net/smc/af_smc.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  *  Shared Memory Communications over RDMA (SMC-R) and RoCE
   4  *
   5  *  AF_SMC protocol family socket handler keeping the AF_INET sock address type
   6  *  applies to SOCK_STREAM sockets only
   7  *  offers an alternative communication option for TCP-protocol sockets
   8  *  applicable with RoCE-cards only
   9  *
  10  *  Initial restrictions:
  11  *    - support for alternate links postponed
  12  *
  13  *  Copyright IBM Corp. 2016, 2018
  14  *
  15  *  Author(s):  Ursula Braun <ubraun@linux.vnet.ibm.com>
  16  *              based on prototype from Frank Blaschka
  17  */
  18
  19 #define KMSG_COMPONENT "smc"
  20 #define pr_fmt(fmt) KMSG_COMPONENT ": " fmt
  21
  22 #include <linux/module.h>
  23 #include <linux/socket.h>
  24 #include <linux/workqueue.h>
  25 #include <linux/in.h>
  26 #include <linux/sched/signal.h>
  27 #include <linux/if_vlan.h>
  28 #include <linux/rcupdate_wait.h>
  29
  30 #include <net/sock.h>
  31 #include <net/tcp.h>
  32 #include <net/smc.h>
  33 #include <asm/ioctls.h>
  34
  35 #include <net/net_namespace.h>
  36 #include <net/netns/generic.h>
  37 #include "smc_netns.h"
  38
  39 #include "smc.h"
  40 #include "smc_clc.h"
  41 #include "smc_llc.h"
  42 #include "smc_cdc.h"
  43 #include "smc_core.h"
  44 #include "smc_ib.h"
  45 #include "smc_ism.h"
  46 #include "smc_pnet.h"
  47 #include "smc_tx.h"
  48 #include "smc_rx.h"
  49 #include "smc_close.h"
  50
  51 static DEFINE_MUTEX(smc_server_lgr_pending);    /* serialize link group
  52                                                  * creation on server
  53                                                  */
  54 static DEFINE_MUTEX(smc_client_lgr_pending);    /* serialize link group
  55                                                  * creation on client
  56                                                  */
  57
  58 static void smc_tcp_listen_work(struct work_struct *);
  59 static void smc_connect_work(struct work_struct *);
  60
  61 static void smc_set_keepalive(struct sock *sk, int val)
  62 {
  63         struct smc_sock *smc = smc_sk(sk);
  64
  65         smc->clcsock->sk->sk_prot->keepalive(smc->clcsock->sk, val);
  66 }
  67
  68 static struct smc_hashinfo smc_v4_hashinfo = {
  69         .lock = __RW_LOCK_UNLOCKED(smc_v4_hashinfo.lock),
  70 };
  71
  72 static struct smc_hashinfo smc_v6_hashinfo = {
  73         .lock = __RW_LOCK_UNLOCKED(smc_v6_hashinfo.lock),
  74 };
  75
  76 int smc_hash_sk(struct sock *sk)
  77 {
  78         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  79         struct hlist_head *head;
  80
  81         head = &h->ht;
  82
  83         write_lock_bh(&h->lock);
  84         sk_add_node(sk, head);
  85         sock_prot_inuse_add(sock_net(sk), sk->sk_prot, 1);
  86         write_unlock_bh(&h->lock);
  87
  88         return 0;
  89 }
  90 EXPORT_SYMBOL_GPL(smc_hash_sk);
  91
  92 void smc_unhash_sk(struct sock *sk)
  93 {
  94         struct smc_hashinfo *h = sk->sk_prot->h.smc_hash;
  95
  96         write_lock_bh(&h->lock);
  97         if (sk_del_node_init(sk))
  98                 sock_prot_inuse_add(sock_net(sk), sk->sk_prot, -1);
  99         write_unlock_bh(&h->lock);
 100 }
 101 EXPORT_SYMBOL_GPL(smc_unhash_sk);
 102
 103 struct proto smc_proto = {
 104         .name           = "SMC",
 105         .owner          = THIS_MODULE,
 106         .keepalive      = smc_set_keepalive,
 107         .hash           = smc_hash_sk,
 108         .unhash         = smc_unhash_sk,
 109         .obj_size       = sizeof(struct smc_sock),
 110         .h.smc_hash     = &smc_v4_hashinfo,
 111         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 112 };
 113 EXPORT_SYMBOL_GPL(smc_proto);
 114
 115 struct proto smc_proto6 = {
 116         .name           = "SMC6",
 117         .owner          = THIS_MODULE,
 118         .keepalive      = smc_set_keepalive,
 119         .hash           = smc_hash_sk,
 120         .unhash         = smc_unhash_sk,
 121         .obj_size       = sizeof(struct smc_sock),
 122         .h.smc_hash     = &smc_v6_hashinfo,
 123         .slab_flags     = SLAB_TYPESAFE_BY_RCU,
 124 };
 125 EXPORT_SYMBOL_GPL(smc_proto6);
 126
 127 static void smc_restore_fallback_changes(struct smc_sock *smc)
 128 {
 129         if (smc->clcsock->file) { /* non-accepted sockets have no file yet */
 130                 smc->clcsock->file->private_data = smc->sk.sk_socket;
 131                 smc->clcsock->file = NULL;
 132         }
 133 }
 134
 135 static int __smc_release(struct smc_sock *smc)
 136 {
 137         struct sock *sk = &smc->sk;
 138         int rc = 0;
 139
 140         if (!smc->use_fallback) {
 141                 rc = smc_close_active(smc);
 142                 sock_set_flag(sk, SOCK_DEAD);
 143                 sk->sk_shutdown |= SHUTDOWN_MASK;
 144         } else {
 145                 if (sk->sk_state != SMC_LISTEN && sk->sk_state != SMC_INIT)
 146                         sock_put(sk); /* passive closing */
 147                 if (sk->sk_state == SMC_LISTEN) {
 148                         /* wake up clcsock accept */
 149                         rc = kernel_sock_shutdown(smc->clcsock, SHUT_RDWR);
 150                 }
 151                 sk->sk_state = SMC_CLOSED;
 152                 sk->sk_state_change(sk);
 153                 smc_restore_fallback_changes(smc);
 154         }
 155
 156         sk->sk_prot->unhash(sk);
 157
 158         if (sk->sk_state == SMC_CLOSED) {
 159                 if (smc->clcsock) {
 160                         release_sock(sk);
 161                         smc_clcsock_release(smc);
 162                         lock_sock(sk);
 163                 }
 164                 if (!smc->use_fallback)
 165                         smc_conn_free(&smc->conn);
 166         }
 167
 168         return rc;
 169 }
 170
 171 static int smc_release(struct socket *sock)
 172 {
 173         struct sock *sk = sock->sk;
 174         struct smc_sock *smc;
 175         int rc = 0;
 176
 177         if (!sk)
 178                 goto out;
 179
 180         sock_hold(sk); /* sock_put below */
 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         flush_work(&smc->connect_work);
 187
 188         if (sk->sk_state == SMC_LISTEN)
 189                 /* smc_close_non_accepted() is called and acquires
 190                  * sock lock for child sockets again
 191                  */
 192                 lock_sock_nested(sk, SINGLE_DEPTH_NESTING);
 193         else
 194                 lock_sock(sk);
 195
 196         rc = __smc_release(smc);
 197
 198         /* detach socket */
 199         sock_orphan(sk);
 200         sock->sk = NULL;
 201         release_sock(sk);
 202
 203         sock_put(sk); /* sock_hold above */
 204         sock_put(sk); /* final sock_put */
 205 out:
 206         return rc;
 207 }
 208
 209 static void smc_destruct(struct sock *sk)
 210 {
 211         if (sk->sk_state != SMC_CLOSED)
 212                 return;
 213         if (!sock_flag(sk, SOCK_DEAD))
 214                 return;
 215
 216         sk_refcnt_debug_dec(sk);
 217 }
 218
 219 static struct sock *smc_sock_alloc(struct net *net, struct socket *sock,
 220                                    int protocol)
 221 {
 222         struct smc_sock *smc;
 223         struct proto *prot;
 224         struct sock *sk;
 225
 226         prot = (protocol == SMCPROTO_SMC6) ? &smc_proto6 : &smc_proto;
 227         sk = sk_alloc(net, PF_SMC, GFP_KERNEL, prot, 0);
 228         if (!sk)
 229                 return NULL;
 230
 231         sock_init_data(sock, sk); /* sets sk_refcnt to 1 */
 232         sk->sk_state = SMC_INIT;
 233         sk->sk_destruct = smc_destruct;
 234         sk->sk_protocol = protocol;
 235         smc = smc_sk(sk);
 236         INIT_WORK(&smc->tcp_listen_work, smc_tcp_listen_work);
 237         INIT_WORK(&smc->connect_work, smc_connect_work);
 238         INIT_DELAYED_WORK(&smc->conn.tx_work, smc_tx_work);
 239         INIT_LIST_HEAD(&smc->accept_q);
 240         spin_lock_init(&smc->accept_q_lock);
 241         spin_lock_init(&smc->conn.send_lock);
 242         sk->sk_prot->hash(sk);
 243         sk_refcnt_debug_inc(sk);
 244         mutex_init(&smc->clcsock_release_lock);
 245
 246         return sk;
 247 }
 248
 249 static int smc_bind(struct socket *sock, struct sockaddr *uaddr,
 250                     int addr_len)
 251 {
 252         struct sockaddr_in *addr = (struct sockaddr_in *)uaddr;
 253         struct sock *sk = sock->sk;
 254         struct smc_sock *smc;
 255         int rc;
 256
 257         smc = smc_sk(sk);
 258
 259         /* replicate tests from inet_bind(), to be safe wrt. future changes */
 260         rc = -EINVAL;
 261         if (addr_len < sizeof(struct sockaddr_in))
 262                 goto out;
 263
 264         rc = -EAFNOSUPPORT;
 265         if (addr->sin_family != AF_INET &&
 266             addr->sin_family != AF_INET6 &&
 267             addr->sin_family != AF_UNSPEC)
 268                 goto out;
 269         /* accept AF_UNSPEC (mapped to AF_INET) only if s_addr is INADDR_ANY */
 270         if (addr->sin_family == AF_UNSPEC &&
 271             addr->sin_addr.s_addr != htonl(INADDR_ANY))
 272                 goto out;
 273
 274         lock_sock(sk);
 275
 276         /* Check if socket is already active */
 277         rc = -EINVAL;
 278         if (sk->sk_state != SMC_INIT || smc->connect_nonblock)
 279                 goto out_rel;
 280
 281         smc->clcsock->sk->sk_reuse = sk->sk_reuse;
 282         rc = kernel_bind(smc->clcsock, uaddr, addr_len);
 283
 284 out_rel:
 285         release_sock(sk);
 286 out:
 287         return rc;
 288 }
 289
 290 static void smc_copy_sock_settings(struct sock *nsk, struct sock *osk,
 291                                    unsigned long mask)
 292 {
 293         /* options we don't get control via setsockopt for */
 294         nsk->sk_type = osk->sk_type;
 295         nsk->sk_sndbuf = osk->sk_sndbuf;
 296         nsk->sk_rcvbuf = osk->sk_rcvbuf;
 297         nsk->sk_sndtimeo = osk->sk_sndtimeo;
 298         nsk->sk_rcvtimeo = osk->sk_rcvtimeo;
 299         nsk->sk_mark = osk->sk_mark;
 300         nsk->sk_priority = osk->sk_priority;
 301         nsk->sk_rcvlowat = osk->sk_rcvlowat;
 302         nsk->sk_bound_dev_if = osk->sk_bound_dev_if;
 303         nsk->sk_err = osk->sk_err;
 304
 305         nsk->sk_flags &= ~mask;
 306         nsk->sk_flags |= osk->sk_flags & mask;
 307 }
 308
 309 #define SK_FLAGS_SMC_TO_CLC ((1UL << SOCK_URGINLINE) | \
 310                              (1UL << SOCK_KEEPOPEN) | \
 311                              (1UL << SOCK_LINGER) | \
 312                              (1UL << SOCK_BROADCAST) | \
 313                              (1UL << SOCK_TIMESTAMP) | \
 314                              (1UL << SOCK_DBG) | \
 315                              (1UL << SOCK_RCVTSTAMP) | \
