if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
 478                 smc->clcsock->file = smc->sk.sk_socket->file;
 479                 smc->clcsock->file->private_data = smc->clcsock;
 480                 smc->clcsock->wq.fasync_list =
 481                         smc->sk.sk_socket->wq.fasync_list;
 482
 483                 /* There may be some entries remaining in
 484                  * smc socket->wq, which should be removed
 485                  * to clcsocket->wq during the fallback.
 486                  */
 487                 spin_lock_irqsave(&smc_wait->lock, flags);
 488                 spin_lock_nested(&clc_wait->lock, SINGLE_DEPTH_NESTING);
 489                 list_splice_init(&smc_wait->head, &clc_wait->head);
 490                 spin_unlock(&clc_wait->lock);
 491                 spin_unlock_irqrestore(&smc_wait->lock, flags);
 492         }
 493 }
 494
 495 /* fall back during connect */
 496 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
 497 {
 498         smc_switch_to_fallback(smc);
 499         smc->fallback_rsn = reason_code;
 500         smc_copy_sock_settings_to_clc(smc);
 501         smc->connect_nonblock = 0;
 502         if (smc->sk.sk_state == SMC_INIT)
 503                 smc->sk.sk_state = SMC_ACTIVE;
 504         return 0;
 505 }
 506
 507 /* decline and fall back during connect */
 508 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
 509 {
 510         int rc;
 511
 512         if (reason_code < 0) { /* error, fallback is not possible */
 513                 if (smc->sk.sk_state == SMC_INIT)
 514                         sock_put(&smc->sk); /* passive closing */
 515                 return reason_code;
 516         }
 517         if (reason_code != SMC_CLC_DECL_PEERDECL) {
 518                 rc = smc_clc_send_decline(smc, reason_code);
 519                 if (rc < 0) {
 520                         if (smc->sk.sk_state == SMC_INIT)
 521                                 sock_put(&smc->sk); /* passive closing */
 522                         return rc;
 523                 }
 524         }
 525         return smc_connect_fallback(smc, reason_code);
 526 }
 527
 528 /* abort connecting */
 529 static int smc_connect_abort(struct smc_sock *smc, int reason_code,
 530                              int local_contact)
 531 {
 532         if (local_contact == SMC_FIRST_CONTACT)
 533                 smc_lgr_forget(smc->conn.lgr);
 534         if (smc->conn.lgr->is_smcd)
 535                 /* there is only one lgr role for SMC-D; use server lock */
 536                 mutex_unlock(&smc_server_lgr_pending);
 537         else
 538                 mutex_unlock(&smc_client_lgr_pending);
 539
 540         smc_conn_free(&smc->conn);
 541         smc->connect_nonblock = 0;
 542         return reason_code;
 543 }
 544
 545 /* check if there is a rdma device available for this connection. */
 546 /* called for connect and listen */
 547 static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
 548 {
 549         /* PNET table look up: search active ib_device and port
 550          * within same PNETID that also contains the ethernet device
 551          * used for the internal TCP socket
 552          */
 553         smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
 554         if (!ini->ib_dev)
 555                 return SMC_CLC_DECL_NOSMCRDEV;
 556         return 0;
 557 }
 558
 559 /* check if there is an ISM device available for this connection. */
 560 /* called for connect and listen */
 561 static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
 562 {
 563         /* Find ISM device with same PNETID as connecting interface  */
 564         smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
 565         if (!ini->ism_dev)
 566                 return SMC_CLC_DECL_NOSMCDDEV;
 567         return 0;
 568 }
 569
 570 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
 571 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
 572                                       struct smc_init_info *ini)
 573 {
 574         if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev, ini->vlan_id))
 575                 return SMC_CLC_DECL_ISMVLANERR;
 576         return 0;
 577 }
 578
 579 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
 580  * used, the VLAN ID will be registered again during the connection setup.
 581  */
 582 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc, bool is_smcd,
 583                                         struct smc_init_info *ini)
 584 {
 585         if (!is_smcd)
 586                 return 0;
 587         if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev, ini->vlan_id))
 588                 return SMC_CLC_DECL_CNFERR;
 589         return 0;
 590 }
 591
 592 /* CLC handshake during connect */
 593 static int smc_connect_clc(struct smc_sock *smc, int smc_type,
 594                            struct smc_clc_msg_accept_confirm *aclc,
 595                            struct smc_init_info *ini)
 596 {
 597         int rc = 0;
 598
 599         /* do inband token exchange */
 600         rc = smc_clc_send_proposal(smc, smc_type, ini);
 601         if (rc)
 602                 return rc;
 603         /* receive SMC Accept CLC message */
 604         return smc_clc_wait_msg(smc, aclc, sizeof(*aclc), SMC_CLC_ACCEPT,
 605                                 CLC_WAIT_TIME);
 606 }
 607
 608 /* setup for RDMA connection of client */
 609 static int smc_connect_rdma(struct smc_sock *smc,
 610                             struct smc_clc_msg_accept_confirm *aclc,
 611                             struct smc_init_info *ini)
 612 {
 613         struct smc_link *link;
 614         int reason_code = 0;
 615
 616         ini->is_smcd = false;
 617         ini->ib_lcl = &aclc->lcl;
 618         ini->ib_clcqpn = ntoh24(aclc->qpn);
 619         ini->srv_first_contact = aclc->hdr.flag;
 620
 621         mutex_lock(&smc_client_lgr_pending);
 622         reason_code = smc_conn_create(smc, ini);
 623         if (reason_code) {
 624                 mutex_unlock(&smc_client_lgr_pending);
 625                 return reason_code;
 626         }
 627         link = &smc->conn.lgr->lnk[SMC_SINGLE_LINK];
 628
 629         smc_conn_save_peer_info(smc, aclc);
 630
 631         /* create send buffer and rmb */
 632         if (smc_buf_create(smc, false))
 633                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM,
 634                                          ini->cln_first_contact);
 635
 636         if (ini->cln_first_contact == SMC_FIRST_CONTACT)
 637                 smc_link_save_peer_info(link, aclc);
 638
 639         if (smc_rmb_rtoken_handling(&smc->conn, aclc))
 640                 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RTOK,
 641                                          ini->cln_first_contact);
 642
 643         smc_close_init(smc);
 644         smc_rx_init(smc);
 645
 646         if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
 647                 if (smc_ib_ready_link(link))
 648                         return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RDYLNK,
 649                                                  ini->cln_first_contact);
 650         } else {
 651                 if (smc_reg_rmb(link, smc->conn.rmb_desc, true))
 652                         return smc_connect_abort(smc, SMC_CLC_DECL_ERR_REGRMB,
 653                                                  ini->cln_first_contact);
 654         }
 655         smc_rmb_sync_sg_for_device(&smc->conn);
 656
 657         reason_code = smc_clc_send_confirm(smc);
 658         if (reason_code)
 659                 return smc_connect_abort(smc, reason_code,
 660                                          ini->cln_first_contact);
 661
 662         smc_tx_init(smc);
 663
 664         if (ini->cln_first_contact == SMC_FIRST_CONTACT) {
 665                 /* QP confirmation over RoCE fabric */
 666                 reason_code = smc_clnt_conf_first_link(smc);
 667                 if (reason_code)
 668                         return smc_connect_abort(smc, reason_code,
 669                                                  ini->cln_first_contact);
 670         }
 671         mutex_unlock(&smc_client_lgr_pending);