 316                              (1UL << SOCK_RCVTSTAMPNS) | \
 317                              (1UL << SOCK_LOCALROUTE) | \
 318                              (1UL << SOCK_TIMESTAMPING_RX_SOFTWARE) | \
 319                              (1UL << SOCK_RXQ_OVFL) | \
 320                              (1UL << SOCK_WIFI_STATUS) | \
 321                              (1UL << SOCK_NOFCS) | \
 322                              (1UL << SOCK_FILTER_LOCKED) | \
 323                              (1UL << SOCK_TSTAMP_NEW))
 324 /* copy only relevant settings and flags of SOL_SOCKET level from smc to
 325  * clc socket (since smc is not called for these options from net/core)
 326  */
 327 static void smc_copy_sock_settings_to_clc(struct smc_sock *smc)
 328 {
 329         smc_copy_sock_settings(smc->clcsock->sk, &smc->sk, SK_FLAGS_SMC_TO_CLC);
 330 }
 331
 332 #define SK_FLAGS_CLC_TO_SMC ((1UL << SOCK_URGINLINE) | \
 333                              (1UL << SOCK_KEEPOPEN) | \
 334                              (1UL << SOCK_LINGER) | \
 335                              (1UL << SOCK_DBG))
 336 /* copy only settings and flags relevant for smc from clc to smc socket */
 337 static void smc_copy_sock_settings_to_smc(struct smc_sock *smc)
 338 {
 339         smc_copy_sock_settings(&smc->sk, smc->clcsock->sk, SK_FLAGS_CLC_TO_SMC);
 340 }
 341
 342 /* register the new rmb on all links */
 343 static int smcr_lgr_reg_rmbs(struct smc_link *link,
 344                              struct smc_buf_desc *rmb_desc)
 345 {
 346         struct smc_link_group *lgr = link->lgr;
 347         int i, rc = 0;
 348
 349         rc = smc_llc_flow_initiate(lgr, SMC_LLC_FLOW_RKEY);
 350         if (rc)
 351                 return rc;
 352         /* protect against parallel smc_llc_cli_rkey_exchange() and
 353          * parallel smcr_link_reg_rmb()
 354          */
 355         mutex_lock(&lgr->llc_conf_mutex);
 356         for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 357                 if (!smc_link_active(&lgr->lnk[i]))
 358                         continue;
 359                 rc = smcr_link_reg_rmb(&lgr->lnk[i], rmb_desc);
 360                 if (rc)
 361                         goto out;
 362         }
 363
 364         /* exchange confirm_rkey msg with peer */
 365         rc = smc_llc_do_confirm_rkey(link, rmb_desc);
 366         if (rc) {
 367                 rc = -EFAULT;
 368                 goto out;
 369         }
 370         rmb_desc->is_conf_rkey = true;
 371 out:
 372         mutex_unlock(&lgr->llc_conf_mutex);
 373         smc_llc_flow_stop(lgr, &lgr->llc_flow_lcl);
 374         return rc;
 375 }
 376
 377 static int smcr_clnt_conf_first_link(struct smc_sock *smc)
 378 {
 379         struct smc_link *link = smc->conn.lnk;
 380         struct smc_llc_qentry *qentry;
 381         int rc;
 382
 383         /* receive CONFIRM LINK request from server over RoCE fabric */
 384         qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
 385                               SMC_LLC_CONFIRM_LINK);
 386         if (!qentry) {
 387                 struct smc_clc_msg_decline dclc;
 388
 389                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 390                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 391                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
 392         }
 393         smc_llc_save_peer_uid(qentry);
 394         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_REQ);
 395         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
 396         if (rc)
 397                 return SMC_CLC_DECL_RMBE_EC;
 398
 399         rc = smc_ib_modify_qp_rts(link);
 400         if (rc)
 401                 return SMC_CLC_DECL_ERR_RDYLNK;
 402
 403         smc_wr_remember_qp_attr(link);
 404
 405         if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
 406                 return SMC_CLC_DECL_ERR_REGRMB;
 407
 408         /* confirm_rkey is implicit on 1st contact */
 409         smc->conn.rmb_desc->is_conf_rkey = true;
 410
 411         /* send CONFIRM LINK response over RoCE fabric */
 412         rc = smc_llc_send_confirm_link(link, SMC_LLC_RESP);
 413         if (rc < 0)
 414                 return SMC_CLC_DECL_TIMEOUT_CL;
 415
 416         smc_llc_link_active(link);
 417         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
 418
 419         /* optional 2nd link, receive ADD LINK request from server */
 420         qentry = smc_llc_wait(link->lgr, NULL, SMC_LLC_WAIT_TIME,
 421                               SMC_LLC_ADD_LINK);
 422         if (!qentry) {
 423                 struct smc_clc_msg_decline dclc;
 424
 425                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
 426                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
 427                 if (rc == -EAGAIN)
 428                         rc = 0; /* no DECLINE received, go with one link */
 429                 return rc;
 430         }
 431         smc_llc_flow_qentry_clr(&link->lgr->llc_flow_lcl);
 432         smc_llc_cli_add_link(link, qentry);
 433         return 0;
 434 }
 435
 436 static void smcr_conn_save_peer_info(struct smc_sock *smc,
 437                                      struct smc_clc_msg_accept_confirm *clc)
 438 {
 439         int bufsize = smc_uncompress_bufsize(clc->r0.rmbe_size);
 440
 441         smc->conn.peer_rmbe_idx = clc->r0.rmbe_idx;
 442         smc->conn.local_tx_ctrl.token = ntohl(clc->r0.rmbe_alert_token);
 443         smc->conn.peer_rmbe_size = bufsize;
 444         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 445         smc->conn.tx_off = bufsize * (smc->conn.peer_rmbe_idx - 1);
 446 }
 447
 448 static void smcd_conn_save_peer_info(struct smc_sock *smc,
 449                                      struct smc_clc_msg_accept_confirm *clc)
 450 {
 451         int bufsize = smc_uncompress_bufsize(clc->d0.dmbe_size);
 452
 453         smc->conn.peer_rmbe_idx = clc->d0.dmbe_idx;
 454         smc->conn.peer_token = clc->d0.token;
 455         /* msg header takes up space in the buffer */
 456         smc->conn.peer_rmbe_size = bufsize - sizeof(struct smcd_cdc_msg);
 457         atomic_set(&smc->conn.peer_rmbe_space, smc->conn.peer_rmbe_size);
 458         smc->conn.tx_off = bufsize * smc->conn.peer_rmbe_idx;
 459 }
 460
 461 static void smc_conn_save_peer_info(struct smc_sock *smc,
 462                                     struct smc_clc_msg_accept_confirm *clc)
 463 {
 464         if (smc->conn.lgr->is_smcd)
 465                 smcd_conn_save_peer_info(smc, clc);
 466         else
 467                 smcr_conn_save_peer_info(smc, clc);
 468 }
 469
 470 static void smc_link_save_peer_info(struct smc_link *link,
 471                                     struct smc_clc_msg_accept_confirm *clc)
 472 {
 473         link->peer_qpn = ntoh24(clc->r0.qpn);
 474         memcpy(link->peer_gid, clc->r0.lcl.gid, SMC_GID_SIZE);
 475         memcpy(link->peer_mac, clc->r0.lcl.mac, sizeof(link->peer_mac));
 476         link->peer_psn = ntoh24(clc->r0.psn);
 477         link->peer_mtu = clc->r0.qp_mtu;
 478 }
 479
 480 static void smc_switch_to_fallback(struct smc_sock *smc)
 481 {
 482         smc->use_fallback = true;
 483         if (smc->sk.sk_socket && smc->sk.sk_socket->file) {
 484                 smc->clcsock->file = smc->sk.sk_socket->file;
 485                 smc->clcsock->file->private_data = smc->clcsock;
 486                 smc->clcsock->wq.fasync_list =
 487                         smc->sk.sk_socket->wq.fasync_list;
 488         }
 489 }
 490
 491 /* fall back during connect */
 492 static int smc_connect_fallback(struct smc_sock *smc, int reason_code)
 493 {
 494         smc_switch_to_fallback(smc);
 495         smc->fallback_rsn = reason_code;
 496         smc_copy_sock_settings_to_clc(smc);
 497         smc->connect_nonblock = 0;
 498         if (smc->sk.sk_state == SMC_INIT)
 499                 smc->sk.sk_state = SMC_ACTIVE;
 500         return 0;
 501 }
 502
 503 /* decline and fall back during connect */
 504 static int smc_connect_decline_fallback(struct smc_sock *smc, int reason_code)
 505 {
 506         int rc;
 507
 508         if (reason_code < 0) { /* error, fallback is not possible */
 509                 if (smc->sk.sk_state == SMC_INIT)
 510                         sock_put(&smc->sk); /* passive closing */
 511                 return reason_code;
 512         }
 513         if (reason_code != SMC_CLC_DECL_PEERDECL) {
 514                 rc = smc_clc_send_decline(smc, reason_code);
 515                 if (rc < 0) {
 516                         if (smc->sk.sk_state == SMC_INIT)
 517                                 sock_put(&smc->sk); /* passive closing */
 518                         return rc;
 519                 }
 520         }
 521         return smc_connect_fallback(smc, reason_code);
 522 }
 523
 524 /* abort connecting */
 525 static int smc_connect_abort(struct smc_sock *smc, int reason_code,
 526                              int local_first)
 527 {
 528         bool is_smcd = smc->conn.lgr->is_smcd;
 529
 530         if (local_first)
 531                 smc_lgr_cleanup_early(&smc->conn);
 532         else
 533                 smc_conn_free(&smc->conn);
 534         if (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->connect_nonblock = 0;
 541         return reason_code;
 542 }
 543
 544 /* check if there is a rdma device available for this connection. */
 545 /* called for connect and listen */
 546 static int smc_find_rdma_device(struct smc_sock *smc, struct smc_init_info *ini)
 547 {
 548         /* PNET table look up: search active ib_device and port
 549          * within same PNETID that also contains the ethernet device
 550          * used for the internal TCP socket
 551          */
 552         smc_pnet_find_roce_resource(smc->clcsock->sk, ini);
 553         if (!ini->ib_dev)
 554                 return SMC_CLC_DECL_NOSMCRDEV;
 555         return 0;
 556 }
 557
 558 /* check if there is an ISM device available for this connection. */
 559 /* called for connect and listen */
 560 static int smc_find_ism_device(struct smc_sock *smc, struct smc_init_info *ini)
 561 {
 562         /* Find ISM device with same PNETID as connecting interface  */
 563         smc_pnet_find_ism_resource(smc->clcsock->sk, ini);
 564         if (!ini->ism_dev)
 565                 return SMC_CLC_DECL_NOSMCDDEV;
 566         return 0;
 567 }
 568
 569 /* Check for VLAN ID and register it on ISM device just for CLC handshake */
 570 static int smc_connect_ism_vlan_setup(struct smc_sock *smc,
 571                                       struct smc_init_info *ini)
 572 {
 573         if (ini->vlan_id && smc_ism_get_vlan(ini->ism_dev, ini->vlan_id))
 574                 return SMC_CLC_DECL_ISMVLANERR;
 575         return 0;
 576 }
 577
 578 /* cleanup temporary VLAN ID registration used for CLC handshake. If ISM is
 579  * used, the VLAN ID will be registered again during the connection setup.
 580  */
 581 static int smc_connect_ism_vlan_cleanup(struct smc_sock *smc, bool is_smcd,
 582                                         struct smc_init_info *ini)
 583 {
 584         if (!is_smcd)
 585                 return 0;
 586         if (ini->vlan_id && smc_ism_put_vlan(ini->ism_dev, ini->vlan_id))
 587                 return SMC_CLC_DECL_CNFERR;
 588         return 0;
 589 }
 590
 591 /* CLC handshake during connect */
 592 static int smc_connect_clc(struct smc_sock *smc, int smc_type,
 593                            struct smc_clc_msg_accept_confirm *aclc,
 594                            struct smc_init_info *ini)
 595 {
 596         int rc = 0;
 597
 598         /* do inband token exchange */
 599         rc = smc_clc_send_proposal(smc, smc_type, ini);
 600         if (rc)
 601                 return rc;
 602         /* receive SMC Accept CLC message */
 603         return smc_clc_wait_msg(smc, aclc, sizeof(*aclc), SMC_CLC_ACCEPT,
 604                                 CLC_WAIT_TIME);
 605 }
 606
 607 /* setup for RDMA connection of client */
 608 static int smc_connect_rdma(struct smc_sock *smc,
 609                             struct smc_clc_msg_accept_confirm *aclc,
 610                             struct smc_init_info *ini)
 611 {
 612         int i, reason_code = 0;
 613         struct smc_link *link;
 614
 615         ini->is_smcd = false;
 616         ini->ib_lcl = &aclc->r0.lcl;
 617         ini->ib_clcqpn = ntoh24(aclc->r0.qpn);
 618         ini->first_contact_peer = aclc->hdr.flag;
 619
 620         mutex_lock(&smc_client_lgr_pending);
 621         reason_code = smc_conn_create(smc, ini);
 622         if (reason_code) {
 623                 mutex_unlock(&smc_client_lgr_pending);
 624                 return reason_code;
 625         }
 626
 627         smc_conn_save_peer_info(smc, aclc);
 628
 629         if (ini->first_contact_local) {
 630                 link = smc->conn.lnk;
 631         } else {
 632                 /* set link that was assigned by server */
 633                 link = NULL;
 634                 for (i = 0; i < SMC_LINKS_PER_LGR_MAX; i++) {
 635                         struct smc_link *l = &smc->conn.lgr->lnk[i];
 636
 637                         if (l->peer_qpn == ntoh24(aclc->r0.qpn) &&
 638                             !memcmp(l->peer_gid, &aclc->r0.lcl.gid,
 639                                     SMC_GID_SIZE) &&
 640                             !memcmp(l->peer_mac, &aclc->r0.lcl.mac,
 641                                     sizeof(l->peer_mac))) {
 642                                 link = l;
 643                                 break;
 644                         }
 645                 }
 646                 if (!link)
 647                         return smc_connect_abort(smc, SMC_CLC_DECL_NOSRVLINK,
 648                                                  ini->first_contact_local);
 649                 smc->conn.lnk = link;
 650         }
 651
 652         /* create send buffer and rmb */
 653         if (smc_buf_create(smc, false))
 654                 return smc_connect_abort(smc, SMC_CLC_DECL_MEM,
 655                                          ini->first_contact_local);
 656
 657         if (ini->first_contact_local)
 658                 smc_link_save_peer_info(link, aclc);
 659
 660         if (smc_rmb_rtoken_handling(&smc->conn, link, aclc))
 661                 return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RTOK,
 662                                          ini->first_contact_local);
 663
 664         smc_close_init(smc);
 665         smc_rx_init(smc);
 666
 667         if (ini->first_contact_local) {
 668                 if (smc_ib_ready_link(link))
 669                         return smc_connect_abort(smc, SMC_CLC_DECL_ERR_RDYLNK,
 670                                                  ini->first_contact_local);
 671         } else {
 672                 if (smcr_lgr_reg_rmbs(link, smc->conn.rmb_desc))
 673                         return smc_connect_abort(smc, SMC_CLC_DECL_ERR_REGRMB,
 674                                                  ini->first_contact_local);
 675         }
 676         smc_rmb_sync_sg_for_device(&smc->conn);
 677
 678         reason_code = smc_clc_send_confirm(smc);
 679         if (reason_code)
 680                 return smc_connect_abort(smc, reason_code,
 681                                          ini->first_contact_local);
 682
 683         smc_tx_init(smc);
 684
 685         if (ini->first_contact_local) {
 686                 /* QP confirmation over RoCE fabric */
 687                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
 688                 reason_code = smcr_clnt_conf_first_link(smc);
 689                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
 690                 if (reason_code)
 691                         return smc_connect_abort(smc, reason_code,
 692                                                  ini->first_contact_local);
 693         }
 694         mutex_unlock(&smc_client_lgr_pending);
 695
 696         smc_copy_sock_settings_to_clc(smc);
 697         smc->connect_nonblock = 0;
 698         if (smc->sk.sk_state == SMC_INIT)
 699                 smc->sk.sk_state = SMC_ACTIVE;
 700
 701         return 0;
 702 }
 703
 704 /* setup for ISM connection of client */
 705 static int smc_connect_ism(struct smc_sock *smc,
 706                            struct smc_clc_msg_accept_confirm *aclc,
 707                            struct smc_init_info *ini)
 708 {
 709         int rc = 0;
 710
 711         ini->is_smcd = true;
 712         ini->ism_peer_gid = aclc->d0.gid;
 713         ini->first_contact_peer = aclc->hdr.flag;
 714
 715         /* there is only one lgr role for SMC-D; use server lock */
 716         mutex_lock(&smc_server_lgr_pending);
 717         rc = smc_conn_create(smc, ini);
 718         if (rc) {
 719                 mutex_unlock(&smc_server_lgr_pending);
 720                 return rc;
 721         }
 722
 723         /* Create send and receive buffers */
 724         rc = smc_buf_create(smc, true);
 725         if (rc)
 726                 return smc_connect_abort(smc, (rc == -ENOSPC) ?
 727                                               SMC_CLC_DECL_MAX_DMB :
 728                                               SMC_CLC_DECL_MEM,
 729                                          ini->first_contact_local);
 730
 731         smc_conn_save_peer_info(smc, aclc);
 732         smc_close_init(smc);
 733         smc_rx_init(smc);
 734         smc_tx_init(smc);
 735
 736         rc = smc_clc_send_confirm(smc);
 737         if (rc)
 738                 return smc_connect_abort(smc, rc, ini->first_contact_local);
 739         mutex_unlock(&smc_server_lgr_pending);
 740
 741         smc_copy_sock_settings_to_clc(smc);
 742         smc->connect_nonblock = 0;
 743         if (smc->sk.sk_state == SMC_INIT)
 744                 smc->sk.sk_state = SMC_ACTIVE;
 745
 746         return 0;
 747 }
 748
 749 /* perform steps before actually connecting */
 750 static int __smc_connect(struct smc_sock *smc)
 751 {
 752         bool ism_supported = false, rdma_supported = false;
 753         struct smc_clc_msg_accept_confirm aclc;
 754         struct smc_init_info ini = {0};
 755         int smc_type;
 756         int rc = 0;
 757
 758         if (smc->use_fallback)
 759                 return smc_connect_fallback(smc, smc->fallback_rsn);
 760
 761         /* if peer has not signalled SMC-capability, fall back */
 762         if (!tcp_sk(smc->clcsock->sk)->syn_smc)
 763                 return smc_connect_fallback(smc, SMC_CLC_DECL_PEERNOSMC);
 764
 765         /* IPSec connections opt out of SMC-R optimizations */
 766         if (using_ipsec(smc))
 767                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_IPSEC);
 768
 769         /* get vlan id from IP device */
 770         if (smc_vlan_by_tcpsk(smc->clcsock, &ini))
 771                 return smc_connect_decline_fallback(smc,
 772                                                     SMC_CLC_DECL_GETVLANERR);
 773
 774         /* check if there is an ism device available */
 775         if (!smc_find_ism_device(smc, &ini) &&
 776             !smc_connect_ism_vlan_setup(smc, &ini)) {
 777                 /* ISM is supported for this connection */
 778                 ism_supported = true;
 779                 smc_type = SMC_TYPE_D;
 780         }
 781
 782         /* check if there is a rdma device available */
 783         if (!smc_find_rdma_device(smc, &ini)) {
 784                 /* RDMA is supported for this connection */
 785                 rdma_supported = true;
 786                 if (ism_supported)
 787                         smc_type = SMC_TYPE_B; /* both */
 788                 else
 789                         smc_type = SMC_TYPE_R; /* only RDMA */
 790         }
 791
 792         /* if neither ISM nor RDMA are supported, fallback */
 793         if (!rdma_supported && !ism_supported)
 794                 return smc_connect_decline_fallback(smc, SMC_CLC_DECL_NOSMCDEV);
 795
 796         /* perform CLC handshake */
 797         rc = smc_connect_clc(smc, smc_type, &aclc, &ini);
 798         if (rc) {
 799                 smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
 800                 return smc_connect_decline_fallback(smc, rc);
 801         }
 802
 803         /* depending on previous steps, connect using rdma or ism */
 804         if (rdma_supported && aclc.hdr.path == SMC_TYPE_R)
 805                 rc = smc_connect_rdma(smc, &aclc, &ini);
 806         else if (ism_supported && aclc.hdr.path == SMC_TYPE_D)
 807                 rc = smc_connect_ism(smc, &aclc, &ini);
 808         else
 809                 rc = SMC_CLC_DECL_MODEUNSUPP;
 810         if (rc) {
 811                 smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
 812                 return smc_connect_decline_fallback(smc, rc);
 813         }
 814
 815         smc_connect_ism_vlan_cleanup(smc, ism_supported, &ini);
 816         return 0;
 817 }
 818
 819 static void smc_connect_work(struct work_struct *work)
 820 {
 821         struct smc_sock *smc = container_of(work, struct smc_sock,
 822                                             connect_work);
 823         long timeo = smc->sk.sk_sndtimeo;
 824         int rc = 0;
 825
 826         if (!timeo)
 827                 timeo = MAX_SCHEDULE_TIMEOUT;
 828         lock_sock(smc->clcsock->sk);
 829         if (smc->clcsock->sk->sk_err) {
 830                 smc->sk.sk_err = smc->clcsock->sk->sk_err;
 831         } else if ((1 << smc->clcsock->sk->sk_state) &
 832                                         (TCPF_SYN_SENT | TCP_SYN_RECV)) {
 833                 rc = sk_stream_wait_connect(smc->clcsock->sk, &timeo);
 834                 if ((rc == -EPIPE) &&
 835                     ((1 << smc->clcsock->sk->sk_state) &
 836                                         (TCPF_ESTABLISHED | TCPF_CLOSE_WAIT)))
 837                         rc = 0;
 838         }
 839         release_sock(smc->clcsock->sk);