 672
 673         smc_copy_sock_settings_to_clc(smc);
 674         smc->connect_nonblock = 0;
 675         if (smc->sk.sk_state == SMC_INIT)
 676                 smc->sk.sk_state = SMC_ACTIVE;
 677
 678         return 0;
 679 }
 680
 681 /* setup for ISM connection of client */
 682 static int smc_connect_ism(struct smc_sock *smc,
 683                            struct smc_clc_msg_accept_confirm *aclc,
 684                            struct smc_init_info *ini)
 685 {
 686         int rc = 0;
 687
 688         ini->is_smcd = true;
 689         ini->ism_gid = aclc->gid;
 690         ini->srv_first_contact = aclc->hdr.flag;
 691
 692         /* there is only one lgr role for SMC-D; use server lock */
 693         mutex_lock(&smc_server_lgr_pending);
 694         rc = smc_conn_create(smc, ini);
 695         if (rc) {
 696                 mutex_unlock(&smc_server_lgr_pending);
 697                 return rc;
 698         }
 699
 700         /* Create send and receive buffers */
 701         if (smc_buf_create(smc, true))
 702                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM,
 703                                          ini->cln_first_contact);
 704
 705         smc_conn_save_peer_info(smc, aclc);
 706         smc_close_init(smc);
 707         smc_rx_init(smc);
 708         smc_tx_init(smc);
 709
 710         rc = smc_clc_send_confirm(smc);
 711         if (rc)
 712                 return smc_connect_abort(smc, rc, ini->cln_first_contact);
 713         mutex_unlock(&smc_server_lgr_pending);
 714
 715         smc_copy_sock_settings_to_clc(smc);
 716         smc->connect_nonblock = 0;
 717         if (smc->sk.sk_state == SMC_INIT)
 718                 smc->sk.sk_state = SMC_ACTIVE;
 719
 720         return 0;
 721 }
 722
 723 /* perform steps before actually connecting */
 724 static int __smc_connect(struct smc_sock *smc)
 725 {
 726         bool ism_supported = false, rdma_supported = false;
 727         struct smc_clc_msg_accept_confirm aclc;
 728         struct smc_init_info ini = {0};
 729         int smc_type;
 730         int rc = 0;
 731
 732         if (smc->use_fallback)
 733                 return smc_connect_fallback(smc, smc->fallback_rsn);
 734
 735         /* if peer has not signalled SMC-capability, fall back */
 736         if (!tcp_sk(smc->clcsock->sk)->syn_smc)
 737                 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
 738
 739         /* IPSec connections opt out of SMC-R optimizations */
 740         if (using_ipsec(smc))
 741                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
 742
 743         /* get vlan id from IP device */
 744         if (smc_vlan_by_tcpsk(smc->clcsock, &ini))
 745                 return smc_connect_decline_fallback(smc,
 746                                                     SMC_CLC_DECL_GETVLANERR);
 747
 748         /* check if there is an ism device available */
 749         if (!smc_find_ism_device(smc, &ini) &&
 750             !smc_connect_ism_vlan_setup(smc, &ini)) {
 751                 /* ISM is supported for this connection */
 752                 ism_supported = true;
 753                 smc_type = SMC_TYPE_D;
 754         }
 755
 756         /* check if there is a rdma device available */
 757         if (!smc_find_rdma_device(smc, &ini)) {
 758                 /* RDMA is supported for this connection */
 759                 rdma_supported = true;
 760                 if (ism_supported)
 761                         smc_type = SMC_TYPE_B; /* both */
 762                 else
 763                         smc_type = SMC_TYPE_R; /* only RDMA */
 764         }
 765
 766         /* if neither ISM nor RDMA are supported, fallback */
 767         if (!rdma_supported && !ism_supported)
 768                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_NOSMCDEV);
 769
 770         /* perform CLC handshake */
 771         rc = smc_connect_clc(smc, smc_type, &aclc, &ini);
 772         if (rc) {
 773                 smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
 774                 return smc_connect_decline_fallback(smc, rc);
 775         }
 776
 777         /* depending on previous steps, connect using rdma or ism */
 778         if (rdma_supported && aclc.hdr.path == SMC_TYPE_R)
 779                 rc = smc_connect_rdma(smc, &aclc, &ini);
 780         else if (ism_supported && aclc.hdr.path == SMC_TYPE_D)
 781                 rc = smc_connect_ism(smc, &aclc, &ini);
 782         else
 783                 rc = SMC_CLC_DECL_MODEUNSUPP;
 784         if (rc) {
 785                 smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
 786                 return smc_connect_decline_fallback(smc, rc);
 787         }
 788
 789         smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
 790         return 0;
 791 }
 792
 793 static void smc_connect_work(struct work_struct *work)
 794 {
 795         struct smc_sock *smc = container_of(work, struct smc_sock,
 796                                             connect_work);
 797         long timeo = smc->sk.sk_sndtimeo;
 798         int rc = 0;
 799
 800         if (!timeo)
 801                 timeo = MAX_SCHEDULE_TIMEOUT;
 802         lock_sock(smc->clcsock->sk);
 803         if (smc->clcsock->sk->sk_err) {
 804                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
 805         } else if ((1 << smc->clcsock->sk->sk_state) &
 806                                         (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
 807                 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
 808                 if ((rc == -EPIPE) &&
 809                     ((1 << smc->clcsock->sk->sk_state) &
 810                                         (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
 811                         rc = 0;
 812         }
 813         release_sock(smc->clcsock->sk);
 814         lock_sock(&smc->sk);
 815         if (rc != 0 || smc->sk.sk_err) {
 816                 smc->sk.sk_state = SMC_CLOSED;
 817                 if (rc == -EPIPE || rc == -EAGAIN)
 818                         smc->sk.sk_err = EPIPE;
 819                 else if (rc == -ECONNREFUSED)
 820                         smc->sk.sk_err = ECONNREFUSED;
 821                 else if (signal_pending(current))
 822                         smc->sk.sk_err = -sock_intr_errno(timeo);
 823                 sock_put(&smc->sk); /* passive closing */
 824                 goto out;
 825         }
 826
 827         rc = __smc_connect(smc);
 828         if (rc < 0)
 829                 smc->sk.sk_err = -rc;
 830
 831 out:
 832         if (!sock_flag(&smc->sk, SOCK_DEAD)) {
 833                 if (smc->sk.sk_err) {
 834                         smc->sk.sk_state_change(&smc->sk);
 835                 } else { /* allow polling before and after fallback decision */
 836                         smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
 837                         smc->sk.sk_write_space(&smc->sk);
 838                 }
 839         }
 840         release_sock(&smc->sk);
 841 }
 842
 843 static int smc_connect(struct socket *sock, struct sockaddr *addr,
 844                        int alen, int flags)
 845 {
 846         struct sock *sk = sock->sk;
 847         struct smc_sock *smc;
 848         int rc = -EINVAL;
 849
 850         smc = smc_sk(sk);
 851
 852         /* separate smc parameter checking to be safe */
 853         if (alen < sizeof(addr->sa_family))
 854                 goto out_err;
 855         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
 856                 goto out_err;
 857
 858         lock_sock(sk);
 859         switch (sk->sk_state) {
 860         default:
 861                 goto out;
 862         case SMC_ACTIVE:
 863                 rc = -EISCONN;
 864                 goto out;
 865         case SMC_INIT:
 866                 rc = 0;
 867                 break;
 868         }
 869
 870         smc_copy_sock_settings_to_clc(smc);
 871         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
 872         if (smc->connect_nonblock) {
 873                 rc = -EALREADY;
 874                 goto out;
 875         }
 876         rc = kernel_connect(smc->clcsock, addr, alen, flags);