 840         lock_sock(&smc->sk);
 841         if (rc != 0 || smc->sk.sk_err) {
 842                 smc->sk.sk_state = SMC_CLOSED;
 843                 if (rc == -EPIPE || rc == -EAGAIN)
 844                         smc->sk.sk_err = EPIPE;
 845                 else if (signal_pending(current))
 846                         smc->sk.sk_err = -sock_intr_errno(timeo);
 847                 sock_put(&smc->sk); /* passive closing */
 848                 goto out;
 849         }
 850
 851         rc = __smc_connect(smc);
 852         if (rc < 0)
 853                 smc->sk.sk_err = -rc;
 854
 855 out:
 856         if (!sock_flag(&smc->sk, SOCK_DEAD)) {
 857                 if (smc->sk.sk_err) {
 858                         smc->sk.sk_state_change(&smc->sk);
 859                 } else { /* allow polling before and after fallback decision */
 860                         smc->clcsock->sk->sk_write_space(smc->clcsock->sk);
 861                         smc->sk.sk_write_space(&smc->sk);
 862                 }
 863         }
 864         release_sock(&smc->sk);
 865 }
 866
 867 static int smc_connect(struct socket *sock, struct sockaddr *addr,
 868                        int alen, int flags)
 869 {
 870         struct sock *sk = sock->sk;
 871         struct smc_sock *smc;
 872         int rc = -EINVAL;
 873
 874         smc = smc_sk(sk);
 875
 876         /* separate smc parameter checking to be safe */
 877         if (alen < sizeof(addr->sa_family))
 878                 goto out_err;
 879         if (addr->sa_family != AF_INET && addr->sa_family != AF_INET6)
 880                 goto out_err;
 881
 882         lock_sock(sk);
 883         switch (sk->sk_state) {
 884         default:
 885                 goto out;
 886         case SMC_ACTIVE:
 887                 rc = -EISCONN;
 888                 goto out;
 889         case SMC_INIT:
 890                 rc = 0;
 891                 break;
 892         }
 893
 894         smc_copy_sock_settings_to_clc(smc);
 895         tcp_sk(smc->clcsock->sk)->syn_smc = 1;
 896         if (smc->connect_nonblock) {
 897                 rc = -EALREADY;
 898                 goto out;
 899         }
 900         rc = kernel_connect(smc->clcsock, addr, alen, flags);
 901         if (rc && rc != -EINPROGRESS)
 902                 goto out;
 903
 904         sock_hold(&smc->sk); /* sock put in passive closing */
 905         if (smc->use_fallback)
 906                 goto out;
 907         if (flags & O_NONBLOCK) {
 908                 if (schedule_work(&smc->connect_work))
 909                         smc->connect_nonblock = 1;
 910                 rc = -EINPROGRESS;
 911         } else {
 912                 rc = __smc_connect(smc);
 913                 if (rc < 0)
 914                         goto out;
 915                 else
 916                         rc = 0; /* success cases including fallback */
 917         }
 918
 919 out:
 920         release_sock(sk);
 921 out_err:
 922         return rc;
 923 }
 924
 925 static int smc_clcsock_accept(struct smc_sock *lsmc, struct smc_sock **new_smc)
 926 {
 927         struct socket *new_clcsock = NULL;
 928         struct sock *lsk = &lsmc->sk;
 929         struct sock *new_sk;
 930         int rc = -EINVAL;
 931
 932         release_sock(lsk);
 933         new_sk = smc_sock_alloc(sock_net(lsk), NULL, lsk->sk_protocol);
 934         if (!new_sk) {
 935                 rc = -ENOMEM;
 936                 lsk->sk_err = ENOMEM;
 937                 *new_smc = NULL;
 938                 lock_sock(lsk);
 939                 goto out;
 940         }
 941         *new_smc = smc_sk(new_sk);
 942
 943         mutex_lock(&lsmc->clcsock_release_lock);
 944         if (lsmc->clcsock)
 945                 rc = kernel_accept(lsmc->clcsock, &new_clcsock, SOCK_NONBLOCK);
 946         mutex_unlock(&lsmc->clcsock_release_lock);
 947         lock_sock(lsk);
 948         if  (rc < 0 && rc != -EAGAIN)
 949                 lsk->sk_err = -rc;
 950         if (rc < 0 || lsk->sk_state == SMC_CLOSED) {
 951                 new_sk->sk_prot->unhash(new_sk);
 952                 if (new_clcsock)
 953                         sock_release(new_clcsock);
 954                 new_sk->sk_state = SMC_CLOSED;
 955                 sock_set_flag(new_sk, SOCK_DEAD);
 956                 sock_put(new_sk); /* final */
 957                 *new_smc = NULL;
 958                 goto out;
 959         }
 960
 961         /* new clcsock has inherited the smc listen-specific sk_data_ready
 962          * function; switch it back to the original sk_data_ready function
 963          */
 964         new_clcsock->sk->sk_data_ready = lsmc->clcsk_data_ready;
 965         (*new_smc)->clcsock = new_clcsock;
 966 out:
 967         return rc;
 968 }
 969
 970 /* add a just created sock to the accept queue of the listen sock as
 971  * candidate for a following socket accept call from user space
 972  */
 973 static void smc_accept_enqueue(struct sock *parent, struct sock *sk)
 974 {
 975         struct smc_sock *par = smc_sk(parent);
 976
 977         sock_hold(sk); /* sock_put in smc_accept_unlink () */
 978         spin_lock(&par->accept_q_lock);
 979         list_add_tail(&smc_sk(sk)->accept_q, &par->accept_q);
 980         spin_unlock(&par->accept_q_lock);
 981         sk_acceptq_added(parent);
 982 }
 983
 984 /* remove a socket from the accept queue of its parental listening socket */
 985 static void smc_accept_unlink(struct sock *sk)
 986 {
 987         struct smc_sock *par = smc_sk(sk)->listen_smc;
 988
 989         spin_lock(&par->accept_q_lock);
 990         list_del_init(&smc_sk(sk)->accept_q);
 991         spin_unlock(&par->accept_q_lock);
 992         sk_acceptq_removed(&smc_sk(sk)->listen_smc->sk);
 993         sock_put(sk); /* sock_hold in smc_accept_enqueue */
 994 }
 995
 996 /* remove a sock from the accept queue to bind it to a new socket created
 997  * for a socket accept call from user space
 998  */
 999 struct sock *smc_accept_dequeue(struct sock *parent,
1000                                 struct socket *new_sock)
1001 {
1002         struct smc_sock *isk, *n;
1003         struct sock *new_sk;
1004
1005         list_for_each_entry_safe(isk, n, &smc_sk(parent)->accept_q, accept_q) {
1006                 new_sk = (struct sock *)isk;
1007
1008                 smc_accept_unlink(new_sk);
1009                 if (new_sk->sk_state == SMC_CLOSED) {
1010                         new_sk->sk_prot->unhash(new_sk);
1011                         if (isk->clcsock) {
1012                                 sock_release(isk->clcsock);
1013                                 isk->clcsock = NULL;
1014                         }
1015                         sock_put(new_sk); /* final */
1016                         continue;
1017                 }
1018                 if (new_sock) {
1019                         sock_graft(new_sk, new_sock);
1020                         if (isk->use_fallback) {
1021                                 smc_sk(new_sk)->clcsock->file = new_sock->file;
1022                                 isk->clcsock->file->private_data = isk->clcsock;
1023                         }
1024                 }
1025                 return new_sk;
1026         }
1027         return NULL;
1028 }
1029
1030 /* clean up for a created but never accepted sock */
1031 void smc_close_non_accepted(struct sock *sk)
1032 {
1033         struct smc_sock *smc = smc_sk(sk);
1034
1035         sock_hold(sk); /* sock_put below */
1036         lock_sock(sk);
1037         if (!sk->sk_lingertime)
1038                 /* wait for peer closing */
1039                 sk->sk_lingertime = SMC_MAX_STREAM_WAIT_TIMEOUT;
1040         __smc_release(smc);
1041         release_sock(sk);
1042         sock_put(sk); /* sock_hold above */
1043         sock_put(sk); /* final sock_put */
1044 }
1045
1046 static int smcr_serv_conf_first_link(struct smc_sock *smc)
1047 {
1048         struct smc_link *link = smc->conn.lnk;
1049         struct smc_llc_qentry *qentry;
1050         int rc;
1051
1052         if (smcr_link_reg_rmb(link, smc->conn.rmb_desc))
1053                 return SMC_CLC_DECL_ERR_REGRMB;
1054
1055         /* send CONFIRM LINK request to client over the RoCE fabric */
1056         rc = smc_llc_send_confirm_link(link, SMC_LLC_REQ);
1057         if (rc < 0)
1058                 return SMC_CLC_DECL_TIMEOUT_CL;
1059
1060         /* receive CONFIRM LINK response from client over the RoCE fabric */
1061         qentry = smc_llc_wait(link->lgr, link, SMC_LLC_WAIT_TIME,
1062                               SMC_LLC_CONFIRM_LINK);
1063         if (!qentry) {
1064                 struct smc_clc_msg_decline dclc;
1065
1066                 rc = smc_clc_wait_msg(smc, &dclc, sizeof(dclc),
1067                                       SMC_CLC_DECLINE, CLC_WAIT_TIME_SHORT);
1068                 return rc == -EAGAIN ? SMC_CLC_DECL_TIMEOUT_CL : rc;
1069         }
1070         smc_llc_save_peer_uid(qentry);
1071         rc = smc_llc_eval_conf_link(qentry, SMC_LLC_RESP);
1072         smc_llc_flow_qentry_del(&link->lgr->llc_flow_lcl);
1073         if (rc)
1074                 return SMC_CLC_DECL_RMBE_EC;
1075
1076         /* confirm_rkey is implicit on 1st contact */
1077         smc->conn.rmb_desc->is_conf_rkey = true;
1078
1079         smc_llc_link_active(link);
1080         smcr_lgr_set_type(link->lgr, SMC_LGR_SINGLE);
1081
1082         /* initial contact - try to establish second link */
1083         smc_llc_srv_add_link(link);
1084         return 0;
1085 }
1086
1087 /* listen worker: finish */
1088 static void smc_listen_out(struct smc_sock *new_smc)
1089 {
1090         struct smc_sock *lsmc = new_smc->listen_smc;
1091         struct sock *newsmcsk = &new_smc->sk;
1092
1093         if (lsmc->sk.sk_state == SMC_LISTEN) {
1094                 lock_sock_nested(&lsmc->sk, SINGLE_DEPTH_NESTING);
1095                 smc_accept_enqueue(&lsmc->sk, newsmcsk);
1096                 release_sock(&lsmc->sk);
1097         } else { /* no longer listening */
1098                 smc_close_non_accepted(newsmcsk);
1099         }
1100
1101         /* Wake up accept */
1102         lsmc->sk.sk_data_ready(&lsmc->sk);
1103         sock_put(&lsmc->sk); /* sock_hold in smc_tcp_listen_work */
1104 }
1105
1106 /* listen worker: finish in state connected */
1107 static void smc_listen_out_connected(struct smc_sock *new_smc)
1108 {
1109         struct sock *newsmcsk = &new_smc->sk;
1110
1111         sk_refcnt_debug_inc(newsmcsk);
1112         if (newsmcsk->sk_state == SMC_INIT)
1113                 newsmcsk->sk_state = SMC_ACTIVE;
1114
1115         smc_listen_out(new_smc);
1116 }
1117
1118 /* listen worker: finish in error state */
1119 static void smc_listen_out_err(struct smc_sock *new_smc)
1120 {
1121         struct sock *newsmcsk = &new_smc->sk;
1122
1123         if (newsmcsk->sk_state == SMC_INIT)
1124                 sock_put(&new_smc->sk); /* passive closing */
1125         newsmcsk->sk_state = SMC_CLOSED;
1126
1127         smc_listen_out(new_smc);
1128 }
1129
1130 /* listen worker: decline and fall back if possible */
1131 static void smc_listen_decline(struct smc_sock *new_smc, int reason_code,
1132                                bool local_first)
1133 {
1134         /* RDMA setup failed, switch back to TCP */
1135         if (local_first)
1136                 smc_lgr_cleanup_early(&new_smc->conn);
1137         else
1138                 smc_conn_free(&new_smc->conn);
1139         if (reason_code < 0) { /* error, no fallback possible */
1140                 smc_listen_out_err(new_smc);
1141                 return;
1142         }
1143         smc_switch_to_fallback(new_smc);
1144         new_smc->fallback_rsn = reason_code;
1145         if (reason_code && reason_code != SMC_CLC_DECL_PEERDECL) {
1146                 if (smc_clc_send_decline(new_smc, reason_code) < 0) {
1147                         smc_listen_out_err(new_smc);
1148                         return;
1149                 }
1150         }
1151         smc_listen_out_connected(new_smc);
1152 }
1153
1154 /* listen worker: check prefixes */
1155 static int smc_listen_prfx_check(struct smc_sock *new_smc,
1156                                  struct smc_clc_msg_proposal *pclc)
1157 {
1158         struct smc_clc_msg_proposal_prefix *pclc_prfx;
1159         struct socket *newclcsock = new_smc->clcsock;
1160
1161         pclc_prfx = smc_clc_proposal_get_prefix(pclc);
1162         if (smc_clc_prfx_match(newclcsock, pclc_prfx))
1163                 return SMC_CLC_DECL_DIFFPREFIX;
1164
1165         return 0;
1166 }
1167
1168 /* listen worker: initialize connection and buffers */
1169 static int smc_listen_rdma_init(struct smc_sock *new_smc,
1170                                 struct smc_init_info *ini)
1171 {
1172         int rc;
1173
1174         /* allocate connection / link group */
1175         rc = smc_conn_create(new_smc, ini);
1176         if (rc)
1177                 return rc;
1178
1179         /* create send buffer and rmb */
1180         if (smc_buf_create(new_smc, false))
1181                 return SMC_CLC_DECL_MEM;
1182
1183         return 0;
1184 }
1185
1186 /* listen worker: initialize connection and buffers for SMC-D */
1187 static int smc_listen_ism_init(struct smc_sock *new_smc,
1188                                struct smc_clc_msg_proposal *pclc,
1189                                struct smc_init_info *ini)
1190 {
1191         struct smc_clc_msg_smcd *pclc_smcd;
1192         int rc;
1193
1194         pclc_smcd = smc_get_clc_msg_smcd(pclc);
1195         ini->ism_peer_gid = pclc_smcd->gid;
1196         rc = smc_conn_create(new_smc, ini);
1197         if (rc)
1198                 return rc;
1199
1200         /* Check if peer can be reached via ISM device */
1201         if (smc_ism_cantalk(new_smc->conn.lgr->peer_gid,
1202                             new_smc->conn.lgr->vlan_id,
1203                             new_smc->conn.lgr->smcd)) {
1204                 if (ini->first_contact_local)
1205                         smc_lgr_cleanup_early(&new_smc->conn);
1206                 else
1207                         smc_conn_free(&new_smc->conn);
1208                 return SMC_CLC_DECL_SMCDNOTALK;
1209         }
1210
1211         /* Create send and receive buffers */
1212         rc = smc_buf_create(new_smc, true);
1213         if (rc) {
1214                 if (ini->first_contact_local)
1215                         smc_lgr_cleanup_early(&new_smc->conn);
1216                 else
1217                         smc_conn_free(&new_smc->conn);
1218                 return (rc == -ENOSPC) ? SMC_CLC_DECL_MAX_DMB :
1219                                          SMC_CLC_DECL_MEM;
1220         }
1221
1222         return 0;
1223 }
1224
1225 /* listen worker: register buffers */
1226 static int smc_listen_rdma_reg(struct smc_sock *new_smc, bool local_first)
1227 {
1228         struct smc_connection *conn = &new_smc->conn;
1229
1230         if (!local_first) {
1231                 if (smcr_lgr_reg_rmbs(conn->lnk, conn->rmb_desc))
1232                         return SMC_CLC_DECL_ERR_REGRMB;
1233         }
1234         smc_rmb_sync_sg_for_device(&new_smc->conn);
1235
1236         return 0;
1237 }
1238
1239 /* listen worker: finish RDMA setup */
1240 static int smc_listen_rdma_finish(struct smc_sock *new_smc,
1241                                   struct smc_clc_msg_accept_confirm *cclc,
1242                                   bool local_first)
1243 {
1244         struct smc_link *link = new_smc->conn.lnk;
1245         int reason_code = 0;
1246
1247         if (local_first)
1248                 smc_link_save_peer_info(link, cclc);
1249
1250         if (smc_rmb_rtoken_handling(&new_smc->conn, link, cclc)) {
1251                 reason_code = SMC_CLC_DECL_ERR_RTOK;
1252                 goto decline;
1253         }
1254
1255         if (local_first) {
1256                 if (smc_ib_ready_link(link)) {
1257                         reason_code = SMC_CLC_DECL_ERR_RDYLNK;
1258                         goto decline;
1259                 }
1260                 /* QP confirmation over RoCE fabric */
1261                 smc_llc_flow_initiate(link->lgr, SMC_LLC_FLOW_ADD_LINK);
1262                 reason_code = smcr_serv_conf_first_link(new_smc);
1263                 smc_llc_flow_stop(link->lgr, &link->lgr->llc_flow_lcl);
1264                 if (reason_code)
1265                         goto decline;
1266         }
1267         return 0;
1268
1269 decline:
1270         smc_listen_decline(new_smc, reason_code, local_first);
1271         return reason_code;
1272 }
1273
1274 /* setup for RDMA connection of server */
1275 static void smc_listen_work(struct work_struct *work)
1276 {
1277         struct smc_sock *new_smc = container_of(work, struct smc_sock,
1278                                                 smc_listen_work);
1279         struct socket *newclcsock = new_smc->clcsock;
1280         struct smc_clc_msg_accept_confirm cclc;
1281         struct smc_clc_msg_proposal_area *buf;
1282         struct smc_clc_msg_proposal *pclc;
1283         struct smc_init_info ini = {0};
1284         bool ism_supported = false;
1285         int rc = 0;
1286
1287         if (new_smc->listen_smc->sk.sk_state != SMC_LISTEN)
1288                 return smc_listen_out_err(new_smc);
1289
1290         if (new_smc->use_fallback) {
1291                 smc_listen_out_connected(new_smc);
1292                 return;
1293         }
1294
1295         /* check if peer is smc capable */
1296         if (!tcp_sk(newclcsock->sk)->syn_smc) {
1297                 smc_switch_to_fallback(new_smc);
1298                 new_smc->fallback_rsn = SMC_CLC_DECL_PEERNOSMC;
1299                 smc_listen_out_connected(new_smc);
1300                 return;
1301         }
1302
1303         /* do inband token exchange -
1304          * wait for and receive SMC Proposal CLC message
1305          */
1306         buf = kzalloc(sizeof(*buf), GFP_KERNEL);
1307         if (!buf) {
1308                 rc = SMC_CLC_DECL_MEM;
1309                 goto out_decl;
1310         }
1311         pclc = (struct smc_clc_msg_proposal *)buf;
1312         rc = smc_clc_wait_msg(new_smc, pclc, sizeof(*buf),
1313                               SMC_CLC_PROPOSAL, CLC_WAIT_TIME);
1314         if (rc)
1315                 goto out_decl;
1316
1317         /* IPSec connections opt out of SMC-R optimizations */
1318         if (using_ipsec(new_smc)) {
1319                 rc = SMC_CLC_DECL_IPSEC;
1320                 goto out_decl;
1321         }
1322
1323         /* check for matching IP prefix and subnet length */
1324         rc = smc_listen_prfx_check(new_smc, pclc);
1325         if (rc)
1326                 goto out_decl;
1327
1328         /* get vlan id from IP device */
1329         if (smc_vlan_by_tcpsk(new_smc->clcsock, &ini)) {
1330                 rc = SMC_CLC_DECL_GETVLANERR;
1331                 goto out_decl;
1332         }
1333
1334         mutex_lock(&smc_server_lgr_pending);
1335         smc_close_init(new_smc);
1336         smc_rx_init(new_smc);
1337         smc_tx_init(new_smc);
1338
1339         /* check if ISM is available */
1340         if (pclc->hdr.path == SMC_TYPE_D || pclc->hdr.path == SMC_TYPE_B) {
1341                 ini.is_smcd = true; /* prepare ISM check */
1342                 rc = smc_find_ism_device(new_smc, &ini);
1343                 if (!rc)
1344                         rc = smc_listen_ism_init(new_smc, pclc, &ini);
1345                 if (!rc)
1346                         ism_supported = true;
1347                 else if (pclc->hdr.path == SMC_TYPE_D)
1348                         goto out_unlock; /* skip RDMA and decline */
1349         }
1350
1351         /* check if RDMA is available */
1352         if (!ism_supported) { /* SMC_TYPE_R or SMC_TYPE_B */
1353                 /* prepare RDMA check */
1354                 ini.is_smcd = false;
1355                 ini.ism_dev = NULL;
1356                 ini.ib_lcl = &pclc->lcl;
1357                 rc = smc_find_rdma_device(new_smc, &ini);
1358                 if (rc) {
1359                         /* no RDMA device found */
1360                         if (pclc->hdr.path == SMC_TYPE_B)
1361                                 /* neither ISM nor RDMA device found */
1362                                 rc = SMC_CLC_DECL_NOSMCDEV;
1363                         goto out_unlock;
1364                 }
1365                 rc = smc_listen_rdma_init(new_smc, &ini);
1366                 if (rc)
1367                         goto out_unlock;
1368                 rc = smc_listen_rdma_reg(new_smc, ini.first_contact_local);
1369                 if (rc)
1370                         goto out_unlock;
1371         }
1372
1373         /* send SMC Accept CLC message */
1374         rc = smc_clc_send_accept(new_smc, ini.first_contact_local);
1375         if (rc)
1376                 goto out_unlock;
1377
1378         /* SMC-D does not need this lock any more */
1379         if (ism_supported)
1380                 mutex_unlock(&smc_server_lgr_pending);
1381
1382         /* receive SMC Confirm CLC message */
1383         rc = smc_clc_wait_msg(new_smc, &cclc, sizeof(cclc),
1384                               SMC_CLC_CONFIRM, CLC_WAIT_TIME);
1385         if (rc) {
1386                 if (!ism_supported)
1387                         goto out_unlock;
1388                 goto out_decl;
1389         }
1390
1391         /* finish worker */
1392         kfree(buf);
1393         if (!ism_supported) {
1394                 rc = smc_listen_rdma_finish(new_smc, &cclc,
1395                                             ini.first_contact_local);
1396                 mutex_unlock(&smc_server_lgr_pending);
1397                 if (rc)
1398                         return;
1399         }
1400         smc_conn_save_peer_info(new_smc, &cclc);
1401         smc_listen_out_connected(new_smc);
1402         return;
1403
1404 out_unlock:
1405         mutex_unlock(&smc_server_lgr_pending);
1406 out_decl:
1407         smc_listen_decline(new_smc, rc, ini.first_contact_local);
1408         kfree(buf);
1409 }