 877         if (rc && rc != -EINPROGRESS)
 878                 goto out;
 879
 880         if (smc->use_fallback)
 881                 goto out;
 882         sock_hold(&smc->sk); /* sock put in passive closing */
 883         if (flags & O_NONBLOCK) {
 884                 if (schedule_work(&smc->connect_work))
 885                         smc->connect_nonblock = 1;
 886                 rc = -EINPROGRESS;
 887         } else {
 888                 rc = __smc_connect(smc);
 889                 if (rc < 0)
 890                         goto out;
 891                 else
 892                         rc = 0; /* success cases including fallback */
 893         }
 894
 895 out:
 896         release_sock(sk);
 897 out_err:
 898         return rc;
 899 }
 900
 901 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 902 {
 903         struct socket *new_clcsock = NULL;
 904         struct sock *lsk = &lsmc->sk;
 905         struct sock *new_sk;
 906         int rc = -EINVAL;
 907
 908         release_sock(lsk);
 909         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
 910         if (!new_sk) {
 911                 rc = -ENOMEM;
 912                 lsk->sk_err = ENOMEM;
 913                 *new_smc = NULL;
 914                 lock_sock(lsk);
 915                 goto out;
 916         }
 917         *new_smc = smc_sk(new_sk);
 918
 919         mutex_lock(&lsmc->clcsock_release_lock);
 920         if (lsmc->clcsock)
 921                 rc = kernel_accept(lsmc->clcsock, &new_clcsock, 0);
 922         mutex_unlock(&lsmc->clcsock_release_lock);
 923         lock_sock(lsk);
 924         if  (rc < 0)
 925                 lsk->sk_err = -rc;
 926         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
 927                 new_sk->sk_prot->unhash(new_sk);
 928                 if (new_clcsock)
 929                         sock_release(new_clcsock);
 930                 new_sk->sk_state = SMC_CLOSED;
 931                 sock_set_flag(new_sk, SOCK_DEAD);
 932                 sock_put(new_sk); /* final */
 933                 *new_smc = NULL;
 934                 goto out;
 935         }
 936
 937         (*new_smc)->clcsock = new_clcsock;
 938 out:
 939         return rc;
 940 }
 941
 942 /* add a just created sock to the accept queue of the listen sock as
 943  * candidate for a following socket accept call from user space
 944  */
 945 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
 946 {
 947         struct smc_sock *par = smc_sk(parent);
 948
 949         sock_hold(sk); /* sock_put in smc_accept_unlink () */
 950         spin_lock(&par->accept_q_lock);
 951         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
 952         spin_unlock(&par->accept_q_lock);
 953         sk_acceptq_added(parent);
 954 }
 955
 956 /* remove a socket from the accept queue of its parental listening socket */
 957 static void smc_accept_unlink(struct sock *sk)
 958 {
 959         struct smc_sock *par = smc_sk(sk)->listen_smc;
 960
 961         spin_lock(&par->accept_q_lock);
 962         list_del_init(&smc_sk(sk)->accept_q);
 963         spin_unlock(&par->accept_q_lock);
 964         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
 965         sock_put(sk); /* sock_hold in smc_accept_enqueue */
 966 }
 967
 968 /* remove a sock from the accept queue to bind it to a new socket created
 969  * for a socket accept call from user space
 970  */
 971 struct sock *smc_accept_dequeue(struct sock *parent,
 972                                 struct socket *new_sock)
 973 {
 974         struct smc_sock *isk, *n;
 975         struct sock *new_sk;
 976
 977         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
 978                 new_sk = (struct sock *)isk;
 979
 980                 smc_accept_unlink(new_sk);
 981                 if (new_sk->sk_state == SMC_CLOSED) {
 982                         new_sk->sk_prot->unhash(new_sk);
 983                         if (isk->clcsock) {
 984                                 sock_release(isk->clcsock);
 985                                 isk->clcsock = NULL;
 986                         }
 987                         sock_put(new_sk); /* final */
 988                         continue;
 989                 }
 990                 if (new_sock) {
 991                         sock_graft(new_sk, new_sock);
 992                         if (isk->use_fallback) {
 993                                 smc_sk(new_sk)->clcsock->file = new_sock->file;
 994                                 isk->clcsock->file->private_data = isk->clcsock;
 995                         }
 996                 }
 997                 return new_sk;
 998         }
 999         return NULL;
1000 }
1001
1002 /* clean up for a created but never accepted sock */
1003 void smc_close_non_accepted(struct sock *sk)
1004 {
1005         struct smc_sock *smc = smc_sk(sk);
1006
1007         lock_sock(sk);
1008         if (!sk->sk_lingertime)
1009                 /* wait for peer closing */
1010                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
1011         __smc_release(smc);
1012         release_sock(sk);
1013         sock_put(sk); /* final sock_put */
1014 }
1015
1016 static int smc_serv_conf_first_link(struct smc_sock *smc)
1017 {
1018         struct net *net = sock_net(smc->clcsock->sk);
1019         struct smc_link_group *lgr = smc->conn.lgr;
1020         struct smc_link *link;
1021         int rest;
1022         int rc;
1023
1024         link = &lgr->lnk[SMC_SINGLE_LINK];
1025
1026         if (smc_reg_rmb(link, smc->conn.rmb_desc, false))
1027                 return SMC_CLC_DECL_ERR_REGRMB;
1028
1029         /* send CONFIRM LINK request to client over the RoCE fabric */
1030         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1031         if (rc < 0)
1032                 return SMC_CLC_DECL_TIMEOUT_CL;
1033
1034         /* receive CONFIRM LINK response from client over the RoCE fabric */
1035         rest = wait_for_completion_interruptible_timeout(
1036                 &link->llc_confirm_resp,
1037                 SMC_LLC_WAIT_FIRST_TIME);
1038         if (rest <= 0) {
1039                 struct smc_clc_msg_decline dclc;
1040
1041                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1042                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1043                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1044         }
1045
1046         if (link->llc_confirm_resp_rc)
1047                 return SMC_CLC_DECL_RMBE_EC;
1048
1049         /* send ADD LINK request to client over the RoCE fabric */
1050         rc = smc_llc_send_add_link(link,
1051                                    link->smcibdev->mac[link->ibport - 1],
1052                                    link->gid, SMC_LLC_REQ);
1053         if (rc < 0)
1054                 return SMC_CLC_DECL_TIMEOUT_AL;
1055
1056         /* receive ADD LINK response from client over the RoCE fabric */
1057         rest = wait_for_completion_interruptible_timeout(&link->llc_add_resp,
1058                                                          SMC_LLC_WAIT_TIME);
1059         if (rest <= 0) {
1060                 struct smc_clc_msg_decline dclc;
1061
1062                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1063                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1064                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_AL : rc;
1065         }
1066
1067         smc_llc_link_active(link, net->ipv4.sysctl_tcp_keepalive_time);
1068
1069         return 0;
1070 }
1071
1072 /* listen worker: finish */
1073 static void smc_listen_out(struct smc_sock *new_smc)
1074 {
1075         struct smc_sock *lsmc = new_smc->listen_smc;
1076         struct sock *newsmcsk = &new_smc->sk;
1077
1078         if (lsmc->sk.sk_state == SMC_LISTEN) {
1079                 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1080                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1081                 release_sock(&lsmc->sk);
1082         } else { /* no longer listening */
1083                 smc_close_non_accepted(newsmcsk);
1084         }
1085
1086         /* Wake up accept */
1087         lsmc->sk.sk_data_ready(&lsmc->sk);