1410
1411 static void smc_tcp_listen_work(struct work_struct *work)
1412 {
1413         struct smc_sock *lsmc = container_of(work, struct smc_sock,
1414                                              tcp_listen_work);
1415         struct sock *lsk = &lsmc->sk;
1416         struct smc_sock *new_smc;
1417         int rc = 0;
1418
1419         lock_sock(lsk);
1420         while (lsk->sk_state == SMC_LISTEN) {
1421                 rc = smc_clcsock_accept(lsmc, &new_smc);
1422                 if (rc) /* clcsock accept queue empty or error */
1423                         goto out;
1424                 if (!new_smc)
1425                         continue;
1426
1427                 new_smc->listen_smc = lsmc;
1428                 new_smc->use_fallback = lsmc->use_fallback;
1429                 new_smc->fallback_rsn = lsmc->fallback_rsn;
1430                 sock_hold(lsk); /* sock_put in smc_listen_work */
1431                 INIT_WORK(&new_smc->smc_listen_work, smc_listen_work);
1432                 smc_copy_sock_settings_to_smc(new_smc);
1433                 new_smc->sk.sk_sndbuf = lsmc->sk.sk_sndbuf;
1434                 new_smc->sk.sk_rcvbuf = lsmc->sk.sk_rcvbuf;
1435                 sock_hold(&new_smc->sk); /* sock_put in passive closing */
1436                 if (!schedule_work(&new_smc->smc_listen_work))
1437                         sock_put(&new_smc->sk);
1438         }
1439
1440 out:
1441         release_sock(lsk);
1442         sock_put(&lsmc->sk); /* sock_hold in smc_clcsock_data_ready() */
1443 }
1444
1445 static void smc_clcsock_data_ready(struct sock *listen_clcsock)
1446 {
1447         struct smc_sock *lsmc;
1448
1449         lsmc = (struct smc_sock *)
1450                ((uintptr_t)listen_clcsock->sk_user_data & ~SK_USER_DATA_NOCOPY);
1451         if (!lsmc)
1452                 return;
1453         lsmc->clcsk_data_ready(listen_clcsock);
1454         if (lsmc->sk.sk_state == SMC_LISTEN) {
1455                 sock_hold(&lsmc->sk); /* sock_put in smc_tcp_listen_work() */
1456                 if (!schedule_work(&lsmc->tcp_listen_work))
1457                         sock_put(&lsmc->sk);
1458         }
1459 }
1460
1461 static int smc_listen(struct socket *sock, int backlog)
1462 {
1463         struct sock *sk = sock->sk;
1464         struct smc_sock *smc;
1465         int rc;
1466
1467         smc = smc_sk(sk);
1468         lock_sock(sk);
1469
1470         rc = -EINVAL;
1471         if ((sk->sk_state != SMC_INIT && sk->sk_state != SMC_LISTEN) ||
1472             smc->connect_nonblock)
1473                 goto out;
1474
1475         rc = 0;
1476         if (sk->sk_state == SMC_LISTEN) {
1477                 sk->sk_max_ack_backlog = backlog;
1478                 goto out;
1479         }
1480         /* some socket options are handled in core, so we could not apply
1481          * them to the clc socket -- copy smc socket options to clc socket
1482          */
1483         smc_copy_sock_settings_to_clc(smc);
1484         if (!smc->use_fallback)
1485                 tcp_sk(smc->clcsock->sk)->syn_smc = 1;
1486
1487         /* save original sk_data_ready function and establish
1488          * smc-specific sk_data_ready function
1489          */
1490         smc->clcsk_data_ready = smc->clcsock->sk->sk_data_ready;
1491         smc->clcsock->sk->sk_data_ready = smc_clcsock_data_ready;
1492         smc->clcsock->sk->sk_user_data =
1493                 (void *)((uintptr_t)smc | SK_USER_DATA_NOCOPY);
1494         rc = kernel_listen(smc->clcsock, backlog);
1495         if (rc)
1496                 goto out;
1497         sk->sk_max_ack_backlog = backlog;
1498         sk->sk_ack_backlog = 0;
1499         sk->sk_state = SMC_LISTEN;
1500
1501 out:
1502         release_sock(sk);
1503         return rc;
1504 }
1505
1506 static int smc_accept(struct socket *sock, struct socket *new_sock,
1507                       int flags, bool kern)
1508 {
1509         struct sock *sk = sock->sk, *nsk;
1510         DECLARE_WAITQUEUE(wait, current);
1511         struct smc_sock *lsmc;
1512         long timeo;
1513         int rc = 0;
1514
1515         lsmc = smc_sk(sk);
1516         sock_hold(sk); /* sock_put below */
1517         lock_sock(sk);
1518
1519         if (lsmc->sk.sk_state != SMC_LISTEN) {
1520                 rc = -EINVAL;
1521                 release_sock(sk);
1522                 goto out;
1523         }
1524
1525         /* Wait for an incoming connection */
1526         timeo = sock_rcvtimeo(sk, flags & O_NONBLOCK);
1527         add_wait_queue_exclusive(sk_sleep(sk), &wait);
1528         while (!(nsk = smc_accept_dequeue(sk, new_sock))) {
1529                 set_current_state(TASK_INTERRUPTIBLE);
1530                 if (!timeo) {
1531                         rc = -EAGAIN;
1532                         break;
1533                 }
1534                 release_sock(sk);
1535                 timeo = schedule_timeout(timeo);
1536                 /* wakeup by sk_data_ready in smc_listen_work() */
1537                 sched_annotate_sleep();
1538                 lock_sock(sk);
1539                 if (signal_pending(current)) {
1540                         rc = sock_intr_errno(timeo);
1541                         break;
1542                 }
1543         }
1544         set_current_state(TASK_RUNNING);
1545         remove_wait_queue(sk_sleep(sk), &wait);
1546
1547         if (!rc)
1548                 rc = sock_error(nsk);
1549         release_sock(sk);
1550         if (rc)
1551                 goto out;
1552
1553         if (lsmc->sockopt_defer_accept && !(flags & O_NONBLOCK)) {
1554                 /* wait till data arrives on the socket */
1555                 timeo = msecs_to_jiffies(lsmc->sockopt_defer_accept *
1556                                                                 MSEC_PER_SEC);
1557                 if (smc_sk(nsk)->use_fallback) {
1558                         struct sock *clcsk = smc_sk(nsk)->clcsock->sk;
1559
1560                         lock_sock(clcsk);
1561                         if (skb_queue_empty(&clcsk->sk_receive_queue))
1562                                 sk_wait_data(clcsk, &timeo, NULL);
1563                         release_sock(clcsk);
1564                 } else if (!atomic_read(&smc_sk(nsk)->conn.bytes_to_rcv)) {
1565                         lock_sock(nsk);
1566                         smc_rx_wait(smc_sk(nsk), &timeo, smc_rx_data_available);
1567                         release_sock(nsk);
1568                 }
1569         }
1570
1571 out:
1572         sock_put(sk); /* sock_hold above */
1573         return rc;
1574 }
1575
1576 static int smc_getname(struct socket *sock, struct sockaddr *addr,
1577                        int peer)
1578 {
1579         struct smc_sock *smc;
1580
1581         if (peer && (sock->sk->sk_state != SMC_ACTIVE) &&
1582             (sock->sk->sk_state != SMC_APPCLOSEWAIT1))
1583                 return -ENOTCONN;
1584
1585         smc = smc_sk(sock->sk);
1586
1587         return smc->clcsock->ops->getname(smc->clcsock, addr, peer);
1588 }
1589
1590 static int smc_sendmsg(struct socket *sock, struct msghdr *msg, size_t len)
1591 {
1592         struct sock *sk = sock->sk;
1593         struct smc_sock *smc;
1594         int rc = -EPIPE;
1595
1596         smc = smc_sk(sk);
1597         lock_sock(sk);
1598         if ((sk->sk_state != SMC_ACTIVE) &&
1599             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1600             (sk->sk_state != SMC_INIT))
1601                 goto out;
1602
1603         if (msg->msg_flags & MSG_FASTOPEN) {
1604                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1605                         smc_switch_to_fallback(smc);
1606                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1607                 } else {
1608                         rc = -EINVAL;
1609                         goto out;
1610                 }
1611         }
1612
1613         if (smc->use_fallback)
1614                 rc = smc->clcsock->ops->sendmsg(smc->clcsock, msg, len);
1615         else
1616                 rc = smc_tx_sendmsg(smc, msg, len);
1617 out:
1618         release_sock(sk);
1619         return rc;
1620 }
1621
1622 static int smc_recvmsg(struct socket *sock, struct msghdr *msg, size_t len,
1623                        int flags)
1624 {
1625         struct sock *sk = sock->sk;
1626         struct smc_sock *smc;
1627         int rc = -ENOTCONN;
1628
1629         smc = smc_sk(sk);
1630         lock_sock(sk);
1631         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1632                 /* socket was connected before, no more data to read */
1633                 rc = 0;
1634                 goto out;
1635         }
1636         if ((sk->sk_state == SMC_INIT) ||
1637             (sk->sk_state == SMC_LISTEN) ||
1638             (sk->sk_state == SMC_CLOSED))
1639                 goto out;
1640
1641         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1642                 rc = 0;
1643                 goto out;
1644         }
1645
1646         if (smc->use_fallback) {
1647                 rc = smc->clcsock->ops->recvmsg(smc->clcsock, msg, len, flags);
1648         } else {
1649                 msg->msg_namelen = 0;
1650                 rc = smc_rx_recvmsg(smc, msg, NULL, len, flags);
1651         }
1652
1653 out:
1654         release_sock(sk);
1655         return rc;
1656 }
1657
1658 static __poll_t smc_accept_poll(struct sock *parent)
1659 {
1660         struct smc_sock *isk = smc_sk(parent);
1661         __poll_t mask = 0;
1662
1663         spin_lock(&isk->accept_q_lock);
1664         if (!list_empty(&isk->accept_q))
1665                 mask = EPOLLIN | EPOLLRDNORM;
1666         spin_unlock(&isk->accept_q_lock);
1667
1668         return mask;
1669 }
1670
1671 static __poll_t smc_poll(struct file *file, struct socket *sock,
1672                              poll_table *wait)
1673 {
1674         struct sock *sk = sock->sk;
1675         struct smc_sock *smc;
1676         __poll_t mask = 0;
1677
1678         if (!sk)
1679                 return EPOLLNVAL;
1680
1681         smc = smc_sk(sock->sk);
1682         if (smc->use_fallback) {
1683                 /* delegate to CLC child sock */
1684                 mask = smc->clcsock->ops->poll(file, smc->clcsock, wait);
1685                 sk->sk_err = smc->clcsock->sk->sk_err;
1686         } else {
1687                 if (sk->sk_state != SMC_CLOSED)
1688                         sock_poll_wait(file, sock, wait);
1689                 if (sk->sk_err)
1690                         mask |= EPOLLERR;
1691                 if ((sk->sk_shutdown == SHUTDOWN_MASK) ||
1692                     (sk->sk_state == SMC_CLOSED))
1693                         mask |= EPOLLHUP;
1694                 if (sk->sk_state == SMC_LISTEN) {
1695                         /* woken up by sk_data_ready in smc_listen_work() */
1696                         mask |= smc_accept_poll(sk);
1697                 } else if (smc->use_fallback) { /* as result of connect_work()*/
1698                         mask |= smc->clcsock->ops->poll(file, smc->clcsock,
1699                                                            wait);
1700                         sk->sk_err = smc->clcsock->sk->sk_err;
1701                 } else {
1702                         if ((sk->sk_state != SMC_INIT &&
1703                              atomic_read(&smc->conn.sndbuf_space)) ||
1704                             sk->sk_shutdown & SEND_SHUTDOWN) {
1705                                 mask |= EPOLLOUT | EPOLLWRNORM;
1706                         } else {
1707                                 sk_set_bit(SOCKWQ_ASYNC_NOSPACE, sk);
1708                                 set_bit(SOCK_NOSPACE, &sk->sk_socket->flags);
1709                         }
1710                         if (atomic_read(&smc->conn.bytes_to_rcv))
1711                                 mask |= EPOLLIN | EPOLLRDNORM;
1712                         if (sk->sk_shutdown & RCV_SHUTDOWN)
1713                                 mask |= EPOLLIN | EPOLLRDNORM | EPOLLRDHUP;
1714                         if (sk->sk_state == SMC_APPCLOSEWAIT1)
1715                                 mask |= EPOLLIN;
1716                         if (smc->conn.urg_state == SMC_URG_VALID)
1717                                 mask |= EPOLLPRI;
1718                 }
1719         }
1720
1721         return mask;
1722 }
1723
1724 static int smc_shutdown(struct socket *sock, int how)
1725 {
1726         struct sock *sk = sock->sk;
1727         struct smc_sock *smc;
1728         int rc = -EINVAL;
1729         int rc1 = 0;
1730
1731         smc = smc_sk(sk);
1732
1733         if ((how < SHUT_RD) || (how > SHUT_RDWR))
1734                 return rc;
1735
1736         lock_sock(sk);
1737
1738         rc = -ENOTCONN;
1739         if ((sk->sk_state != SMC_ACTIVE) &&
1740             (sk->sk_state != SMC_PEERCLOSEWAIT1) &&
1741             (sk->sk_state != SMC_PEERCLOSEWAIT2) &&
1742             (sk->sk_state != SMC_APPCLOSEWAIT1) &&
1743             (sk->sk_state != SMC_APPCLOSEWAIT2) &&
1744             (sk->sk_state != SMC_APPFINCLOSEWAIT))
1745                 goto out;
1746         if (smc->use_fallback) {
1747                 rc = kernel_sock_shutdown(smc->clcsock, how);
1748                 sk->sk_shutdown = smc->clcsock->sk->sk_shutdown;
1749                 if (sk->sk_shutdown == SHUTDOWN_MASK)
1750                         sk->sk_state = SMC_CLOSED;
1751                 goto out;
1752         }
1753         switch (how) {
1754         case SHUT_RDWR:         /* shutdown in both directions */
1755                 rc = smc_close_active(smc);
1756                 break;
1757         case SHUT_WR:
1758                 rc = smc_close_shutdown_write(smc);
1759                 break;
1760         case SHUT_RD:
1761                 rc = 0;
1762                 /* nothing more to do because peer is not involved */
1763                 break;
1764         }
1765         if (smc->clcsock)
1766                 rc1 = kernel_sock_shutdown(smc->clcsock, how);
1767         /* map sock_shutdown_cmd constants to sk_shutdown value range */
1768         sk->sk_shutdown |= how + 1;
1769
1770 out:
1771         release_sock(sk);
1772         return rc ? rc : rc1;
1773 }
1774
1775 static int smc_setsockopt(struct socket *sock, int level, int optname,
1776                           sockptr_t optval, unsigned int optlen)
1777 {
1778         struct sock *sk = sock->sk;
1779         struct smc_sock *smc;
1780         int val, rc;
1781
1782         smc = smc_sk(sk);
1783
1784         /* generic setsockopts reaching us here always apply to the
1785          * CLC socket
1786          */
1787         if (unlikely(!smc->clcsock->ops->setsockopt))
1788                 rc = -EOPNOTSUPP;
1789         else
1790                 rc = smc->clcsock->ops->setsockopt(smc->clcsock, level, optname,
1791                                                    optval, optlen);
1792         if (smc->clcsock->sk->sk_err) {
1793                 sk->sk_err = smc->clcsock->sk->sk_err;
1794                 sk->sk_error_report(sk);
1795         }
1796
1797         if (optlen < sizeof(int))
1798                 return -EINVAL;
1799         if (copy_from_sockptr(&val, optval, sizeof(int)))
1800                 return -EFAULT;
1801
1802         lock_sock(sk);
1803         if (rc || smc->use_fallback)
1804                 goto out;
1805         switch (optname) {
1806         case TCP_ULP:
1807         case TCP_FASTOPEN:
1808         case TCP_FASTOPEN_CONNECT:
1809         case TCP_FASTOPEN_KEY:
1810         case TCP_FASTOPEN_NO_COOKIE:
1811                 /* option not supported by SMC */
1812                 if (sk->sk_state == SMC_INIT && !smc->connect_nonblock) {
1813                         smc_switch_to_fallback(smc);
1814                         smc->fallback_rsn = SMC_CLC_DECL_OPTUNSUPP;
1815                 } else {
1816                         rc = -EINVAL;
1817                 }
1818                 break;
1819         case TCP_NODELAY:
1820                 if (sk->sk_state != SMC_INIT &&
1821                     sk->sk_state != SMC_LISTEN &&
1822                     sk->sk_state != SMC_CLOSED) {
1823                         if (val)
1824                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1825                                                  0);
1826                 }
1827                 break;
1828         case TCP_CORK:
1829                 if (sk->sk_state != SMC_INIT &&
1830                     sk->sk_state != SMC_LISTEN &&
1831                     sk->sk_state != SMC_CLOSED) {
1832                         if (!val)
1833                                 mod_delayed_work(system_wq, &smc->conn.tx_work,
1834                                                  0);
1835                 }
1836                 break;
1837         case TCP_DEFER_ACCEPT:
1838                 smc->sockopt_defer_accept = val;
1839                 break;
1840         default:
1841                 break;
1842         }
1843 out:
1844         release_sock(sk);
1845
1846         return rc;
1847 }
1848
1849 static int smc_getsockopt(struct socket *sock, int level, int optname,
1850                           char __user *optval, int __user *optlen)
1851 {
1852         struct smc_sock *smc;
1853
1854         smc = smc_sk(sock->sk);
1855         /* socket options apply to the CLC socket */
1856         if (unlikely(!smc->clcsock->ops->getsockopt))
1857                 return -EOPNOTSUPP;
1858         return smc->clcsock->ops->getsockopt(smc->clcsock, level, optname,
1859                                              optval, optlen);
1860 }
1861
1862 static int smc_ioctl(struct socket *sock, unsigned int cmd,
1863                      unsigned long arg)
1864 {
1865         union smc_host_cursor cons, urg;
1866         struct smc_connection *conn;
1867         struct smc_sock *smc;
1868         int answ;
1869
1870         smc = smc_sk(sock->sk);
1871         conn = &smc->conn;
1872         lock_sock(&smc->sk);
1873         if (smc->use_fallback) {
1874                 if (!smc->clcsock) {
1875                         release_sock(&smc->sk);
1876                         return -EBADF;
1877                 }
1878                 answ = smc->clcsock->ops->ioctl(smc->clcsock, cmd, arg);
1879                 release_sock(&smc->sk);
1880                 return answ;
1881         }
1882         switch (cmd) {
1883         case SIOCINQ: /* same as FIONREAD */
1884                 if (smc->sk.sk_state == SMC_LISTEN) {
1885                         release_sock(&smc->sk);
1886                         return -EINVAL;
1887                 }
1888                 if (smc->sk.sk_state == SMC_INIT ||
1889                     smc->sk.sk_state == SMC_CLOSED)
1890                         answ = 0;
1891                 else
1892                         answ = atomic_read(&smc->conn.bytes_to_rcv);
1893                 break;
1894         case SIOCOUTQ:
1895                 /* output queue size (not send + not acked) */
1896                 if (smc->sk.sk_state == SMC_LISTEN) {
1897                         release_sock(&smc->sk);
1898                         return -EINVAL;
1899                 }
1900                 if (smc->sk.sk_state == SMC_INIT ||
1901                     smc->sk.sk_state == SMC_CLOSED)
1902                         answ = 0;
1903                 else
1904                         answ = smc->conn.sndbuf_desc->len -
1905                                         atomic_read(&smc->conn.sndbuf_space);
1906                 break;
1907         case SIOCOUTQNSD:
1908                 /* output queue size (not send only) */
1909                 if (smc->sk.sk_state == SMC_LISTEN) {
1910                         release_sock(&smc->sk);
1911                         return -EINVAL;
1912                 }
1913                 if (smc->sk.sk_state == SMC_INIT ||
1914                     smc->sk.sk_state == SMC_CLOSED)
1915                         answ = 0;
1916                 else
1917                         answ = smc_tx_prepared_sends(&smc->conn);
1918                 break;
1919         case SIOCATMARK:
1920                 if (smc->sk.sk_state == SMC_LISTEN) {
1921                         release_sock(&smc->sk);
1922                         return -EINVAL;
1923                 }
1924                 if (smc->sk.sk_state == SMC_INIT ||
1925                     smc->sk.sk_state == SMC_CLOSED) {
1926                         answ = 0;
1927                 } else {
1928                         smc_curs_copy(&cons, &conn->local_tx_ctrl.cons, conn);
1929                         smc_curs_copy(&urg, &conn->urg_curs, conn);
1930                         answ = smc_curs_diff(conn->rmb_desc->len,
1931                                              &cons, &urg) == 1;
1932                 }
1933                 break;
1934         default:
1935                 release_sock(&smc->sk);
1936                 return -ENOIOCTLCMD;
1937         }
1938         release_sock(&smc->sk);
1939
1940         return put_user(answ, (int __user *)arg);
1941 }
1942
1943 static ssize_t smc_sendpage(struct socket *sock, struct page *page,
1944                             int offset, size_t size, int flags)
1945 {
1946         struct sock *sk = sock->sk;
1947         struct smc_sock *smc;
1948         int rc = -EPIPE;
1949
1950         smc = smc_sk(sk);
1951         lock_sock(sk);
1952         if (sk->sk_state != SMC_ACTIVE) {
1953                 release_sock(sk);
1954                 goto out;
1955         }
1956         release_sock(sk);
1957         if (smc->use_fallback)
1958                 rc = kernel_sendpage(smc->clcsock, page, offset,
1959                                      size, flags);
1960         else
1961                 rc = sock_no_sendpage(sock, page, offset, size, flags);
1962
1963 out:
1964         return rc;
1965 }
1966
1967 /* Map the affected portions of the rmbe into an spd, note the number of bytes
1968  * to splice in conn->splice_pending, and press 'go'. Delays consumer cursor
1969  * updates till whenever a respective page has been fully processed.