1088         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1089 }
1090
1091 /* listen worker: finish in state connected */
1092 static void smc_listen_out_connected(struct smc_sock *new_smc)
1093 {
1094         struct sock *newsmcsk = &new_smc->sk;
1095
1096         if (newsmcsk->sk_state == SMC_INIT)
1097                 newsmcsk->sk_state = SMC_ACTIVE;
1098
1099         smc_listen_out(new_smc);
1100 }
1101
1102 /* listen worker: finish in error state */
1103 static void smc_listen_out_err(struct smc_sock *new_smc)
1104 {
1105         struct sock *newsmcsk = &new_smc->sk;
1106
1107         if (newsmcsk->sk_state == SMC_INIT)
1108                 sock_put(&new_smc->sk); /* passive closing */
1109         newsmcsk->sk_state = SMC_CLOSED;
1110         smc_conn_free(&new_smc->conn);
1111
1112         smc_listen_out(new_smc);
1113 }
1114
1115 /* listen worker: decline and fall back if possible */
1116 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1117                                int local_contact)
1118 {
1119         /* RDMA setup failed, switch back to TCP */
1120         if (local_contact == SMC_FIRST_CONTACT)
1121                 smc_lgr_forget(new_smc->conn.lgr);
1122         if (reason_code < 0) { /* error, no fallback possible */
1123                 smc_listen_out_err(new_smc);
1124                 return;
1125         }
1126         smc_conn_free(&new_smc->conn);
1127         smc_switch_to_fallback(new_smc);
1128         new_smc->fallback_rsn = reason_code;
1129         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1130                 if (smc_clc_send_decline(new_smc, reason_code) < 0) {
1131                         smc_listen_out_err(new_smc);
1132                         return;
1133                 }
1134         }
1135         smc_listen_out_connected(new_smc);
1136 }
1137
1138 /* listen worker: check prefixes */
1139 static int smc_listen_prfx_check(struct smc_sock *new_smc,
1140                                  struct smc_clc_msg_proposal *pclc)
1141 {
1142         struct smc_clc_msg_proposal_prefix *pclc_prfx;
1143         struct socket *newclcsock = new_smc->clcsock;
1144
1145         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1146         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1147                 return SMC_CLC_DECL_DIFFPREFIX;
1148
1149         return 0;
1150 }
1151
1152 /* listen worker: initialize connection and buffers */
1153 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1154                                 struct smc_init_info *ini)
1155 {
1156         int rc;
1157
1158         /* allocate connection / link group */
1159         rc = smc_conn_create(new_smc, ini);
1160         if (rc)
1161                 return rc;
1162
1163         /* create send buffer and rmb */
1164         if (smc_buf_create(new_smc, false))
1165                 return SMC_CLC_DECL_MEM;
1166
1167         return 0;
1168 }
1169
1170 /* listen worker: initialize connection and buffers for SMC-D */
1171 static int smc_listen_ism_init(struct smc_sock *new_smc,
1172                                struct smc_clc_msg_proposal *pclc,
1173                                struct smc_init_info *ini)
1174 {
1175         struct smc_clc_msg_smcd *pclc_smcd;
1176         int rc;
1177
1178         pclc_smcd = smc_get_clc_msg_smcd(pclc);
1179         ini->ism_gid = pclc_smcd->gid;
1180         rc = smc_conn_create(new_smc, ini);
1181         if (rc)
1182                 return rc;
1183
1184         /* Check if peer can be reached via ISM device */
1185         if (smc_ism_cantalk(new_smc->conn.lgr->peer_gid,
1186                             new_smc->conn.lgr->vlan_id,
1187                             new_smc->conn.lgr->smcd)) {
1188                 if (ini->cln_first_contact == SMC_FIRST_CONTACT)
1189                         smc_lgr_forget(new_smc->conn.lgr);
1190                 smc_conn_free(&new_smc->conn);
1191                 return SMC_CLC_DECL_SMCDNOTALK;
1192         }
1193
1194         /* Create send and receive buffers */
1195         if (smc_buf_create(new_smc, true)) {
1196                 if (ini->cln_first_contact == SMC_FIRST_CONTACT)
1197                         smc_lgr_forget(new_smc->conn.lgr);
1198                 smc_conn_free(&new_smc->conn);
1199                 return SMC_CLC_DECL_MEM;
1200         }
1201
1202         return 0;
1203 }
1204
1205 /* listen worker: register buffers */
1206 static int smc_listen_rdma_reg(struct smc_sock *new_smc, int local_contact)
1207 {
1208         struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
1209
1210         if (local_contact != SMC_FIRST_CONTACT) {
1211                 if (smc_reg_rmb(link, new_smc->conn.rmb_desc, true))
1212                         return SMC_CLC_DECL_ERR_REGRMB;
1213         }
1214         smc_rmb_sync_sg_for_device(&new_smc->conn);
1215
1216         return 0;
1217 }
1218
1219 /* listen worker: finish RDMA setup */
1220 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1221                                   struct smc_clc_msg_accept_confirm *cclc,
1222                                   int local_contact)
1223 {
1224         struct smc_link *link = &new_smc->conn.lgr->lnk[SMC_SINGLE_LINK];
1225         int reason_code = 0;
1226
1227         if (local_contact == SMC_FIRST_CONTACT)
1228                 smc_link_save_peer_info(link, cclc);
1229
1230         if (smc_rmb_rtoken_handling(&new_smc->conn, cclc)) {
1231                 reason_code = SMC_CLC_DECL_ERR_RTOK;
1232                 goto decline;
1233         }
1234
1235         if (local_contact == SMC_FIRST_CONTACT) {
1236                 if (smc_ib_ready_link(link)) {
1237                         reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1238                         goto decline;
1239                 }
1240                 /* QP confirmation over RoCE fabric */
1241                 reason_code = smc_serv_conf_first_link(new_smc);
1242                 if (reason_code)
1243                         goto decline;
1244         }
1245         return 0;
1246
1247 decline:
1248         smc_listen_decline(new_smc, reason_code, local_contact);
1249         return reason_code;
1250 }
1251
1252 /* setup for RDMA connection of server */
1253 static void smc_listen_work(struct work_struct *work)
1254 {
1255         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1256                                                 smc_listen_work);
1257         struct socket *newclcsock = new_smc->clcsock;
1258         struct smc_clc_msg_accept_confirm cclc;
1259         struct smc_clc_msg_proposal *pclc;
1260         struct smc_init_info ini = {0};
1261         bool ism_supported = false;
1262         u8 buf[SMC_CLC_MAX_LEN];
1263         int rc = 0;
1264
1265         if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1266                 return smc_listen_out_err(new_smc);
1267
1268         if (new_smc->use_fallback) {
1269                 smc_listen_out_connected(new_smc);
1270                 return;
1271         }
1272
1273         /* check if peer is smc capable */
1274         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1275                 smc_switch_to_fallback(new_smc);
1276                 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1277                 smc_listen_out_connected(new_smc);
1278                 return;
1279         }
1280
1281         /* do inband token exchange -
1282          * wait for and receive SMC Proposal CLC message
1283          */
1284         pclc = (struct smc_clc_msg_proposal *)&buf;
1285         rc = smc_clc_wait_msg(new_smc, pclc, SMC_CLC_MAX_LEN,
1286                               SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1287         if (rc)
1288                 goto out_decl;
1289
1290         /* IPSec connections opt out of SMC-R optimizations */
1291         if (using_ipsec(new_smc)) {
1292                 rc = SMC_CLC_DECL_IPSEC;
1293                 goto out_decl;
1294         }
1295
1296         /* check for matching IP prefix and subnet length */
1297         rc = smc_listen_prfx_check(new_smc, pclc);
1298         if (rc)
1299                 goto out_decl;
1300
1301         /* get vlan id from IP device */
1302         if (smc_vlan_by_tcpsk(new_smc->clcsock, &ini)) {
1303                 rc = SMC_CLC_DECL_GETVLANERR;
1304                 goto out_decl;
1305         }
1306
1307         mutex_lock(&smc_server_lgr_pending);
1308         smc_close_init(new_smc);
1309         smc_rx_init(new_smc);
1310         smc_tx_init(new_smc);
1311
1312         /* check if ISM is available */
1313         if (pclc->hdr.path == SMC_TYPE_D || pclc->hdr.path == SMC_TYPE_B) {
1314                 ini.is_smcd = true; /* prepare ISM check */
1315                 rc = smc_find_ism_device(new_smc, &ini);
1316                 if (!rc)
1317                         rc = smc_listen_ism_init(new_smc, pclc, &ini);
1318                 if (!rc)
1319                         ism_supported = true;
1320                 else if (pclc->hdr.path == SMC_TYPE_D)
1321                         goto out_unlock; /* skip RDMA and decline */
1322         }
1323
1324         /* check if RDMA is available */
1325         if (!ism_supported) { /* SMC_TYPE_R or SMC_TYPE_B */
1326                 /* prepare RDMA check */
1327                 ini.is_smcd = false;
1328                 ini.ism_dev = NULL;
1329                 ini.ib_lcl = &pclc->lcl;
1330                 rc = smc_find_rdma_device(new_smc, &ini);
1331                 if (rc) {
1332                         /* no RDMA device found */
1333                         if (pclc->hdr.path == SMC_TYPE_B)
1334                                 /* neither ISM nor RDMA device found */
1335                                 rc = SMC_CLC_DECL_NOSMCDEV;
1336                         goto out_unlock;
1337                 }
1338                 rc = smc_listen_rdma_init(new_smc, &ini);
1339                 if (rc)
1340                         goto out_unlock;
1341                 rc = smc_listen_rdma_reg(new_smc, ini.cln_first_contact);
1342                 if (rc)
1343                         goto out_unlock;
1344         }
1345
1346         /* send SMC Accept CLC message */
1347         rc = smc_clc_send_accept(new_smc, ini.cln_first_contact);
1348         if (rc)
1349                 goto out_unlock;
1350
1351         /* SMC-D does not need this lock any more */
1352         if (ism_supported)
1353                 mutex_unlock(&smc_server_lgr_pending);
1354
1355         /* receive SMC Confirm CLC message */
1356         rc = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
1357                               SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1358         if (rc) {
1359                 if (!ism_supported)
1360                         goto out_unlock;
1361                 goto out_decl;
1362         }
1363
1364         /* finish worker */
1365         if (!ism_supported) {
1366                 rc = smc_listen_rdma_finish(new_smc, &cclc,
1367                                             ini.cln_first_contact);
1368                 mutex_unlock(&smc_server_lgr_pending);
1369                 if (rc)
1370                         return;
1371         }
1372         smc_conn_save_peer_info(new_smc, &cclc);
1373         smc_listen_out_connected(new_smc);
1374         return;
1375
1376 out_unlock:
1377         mutex_unlock(&smc_server_lgr_pending);
1378 out_decl:
1379         smc_listen_decline(new_smc, rc, ini.cln_first_contact);
1380 }
1381
1382 static void smc_tcp_listen_work(struct work_struct *work)
1383 {
1384         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1385                                              tcp_listen_work);
1386         struct sock *lsk = &lsmc->sk;
1387         struct smc_sock *new_smc;
1388         int rc = 0;
1389
1390         lock_sock(lsk);
1391         while (lsk->sk_state == SMC_LISTEN) {
1392                 rc = smc_clcsock_accept(lsmc, &new_smc);
1393                 if (rc)
1394                         goto out;
1395                 if (!new_smc)
1396                         continue;
1397
1398                 new_smc->listen_smc = lsmc;
1399                 new_smc->use_fallback = lsmc->use_fallback;
1400                 new_smc->fallback_rsn = lsmc->fallback_rsn;
1401                 sock_hold(lsk); /* sock_put in smc_listen_work */
1402                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1403                 smc_copy_sock_settings_to_smc(new_smc);
1404                 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1405                 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1406                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1407                 if (!schedule_work(&new_smc->smc_listen_work))
1408                         sock_put(&new_smc->sk);
1409         }
1410
1411 out:
1412         release_sock(lsk);
1413         sock_put(&lsmc->sk); /* sock_hold in smc_listen */
1414 }
1415
1416 static int smc_listen(struct socket *sock, int backlog)
1417 {
1418         struct sock *sk = sock->sk;
1419         struct smc_sock *smc;
1420         int rc;
1421
1422         smc = smc_sk(sk);
1423         lock_sock(sk);
1424
1425         rc = -EINVAL;
1426         if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
1427             smc->connect_nonblock)
1428                 goto out;
1429
1430         rc = 0;
1431         if (sk->sk_state == SMC_LISTEN) {
1432                 sk->sk_max_ack_backlog = backlog;
1433                 goto out;
1434         }
1435         /* some socket options are handled in core, so we could not apply
1436          * them to the clc socket -- copy smc socket options to clc socket
1437          */
1438         smc_copy_sock_settings_to_clc(smc);
1439         if (!smc->use_fallback)
1440                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1441
1442         rc = kernel_listen(smc->clcsock, backlog);
1443         if (rc)
1444                 goto out;
1445         sk->sk_max_ack_backlog = backlog;
1446         sk->sk_ack_backlog = 0;
1447         sk->sk_state = SMC_LISTEN;
1448         sock_hold(sk); /* sock_hold in tcp_listen_worker */
1449         if (!schedule_work(&smc->tcp_listen_work))
1450                 sock_put(sk);
1451
1452 out:
1453         release_sock(sk);
1454         return rc;
1455 }
1456
1457 static int smc_accept(struct socket *sock, struct socket *new_sock,
1458                       int flags, bool kern)
1459 {
1460         struct sock *sk = sock->sk, *nsk;
1461         DECLARE_WAITQUEUE(wait, current);
1462         struct smc_sock *lsmc;
1463         long timeo;
1464         int rc = 0;
1465
1466         lsmc = smc_sk(sk);
1467         sock_hold(sk); /* sock_put below */
1468         lock_sock(sk);
1469
1470         if (lsmc->sk.sk_state != SMC_LISTEN) {
1471                 rc = -EINVAL;
1472                 release_sock(sk);
1473                 goto out;
1474         }
1475
1476         /* Wait for an incoming connection */
1477         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1478         add_wait_queue_exclusive(sk_sleep(sk), &wait);
1479         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1480                 set_current_state(TASK_INTERRUPTIBLE);
1481                 if (!timeo) {
1482                         rc = -EAGAIN;
1483                         break;
1484                 }
1485                 release_sock(sk);
1486                 timeo = schedule_timeout(timeo);
1487                 /* wakeup by sk_data_ready in smc_listen_work() */
1488                 sched_annotate_sleep();
1489                 lock_sock(sk);
1490                 if (signal_pending(current)) {
1491                         rc = sock_intr_errno(timeo);
1492                         break;
1493                 }
1494         }
1495         set_current_state(TASK_RUNNING);
1496         remove_wait_queue(sk_sleep(sk), &wait);
1497
1498         if (!rc)
1499                 rc = sock_error(nsk);
1500         release_sock(sk);
1501         if (rc)
1502                 goto out;
1503
1504         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1505                 /* wait till data arrives on the socket */
1506                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1507                                                                 MSEC_PER_SEC);
1508                 if (smc_sk(nsk)->use_fallback) {