1970  * Note that subsequent recv() calls have to wait till all splice() processing
1971  * completed.
1972  */
1973 static ssize_t smc_splice_read(struct socket *sock, loff_t *ppos,
1974                                struct pipe_inode_info *pipe, size_t len,
1975                                unsigned int flags)
1976 {
1977         struct sock *sk = sock->sk;
1978         struct smc_sock *smc;
1979         int rc = -ENOTCONN;
1980
1981         smc = smc_sk(sk);
1982         lock_sock(sk);
1983         if (sk->sk_state == SMC_CLOSED && (sk->sk_shutdown & RCV_SHUTDOWN)) {
1984                 /* socket was connected before, no more data to read */
1985                 rc = 0;
1986                 goto out;
1987         }
1988         if (sk->sk_state == SMC_INIT ||
1989             sk->sk_state == SMC_LISTEN ||
1990             sk->sk_state == SMC_CLOSED)
1991                 goto out;
1992
1993         if (sk->sk_state == SMC_PEERFINCLOSEWAIT) {
1994                 rc = 0;
1995                 goto out;
1996         }
1997
1998         if (smc->use_fallback) {
1999                 rc = smc->clcsock->ops->splice_read(smc->clcsock, ppos,
2000                                                     pipe, len, flags);
2001         } else {
2002                 if (*ppos) {
2003                         rc = -ESPIPE;
2004                         goto out;
2005                 }
2006                 if (flags & SPLICE_F_NONBLOCK)
2007                         flags = MSG_DONTWAIT;
2008                 else
2009                         flags = 0;
2010                 rc = smc_rx_recvmsg(smc, NULL, pipe, len, flags);
2011         }
2012 out:
2013         release_sock(sk);
2014
2015         return rc;
2016 }
2017
2018 /* must look like tcp */
2019 static const struct proto_ops smc_sock_ops = {
2020         .family         = PF_SMC,
2021         .owner          = THIS_MODULE,
2022         .release        = smc_release,
2023         .bind           = smc_bind,
2024         .connect        = smc_connect,
2025         .socketpair     = sock_no_socketpair,
2026         .accept         = smc_accept,
2027         .getname        = smc_getname,
2028         .poll           = smc_poll,
2029         .ioctl          = smc_ioctl,
2030         .listen         = smc_listen,
2031         .shutdown       = smc_shutdown,
2032         .setsockopt     = smc_setsockopt,
2033         .getsockopt     = smc_getsockopt,
2034         .sendmsg        = smc_sendmsg,
2035         .recvmsg        = smc_recvmsg,
2036         .mmap           = sock_no_mmap,
2037         .sendpage       = smc_sendpage,
2038         .splice_read    = smc_splice_read,
2039 };
2040
2041 static int smc_create(struct net *net, struct socket *sock, int protocol,
2042                       int kern)
2043 {
2044         int family = (protocol == SMCPROTO_SMC6) ? PF_INET6 : PF_INET;
2045         struct smc_sock *smc;
2046         struct sock *sk;
2047         int rc;
2048
2049         rc = -ESOCKTNOSUPPORT;
2050         if (sock->type != SOCK_STREAM)
2051                 goto out;
2052
2053         rc = -EPROTONOSUPPORT;
2054         if (protocol != SMCPROTO_SMC && protocol != SMCPROTO_SMC6)
2055                 goto out;
2056
2057         rc = -ENOBUFS;
2058         sock->ops = &smc_sock_ops;
2059         sk = smc_sock_alloc(net, sock, protocol);
2060         if (!sk)
2061                 goto out;
2062
2063         /* create internal TCP socket for CLC handshake and fallback */
2064         smc = smc_sk(sk);
2065         smc->use_fallback = false; /* assume rdma capability first */
2066         smc->fallback_rsn = 0;
2067         rc = sock_create_kern(net, family, SOCK_STREAM, IPPROTO_TCP,
2068                               &smc->clcsock);
2069         if (rc) {
2070                 sk_common_release(sk);
2071                 goto out;
2072         }
2073         smc->sk.sk_sndbuf = max(smc->clcsock->sk->sk_sndbuf, SMC_BUF_MIN_SIZE);
2074         smc->sk.sk_rcvbuf = max(smc->clcsock->sk->sk_rcvbuf, SMC_BUF_MIN_SIZE);
2075
2076 out:
2077         return rc;
2078 }
2079
2080 static const struct net_proto_family smc_sock_family_ops = {
2081         .family = PF_SMC,
2082         .owner  = THIS_MODULE,
2083         .create = smc_create,
2084 };
2085
2086 unsigned int smc_net_id;
2087
2088 static __net_init int smc_net_init(struct net *net)
2089 {
2090         return smc_pnet_net_init(net);
2091 }
2092
2093 static void __net_exit smc_net_exit(struct net *net)
2094 {
2095         smc_pnet_net_exit(net);
2096 }
2097
2098 static struct pernet_operations smc_net_ops = {
2099         .init = smc_net_init,
2100         .exit = smc_net_exit,
2101         .id   = &smc_net_id,
2102         .size = sizeof(struct smc_net),
2103 };
2104
2105 static int __init smc_init(void)
2106 {
2107         int rc;
2108
2109         rc = register_pernet_subsys(&smc_net_ops);
2110         if (rc)
2111                 return rc;
2112
2113         rc = smc_pnet_init();
2114         if (rc)
2115                 goto out_pernet_subsys;
2116
2117         rc = smc_core_init();
2118         if (rc) {
2119                 pr_err("%s: smc_core_init fails with %d\n", __func__, rc);
2120                 goto out_pnet;
2121         }
2122
2123         rc = smc_llc_init();
2124         if (rc) {
2125                 pr_err("%s: smc_llc_init fails with %d\n", __func__, rc);
2126                 goto out_core;
2127         }
2128
2129         rc = smc_cdc_init();
2130         if (rc) {
2131                 pr_err("%s: smc_cdc_init fails with %d\n", __func__, rc);
2132                 goto out_core;
2133         }
2134
2135         rc = proto_register(&smc_proto, 1);
2136         if (rc) {
2137                 pr_err("%s: proto_register(v4) fails with %d\n", __func__, rc);
2138                 goto out_core;
2139         }
2140
2141         rc = proto_register(&smc_proto6, 1);
2142         if (rc) {
2143                 pr_err("%s: proto_register(v6) fails with %d\n", __func__, rc);
2144                 goto out_proto;
2145         }
2146
2147         rc = sock_register(&smc_sock_family_ops);
2148         if (rc) {
2149                 pr_err("%s: sock_register fails with %d\n", __func__, rc);
2150                 goto out_proto6;
2151         }
2152         INIT_HLIST_HEAD(&smc_v4_hashinfo.ht);
2153         INIT_HLIST_HEAD(&smc_v6_hashinfo.ht);
2154
2155         rc = smc_ib_register_client();
2156         if (rc) {
2157                 pr_err("%s: ib_register fails with %d\n", __func__, rc);
2158                 goto out_sock;
2159         }
2160
2161         static_branch_enable(&tcp_have_smc);
2162         return 0;
2163
2164 out_sock:
2165         sock_unregister(PF_SMC);
2166 out_proto6:
2167         proto_unregister(&smc_proto6);
2168 out_proto:
2169         proto_unregister(&smc_proto);
2170 out_core:
2171         smc_core_exit();
2172 out_pnet:
2173         smc_pnet_exit();
2174 out_pernet_subsys:
2175         unregister_pernet_subsys(&smc_net_ops);
2176
2177         return rc;
2178 }
2179
2180 static void __exit smc_exit(void)
2181 {
2182         static_branch_disable(&tcp_have_smc);
2183         sock_unregister(PF_SMC);
2184         smc_core_exit();
2185         smc_ib_unregister_client();
2186         proto_unregister(&smc_proto6);
2187         proto_unregister(&smc_proto);
2188         smc_pnet_exit();
2189         unregister_pernet_subsys(&smc_net_ops);
2190         rcu_barrier();
2191 }
2192
2193 module_init(smc_init);
2194 module_exit(smc_exit);
2195
2196 MODULE_AUTHOR("Ursula Braun <ubraun@linux.vnet.ibm.com>");
2197 MODULE_DESCRIPTION("smc socket address family");
2198 MODULE_LICENSE("GPL");
2199 MODULE_ALIAS_NETPROTO(PF_SMC);