1509                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1510
1511                         lock_sock(clcsk);
1512                         if (skb_queue_empty(&clcsk->sk_receive_queue))
1513                                 sk_wait_data(clcsk, &timeo, NULL);
1514                         release_sock(clcsk);
1515                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1516                         lock_sock(nsk);
1517                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1518                         release_sock(nsk);
1519                 }
1520         }
1521
1522 out:
1523         sock_put(sk); /* sock_hold above */
1524         return rc;
1525 }
1526
1527 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1528                        int peer)
1529 {
1530         struct smc_sock *smc;
1531
1532         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1533             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1534                 return -ENOTCONN;
1535
1536         smc = smc_sk(sock->sk);
1537
1538         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1539 }
1540
1541 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1542 {
1543         struct sock *sk = sock->sk;
1544         struct smc_sock *smc;
1545         int rc;
1546
1547         smc = smc_sk(sk);
1548         lock_sock(sk);
1549
1550         /* SMC does not support connect with fastopen */
1551         if (msg->msg_flags & MSG_FASTOPEN) {
1552                 /* not connected yet, fallback */
1553                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1554                         smc_switch_to_fallback(smc);
1555                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1556                 } else {
1557                         rc = -EINVAL;
1558                         goto out;
1559                 }
1560         } else if ((sk->sk_state != SMC_ACTIVE) &&
1561                    (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1562                    (sk->sk_state != SMC_INIT)) {
1563                 rc = -EPIPE;
1564                 goto out;
1565         }
1566
1567         if (smc->use_fallback)
1568                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1569         else
1570                 rc = smc_tx_sendmsg(smc, msg, len);
1571 out:
1572         release_sock(sk);
1573         return rc;
1574 }
1575
1576 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1577                        int flags)
1578 {
1579         struct sock *sk = sock->sk;
1580         struct smc_sock *smc;
1581         int rc = -ENOTCONN;
1582
1583         smc = smc_sk(sk);
1584         lock_sock(sk);
1585         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1586                 /* socket was connected before, no more data to read */
1587                 rc = 0;
1588                 goto out;
1589         }
1590         if ((sk->sk_state == SMC_INIT) ||
1591             (sk->sk_state == SMC_LISTEN) ||
1592             (sk->sk_state == SMC_CLOSED))
1593                 goto out;
1594
1595         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1596                 rc = 0;
1597                 goto out;
1598         }
1599
1600         if (smc->use_fallback) {
1601                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1602         } else {
1603                 msg->msg_namelen = 0;
1604                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1605         }
1606
1607 out:
1608         release_sock(sk);
1609         return rc;
1610 }
1611
1612 static __poll_t smc_accept_poll(struct sock *parent)
1613 {
1614         struct smc_sock *isk = smc_sk(parent);
1615         __poll_t mask = 0;
1616
1617         spin_lock(&isk->accept_q_lock);
1618         if (!list_empty(&isk->accept_q))
1619                 mask = EPOLLIN | EPOLLRDNORM;
1620         spin_unlock(&isk->accept_q_lock);
1621
1622         return mask;
1623 }
1624
1625 static __poll_t smc_poll(struct file *file, struct socket *sock,
1626                              poll_table *wait)
1627 {
1628         struct sock *sk = sock->sk;
1629         struct smc_sock *smc;
1630         __poll_t mask = 0;
1631
1632         if (!sk)
1633                 return EPOLLNVAL;
1634
1635         smc = smc_sk(sock->sk);
1636         if (smc->use_fallback) {
1637                 /* delegate to CLC child sock */
1638                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1639                 sk->sk_err = smc->clcsock->sk->sk_err;
1640         } else {
1641                 if (sk->sk_state != SMC_CLOSED)
1642                         sock_poll_wait(file, sock, wait);
1643                 if (sk->sk_err)
1644                         mask |= EPOLLERR;
1645                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1646                     (sk->sk_state == SMC_CLOSED))
1647                         mask |= EPOLLHUP;
1648                 if (sk->sk_state == SMC_LISTEN) {
1649                         /* woken up by sk_data_ready in smc_listen_work() */
1650                         mask |= smc_accept_poll(sk);
1651                 } else if (smc->use_fallback) { /* as result of connect_work()*/
1652                         mask |= smc->clcsock->ops->poll(file, smc->clcsock,
1653                                                            wait);
1654                         sk->sk_err = smc->clcsock->sk->sk_err;
1655                 } else {
1656                         if ((sk->sk_state != SMC_INIT &&
1657                              atomic_read(&smc->conn.sndbuf_space)) ||
1658                             sk->sk_shutdown & SEND_SHUTDOWN) {
1659                                 mask |= EPOLLOUT | EPOLLWRNORM;
1660                         } else {
1661                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1662                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1663                         }
1664                         if (atomic_read(&smc->conn.bytes_to_rcv))
1665                                 mask |= EPOLLIN | EPOLLRDNORM;
1666                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1667                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1668                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
1669                                 mask |= EPOLLIN;
1670                         if (smc->conn.urg_state == SMC_URG_VALID)
1671                                 mask |= EPOLLPRI;
1672                 }
1673         }
1674
1675         return mask;
1676 }
1677
1678 static int smc_shutdown(struct socket *sock, int how)
1679 {
1680         struct sock *sk = sock->sk;
1681         bool do_shutdown = true;
1682         struct smc_sock *smc;
1683         int rc = -EINVAL;
1684         int old_state;
1685         int rc1 = 0;
1686
1687         smc = smc_sk(sk);
1688
1689         if ((how < SHUT_RD) || (how > SHUT_RDWR))
1690                 return rc;
1691
1692         lock_sock(sk);
1693
1694         rc = -ENOTCONN;
1695         if ((sk->sk_state != SMC_ACTIVE) &&
1696             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1697             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1698             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1699             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1700             (sk->sk_state != SMC_APPFINCLOSEWAIT))
1701                 goto out;
1702         if (smc->use_fallback) {
1703                 rc = kernel_sock_shutdown(smc->clcsock, how);
1704                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1705                 if (sk->sk_shutdown == SHUTDOWN_MASK) {
1706                         sk->sk_state = SMC_CLOSED;
1707                         sock_put(sk);
1708                 }
1709                 goto out;
1710         }
1711         switch (how) {
1712         case SHUT_RDWR:         /* shutdown in both directions */
1713                 old_state = sk->sk_state;
1714                 rc = smc_close_active(smc);
1715                 if (old_state == SMC_ACTIVE &&
1716                     sk->sk_state == SMC_PEERCLOSEWAIT1)
1717                         do_shutdown = false;
1718                 break;
1719         case SHUT_WR:
1720                 rc = smc_close_shutdown_write(smc);
1721                 break;
1722         case SHUT_RD:
1723                 rc = 0;
1724                 /* nothing more to do because peer is not involved */
1725                 break;
1726         }
1727         if (do_shutdown && smc->clcsock)
1728                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1729         /* map sock_shutdown_cmd constants to sk_shutdown value range */
1730         sk->sk_shutdown |= how + 1;
1731
1732 out:
1733         release_sock(sk);
1734         return rc ? rc : rc1;
1735 }
1736
1737 static int smc_setsockopt(struct socket *sock, int level, int optname,
1738                           char __user *optval, unsigned int optlen)
1739 {
1740         struct sock *sk = sock->sk;
1741         struct smc_sock *smc;
1742         int val, rc;
1743
1744         if (level == SOL_TCP && optname == TCP_ULP)
1745                 return -EOPNOTSUPP;
1746
1747         smc = smc_sk(sk);
1748
1749         /* generic setsockopts reaching us here always apply to the
1750          * CLC socket
1751          */
1752         rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1753                                            optval, optlen);
1754         if (smc->clcsock->sk->sk_err) {
1755                 sk->sk_err = smc->clcsock->sk->sk_err;
1756                 sk->sk_error_report(sk);
1757         }
1758
1759         if (optlen < sizeof(int))
1760                 return -EINVAL;
1761         if (get_user(val, (int __user *)optval))
1762                 return -EFAULT;
1763
1764         lock_sock(sk);
1765         if (rc || smc->use_fallback)
1766                 goto out;
1767         switch (optname) {
1768         case TCP_FASTOPEN:
1769         case TCP_FASTOPEN_CONNECT:
1770         case TCP_FASTOPEN_KEY:
1771         case TCP_FASTOPEN_NO_COOKIE:
1772                 /* option not supported by SMC */
1773                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1774                         smc_switch_to_fallback(smc);
1775                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1776                 } else {
1777                         rc = -EINVAL;
1778                 }
1779                 break;
1780         case TCP_NODELAY:
1781                 if (sk->sk_state != SMC_INIT &&
1782                     sk->sk_state != SMC_LISTEN &&
1783                     sk->sk_state != SMC_CLOSED) {
1784                         if (val)
1785                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1786                                                  0);
1787                 }
1788                 break;
1789         case TCP_CORK:
1790                 if (sk->sk_state != SMC_INIT &&
1791                     sk->sk_state != SMC_LISTEN &&
1792                     sk->sk_state != SMC_CLOSED) {
1793                         if (!val)
1794                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1795                                                  0);
1796                 }
1797                 break;
1798         case TCP_DEFER_ACCEPT:
1799                 smc->sockopt_defer_accept = val;
1800                 break;
1801         default:
1802                 break;
1803         }
1804 out:
1805         release_sock(sk);
1806
1807         return rc;
1808 }
1809
1810 static int smc_getsockopt(struct socket *sock, int level, int optname,
1811                           char __user *optval, int __user *optlen)
1812 {
1813         struct smc_sock *smc;
1814
1815         smc = smc_sk(sock->sk);
1816         /* socket options apply to the CLC socket */
1817         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1818                                              optval, optlen);
1819 }
1820
1821 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1822                      unsigned long arg)
1823 {
1824         union smc_host_cursor cons, urg;
1825         struct smc_connection *conn;
1826         struct smc_sock *smc;
1827         int answ;
1828
1829         smc = smc_sk(sock->sk);
1830         conn = &smc->conn;
1831         lock_sock(&smc->sk);
1832         if (smc->use_fallback) {
1833                 if (!smc->clcsock) {
1834                         release_sock(&smc->sk);
1835                         return -EBADF;
1836                 }
1837                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1838                 release_sock(&smc->sk);
1839                 return answ;
1840         }
1841         switch (cmd) {
1842         case SIOCINQ: /* same as FIONREAD */
1843                 if (smc->sk.sk_state == SMC_LISTEN) {
1844                         release_sock(&smc->sk);
1845                         return -EINVAL;
1846                 }
1847                 if (smc->sk.sk_state == SMC_INIT ||
1848                     smc->sk.sk_state == SMC_CLOSED)
1849                         answ = 0;
1850                 else
1851                         answ = atomic_read(&smc->conn.bytes_to_rcv);
1852                 break;
1853         case SIOCOUTQ:
1854                 /* output queue size (not send + not acked) */
1855                 if (smc->sk.sk_state == SMC_LISTEN) {
1856                         release_sock(&smc->sk);
1857                         return -EINVAL;
1858                 }
1859                 if (smc->sk.sk_state == SMC_INIT ||
1860                     smc->sk.sk_state == SMC_CLOSED)
1861                         answ = 0;
1862                 else
1863                         answ = smc->conn.sndbuf_desc->len -
1864                                         atomic_read(&smc->conn.sndbuf_space);
1865                 break;
1866         case SIOCOUTQNSD:
1867                 /* output queue size (not send only) */
1868                 if (smc->sk.sk_state == SMC_LISTEN) {
1869                         release_sock(&smc->sk);
1870                         return -EINVAL;
1871                 }
1872                 if (smc->sk.sk_state == SMC_INIT ||
1873                     smc->sk.sk_state == SMC_CLOSED)
1874                         answ = 0;
1875                 else
1876                         answ = smc_tx_prepared_sends(&smc->conn);
1877                 break;
1878         case SIOCATMARK:
1879                 if (smc->sk.sk_state == SMC_LISTEN) {
1880                         release_sock(&smc->sk);
1881                         return -EINVAL;
1882                 }
1883                 if (smc->sk.sk_state == SMC_INIT ||
1884                     smc->sk.sk_state == SMC_CLOSED) {
1885                         answ = 0;
1886                 } else {
1887                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
1888                         smc_curs_copy(&urg, &conn->urg_curs, conn);
1889                         answ = smc_curs_diff(conn->rmb_desc->len,
1890                                              &cons, &urg) == 1;
1891                 }
1892                 break;
1893         default:
1894                 release_sock(&smc->sk);
1895                 return -ENOIOCTLCMD;
1896         }
1897         release_sock(&smc->sk);
1898
1899         return put_user(answ, (int __user *)arg);
1900 }
1901
1902 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1903                             int offset, size_t size, int flags)
1904 {
1905         struct sock *sk = sock->sk;
1906         struct smc_sock *smc;
1907         int rc = -EPIPE;
1908
1909         smc = smc_sk(sk);
1910         lock_sock(sk);
1911         if (sk->sk_state != SMC_ACTIVE) {
1912                 release_sock(sk);
1913                 goto out;
1914         }
1915         release_sock(sk);
1916         if (smc->use_fallback)
1917                 rc = kernel_sendpage(smc->clcsock, page, offset,
1918                                      size, flags);
1919         else
1920                 rc = sock_no_sendpage(sock, page, offset, size, flags);
1921
1922 out:
1923         return rc;
1924 }
1925
1926 /* Map the affected portions of the rmbe into an spd, note the number of bytes
1927  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
1928  * updates till whenever a respective page has been fully processed.
1929  * Note that subsequent recv() calls have to wait till all splice() processing
1930  * completed.
1931  */
1932 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1933                                struct pipe_inode_info *pipe, size_t len,
1934                                unsigned int flags)
1935 {
1936         struct sock *sk = sock->sk;
1937         struct smc_sock *smc;
1938         int rc = -ENOTCONN;
1939
1940         smc = smc_sk(sk);
1941         lock_sock(sk);
1942         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1943                 /* socket was connected before, no more data to read */
1944                 rc = 0;
1945                 goto out;
1946         }
1947         if (sk->sk_state == SMC_INIT ||
1948             sk->sk_state == SMC_LISTEN ||
1949             sk->sk_state == SMC_CLOSED)
1950                 goto out;
1951
1952         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1953                 rc = 0;
1954                 goto out;
1955         }
1956
1957         if (smc->use_fallback) {
1958                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
1959                                                     pipe, len, flags);
1960         } else {
1961                 if (*ppos) {
1962                         rc = -ESPIPE;
1963                         goto out;
1964                 }
1965                 if (flags & SPLICE_F_NONBLOCK)
1966                         flags = MSG_DONTWAIT;
1967                 else
1968                         flags = 0;
1969                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
1970         }
1971 out:
1972         release_sock(sk);
1973
1974         return rc;
1975 }
1976
1977 /* must look like tcp */
1978 static const struct proto_ops smc_sock_ops = {
1979         .family         = PF_SMC,
1980         .owner          = THIS_MODULE,
1981         .release        = smc_release,
1982         .bind           = smc_bind,
1983         .connect        = smc_connect,
1984         .socketpair     = sock_no_socketpair,
1985         .accept         = smc_accept,
1986         .getname        = smc_getname,
1987         .poll           = smc_poll,
1988         .ioctl          = smc_ioctl,
1989         .listen         = smc_listen,
1990         .shutdown       = smc_shutdown,
1991         .setsockopt     = smc_setsockopt,
1992         .getsockopt     = smc_getsockopt,
1993         .sendmsg        = smc_sendmsg,
1994         .recvmsg        = smc_recvmsg,
1995         .mmap           = sock_no_mmap,
1996         .sendpage       = smc_sendpage,
1997         .splice_read    = smc_splice_read,
1998 };
1999
2000 static int smc_create(struct net *net, struct socket *sock, int protocol,
2001                       int kern)
2002 {
2003         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
2004         struct smc_sock *smc;
2005         struct sock *sk;
2006         int rc;
2007
2008         rc = -ESOCKTNOSUPPORT;
2009         if (sock->type != SOCK_STREAM)
2010                 goto out;
2011
2012         rc = -EPROTONOSUPPORT;
2013         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
2014                 goto out;
2015
2016         rc = -ENOBUFS;
2017         sock->ops = &smc_sock_ops;
2018         sk = smc_sock_alloc(net, sock, protocol);
2019         if (!sk)
2020                 goto out;
2021
2022         /* create internal TCP socket for CLC handshake and fallback */
2023         smc = smc_sk(sk);
2024         smc->use_fallback = false; /* assume rdma capability first */
2025         smc->fallback_rsn = 0;
2026         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2027                               &smc->clcsock);
2028         if (rc) {
2029                 sk_common_release(sk);
2030                 goto out;
2031         }
2032         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2033         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
2034
2035 out:
2036         return rc;
2037 }
2038
2039 static const struct net_proto_family smc_sock_family_ops = {
2040         .family = PF_SMC,
2041         .owner  = THIS_MODULE,
2042         .create = smc_create,
2043 };
2044
2045 unsigned int smc_net_id;
2046
2047 static __net_init int smc_net_init(struct net *net)
2048 {
2049         return smc_pnet_net_init(net);
2050 }
2051
2052 static void __net_exit smc_net_exit(struct net *net)
2053 {
2054         smc_pnet_net_exit(net);
2055 }
2056
2057 static struct pernet_operations smc_net_ops = {
2058         .init = smc_net_init,
2059         .exit = smc_net_exit,
2060         .id   = &smc_net_id,
2061         .size = sizeof(struct smc_net),
2062 };
2063
2064 static int __init smc_init(void)
2065 {
2066         int rc;
2067
2068         rc = register_pernet_subsys(&smc_net_ops);
2069         if (rc)
2070                 return rc;
2071
2072         rc = smc_pnet_init();
2073         if (rc)
2074                 goto out_pernet_subsys;
2075
2076         rc = smc_llc_init();
2077         if (rc) {
2078                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2079                 goto out_pnet;
2080         }
2081
2082         rc = smc_cdc_init();
2083         if (rc) {
2084                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2085                 goto out_pnet;
2086         }
2087
2088         rc = proto_register(&smc_proto, 1);
2089         if (rc) {
2090                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2091                 goto out_pnet;
2092         }
2093
2094         rc = proto_register(&smc_proto6, 1);
2095         if (rc) {
2096                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2097                 goto out_proto;
2098         }
2099
2100         rc = sock_register(&smc_sock_family_ops);
2101         if (rc) {
2102                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2103                 goto out_proto6;
2104         }
2105         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2106         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2107
2108         rc = smc_ib_register_client();
2109         if (rc) {
2110                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2111                 goto out_sock;
2112         }
2113
2114         static_branch_enable(&tcp_have_smc);
2115         return 0;
2116
2117 out_sock:
2118         sock_unregister(PF_SMC);
2119 out_proto6:
2120         proto_unregister(&smc_proto6);
2121 out_proto:
2122         proto_unregister(&smc_proto);
2123 out_pnet:
2124         smc_pnet_exit();
2125 out_pernet_subsys:
2126         unregister_pernet_subsys(&smc_net_ops);
2127
2128         return rc;
2129 }
2130
2131 static void __exit smc_exit(void)
2132 {
2133         smc_core_exit();
2134         static_branch_disable(&tcp_have_smc);
2135         smc_ib_unregister_client();
2136         sock_unregister(PF_SMC);
2137         proto_unregister(&smc_proto6);
2138         proto_unregister(&smc_proto);
2139         smc_pnet_exit();
2140         unregister_pernet_subsys(&smc_net_ops);
2141 }
2142
2143 module_init(smc_init);
2144 module_exit(smc_exit);
2145
2146 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2147 MODULE_DESCRIPTION("smc socket address family");
2148 MODULE_LICENSE("GPL");
2149 MODULE_ALIAS_NETPROTO(PF_SMC);